Professional Documents
Culture Documents
Lecture Notes
Lecture Notes
Anthony Horsley
Andrew J. Wrobel
The Short-Run
Approach to Long-
Run Equilibrium
in Competitive
Markets
A General Theory with Application to
Peak-Load Pricing with Storage
Lecture Notes in Economics
and Mathematical Systems 684
Founding Editors:
M. Beckmann
H.P. Künzi
Managing Editors:
Prof. Dr. G. Fandel
Fachbereich Wirtschaftswissenschaften
Fernuniversität Hagen
Hagen, Germany
and
Editorial Board:
H. Dawid, D. Dimitrov, A. Gerber, C.-J. Haake, C. Hofmann, T. Pfeiffer,
R. Slowiński, W.H.M. Zijm
More information about this series at http://www.springer.com/series/300
Anthony Horsley • Andrew J. Wrobel
123
Anthony Horsley (1939-2006) Andrew J. Wrobel
Watford, Hertfordshire, UK Warsaw, Poland
Completed in August 2015, this book is a revised and restructured version of the STICERD Discussion
Paper TE/05/490 “Characterizations of long-run producer optima and the short-run approach to long-run
market equilibrium: a general theory with applications to peak-load pricing” © Anthony Horsley and
Andrew J. Wrobel (London, LSE, 2005).
v
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
2 Peak-Load Pricing with Cross-Price Independent
Demands: A Simple Illustration .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 15
2.1 Short-Run Approach to Simplest Peak-Load Pricing Problem .. . . . . 15
2.2 Reinterpreting Cost Recovery as a Valuation Condition .. . . . . . . . . . . . 17
2.3 Equilibrium Prices for the Single-Consumer Case . . . . . . . . . . . . . . . . . . . 18
3 Characterizations of Long-Run Producer Optimum . . . . . . . . . . . . . . . . . . . . 21
3.1 Cost and Profit as Values of Programmes with Quantity
Decisions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 21
3.2 Split SRP Optimization: A Primal-Dual System
for the Short-Run Approach.. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 25
3.3 Duality: Cost and Profit as Values of Programmes
with Shadow-Price Decisions . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 26
3.4 SRP and SRC Optimization Systems . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 38
3.5 SRC/P Partial Differential System for the Short-Run Approach . . . . 40
3.6 Other Differential Systems . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 42
3.7 Transformations of Differential Systems by Using SSL or PIR . . . . . 43
3.8 Summary of Systems Characterizing Long-Run
Producer Optimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
3.9 Extended Wong-Viner Theorem and Other
Transcriptions from SRP to LRC . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 47
3.10 Derivation of Dual Programmes . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 52
3.11 Shephard-Hotelling Lemmas and Their Dual Counterparts . . . . . . . . . 53
3.12 Duality for Linear Programmes with Nonstandard
Parameters in Constraints . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 62
4 Short-Run Profit Approach to Long-Run Market Equilibrium . . . . . . . . 73
4.1 Outline of the Short-Run Approach.. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
4.2 Detailed Framework for Short-Run Profit Approach . . . . . . . . . . . . . . . . 80
vii
viii Contents
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 193
List of Figures
ix
x List of Figures
Fig. 5.1 Trajectories of: (a) shadow price of stock O , and (b)
output of pumped-storage plant (optimum storage
policy) yO PS in Sect. 5.2, and in
Theorem
5.3.1.
0 Unitrent
00
C O O
for storage capacity is Varc D d C dO ,
the sum of rises of ˇ O . Unit rent for conversion capacity
R T ˇˇ ˇ
is 0 ˇp .t/ O .t/ˇ dt, the sum of grey areas. By
This is a new formal framework for the theory of perfectly competitive equilibrium
and its industrial applications. The “short-run approach” is a scheme for calculating
long-run producer optimum and market equilibrium by building on short-run
solutions to the producer’s profit maximization problem, in which capital inputs
and natural resources are treated as fixed. These fixed inputs are valued at their
marginal contributions to the operating profit and, where possible, their levels are
then adjusted accordingly.1 Since short-run profit is a concave but generally non-
differentiable function of the fixed inputs, their marginal values are defined as the
generally nonunique supergradient vectors. Also, they usually have to be obtained
as solutions to the dual programme of fixed-input valuation because there is rarely
an explicit formula for the operating profit. The key property of the dual solution
is therefore its marginal interpretation, but this requires the use of a generalized,
multi-valued derivative of a convex function—viz., the subdifferential—because an
optimal-value function, such as cost or profit, is commonly nondifferentiable.
Despite being essential for applications, differential calculus has been purged
from geometric and topological treatments of the Arrow-Debreu model, which
are limited to equilibrium existence and Pareto optimality results. But the use of
subgradients restores calculus as a rigorous method for equilibrium theory. The
mathematical tools employed here—convex programmes and subdifferentials—
make it possible to reformulate some basic microeconomic results. In addition to
statements of known subdifferential versions of the Shephard-Hotelling Lemmas,
a subdifferential version of the Wong-Viner Envelope Theorem is devised here for
the short-run approach especially (Sect. 3.9). This facilitates economic analysis and
1
When carried out by iterations, the calculations might also be seen as modelling the real processes
of price and quantity adjustments.
2
The theory of differentiable convex functions is, of course, included in subdifferential calculus as
a special case. Furthermore, the subgradient concept can also be used to prove, rather than assume,
that a convex function is differentiable—by showing that it has a unique subgradient. This method
is used in [21, 23], [27, Section 9] and [30, Section 9].
3
From Sect. 3.2 on, short-run cost minimization is split off as a subprogramme, whose solution is
denoted by vL .y; k; w/. In these terms, vO . p; k; w/ D vL .Oy . p; k; w/ ; k; w/.
1 Introduction 3
as a practical method for calculating market equilibria. For this, with the input
prices r and w taken as fixed for simplicity, the short-run profit-maximizing supply
yO . p; k; w/ is equated to demand for the products xO . p/ to determine the short-
run equilibrium output prices p?SR .k; w/. The imputed capacity values rO . p; k; w/,
evaluated at p D p?SR .k; w/ together with the given k and w, are only then equated
to the given capacity prices r to determine the long-run equilibrium
capacities
k? .r; w/—by solving for k the equation rO p?SR .k; w/ ; k; w D r. Finally, the
long-run equilibrium output prices and quantities are determined by substituting
k? .r; w/ for k in the short-run equilibrium solution.4 In other words, determination
of investment is postponed until after the equilibrium in the product markets has
been found: the producer’s long-run problem is split into two problems—that of
operation and that of investment—and the short-run market equilibrium problem is
“inserted” in between. Since the operating solutions usually have relatively simple
forms, doing things in this order can greatly ease the fixed-point problem of solving
for equilibrium: indeed, the problem can even become elementary when approached
in this way (Chap. 2). Furthermore, unlike the optimal investment of the pure
producer problem, the equilibrium investment k? has a definite scale
(determined
by demand for the products). Put another way: rO p?SR .k; w/ ; k; w , the value to
be equated to r, is not homogeneous of degree zero in k like rO . p; k; w/. Thus
one can keep mostly to single-valued maps and avoid dealing with multi-valued
correspondences—even when the returns to scale are constant. Last but not least,
like the short-run producer optimum, the short-run general equilibrium is of interest
in itself.
This exposition comes in six chapters (not counting the Introduction, Conclu-
sions, or Appendices), which can be divided into three parts. The first and main
part (Chaps. 2–5) gives various characterizations of long-run producer optimum
(Chap. 3), but its final objective is a framework for the short-run approach to
long-run general-equilibrium pricing of a range of commodities with joint costs
of production (Chap. 4), which is applied to peak-load pricing of electricity
generated by a variety of techniques (Chap. 5). A much simplified version of
the electricity pricing problem serves also as an introductory example (Chap. 2).
The characterizations of producer optimum (which are needed for the short-run
approach) are complemented by conditions for existence of the optimal quantities
and shadow prices in the short-run profit maximization and cost minimization
problems, and for equality of the total values of the variable quantities and of the
fixed quantities—i.e., for absence of a gap between the primal and dual solutions.
These results form the second part (Chap. 6). The third and last part (Chap. 7)
introduces the concept of technologies with conditionally fixed coefficients, and
specializes the preceding general analysis to this class (which includes, e.g.,
4
A short-run approach to equilibrium might also be based on short-run cost minimization, in which
not only the capital inputs (k) but also the outputs (y) are kept fixed and are shadow-priced in the
dual problem, but such cost-based calculations are usually much more complicated than those
using profit maximization: see Sect. 4.1.
4 1 Introduction
thermal generation of electricity and pumped energy storage, but not hydroelectric
generation). Appendix A gives a contextual example of a duality gap—a possible but
rather exceptional mathematical complication in convex programming. Sections B.1
to B.7 of Appendix B give the required standard results of convex calculus—
including one innovation, viz., Lemma B.7.2 on subdifferential sections (the SSL),
which underlies the Extended Wong-Viner Envelope Theorem (3.9.1). The typical
mathematical obstacle that necessitates the extension—viz., nonfactorability of joint
subdifferentials for nondifferentiable bivariate convex functions—is looked at in
more detail in Sect. B.8.
First of all, for a simple but instructive introduction to the short-run approach to
long-run equilibrium, Boiteux’s treatment of the simplest peak-load pricing problem
is rehearsed: this is the problem of pricing the services of a homogeneous capacity
that produces a nonstorable good with cyclic demands (such as electricity). A
direct calculation of long-run equilibrium poses a fixed-point problem, but, with
cross-price independent demands, short-run equilibrium can be determined by the
elementary method of intersecting the supply and demand curves for each time
instant separately. At each time t, the short-run equilibrium output price p?SR .t/ is the
sum of the unit operating cost w and a capacity charge ?SR .t/ 0 that is nonzero
only at the times of full capacity utilization, i.e., when the output rate y?SR .t/ equals
the given capacity k. Finally, long-run equilibrium is found by adjusting the capacity
k so that its unit cost r equals its unit value defined as the unit operating profit, which
RT RT
equals the total capacity charge over the cycle, 0 ?SR .t/ dt D 0 p?SR .t/ w dt.
This solution is given by Boiteux with discretized time [9, 3.2–3.3].5 Its continuous-
time version is given in Chap. 2.
Boiteux’s idea is developed here into a frame for analysis of investment and
pricing by an industry that supplies a range of commodities—such as a good
differentiated over time, locations or events (Chap. 4). In Chap. 5, this is applied to
augment the rudimentary one-station model to a continuous-time equilibrium model
of electricity pricing with a diverse technology, including energy storage and hydro
as well as thermal generation. Such a plant mix makes supply cross-price dependent,
even in the short run (i.e., with the capacities fixed). Demand, too, is allowed to be
cross-price dependent.
The setting up of the short-run approach to pricing and investment (Chap. 4) is
the most novel part of this work. Unlike the characterizations of producer optimum,
and the existence results on it, this part of the study is not fully formalized into
mathematical theorems: it is assumed, rather than proved, that the short-run equilib-
rium is indeed unique, and as for its existence it is merely noted that this cannot be
5
Boiteux’s work is also presented by Drèze [15, pp. 10–16], but the short-run character of the
approach is more evident from the original [9, 3.2–3.3] because Boiteux discusses the short-run
equilibrium first, before using it as part of the long-run equilibrium system. When Drèze mentions
short-run equilibrium on its own, it is only as an afterthought [15, p. 16].
1 Introduction 5
guaranteed unless the fixed capacities are all positive (i.e., unless k 0).6 Also the
question of methods for computing short-run market equilibria is only touched upon,
in Fig. 4.1, where the use of Walrasian tâtonnement is suggested.7 And no qualitative
properties
are established of the valuation condition for long-run equilibrium—
that rO p?SR .k; w/ ; k; w D r—as an equation for the investment k.8 But it is
shown that the SRP Programme-Based System, consisting of Conditions (4.2.12)–
(4.2.16) together with (4.2.19)–(4.2.20), is a full characterization of long-run market
equilibrium. And, as is seen already from the introductory example of Chap. 2,
the short-run approach can greatly simplify the problem of solving for long-run
equilibrium (as well as finding the short-run equilibrium on the way). It seems
clear that the approach is worth applying not only to the case of electricity but
also to the supply of other time-differentiated commodities (such as water, natural
gas, telecommunications, and so on). The questions of uniqueness, stability and
iterative computation of equilibria, although important, are not specific to the short-
run approach; also, they have been much studied and are well understood (at least
for finite-dimensional commodity spaces). The central and distinctive quantitative
elements of the approach are valuation and operation of plants; these problems have
been fully solved for the various types of plant in the ESI (see Sect. 5.2 and its
references). The priorities in developing the short-run approach are: (i) to analyze
the valuation and operation problems for other technologies and industries, and
(ii) to compute numerical solutions from real data by using, at least to start with, the
standard methods (viz., linear programming for producer optima and tâtonnement
for market equilibria). It would seem sensible to address the theoretical questions of
equilibrium uniqueness and stability in the light of future computational experience
(in which more elaborate iterative methods could be employed if necessary). These
questions are potentially important for practice as well as for completing the theory,
but they are not priorities for this study, and are left for further research.
The bulk of Chap. 3, between the introductory example and the setup for the
short-run approach, gives various characterizations of long-run producer optimum
(Sects. 3.1 to 3.11). Each of these is either an optimization system or a differential
system , i.e., it is a set of conditions formulated in terms of either the marginal
optimal values or the optimal solutions to a primal-dual pair of programmes
(although the two kinds of condition can also be mixed in one system). Though
equivalent, the various systems are not equally usable, and the best choice of system
depends on one’s purpose as well as on the available mathematical description
of the technology. In the application to electricity pricing with non-thermal as
6
This is not an unacceptable condition, but some capacities can of course be zero in long-run
equilibrium. The long-run model meets the usual adequacy assumption, as does the short-run
model with positive capacities, and so existence of an equilibrium follows from Bewley’s result
[7, Theorem 1], which is amplified in [31, Section 3] and [29] by a proof based on continuity of
demand in prices.
7
As is well known, this process does not always converge, but there are other iterative methods.
8
In general, this is an inclusion rather than an equality: see (4.2.19).
6 1 Introduction
9
By contrast, SRC minimization for a system of plants can be difficult because it involves
allocating the system’s given output among the plants. Its complexity shows in, e.g., the case
of a hydro-thermal electricity-generating system studied by Koopmans [35]. The decentralized
approach taken here (Chaps. 4 and 5 with their references) avoids having to deal directly
with the formidable problem of minimizing the entire system’s cost: see the Comments with
Formulae (4.1.3) and (4.1.4).
10
This shows how mistaken is the widespread but unexamined view that nondifferentiabilities of
convex functions are of little consequence: the very points which, in a sense, are exceptional a priori
turn out to be the rule rather than the exception in equilibrium. Also, it is only on finite-dimensional
spaces that convex functions are “generically smooth” or, more precisely, twice differentiable
almost everywhere with respect to the Lebesgue measure (Alexandroff’s Theorem). On an infinite-
dimensional space, a convex function can be nondifferentiable everywhere.
1 Introduction 7
suppressed from the notation). For differentiable costs, the inclusion reduces to the
Wong-Viner equality of gradient vectors: r y CLR D r y CSR (when the capital inputs
are at an optimum). But for nondifferentiable costs, all it shows is that each LRMC
is an SRMC—which is the reverse of what is required for the short-run approach.
The way out of this difficulty is to bring in the SRP function, …SR , and require that
the given prices for the capital inputs are equal to their profit-imputed values, i.e.,
that r D r k …SR . p; k/ or, should the gradient not exist, that r 2 @O k …SR (which
is the superdifferential a.k.a. the supergradient set). This condition is stronger than
cost-optimality of the fixed inputs, when the output price system p is an SRMC:
i.e., if p 2 @y CSR .y; k/ then @O k …SR . p; k/ @k CSR .y; k/, and the inclusion is
generally strict (indeed, r k …SR can exist also when r k CSR does not, in which
case r k …SR 2 @k CSR ). But the new condition—that r 2 @O k …SR . p; k/—is no
stronger than it need be: it is just strong enough to do the job and guarantee that if
p 2 @y CSR .y; k/ then p 2 @y CLR .y; r/.
The present analysis of the relationship between SRMC and LRMC bears out,
amplifies and develops Boiteux’s ideas—which, at the time, he allowed, with a
hint of exasperation, were “false in the theoretical general case, but more or less
true of ordinary industrial plant”. Both cases are thus accommodated: the industrial
reality of fixed coefficients and rigid capacities as well as the rather unrealistic
textbook supposition of smooth costs. The gap is bridged between the inadequate
existing theory and its intended applications, and an end is put to its disturbing and
unnecessary divorce from reality. This allows peak-load pricing to be put, for the
first time, on a sound and rigorous theoretical basis (Chap. 5).
R From the new perspective, Boiteux’s condition for long-run optimality (r D
. p .t/ w/ dt) should be viewed as a special case, for the one-station technology,
of the equation r D r k …SR . But staying within the confines of this particular
example, Boiteux interprets his condition merely as recovery of the total production
cost, including the capital cost [9, 3.4.2: (2) and Conclusions: 4]. This is correct,
but only in the case of a single capital input, and it cannot provide a basis for
dealing with a production technique that uses a number of interdependent capital
inputs.11 In such a case, the present generalization of Boiteux’s condition for long-
run optimality is stronger than capital-cost recovery: under constant returns to scale,
if r 2 @O k …SR (or r D r k …SR ) then r k D …SR , but not vice versa if there are two
or more capital inputs (though also the converse is of course true when, with just
one capital input, k is a nonzero scalar). It is a dead end to think purely in terms of
11
Capital inputs are called independent if the SRP function (…SR ) is linear in the capital-
input bundle k D .k1 ; k2 ; : : :/; an example is the multi-station technology of thermal electricity
generation. Such a technology in effect separates into a number of production techniques with
a single capital input each, and so Boiteux’s analysis applies readily: to ensure that a short-run
equilibrium is a long-run equilibrium, it suffices to require cost recovery for each production
technique with k > 0, although one must also remember to check that any unutilized
production
R technique (one with k D 0) is unprofitable at the equilibrium prices (e.g., that
r . p .t/ w / dt for any unbuilt type of thermal station, with unit capital cost r and
unit fuel cost w ).
1 Introduction 9
marginal costs and cost recovery: with multiple capital inputs, cost recovery is not
sufficient to guarantee that a short-run equilibrium is also a long-run equilibrium or,
equivalently, that an SRMC tariff is also an LRMC tariff. The SRP function with its
marginals (derivatives w.r.t. k), or the SRP programme with the dual solution, must
be brought into the short-run approach. This is done here for the first time.
In mathematical terms, the Extended Wong-Viner Theorem (3.9.1) comes from
what is named the Subdifferential Sections Lemma (SSL), which gives the joint
subdifferential of a bivariate convex function, @y;k C, in terms of one of its partial
subdifferentials, @y C, and of a partial superdifferential, @O k … . p; k/, of the relevant
partial conjugate of C (denoted by …, it is a saddle function)—see (3.7.3), and
Lemma B.7.2 in Appendix B. For the extension (3.9.1), the SSL is applied twice:
to either the SRP or the LRC as a saddle function obtained by partial conjugacy
from the SRC, which is a jointly convex function (C) of the output bundle y and
the fixed-input bundle k, with the variable-input prices w kept fixed (Sect. 3.9). In
the wider context of convex calculus and its applications, the SSL can be usefully
regarded as a direct precursor of the Partial Inversion Rule (PIR), a well-known
result that relates the partial sub/super-differentials of a saddle function (@p … and
@O k …) to the joint subdifferential of its bivariate convex “parent” function (@y;k C):
see Lemmas B.7.3 and B.7.5 (whose proofs do derive the PIR from the SSL).
One well-known application of this fundamental principle is the equivalence of
two conditions for optimality, viz., the parametric version of Fermat’s rule and
the Kuhn-Tucker characterization of primal and dual optima as a saddle point of
the Lagrange function: see, e.g., [45, 11.39 (d) and 11.50]. Another well-known
use of the PIR establishes the equivalence of Hamiltonian and Lagrangian systems
in convex variational calculus; when the Lagrange integrand is nondifferentiable,
this usefully splits the Euler-Lagrange differential inclusion (a generalized equation
system) into the pair of Hamiltonian differential inclusions, and it may even
transform the inclusion into ordinary equations because the Hamiltonian can be
differentiable also when the Lagrangian is not: see, e.g., [44, (10.38) and (10.40)],
[43, Theorem 6] or [4, 4.8.2].12 The present application of the PIR or the SSL
relates the marginal optimal values for a programme to those of a subprogramme,
to put it in general terms. In the specific context of extending the Wong-Viner
Theorem, SRC minimization is a subprogramme both of SRP maximization and of
LRC minimization (their optimal values are CSR .y; k/, …SR . p; k/ and CLR .y; r/,
respectively). This is a new use of what is, in Rockafellar’s words, “a striking
relationship: : :at the heart of programming theory” [41, p. 604].
One half of this argument—the application of the SSL to the saddle function
…SR as a partial conjugate of the bivariate convex function CSR to prove the first
equivalence in (3.9.1)—is given already in Sect. 3.7. It comes along with other
applications of the PIR and the SSL that establish the equivalence of the partial
12
To distinguish the two quite different meanings of the word “Lagrangian”, it shall be occasionally
expanded into either “Lagrange function” (in the multiplier method of optimization) or “Lagrange
integrand” (in the calculus of variations only).
10 1 Introduction
differential systems to the saddle differential systems of Sect. 3.6 (which use joint
subdifferentials).
Like all optimization, economic theory has to deal with the nondifferentiability
of optimal values that commonly arises even when the programmes’ objective and
constraint functions are all smooth. This has led to the eschewing of marginal
concepts in rigorous equilibrium analysis, but any need for this disappeared with
the advent of nonsmooth calculus. Of course, in using generalized derivatives
such as the subdifferential, one cannot expect to transcribe familiar theorems from
the smooth to the subdifferentiable case simply by replacing the ordinary single
gradients with multi-valued subdifferentials—proper subdifferential calculus must
be applied. This not only extends the scope for marginal analysis, but also leads to
a rethinking and reinterpretation that can give a new economic content to known
results. The Wong-Viner Theorem is a case in point: a useful extension depends on
recasting its fixed-input optimality assumption in terms of profit-based valuations
(i.e., on restating optimality of the fixed inputs as equality of their rental prices
to their marginal contributions to the operating profit). After this reformulation of
optimality in terms of marginal SRP—but not before—the “smooth” version of the
theorem can be transcribed to the case of subdifferentiable costs (by replacing each
r with a @). Without this preparatory step, all extension attempts are doomed: a
direct transcription of the original Wong-Viner equality of SRMC and LRMC to
the subdifferentiable case is plainly false, and although it can be changed to a true
inclusion without bringing in the SRP function, that kind of result fails to attain the
goal of identifying an SRMC as an LRMC.13
One well-known condition for optimality is, perhaps, conspicuous by its near-
absence from the main part of this analysis. The Lagrangian Saddle-Point Condi-
tions of Kuhn and Tucker are central to the duality theory of convex programmes
(CPs)—and they are used in the studies of hydro and energy storage [21, 23, 27] and
[30], which feed the application of the short-run framework to electricity supply in
Sects. 5.1 to 5.3—but here the Kuhn-Tucker Conditions are not used before the
study of technologies with conditionally fixed coefficients (in Chap. 7), although
they do appear earlier on the margin (in Comments in Sects. 3.3 and 4.1). Instead
of the Kuhn-Tucker Conditions, for a general analysis with an abstract production
cone it is preferred here to use the Complementarity Conditions (3.1.5) on the
price system and the input-output bundle. This system is a case of what will be
called the FFE Conditions, which consist of primal feasibility, dual feasibility and
equality of the primal and dual objectives (at the feasible points in question). The
FFE Conditions form an effective system whenever the dual programme can be
worked out from the primal explicitly. This is always so, in principle at least, with
the profit and cost problems because they become linear programmes (LPs) once the
13
Without involving …SR , the inclusion (@y CLR @y CSR ) can be improved only by making it
more precise but no more useful: @y CSR .y; k/ can be shown to equal the union of @y CLR .y; r/ over
r 2 @k CSR .y; k/, i.e., over all those fixed-input price systems r for which k is an optimal fixed-
input bundle for the output bundle y (given also the suppressed variable-input price system w):
see (3.9.11).
1 Introduction 11
production cone has been represented by linear inequalities. For an LP, the system
of FFE Conditions is linear in the primal and dual variables jointly—unlike the
system of Kuhn-Tucker Conditions (which is nonlinear because of the quadratic
term in the Complementary Slackness Condition): compare (3.3.3) with (3.3.2). And
a linear system (i.e., a system of linear equalities and inequalities) is much simpler
to deal with; in particular, it can be solved numerically by the simplex method (or
another LP algorithm). The problem’s size is smaller, though, when the method
is applied directly to the relevant LP (or to its dual), rather than to its system of
FFE Conditions.14 Either way, there is no need for the Kuhn-Tucker Conditions in
solving the SRP programmes with their fixed-input valuation duals—although they
are instrumental in proving uniqueness of solutions, as in [21, 23, 27] and [30].
In the LP formulation of a profit or cost programme, the fixed quantities are
primal parameters but need not be the same as the standard “right-hand side”
parameters—and so their shadow prices, which are the dual variables, need not be
identical to the standard dual variables. Yet the usual theory of linear programming
works with the standard parameterization, and it is the standard dual solution
that the simplex method provides along with the primal solution. But, as is
shown in Sect. 3.12, this is not much of a complication because any other dual
variables can be expressed in terms of the standard dual variables a.k.a. the usual
Lagrange multipliers of the constraints. This is used in valuing the fixed inputs for
electricity generation (Sect. 5.2). The principle has also a counterpart beyond the
linear or convex duality framework: it is the Generalized Envelope Theorem for
smooth optimization, whereby the marginal values of all parameters, including any
nonstandard ones, are equal to the corresponding partial derivatives of the ordinary
Lagrangian—and are thus expressed in terms of the constraints’ multipliers. See [1,
(10.8)] or [47, 1.F.b].
The exposition of producer optimum pauses for “stock-taking” in Sect. 3.8. In
particular, Tables 3.1 and 3.2 summarize the various characterizations of long-run
optimum, though not their “mirror images” which result from a formal substitution
of the LRC for the SRP. These tables record also the methods employed to transform
these systems into one another. This shows a unity: the same methods are applied
to systems of the same type, even though this exposition gives special places to
the two systems of most importance for the application of the short-run approach
to the ESI (in Chap. 5)—viz., the Split SRP Optimization System of Sect. 3.2
and the SRC/P Partial Differential System of Sect. 3.5. The latter system’s “mirror
image”, the L/SRC Partial Differential System of Sect. 3.9, is also directly involved
in applications when its conditions of LRMC pricing and LRC minimization serve
as the definition of long-run optimum—as is often the case in public utility pricing,
including Boiteux’s work and the account thereof in Chap. 2. The other fourteen
systems are not applied here, but any of them may be the best tool (for the short-run
approach as for other purposes) when the technology is described most simply in
14
For a count of variables and constraints, see the last Comment in Sect. 3.12 before For-
mula (3.12.15).
12 1 Introduction
the system’s own terms; see also the Comments at the end of Sect. 4.1. In particular,
one should not be trapped by the language into thinking that a system using the
LRC programme or function is somehow inherently unsuitable for the short-run
approach.
The summarizing Sect. 3.8 ends by noting that some of the systems—including
the two “special” ones—can be partitioned into a short-run subsystem (which
characterizes SRP maxima) and a valuation condition that generalizes Boiteux’s
condition for long-run optimality and requires that investment be at a profit
maximum.
A complete formalization of all the duality-based systems is deferred to
Sects. 3.10 and 3.11, in which the programmes’ duality and the systems’ equivalence
are cast as rigorous results with proofs. To this end, Sect. 3.11 restates formally
the subdifferential versions of the Shephard-Hotelling Lemmas (some of which are
announced earlier in Sect. 3.6). As has long been known [14, pp. 555 and 583],
these are cases of the derivative property of optimal value, which transcribes to the
subdifferentiable case directly (by replacing r with @).
The characterizations of long-run producer optimum are complemented by
results on solubility of the primal and dual programmes and on equality of their
values (absence of a duality gap). Such an analysis is given in Sects. 6.1 to 6.4; it
yields sufficient conditions for existence of a pair of solutions with equal values.
First, it is recalled from the general theory of CPs that absence of a duality gap is
equivalent to semicontinuity of either optimal value, and this is spelt out for the
profit and cost programmes (Sect. 6.1). To make this criterion applicable, Sect. 6.2
gives some sufficient conditions for the required semicontinuity of SRP as well
as of LRC and SRC, as functions of the programmes’ quantity data (the fixed
quantities). When the commodity space for either the fixed or the variable quantities
(the programme’s quantity data or its decision variables) is infinite-dimensional,
these criteria use its weak* topology as well as its vector order. The commodity
spaces are therefore taken to be dual Banach lattices (i.e., the duals of completely
normed vector lattices). One example is L1 Œ0; T, which serves here as the output
space in the application to peak-load pricing. With this or any other nonreflexive
commodity space, these semicontinuity criteria for profit or cost (as a function
of the fixed quantities) apply only when the given price system (for the variable
quantities) lies not just in the Banach dual of the commodity space but actually
in the smaller predual space. Such a criterion is therefore adequate for the short-
run approach to general equilibrium (and other analysis thereof) only when the
equilibrium price system is known to lie in the predual space—as is the case for the
commodity space L1 and its predual L1 under Bewley’s assumptions [7], which are
weakened in [26] to make his density representation of the price system apply to at
least some continuous-time problems. Unavoidably, even the weakened assumption
is restrictive: it requires that brief interruptions of a consumption or input flow cause
only small losses of utility or output (i.e., interruptibility of consumer demand and
of input demand). When this is not so and the programme’s given price system
cannot be taken to lie in L1 Œ0; T—or in whatever price space is the predual of
the commodity space in some other economic context—a duality gap can still be
1 Introduction 13
15
See [21] and [23] for examples of an SRP programme in which the output space is L1 Œ0; T and
a “singular” price term places the price system outside the predual L1 Œ0; T, but it is the timing of
the singularity, and not just its presence, that determines whether the programme is soluble or not.
Chapter 2
Peak-Load Pricing with Cross-Price
Independent Demands: A Simple Illustration
The short-run approach to solving for long-run market equilibrium is next illustrated
with the example of pricing, over the demand cycle, the services of a homogeneous
productive capacity with a unit capital cost r and a unit running cost w. The
technology can be interpreted as, e.g., electricity generation from a single type of
thermal station with a fuel cost w (in $/kWh) and a capacity cost r (in $/kW) per
period. The cycle is represented by a continuous time interval Œ0; T. Demand for the
time-differentiated, nonstorable product, Dt .p/, is assumed to depend only on the
time t and on the current price p (a scalar). As a result, the short-run equilibrium can
be found separately at each instant t, by intersecting the demand and supply curves
in the price-quantity plane. This is because, with this technology, short-run supply
is cross-price independent: given a capacity k, the supply is
8
<0 for p < w
S .p; k; w/ D Œ0; k for p D w (2.1.1)
:
k for p > w
where p is the current price. That is, given a time-of-use (TOU) tariff p—i.e., given
a price p .t/ as a function of time t 2 Œ0; T—the set of all the profit-maximizing
output trajectories, YO . p; k; w/, consists of selections from the correspondence t 7!
S . p .t/ ; k; w/. When Dt .w/ > k, the short-run equilibrium TOU price, p?SR .t; k; w/,
exceeds w by whatever is required to bring the demand down to k (Fig. 2.1a). The
total of this excess, or “capacity premium”, over the cycle is the unit operating profit,
which in the long run should equal the unit capacity cost r. That is, the long-run
Fig. 2.1 Short-run approach to long-run equilibrium of supply and (cross-price independent)
demand for thermally generated electricity: (a) determination of the short-run equilibrium price
and output for each instant t, given a capacity k; (b) and (d) trajectories of the short-run equilibrium
price and output; (c) the short-run cost curve. When k is such that the shaded area in (b) equals r,
the short-run equilibrium is the long-run equilibrium
equilibrium capacity, k? .r; w/, can be determined by solving for k the equation
Z T
rD . p?SR .t; k; w/ w/C dt (2.1.2)
0
This poses a fixed-point problem that, unlike the short-run approach, is not much
simplified by cross-price independence of demands.1
is nondifferentiable whenever the output y has multiple peaks: indeed, for every
satisfying (2.1.4), the function p D w C r represents a subgradient of CLR with
respect to y (w.r.t. y). And multiple peaks are more the rule than the exception in
equilibrium—note the peak output plateau in Fig. 2.1d here, and see [26] for an
extension to the case of cross-price dependent demands. Similarly, the short-run
1
In terms of the subdifferential, @C, of the long-run cost (2.2.1) as a function of output, the fixed-
point problem is to find a function p such that p 2 @CLR .D . p//, where D . p/ .t/ D Dt . p .t// if
demands are cross-price independent.
18 2 Peak-Load Pricing with Cross-Price Independent Demands: A Simple Illustration
cost
( RT
w 0 y .t/ dt if 0 y k
CSR .y ./ ; k; w/ D (2.2.2)
C1 otherwise
2
The SRMC and the short-run supply correspondences are inverse to each other, i.e., they have the
same graph: in Fig. 2.1a, the broken line is both the supply curve and the SRMC curve.
3
This condition (r D r k …SR ) is stronger than cost-optimality of the fixed inputs (when p is an
SRMC).
2.3 Equilibrium Prices for the Single-Consumer Case 19
over x ./ 0 and m 0, the consumptions of the nonstorable good and of the
numeraire, subject to the budget constraint
Z T
mC p .t/ x .t/ dt M
0
where M is the income and p ./ is a TOU price in terms of the numeraire
(which represents all the other goods and thus closes the model). For each t,
the instantaneous utility u .t; x/ is taken to be a strictly concave, increasing and
differentiable function of the consumption rate x 2 RC , with .@u=@x/ .t; 0/ > w (to
ensure that, in a short-run equilibrium with k > 0, consumption is positive at every
t). The household’s income consists of an endowment of the numeraire (mEn ) and
the pure profit from electricity sales, i.e.,
Z T
M Dm Ck En
. p .t/ w/C dt rk.
0
To guarantee a positive demand for the numeraire, assume that mEn > .Tw C r/ k.
Then, at any time t, demand for the good depends only on the current price p .t/,
and it is determined from the equation
@u
.t; x .t// D p .t/ .
@x
In other words, Dt .p/ D ..@u=@x/ .t; //1 .p/. When w < .@u=@x/ .t; k/, this value
of @u=@x is the price needed to equate demand to k. So the short-run equilibrium
price is
C
@u
p?SR .t; k; w/ D w C .t; k/ w . (2.3.1)
@x
Costs and profits of a price-taking producer are, by definition, the optimal values of
programmes with quantities as decision variables. With several decision variables,
it can be easier to solve the mathematical problem in stages, by fixing some of
the variables and dealing with the resulting subproblem first. The subproblem may
also be of independent interest, especially if it corresponds to a stage in a practical
implementation of the complete solution. In the context of production, the decision
on how to operate a plant with a fixed equipment corresponds to short-run profit
maximization as a subproblem of long-run profit maximization: although plant
operation is usually planned along with the investment, the producer is still free
to make operating decisions after constructing the plant. In other words, his final
choices of the outputs y and the variable inputs v are made only after fixing the
capital inputs k. Such a multi-stage problem can be solved in the reverse order: the
decisions to be implemented last are determined first, but are made contingent on
the decisions to be implemented earlier, and the complete solution is put together by
back substitution. For the producer, this means first choosing y and v to maximize
the short-run profit, given an arbitrary k as well as the prices, p and w, for the
variable commodities. Even within the confines of the purely periodic (or static)
problems considered here, this approach has a couple of analytical advantages. First,
in addition to being of independent interest, the short-run equilibrium (with a fixed
k) may be easier to find than the long-run equilibrium—as in Chap. 2. Second, when
there is a number of technologies, the short-run equilibrium is usually easier to
find by solving the profit maximization programmes to determine the total short-run
supply (and then equate it to demand) than by solving the duals of cost minimization
programmes to determine the SRMCs (which would then have to be equated both
to one another and to inverse demand). This profit approach is simpler than the cost
approach in two ways: it gives unique solutions to the producer problem and its dual,
and it reduces the number of unknowns in the subsequent equilibrium problem (see
Sect. 4.1).
A third advantage of the short-run approach emerges when the framework—
unlike this one—takes account of non-periodic demand and price uncertainty. Prices
for the variable commodities . p; w/, or their probability distribution in a stochastic
model, will change in unforeseen ways between the planning and the completion of
plants, and will keep shifting thereafter also. As a result, both the plant mix and the
design of individual plants will become suboptimal. But whether a plant is optimal
or not, it should be operated optimally, and a solution to this problem is a part of the
short-run approach.
The above considerations are what makes short-run profit maximization the
subproblem of central interest to us. It, too, may be solved in two stages, though
this time the order in which the decision variables (y and v) are determined is only a
matter of convenience—it is usually best to start with the simpler subproblem. Here,
it is assumed that short-run cost minimization (finding v given k and y) is easier than
revenue maximization (finding y given k and v). The solution sequence (first v, then
y and finally k) corresponds to a chain of three problems: (i) the “small” one of
short-run cost minimization (with k and y as data, v as a decision variable), (ii) an
“intermediate” problem of short-run profit maximization (with k as a datum, y and v
as decision variables), and (iii) the “large” problem of long-run profit maximization
(with k, y and v as decision variables).
A fourth problem—another intermediate one—is that of long-run cost minimiza-
tion (with y as a datum, k and v as decision variables). It is in terms of this problem
and its value function that public utilities usually formulate their welfare-promoting
principles of meeting the demand at a minimum operating cost, optimizing their
capital stocks, and pricing their outputs at LRMC. Together, these policies result in
long-run profit maximization and competitive equilibrium in the products’ markets.
Although the separate aims are usually stated in terms of long-run costs, as LRMC
pricing and LRC minimization, their combination is best achieved through short-run
calculations—for the reasons outlined above and detailed in Sect. 4.1.
Each of the four problems, when formulated as one of optimization constrained
by a convex (and nonempty) production set Y, has a linear objective function.1 This
has several implications. One is that each problem (SRC or LRC minimization,
or SRP or LRP maximization) can be formulated as a linear programme (LP), by
representing Y as the intersection of a finite or infinite set of half-spaces. This
is discussed further in Sect. 3.12. What matters for now is that in passing to a
subproblem, once a decision variable has become a datum (like k in passing from
the long to the short run), the corresponding term of the linear optimand (r k)
1
Even if the objective were nonlinear, it could always be replaced by a linear one with an extra
scalar variable, subject to an extra nonlinear constraint: as is noted in [12, p. 48], minimization of
f .y/ over y is equivalent to minimization of % over y and % subject to: % f .y/, in addition to any
original constraints on y.
3.1 Cost and Profit as Values of Programmes with Quantity Decisions 23
can be dropped, since it is fixed. Its coefficient (r) can then be removed from the
subproblem’s data (which include k).2
The commodity spaces for outputs, fixed inputs and variable inputs are denoted
by Y, K and V, respectively. These are paired with price spaces P, R and W
by bilinear forms (a.k.a. scalar products) denoted by h p j yi, etc.; the alternative
notation p y is employed to mean pT y when both P and Y are equal to the finite-
dimensional space Rn (where pT is the row vector obtained by transposing a column
p). Unless specified, the range of a decision variable (say y) is the whole space (Y).
With p, r and w denoting the prices for outputs, fixed inputs and variable inputs
(y, k and v, respectively), the long-run profit maximization programme is:
Its optimal value, the maximum LRP as a function of the data, is denoted by
…LR . p; r; w/. By definition, .y; k; v/ solves (3.1.1)–(3.1.2) if and only if
In the central case of constant returns to scale (c.r.t.s.), the production set Y is a
cone, and …LR is the 0-1 indicator of the polar cone
2
More generally, this is so whenever the optimand separates into a function of .r; k/ plus terms
independent of r and k.
24 3 Characterizations of Long-Run Producer Optimum
w
…LR
r p
% -
k y
w …SR CLR w
p r
- %
y k
CSR
w . (3.1.12)
For example, the arrow from the y next to CSR to the p next to …SR indicates that
…SR is, as a function of p, the Fenchel-Legendre convex conjugate of CSR as a
function of y, with .k; w/ fixed—i.e., by definition,
The right half of the diagram (3.1.12) represents similar relationships between CLR
and CSR or …LR . Details such as the signs and convexity or concavity are omitted.
3.2 Split SRP Optimization: A Primal-Dual System for the Short-Run Approach 25
As is spelt out next, those y and k which yield the suprema in (3.1.13) and (3.1.14)
are parts of an input-output bundle that maximizes the long-run profit.
and the bundle .y; v/ maximizes short-run profit (given k) at prices . p; w/ or,
equivalently,
The system (3.2.1)–(3.2.3) shall be called the Split LRP Optimization System. Its
SRC subprogramme for v in (3.2.3) is taken to be readily soluble. By contrast, the
reduced SRP programme for y in (3.2.2) may require the duality approach. This
consists in pricing the constraining parameters and solving the dual programme of
valuation together with the primal programme of operation (when there is no duality
gap). For the SRP programme as the primal, this means valuing the fixed inputs k: a
dual solution (with no duality gap) is a shadow-price system r such that
The duality scheme that produces the programme in (3.2.4) or (3.2.6) as the dual of
SRP maximization is set out in detail in Sect. 3.3.
As well as helping to solve the operation problem in (3.2.2), the dual solution
can be used to check the investment for optimality, i.e., (3.2.1) is equivalent
to (3.2.4)–(3.2.5).3 The system (3.2.2)–(3.2.5) is therefore equivalent to (3.2.1)–
(3.2.3), and hence also to LRP maximization (3.1.3), and to Complementarity
Conditions (3.1.5) under c.r.t.s. It is, however, put entirely in terms of solutions
to the SRP programme for .y; v/ and its dual programme for r, with the primal split
into the SRC programme (for v) and the reduced SRP programme (for y). Therefore,
(3.2.2)–(3.2.5) shall be called the Split SRP Optimization System. It is likely to
be the best basis for the short-run approach when the technology is specified by
means of a production set (as is usual in an engineer’s description of a technology)
and, in addition, the SRC is readily calculable. Alternative systems are presented in
Sects. 3.4 to 3.6, 3.8 and 3.9.
Unless there are duality gaps, short-run and long-run cost and profit are also the
optimal values of programmes that are dual to those of Sect. 3.1. The scheme
producing the duals is an application of the usual duality framework for convex
3
Formally, this follows from the definitional conjugacy relationship (3.1.14) between …SR and …LR
(as functions of k and r, respectively) by using the first-order condition (B.5.5) and the Inversion
Rule (B.6.2) of Appendix B.
3.3 Duality: Cost and Profit as Values of Programmes with Shadow-Price. . . 27
programmes (CPs), expounded in, e.g., [44] and [36, Chapter 7]. The present scheme
is, however, a little different in that it starts not from a single programme, yet to be
perturbed, but from a family of programmes that depend on a set of data, whose
particular values complete a programme’s specification. So, one way to perturb
a programme is simply to add an increment to its data point, thus “shifting” it
within the given family. Some, possibly all, of the scheme’s primal perturbations
are therefore increments to some—though typically not all—of the data. The same
goes for dual perturbations.
Before the duality scheme is applied to the profit and cost programmes, it is
briefly discussed and illustrated in the framework of linear programming. A central
idea is that the dual programme depends on the choice of perturbations of the primal
programme: different perturbation schemes produce different duals. Theoretical
expositions of duality usually start from a programme without any data variables
whose increments might serve as primal perturbations: say, f .y/ is to be maximized
over y subject to a number of inequalities G1 .y/ 0, G2 .y/ 0; : : :, abbreviated
to G .y/ 0. In such a case, any perturbations must first be introduced, and the
standard choice is to add D . 1 ; 2 ; : : :/ to the zeros on the right-hand sides
(r.h.s.’s)—thus perturbing the original constraints G .y/ 0 to G .y/ . The
original programme has no data other than the functions f and G themselves,
and the increments f and G (which change the programme to maximization
of . f C f / .y/ over y subject to .G C G/ .y/ 0) can never serve as primal
perturbations—not even if they were taken to be linear, i.e., if f and G were a
vector and a matrix of coefficients of the primal variables y D .y1 ; y2 ; : : :/. This
is because the perturbed constrained maximand must be jointly concave in the
decision variables and the perturbations,4 but the bilinear form f y is neither concave
nor convex in f and y jointly.5
But in applications, the primal programme usually comes with a set of data that it
depends on, and increments to some of the programme’s data can commonly serve
as primal perturbations. Such data shall be called the intrinsic primal parameters ;
some or all of the other data will turn out to be dual parameters. For example, in
SRP maximization (3.1.6)–(3.1.7), the fixed-input bundle k is a primal parameter
because, since the production set Y is convex, the constrained maximand is a
concave function of .y; k; v/: it is
h p j yi hw j vi ı .y; k; v j Y/
4
This is equivalent to joint convexity of the constrained minimand (which is the sum of the
minimand and of the 0-1 indicator function of the constraint set). In [44] it is called “the
minimand” for brevity.
5
A linear change of variables makes it a saddle function: 4f y D . f C y/. f C y/. f y/. f y/
is convex in f C y and concave in f y.
28 3 Characterizations of Long-Run Producer Optimum
6
In this as in other contexts, it can be convenient to think of extrinsic perturbations either as
(i) complementing the intrinsic perturbations (which are increments to the fixed inputs) by varying
some aspects of the technology (such as nonnegativity constraints), or as (ii) replacing the intrinsic
perturbations with finer, more varied increments (to the fixed inputs). For example, the time-
constant capacity k in (5.2.3) is an intrinsic primal parameter. The corresponding perturbation
is a constant increment k , which can be refined to a time-varying increment k ./. The
perturbation k ./ is complemented by the increment n ./ to the zero floor for the output rate
y ./ in (5.2.3). The same goes for all the occurrences of k and n in the context of pumped
storage and hydro (where
is another complementary extrinsic perturbation, of the balance
constraint (5.2.15) or (5.2.35)).
3.3 Duality: Cost and Profit as Values of Programmes with Shadow-Price. . . 29
to maximization). When all the primal perturbations are intrinsic, the resulting dual
programme is called the intrinsic dual .
Some or possibly all of the dual perturbations may turn out to perturb the dual
programme just like increments to some of the data—which are thus identified as
the intrinsic dual parameters . Any other dual perturbations are called extrinsic, and
these can be thought of as increments to extrinsic dual parameters (whose original,
unperturbed values are set at zeros). However, in the profit or cost programmes, all
the dual parameters are price data (and are therefore intrinsic).
In the reduced formulations of the profit or cost problems, some of the price
data are not dual parameters because the corresponding quantities have been solved
for in the reduction process, and have thus ceased to be decision variables: e.g.,
the variable-input price w is not a dual parameter of the reduced SRP programme
in (3.2.2) because the corresponding input bundle v has been found in SRC
minimization (and so it is no longer a decision variable). But in the full (not reduced)
formulations, all the price data are dual parameters, and thus the programme’s data
(other than the technology itself) are partitioned into the primal parameters (the
quantity data) and dual parameters (the price data).
The primal and dual optimal values can differ at some “degenerate” parameter
points (see Appendix A), but such duality gaps are exceptional, and they do not
occur when the primal or dual value is semicontinuous in, respectively, the primal or
dual parameters (Sect. 6.1). Note that both optimal values, primal and dual, depend
on the data, which are the same for both programmes. So, in this scheme, either of
the optimal values (primal or dual) is a function of both primal and dual parameters,
and so it can have two types of continuity and of derivatives (marginal values):
• continuity/derivative of Type One is that of the primal value with respect to the
primal parameters, or of the dual value w.r.t. the dual parameters;
• continuity/derivative of Type Two is that of the dual value w.r.t. the primal
parameters, or of the primal value w.r.t. the dual parameters.
This useful distinction cannot be articulated when, as in [44] and [36], the primal
and dual values are considered only as functions of either the primal or the dual
parameters, respectively.
Comments (Parameters and Their Marginal Values, Dual Programme and the
FFE Conditions, the Lagrangian and the Kuhn-Tucker Conditions for LPs)
• Let the primal linear programme be: Given any p 2 Rn and s 2 Rm , and an
m n matrix A, maximize p y over y 2 Rn subject to Ay s. Here, the only
intrinsic primal parameter is the standard parameter s. There is no obviously
useful candidate for an extrinsic primal parameter, and if none is introduced,
then the dual is the standard dual LP: Given p and s (and A), minimize s over
30 3 Characterizations of Long-Run Producer Optimum
Ay s; 0; pT D T A and p y D s (3.3.3)
7
The dual constraint AT D p must be changed to AT p if y 0 is adjoined as another primal
constraint (in which case the primal LP can be interpreted as, e.g., revenue maximization—given
a resource bundle s, an output-price system p and a Leontief technology defined by an input-
coefficient matrix A).
8
Without a proof of value differentiability, the Generalized Envelope Theorem is given also in, e.g.,
[47, 1.F.b].
9
These are the Lagrangian Saddle-Point Conditions (0 2 @ L and 0 2 @O y L) for the present LP
case.
10
In this case, equivalence of the Kuhn-Tucker Conditions and the FFE Conditions can be seen
directly, but it holds always (since, by the general theory of CPs, each system is equivalent to the
conjunction of primal and dual optimality together with absence of a duality gap).
3.3 Duality: Cost and Profit as Values of Programmes with Shadow-Price. . . 31
which is linear in .y; /. This makes it simpler to solve than the system of Kuhn-
Tucker Conditions (3.3.2). The FFE system (3.3.3) is so effective because, in
linear programming, the dual programme can be worked out from the primal
explicitly.
• But the dual of a general CP cannot be given explicitly (i.e., without leaving
an unevaluated extremum in the formula for the dual constrained objective
function in terms of the Lagrangian).11 That is why, as a general solution method
for convex programming, the Kuhn-Tucker Conditions are better than the FFE
Conditions, although the latter system is simpler in some important specific
cases (such as linear programming). Whereas using the FFE Conditions requires
forming the dual from the primal to start with, using the Kuhn-Tucker Conditions
requires only the Lagrangian. Thus the latter Kuhn-Tucker Conditions offer a
workable general method of solving the primal-dual programme pair, and this
matters more than an explicit expression for the dual programme. The FFE
Conditions can, however, be simpler in the case of a specific CP that, like an
LP, has an explicit dual.
The duality scheme is next applied to all four of the profit and cost programmes
of Sect. 3.1; the one of most importance in the context of a decentralized industry
(such as the ESI of Sects. 5.1 to 5.3) is the programme of SRP maximization. The
duals are shown to consist in shadow-pricing the given quantities—and so their
subprogramme relationship is the reverse of that between the primals: the more
quantities that are fixed, the more commodities there are to shadow-price. (In other
words, the fewer primal variables, the more primal parameters, and hence more
dual variables.) For this reason, the duals are listed, below, in the order reverse to
that in which the primals are listed in Sect. 3.1. See also Fig. 3.1, in which the
large single arrows point from primal programmes to their subprogrammes, and the
double arrows point from the dual programmes to their subprogrammes. Each of
the four middle boxes gives the data for the pair of programmes represented by the
two adjacent boxes (the outer box for the primal and the inner box for the dual); the
data are partitioned into the primal parameters (the given quantities) and the dual
parameters (the given prices). There are no other parameters in this scheme (i.e., it
has no extrinsic parameters).
11
The standard dual to the ordinary CP of maximizing a concave function f .y/ over y subject
to G .y/ s (where G1 , G2 , etc., are convex functions) is to minimize supy L .y; / WD
supy . f .y/ C .s G .y/// over 0 (the standard dual variables, which are the Lagrange
multipliers for the primal constraints): see, e.g., [44, (5.1)]. And supy L (the Lagrangian’s
supremum over the primal variables) cannot be evaluated without assuming a specific form for
f and G (the primal objective and constraint functions).
32 3 Characterizations of Long-Run Producer Optimum
Fig. 3.1 Decision variables and parameters for primal programmes (optimization of: long-run
profit, short-run profit, long-run cost, short-run cost) and for dual programmes (price consistency
check, optimization of: fixed-input value, output value, output value less fixed-input value). In
each programme pair, the same prices and quantities—. p; y/ for outputs, .r; k/ for fixed inputs,
and .w; v/ for variable inputs—are differently partitioned into decision variables and data (which
are subdivided into primal and dual parameters). Arrows lead from programmes to subprogrammes
Its optimal value is denoted by C SR .y; k; w/ CSR .y; k; w/, with equality when
Sect. 6.2 applies. The dual parameter is w.
In the LRC minimization programme (3.1.8)–(3.1.9), only y can serve as a
primal parameter, and perturbation by the increment y yields the following dual
12
Since the minimand hw j vi is not jointly convex in w and v, w cannot serve as a primal parameter
(it will turn out to be a dual parameter).
3.3 Duality: Cost and Profit as Values of Programmes with Shadow-Price. . . 33
Its optimal value is denoted by CLR .y; r; w/ CLR .y; r; w/, with equality when
Sect. 6.2 or 6.4 applies. The dual parameters are r and w.
In the SRP maximization programme (3.1.6)–(3.1.7), only k can serve as a
primal parameter, and perturbation by the increment k yields the following dual
programme of shadow-pricing the fixed inputs:
Its optimal value is denoted by …SR . p; k; w/ …SR . p; k; w/, with equality when
Sect. 6.2 or 6.4 applies.13 The dual parameters are p and w.
The same programme for r—viz., (3.3.6) or (3.3.13)–(3.3.14) under c.r.t.s.—is
also the dual of the reduced SRP programme in (3.2.2), again with k as the primal
parameter. That is, the reduced and the full primal programmes have the same primal
parameters and the same dual programme. Of course, the two duality relationships
cannot be exactly the same because the two dual parameterizations are different: as
has already been pointed out, the reduced primal programme has fewer variables,
and hence fewer dual parameters, than the full programme, whose data are its primal
and dual parameters. Since both programmes have the same data, it follows that
the reduced one has a datum that is neither a primal nor a dual parameter. In the
case of the reduced SRP programme in (3.2.2), such a datum is w: the only primal
parameter is k, and the only dual parameter is p (since y is the only primal variable).
For comparison, in the full SRP programme (3.1.6)–(3.1.7) both p and w are dual
parameters (paired to the primal variables y and v).14
The LRP maximization programme (3.1.1)–(3.1.2) is, in this context, unusual
because none of its data (p, r, w) can serve as a primal parameter—all of the data
are dual parameters. This means that the intrinsic dual has no decision variable;
formally, it is: Given . p; r; w/, minimize …LR . p; r; w/. Having no variable, the dual
minimand is a constant, and it equals the primal value (…LR ): since the dual is trivial,
there can be no question of a duality gap in this case.
By contrast, the other programme pairs can have duality gaps, especially when
the spaces are infinite-dimensional. But even then a gap can appear only at
an exceptional data point: the primal and dual values are always equal under
Slater’s Condition, as generalized in [44, (8.12)], or the compactness-and-continuity
13
As the notation indicates, … and C are thought of mainly as dual expressions for … and C
(although duality of programmes is fully symmetric).
14
A similar remark applies to the full and reduced shadow-pricing programmes, (3.3.4) for . p; r/
and the one in (3.4.7) for p alone. Taken as the primal parameterized by w, each has the same dual,
viz., the SRC programme (3.1.10)–(3.1.11). And both of the other vector data (y and k) are dual
parameters for the full programme (3.3.4). But the datum k is neither a dual nor a primal parameter
for the reduced programme in (3.4.7).
34 3 Characterizations of Long-Run Producer Optimum
conditions of [44, Example 4’ after (5.13)] and [44, Theorem 18’ (d) or (e)]. In the
problem of profit-maximizing operation of a plant with capacity constraints, Slater’s
Condition requires only that the capacities be strictly positive, i.e., that k 0; in
other words, it is always met unless the plant lacks a component. See Lemma 6.4.1
and Proposition 7.4.2 for details, and Appendix A for a counterexample when k is
not strictly positive.
The partial conjugacy relationships between the dual value functions (CSR , CLR ,
…SR , and …LR D …LR ) can be summarized in a diagram like that in (3.1.12) for the
primal values, but with the arrows reversed (and with bars added to the symbols …
and C):
w
…LR
r p
. &
k y
w …SR C LR w
p r
& .
y k
CSR
w . (3.3.7)
For example, the arrow from the p next to …SR to the y next to C SR indicates that
CSR is, as a function of y, the convex conjugate of …SR as a function of p (with k
and w fixed): i.e., by definition,
˚
CSR .y; k; w/ D sup h p j yi …SR . p; k; w/ . (3.3.8)
p
The dual of LRP maximization has no decision variable, and, with c.r.t.s., it may be
thought of as a price consistency check : its value is 0 if . p; r; w/ 2 Yı , and C1
otherwise. Formally, the dual is:
Thus, with c.r.t.s., the dual objectives are “automatic”, and formation of the dual
programmes boils down to working out Yı from a specific cone Y. Two frameworks
for this are provided in Sects. 3.12 and 7.2.
Like the primals, the three duals (of the SRC and LRC and SRP programmes) are
henceforth named after their objectives: OFIV, OV and FIV. Strictly speaking, this
terminology fits only the case of c.r.t.s. for the long run (i.e., the case of a production
cone). But it will be used also when c.r.t.s. are not assumed (in Fig. 3.1, Sect. 3.4
and Tables 3.1 and 3.2).
Comments (Dual of a CP More General Than the Profit and Cost
Programmes)
• The dual can be similarly spelt out for a programme of a more general form, with
a parametric primal maximand
Table 3.1 The SRP optimization system with its split form, and four derived differential systems
(three of which are derived directly by applying the DP and FOC, and one indirectly by using also
the SSL)
First-Order Condition
Deriv. Prop. of Opt. Val. (twice)
m m Deriv. Prop. of Opt. Val. (twice)
Absorption of No-Gap Cond.
Absorption of No-Gap Cond.
SRP Opt. Sys. Split SRP Opt. Sys.
(3.4.1)–(3.4.3) Two-stage (3.2.2)–(3.2.5)
.y; v/ maxi’es short-run profit solving y maximizes revenue less CSR
”
v minimizes short-run cost
r minimizes fixed-input value r minimizes fixed-input value
…SR D …SR at . p; k; w/ …SR D …SR at . p; k; w/
Deriv. Prop. of Opt. Val. (twice)
m
Absorption of No-Gap Cond.
FIV Saddle Diff. Sys. O-FIV Part. Diff. Sys.
(3.6.6)–(3.6.7) Subdiff. (3.6.1)–(3.6.3)
.y; v/ 2 @p;w …SR (Type One) Sect. Lem. y 2 @p …SR
”
v 2 @O w CSR (Type One)
r 2 @O k …SR (Type Two) r 2 @O k …SR
minimand is then
hr j ki C I # . p; r/ (3.3.17)
Table 3.2 The SRC optimization system with its split form, and four derived differential systems
(three of which are derived directly by applying the DP and FOC, and one indirectly by using also
the SSL)
First-Order Condition
Deriv. Prop. of Opt. Val. (twice)
m m Deriv. Prop. of Opt. Val. (twice)
Absorption of No-Gap Cond.
Absorption of No-Gap Cond.
SRC Opt. Sys. Split SRC Opt. Sys.
(3.4.4)–(3.4.6) Two-stage (3.4.5)–(3.4.8)
solving p maximizes revenue less …SR
”
v minimizes short-run cost v minimizes short-run cost
. p; r/ maxs rev. fix.-inp. val. r minimizes fixed-input value
CSR D CSR at .y; k; w/ CSR D CSR at .y; k; w/
Deriv. Prop. of Opt. Val. (twice)
m
Absorption of No-Gap Cond.
SRC Saddle Diff. Sys. SRC/P Part. Diff. Sys.
(3.6.8)–(3.6.9) Subdiff. (3.5.1)–(3.5.3)
Sect. Lem. p 2 @y CSR
”
v 2 @O w CSR (Type Two) v 2 @O w CSR
. p; r/ 2 @y;k CSR (Type One) r 2 @O k …SR (Type One)
• The case of a finite LP, parameterized in the standard way, is obtained when
˚
YD f.y; k/ 2 Rn Rm W Ay kg , so Yı D . p; r/ 2 Rn Rm W pDAT r; r 0
The conjugacy (3.1.14) between the SRP and the LRP has yielded a characterization
of profit-maximizing investments in terms of their imputed values—i.e., it has
served to reformulate the investment-optimality condition (3.2.1) as the valuation
condition (3.2.4) together with the no-gap condition (3.2.4). The valuation pro-
gramme in (3.2.4) has subsequently been obtained as the dual (3.3.6), or (3.3.13)–
(3.3.14) under c.r.t.s., to the short-run profit maximization programme (3.1.6)–
(3.1.7), which appears in a split form in (3.2.2)–(3.2.3). Thus the use of conjugacy
has produced the system (3.2.2)–(3.2.5) of optimality conditions on y, v and r,
and the use of duality has shown that this system means that .y; v/ and r form a
pair of solutions to the SRP programme and its dual.15 Similar arguments lead to
characterizations of optimality in terms of the SRC programme with its dual. So,
either of the following two systems of conditions is equivalent to maximization of
long-run profit at prices . p; r; w/ by an input-output bundle .y; k; v/.
The SRP Optimization System is: .y; v/ maximizes the short-run profit at
prices . p; w/, and r minimizes the value of the fixed-input k (plus maximum LRP
if r.t.s. are decreasing), and the two optimal values are equal (i.e., under c.r.t.s.,
maximum SRP equals minimum FIV). Formally, it is: given . p; k; w/,
The SRC Optimization System is: v minimizes the short-run cost at price w,
and . p; r/ maximizes the value of output y less that of fixed-input k (and less
maximum LRP under d.r.t.s.), and the two optimal values are equal (i.e., under
15
These arguments use the subprogramme and duality concepts to view …SR in two ways: (i) as
the value of a subprogramme, and (ii) as the primal value—and thus arrive in two ways at
the FIV programme for r in (3.2.4). In detail: since …SR is the value of the subprogramme of
LRP maximization obtained by fixing k, its (concave) conjugate w.r.t. k is …LR as a function
of r. It follows—by (B.5.5) and (B.6.2)—that k solves the “conjugacy programme” in (3.2.1)
if and only if: r solves the “reverse conjugacy programme” in (3.2.4) and (3.2.5) holds. Also
by using the conjugacy between …SR and …LR , the same programme for r can be derived as
the dual of SRP maximization parameterized by k. This is done in Proposition 3.10.1 (which
additionally identifies p and w as the dual parameters): it is the formal foundation of duality for
CPs that the dual minimand is the primal parameter times the dual variable, minus the (concave)
conjugate of the primal maximum value (as a function of the primal parameter)—and here, the
sum is hr j ki C …LR .r/. See, e.g., [44, Theorem 7], which here must be applied to the function
k 7! …SR .k C k/ as Rockafellar’s primal value (his is a function of the parameter increment,
rather than of the parameter point as here, and this is what shifts the argument by k and adds the
term hr j ki to minus the conjugate).
3.4 SRP and SRC Optimization Systems 39
c.r.t.s., minimum SRC equals maximum OFIV). Formally, it is: given .y; k; w/,
Additionally, one can split both of the joint programmes for two decision
variables: just as (3.1.6)–(3.1.7) has been split into (3.2.2) and (3.2.3), so the joint
programme (3.3.4) for . p; r/ can be replaced by two programmes for p and r
separately. Condition (3.4.4) is therefore equivalent to16 :
16
The maximum value in (3.4.7) is CSR .y; k; w/, by the definitions of …SR and CSR as the optimal
values of (3.3.6) and (3.3.4).
40 3 Characterizations of Long-Run Producer Optimum
and of the derivative property of CSR as the optimal value of (3.2.3), i.e., into the
system
r 2 @O k …SR . p; k; w/ (3.5.1)
p 2 @y CSR .y; k; w/ (3.5.2)
It shall be called the SRC/P Partial Differential System because it uses the partial
sub/super-differentials @y and @O w of CSR (the SRC) as a saddle (convex-concave)
function of y and w, in addition to using the partial superdifferential @O k of …SR (the
SRP). A similar use of CSR , but as a saddle function of .k; w/, arises later in the
L/SRC Partial Differential System (3.9.8)–(3.9.10); the affices “P” and “L” in the
two names stand for “profit” and “long-run”.
Comments (Absorption of a No-Gap Condition in a Differential Condition)
• The system (3.5.1)–(3.5.3) can be derived also from the Split SRP Optimization
System (3.2.2)–(3.2.5). The FOC for (3.2.2) and the derivative property of CSR
as the value function for (3.2.3) are used just as before. But, instead of the
FOC for (3.2.1), this time the third condition is the derivative property of …SR
as the value function for (3.2.4) or (3.3.6), i.e., that r 2 @O k …SR . p; k; w/. This
and (3.2.5) together mean exactly that r 2 @O k …SR , since (3.2.5) means that
…SR D …SR , at . p; k; w/.
• The last argument is a case of absorbing a no-gap condition in a subdifferential
condition by changing the derivative from Type Two (here, @O k …SR ) to Type
One (@O k …SR ). In this, the value function is changed either from the dual to the
primal (if the parameter in question is primal like the k here), or vice versa . The
optimal solution is always equal to the marginal value of the programme being
solved; this is a derivative of Type Two (see the Shephard-Hotelling Lemmas in
Sect. 3.11). The derivative is actually of Type One—i.e., it is the marginal value
of the programme dual to that being solved—if there is no duality gap. But if
there is a gap, then the Type One derivative does not exist (i.e., the sub/super-
differential is empty). In the context of fixed-input valuation: the set of solutions,
for r, of (3.2.4) or (3.3.6) is always identical to @O k …SR . p; k; w/, which is a
derivative of Type Two—this is the Dual of Hotelling’s Lemma (Lemma 3.11.2).
The solution set equals @O k …SR (a derivative of Type One) if …SR D …SR , at
the given . p; k; w/: see Remark 3.11.8. But if …SR ¤ …SR at . p; k; w/, then
@O k … . p; k; w/ D ; (the empty set); so if r 2 @O k …SR then …SR D …SR (at the
given p, k and w), i.e., there is no duality gap (between SRP maximization and
FIV minimization).
42 3 Characterizations of Long-Run Producer Optimum
Applied to the Split SRC Optimization System (3.4.5)–(3.4.8), the same methods
yield the partial differential system that consists of the FOC for (3.4.7) and the
derivative properties of CSR and …SR as the value functions for (3.4.5) and (3.4.8),
with @w CSR changed to @w C SR to absorb the no-gap condition (3.4.6)—i.e., the
system:
r 2 @O k …SR . p; k; w/ (3.6.1)
y 2 @p …SR . p; k; w/ (3.6.2)
It shall be called the O-FIV Partial Differential System because it uses the partial
sub/super-differentials @p and @O k of …SR (the FIV) as a saddle function of p and k,
in addition to using the partial superdifferential @O w of CSR (the OFIV). Thus it uses
only the dual value functions (…SR and CSR ), whilst the system (3.5.1)–(3.5.3) uses
only the primal value functions (…SR and CSR ).
The derivative property of the optimal value can be used also to transform the
“unsplit” optimization systems of Sect. 3.4 into differential systems. For example,
by the derivative property applied twice, the SRP Optimization System (3.4.1)–
(3.4.3) is equivalent to:
r 2 @O k …SR . p; k; w/ (3.6.5)
which is so named because it uses the (joint) subdifferential @p;w and the superdiffer-
ential @O k of …SR as a saddle (convex-concave) function of . p; w/ and k. It produces
also the FIV Saddle Differential System
r 2 @O k …SR . p; k; w/ . (3.6.7)
v 2 @O w CSR .y; k; w/ ; . p; r/ 2 @y;k CSR .y; k; w/ and CSR .y; k; w/ D CSR .y; k; w/
3.7 Transformations of Differential Systems by Using SSL or PIR 43
So far, all the differential systems have been derived from optimization systems—
and this has to be so in convex analysis because it uses the FOC for minimization as
the very definition of the subdifferential: see (B.3.2). But this definition can be used
also to transform one subdifferential condition into another. Once formulated, such
results can be applied to transform one differential system into another directly, i.e.,
without passing through the FOC explicitly. In particular, the partial differential
systems can be derived from the saddle differential systems, which use joint
subdifferentials: a condition involving a subdifferential taken jointly in two groups
of variables—such as @y;k CSR in (3.6.9) or @p;w …SR in (3.6.6)—can be recast in
terms of partial subdifferentials (@y , @k , @p , @w ). This cannot, however, be achieved
simply by splitting the joint derivative into the partials (as in the differentiable case)
because a joint subdifferential does not usually factorize into the Cartesian product
of the partials: it is a general convex set, not a product set. In other words, the
obvious inclusions17
17
Being fixed, the third variable (w or k) is suppressed from (3.7.1) and (3.7.2).
44 3 Characterizations of Long-Run Producer Optimum
are usually strict: see Sect. B.8 of Appendix B for further explanation and examples.
But the two variables of differentiation can be split from each other in a different
way—one that parallels, and derives from, the staged approach to optimization
(introduced in Sect. 3.2). First, the joint subdifferential is used to formulate a FOC
for simultaneous optimization over the two variables. This programme is then split
into two successive optimization programmes with one variable each—and each of
the two has a separate FOC that uses a partial subdifferential. In the case of @y;k CSR ,
this argument consists in stating (i) the FOC for maximizing the LRP over y and k
simultaneously and (ii) the FOCs for maximizing it over y and k successively. The
FOC for a maximum of h p j yi hr j ki CSR over .y; k/ is that . p; r/ 2 @y;k CSR .
The FOC for a maximum of h p j yi CSR .y; k; w/ over y is that p 2 @y CSR ; the
maximum value is …SR , and the FOC for a maximum of …SR . p; k; w/ hr j ki over
k is that r 2 @O k …SR . Since the “joint” FOC is equivalent to the two “partial” FOCs
together,18
. p; r/ 2 @y;k CSR .y; k; w/ , p 2 @y CSR .y; k; w/ and r 2 @O k …SR . p; k; w/ .
(3.7.3)
This is the Subdifferential Sections Lemma (SSL) for this context; it requires
bringing in another function (…SR ), which is linked to the original function (CSR )
by partial conjugacy. This result is formalized fully in Appendix B (Lemma B.7.2).
The SSL is the basic tool for “splitting” joint subdifferentials, but there is also
a couple of derived techniques, viz., the Partial Inversion Rule and its dual variant
(PIR and DPIR, i.e., Corollaries B.7.3 and B.7.5). These can be applied to the joint
subdifferentials of Sect. 3.6:
1. With k fixed, the DPIR applies to CSR .; k; / as a saddle function on Y W which
is a partial conjugate of the 0-1 indicator of the short-run production set YSR .k/,
defined formally by (6.2.1). The indicator is a convex function on Y V, and its
total conjugate is …SR .; k; /, a convex function on P W. It follows that the
condition .y; v/ 2 @p;w …SR can be replaced by: p 2 @y CSR and v 2 @O w CSR .
Thus the SRP Saddle Differential System (3.6.4)–(3.6.5) can be transformed into
the SRC/P Partial Differential System (3.5.1)–(3.5.3).19
The SRC/P Partial Differential System, (3.5.1)–(3.5.3), can be derived also
from the SRC Saddle Differential System (3.6.8)–(3.6.9). This is what (3.7.3)
shows: with w fixed, the SSL applies to …SR .; ; w/ as a saddle function on PK
which is (by definition) a partial conjugate of CSR .; ; w/, a convex function on
18
Dually, (3.6.6) is equivalent to (3.6.2)–(3.6.3), i.e., .y; v/ 2 @p;w …SR if and only if both y 2
@p …SR and v 2 @O w CSR .
19
The PIR would give the same result, but it would require establishing first that CSR .; k; w/ is
l.s.c. to invert the conjugacy relationship (3.1.13), i.e., to show that the saddle function CSR .; k; /
is indeed a partial conjugate of the bivariate convex function …SR .; k; /. This can be problematic
(as is noted in the Comment after Corollary B.7.5).
3.8 Summary of Systems Characterizing Long-Run Producer Optimum 45
Y K. So the condition . p; r/ 2 @y;k CSR can be replaced by: p 2 @y CSR and
r 2 @O k …SR .
2. Similarly, with w fixed, the DPIR applies to …SR .; ; w/ as a saddle function
on P K which is a partial conjugate of …LR .; ; w/. When Y is a cone, the
latter function is the indicator of Yıw , the section of Yı through w. In any case,
it is a convex function on P R, and its total conjugate is C SR .; ; w/, a convex
function on Y K. This shows that the condition . p; r/ 2 @y;k CSR .y; k; w/ can
be replaced by: y 2 @p …SR and r 2 @O k …SR . Thus the OFIV Saddle Differential
System (3.6.10)–(3.6.11) can be transformed into the O-FIV Partial Differential
System (3.6.1)–(3.6.3).20
The O-FIV Partial Differential System (3.6.1)–(3.6.3) can be derived also
from the FIV Saddle Differential System (3.6.6)–(3.6.7). This is because, with
k fixed, the SSL applies to C SR .; k; / as a saddle function on Y W which is (by
definition) a partial conjugate of …SR .; k; /, a convex function on P W. So the
condition .y; v/ 2 @p;w …SR can be replaced by: y 2 @p …SR and v 2 @O w CSR .
Comment (Partial Subdifferentials as Projections of the Joint One) Although
the inclusion (3.7.1)—that @y;k CSR @y CSR @k CSR —is generally strict, it is usually
tight in the sense that @y CSR .y; k/ is equal to the projection of @y;k CSR .y; k/ onto P
if (and only if) every p 2 @y CSR .y; k/ can be complemented to some . p; r/ 2
@y;k CSR .y; k/. A similar result holds for @k CSR (if each of its elements can be so
complemented). For the existence of a complementary subgradient, see Sect. B.8 of
Appendix B.
The ten duality-based systems presented so far (Sects. 3.2, 3.4, 3.5 and 3.6) and the
proofs of their equivalence (detailed in Sect. 3.11) are summarized in Tables 3.1
and 3.2. Since the top right entry of the one table is identical to the bottom right of
the other, the twelve entries include two repetitions. The ten distinct entries are all
the duality-based systems given so far. Seven more systems, to appear in Sect. 3.9,
use the LRC programme and its dual or their value functions. They are mirror
images of the systems shown in the two tables, from which they can be obtained
by replacing …SR . p; k/ with CLR .y; r/ and changing the signs where needed. Thus
three of the seven correspond to the systems on the left in Table 3.1, and the other
20
The PIR would give the same result, but it would require establishing first that …SR .; k; w/ is
l.s.c. to invert the conjugacy relationship (3.3.8), i.e., to show that the saddle function …SR .; ; w/
is indeed a partial conjugate of the bivariate convex function CSR .; ; w/. This can be problematic
(as is noted in the Comment after Corollary B.7.5).
46 3 Characterizations of Long-Run Producer Optimum
four come from the distinct systems on the right in Tables 3.1 and 3.2.21 In other
words, Tables 3.1 and 3.2 deal explicitly with the values and programmes in the
left halves of the conjugacy diagrams (3.1.12) and (3.3.7), but the analysis applies
equally to the right halves.
In the differential systems, the Type One derivatives whose existence rules out
duality gaps are identified. In optimization systems, the various dual programmes
are referred to as “optimization of the fixed quantities’ value”, although this name
fully fits only the case of c.r.t.s. (which need not be assumed). The constraint sets
(Y, and Yı under c.r.t.s.) are not shown in the summarizing tables.
Comments (Partition into a Short-Run Subsystem and a Valuation Condi-
tion)
• All but three of the ten systems shown in Tables 3.1 and 3.2—all except for the
three on the left in Table 3.2, viz., (3.4.4)–(3.4.6), (3.6.8)–(3.6.9) and (3.6.10)–
(3.6.11)—contain a valuation condition on r and . p; k; w/ that, together with
the no-gap condition …SR D …SR , is equivalent to the investment being at a
profit maximum (3.2.1), i.e., to k being a profit-maximizing investment at the
output/fixed-input/variable-input prices . p; r; w/ for the outputs and the fixed
and variable inputs. The condition in question is either “r minimizes the FIV”,
or r 2 @O k …SR , or r 2 @O k …SR (the last of which by itself rules out the duality
gap). Put together, the system’s other conditions (on p, y, w, v and k) are then
equivalent either to (3.2.2)–(3.2.3), or to (3.2.2)–(3.2.3) with (3.2.5)—i.e., to
.y; v/ being a short-run profit-maximizing input-output bundle at prices . p; w/
and given capital inputs k (either with or without …SR being equal to …SR ). This
short-run subsystem is to be solved for v and either y or p—given w and either
p or y, as well as k. The subsystem may be so simple that, as in Chap. 2, it
can be solved on its own, separately from the remaining valuation condition and
without recourse to duality. In such a case, calculation of the imputed values
can be postponed until, in the last stage of the short-run approach to long-run
equilibrium, they are to be equated to the fixed-input prices. As well as being
handy in such simple cases, the system’s partition (into a short-run subsystem
and a valuation condition involving r) is worth examining in detail to clarify the
various ways in which the complete systems rule out duality gaps. Most do so
within the short-run subsystem, but some rely on the valuation condition—when
it takes the form r 2 @O k …SR (a Type One derivative). Therefore, the various short-
run subsystems describe two “grades” of short-run profit maximum: the “plain”
one, and the one without a duality gap. Only the latter kind can be a long-run
profit maximum (for some choice of capital-input prices).
• More formally: given . p; w/ and k, a potential long-run profit-maximizing bundle
is a .y; v/ such that .y; k; v/ maximizes long-run profit at . p; w/ and some r.
21
The three systems on the left in Table 3.2 do not yield new ones (when …SR is replaced by
CLR ) simply because they do not involve …SR at all. So there are not ten but seven of the “mirror
images”.
3.9 Extended Wong-Viner Theorem and Other Transcriptions from SRP to LRC 47
22
The subsystem’s condition that CSR D CSR at .y; k; w/ rules out a different duality gap, and
on its own it does not imply that …SR D …SR at . p; k; w/ when y maximizes short-run profit at
. p; k; w/: see Appendix A for an example (in which CSR D CSR trivally because the technology
has no variable input).
48 3 Characterizations of Long-Run Producer Optimum
This is the Extended Wong-Viner Theorem . Note that the condition that r 2 @k CSR
is the FOC for k to yield the infimum in the definitional formula
(which means that CLR is, as a function of r, the concave conjugate of CSR as a
function of k, with y and w fixed).
For comparison, the usual Wong-Viner Envelope Theorem for differentiable
costs gives
p D r y CSR .y; k/
H) p D r y CLR .y; r/ .
r D r k CSR .y; k/ i.e., k yields the inf in (3.9.2)
(3.9.3)
Comparisons with the two “outer” systems in (3.9.1) show that their equivalence is
indeed an extension of (3.9.3). This is because
i.e., when y yields the supremum in (3.1.13).23 In the differentiable case, the
inclusion (3.9.4) reduces to the equality r k …SR D r k CSR (when p D r y CSR ),
and thus (3.9.1) becomes:
if r D r k CSR .y; k/ then p D r y CSR .y; k/ , p D r y CLR .y; r/ (3.9.5)
23
The inclusion (3.9.4) follows directly from (3.1.13) by Remark B.7.4 (applied to the saddle
function …SR as a partial conjugate of CSR ).
24
The inclusion (3.9.7) follows directly from (3.9.2) by Remark B.7.4 (applied to the saddle
function CLR as a partial conjugate of CSR ).
3.9 Extended Wong-Viner Theorem and Other Transcriptions from SRP to LRC 49
which shall be called the L/SRC Partial Differential System because it uses the
partial sub/super-differentials @k and @O w of CSR (the SRC) as a saddle (convex-
concave) function of k and w, in addition to using the partial subdifferential @y
of CLR (the LRC). It is the “mirror image” or transcript of the SRC/P Partial
Differential System (3.5.1)–(3.5.3), with SRP replaced by LRC and with the
variables suitably swapped.25
When the producer is a public utility, LRMC pricing and LRC minimization—
i.e., Conditions (3.9.8) to (3.9.10)—are often taken as the definition of a long-run
producer optimum. If the SRC function is simpler than the LRC function (as is
usually the case), and the SRP function is simple too, then the Extended Wong-
Viner Theorem (3.9.1) can facilitate the short-run approach by characterizing
long-run optimality in terms of the SRC and SRP functions—and this has been
used in the introductory peak-load pricing example (Chap. 2). In that problem, the
cost-minimizing inputs were obvious, but the question was how to ensure, by a
simple condition put in terms of a short-run value function, that an SRMC output
price p was actually an LRMC price, i.e., that p met (3.9.8). This was achieved
25
In detail, the transcript is obtained by swapping p with r and y with k, and by replacing
the function . p; k/ 7! …SR . p; k/ with the function .y; r/ 7! CLR .y; r/: compare (3.1.13)
with (3.9.2).
50 3 Characterizations of Long-Run Producer Optimum
The corresponding result for …SR instead of CLR shows that the inclusion (3.9.4) is
tight in the same sense, i.e.,
[
@k CSR .y; k/ D @O k …SR . p; k/
p2@y CSR .y;k/
.k; v/ 2 @O r;w CLR .y; r; w/ ; p 2 @y C LR .y; r; w/ and CLR .y; r; w/ D CLR .y; r; w/
and, hence, also to the LRC Saddle Differential System (transcript of the SRP
Saddle Differential System (3.6.4)–(3.6.5)), which is:
The FI-OV Partial Differential System (the transcript of the O-FIV Partial
Differential System (3.6.1)–(3.6.3)), which is:
26
Here, two-stage solving means first minimizing hw j vi over v (subject to .y; k; v/ 2 Y)
to find the solution vL and the minimum value CSR D hw j vi L as functions of .y; k; w/, and then
minimizing hr j ki C CSR .y; k; w/ over k to find the solution L
k .y; r; w/. This gives the complete
solution (in terms of y, r and w) as the pair kL .y; r; w/ and vL y; kL .y; r; w/ ; w .
52 3 Characterizations of Long-Run Producer Optimum
This system is called “reverse-split” to distinguish it from the Split SRC Optimiza-
tion System (3.4.5)–(3.4.8), of which it is the transcript. (The two systems differ
only in the order in which p and r are optimized when the joint programme (3.3.4)
is split into two stages: in (3.4.5)–(3.4.8), the first stage is to find r in terms of p and
calculate …SR , whereas in (3.9.26)–(3.9.29), the first stage is to find p in terms of r
and calculate CLR .)
This section gives proofs for Sect. 3.3. That is, the dual programmes of Sect. 3.3 are
next derived formally, by using the framework of [44].
Proposition 3.10.1 (Dual to SRP Programme) The dual of the short-run profit
maximization programme (3.1.6)–(3.1.7), with k as the primal parameter ranging
over the space K paired with R as the range for the dual variable r, is the fixed-
input shadow-pricing programme (3.3.6), or equivalently (3.3.13)–(3.3.14) when
the production set Y is a cone. The dual parameter is . p; w/.
Proof Given . p; k; w/, the parametric primal constrained maximand is h p j yi
hw j vi minus the 0-1 indicator ı .y; k; v j Y/, where y and v are the primal
decision variables, and k is the primal parameter (paired with the dual decision
variable r). Let d0 and d00 denote the dual perturbations (paired with y and v).
By [44, (4.17)] with the primal problem reoriented to maximization, the (perturbed)
dual constrained minimand—a function of r and .d0 ; d00 / as well as . p; k; w/—is
˚˝ ˛
sup d0 ; d00 j y; v hr j ki C h p j yi hw j vi ı .y; k k; v j Y/
y;vI k
˚˝ ˛
D hr j ki C sup p C d0 ; r; w C d00 j y; k k; v W .y; k k; v/ 2 Y
y;v; k
˚˝ ˛
D hr j ki C sup p C d0 ; r; w C d00 j y; k; v W .y; k; v/ 2 Y
y;v;k
D hr j ki C …LR p C d0 ; r; w C d00 .
3.11 Shephard-Hotelling Lemmas and Their Dual Counterparts 53
So, by setting d0 D 0 and d00 D 0, the dual programme is (3.3.6); and when Y
is a cone, the dual is to minimize hr j ki C ı . p; r; w j Yı / over r (since …LR D
ı # . j Y/ D ı . j Yı /, i.e., the support function of a cone is the 0-1 indicator
function of the polar cone). Finally, d0 and d00 perturb the dual like increments to p
and w—which therefore are the dual parameters (and so d0 and d00 may be renamed
to p and w).
The other duals in Sect. 3.3 are derived in the same way; the dual of the SRC
programme is spelt out.
Proposition 3.10.2 (Dual to SRC Programme) The dual of the short-run cost
minimization programme (3.1.10)–(3.1.11), with .y; k/ as the primal parameter
ranging over the space Y K paired with P R as the range for the dual
variable . p; r/, is the output-and-fixed-input pricing programme (3.3.4), or equiv-
alently (3.3.9)–(3.3.10) when the production set Y is a cone. The dual parameter
is w.
This section expands, and gives proofs for, Sects. 3.2, 3.4, 3.5, 3.6, 3.9.
Programme solutions can always be characterized as marginal values of Type
Two, i.e., the primal solution set is equal to the subdifferential of the primal optimal
value as a (convex) function of the dual parameter. Likewise, the dual solution set
is equal to the superdifferential of the dual optimal value as a (concave) function of
the primal parameter (when the primal programme is to maximize, and the dual is to
minimize). This derivative property is next stated for the profit and cost optimization
programmes and their duals. All six results are obtained either by applying the
Inversion Rule (Theorem B.6.1) and the FOC (B.3.2) or (B.5.5), or by applying
the Derivative Property of the Conjugate (B.6.3)—which combines the Inversion
Rule and the FOC. The same techniques apply to the reduced programmes: see the
end of this section.
Notation and Terminology (Conjugate Function) As a superscript, the symbol #
indicates the Fenchel-Legendre convex conjugate (of a convex function), defined
by (B.2.1) in Appendix B. As a subscript, # indicates the concave conjugate
(of a concave function), defined by (B.5.1). In either position, # means the total
conjugate , i.e., the conjugate w.r.t. all of the function’s arguments (except for those
indicated as fixed). Partial conjugates w.r.t. one variable (say, the first or the second
variable) are denoted by #1 , #2 , etc.; these are defined by (B.2.6). The partial
conjugate w.r.t. the first and second variables together is denoted by #1;2 (for a
bivariate function, this means the same as #).
54 3 Characterizations of Long-Run Producer Optimum
Lemma 3.11.1 (Hotelling’s Lemma, Short Run) Assume that the production set
Y is closed. Then .y; v/ 2 @p;w …SR . p; k; w/ if and only if .y; v/ solves the short-
run profit maximization programme (3.1.6)–(3.1.7).
Proof By definition, …SR .; k; / is ı # . j YSR .k//, i.e., it is the support function of
the section of Y through k. This is a closed convex subset of Y V; so if it is also
nonempty, then
by (B.6.6). Even in the degenerate case of YSR .k/ D ;, the subdifferential and
the solution set are equal: both are Y V (since every vector is then a subgradient
of … .; k; / D 1, and since every point solves, albeit improperly,27 the then
infeasible programme (3.1.6)–(3.1.7)).
Lemma 3.11.2 (Dual of Hotelling’s Lemma, Short Run) r 2 @O k …SR . p; k; w/ if
and only if r solves the fixed-input pricing programme (3.3.6).
Proof By the definition of …SR as the optimal value of (3.3.6), and by (3.1.14),
r 2 @O k …SR . p; k; w/ , k 2 @r …LR . p; r; w/
, r minimizes …LR . p; ; w/ C h j ki
by the FOC (B.3.2). Alternatively, apply the Derivative Property (B.6.3) to conflate
the two steps.
Alternative Proof of Lemma 3.11.2 (under c.r.t.s.) If Y is a cone, this can be proved
like Lemma 3.11.1: …SR . p; ; w/ is then the inf-support function of the polar cone’s
section
Yıp;w WD fr W . p; r; w/ 2 Yı g (3.11.2)
27
See the Comment on proper and improper solutions in Sect. B.6 of Appendix B.
3.11 Shephard-Hotelling Lemmas and Their Dual Counterparts 55
Lemma 3.11.3 (Shephard’s Lemma, Long Run) Assume that the production set
Y is closed. Then .k; v/ 2 @O r;w CLR .y; r; w/ if and only if .k; v/ solves the long-run
cost minimization programme (3.1.8)–(3.1.9).
Proof Like Lemma 3.11.1, this is a case of differentiating a support function:
CLR .y; ; / is the inf-support function of the section of Y through y, so (B.6.8)
applies.
Lemma 3.11.4 (Dual of Shephard’s Lemma, Long Run) p 2 @y CLR .y; r; w/ if
and only if p solves the long-run output-pricing programme (3.3.5).
Proof Like Lemma 3.11.2, this follows from the two definitional conjugacy rela-
tionships between …LR and either the value function being differentiated in the
lemma (CLR ) or the value dual to it (CLR )—i.e., from
C LR D …#LR
1 #1
and …LR D CLR (3.11.3)
(in other words, from C LR .; r; w/ D …LR .; r; w/# and …LR .; r; w/ D
CLR .; r; w/# ) by applying the Inversion Rule (B.6.1) and the FOC (B.3.2).
Alternatively, apply the Derivative Property (B.6.3) to conflate the two steps.
Alternative Proof of Lemma 3.11.4 (under c.r.t.s.) If Y is a cone then CLR .; r; w/ is
the support function of fp W . p; r; w/ 2 Yı g, and so (B.6.6) applies.
Lemma 3.11.5 (Shephard’s Lemma, Short Run) Assume that the production set
Y is closed. Then v 2 @O w CSR .y; k; w/ if and only if v solves the short-run cost
minimization programme (3.1.10)–(3.1.11).
Proof Like Lemmas 3.11.1 and 3.11.3, this is a case of differentiating a support
function: here, CSR .y; k; / is the inf-support function of the section of Y through
.y; k/, so (B.6.8) applies.
Lemma 3.11.6 (Dual of Shephard’s Lemma, Short Run) . p; r/ 2 @y;k CSR
.y; k; w/ if and only if . p; r/ solves the output-and-fixed-input pricing pro-
gramme (3.3.4).
Proof Like Lemmas 3.11.2 and 3.11.4, this follows from the two definitional con-
jugacy relationships between …LR and either the value function being differentiated
in the lemma (CSR ) or the value dual to it (CSR )—i.e., from
#
1;2 #
C SR .y; k; w/ D …LR .y; k; w/ and …LR . p; r; w/ D CSR1;2 . p; r; w/ (3.11.4)
by applying the Inversion Rule (B.6.1) and the FOC (B.3.2). Alternatively, apply the
Derivative Property (B.6.3) to conflate the two steps.
Alternative Proof of Lemma 3.11.6 (under c.r.t.s.) If Y is a cone then C SR .; ; w/ is
the support function of f. p; r/ W . p; r; w/ 2 Yı g, and so (B.6.6) applies.
56 3 Characterizations of Long-Run Producer Optimum
Both marginal values of Type Two (the marginals of primal value w.r.t. dual
parameters and of dual value w.r.t. primal ones) are actually of Type One (are
marginals of dual value w.r.t. dual parameters and of primal value w.r.t. primal ones)
if (and only if) there is no duality gap. This is next applied, thrice, to complement
the preceding six lemmas.
Remark 3.11.7 .y; v/ 2 @p;w …SR . p; k; w/ if and only if: .y; v/ 2
@p;w …SR . p; k; w/ and …SR . p; k; w/ D …SR . p; k; w/.
Remark 3.11.8 r 2 @O k …SR . p; k; w/ if and only if: r 2 @O k …SR . p; k; w/ and
…SR . p; k; w/ D …SR . p; k; w/.
Remark 3.11.9 .k; v/ 2 @O r;w CLR .y; r; w/ if and only if: .k; v/ 2 @O r;w CLR .y; r; w/
and C LR .y; r; w/ D CLR .y; r; w/.
Remark 3.11.10 p 2 @y CLR .y; r; w/ if and only if: p 2 @y CLR .y; r; w/ and
CLR .y; r; w/ D CLR .y; r; w/.
Remark 3.11.11 v 2 @O w CSR .y; k; w/ if and only if: v 2 @O w CSR .y; k; w/ and
CSR .y; k; w/ D CSR .y; k; w/.
Remark 3.11.12 . p; r/ 2 @y;k CSR .y; k; w/ if and only if: . p; r/ 2
@y;k CSR .y; k; w/ and C SR .y; k; w/ D CSR .y; k; w/.
Since the primal and dual values are assumed to be equal only at a particular data
point (and not on a whole neighbourhood of it), Remarks 3.11.7–3.11.12 do require
a proof. This can be based on (B.3.9), i.e., on the equality of the subdifferentials of
a convex function and of its second conjugate (at a point where the two functions
are equal). This applies because the dual value (in this context, and under c.r.t.s.,
the imputed value of the given quantities) is the second conjugate of the primal
value (profit or cost) as a function of the primal parameters (the quantity data).
Likewise, the primal value is the second conjugate of the dual value as a function
of the dual parameters (the price data). For example, …SR is the second concave
conjugate of …SR as a function of k, with . p; w/ fixed. Dually, …SR is the second
convex conjugate of …SR as a function . p; w/, with k fixed. (Similarly, CLR and
CSR are the second convex conjugates of CLR and CSR as functions of, respectively,
y and .y; k/, with .r; w/ or w fixed. Dually, CLR and CSR are the second concave
conjugates of C LR and CSR as functions of, respectively, .r; w/ and w, with y or
.y; k/ fixed.) These bi-conjugacy relationships are next recorded for use in proving
Remarks 3.11.7–3.11.12.
#1;3 #1;3
Lemma 3.11.13 If Y is closed then …SR D …SR (i.e., …SR .; k; / D
…SR .; k; /## on Y W for each k 2 K).
Proof Since …SR .; k; / is by definition the conjugate of the 0-1 indicator
ı . j YSR .k//, it suffices to show that this is, in turn, the conjugate of …SR .; k; /.
Since Y is closed (and convex), the definitional relationship …LR WD ı # . j Y/ can
3.11 Shephard-Hotelling Lemmas and Their Dual Counterparts 57
be inverted to give
ı .y; v j YSR .k// WD ı .y; k; v j Y/ D …#LR .y; k; v/ (3.11.5)
WD sup .h p j yi hr j ki hw j vi …LR . p; r; w//
p;r;w
D sup h p j yi hw j vi …SR . p; k; w/
p;w
by (B.3.9).
58 3 Characterizations of Long-Run Producer Optimum
Proof of Remark 3.11.8 Fix any . p; w/, and abbreviate …SR . p; ; w/ to …. Then
… D …## (on K) by Lemma 3.11.14. So, for each k,
O .k/ , r 2 @…
r 2 @… O .k/ and … .k/ D … .k/
by (B.5.7).
Proof of Remark 3.11.9 Fix any y, and abbreviate CLR .y; ; / to C. Then C D C##
(on R W) by Lemma 3.11.15. So, for each .r; w/,
O .r; w/ , .k; v/ 2 @C
.k; v/ 2 @C O .r; w/ and C .r; w/ D C .r; w/
by (B.5.7).
Proof of Remark 3.11.10 Fix any .r; w/, and abbreviate CLR .; r; w/ to C. Then C D
C## (on Y) by Lemma 3.11.16. So, for each y,
by (B.3.9).
Proof of Remark 3.11.11 Fix any .y; k/, and abbreviate CSR .y; k; / to C. Then C D
C## (on W) by Lemma 3.11.17. So, for each w,
by (B.5.7).
Proof of Remark 3.11.12 Fix any w, and abbreviate CSR .; ; w/ to C. Then C D C##
(on Y K) by Lemma 3.11.18. So, for each .y; k/,
by (B.3.9).
When there is no duality gap, programme solutions are therefore equal to
marginal values of Type One: the dual solution is then equal to the primal value’s
derivative w.r.t. the primal parameter, and, similarly, the primal solution is the dual
value’s derivative w.r.t. the dual parameter. A pair of solutions with equal values is
therefore the same as a pair of sub/super-gradients, w.r.t. primal and dual parameters,
of just one of the two value functions (either primal or dual). Here, this means that
… and C can replace … and C in Lemmas 3.11.2, 3.11.4 and 3.11.6—which can
then be combined with Lemmas 3.11.1, 3.11.3 and 3.11.5 (respectively) to form
saddle differential systems purely in terms of either … or C (i.e., without … or C).
Similarly, … and C can replace … and C in Lemmas 3.11.1, 3.11.3 and 3.11.5—
which can then be combined with Lemmas 3.11.2, 3.11.4 and 3.11.6 to form saddle
3.11 Shephard-Hotelling Lemmas and Their Dual Counterparts 59
quantities of outputs, alone, and a version of the Dual of Shephard’s Lemma for
the short run that is limited to prices of outputs.
Lemma 3.11.21 (Hotelling’s Lemma for Short-Run Outputs Only) The follow-
ing conditions are equivalent to each other:
1. y 2 YO . p; k; w/, i.e., y yields the supremum in (3.1.13), which is …SR . p; k; w/.
2. y 2 @p …SR . p; k; w/ and CSR .y; k; w/ D …#SR 1
.y; k; w/, i.e., CSR D CSR#1 #1
at
.y; k; w/.
The equality in Condition 2 holds if CSR D CSR at .y; k; w/. Also, if …SR D …SR
at . p; k; w/ and y 2 @p …SR . p; k; w/ then y 2 @p …SR . p; k; w/.
Proof Apply the Inversion Rule (Theorem B.6.1) and the FOC (B.3.2) to CSR and its
conjugate …SR as functions of y and p (with k and w fixed); alternatively, apply the
Derivative Property (B.6.3) to conflate the two steps. This shows that Conditions 1
and 2 are equivalent.
# #
Fix any w and recall that CSR D CSR1;2 1;2 by Lemma 3.11.18. So CSR D CSR at
#1;2 #1;2 #1 #1
.y; k; w/ if and only if CSR D CSR at .y; k; w/, and then a fortiori CSR D CSR at
.y; k; w/ by Remark B.2.1.28
Finally, recall that …SR …SR everywhere (on P K W). So if …SR D …SR at
. p; k; w/ and y 2 @p …SR . p; k; w/ then also y 2 @p …SR . p; k; w/ by the subgradient
inequality (B.3.1).
Lemma 3.11.22 (Dual of SR Shephard’s Lemma for Outputs Only) The follow-
ing conditions are equivalent to each other:
1. p 2 PL .y; k; w/, i.e., p yields the supremum in (3.3.8), which is C SR .y; k; w/.
#1 #1
2. p 2 @y CSR .y; k; w/ and …SR . p; k; w/ D C#SR 1
. p; k; w/, i.e., …SR D …SR at
. p; k; w/.
The equality in Condition 2 holds if …SR D …SR at . p; k; w/. Also, if CSR D CSR
at .y; k; w/ and p 2 @y CSR .y; k; w/ then p 2 @y CSR .y; k; w/.
Proof Being a “mirror image” of Lemma 3.11.21, this is proved by the same
arguments, with CSR .y/, …SR . p/ and …SR . p/ in place of …SR . p/, CSR .y/ and
CSR .y/, respectively. To spell this out, apply the Inversion Rule (Theorem B.6.1)
and the FOC (B.3.2) to …SR and its conjugate C SR as functions of p and y (with k
and w fixed); alternatively, apply the Derivative Property (B.6.3) to conflate the two
steps. This shows that Conditions 1 and 2 are equivalent.
#1;3 #1;3
Fix any k and recall that …SR D …SR by Lemma 3.11.13. So …SR D …SR at
#1;3 #1;3 #1 #1
. p; w/ if and only if …SR D …SR at . p; w/, and then a fortiori …SR . p; w/ D
…SR . p; w/ by Remark B.2.1.
#1
28
For another proof of this, note that: (i) by (3.3.8), CSR D …SR …#SR1 D CSR
#1 #1
by (3.1.13), with
#1 #1
the inequality holding because …SR …SR , and (ii) CSR CSR by (B.2.4) without the middle
#1 #1
term. So CSR CSR CSR everywhere (and it follows that all three are equal if the outer two
are).
62 3 Characterizations of Long-Run Producer Optimum
Finally, recall that CSR CSR everywhere (on Y K W). So if CSR D CSR
at .y; k; w/ and p 2 @y CSR .y; k; w/ then also p 2 @y CSR .y; k; w/ by the subgradient
inequality (B.3.1).
Corollary 3.11.23 Assume both that CSR D CSR at .y; k; w/ and that …SR D …SR
at . p; k; w/. Then the following conditions are equivalent to one another:
1. y 2 YO . p; k; w/, i.e., y yields the supremum in (3.1.13), which is …SR . p; k; w/.
2. p 2 @y CSR .y; k; w/.
3. y 2 @p …SR . p; k; w/.
4. p 2 PL .y; k; w/, i.e., p yields the supremum in (3.3.8), which is C SR .y; k; w/.
5. y 2 @p …SR . p; k; w/.
6. p 2 @y CSR .y; k; w/.
Proof Lemmas 3.11.21 and 3.11.22 state that Conditions 1, 3, 4 and 6 are
equivalent. As for Conditions 2 and 5, recall from the Proofs that these are the FOCs
for the optima in Conditions 1 and 4.
bundle y can always be reinterpreted as the bundle of all the variable commodities
(i.e., the outputs and the variable inputs).
For now, then, Y is a polyhedral cone in the finite-dimensional space Y K D
RT Rˆ , where T and ˆ are the sets of the output and the fixed-input commodities.
Its polar cone, Yı , is therefore a finitely generated convex cone in the price space
P R D RT Rˆ . It can be represented as the sum of a linear subspace spanned by
a finite set G 00 and a line-free convex cone generated by a finite set G 0 , i.e.,
for some positively independent, finite set G 0 and another finite set G 00 (which can be
chosen to be linearly independent).29 The generators (G 0 ) and the spanning vectors
(G 00 ) of the polar cone Yı can serve as the rows of partitioned matrices ŒA0 B0 and
ŒA00 B00 that represent the original cone as
˚
Y D .y; k/ 2 RT Rˆ W A0 y B0 k 0 and A00 y B00 k D 0 . (3.12.1)
Its optimal value is denoted by …SR . p; k/, abbreviated to … . p; k/. As in Sect. 3.3,
the vector datum k is called an intrinsic primal parameter, and its increment k is
an intrinsic perturbation of (3.12.2)–(3.12.4).
The corresponding standard parametric LP has primal parameters s0 and s00 ,
0 00
ranging over RG and RG , instead of the B0 k and B00 k in (3.12.3)–(3.12.4). Its
optimal value is the standard primal value , denoted by …Q . p; s/, where s D .s0 ; s00 /.
29
Although it follows that Yı is the convex cone generated by G 0 [ G 00 [ .G 00 /, it is better to
keep G 0 and G 00 separate when it comes to parameterizing the programme (3.12.2)–(3.12.4) in the
standard way: for this purpose, an equality constraint should not be converted to a pair of opposite
inequalities. To do so would complicate the dual solution by making it nonunique and unbounded:
00
a primal equality constraint (say a y D 0) may have a unique multiplier O , but if it were replaced
by apair ofinequalities (a y 0 and a y 0), then a corresponding multiplier pair would be
0 0 0 0 00
any O 1 ; O 2 0 with O 1 O 2 D O , i.e., it would be any point of a half-line. Its unboundedness
expresses the fact that the programme would become infeasible if one inequality constraint of the
pair were tightened without relaxing the other by the same amount (i.e., if the constraints were
perturbed to a y 1 and a y 2 for 1 < 2 ).
30
Formally, A0 and B0 are the G 0 T and G 0 ˆ matrices with entries A0gt D gt and B0g D g for
t 2 T, 2 ˆ and g 2 G 0 (and the same goes for 00 instead of 0 ). For Farkas’s Lemma, see, e.g.,
[12, 2.2.6], [42, 22.3.1], [45, 6.45] or [48, 4.19].
64 3 Characterizations of Long-Run Producer Optimum
So, by definition,
Q . p; Bk/ B0
… . p; k/ D … where B WD (3.12.5)
B00
for every . p; k/. The standard perturbation consists in relaxing (or tightening) the
0 00
inequality constraints by adding an arbitrary vector s D . s0 ; s00 / 2 RG RG
to the r.h.s. of (3.12.3)–(3.12.4), i.e., it uses a separate scalar increment for each
constraint. This produces the standard dual of (3.12.2)–(3.12.4), which is: Given
the same . p; k/ 2 RT Rˆ ,
0 00
minimize T Bk D 0T B0 k C 00T B00 k over D 0 ; 00 2 RG RG
(3.12.6)
subject to: 0 0 (3.12.7)
p D AT WD A0T 0 C A00T 00 (3.12.8)
Q . p; Bk/ D … . p; k/
… for every . p; k/ . (3.12.9)
In other words, the standard dual LP has the same value as the intrinsic dual; here,
O
the two duals are (3.12.6)–(3.12.8) and (3.3.13)–(3.3.14). For their solution sets, †
O it follows that
and R,
ˇ
Q . p; s/ˇˇ
RO . p; k/ D @O k … . p; k/ D BT @O s … O . p; Bk/
D BT † (3.12.10)
sDBk
n o
WD BT W 2 † O . p; Bk/
31
The identity (3.12.9) reduces to (3.12.5) when the primal and dual values are equal, i.e., when
…Q D … Q and … D … at . p; k/. This always applies to (feasible) finite LPs, but not always to
infinite LPs. To prove (3.12.9) without relying on absence of a duality gap, note that the change
of variables from r to by r D BT transforms (3.3.13)–(3.3.14) into (3.12.6)–(3.12.8). This is
detailed in the first Comment after (3.12.14). The argument extends to infinite LPs (and it applies
also when there is a duality gap).
3.12 Duality for Linear Programmes with Nonstandard Parameters in Constraints 65
by applying the Chain Rule to (3.12.9),32 and by using (twice) the identity of the
dual solution and the marginal value of Type Two.33 Thus the intrinsic dual solution
O is expressed as the linear image of the standard dual solution (†)
(R) O under the
adjoint (BT ) of the linear operation that maps the intrinsic to the standard primal
parameters (s D Bk).
When … D … at . p; k/, the marginal value is actually of Type One by
Remark 3.11.8, i.e.,
@O k … . p; k/ D @O k … . p; k/ D BT †
O . p; Bk/ . (3.12.11)
This always applies to finite LPs because their primal and dual values are equal,
unless both programmes are infeasible (in which case their values are oppositely
infinite).34 If additionally the standard dual solution is unique, then so is the intrinsic
O is a singleton fO g), and
one (i.e., RO is a singleton fOrg if †
rO . p; k/ D r k … . p; k/ D BT O . p; Bk/ . (3.12.12)
This gives the intrinsic dual solution (Or)—equal to the marginal values of the
generally nonstandard intrinsic parameters (k)—in terms of the standard dual
solution ().
O
A variant of the standard dual is obtained by including (in addition to the standard
dual variable ) also the intrinsic dual variable r paired with k. Then r is constrained
to equal BT (and is thus wholly dependent on ). The objective, T Bk, can be
rewritten concisely as r k. This produces the following LP: Given . p; k/ 2 RT Rˆ ,
0 00
minimize r k over r 2 Rˆ and D 0 ; 00 2 RG RG (3.12.13)
subject to: 0 0; p D AT and r D BT . (3.12.14)
This may be called the inclusive standard dual—formally an LP for both r and .
It is the dual that derives from simultaneous standard and intrinsic perturbations,
i.e., from perturbing Bk on the r.h.s. of (3.12.2)–(3.12.4) to s C B .k C k/. Its
solution gives explicitly both sets of marginal values (O and rO ), but in substance
it is equivalent to the standard dual solution O alone (since rO D BT ).
O It can be
more convenient to use a partly inclusive form of the standard dual—one which
includes only some of the intrinsic dual variables (r), leaving out those entries of
r which correspond to “the simplest” columns of B—e.g., to the columns with 0-1
32
For the Chain Rule for subdifferentials, see, e.g., [4, 4.3.6 a], [32, 4.2: Theorem 2], [42, 23.9] or
[44, Theorem 19].
33
First noted at the end of Sect. 3.5, the identity is detailed in Lemma 3.11.2.
34
See, e.g., [11, 5.1 and 9.1] or [44, Example 1’, p. 24] for proofs based on the simplex algorithm or
on polyhedral convexity, respectively. This is not so with a pair of infinite LPs: both can be feasible
without having the same value (i.e., the primal and dual values can both be finite but unequal). See
Appendix A for an example.
66 3 Characterizations of Long-Run Producer Optimum
entries as in the next Comment. For example, the programme of valuing the hydro
inputs (5.2.37)–(5.2.43) includes the TOU shadow price of water but not the total
capacity values rSt and rTu , which are simply the totals of the standard dual variables
St and Tu .
Comment (on Standard and Intrinsic Perturbations) If B were the unit matrix
I, the two perturbation schemes would obviously be the same (and s could
be renamed to k). This would be so if the short-run production constraints
corresponded, one-to-one, to the fixed inputs, i.e., if Y were defined by a system
of inequalities (or equalities) of the form .Ay/ k , one for each 2 ˆ. But
such a correspondence generally fails to exist, for three reasons. First, two fixed
inputs may appear in one constraint (say a y k1 C k2 ). Second, a constraint
may involve only the outputs (a y 0, e.g., yt 0). Third, each fixed quantity
k may impose more than one constraint on y (say .Ay/1 k , .Ay/2 k , . . . ).
Indeed, this is so whenever k is a capacity: staying constant over a time period, it
is a scalar but it imposes as many inequality constraints as there are time instants
(e.g., yt k for each t).35 In such a case, B is a 0-1 matrix whose unit entries
appear just once in a row, but more than once in a column. When additionally k
T
is a scalar, B is the single column 1 : : : 1 ; and an intrinsic perturbation of the
T
constraint system Ay k k : : : relaxes all the constraints by the same amount,
T
to Ay k C k k C k : : : . By contrast, a standard perturbation relaxes each
T
constraint by a different amount, to Ay k C s1 k C s2 : : : . In this sense,
the standard perturbation scheme is the finest, whereas the intrinsic perturbation
scheme is the coarsest (with this particular B). Once the scalar k is identified with
the vector .k; k; : : :/, the standard value function … Q . p; / becomes an extension of
the intrinsic value function … . p; / from the subspace of constant tuples to all of
0 00
RG RG (with G 00 empty if there is no equality constraint), and the intrinsic
dual
solution(a scalar) is simply the total sum of the standard dual solution, i.e.,
rO D 1 : : : 1 O D O 1 C O 2 C : : :. In other words, the scalar parameter’s marginal
value is the sum of the marginal values of relaxing all the constraints in which
it appears. This arises in the peak-load pricing application: total capacity values
are the integrals of the rent flows over the period—see (5.2.10), (5.2.23)–(5.2.24),
and (5.2.44)–(5.2.45). Also, since … Q is an extension of …, it can be convenient to
use the same letter k as the second variable of both functions (i.e., to use k instead
of the s in … Q . p; s/), provided that it is always made clear whether k is a scalar
or a vector. This is done here in the context of hydro and energy storage (where s
signifies the water or energy stock and is not a parameter).
Comment (Derivation of the Standard Dual LP from the Intrinsic Dual) The
standard dual variables D . 0 ; 00 / are the coefficients in the representation of
a price vector . p; r/ 2 Yı as a linear combination of the polar cone’s generators
35
Also, the nonnegativity constraint on k will make it appear a second time even if k imposes
just one constraint on y (i.e., 0 k in addition to a y k for some a ¤ 0).
3.12 Duality for Linear Programmes with Nonstandard Parameters in Constraints 67
and spanning vectors—i.e., a combination of the rows of the matrices ŒA0 B0 and
ŒA00 B00 . That is, the standard dual LP (3.12.6)–(3.12.8) can be obtained from the
intrinsic dual (3.3.13)–(3.3.14) by the change of variables, from r to , defined by:
r D BT WD B0T 0 C B00T 00 , for such that p D AT WD A0T 0 C A00T 00 and
0 0.36
Comment (Representation of the Intrinsic Dual as an LP with the Same
Variables r) The intrinsic dual can be formulated as an LP also without changing
the decision variables from r to . This requires rewriting the parametric equations
of Yı —which are pT D T A and rT D T B for some with 0 0—as ordinary
linear equations and inequalities in the variables p and r. This is straightforward
when p and r fully determine , i.e., when the matrix M WD ŒA B has a right
inverse—or, in other words, when MM T is invertible.37 But in general the standard
dual solution can be “finer” than the intrinsic one, i.e., p and r need not determine
(although determines both r and p). However, the parametric linear equations (of a
finitely generated convex cone or, more generally, a convex set) can actually always
be converted to ordinary equations and inequalities (and vice versa ): a procedure
is given in, e.g., [11, Chapters 16 and 18]. In geometric terms, it means switching
between two representations of a polyhedron: (i) by vertices and rays, and (ii) as an
intersection of half-spaces and hyperplanes.
Comments (Methods of Solving the FFE System in the Form of Linear
Inequalities)
• Once both the primal and the dual have been formulated as LPs, the FFE
Conditions form a finite system of linear equations and inequalities in finitely
many variables. With the standard dual (3.12.6)–(3.12.8) of the LP (3.12.2)–
(3.12.4), the FFE Conditions on y and (given p and k) are: Ay Bk 0,
pT D T A and 0 0, and pT y D T Bk (or, equivalently, pT y T Bk). With
the intrinsic dual (3.3.13)–(3.3.14), the FFE Conditions are the Complementarity
Conditions (3.1.5) on the quantity and price variables y and r (given p and k),
which can be formulated also as a system of linear equations and inequalities:
those describing Y (viz., (3.12.3) and (3.12.4)), those describing Yı as in the
preceding Comment, and the equality of values (p y D r k). Like any linear
system, that of the FFE Conditions can be solved either (i) directly by the
Fourier-Motzkin elimination (which gives all the solutions) or (ii) indirectly by
converting it into an auxiliary LP and applying the simplex method (or another
36
First, the intrinsic dual’s constraint . p; r/ 2 Yı is rewritten as: p D AT and r D BT for
some D . 0 ; 00 / with 0 0. Then is made an explicit decision variable alongside r (and so
the existential quantifier on is dropped); this produces the inclusive standard dual LP (3.12.13)–
(3.12.14). Finally, r is replaced by BT ; this produces the standard dual LP (3.12.6)–(3.12.8).
1
37
In this case, the matrix
M R WD M T MM T is a right inverse of M. The parametric equations
of Yı imply that T D pT rT M R , and, after partitioning M R into ŒR 0 00
R to match the
partition
D . 0 ; 00 /, the required system of equations and inequalities is: pT rT M R M I D 0 and
T T 0
p r R 0.
68 3 Characterizations of Long-Run Producer Optimum
LP algorithm) to find at least one solution, and thus also the common value of
the original programme and its dual; any other solutions can then also be found.
For the auxiliary LP, see, e.g., [11, (16.2), p. 240].
• However, it seems somewhat better to deal with the original LP with its standard
dual than to solve the system of FFE Conditions by either method. First, Fourier-
Motzkin elimination is far less efficient than the simplex method (applied to the
auxiliary LP); this is noted in, e.g., [11, p. 242]. Second, the original LP is smaller
in size than the auxiliary LP.38
Finally, this analysis and the expression (3.12.12) for the intrinsic dual variables
(r) in terms of the standard ones () can be extended to infinite LPs. This requires
using suitable cones in infinite-dimensional spaces of variables and parameters to
formulate infinite systems of constraints on, generally, an infinity of variables. Such
a framework is provided in, e.g., [12, 4.2], [36, 7.9] and [44, Examples 4, 4’,
4”]. The assumptions made here to adapt it are not the weakest possible; they are
selected for their simplicity and adequacy to the applications in Sect. 5.2. The output
and fixed-input spaces, Y and K, are now taken to be general Banach spaces, i.e.,
complete normed spaces (instead of RT and Rˆ ). Their norm-duals, Y and K ,
serve as the corresponding price spaces, P and R. For the primal programme of SRP
maximization, Y is the primal-variable space paired with the dual parameter space
P, and K is the primal-parameter space paired with the dual-variable space R. The
production cone Y is given by (3.12.1) in terms of two norm-to-norm continuous
linear operations: (i) A0 W Y ! L and B0 W K ! L, whose common codomain L is a
Banach lattice (with a vector order and the corresponding nonnegative cone LC ),
and (ii) A00 W Y ! X and B00 W K ! X, whose common codomain X is a Banach space.
0 00
The spaces L and X replace RG and RG as the spaces for standard perturbations
( s and s ). Their norm-duals, L and X , serve as the spaces for standard dual
0 00
variables ( 0 and 00 ). It is best to keep L and X small, but obviously L must contain
the ranges both of A0 on Y and of B0 on K (similarly, X must contain both A00 Y and
B00 K ).
38
To see this, let the original primal LP be to maximize p y over y 2 Rn subject to Ay k, given
arbitrary vectors p 2 Rn and k 2 Rm , and given an m n matrix A (i.e., assume for simplicity that
A D A0 , B D B0 D I, and so D 0 D r and the standard and the intrinsic duals are the same).
The dual LP is to minimize r k over r 0 subject to rT A D pT . The FFE (Complementarity)
Conditions on .y; r/ are: Ay k, r 0, rT A D pT and p y r k (or, equivalently, p y D r k).
This is a system with n C m variables and 2m C 2n C 1 inequalities (counting an equality as two
inequalities). Its auxiliary LP has n C m C 1 decision variables (viz., y, r and an artificial variable,
say z 0, as the minimand, whose minimum value is zero if and only if the FFE system is soluble)
and 2 .m C n C 1/ inequality constraints (viz., z 0 and all the complementarity inequalities but
with z subtracted from the lesser side, i.e., p y r k z, etc.): see [11, (16.2), p. 240]. So the
auxiliary LP has one more variable and one more constraint than the original primal and dual LPs
together . Solving the auxiliary LP by a primal-dual algorithm (such as the simplex method) gives
a solution to the original LP “in duplicate”.
3.12 Duality for Linear Programmes with Nonstandard Parameters in Constraints 69
As for the choices of topologies, these must be consistent with the pairings of
spaces. Furthermore, the norm topology has to be put on the primal parameter
space L if Slater’s Condition as generalized in [44, (8.12)] is to be met for the SRP
programme (3.12.2)–(3.12.4), i.e., if a y is to exist such that A0 y B0 k 2 int .LC /
and A00 y B00 k D 0X . Topologies on Y, K, L and X must make the maximand upper
semicontinuous (u.s.c.) and the constraint relations closed; here, this means making
h p j i, A and B continuous. So the norm topologies on Y (the primal-variable space)
and on K, L and X (the primal-parameter spaces) will do. On the dual-variable spaces
K , L and X , the weak* topologies will do.39 On Y (the dual parameter space),
the Mackey topology m .Y ; Y/ is the best choice if continuity of the dual value
function is sought. When Y has a Banach predual Y 0 , it can be useful to pair Y also
with Y 0 as a dual parameter space that is generally smaller than Y ; the restriction
of m .Y ; Y/ to Y 0 is the norm topology of Y 0 . The pairing of Y with Y 0 is adequate
when p 2 Y 0 , but not when p 2 Y n Y 0 .
For a specific Y, its original definition may be of the form (3.12.1)—in which
case the linear operations A and B can simply be read off. This is so in the application
to electricity supply: the production sets (5.1.1), (5.1.3) or (5.1.4) are all of the
form (3.12.1).40 There is an alternative construction of A and B which sometimes,
but not always, will give the same result as the original definition (when both
approaches are available). Namely, A0 and B0 (with no A00 or B00 , i.e., with the zero
space as X) can be constructed from a weakly* compact convex base,
, for the
polar cone Yı P R D Y K ; such a base exists if and only if Y is solid (i.e.,
has a nonempty for the norm on Y K: see, e.g., [3, Theorem 3.16]. An
interior)
interior point yS ; kS defines the base
˚ ˝ ˛ ˝ ˛
WD . p; r/ 2 Yı W p j yS r j kS D 1 . (3.12.15)
Such a
(or ext
if it is closed) can serve as a replacement for the finite set G 0
that generates Yı when Y is a solid polyhedral cone in a finite-dimensional space.
The Banach lattice of all weakly* continuous functions on
, denoted by C .
/,
0
replaces RG and serves as the codomain (L) for the operations A0 and B0 . These are
39
The weak topologies do not enter the analysis explicitly, but they make the adjoint operators
continuous: see, e.g., [18, 16C].
40
The output space is Y D L1 Œ0; T, which has a Banach predual Y 0 D L1 Œ0; T. The fixed-
input space K depends on the technique: it is either R for a thermal technique, or R2 for pumped
storage, or R2 L1 Œ0; T for hydro. As for L (the space of standard perturbations of the inequality
constraints), it is either L1 Œ0; T or its Cartesian product with C Œ0; T when, in the case of an
energy storage technique, there are reservoir constraints in addition to the generation constraints.
And the balance constraint of a storage technique has R as X (the space of standard perturbations
of the equality constraint).
70 3 Characterizations of Long-Run Producer Optimum
specified by41
A0 y B0 k . p; r/ WD h p j yi hr j ki for . p; r/ 2
. (3.12.16)
41
Formula (3.12.16) adapts [12, p. 154, line 11 f.b.], where the construction is mistakenly proposed
as a possible way of dealing with a non-solid cone (in such a case the polar cannot have a compact
base, so the analysis does not apply). The construction can, however, be extended to the case that
Y is only relatively solid, i.e., has a nonempty interior in the linear subspace Y Y (assumed to
be closed in Y K); the polar Yı is then the sum of the annihilator .Y Y/? and a cone with a
compact base.
3.12 Duality for Linear Programmes with Nonstandard Parameters in Constraints 71
where "t is the Dirac measure at t (i.e., a unit mass concentrated at the single
point t). With
replaced by ext
' Œ0; T, Formula (3.12.16) gives
A0 y B0 k ."t ; 1/ WD h"t j yi k D y .t/ k for t 2 Œ0; T
and thus, upon identifying each "t with t itself, it reproduces the original
operations A0 and B0 exactly .
– This is no longer quite so once the space Y in (3.12.17) is enlarged from
C Œ0; T to L1 Œ0; T, the space of all essentially bounded
functions.
Although
the
constructed from the same interior point 0Œ0;T ; 1 is still the
nonnegative part of the unit sphere, the sphere is now that of L1 rather than
of M as in (3.12.18). In either case, its extreme points can be characterized
as scalar-valued lattice-homomorphisms on L1 or C (into R) of unit norm,
and also as nonzero multiplicative linear functionals (i.e., scalar-valued
algebra-homomorphisms) on L1 or C: see, e.g., [2, 12.27] and [46, 11.32],
respectively. More precisely, ext
D H f1g, where H is the set of all such
homomorphisms on either L1 or C. But the homomorphisms on L1 Œ0; T are
not as simple as those on C Œ0; T, which, being the Dirac measures, correspond
to the points of Œ0; T. In the case of L1 , H is an extremally disconnected
compact subset of L1 (with the weak* topology), and C .H/ is isomorphic
(both as a normed lattice and as a normed vector algebra) to L1 Œ0; T. In
other words, the construction amounts to representing the equivalence classes
of bounded measurable functions on Œ0; T as continuous functions on another,
much more complicated, compact set H. The “almost everywhere” inequality
constraint, y .t/ k for a.e. t, is thus replaced by the system of scalar
inequalities h p j yi k for every p 2 H. Since the indexing set H is far
from simple, such a reformulation may not be worthwhile.
Chapter 4
Short-Run Profit Approach to Long-Run
Market Equilibrium
The preceding characterizations of long-run producer optimum are used here for the
short-run approach to long-run general equilibrium (LRGE). This approach means
that the capital inputs k are kept fixed at the stage of calculating the equilibrium in
the products’ market. The variable-input prices w are assumed to be fixed throughout
the analysis (although this is not at all essential, and w might instead be determined
in equilibrium just like the output prices p). This still leaves two alternative ways to
handle the supply side of the short-run general equilibrium (SRGE) problem, and
hence two varieties of the short-run approach:
• In the short-run profit approach , the output and variable-input quantities yO
and v, L and the fixed-input values rO , are derived from any given p, k and w
(usually, though not always, by solving the SRP problem (3.1.6)–(3.1.7) and its
dual (3.3.6), or (3.3.13)–(3.3.14) under c.r.t.s.). The supply yO . p; k; w/ is then
equated to demand xO . p/ to determine the SRGE price system p?SR .k/, which
depends also on w. This stage corresponds to the inner loop in Fig. 4.1, if an
iterative method (such as the Walrasian tâtonnement) is used to solve the demand-
supply equation for p. The capital inputs’ marginal values rO p?SR .k; w/ ; k; w ,
imputed at the SRGE prices, are then equated to their given , fixed rental prices rF
? F
to determine, by solving for k, the
LRGE k r ; w . This gives also the
capacities
LRGE price system p?LR rF ; w D p?SR k? rF ; w ; w . This stage corresponds to
the outer loop in Fig. 4.1, if an iterative method is used to solve the price-value
equation for k.
• In the alternative short-run cost approach , the variable-input quantities v, L and
the shadow prices for outputs and fixed inputs—i.e., a typically nonunique p 2
PL .y; k; w/ with the associated, typically unique rO . p; k; w/—are derived from any
given y, k and w (usually by solving the SRC problem (3.1.10)–(3.1.11) and its
Fig. 4.1 Flow chart for an iterative implementation of the short-run profit approach to long-
run market equilibrium. For simplicity, all demand for the industry’s outputs is assumed to be
consumer demand that is independent of profit income, and all input prices are fixed (in terms of
the numeraire). Absence of duality gap and existence of the optima (Or , yO ) can be ensured by using
the results of Sects. 6.1 to 6.4
dual (3.3.4), or (3.3.9)–(3.3.10) under c.r.t.s.). To find the SRGE by this approach,
inverse demand is then required to equal one of the typically nonunique output
price systems that solve the short-run output-pricing programme in (3.4.7), which
is a subprogramme of (3.3.4). Its solution set PL .y; k; w/ consists essentially of
SRMCs: see (4.1.3) for details. Finally, the LRGE capacities, and hence also the
output prices, are found just as in the profit approach.
In principle, the duality theory of convex programming can be brought to bear
however the commodities are divided into the “variable” quantities with given prices
4.1 Outline of the Short-Run Approach 75
and the “fixed”, as-yet unpriced quantities: in studying the producer optimum, the
roles of prices and quantities are formally symmetric. At an abstract level, therefore,
there is no reason to prefer any particular programme pair or the associated
representation of the technology by the value functions (…SR , CLR or CSR , etc.).
But usually the classification of commodities as “fixed in the short run” is not
arbitrary and nominal but real and objective: these are the capital goods and the
natural resources. Their quantities (k) must be taken as known in any short-run
analysis. In addition, also some of the quantities to be determined in the SRGE,
such as the outputs (y), might be taken as known at the earlier stage of finding
the short-run producer optimum and the shadow prices: this would mean solving
the SRC programme (for v) together with its dual (for p and r). But this would be
analytically disadvantageous because, when the capital inputs (k) impose capacity
constraints on a cyclically varying output (y), this SRC approach results in dual
solutions so indeterminate that they form an unbounded set: if not only r but also p
is an unknown, then nearly nothing can be said about the capacity charges (which
are terms of p, and give r as their total over the cycle).
Another disadvantage of the SRC approach—one that emerges only at the equi-
librium stage, and especially in the case of an industry with multiple producers—is
that the cost approach entails working out the inverse supply map (PL ) of every
individual producer and equating each of these to the inverse demand to find the
SRGE output bundle (y?SR ) of each producer, from the inclusion a.k.a. “generalized
equation” (4.1.2) below. This is usually much harder than simply to add up all
the (direct) supply maps (YO ), equate their sum to demand, and solve (4.1.1) for
the single market price system (p?SR )—which is what the SRP approach requires.
Also, unlike the typically multi-valued inverse supply map (PL ), the supply map
may be single-valued (Oy )—in which case the inclusion (4.1.1) to be solved for p is
an ordinary equation.
In summary: it is better not to fix any more quantities than is necessary—and
this means using the SRP rather than the SRC approach. The profit approach is
likely to be more workable because it has two advantages over the cost approach:
(i) determinacy, or uniqueness, of solutions to the short-run producer problem and
its dual, and (ii) reduction of the number of unknowns in the subsequent equilibrium
problem. Both are detailed next.
The first of these advantages is simply the convenience of dealing mostly with
single-valued maps rather than multi-valued correspondences. Solutions for . p; r/
to the dual (3.3.4) of the SRC problem are typically nonunique: indeed, the set of
optimal . p; r/’s is unbounded. Typically this is because, in pure SRC calculations,
the capacity premium is completely indeterminate (except when it vanishes because
there is excess capacity). But the r associated with a particular p can still be unique,
and so can y and v (as has been tacitly assumed by using the notation rO and yO in
describing the short-run approach). That is, solutions for r and .y; v/ to the SRP
problem (3.1.6)–(3.1.7) and its dual (3.3.6) can both be expected to be unique or, at
the very least, to form bounded sets. This can be illustrated with an elementary but
instructive example. Suppose for simplicity that there is no variable input, and that
76 4 Short-Run Profit Approach to Long-Run Market Equilibrium
In the profit approach (given p and k), both inclusions are useful in solving
this system for y and r. But in the cost approach (given y and k), the first
inclusion restricts only the data—so, when it is met, ˚ it is of no2 help at all in
solving˚for p and r. The simplest
example is Y D .y; k/ 2 R W y D k ; then
Yı D . p; r/ 2 R2 W p D r . In the cost approach the level of . p; r/ is therefore
indeterminate, but in the profit approach both solutions are unique: .Oy; rO / D .k; p/.1
This pattern is present also in complex applications such as peak-load pricing with
storage, in which the optimum rO . p; k; w/ or yO . p; k; w/ is shown to be unique if
the TOU price p is, respectively, a continuous or plateau-less function of time: see
Sect. 5.2 here and, for detailed analysis, [21, 23, 27] and [30].
The other—and the more significant—advantage of the SRP approach over the
SRC approach emerges at the equilibrium stage. Usually there is a number of
producers, each with his own technology Y , for 2 ‚. In the profit approach,
the SRGEP is found by equating the demand xO . p/ to the profit-maximizing total
output yO . p; k ; w/ and solving for p; when the optimal output is nonunique,
one solves for p the inclusion
X
xO . p/ 2 YO . p; k ; w/ (4.1.1)
2‚
where YO is the solution set for the reduced SRP programme in (3.1.13) and (3.2.2).
For comparison, the cost approach requires solving, for the output bundles .y /, the
inclusion
!
X \
pQ y 2 PL .y ; k ; w/ (4.1.2)
2‚ 2‚
where pQ is the inverse demand map and PL .y ; k ; w/ is the solution set for the short-
run output-pricing programme in (3.4.7)—i.e., PL is essentially the same as @y CSR
,
the multi-valued SRMC of an individual producer : see (4.1.3). This route is likely
to be more difficult because, with multiple producers, it means having to solve for a
1
When there are variable inputs whose cost-minimizing quantities vL are known functions of the
data .y; k; w/, the condition .y; k; v/ 2 Y in (3.1.5) boils down to .y; k; vL .y; k; w// 2 Y,
which is again a pure restriction on the data with no information about the unknowns p and
r. Of course, the profit approach would have a similar comparative weakness in the condition
. p; r; w/ 2 Yı if the fixed-input values rO were easily calculated functions of the data . p; k; w/. But
the programme that we take to be readily soluble, without using duality, is the SRC programme
(for v), and not the dual of the SRP programme (for r).
4.1 Outline of the Short-Run Approach 77
number of variables .y / instead of the single variable p, as well as having to work
out and intersect the price sets PL to start with. And these are large, unbounded
sets (whenever the fixed inputs impose capacity constraints).
Comments (The Relative Complexity of the Cost Approach)
T
• It is not easy even to identify those output allocations .y /2‚ with PL ¤
; in (4.1.2), since this involves splitting the industry’s total output among the
plants in a cost-minimizing way, which can be a difficult problem—known in
the context of electricity generation as optimal system despatch. Henceforth,
means a particular type of plant rather than a producer (who may own plants of
various types), but actually there is no loss of generality in identifying producers
with plant types: see the remarks about T (4.2.4).
• To see in detail that a .y / with L
P ¤ ; is necessarily a cost-minimizing
P 2‚
split of y , note first that
if CSR D CSR
at .y ; k ; w/ then PL @y CSR
at .y ; k ; w/ (4.1.3)
i.e., it is the infimal convolution of the individual plants’ operating cost functions
CSR .; k ; w/, abbreviated
T here to C . With 4 denoting the convolution
P operator,
one has p 2
P @C
P
.y / if and only if both p 2 @ 4 C y and
4 C y T D C .y /: see, e.g., [36, 6.6.3 and 6.6.4]. The “only if”
part shows that if @CP .y / ¤ ;, then .y / 2‚ is a cost-minimizing split of
the industry’s total output y among the plants with the given capacities .k /
and technologies .Y /.
• This means that competitive profit maximization, by the choice of the outputs
.y / at a common output price p, leads to an optimal (cost-minimizing) allocation
of the total output among the plants. Thus the decentralized, plant-by-plant
determination of the industry’s total output (given a common output price p)
bypasses the problem of cost-minimizing allocation of a given total output yTo ,
which is usually much more complex than the individual profit-maximizing
operation problems. For example, cost-minimizing despatch of a hydro-thermal
electricity-generating system necessitates solving a CP that has no simple form
L
nEven if P @y CSR at .yo ; k ; w/, the two sets have the same intersection with the set
2
p W …SR D …SR at . p; k ; w/ , by Corollary 3.11.23.
78 4 Short-Run Profit Approach to Long-Run Market Equilibrium
for either the primal or the dual solution: see the policy construction in [35,
pp. 201–219]. By contrast, the problem of profit-maximizing operation for a
hydro plant (or a pumped-storage plant) can be set up as an LP whose solution
has a relatively simple structure: see Sect. 5.2 here, [21] or [27, Sections 4 and 8],
and [23] or [30, Sections 4 and 8].
The above outline of either variety, SRP or SRC, of the short-run approach
assumes the use of, respectively, the SRP or the SRC Optimization System (possibly
in the split form). Of the optimization systems, this is the one directly suited to the
purpose; and when the technology is given by a production set—as in an engineer’s
description—there may be no tractable formulae for the value functions, and hence
no usable alternative among the differential systems. A differential system is likely
to be useful only when each of the profit or cost functions it uses is either easy to
calculate (by solving the relevant programme), or is simply given as a definition
of the technology (as in the econometric uses of duality). These remarks can be
expanded as follows.
Comments (on Choosing a System for a Short-Run Approach)
• What defines a particular approach to the producer problem is which of its
price and quantity variables are treated as known and which as unknown . With
three groups of commodities, there are eight (23 ) possibilities: the knowns-
unknowns patterns of the SRP approach is . p; k; w/-.y; r; v/, whilst that of the
SRC approach is .y; k; w/-. p; r; v/. Either approach may use its “own” (SRP or
SRC) Optimization System, but alternatively it might use the LRC Optimization
System for the same purpose—viz., to determine r and v and either y or p from
any known k, w and either p or y (thus solving not a long-run problem, but a short-
run profit or cost problem with its dual). Indeed, either variety of the short-run
approach may use whichever of all the equivalent systems is most convenient: in
principle, it need not matter whether producer optimum is characterized in terms
of short or long run, profit or cost, programme solutions or marginal values.
• Within the optimization systems, every choice leads to exactly the same analysis
when duality is used: all the systems lead to the same FFE Conditions (viz.,
the Complementarity Conditions (3.1.5)) and, also, to the same Kuhn-Tucker
Conditions (once the constraint sets Y and Yı are represented by systems of
inequalities).3 When analyzed by either of the two duality methods (Kuhn-
Tucker’s or FFE), all the optimization systems become therefore identical—but
even so it simplifies the terminology to start from the approach’s “own” system,
i.e., the one whose programme data and decisions are, respectively, the knowns
and the unknowns of the chosen approach. (In the short-run profit approach, this
means using the SRP Optimization System, as is done next, in Sect. 4.2.) Then
“solving the programmes for their decision variables” means exactly the same as
3
For production techniques with conditionally fixed coefficients, the Kuhn-Tucker Conditions are
spelt out in Sect. 7.1.
4.1 Outline of the Short-Run Approach 79
“solving the system for the unknowns of the approach” (which is the task to be
done).
• If a different, “non-own” pair of programmes were solved—for its decisions in
terms of its data—then the whole solution correspondence (data-to-decisions)
would have to be obtained, and then part-inverted to express the unknowns in
terms of the knowns (thus compensating for the original mismatch between these
and the data and decisions). This might be worthwhile, but only when the “non-
own” programme in question is particularly easy to solve without using duality
(the use of duality leads from any programme pair to the same Kuhn-Tucker
Conditions and to the same FFE Conditions).
• When there is such a readily soluble programme and its value function is easy
not only to calculate but also to differentiate, it may well be best to use the
corresponding differential system. This may be a “non-own” system, i.e., one
in which the arguments and the derivatives of the function do not correspond
to the knowns and the unknowns of the chosen approach. In such a case,
after calculating the subdifferential correspondence, one must part-invert it as
required. The method can be useful when there is no explicit formula for
the chosen approach’s “own” function (whose arguments and derivatives are,
respectively, the knowns and unknowns of the approach), but there is a formula
for another value function. For example, there is no general formula for the SRP
of a production technique with conditionally fixed coefficients (c.f.c., a concept
introduced in Sect. 7.1), but the SRP approach can be based on the formula for
the LRC (7.1.3) or the SRC (7.1.23). However, this is worthwhile only if the
input requirement functions (kL and v)
L are simple enough. When they are not, it is
better to use an optimization system.4
• Finally, it might seem preferable always to use one of those seven systems which
can be partitioned in the way discussed in the Comments in Sect. 3.8. When such
a system is used for the SRP approach to LRGE, the calculation of SRGE requires
only the subsystem but not the valuation condition—i.e., this stage requires
solving the SRP programme (3.1.6)–(3.1.7) for .y; v/, but it need not include the
shadow-pricing of the fixed inputs (by solving for r the dual programme (3.3.6),
or (3.3.13)–(3.3.14) under c.r.t.s., or possibly by differentiating …SR or …SR
w.r.t. k). However, this does not save on computation if, as is common, the SRP
programme has to be solved by a duality method: the dual is then being solved
together with the primal anyway.
4
For example, this is how the pumped-storage technique of electricity generation (5.1.3) is
dealt with, in Sects. 5.2 and 5.3. And this is because the subdifferential of the storage capacity
requirement function (5.1.8)—calculated in [21]—is far from simple, even under the simplifying
assumption of perfect energy conversion.
80 4 Short-Run Profit Approach to Long-Run Market Equilibrium
Y
Y R ˆ R „ . (4.2.1)
(which is the SI’s set of fixed inputs). When the sets fˆ g2‚ are pairwise disjoint,
the summation in (4.2.2) reduces to a single term (for each ), and the notation can
be simplified accordingly: see the Comment following (4.2.21).
The set of all the fixed inputs of the SI, ˆ‚ , is partitioned into two subsets: ˆF‚
consisting of those with given prices, and ˆE‚ consisting of those whose prices are
determined in long-run equilibrium. For a particular technique 2 ‚, its set of
fixed inputs ˆ is thus partitioned into two subsets
S S
(where 2‚ ˆE D ˆE‚ and 2‚ ˆF D ˆF‚ of course). An input 2 ˆF‚ is
supplied at a fixed unit cost rF (in terms of the numeraire), and so its total supply
cost is linear. By contrast, the total supply cost of an input 2 ˆE‚ is given by a
convex function, G , of the supplied quantity q
. Typically,
G is a strictly convex
and increasing, finite function on an interval 0; q , with G .0/ D 0. But the
case of an input in a fixed supply q (without free disposal) is captured by setting
G q equal to 0 for q D q and to C1 otherwise (in which case the long-
run equilibrium condition that r 2 @G q means merely that q D q ). For
examples in modelling the Electricity Supply Industry (ESI), see Sect. 5.3 here, and
[21] or [27, Section 11], and [23] or [30, Section 9].
This classification of inputs will not always be clear-cut, but as a rough-and-ready
rule, for an industry supplying a good with a cyclical demand, its fixed inputs are
those which cannot be adjusted within a demand cycle because of the cost and the
time it takes. For example, there is usually no question of adjusting plant capacity
to demand even if the cycle is as long as a year. The variable inputs are those
which can be adjusted quickly, at negligible cost, to the time-varying output rate
y .t/. For example, in the model of thermal electricity generation in (5.1.1), the fuel
inputs are assumed to be adjustable
instantaneously. The variable inputs are regarded
as having fixed prices w , usually by reason of being internationally traded.
Likewise, a typical fix-priced capital input 2 ˆF‚ is an internationally traded
kind of equipment, and its rental price rF is the annuity consisting of depreciation
and interest on the purchase price.5 By contrast, an equilibrium-priced capital input
2 ˆE‚ —whose rental price rE is determined only in long-run equilibrium—is
typically a factor that can be supplied only locally and at an increasing marginal
cost, as a result of the fixity of some assets required for its supply (such as special
geological sites or other natural resources). Constancy of returns to scale, assumed
for the SI’s technology, need not extend to the supply of its inputs, and in the
application to peak-load pricing with storage the reservoir capacity has an increasing
marginal cost (Sect. 5.3).
For simplicity, all input demand for the SI’s products is taken to come from
a single Industrial User (IU), who produces a final good from inputs of the
differentiated good and of the numeraire. The user’s production function FW YC
RC ! R, assumed to be strictly concave and increasing, defines his production set
where YC is a convex cone that is P-closed (i.e., closed for some, and hence for
every, locally convex topology on Y that yields P as the continuous dual space).
When, as in superdifferentiation at the algebraic boundary points (non-core points)
5
Formally, the fixed prices rF and w are built into the standard competitive equilibrium model
by
introducing a linear production set equal to the hyperplane perpendicular to the vector rF ; w; 1
and passing through the origin in the space of the supplier’s fix-priced inputs and the numeraire.
82 4 Short-Run Profit Approach to Long-Run Market Equilibrium
S
where „‚ WD 2‚ „ . There is a finite set, Ho, of households. For each h 2 Ho,
its utility is a concave nondecreasing function Uh on the consumption set YC R2C 7 ;
it is assumed to be nonsatiated in each of the two homogeneous goods (the IU’s
product and the numeraire), i.e., Uh .xI '; m/ is increasing in ' and in m. This
guarantees that both prices are positive in equilibrium. Each household’s initial
endowment is a quantity mEn h > 0 of the numeraire Ponly, and its share of profit
from the supply of input 2 ˆE‚ is & h 0 (with h & h D 1). Similarly, & h IU
denotes household h’s share in the Industrial User’s profit.
The Supply Industry’s pure profit vanishes in long-run equilibrium (because
of c.r.t.s.), but short-run analysis requires specifying the households’ shares in
the operating profits from the SI’s plants if it is to be exact. This is because the
operating profit …SR in (4.2.11) is only approximately offset by the liabilities rjEF
k ,
which represent plant depreciation and interest (on the debt from which the plant
is assumed to have been financed). A plant is specified by its type and by its
capacities (or, more generally, its quantities of the fixed inputs) k , for 2 ˆ .
It is assumed that every plant of a particular type has the same capacity ratios
(k 0 W k 00 W ). With c.r.t.s., this amounts to assuming—as is now done—that
there is no more than one “totalized” plant of each type (with capacity equal to
the total capacity of plants of this type). Though in reality this is rarely so in an
industry that has evolved over time, the equal-ratios condition is met in long-run
equilibrium—calculation of which is the main use for the short-run model here. It
makes
P sense, then, to speak of profit shares in a technique : denoted by $ h (with
h $ h D 1), household h’s share in the operating profit from technique is
X
$ h WD ˇ hi ˛ i (4.2.4)
i
where ˇ hi is h’s share in producer i, and ˛ i is i’s share in the plant of type .
In other words, one can regard the concepts of “producers” and “plant types” as
identical (since, in perfectly competitive equilibrium analysis, decomposition of the
total production set into the sum of individual producers’ sets matters only insofar
as the households’ shares of profits may differ from one producer to another).
Notation (Restrictions of Input Price Systems) The restriction, to „ , of a
wW „‚ ! R is wj„ , abbreviated to wj . Similarly, rj
E F
and rj mean the restrictions
toEˆFand
to ˆ of an r W ˆ‚ ! R and an r W ˆ‚ ! R, respectively. Also, the pair
E F E E F F
8
The abbreviation is unambiguous because there is no need for the notation rF? (it would mean the
same as rF ).
84 4 Short-Run Profit Approach to Long-Run Market Equilibrium
i.e., the equilibrium quantities and prices meet the Complementarity Condi-
tions (3.1.5), or any of the equivalent systems of conditions. In other words,
?
y ; k? ; v? maximizes, to zero, the pure profit at prices p? ; rj?
; rj
F
; wj .
2. Producer optimum in User Industry : . p? ; 1/ 2 %? @F O .z? ; n? /.
? ? ?
3. Consumer utility maximization : For each h, xh ; ' h ; mh maximizes Uh on the
budget set B p? ; %? ; MO LR h . p? ; r? ; %? / , where
P P P
4. Market clearance : y? D z? C ?
h xh ' ?h .
and F .z? ; n? / D
P ?
h
5. Marginal cost pricing of Supply Industry’s fixed inputs : r? 2 @G k for
each 2 ˆE‚ .9
Comment This is an instance of the usual equilibrium concept, except for being
specialized to the case of nonzero prices (%? and 1) for the two composite goods
(the above characterization of the IU’s profit maximum, Condition 2, relies on
the positivity of the output price %? ). The usual definition of general equilibrium
captures also the case of zero prices, but here this cannot arise because of the
nonsatiation assumptions. In other words, price positivity is actually a property of
any equilibrium (and not part of the concept itself).
The short-run profit (SRP) approach to solving the long-run equilibrium system
starts by fixing the SI’s capital inputs .k / 2‚ . Given these quantities as well as
arbitrary prices . p; w/ for the SI’s variable commodities, a suitably chosen system of
conditions characterizing long-run producer optimum is then solved for: the plants’
outputs .y /, their variable inputs .v /, and the values .r / imputed
P to
the fixed
inputs in the plant of each type . The total of the optimal outputs yO p; k ; wj
is then equated to the demand for the SI’s products to find their short-run equilibrium
9
The subdifferential @G is an interval if the left
and right derivatives of G differ; this can
be the case only on a countable subset of 0; q . Also, @G .0/ D 0; dG =dq .0C/ and
h
@G q D dG =dq q ; C1 .
4.2 Detailed Framework for Short-Run Profit Approach 85
prices p?SR —which depend on the k ’s.10 Finally, to determine the capacities .k /
andthe prices rE of any equilibrium-priced capital inputs, every imputed value
rO p; k ; wj is equated either to the given price rF (for 2 ˆF‚ ) or to the
P
marginal supply cost dG =dq at q D k (for 2 ˆ‚ ). As part of this
E
X
MO SR h pI rE ; rF I w; % j k WD mEn
h C $ h …SR p; k ; wj rj
EF
k
2‚
0 0 11
X X X
C & h @rE k G @ k AA C & h IU …IU . p; %/ . (4.2.11)
2ˆE‚ W 2ˆE W 2ˆE
Given a k as well as rEF and w, the short-run general equilibrium (SRGE) system
to be solved consists of the following conditions on the other variables (which are
prices p paired with quantities y and xh and z, price % paired with quantity ' h ,
quantities v , and amounts of numeraire mh and n):
10
The corresponding demand for the variable inputs, vL yO p; k ; wj ; k ; wj , would similarly
have to be equated to their supply if the supply were not taken to be perfectly elastic (i.e., if the
variable-input prices w were not fixed, and they too had to be determined in equilibrium).
86 4 Short-Run Profit Approach to Long-Run Market Equilibrium
.xh ; ' h ; mh / maximizes Uh on B p; %; MO SR h p; rEF ; w; % j k (4.2.15)
X X X
y D z C xh and F .z; n/ D ' h. (4.2.16)
2‚ h2Ho h2Ho
11
This programme is an LP if vL (including the constraints Rthat describe its domain) is linear in y .
For example, in thermal electricity generation vL .y / D y .t/ dt for y k , and so (5.2.1)–
(5.2.3) is an LP.
12
The short-run equilibrium is assumed to be unique, to simplify the notation.
4.2 Detailed Framework for Short-Run Profit Approach 87
of the short-run equilibrium problem’s data k and rE (as well as depending on the
fixed prices rF and w).13
The second and final stage is to determine the long-run equilibrium, i.e., the
equilibrium capacities and the prices of any equilibrium-priced capital inputs (the
inputs in ˆE‚ ). In the SRP approach, optimal choices of the investments .k /
are made by meeting the other conditions of the SRP Optimization System—
i.e., (3.2.6)–(3.2.7)
for each technique’s production set Y . For this, the solution
set RO p; k ; wj of the FIV minimization programme (3.3.13)–(3.3.14) with Y
in place
of Y,
or the solution rO if it is indeed unique, is calculated at p D
p?SR k; rEF ; w . Usually rO will already have been found as the dual solution in the
process of solving the SRP programme for y by a duality method (as a by-product
of Step 2 in solving (4.2.12)–(4.2.16)). Finally, the long-run equilibrium system of
valuation conditions:
E
rj ; rj
F
2 RO p?SR kI rEF I w ; k ; wj EF
i.e., rj satisfies (3.2.6) for each 2 ‚
(4.2.19)
0 1
X
rE 2 @G @ k A for each 2 ˆE‚ (4.2.20)
W 2ˆ
13
The SRGE analysis simplifies when there is no income effect on consumer demand for the
differentiated
? good (i.e., when xO h is independent of M, in the relevant range): the solution
pSR ; %?SR to (4.2.17)–(4.2.18) is then independent of rEF (as in Chap. 2).
14
As a basic check, note that the number of “generalized equations” in this system (each d-
dimensional
P vector inclusion counting as d “equations”) is the same as the number of unknowns
(viz., 2‚ card ˆ C card ˆE‚ ).
88 4 Short-Run Profit Approach to Long-Run Market Equilibrium
The SRGE System (4.2.12)–(4.2.16) together with the system of valuation condi-
tions (4.2.19)–(4.2.20) may be called the SRP Programme-Based LRGE System.
Comments (on the Technology Model and the Valuation Conditions)
• A production technique can usually be identified by its set of fixed inputs, i.e.,
ˆ 0 ¤ ˆ 00 for 0 ¤ 00 . Under the stronger assumption that different techniques
use disjoint sets of fixed inputs, i.e., that
ˆ 0 \ ˆ 00 D ; for 0 ¤ 00 , (4.2.21)
the SI’s total investment in fixed input is simply k for the one such that
ˆ 3 . In other words, it is the case-function (of ) defined, piecewise, as equal
to the function k on each ˆ . Thus it can be identified with k D .k / 2‚ itself.
So, under (4.2.21), the total investment can be denoted by kW ˆ‚ ! R. The
investment in technique is then the restriction of k to ˆ , which is denoted by
kjˆ , abbreviated to kj . The investment in fixed input is k (i.e., q D k in
this case). This is so in the model of the ESI presented in Sect. 5.1.
• Assume that: (i) different techniques use disjoint sets of capital inputs,
i.e., (4.2.21) holds, (ii) each input-cost, G k , is a differentiable
function
of k 2 RCC , and (iii) a unique shadow price system rO p; kj ; wj exists at
every k 0 and every p in a subspace of P that is known to contain p?SR . (As
is shown in [28] for a class of problems that includes peak-load pricing with
storage, this is so for the space of continuous real-valued functions C Œ0; T, as a
price subspace of P D L1 Œ0; T.) If a long-run equilibrium with k? 0 is sought,
then Conditions (4.2.19)–(4.2.20) on k simplify to the following equations for k
(a strictly positive vector in Rˆ‚ ):
! !
dG
rO p?SR kI k ; r I w ; kj ; wj
F
dk 2ˆE‚
( dG
k if 2 ˆE
D dk
(4.2.22)
rF if 2 ˆF
15
Formally, this is because in long-run equilibrium rj EF
D rO as per (4.2.19), and because
ı
p; rO ; wj 2 Y by the dual constraint on r . For the same reason, in calculating the long-
run
equilibrium, attention can be restricted, already at the short-run stage, to those rEF ’s with
EF
p; rj ; wj 2 Yı for each .
Chapter 5
Short-Run Approach to Electricity Pricing
in Continuous Time
1
A steam plant’s efficiency is the product of the boiler’s and turbine-generator’s efficiencies, which
is about 0:85 0:45 38 % (i.e., the heat rate is about 1=0:38 3600 kJ/kWh 9500 k J/kWh).
2
In reality, the minimum operating load is 10 % to 25 % of the maximum, and the incremental rate
rises with load by up to 5 % to 15 %. Also, there is a no-load heat input (which is a sunk operating
cost per unit time of being on line). See, e.g., [38, Figures 8.2 and 8.3, and Table 8.3].
3
In reality, a turbine’s efficiency varies with the load (from about 85 % to 95 % for movable-blade
types, or from 70 % to 95 % for fixed-blade types). Also, a plant’s head varies with the water stock.
The variation tends to be larger in lower-head plants, but it much depends on the particular plant:
e.g., with a typical medium head (say about 150 m), the variation is 3 % of the maximum in some
plants, but over 30 % in others. For a variable-head plant, the operation and valuation problems are
studied in [49].
4
In reality, the round-trip conversion efficiency Ro is close to 1 in SMES (it is over 95 %). In
PWES and CAES, Ro is around 70 % to 75 % (i.e., 0:7 to 0:75 kWh of electricity is recovered
from every kWh used up). The case of Ro < 1 is included in the model of pumped storage in [21],
as are the cases of converter asymmetry or nonreversibility (although reversibility is common,
some high-head PWES plants do use nonreversible multi-stage pumps).
5
In reality, startup times range from nearly zero for some energy storage plants (SMES coils and
batteries can switch from charging to discharging in 4 to 20 miliseconds), through a few (1 to 10)
minutes for other storage plants (PWES or CAES) as well as for gas turbines and hydro plants, to
hours for nuclear or fossil (coal, oil, gas) steam-plants (whose long startup times must of course
be distinguished from the very much shorter loading times of the spinning reserves): see, e.g., [38,
Table 8.2] and [40] or [10].
6
For the one-station technology, the thermal SRC and LRC are given by (2.2.1) and (2.2.2). For
extension of these formulae to a multi-station technology, see, e.g., [24, (22)–(26)] and [23].
7
See the last Comment in Sect. 7.1.
5.1 Technologies for Electricity Generation and Energy Storage 93
and the energy converter) do have well-defined and separate profit-imputed marginal
values: see [21] or [27, Section 9]. In the case of a hydro plant, it guarantees also that
the river flows have well-defined marginal values (as do the reservoir and turbine
capacities): see [23] or [30].
But the assumption of no startup costs can be rather less distorting than it may
seem. This is because the slow-starting plants tend to have low unit running costs,
and the quick-starting plants tend to have high unit running costs. To minimize the
operating cost, one allocates the base load to the lowest-cost plants, and the near-
peak loads to the highest-cost plants. Thus the slow starters end up serving mainly
the constant load levels (the base load), and the quick starters end up serving the
most intermittent load levels (the near-peaks)—even if the differences in startup
times are disregarded in making the despatch policy.
The complete generating technology consists, then, of the various thermal,
hydro and pumped-storage techniques, which form three sets: ‚Th , ‚H and ‚PS .
However, what is considered here is a smaller model with a number of thermal
techniques and just one other, viz., either a pumped-storage technique or a hydro
technique. So the single non-thermal technique can be denoted simply by PS or
H, and the set of thermal techniques by f1; 2; : : : ; ‚g, where ‚ means the number
of thermal techniques. In other words, the ESI’s set of techniques is henceforth
either f1; 2; : : : ; ‚I PSg or f1; 2; : : : ; ‚I Hg. It plays the role of the abstract set ‚ of
Sects. 4.1 and 4.2.
The output space Y is here L1 Œ0; T, which is the vector space of all essentially
bounded real-valued functions on the interval Œ0; T that represents the cycle. Func-
tions equal almost everywhere, w.r.t. the Lebesgue measure (meas), are identified
with one another. With the usual order and the supremum norm
the space L1 is a dual Banach lattice.8 Its Banach predual is L1 Œ0; T, the space
of all integrable functions. When, as here, it serves as the price space P, a TOU
electricity price is a density function, i.e., a time-dependent rate p .t/ in $/kWh.
The price space L1 Œ0; T is sufficient in the case of interruptible demand because
capacity charges are then spread out over a flattened peak: see [26]. A larger price
space is needed to accommodate the instantaneous capacity charge that arises in the
case of a firm, pointed peak.9
8
For Banach-lattice theory, see, e.g., [2, Chapter 4], [8, XV.12], [33, Chapter X] and [39].
9
An instantaneous charge can be represented by a point measure; in the context of electricity
pricing, this is a capacity charge in $ per kW of power taken at a peak instant, and it is additional
to the marginal fuel charge, which is a price density in $ per kWh of energy . A general singular
measure can be interpreted as a concentrated charge. As is pointed out in [26, Sections 1 and 2],
the Banach dual L1 can be useful in arriving at such a price representation when the equilibrium
allocation lies actually in the space of continuous functions C Œ0; T
L1 Œ0; T. This is because
the restriction, to C , of a linear functional p 2 L1 has the Riesz representation by a (countably
94 5 Short-Run Approach to Electricity Pricing in Continuous Time
1 T
YTh WD .yI k; v/ 2 L1
C R2 W y k; y .t/ dt v; y 0 (5.1.1)
0
where the constant is the efficiency of energy conversion (the ratio of electricity
output to heat input). The unit fuel cost w (in $ per kWh of electricity output) is the
fuel’s price (in $ per kWh of heat input) times the heat rate 1=. Henceforth, it is
taken to represent all of the unit running cost (a.k.a. operating or variable cost).10
There is a number of thermal techniques D 1, 2; : : : ; ‚. Each has the same
structure (5.1.1), but it uses its own input commodities, viz., the capacity of type
and the suitable type of fuel, : in terms of (4.2.1), ˆ D fg and „ D f g.
Its production set, Y , is formally YTh embedded in the full commodity space
by inserting zeros in the input-output bundle at all the positions other than ,
and the t’s (as in Sect. 4.2). The relevant quantities and prices are indicated with
the subscript : technique generates an output flow y from an input k of the
generating capacity of type and from an input v of the fuel of type . Its unit
fuel cost is its heat rate 1= times its fuel’s price wQ . The unit fuel cost of plant
type is denoted by11
w WD wQ = .
Without loss of generality, one can assume that the thermal techniques are numbered
in the order of increasing unit operating cost (a.k.a. the merit order), i.e., that
additive) measure pC 2 M D C , which can have a singular part as well as a density part. The
failure of L1 itself to have a tractable mathematical form is thus side-stepped without restricting
the analysis to the case of price densities. (The alternative of working entirely within C and M as
the commodity and price spaces is available when all demand is uninterruptible [20]. When, by
contrast, all demand is harmlessly interruptible, the equilibrium price is a pure density [26].)
10
The other components of the unit running cost (extra maintenance, etc.) can be accounted for by
a levy on fuel.
11
Note that each w can be interpreted as the price of the fuel of type if different types of plants
use different fuels (i.e., if 0 ¤ 00 for 0 ¤ 00 ). Each fuel can then be unambiguously measured
in kWh of generated electricity instead of being measured as the heat input (and such measurement
redefines the plant’s efficiency as 1, thus equating its unit fuel cost to the price of its fuel).
5.1 Technologies for Electricity Generation and Energy Storage 95
efficient and symmetrically reversible: this means that, in unit time, a unit converter
can either turn a unit of electricity into a unit of the storable energy, or vice versa .
So the output from storage, y D yC y , equals the rate of energy flow from the
reservoir, Ps D ds=dt (where s .t/ is the energy stock at time t). Energy can be
held in storage at no running cost (or loss of stock). The long-run production set is,
therefore, the convex cone
Z T
YPS WD .yI kSt ; kCo / 2 L1 Œ0; T R2 W jyj kCo ; y .t/ dt D 0
0
Z t
Z T
and 9 2 Œ0; e .y .t/ e .t/ C .t// dt D 0 and
0
Z t
12
Note the two unrelated uses of the symbols s and : here, in Sect. 5.1, in the description of the
pumped storage and hydro techniques, these mean the energy stock (s) and water spillage ( ). But
in the general duality scheme of Sects. 3.3 and 3.12, they mean the standard parameters (s) paired
with the standard dual variables ( ).
96 5 Short-Run Approach to Electricity Pricing in Continuous Time
Y0 D L1
0 Œ0; T WD y 2 L1
W y .t/ dt D 0 (5.1.7)
0
and the requirements for storage capacity and conversion capacity, when the
(signed) output from storage is y 2 L1
0 , are:
Z t Z T
kL St .y/ D max y .t/ dt C max y .t/ dt (5.1.8)
t2Œ0;T 0 t2Œ0;T t
13
Also, though this is only a technicality, the hydro technique has an infinity of input variables
.e .t//t2Œ0;T —unlike a c.f.c. technique as defined in Sect. 7.1.
5.2 Operation and Valuation of Electric Power Plants 97
• In both pumped storage and hydro generation, the flows to and from the reservoir
RT
are required to balance over the cycle ( 0 sP .t/ dt D 0), i.e., the stock must be
a periodic function of time. But its level at the beginning or end of a cycle
is taken to be a costless decision variable, s0 . In other words, when it is first
commissioned, the reservoir comes filled up to any required level at no extra cost,
but its periodic operation thereafter is taken to be a technological constraint. For
a brief comparison with the case of a given s0 , or a variable but costly s0 , see [21].
For each of the plant types described in Sect. 5.1, the problem of profit-maximizing
operation can be formulated as a doubly infinite linear programme for the output
rate y .t/ at each time t (in kW), given a TOU electricity price p .t/ for each time
t in the cycle (in $/kWh). Its dual is the plant valuation programme (either for the
values of capacity services .t/ at each time t, or just for their total value over the
cycle, r). This is an application of the general duality scheme of Sects. 3.3 and 3.12.
Correspondence of notation between that scheme and its present application is spelt
out in Table 5.1: in particular, in both cases r is an intrinsic dual variable, but the
standard dual variable of the general scheme corresponds to and (and ) in
the context of the ESI.14
For a thermal plant of capacity k with a unit running cost w, the operation LP
RT
(reduced by working out the short-run cost as w 0 y dt) is:
i.e., the measurable functions satisfying (5.2.4) form the solution set YO Th . p; k; w/.
RT C
So the thermal plant’s operating profit is …Th
SR . p; k; w/ D k 0 . p .t/ w/ dt, and
Whereas in the context of a hydro plant, means not a shadow price but a quantity (viz., spillage,
14
Table 5.1 Correspondence of notation between the general duality scheme (Sects. 3.3 and 3.12)
and its applications to the ESI (Sects. 5.2 and 5.3). The abbreviations read: (i) in the leftmost
column: GS D general scheme, Th D thermal generation, PS D pumped storage, Hy D hydro
generation; (ii) elsewhere: St D storage reservoir, Co D converter, Pu/Tu D pump/turbine (two
working modes of a reversible PS converter), Tu D hydro turbine. Functions of time are marked
with a ./, and measures on the time interval are marked with a .d/. In the general duality scheme,
s and mean the standard parameters and dual variables Lagrange multipliers). But in the context
of energy storage (both PS and Hy), s means the energy stock (and means spillage in Hy). Also,
the intrinsic parameters and dual variables of the general scheme, r and k, correspond to .r; / and
.k; e/ in the hydro problem
Intrinsic Intrinsic Std Std Relationship between
primal dual primal dual intrinsic and standard
param. vbles param. vbles dual variables
GS k [vect] r [vect] s [vect] [vect] r D BT
Th k ./ ./ R
k [scal] r [scal] r D dt
n ./ ./
PS kSt ./ St .d/
R
kSt [scal] rSt [scal] nSt ./ St .d/ rSt D St .dt/
kTu ./ Tu ./
R
kCo [scal] rCo [scal] kPu ./ Pu ./ rCo D . Tu C Pu / dt
Hy kSt ./ St .d/ R
kSt [scal] rSt [scal] rSt D St .dt/
nSt ./ St .d/
kTu ./ Tu ./ R
kTu [scal] rTu [scal] rTu D Tu dt
nTu ./ Tu ./
Differentiation is the simplest way to value a unit of thermal capacity because the
operation problem is so simple that its solution and value function can be calculated
directly (i.e., without using a duality method). Of course, rOTh can be calculated
also by solving the dual problem of capacity valuation. The standard dual of the
operation LP is the following programme for the flow of rent (whose total for the
cycle is r), with as the Lagrange multiplier for the nonnegativity constraint on y
in (5.2.3):
which is the thermal plant’s unit rental value. The standard dual solution, unique if
k > 0, is
For a pumped-storage plant with capacities .kSt ; kCo /, the operation LP is:
Unlike the case of …Th SR , there is no explicit formula for …SR . p; kSt ; kCo /, the
PS
operating profit of a pumped-storage plant; and both operation and rental valuation
of a storage plant are best approached through the dual problem of capacity
valuation. The standard dual of the operation LP is the following programme for:
(i) the flow of reservoir’s rent St , and (ii) the flow of converter’s rents Co D
Pu C Tu , which it earns in its two modes of work, viz., charging the reservoir
as a “pump” and discharging it as a “turbine”. Their totals for the cycle are the unit
rental values: (i) of the reservoir rSt (in $/kWh), and (ii) of the converter rCo (in
$/kW). The dual variables Pu and Tu range over L1 Œ0; T, like the in (5.2.7).
The space for St is M Œ0; T, the space of Borel measures on Œ0; T, which is the
norm-dual of the space of continuous functions C Œ0; T. This is also the space for
the multiplier St for the nonnegativity constraint in (5.2.16). The multiplier for
100 5 Short-Run Approach to Electricity Pricing in Continuous Time
The standard dual’s inclusive form has also the dependent decision variables
Z T
rSt D St .dt/ D St Œ0; T (5.2.23)
0
Z T
rCo D . Pu .t/ C Tu .t// dt (5.2.24)
0
15
This is done by using the constraints (5.2.21)–(5.2.22) and the disjointness conditions Tu ^
Pu D 0 and St ^ St D 0, which are met by any solution to (5.2.17)–(5.2.22) if kSt > 0 and kCo >
0; i.e., it is not optimal for the dual variables to overlap and partly cancel each other out in (5.2.22).
Note that St and St are disjoint as measures on the circle, i.e., min f St f0; Tg ; St f0; Tgg D 0 in
addition to St ^ St D 0 in the lattice M Œ0; T.
16
The one-sided limits exist at every t and are equal nearly everywhere (n.e.), i.e., everywhere
except for a countable set. Specification of .t/ between .t/ and .tC/ is unnecessary.
5.2 Operation and Valuation of Electric Power Plants 101
is extended by continuity to Œ0; T; i.e., .0/ WD .0C/ and .T/ WD .T/. The
cyclic positive variation of is
VarC C
c . / WD Var . / C . .0/ .T//C (5.2.26)
where VarC . / is the total positive variation (a.k.a. upper variation) of , i.e.,
P C
the supremum of m . m / m over all finite sets of pairwise disjoint
subintervals m ; m of .0; T/: see, e.g., [16, Section 8.1] for details.17
In these terms, the capacity valuation problem (for a pumped-storage plant)
becomes the following unconstrained CP for shadow-pricing the energy stock:
Its main feature is the trade-off between minimizing the variation (which on its own
would require setting at a constant value) and minimizing the integral (which on
its own would require setting equal to p). This trade-off is what determines the
extent to which local peaks of p should be “shaved off” and the troughs “filled in”
to obtain the optimum shadow price function O PS . p; kSt ; kCo /, at least in the case of
a piecewise strictly monotone p. The solution, shown in Fig. 5.1a, is determined by
constancy intervals for O PS around a local peak or trough of p (as a function of t).
Unless kSt =kCo is relatively long, these intervals do not abut, and must all be of that
length.18 The optimal output has the “bang-coast-bang” form
yO PS .tI p; kSt ; kCo / D kCo sgn p .t/ O PS .tI p; kSt ; kCo / (5.2.29)
i.e., yO PS .t/ equals kCo , 0 or kCo if, respectively, p .t/ > O PS .t/, p .t/ D O PS .t/
or p .t/ < O PS .t/: see Fig. 5.1b. The lowercase notation, yO PS or O PS , is used
only when the solution is unique . In general, the solution sets for (5.2.12)–(5.2.16)
and (5.2.27)–(5.2.28) are denoted by YO PS . pI kSt ; kCo / and ‰ O PS . pI kSt ; kCo /. More
Rt
precisely, y 2 YO PS means that y together with s0 D maxt 0 y ./ d (which the
Rt
lowest initial stock needed for the stock s0 0 y ./ d never to fall below 0) solves
(5.2.12)–(5.2.16).
17
The other term, . .0/ .T//C , represents a possible jump of at the instant separating two
consecutive cycles.
18
The matter becomes more complicated when the ratio kSt =kCo is comparable to durations between
the successive local peaks and troughs of p—so that the neighbouring constancy intervals of O PS
start to abut. However, there is a similar optimality rule for such clusters: see [21].
102 5 Short-Run Approach to Electricity Pricing in Continuous Time
Fig. 5.1 Trajectories of: (a) shadow price of stock O , and (b) output of pumped-storage plant
(optimum storage policy) yO PS in Sect. 5.2, and in Theorem 5.3.1. Unit rent for storage capacity
0 00
is VarCc
O D d O C d O , the sum of rises of O . Unit rent for conversion capacity is
R T ˇˇ ˇ
0 ˇp .t/
O .t/ˇˇ dt, the sum of grey areas. By definition, O PS D kSt =kCo
The stock-pricing programme (5.2.27)–(5.2.28) has a solution for every kSt > 0
and kCo > 0 (by Lemma 6.4.1 or Part 2 of Proposition 7.4.2).19 If p is continuous,
i.e., p 2 C Œ0; T, then there is a unique solution O PS . pI kSt ; kCo /. It follows that
the plant’s operating profit …PS SR is differentiable in .kSt ; kCo /; equivalently, with
this technology the programme (3.3.13)–(3.3.14) or (7.1.14)–(7.1.17) has a unique
solution rO . In terms of O PS , the unit rental values of the reservoir and the converter
19
When kSt > 0 but kCo D 0, any constant is a solution. When kCo > 0 but kSt D 0, a solution
exists if and only if p 2 BV, in which case it is unique, viz., D p.
5.2 Operation and Valuation of Electric Power Plants 103
Z Z ˇ ˇ
@…PS T T
ˇ ˇ
SR
D rOCo . p; kSt ; kCo / D .O Pu C O Tu / .t/ dt D ˇp .t/ O PS .t/ˇ dt.
@kCo 0 0
(5.2.31)
(5.2.32)
Z T
maximize p .t/ y .t/ dt over y 2 L1 Œ0; T and s0 2 R (5.2.33)
0
As with pumped storage, there is no explicit formula for the hydro plant’s operating
profit …HSR . pI kSt ; kTu I e/, and both operation and rental valuation of a hydro plant
are best approached through the dual problem of fixed-input valuation, which is an
LP for: (i) the flow of reservoir’s unit rent St , (ii) the flow of turbine’s unit rent Tu ,
and (iii) the river’s unit rent, i.e., the shadow price of water . By including but
not r among the dual variables, this is a partly inclusive form of the standard dual
LP. The fully inclusive form has also rSt and rTu , the rental values of the reservoir (in
$/kWh) and of the turbine (in $/kW), but these are simply the totals of St and Tu
for the cycle. The dual variable Tu ranges over L1 Œ0; T, and the space for St is the
space of measures M Œ0; T, as in pumped storage. The space for can be L1 Œ0; T
formally, but actually is constrained to BV .0; T/ by (5.2.43). The multipliers
for the nonnegativity constraints in (5.2.34) and (5.2.36) are Tu 2 L1 Œ0; T and
St 2 M Œ0; T. The multiplier for the balance constraint (5.2.35) is a scalar . So
104 5 Short-Run Approach to Electricity Pricing in Continuous Time
the LP of fixed-input valuation for a hydro plant (a partly inclusive standard dual of
the operation LP) is:
The dual’s fully inclusive form has also the remaining dependent decision variables
Z T
rSt D St .dt/ (5.2.44)
0
Z T
rTu D Tu .t/ dt. (5.2.45)
0
Recall that VarCc . /, defined by (5.2.26), is the total of all rises of over the cycle.
If kTu > e .t/ > 0 for every t, then the sum of the two integrals in (5.2.47)
has a minimum at (and only at) D p. Therefore, the programme’s main feature
is the trade-off between minimizing the variation (which on its own would require
setting at a constant value) and minimizing the sum of integrals (which on its own
would require setting equal to p). This trade-off is what determines the extent to
5.2 Operation and Valuation of Electric Power Plants 105
which the local peaks of p should be “shaved off” and the troughs “filled in” to
obtain the optimum shadow price function O H . pI kSt ; kTu I e/, at least in the case
that p is piecewise strictly monotone and kTu > e > 0 at all times. The solution
is determined by constancy intervals for O H . If kSt = Sup .e/ and kSt = .kTu Inf .e//,
which are upper bounds on the times needed to fill up and to empty the reservoir,
are sufficiently short, then the constancy intervals do not abut. Around a trough
Rt
of p there is an interval .t; t/ characterized by t e .t/ dt D kSt , on which p .t/ <
O throughout. Around a local peak of p there is an interval .t; t/ characterized by
R tH
O H throughout. The optimal output has
t .kTu e .t// dt D kSt on which p .t/ >
the “bang-coast-bang” form
8
< kTu if p .t/ > O H .tI p; kSt ; kTu ; e/
yO H .tI p; kSt ; kTu ; e/ D e .t/ if p .t/ D O H .tI p; kSt ; kTu ; e/ . (5.2.49)
:
0 if p .t/ < O H .tI p; kSt ; kTu ; e/
kSt > 0 and kTu > EssSup .e/ EssInf .e/ > 0. (5.2.50)
O H . pI kSt ; kTu I e/ D r e …H
SR . pI kSt ; kTu I e/ . (5.2.51)
This is the TOU price of water (unit value of the river flow). It follows that the plant’s
operating profit …H O
SR is differentiable also in .kSt ; kTu /. In terms of H , the unit rental
values of the reservoir and the turbine (in $/kWh and $/kW, respectively) are:
@…H
rOSt . pI kSt ; kTu I e/ D SR
D VarC O H . pI kSt ; kTu I e/ (5.2.52)
@kSt c
Z T C
@…H
rOTu . pI kSt ; kTu I e/ D SR
D p .t/ O H .t/ dt. (5.2.53)
@kTu 0
Comments (Comparison of the Standard and the Intrinsic Duals of the Ther-
mal Plant Operation Programme) 20
• The standard perturbation of the primal LP (5.2.1)–(5.2.3), which produces the
dual LP (5.2.6)–(5.2.9), consists in adding cyclically varying increments ( k .t/,
n .t/) to the constants .k; 0/ 2 R R in (5.2.3). The resource increments,
. k; n/ 2 L1 L1 , are paired with Lagrange multipliers .; / 2 L1 L1 .
• By giving the unit rent’s distribution over time, —rather than only its total
for the cycle, r—the standard dual LP (5.2.6)–(5.2.9) is the “fine” form of the
valuation problemR (in the sense of the first Comment in Sect. 3.12, with the
integral 7! .t/ dt as the adjoint operation 7! BT ). The “coarse” form
of valuation is a case of the intrinsic dual (3.3.13)–(3.3.14). It can be put in a
form specific to production techniques with conditionally fixed coefficients by
substituting the input requirement functions (5.1.5) and (5.1.6) for kL and v,
L and
L1C for Y0 , in either the (constrained) CP (7.1.14)–(7.1.17) or the LP (7.5.6)–
(7.5.9). The latter programme is then a semi-infinite LP for the single variable r
(with an infinity of constraints).
Comments (Comparison of the Standard and the Intrinsic Duals of the
Pumped-Storage Plant Operation Programme)
• The standard perturbation of the primal LP (5.2.12)–(5.2.16), which produces the
dual LP (5.2.17)–(5.2.22), uses cyclically varying increments ( kSt .t/, nSt .t/)
to the constants (kSt , 0) in (5.2.16). It uses also two separate increments ( kPu .t/,
kTu .t/) to the two occurrences of kCo in (5.2.14)—i.e., (5.2.14) is perturbed to:
Additionally, a scalar
is used as an increment to the 0 on the r.h.s. of (5.2.15).
The resource increments kSt 2 C, nSt 2 C, kTu 2 L1 , kPu 2 L1 and
2 R are paired with the Lagrange multipliers St 2 M, St 2 M, Tu 2 L1 ,
Pu 2 L1 and 2 R. This perturbation scheme is described in detail in [21] and
[27, Section 5].
• By giving the distributions of unit rents over time (and over the two conversion
modes), St and Pu C Tu —rather than only their totals for the cycle, rSt and
rCo —the standard dual LP (5.2.17)–(5.2.22) is the “fine” form of the valuation
problem (in the sense of the first Comment in Sect. 3.12). The “coarse” form
of valuation is a case of the intrinsic dual (3.3.13)–(3.3.14). It can be put in a
form specific to production techniques with conditionally fixed coefficients by
L and (5.1.7) for
substituting the input requirement functions (5.1.8)–(5.1.9) for k,
20
Note the two unrelated uses of the symbol n: here, in Sect. 5.2, n and nSt and nTu mean lower
constraint parameters (whose original, unperturbed values are zeros). But in the short-run approach
to equilibrium and its application to electricity pricing, in Sects. 4.2 and 5.3, n means an input of
the numeraire.
5.2 Operation and Valuation of Electric Power Plants 107
21
What is more, a time-continuous optimal output from storage cannot be unique (unless kSt D 0
or kCo D 0). To see this in detail, take any y 2 C Œ0; T \ YO PS . pI kSt ; kCo /. With .kSt ; kCo / 0, if
p is nonconstant on Œ0; T then 0 … YO PS : see [21]. And if p is a constant then y can be chosen to be
nonzero (since every feasible y is then optimal). So the open set ft W 0 < y .t/ < kCo g is nonempty;
let A be one of its component intervals. Then p D D const: on A for each 2 ‰ O PS because:
(i) y .t/ D ˙kCo whenever p .t/ ¤ .t/, and (ii) 0 < s < kSt on A, which implies that D const:
on A. (Both (i) and the implication in (ii) are Complementary Slackness Conditions: see [21] or
0
[27, Section
R 6].) SinceR pjA D const:, y can be modified on A, without loss of optimality, to any y
such that A y0 dt D A y dt and 0 y0 kCo on A (with y0 D y outside of A). A similar argument
applies to the set ft W kCo < y .t/ < 0g.
5.3 Long-Run Equilibrium with Pumped Storage or Hydro Generation of. . . 109
the space for the dual variables paired to perturbations that range over C (viz., for
St and St ). This is because, like CC , the nonnegative cone L1 C has a nonempty
norm-interior—and so the positivity of capacities k, .kSt ; kCo / or .kSt ; kTu /,
together with the strict variant (5.2.50) of the feasibility of “coasting” the
hydro plant, imply that Slater’s Condition, as generalized to infinite-dimensional
inequality constraints in [44, (8.12)], holds with the supremum norm topology
on the primal parameter spaces L1 and C. This ensures the existence of a dual
optimum in the norm-dual spaces (i.e., of O Th and O Th in L1 , O Tu and O Pu in
L1 , O St and O St in M, and of O , O Tu and O Tu in L1 ). Density representation of
these dual variables (other than O St and O St ) comes from the problem’s structure
and the assumption that p is a density: since p 2 L1 , every optimal and (for
a thermal plant) is actually in L1 by (5.2.11), as is every optimal Tu and Pu (for
a storage plant), and every optimal Tu and Tu (for a hydro plant). And every
feasible is in BV
L1 by (5.2.43). This is what justifies the use of L1 (rather
than L1 ) in the above formulations of the dual LPs (when p 2 L1 ).
• In the case of a general p 2 L1 , the space L1 must be used for the relevant
dual variables (those paired to the primal perturbations that range over L1 ), and
then also the generating capacities’ optimal rent flows, O Th and O Tu , are in L1
(although the corresponding O Th and either O Pu or O Tu are in L1 because p 0).
• When p 2 L1 , the degenerate case of zero storage capacity (with a positive
conversion capacity) provides an example of a duality gap in the pumped-storage
problem (Appendix A).
the operation and valuation results for the ESI’s plants in the SRP Programme-
Based LRGE System, (4.2.12)–(4.2.16) with (4.2.19)–(4.2.20). This is done first
for an electricity supply technology that combines thermal generation with pumped
storage.
Except for theF storage capacity, all the ESI’s inputs are taken to have fixed
prices: r1F ; : : : ; r‚ for the thermal generating capacities, .w1 ; : : : ; w‚ / for the
F
corresponding fuels, and rCo for the storage plant’s converter. There is a location
where an energy reservoir of capacity kSt can be constructed at a cost G .kSt /.
The marginal cost is assumed to be increasing,
i.e., the construction cost is a
strictly convex and increasing function GW 0; kSt ! RC with G .0/ D 0. This
is approximately so with the PWES and CAES techniques, which utilize special
geological features.22 In the terminology of Sect. 4.2, the reservoir is the single
equilibrium-priced capital input; all the others have fixed prices. Formally: ˆEPS D
fStg, ˆFPS D fCog, and ˆF D ˆ D fg for each 2 ‚ (the set of thermal plant
types).
All input demand for electricity is taken to come from a single Industrial User,
who produces a final good from the inputs of electricity and of the numeraire, z
and n. His production function, .z; n/ 7! F .z; n/, is assumed to be strictly concave
and increasing, and Mackey continuous, i.e., m L1 R; L1 R -continuous on
L1C Œ0; T RC . One example is the additively separable form for F .; n/, i.e., the
RT
integral functional F .z; n/ D 0 f .t; z .t/ ; n/ dt, where f meets the conditions of [7,
p. 535].23
A complete commodity bundle consists, then, of electricity (differentiated over
time), the ESI’s inputs (viz., the thermal capacities, the fuels, and the storage
and conversion capacities), the produced final good and the numeraire. These
quantities and their prices are always listed in this order, but those which are
irrelevant in a particular context are omitted (as in Sect. 4.2). So a complete
price system is . pI .r / I .w / I rSt ; rCo I %; 1/, but a consumer price system is just
. pI %; 1/ 2 L1 Œ0; T R2 —since a consumption bundle consists of electricity, the
produced final good and the numeraire, denoted by .xI '; m/ 2 L1 Œ0; T R2 .
The utility
function, Uh for household h, also is assumed to be Mackey continuous,
i.e., m L1 R2 ; L1 R2 -continuous on the consumption set L1 2
C Œ0; T RC .
Each household’s initial endowment is a quantity of the numeraire mh > 0. The En
22
A more
general
case is that of initially
decreasing marginal cost: G is then concave on an “initial”
interval 0; kQ , and convex on kQ; k . A limiting case of this arises from a nonzero setup cost
G .0C/ > 0, with G convex on 0; k . Supply (of storage capacity) is then discontinuous at the
price equal to the minimum average cost, which is attained at some k greater than the point of
inflection kQ, i.e., at the price r WD mink .G .k/ =k/ DW G k =k. The profit-maximizing
˚ supply
is 0 at r < r, but it exceeds k at r > r. At r D r, it takes the two values 0; k , but none of the
intermediate values. The total supply curve for this form of marginal and average costs is discussed
in, e.g., [17, 4-4: Figure 4-5].
23
That is, the function t 7! f .t; z; n/ is integrable on Œ0; T for each .z; n/ 2 R2C , and the function
.z; n/ 7! f .t; z; n/ is concave, increasing and continuous on R2C , with f .t; 0; 0/ D 0 for every
t 2 Œ0; T. For a short proof of the Mackey continuity of F, see [25].
5.3 Long-Run Equilibrium with Pumped Storage or Hydro Generation of. . . 111
household’s share in the User Industry’s profit is & h IU 0, and its share of profit
from supplying the storage capacity is & h St 0 (both add up to 1 over h).
By feeding the programming results summarized in Sect. 5.2 into the framework
of Sect. 4.2, long-run market equilibrium is next characterized by optimality of the
ESI’s investments in addition to the SRGE System, which here is either (5.3.4)–
(5.3.9) for pumped storage, or (5.3.14)–(5.3.19) for hydro-thermal generation. For
simplicity, it is assumed that all the equilibrium capacities are positive, i.e., that
every type of plant is built (in reality, some plant types might not be built because
of their costs).
Theorem 5.3.1 (SR Description of LR Equilibrium for ESI with Storage)
Assume that the technology of the Electricity Supply Industry consists of thermal
generation techniques (‚) and a pumped storage technique. Then a price system
made up of:
• a time-continuous electricity tariff p? 2 C Œ0; T
?
• a rental price for storage capacity rSt
?
• a price % > 0 for the produced final good
• the given prices for fuels and the generating capacities (viz., rF for thermal
capacity of type and w for its fuel, and rCo for the converter capacity)
and an allocation made up of:
• an output y? 2 L1
C Œ0; T from the thermal plant of type with
where ? WD O PS p? ; kSt ? ?
; kCo is the optimal price of energy
stock, i.e., the
unique solution to the programme (5.2.27)–(5.2.28) with p? I kSt ? ?
; kCo as the
data.24
(b) (Operating profit maximization by ESI) For each thermal plant type
(whose heat rate is 1= )
8
ˆ
ˆ f0g if p? .t/ < w
ˆ
<
y? .t/ 2 0; k? if p? .t/ D w for a:e: t (5.3.4)
ˆ
ˆ ˚
:̂ k? if p? .t/ > w
Z T
? 1
v D y? .t/ dt. (5.3.5)
0
And, with p? I kSt
? ?
; kCo as the data,
(which implies that the output from pumped storage is y?PS .t/ D kCo when
p? .t/ > ? .t/ and y?PS .t/ D kCo when p? .t/ < ? .t/).
2. (Profit maximization by User Industry) 25
O .z? ; n? / .
. p? ; 1/ 2 %? @F (5.3.7)
3. (Consumer utility maximization) For each h, x?h ; ' ?h ; m?h maximizes Uh on the
budget set
Z
T
.x; '; m/ 0 W p? .t/ x .t/ dt C %? ' C m MO h p? ; rSt
?
; %?
0
24
Since p? 2 C Œ0; T, the optimal is indeed unique: see [21] or [27, Lemma 8].
Since F is taken to be 1 outside of L1 O
C RC , @F contains a term arising from this
25
nonnegativity constraint. To spell this out, assume that F, as a function on its effective domain
L1
C RC , has a Mackey continuous, concave and Gâteaux-differentiable extension F
Ex
defined
1 .z? ? / .1=% ? / ? .z? ?/
on all of L R. Then (5.3.7) means that ; n 0 and p D r z F Ex
; n C
and 1=%? D @FEx =@n .z? ; n? / C for some 2 L1C vanishing a.e. on the set ft W z? .t/ > 0g,
with D 0 if n? > 0. (If p? were in L1 but not in L1 then would be an element of L1 C
concentrated on ft W z? .t/ g for each > 0.)
5.3 Long-Run Equilibrium with Pumped Storage or Hydro Generation of. . . 113
where
MO h . p; rSt ; %/ D mEn
h C & h St sup .rSt kSt G .kSt //
kSt
Z T
C & h IU sup %F .z; n/ p .t/ z .t/ dt n . (5.3.8)
z;n 0
4. (Market clearance)
X X X
y?PS C y? D z? C x?h and F .z? ; n? / D ' ?h . (5.3.9)
h h
Proof Given the results of Sect. 5.2, this is a formality—except for verifying the
absence of a duality gap. Note first that Conditions 2 to 5 of the theorem are simply
specializations, to the ESI case, of the corresponding parts of the definition of a long-
run equilibrium (Sect. 4.2). What has to be shown is the equivalence of the theorem’s
Condition 1 (optimal operation and valuation of the ESI’s plants) to the definition’s
Condition 1 (LRP maximization). As a general principle, this has been established
in Sect. 3.2 and restated in Sect. 3.4 (by taking account of Sect. 3.3). Its substance
is that, in the long run, competitive profit maximization is equivalent—as a system
of conditions on both quantities and prices—to the conjunction of: (i) maximization
of the operating profit a.k.a. short-run profit (which includes minimization of the
operating cost), (ii) minimization of the fixed-input value by shadow pricing (which
is identified as the dual programme), and (iii) equality of the maximum SRP to
the minimum FIV (absence of a duality gap). For each of the ESI’s plants, the SRP
and FIV programmes are spelt out in Sect. 5.2, and it remains only to show that their
values are equal. (Formally: (4.2.5)–(4.2.6) is (3.1.5) at equilibrium prices, which, as
is noted before the Comment in Sect. 3.4, is equivalent to the conjunction of (3.2.2)–
(3.2.3), (3.4.2) and (3.4.3). And, for the ESI’s technology, (3.2.2)–(3.2.3) and (3.4.2)
can be put as (5.3.4)–(5.3.6) and (5.3.1)–(5.3.3). It remains only to prove (3.4.3) for
each of the ESI’s plants.)
To this end, note first that the thermal operation LP (5.2.1)–(5.2.3) and its
dual (5.2.6)–(5.2.9) always have the same value: with in place of Th, the common
RT
value of both LPs is k 0 . p .t/ w /C dt for every . p; k ; w /, by (5.2.4) and
by (5.2.5) or (5.2.10). For pumped storage, however, the equality of values of the
operation LP (5.2.12)–(5.2.16) and its dual—in the form of either the standard
dual LP (5.2.17)–(5.2.22) or the equivalent CP (5.2.27)–(5.2.28)—relies
on the
properties of its data in the general equilibrium, p? I kSt ? ?
; kCo . The equality (the
absence of a duality gap) can be proved in two ways because it follows from either
? ?
of the two assumptions that kSt ; kCo 0 and that p? 2 L1 Œ0; T. Strict positivity
114 5 Short-Run Approach to Electricity Pricing in Continuous Time
where YPS is given by (5.1.3); formally, Lemma 6.2.3 applies.26 (A stronger result
can be obtained by applying the dual-value continuity criterion of [44, Theorem 18’
PS
(e)]: this shows that the convex function …SR .; kSt ; kCo / is norm-continuous on L1 ,
which implies that the concave function …PS SR . p; ; / is u.s.c. at .kSt ; kCo / for each
PS
p 2 L1 , by Lemma 6.1.1.) Finally, Lemma 6.1.1 shows that the equality …SR D
…PS
SR at . pI kSt ; kCo / follows from the upper semicontinuity, and ? a fortiori
from the
? 1 ?
continuity, of …PS SR . p; ; / at .k St ; k Co /. Since p 2 L and k ; k
St Co 0, either
method applies to this data point.
The counterpart result with hydroelectric generation (H) instead of pumped
storage (PS) is given next. The thermal technology remains the same, and its inputs
have fixed prices, r1F ; : : : ; r‚
F
and .w1 ; : : : ; w‚ /. The hydro turbine too has a fixed
F
price, rTu . There is a river with a single location where a dam can be constructed
to create a water reservoir of capacity kSt at a cost G .kSt /. The river has a fixed,
periodic flow, e .t/ at time t 2 Œ0; T, which (on the assumption of a constant head)
means a given energy inflow.27 Its price, .t/ at time t, is determined in long-run
RT
equilibrium. The river’s total rent is 0 e dt, and household h’s share of the rent is
& h Ri 0 (which adds up to 1 over h). Its share of profit from supplying the storage
capacity is & h St . As before, there is a single Industrial User of electricity (whose
production function is F), and the household’s share in his profit is & h IU .
26
For this technology, the Proof of Lemma 6.2.3 simplifies to a direct application of Berge’s
Maximum Theorem given in [6, S VI.3: Theorem 2]. This is because K is the finite-dimensional
space R2 , and because the set k2B YSR .k/ is itself bounded when B is (i.e., the operation vmax
is not needed). More generally, the same applies to c.f.c. techniques: see Proposition 7.4.2.
27
This assumption can be relaxed: it might be possible to improve the watershed to obtain a river
flow e at a cost GRi .e/, a convex function of e. The case of a fixed, unimprovable river flow e is
obtained by setting GRi .e/ equal to 0 for e D e and to C1 otherwise.
5.3 Long-Run Equilibrium with Pumped Storage or Hydro Generation of. . . 115
where ? WD O H p? I kSt? ?
; kTu I e is the optimal price
of water, i.e.,
the unique
solution to the programme (5.2.46)–(5.2.48) with p? I kSt? ?
; kTu I e as the data.
28
The assumption can be dropped, but this complicates the problem and then, as a result, an optimal
water price function need no longer be unique or continuous: see [23].
116 5 Short-Run Approach to Electricity Pricing in Continuous Time
(b) (Operating profit maximization by ESI) For each thermal plant type
(whose heat rate is 1= )
8
ˆ
ˆ f0g if p? .t/ < w
ˆ
<
y? .t/ 2 0; k? if p? .t/ D w for a:e: t (5.3.14)
ˆ
ˆ˚
:̂ k? if p? .t/ > w
Z T
1
v? D y? .t/ dt (5.3.15)
0
? ?
and, with p? I kSt ; kTu I e as the data,
?
(which implies that the hydro output is y?H .t/ D kTu when p? .t/ > .t/
and y?H .t/ D 0 when p? .t/ < ? .t/).
2. (Profit maximization by User Industry)
O .z? ; n? / .
. p? ; 1/ 2 %? @F (5.3.17)
3. (Consumer utility maximization) For each h, x?h ; ' ?h ; m?h maximizes Uh on the
budget set
Z
T
.x; '; m/ 0 W p .t/ x .t/ dt C % ' C m MO h p? ; rSt
? ? ?
; ?
;% ?
0
where
!
MO h . p; rSt ; %; / D mEn
h C & h St sup .rSt kSt G .kSt //
kSt
Z T Z T
C & h IU sup %F .z; n/ p .t/ z .t/ dt n C & h Ri .t/ e .t/ dt.
z;n 0 0
(5.3.18)
4. (Market clearance)
X X X
y?H C y? D z? C x?h and F .z? ; n? / D ' ?h . (5.3.19)
h h
Proof This is proved like Theorem 5.3.1 (taking into account the last of the
Comments on the valuation conditions at the end of Sect. 4.2).
Remark 5.3.3 (Value ? of Site for Reservoir) The rental value of the hydro or storage
? ?
site is rSt kSt G kSt per cycle (the reservoir’s value less its construction cost).
Comments (Multiple Storage Sites) A similar analysis applies when there is a
number of hydro sites or pumped-storage sites with different development cost
functions, Gl at location l. Reservoir capacity is then a good differentiated by its
location, and so is the river flow in the case of hydro. Therefore, some of the long-
run equilibrium prices and quantities may depend on l; the details follow.
? ?
• Consider first the case of pumped storage. Since @…PS SR =@kCo kSt;l ; kCo;l equals
F
rCo , which is independent of l, and since the derivative is homogeneous of degree
? ?
0 in .kSt ; kCo /, the equilibrium capacity ratio kSt;l W kCo;l is independent of l.29
?
Therefore, also the equilibrium price of storage capacity rSt is the same for each
l (since it equals @…PS SR =@kSt , which is homogeneous of degree 0). This is so
because the production technique has just one input whose supply cost depends
on the location (that the technique just two inputs in all is irrelevant here). By
? ?
contrast, the plant’s size does depend on l, since kSt;l meets the condition rSt 2
? ? ?
?
@Gl kSt;l for each l. The site’s rent, too, depends on l: it is rSt kSt;l Gl kSt;l .
• In hydro generation, both the reservoir construction cost function Gl and the fixed
river flow el depend on the location l. So, in hydro, the equilibrium capacity ratio
?
kSt;l ?
=kTu;l ?
, the price of reservoir capacity rSt;l and the shadow price of water ?l .t/
? ? ?
do all depend on l (as do the reservoir’s size kSt;l and the site’s rent rSt kSt;l
?
Gl kSt;l ).
Comment (Optimality of Thermal Output in Terms of the SRMC) Competitive
profit maximization by the thermal plants can be reformulated as SRMC pricing
by the thermal generating system, i.e., by using the thermal system’s instantaneous
SRMC curve. With a finite number of plant types, ‚, the SRMC curve is actually
a “right-angled” broken line30 : under (5.1.2), it consists of (i) the ‚ “horizontal”
segments
29
The ratio kSt W kCo is the (minimum) time it takes the converter to fully charge or discharge the
reservoir (starting from empty or full).
30
In a model with a “continuum” of plant types, the SRMC curve is a general “complete
nondecreasing curve”, in the terminology of [42, 24.3]. But even the continuum model does not
make the SRC curve differentiable: it still has a kink at the peak output, and typically it has offpeak
kinks, too—see [23].
118 5 Short-Run Approach to Electricity Pricing in Continuous Time
(where w‚C1 WD C1, and w0 WD 1 unless free disposal is included). The optimal
thermal output is where the “curve’s slope” equals the current price. In formal terms,
Condition (5.3.4) or (5.3.14) for each is equivalent to:
!
X
?
p .t/ 2 @y cSR y? .t/ I k1? ; : : : ; k‚
?
I w1 ; : : : ; w‚ for a.e. t
where cSR is the thermal system’s instantaneous short-run cost per unit time. With
1A denoting the 0-1 indicator of a set A (equal to 1 on A and to 0 outside), the
system’s instantaneous SRC of generating at a rate y can be given as
Z ‚
yX
cSR .yI k1 ; : : : ; k‚ I w1 ; : : : ; w‚ / WD w 1Œk1 C:::Ck1 ;k1 C:::Ck .q/ dq
0 D1
(5.3.21)
‚1
X
D w1 y C .w C1 w / .y .k1 C : : : C k //C
D1
P‚
if 0 y (otherwise cSR D C1). This his an increasing
D1 k i and convex (and
P‚
piecewise linear) function of the output rate y 2 0; D1 k , with cSR .0/ D 0.
The SRMC curve is the graph of the subdifferential correspondence y 7! @cSR .y/,
in the instantaneous quantity-price plane. When k > 0 for each ,
8
ˆ
ˆ .1; w1 if y D 0
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ fw g if y 2 .k1 C : : : C k1 ; k1 C : : : C k /
ˆ
<
@y cSR .y; .k / ; .w // D Œw ; w C1 if y D k1 C : : : C k and 1 ‚ 1 .
ˆ
ˆ
ˆ Œw ; C1/ if y D P‚ k
ˆ
ˆ
ˆ
ˆ ‚ D1
ˆ P‚
:̂ ; if y > D1 k or y < 0
(5.3.22)
(For the case of ‚ D 1, the SRMC and SRC curves have been used in Chap. 2 and
are shown in Figs. 2.1a and c; the supply and the subdifferential correspondences,
p 7! S .p/ and y 7! @cSR .y/, are inverse to each other.)
Chapter 6
Existence of Optimal Quantities and Shadow
Prices with No Duality Gap
by (B.5.2) and Lemma 3.11.14. The second inequality in (6.1.1) is strict if and only
if its sides are oppositely infinite. So … .k/ D … .k/ if and only if: (i) … .k/ D
usc … .k/, and (ii) … .k/ > 1 or … .k/ < C1 (i.e., it is not the case that both
… .k/ D 1 and … .k/ D C1).
The equivalence of Conditions 1 and 3 is proved similarly: now let … and …
mean …SR .; k; / and …SR .; k; /, which are functions on P W. Then
##
… lsc … … D… (6.1.2)
1
The primal (3.1.6)–(3.1.7) or the dual (3.3.6) is feasible if and only if YSR .k/ ¤ ; or
…LR . p; ; w/ ¤ C1, respectively. When Y is a cone (i.e., under c.r.t.s.), this means that
YSR .k/ ¤ ; or Yıp;w ¤ ;; the two sections are defined by (6.2.1) and (3.11.2).
6.1 Preclusion of Duality Gaps by Semicontinuity of Optimal Values 121
by (B.2.4) and Lemma 3.11.16. The second inequality in (6.1.3) is strict if and only
if its sides are oppositely infinite. So C .y/ D C .y/ if and only if: (i) C .y/ D
lsc C .y/, and (ii) C .y/ < C1 or C .y/ > 1.
The equivalence of Conditions 1 and 3 is proved similarly: now let C and C mean
CLR .y; ; / and CLR .y; ; /, which are functions on R W. Then
by (B.5.2) and Lemma 3.11.15. So C .r; w/ D C .r; w/ if and only if: (i) C .r; w/ D
usc C .r; w/, and (ii) C .r; w/ > 1 or C .r; w/ < C1.
Lemma 6.1.3 (Semicontinuity Conditions for Equality of SRC to Dual Value)
Assume that the production set Y is closed. Then, for each w 2 W, the following
conditions are equivalent to one another:
1. CSR .y; k; w/ D CSR .y; k; w/.
2. The convex function CSR .; ; w/ is lower semicontinuous at .y; k/, and the
primal (3.1.10)–(3.1.11) and the dual (3.3.4) are not both infeasible.
3. The concave function CSR .y; k; / is upper semicontinuous at w, and the pri-
mal (3.1.10)–(3.1.11) and the dual (3.3.4) are not both infeasible.
Proof This can be proved like Lemmas 6.1.1 and 6.1.2: to prove that Conditions 1
and 2 are equivalent, let C and C mean CSR .; ; w/ and C SR .; ; w/, which are
functions on Y K. Recall that
by (B.2.4) and Lemma 3.11.18. The second inequality in (6.1.5) is strict if and only
if its sides are oppositely infinite. So C .y/ D C .y/ if and only if: (i) C .y; k/ D
lsc C .y; k/, and (ii) C .y; k/ < C1 or C .y; k/ > 1.
The equivalence of Conditions 1 and 3 is proved similarly: now let C and C mean
CSR .y; k; / and C SR .y; k; /, which are functions on W. Then
by (B.5.2) and Lemma 3.11.17. So C .w/ D C .w/ if and only if: (i) C .w/ D
usc C .w/, and (ii) C .w/ > 1 or C .w/ < C1.
Comment (Type Two Semicontinuities of Optimal Values) Profit and cost are
always semicontinuous in prices; the dual values are semicontinuous in quantities.
That is, for every p, y, r, k, w and v:
1. (i) …SR .; k; / is l.s.c. convex on PW, (ii) CLR .y; ; / is u.s.c. concave on RW,
and (iii) CSR .y; k; / is u.s.c. concave on W.
2. (i) …SR . p; ; w/ is u.s.c. concave on K, (ii) CLR .; r; w/ is l.s.c. convex on Y, and
(iii) CSR .; ; w/ is l.s.c. convex on Y K.
122 6 Existence of Optimal Quantities and Shadow Prices with No Duality Gap
These results (which are part of Lemmas 3.11.13–3.11.18) follow directly from
the definitions: e.g., …SR .; k; / is the pointwise supremum of a family of contin-
uous (and linear) functions on P W (and likewise …SR . p; ; w/ is the pointwise
infimum of such functions on K). This shows also that …SR is proper convex in
. p; w/, and that …SR is proper concave in k, unless the one or the other is an infinite
constant. (What is more, …SR .; k; / and …SR . p; ; w/ are second conjugates of
…SR .; k; / and …SR . p; ; w/, respectively, by Lemmas 3.11.13 and 3.11.14.)
2
As in [44, p. 38], any point is regarded as an improper solution when the programme is infeasible.
The equality of values of proper solution pairs for finite LPs can be proved also by using the
simplex algorithm: see, e.g., [11, 5.1 and 9.1].
6.2 Semicontinuity of Cost and Profit in Quantity Variables Over Dual Banach. . . 123
Some of the conditions use sections of Y, viz., the short-run production set
3
This is a corollary to the Hahn-Banach Separation Theorem: see, e.g., [18, 12A: Corollary 1].
4
The set projY .Y/ need not be comprehensive downwards (i.e., it need not contain Y WD YC ).
124 6 Existence of Optimal Quantities and Shadow Prices with No Duality Gap
(which is the section of Y through k), and the long-run and short-run input
requirement sets
the sets of all the maximal and the minimal points for the vector order restricted
to Z. Such points represent the efficient output or input bundles. These are next
assumed to form norm-bounded sets—which for brevity are called bounded (but
they need not be order-bounded, unless the space is L1 with the supremum norm
kk1 ).
Production Set Assumptions (PSAs). Five assumptions are maintained from here
on (although not all five are always needed):
1. Y is a cone (with a vertex at 0).
2. Y is convex.
3. Y is weakly* closed, i.e., w.Y K V; Y 0 K 0 V 0 /-closed.
Y includes free disposal of inputs and of producible Q
4. outputs; i.e., if k k,
v v, Q
Q y yQ 2 projY .Y/ and .y; k; v/ 2 Y, then yQ ; k; vQ 2 Y.
5. ILR .y/ KC VC (i.e., Y Y KC VC ).
The following assumptions are made selectively
where needed:
6. For every .k; v/ 2 ILR .y/, there exists a k; Q vQ 2 vmin ILR .y/ with k; Q vQ
.k; v/. S
7. For every bounded set B
Y, also the set y2B vmin ILR .y/ is bounded.5
8. For every .y; v/ 2 YSR .k/, there exists a .Qy; v/ Q 2 vmax YSR .k/ with
.Qy; v/
Q .y; v/. S
9. For every bounded set B
K, also the set k2B vmax YSR .k/ is bounded.
10. For every v 2 ISR .y; k/, there exists a vQ 2 vmin ISR .y;
Sk/ with vQ v.
11. For every bounded set B
Y K, also the set .y;k/2B vmin ISR .y; k/ is
bounded. (This follows from either of PSAs 7 and 9.)
The above PSAs are similar to the conditions of [13, p. 134] and [14, p. 580] for
the finite-dimensional case (see also the end of this section for further comments).
When the commodity space is Rn or a general dual Banach lattice (e.g., L1 or L%
with % > 1), the assumptions that efficient points exist (PSAs 6, 8 and 10) can be
derived from simpler conditions by using the following lemma.
5
If L1 is the space Y in PSA 7, or K in PSA 9, or Y K in PSA 11, then it obviously suffices to
make this assumption for each singleton set (instead of B).
6.2 Semicontinuity of Cost and Profit in Quantity Variables Over Dual Banach. . . 125
for each p 2 L0 . This exhibits yH as the w .L; L0 /-limit of a net in B (the identity
map on H can serve as such a net). So yH 2 B (since B is weakly* closed). Thus
the assumption of Zorn’s Lemma is verified for as a partial order on B (and so a
maximal point exists).
Corollary 6.2.2 (Existence of Efficient Points) Assume PSA 3. Then:
1. PSA 8 holds if the set
which are weakly* closed by PSA 3, and are bounded (even order-bounded) by
PSA 5.
6
This construction is used for proving related but different results in, e.g., [2, 14.11] and [33, X.4:
Theorem 6].
126 6 Existence of Optimal Quantities and Shadow Prices with No Duality Gap
To exploit the relative weak* compactness of the efficient boundary of the short-
run production set (PSA 9), one needs the maximand h p; w j ; i to be weakly*
continuous (i.e., p and w to be in Y 0 and in V 0 ). As is shown next, this guarantees
that …SR is weakly* u.s.c. in k (and obviously the condition on p and w is restrictive
only when the spaces are infinite-dimensional and nonreflexive, i.e., when Y 0 ¤ Y
or V 0 ¤ V ).
Lemma 6.2.3 (Semicontinuity of SRP in Fixed Inputs) Under PSAs 8 and 9 in
addition to PSAs 2 and 3, the concave function …SR . p; ; w/ W K ! R [ f1g is
0 0 7
K 0 -upper semicontinuous (on K) for every . p; w/ 2 YC VC .
Proof Let … mean …SR . p; ; w/. That … .k/ < C1 for every k 2 K follows from
PSAs 8 and 9 for B D fkg. Next, since … is concave, it suffices to prove that
it is u.s.c. for the bounded weak* topology, i.e., that … is weakly* u.s.c. on any
norm-bounded set B
K D K 0 . (This is because every bw -closed convex set is
w -closed, by the Krein-Smulian Theorem: see, e.g., [18, 18E: Corollary 2].) And a
bound on k implies a bound on the efficient combinations of y and v (which are the
only ones that matter because . p; w/ 0). So Berge’s Maximum Theorem applies.
In precise terms: the set
[
A WD clw.YV;Y 0 V 0 / vmax YSR .k/
k2B
7
Also, under PSA 4, if 0 2 Y then …SR . p; ; w/ 0 on KC (outside of KC , it equals 1).
8
Another way to apply here Berge’s Maximum Theorem [6, VI.3: Theorem 2] is to take h p j yi
hw j vi ı .y; k; v j Y/ as the maximand (which is u.s.c. in .y; k; v/) and A as the constraint set
(which is compact and independent of k 2 B).
6.2 Semicontinuity of Cost and Profit in Quantity Variables Over Dual Banach. . . 127
Proof This is proved like Lemma 6.2.3: since the function C WD CLR .; r; w/ is
convex (on Y), it suffices to prove that C is l.s.c. for the bounded weak* topology,
i.e., that C is weakly* l.s.c. on any norm-bounded set B
Y D Y 0 . And a bound
on y implies a bound on the efficient combinations of k and v (which are the only
ones that matter because .r; w/ 0). So Berge’s Maximum Theorem applies.
In precise terms: the set
[
A WD clw.KV;K 0 V 0 / vmin ILR .y/
y2B
is w .V; V 0 /-compact by PSA 11 and the Banach-Alaoglu Theorem. Also, for every
.y; k/ 2 B,
by PSA 10. Since w 2 V 0 , the minimand in (6.2.6) is w .V; V 0 /-l.s.c. (and actually
continuous) in v. In addition, since Y is w.Y K V; Y 0 K 0 V 0 /-closed by
PSA 3 (and since A is compact), the constraint correspondence .y; k/ 7! ISR .y; k/ \
128 6 Existence of Optimal Quantities and Shadow Prices with No Duality Gap
9
Both superscripts can of course be suppressed when the quantity space is reflexive (i.e., when its
predual equals its dual, e.g., when it is finite-dimensional).
6.2 Semicontinuity of Cost and Profit in Quantity Variables Over Dual Banach. . . 129
0
• Finally, let …SR . p; k; w/ and …SR . p; k; w/ denote the optimal values of (3.3.6)
0
with, respectively, K 0 or K as the range for r; like …SR , also …SR is defined for
0
every p 2 Y and w 2 V (and every k 2 K), with …SR …SR everywhere
(since K 0 K ).
0
But if p 2 Y 0 and w 2 V 0 then …SR D …SR at . p; k; w/—and there are
corresponding equalities for CLR and CSR . This is because
By the very definition, @0y F D Y 0 \@y F (everywhere) for every convex function F
on Y, say. So the superdifferential @O 0 …SR (or the subdifferential @0 CLR or @0 CSR ) is
k y y;k
always equal to K 0 \ @O k …SR (or to Y 0 \ @y CLR or .Y 0 K 0 / \ @y;k CSR , respectively).
These are, however, marginal values of Type One, which fail to exist when there
10
This is of course true whenever the “primed” and “starred” dual values are equal, whether or not
the common dual value equals the primal value.
11
Either notation, @0 or @ —but not both unless the quantity space is reflexive—may be abbreviated
to @. In [26] @ means @ , but in [27] and [30] @ means @0 (for a functions on L1 , whose Banach
predual is L1 ). The unembellished symbol @ is also used whenever a single, fixed pairing of spaces
is considered (as in Sects. 3.1 to 6.1 and Appendix B).
130 6 Existence of Optimal Quantities and Shadow Prices with No Duality Gap
are duality gaps—and may thus differ from the dual solutions. The latter are always
equal to marginal values of Type Two, by the Dual Shephard-Hotelling Lemmas
(Lemmas 3.11.2, 3.11.4 and 3.11.6). For clarity, note that there are four marginal
values of Type Two that correspond to the four combinations of 0 and in the
subdifferential and in the dual value itself. Their intersections with the predual space
are, however, all equal unless the dual value depends on whether the predual or the
dual (of the quantity space) serves as the range of the dual variable. For example, for
0
the SRP programme with its dual, the four derivatives are: @O 0k …SR , @O k …SR , @O 0k …SR ,
0 0
@O k …SR . If …SR D …SR at . p; k; w/ then (at that data point)
def
@O 0k …SR D K 0 \ @O k …SR D K 0 \ fsolutions to (3.3.6) with R D K g
˚ 0 def 0
D solutions to (3.3.6) with R D K 0 D @O 0k …SR D K 0 \ @O k …SR
0
where the third equality follows from the assumption (that …SR D …SR ), and the
second and the fourth equalities hold by the Dual Hotelling’s Lemma for the short
run (Lemma 3.11.2 with R D K or K 0 , respectively). If additionally …SR D …SR D
0
…SR at . p; k; w/—which holds when p 2 Y 0 and w 2 V 0 by Lemmas 6.2.3 and 6.1.1
together with (6.2.9)—then these marginal values are actually of Type One, being
equal to @O 0k …SR D K 0 \ @O k …SR .
def
can prove only that C is weakly* l.s.c. on every ball B.12 The Krein-Smulian
Theorem upgrades this result to weak* l.s. continuity on Y (and not just on each
B). To obtain a continuity result on Y without this step, one would have to put
the norm topology on Y to make B a neighbourhood—and then the conclusion
would be weaker, viz., only that C is norm-l.s.c. on Y (i.e., that it is Y -l.s.c. but
not that it is Y 0 -l.s.c.).
• The bounded weak* topology can be equally useful in other contexts. In [21] and
[23] the Krein-Smulian Theorem is used to show that the production sets for the
technologies of energy storage and hydroelectric generation are weakly* closed
(in an L1 space). Another mathematical tool for “localization” to bounded sets
is devised in [25], and both tools are combined there for a simple proof that the
additively separable consumer utility function is Mackey continuous on L1 C.
• Since duality of programmes is symmetric, absence of a duality gap can be
proved also by showing that the dual value is semicontinuous in the dual
parameter (instead of showing that the primal value is semicontinuous in the
primal parameter, i.e., by verifying Condition 3, instead of Condition 2, of
Lemmas 6.1.1–6.1.3).
Comments (on the Production Set Assumptions)
• PSA 9 formalizes the notion that fixed inputs impose capacity constraints.13
• Unlike the fixed inputs k, the variable inputs v alone need not impose any bound
on the output’s norm kyk: see (5.1.1).
• Unlike the inputs k and v, which are always nonnegative by PSA 5, the “output”
can be a signed bundle y D yC y , where y˙ are the nonnegative and
nonpositive parts. This is convenient when, e.g., y represents a single good
differentiated over time, and the dated commodities cannot be classed as net
inputs or net outputs a priori . For example, when y is an output from storage,
it is always signed, i.e., yC ¤ 0 ¤ y unless y D 0: see (5.1.3).14
In addition to the semicontinuity of …SR , CLR and CSR (which rules out duality
gaps), the Production Set Assumptions of Sect. 6.2 guarantee also the solubility of
the primal programmes of SRP, LRC and SRC optimization when p, r and w are
12
For Berge’s Theorem to apply, the (efficient) range of the decision variable must be contained in
a weak* compact that is independent of the parameter (y) as it ranges over a set B
Y—and so B
must be bounded. (The result stated in [44, Example 4’ after (5.13)] also applies, but it is a special
case of Berge’s.)
13
PSAs 7, 9 and 11 make it possible (in Lemmas 6.2.3–6.2.5) to prove semicontinuity of profit and
costs at every . p; r; w/ 0 (and not only at strictly positive prices as is done, for finite-dimensional
spaces, in [13] and [14]).
14
By contrast, when a signed y can arise only from free disposal, the good is essentially a net
output.
132 6 Existence of Optimal Quantities and Shadow Prices with No Duality Gap
in the predual spaces (Y 0 , K 0 and V 0 ). This is because the relevant subsets of the
constraint set (Y) are then weakly* compact (and so Weierstrass’s Extreme Value
Theorem applies). This is next recorded formally.
Proposition 6.3.1 (Solubility of SRP Programme) Under PSAs 3, 8 and 9,15 if
0 0
. p; w/ 2 YC VC and the short-run profit maximization programme (3.1.6)–(3.1.7)
is feasible, then it has a solution.
Proof A fixed k imposes a bound on the efficient combinations of y and v (as in the
Proof of Lemma 6.2.3). In precise terms, the set
by PSA 8 (as part of which, E .k/ ¤ ; if YSR .k/ ¤ ;). Since . p; w/ 2 Y 0 V 0 , the
maximand in (6.3.1) is w .Y V; Y 0 V 0 /-u.s.c. (and actually continuous) in .y; v/
jointly. So, by Weierstrass’s Theorem, it attains its supremum on E .k/, which equals
its supremum on YSR .k/ by (6.3.1).
Proposition 6.3.2 (Solubility of LRC Programme) Under PSAs 3, 6 and 7,16 if
0 0
.r; w/ 2 KC VC and the long-run cost minimization programme (3.1.8)–(3.1.9) is
feasible, then it has a solution.
Proof A fixed y imposes a bound on the efficient combinations of k and v (as in the
Proof of Lemma 6.2.4). In precise terms, the set
by PSA 6 (as part of which, E .y/ ¤ ; if ILR .y/ ¤ ;). Since .r; w/ 2 K 0 V 0 , the
minimand in (6.3.2) is w .K V; K 0 V 0 /-l.s.c. (and actually continuous) in .k; v/
jointly. So, by Weierstrass’s Theorem, it attains its infimum on E .y/, which equals
its infimum on ILR .y/ by (6.3.2).
15
Here, it suffices to assume PSA 9 for B D fkg, i.e., that vmax YSR .k/ is bounded for each k 2 K.
16
Here, it suffices to assume PSA 7 for B D fyg, i.e., that vmin ILR .y/ is bounded for each y 2 Y.
6.4 Continuity of Profit and Cost in Quantities and Solubility of Shadow-. . . 133
To ensure Type One continuity—and not just semicontinuity—of the optimal value,
it suffices to impose Slater’s Condition on the programme.18 In turn, Type One
continuity of the primal value guarantees not only that there is no duality gap but
also that a dual solution exists (and can be obtained as a cluster point of any sequence
of approximate optima): see, e.g., [44, Theorem 17]. As is spelt out next, this applies
to the value function …SR . p; ; w/ when its domain, K, carries the norm topology.
(A weaker topology would not do because the effective domain of …SR is typically
KC , and to have a nonempty interior it must carry the norm topology in addition
to having a nonempty core a.k.a. algebraic interior.) A similar result is given for
CLR .; r; w/—but not for CSR .; ; w/ because, without modifications, it would be
vacuous in the cases of most interest: see a Comment at the end of this section.
Lemma 6.4.1 (Solubility of Dual to SRP Programme) Assume PSAs 8 and 9 of
Sect. 6.2. If a k 2 K has a norm-neighbourhood N for which there exists a .y; v/
17
Here, it suffices to assume PSA 7 for B D f.y; k/g, i.e., that vmin ISR .y; k/ is bounded for each
y 2 Y and k 2 K.
18
The continuity derived from Slater’s Condition holds locally —on a neighbourhood of a particular
parameter point—unlike the global semicontinuity obtained in Lemmas 6.2.3–6.2.5.
134 6 Existence of Optimal Quantities and Shadow Prices with No Duality Gap
Q v 2 Y for every kQ 2 N then, for each . p; w/ 2 YC
such that y; k;
VC :
1. The concave function …SR . p; ; w/ W K ! R [ f1g is finite and norm-
continuous at k (and hence K -u.s.c. at k).19
2. So its superdifferential at k is nonempty when K is paired with its norm-dual K
as R, i.e., in the notation (6.2.10),
@O k …SR . p; k; w/ ¤ ;.
19
As in Lemma 6.2.3, …SR < C1 everywhere by PSAs 8 and 9 with B D fkg.
20
As in Lemma 6.2.4, CLR 0 everywhere by PSA 5.
6.4 Continuity of Profit and Cost in Quantities and Solubility of Shadow-. . . 135
Such technologies describe the case of completely rigid industrial plants. Examples
have already been encountered here in the context of electricity (Sect. 5.1): both
thermal generation and pumped storage, though not hydro, are such techniques. By
definition, a production technique has conditionally fixed coefficients (c.f.c.) if the
conditional input demands are price-independent, i.e., if the cost-minimizing input
quantities are functions not of the input prices .r; w/, but of the output bundle y
alone. Denoted by kL .y/ and vL .y/, these are the (least) input requirements for a
fixed input 2 ˆand a variable input 2 „: the input requirement set is an orthant
with kL .y/ ; vL .y/ as its vertex. In addition to being constrained by the available
input quantities, any producible output bundle may R be subject also to a constraint
that applies in the long as in the short run: e.g., y dt D 0 when y is the net flow
from storage, as in (5.1.3). In these terms, the long-run production set for a c.f.c.
technique is the convex cone
n o
Y D .y; k; v/ W kL .y/ k; vL .y/ v; y 2 Y0 (7.1.1)
where each of the (real-valued) functions kL and vL is: (i) sublinear, i.e., convex and
positively linearly homogeneous (p.l.h.) on Y, and (ii) nonnegative on Y0 , which is
a convex cone in the output space Y. Usually
˚ ˝ ˛
Y0 D y W aj j y D 0; bl .y/ 0 for j 2 J; l 2 L (7.1.2)
The finite sets J and L may both be empty (in which case Y0 D Y and Y0ı D f0g).
Note, also, that unless the output y is a scalar, this is not the ordinary fixed-
coefficients technology: see also the Comment at the end of this section.
A direct route to characterizing a long-run producer optimum in terms of the
functions kL and vL is to note that, for r 0 and w 0,
and to use either the LRC Optimization System (3.9.12)–(3.9.14) or its differ-
ential equivalent (3.9.15)–(3.9.16) or, easiest of all, the conjunction of (3.9.12)
and (3.9.16). In the c.f.c. case, it is no problem to split the joint programme (3.1.8)–
(3.1.9) for k and v in (3.9.12): the optimal k’s and v’s can be found separately from
each other (as functions of y); equivalently, @r;w CLR D @r CLR @w CLR . When r 0
and w 0 (and y 2 Y0 ), the unique optima are kL .y/ and vL .y/—so using this
notation in the context of a c.f.c. technique is essentially consistent with the earlier
meaning of vL .y; k; w/ and kL .y; r; w/ for a general technology (Sects. 3.2 and 3.9
after (3.9.22)). When r and w are nonnegative but not strictly positive, the solution
set for (3.9.12) is not just the single point kL .y/ ; vL .y/ : it is the Cartesian product
of the suborthants
n o
KL .y; r/ WD k W k kL .y/ ; r k kL .y/ D 0
N .y j Y0 / WD @ı .y j Y0 / D f 2 Y0ı W h j yi D 0g (7.1.9)
8 9
<X X =
D ˛ j aj O l .y/ W ˛ 2 RJ ; ˇ 2 RL ; ˇ 0; ˇ b .y/ D 0 .
ˇ l @b
: ;
j lW ˇ l ¤0
(7.1.10)
1
At 0, the normal cone to Y0 equals the polar cone Y0ı . When Y0 is a vector subspace of Y, as
in (5.1.7), the normal cone is the same at every y: it is the annihilator space (a.k.a. orthogonal
complement) Y0? .
2
For example, the function kL .y/ WD EssSup .y/ WD ess supt y .t/, for y 2 Y WD L1 Œ0; T, has
no subgradient in P WD L1 Œ0; T at any y with meas ft W y .t/ D EssSup .y/g D 0. This is because
RT
2 L1 \ @a EssSup .y/ if and only if 0, 0 .t/ dt D 1 and D 0 on ft W y .t/ < EssSup .y/g:
see, e.g., [32, 4.5.1: Example 3].
140 7 Production Techniques with Conditionally Fixed Coefficients
The dual programme (3.3.13)–(3.3.14) consists in plant valuation ; this is the stan-
dard dual of (7.1.11)–(7.1.13), and so its variables (r) are the Lagrange multipliers
of the primal inequality constraints. It can be formulated as the following CP:
3
To see that (7.1.18)–(7.1.21) are indeed the FFE Conditions, recall from (3.1.5) that primal and
dual feasibilities mean that .y; k; v/ 2 Y and . p; r; w/ 2 Yı . In the c.f.c. case, the two
feasibility conditions expand into (7.1.18) and (7.1.19)–(7.1.20).
7.1 Producer Optimum When Technical Coefficients Are Conditionally Fixed 141
h p j yi r kL .y/ C w vL .y/ r k C w v
so (7.1.22) and (7.1.6) together are equivalent to (7.1.21), i.e., to the equality of
values at the two feasible points.
• The FFE Conditions on y and r to solve the (reduced) profit-maximizing
operation programme (7.1.11)–(7.1.13) and its dual (7.1.14)–(7.1.17) are the
same as (7.1.18)–(7.1.21), but with v D vL .y/, which makes the last part
of (7.1.18) redundant (also, that w 0 is now an assumption needed for the
reduction).
• The Kuhn-Tucker Conditions (7.1.4)–(7.1.7) can be derived also by using,
instead of the LRC function (7.1.3), the SRC function
w vL .y/ if kL .y/ k and y 2 Y0
CSR .y; k; w/ D (7.1.23)
C1 otherwise
Since the function CSR represents, by (7.1.23), the capacity constraints as well
as the variable cost actually incurred, the sum representing the (multi-valued)
SRMC in (7.1.24) contains capacity premia 2 r @kL .y/, where each r
is nonnegative but otherwise completely undetermined by pure short-run cost
calculations.5 This is the short-run counterpart of the LRMC’s decomposi-
tion (7.1.7).
4
The term corresponding to any inactive capacity constraint must be deleted from the sum
in (7.1.24) as expanded in (7.1.8).
5
This term is an outward normal vector to the intersection of the sublevel sets of kL ’s in (7.1.23):
see, e.g., [42, 23.7.1 and 23.8.1] or [32, 4.3: Propositions 1 and 2].
142 7 Production Techniques with Conditionally Fixed Coefficients
can have ordinary derivatives w.r.t. each input quantity, k or v . Then, a fortiori ,
also the (maximum) SRP function is differentiable in k—and so the capital inputs
have definite and separate marginal values, whose ratio is a well-defined rate of
substitution (@…SR =@k1 W @…SR =@k2 ). This is so with, e.g., the storage technique
(5.1.3) when the good’s price is a continuous function of time.8 Such substitution
between perfectly complementary inputs would, of course, be impossible with a
homogeneous, one-dimensional output good: in such a case the output from an input
bundle .k; v/ could only have the familiar form of the fixed-coefficients production
function, min fk1 ; : : : ; v1 ; : : :g. But with a multi-dimensional, differentiated output
good, perfect complementarity of inputs would imply fixed proportions between
them only if the proportions between the outputs were fixed—and they are not. With
output proportions allowed to vary, it is the output price system p that aggregates an
output bundle y into a scalar revenue; and, given a suitable p, substitution in revenue
terms is possible. With multiple outputs, inputs can be perfect complements without,
like nuts and bolts, having always to be used in a fixed proportion.
This section gives proofs for Sect. 7.1: the polar and normal cones to the production
cone of a technique with conditionally fixed coefficients are calculated, and the
formulae are then used to specialize the dual programmes of Sect. 3.3 to such
technologies. In particular, it is proved that the dual of the SRP programme is
indeed (7.1.14)–(7.1.17), and it is shown that the Kuhn-Tucker Conditions are
indeed (7.1.4)–(7.1.7).
6
This is the borderline case between Hicks-Allen complements and substitutes: see, e.g., [47,
1.F.d].
7
Formally, the multi-valued rate of substitution equals RC D Œ0; C1/.
8
Shown in [21] and in [27, Section 9], the result is summarized and used in Sects. 5.2 and 5.3 here.
7.2 Derivation of Dual Programmes and Kuhn-Tucker Conditions 143
Lemma 7.2.1 (Polar and Normal to Production Cone with c.f.c.) Assume that
the production set Y has the form (7.1.1)–(7.1.2), where:
1. kL W Y ! R and vL W Y ! R are sublinear (p.l.h. convex), aj W Y ! R is linear,
and bl W Y ! R is superlinear (p.l.h. concave), with kL .0/ D 0, vL .0/ D 0 and
bl .0/ D 0 (for each , , j and l ˝in some˛ finite sets ˆ, „, J and L).
2. There exists a y0 2 Y such that aj j y0 D 0 for each j 2 J and bl .y0 / > 0 for
each l 2 L.9
3. All but at most one of the functions kL , vL and bl (for 2 ˆ, 2 „ and l 2 L)
are continuous for m .Y; P/, the strongest locally convex topology that makes P
the continuous dual of Y. All the linear functionals aj belong to P (for j 2 J).10
Then, for every .y; k; v/ 2 Y,
8 9
<X X =
N .y j Y0 / D ˛ j aj O l .y/ W ˛ 2 RJ ; ˇ 2 RL ; ˇ b .y/ D 0
ˇ l @b C
: ;
j lW ˇ l ¤0
(7.2.1)
8 9
<X X =
Y0ı D ˛ j aj O l .0/ W ˛ 2 RJ ; ˇ 2 RL
ˇ l @b (7.2.2)
C
: ;
j lW ˇ l ¤0
˚
N .y; k; v j Y/ D . p; r; w/ 2 P Rˆ „
C RC W .7.1.6/ and .7.1.7/ hold
(7.2.3)
n o
Yı D . p; r; w/ 2 P Rˆ „ L
C RC W p 2 r@k .0/ C w@vL .0/ C Y0ı . (7.2.4)
9
When y is a decision variable, as in the SRP programme, this is Slater’s Condition on the
constraints (inequalities and equations) that define Y0 .
10
For a linear functional aj , its m .Y; P/-continuity is equivalent to its w .Y; P/-continuity (and
it simply means that aj 2 P). But a concave function (bl ) or a convex function (kL or vL ) can
be m .Y; P/-continuous (and hence w .Y; P/-u.s.c. or l.s.c., respectively) without being w .Y; P/-
continuous. The weak and Mackey topologies are briefly discussed in Sect. 6.2.
144 7 Production Techniques with Conditionally Fixed Coefficients
by: (i) the Factorization Lemma (a.k.a. Sard’s Quotient Theorem) given in, e.g.,
[18, 21A] and [32, 0.1.4: Corollary],11 and (ii) the formula for the (outward) normal
cone to a sublevel set of a convex function (reoriented here to a concave function’s
superlevel set), which is given in, e.g., [32, 4.3: Proposition 2], [42, 23.7.1] and [48,
P P
7.8]. In other words, p 2 Na .y j Y0 / if and only if p D j ˛ j aj l ˇ l @O a bl .y/ for
some ˛ and ˇ 0 with ˇ b .y/ D 0. This proves (7.2.1), and (7.2.2) is a case
of (7.2.1) for y D 0 (since Y0ıa D Na .0 j Y0 /).
Note that both the decomposition (7.2.5) and the representation (7.2.6) rely on
Assumption 2. First, it guarantees that all but one of the sets in question T have a
common interior point that lies also in the excepted set: y0 2 ker a D j ker aj ,
and y0 2 fbl > 0g D cor fbl 0g D intY;TSLC fbl 0g by (B.3.4). So (B.3.7) applies
and gives (7.2.5). Second, the existence of a point y0 satisfying the inequality strictly
is what validates the formula for the normal cone, which gives the second term
in (7.2.6).
The same arguments apply to Y, which is the intersection of Y0 Rˆ R„ with the
sublevel sets of kL .y/ k , etc., as functions of .y; k; v/. Their subdifferentials
are: @a kL f.0; : : : ; 1; 0; : : :/gf.0; : : : ; 0/g, etc. And these card ˆCcard „ sublevel
sets do have a common interior point that lies in Y0 Rˆ R„ : e.g., .0; k; v/ with
any .k; v/ 0 will do (since each kL or vL is TSLC -continuous by Assumption 1).
So, for every .y; k; v/ 2Y, one has . p; r; w/ 2 Na .y; k; v j Y/ if and only if:
r 0, w 0, r k kL .y/ D 0, w .v vL .y// D 0 and p 2 r@a kL .y/ C w@a vL .y/ C
Na .y j Y0 /. This proves (7.2.3), and (7.2.4) is a case of (7.2.3) for y D 0, k D 0 and
v D 0.
Now that (7.2.1)–(7.2.4) have been proved for P D Y a , their extension to any P
follows from Assumption 3 and the fact that if vectors p2 ; : : : ; pn all belong to P
then: p1 C : : : C pn 2 P if and only if p1 2 P. By Assumption 3 and (B.3.10), all
but at most one of the algebraic subdifferentials—@akL .y/, @a vL .y/, @O a bl .y/—lie
11
[42, 22.3.1] and [48, 4.19] give only Farkas’s Lemma, but this contains the Factorization Lemma.
7.2 Derivation of Dual Programmes and Kuhn-Tucker Conditions 145
where the last (third) equality holds by definition (@ WD P \ @a ), and the penultimate
(second) equality holds because P \ %D D % .P \ D/ for every D Y a and every
real number % ¤ 0,12 whilst P \ 0D D f0g for every D ¤ ;; this is applied
to D D @a kL .y/, @a vL .y/, @O a bl .y/. This shows that (7.2.3)–(7.2.4) hold as stated
(i.e., also when P ¤ Y a in @ D P \ @a and in N D P \ Na ). Finally, the same
arguments—but with all the functions kL and vL left out, or replaced by zeros—
derive (7.2.1)–(7.2.2) for a general space P from the case of P D Y a .
The formula for Yı is next used to spell out all the dual programmes (when Y is
a c.f.c. technique).
Corollary 7.2.2 (Dual to SRP Programme with c.f.c.) On the assumptions of
Lemma 7.2.1, the dual of the profit-maximizing operation programme (7.1.11)–
(7.1.13), with k as the primal parameter, is the plant valuation programme (7.1.14)–
(7.1.17) with Y0ı given by (7.2.2).
Proof Apply Proposition 3.10.1 and (7.2.4) with (7.2.2).
As has already been noted (in a Comment in Sect. 7.1), the formula (7.2.4) for
Yı shows also that the system (7.1.18)–(7.1.21) is the FFE characterization of a
solution pair (to the SRP or LRC or SRC programme together with its dual). And,
by using Euler’s Theorem, this FFE characterization has been proven equivalent
to (7.1.4)–(7.1.7). What still remains to be shown is that the latter system, which has
already been referred to as the Kuhn-Tucker Conditions, is indeed an expanded form
of the Kuhn-Tucker Lagrangian Saddle-Point Conditions. This is next done for the
SRP programme (the LRC and SRC cases being similar). Identification of (7.1.4)–
(7.1.7) as the Saddle-Point Conditions—which is known, from the general theory
of CPs, to be equivalent to optimality and absence of a duality gap—will also
12
For % D 0, this fails if and only if D ¤ ; but P \ D D ; (in which case P \ 0D D f0g but
0 .P \ D/ D ;).
146 7 Production Techniques with Conditionally Fixed Coefficients
It remains to be shown that this sum can be subdifferentiated term by term (and then
to apply (7.2.1) to expand @ı .y j Y0 / D N .y j Y0 /). This is done in the same way as
in the Proof of Lemma 7.2.1. First, note that
@a r kL C w vL C ı . j Y0 / .y/ D @a r kL .y/ C @a .w v/
L .y/ C @a ı .y j Y0 /
(7.2.8)
D r@a kL .y/ C w@a vL .y/ C Na .y j Y0 /
13
This is a case of the ordinary Lagrangian (with Y0 as an “abstract” constraint, unpriced by L):
see, e.g., [44, (4.4)], where it is derived from the generalized Lagrangian defined in [44, (4.2)].
7.2 Derivation of Dual Programmes and Kuhn-Tucker Conditions 147
for every y 2 Y0 (for y … Y0 , both sides equal ;). This is because each kL or
vL , being a finite convex function, is TSLC -continuous (everywhere on Y). The only
other function, ı . j Y0 /, may have no point of continuity, but this does not matter:
since all but one of these functions are continuous, the algebraic subdifferential @a
is an additive operator by (B.3.6). And it is p.l.h. by (B.3.8) with (B.3.10).
Additivity of @ WD P \ @a follows from that of @a by using Assumption 3 on
the functions involved, as in the Proof of Lemma 7.2.1. This is only sketched. Say,
for simplicity, that L D ;, i.e., Y0 D ker a and so Y0ı D span fag P. All but at
most one of the sets @a kL .y/ and @a vL .y/ lie wholly in P, by (B.3.10). So the sum
of their elements (one from each set) belongs to P if and only if each term does. (If
p2 ; : : : ; pn all belong to P then: p1 C : : : C pn 2 P if and only if p1 2 P.) This means
that (7.2.8) holds also with @a replaced by @ WD P \ @a .
Comments
• Instead of deriving the dual programme (7.1.14)–(7.1.17) straight from the primal
(as in the Proof of Proposition 3.10.1, which is applied in the earlier Proof of
Corollary 7.2.2), one can obtain the dual through the Lagrange function—since
the dual of a maximization programme consists in minimizing, over the dual
variables, the supremum of the Lagrange function over the primal variables: see,
e.g., [44, (4.6) or Example 1’: (5.1)]. Here, this means minimizing supy L .y; r/
over r. Denote14
14
…Exc .y/ is the excess a.k.a. pure profit from an output y (i.e., revenue at prices p less minimum
input cost at prices r and w).
148 7 Production Techniques with Conditionally Fixed Coefficients
To apply the results of Sects. 6.2 to 6.4, which rule out duality gaps and ensure that
both programmes (the primal and the dual) are soluble, one needs to verify the PSAs
of Sect. 6.2. This is next done for techniques with conditionally fixed coefficients,
when the output space Y is the dual of a Banach lattice Y 0 .
15
The partial derivatives @…SR =@
0j and @…SR =@
00l exist if the ˛ j and ˇ l associated by (7.2.1)
with the optimal y are unique. If not, the derivative property still holds for the superdifferential,
i.e., @O
0 ;
00 … contains each .˛; ˇ/ that satisfies (7.2.1) for some optimal y (and hence for every
optimal y).
7.3 Verification of Production Set Assumptions 149
Lemma 7.3.1 (Properties of Production Set with c.f.c.) Assume that Y is given
by (7.1.1), i.e., that .y; k; v/ 2 Y if and only if
L Y ! Rˆ and vW
where16 kW L Y ! R„ are sublinear maps (with kL .0/ D 0 and
vL .0/ D 0) that are nondecreasing and nonnegative on Y0 , which is a convex cone
in Y (and ˆ and „ are finite sets). Then:
1. Y satisfies PSAs 1, 2, 4, 5, 6 and 10.
2. If kL and vL are w .Y; Y 0 /-lower semicontinuous (for each 2 ˆ and 2 „)
and Y0 is w .Y; Y 0 /-closed, then Y satisfies PSA 3 (i.e., it too is weakly* closed).
3. If kL and vL are norm-continuous (on Y), then Y satisfies PSA 7 (and hence also
PSA 11). n o
4. Under the assumptions of Part 2, if the set y 2 Y0 W kL .y/ k is bounded for
each k,17 then Y satisfies PSA 8.
If additionally vL is norm-continuous (for each 2 „),18 then Y satisfies
PSA 9 (and hence also PSA 11).
5. Y satisfies PSA 8 also when either (a) Y0 is a linear subspace with Y0 \YC D f0g,
or (b) for some 2 „, the function vL is increasing on Y0 (i.e., vL .y0 / < vL .y00 /
whenever y0 < y00 , for y0 and y00 in Y0 ).
Proof Part 1 is obvious: PSAs 1 and 2 hold (i.e., Y is a convex cone) because kL
and vL are sublinear and Y0 is a convex cone. PSA 4 holds because kL and vL are
nondecreasing on Y0 (and because projY .Y/ D Y0 ). PSA 5 (with K D Rˆ , V D R„ )
holds because kL and vL are nonnegative on Y0 . Finally, PSAs 6 and 10 are verified at
kQ D kL .y/ and vQ D vL .y/.
Also Part 2 is obvious: the l.s. continuity of kL and vL means that their sublevel sets
are closed. n o
For Part 3, note that vmin ILR .y/ D kL .y/ ; vL .y/ if y 2 Y0 (and if not, then
ILR .y/ D ;). By their norm-continuity, kL and vL are bounded on some ball centered
at the origin. It follows, by their p.l. homogeneity, that kL and vL are bounded on every
ball in Y—i.e., PSA 7 holds.
For Part 4, given a y 2 Y0 with kL .y/ k, take the point .y; vL .y//. It is itself
maximal in YSR .k/ if the set
n o
y0 2 Y0 W y0 y; kL y0 k; vL y0 D vL .y/ (7.3.1)
16
In other words, each kL or vL is a p.l.h. convex finite function.
17
When each kL nis weakly* l.s.c. and o Y0 is weakly* closed as in Part 2, this is equivalent to weak*
L
compactness of y 2 Y0 W k .y/ k .
18
This assumption holds vacuously when „ D ; (i.e., when there are no variable inputs, as with
the storage and the hydro techniques (5.1.3) and (5.1.4)).
150 7 Production Techniques with Conditionally Fixed Coefficients
has no element other than y. But even if it has, the method of Lemma 6.2.1 applies.
This is because the set (7.3.1), after embedding it in Y V by taking its Cartesian
product with fvL .y/g, is here the same as the set (6.2.2) with v D vL .y/, and it is
bounded (being contained in the set that is bounded by assumption)—so Part 1 of
Corollary 6.2.2 applies. So PSA 8 holds. To verify PSA 9, note that every bounded
B
K is bounded from above by some k (since K is finite-dimensional). For each
k 2 B, every point of vmax YSR .k/ has the form .y;n vL .y// for someoy 2 Y0 with
S
kL .y/ k. So k2B vmax YSR .k/ is bounded (since y 2 Y0 W kL .y/ k is bounded
by assumption, and since v, L being p.l.h. and norm-continuous, is bounded on every
bounded set).
In Part 5, the point .y; vL .y// itself is always maximal in YSR .k/: in Case (a),
no point of Y0 is greater than y; and in Case (b), if y0 is a point of Y0 greater than y,
then vL .y0 / is greater than vL .y/.
The foregoing analysis (Sects. 3.11, 6.1, 6.2 and 6.3, 6.4) is next specialized to the
SRP programme and its dual for a technique with conditionally fixed coefficients.
As in the preceding Sect. 7.3 (and in Sects. 6.2, 6.3 and 6.4), the output space Y is
the dual of a Banach lattice Y 0 (and Y is the norm-dual of Y, with Y 0 Y ).
Notation The optimal solution sets for programmes (7.1.11)–(7.1.13) and (7.1.14)–
(7.1.17) are denoted by YO . p; k; w/ and RO . p; k; w/, respectively. The corresponding
lowercase notation, yO or rO , is used only when the solution is known to be unique.
Proposition 7.4.1 (Hotelling’s Lemma, Solubility of SRP Programme with
c.f.c.) Assume that the production set Y is given by (7.1.1), where the input
requirement functions kL W Y ! R and vL W Y ! R are w .Y; Y 0 /-lower
semicontinuous (for each 2 ˆ and 2 „), and the output constraint cone
Y0 is w .Y; Y 0 /-closed (so Y is weakly* closed). Then:
1. For every p 2 Y , k 0 and w 0,19
YO . p; k; w/ D @p …SR . p; k; w/ . (7.4.1)
Also, (6.2.11) applies, i.e., …SR .; k; w/ can be restricted to Y 0 when calculating
@p …SR at a p 2 Y 0 .
19
Formally, (7.4.1) holds also when k 0: in this case, YO D Y D @p …SR (the programme (7.1.11)–
(7.1.13) is then infeasible, so every y is an improper solution, and …SR .; k; w/ D 1).
7.4 Existence of Optimal Operation and Plant Valuation and Their Equality to. . . 151
n o
2. If k 0 and y 2 Y0 W kL .y/ k , the feasible set of the operation pro-
gramme (7.1.13)–(7.1.11), is norm-bounded then …SR . p; k; w/ is finite, for every
p 2 Y and w 0. If additionally p 2 Y 0 then
YO . p; k; w/ ¤ ; (7.4.2)
20
When kL and vL are norm-continuous, the l.s. continuity of CSR (on Y K) can be deduced also by
using Lemma 7.3.1 to verify the assumptions
n of Lemma 6.2.5.
o
21
Being also weakly* closed, the set y 2 Y0 W kL .y/ k is actually weakly* compact by the
Banach-Alaoglu Theorem.
152 7 Production Techniques with Conditionally Fixed Coefficients
Proposition 7.4.2 (Dual Hotelling Lemma, Solubility of FIV Prog. with c.f.c.)
In addition to the assumptions of Proposition 7.4.1 on kL , vL and Y0 (viz., that each
kL and vL is weakly* l.s.c. and that
n Y0 is weakly* closed),
o assume that each vL is
L
norm-continuous, and that the set y 2 Y0 W k .y/ k is norm-bounded. Then:
RO . p; k; w/ D @O k …SR . p; k; w/ . (7.4.3)
RO . p; k; w/ ¤ ; (7.4.4)
for every p 2 Y and w 0 (this means that the fixed-input value minimization
programme (7.1.14)–(7.1.17) has a proper solution, since its value …SR . p; k; w/
is finite).
Proof The operation programme (7.1.11)–(7.1.13) is feasible, i.e., …SR . p; k; w/ >
1 for every k 2 Rˆ 0
C . If p 2 Y then …SR . p; ; w/ is u.s.c. by Lemma 6.2.3;
its assumptions are verified by applying Parts 2 and 4 of Lemma 7.3.1. If k
0 then …SR . p; ; w/ is continuous at k, by Part 1 of Lemma 6.4.1. In either
case, …SR . p; k; w/ D …SR . p; k; w/ by Lemma 6.1.1. So (7.4.3) follows from
Lemma 3.11.2 with Remark 3.11.8 (as in the Proof of Part 1 of Corollary 3.11.19).
This proves Part 1.
For Part 2, since k 0, (7.4.4) follows from (7.4.3) and Part 2 of Lemma 6.4.1.
This means that the valuation programme (7.1.14)–(7.1.17) has a proper solution,
provided that it is feasible, i.e., that …SR . p; k; w/ D …SR . p; k; w/ < C1—which
is the case here (see the Proof of Proposition 7.4.1).
The original description of the production set, Y, of a c.f.c. technique need not be in
terms of input requirement functions as in (7.1.1). Indeed, a sublinear requirement
function kL can arise from summarizing, in a single scalar constraint, a system of
linear inequality constraints (i.e., a multi-dimensional or infinite-dimensional linear
inequality constraint). For example, a capacity k may constrain the output rate to a
y .t/ k at any time t, and this can be summarized as k kL .y/ WD supt y .t/—as
in (5.1.5) for the case of thermal generation. Another example is the storage capacity
requirement kL St .y/ of (5.1.8), which is used in (5.1.10) to summarize the continuum
7.5 Linear Programming for Techniques with Conditionally Fixed Coefficients 153
22
Similarly, if a unit output requires a unit of a costlessly storable variable input, whose total
amount available, v , can be spread as an input flow vQ ./ over the period, then the output rate is
RT
constrained to a nonnegative y .t/ vQ .t/ for some vQ .t/ 0 with 0 vQ .t/ dt D v . This can be
RT
summarized in the single constraint v vL .y/ WD 0 y .t/ dt.
154 7 Production Techniques with Conditionally Fixed Coefficients
of the cone Y0ı . But even after the pruning, the LP (7.5.1)–(7.5.5) may be doubly
infinite: the number of its decision variables is finite if and only if the space Y is
finite-dimensional, and the number of constraints is finite if each @kL .0/ or @vL .0/
is a polytope and the cone Y0ı is finitely generated.
And the plant valuation programme is reformulated as the LP:
(when an input in fixed supply has alternative uses even in the short run, as with
inflowing water that can be either stored or used up immediately).1 As the key to
efficient operation and to transition from the short-run to the long-run solution, the
inputs’ imputed values (shadow prices) are fundamental to this approach. The short-
run programmes provide a workable method for calculating these values, and the
transition to long-run market equilibrium ensures that they truly reflect the inputs’
scarcity. This puts valuation on a sound basis—and this is the long and the short
of it.
1
Valuation matters also for other purposes, e.g., in setting compensation payments for land or
rivers.
Appendix A
Example of Duality Gap Between SRP and FIV
Programmes
Equality of the primal and dual optimal values is equivalent to the semicontinuity
of either value function w.r.t. its “own” parameters, i.e., to Type One semicontinuity
(upper or lower in the case of a maximum or a minimum value, respectively): see
Sect. 6.1. Therefore, any sufficient condition for continuity of the one value rules
out a duality gap and implies that the other value is semicontinuous. It implies also
that the other programme is soluble (Sect. 6.4). In this Appendix, “continuity” of a
value always means Type One continuity (unless it is qualified as Type Two).
Any result for the primal value can be transcribed to the dual value by swapping
the two programmes. The sufficient conditions for continuity to be considered below
are all put entirely in terms of the primal programme. Such a criterion can be classed
by the particular value whose continuity it guarantees: it is either a primal-value
or a dual-value continuity criterion. In other words, it gives, in terms of the one
programme (here, the primal), a condition that guarantees value continuity for either
the same or the other programme of the pair.
There is a salient criterion in either class. A criterion of primal -value continuity
(w.r.t. the primal parameters) is Slater’s Condition on the primal programme,
together with its generalized forms: see [44, (8.12) and Theorem 18 (a)]. A useful
criterion of dual -value continuity (w.r.t. the dual parameters) can be based on
compactness-and-continuity conditions on the primal constraints and optimand: see
[44, Theorem 18’ (e)].1 Its semicontinuity implication for the primal value (w.r.t.
the primal parameters) can be regarded as a version of a part of Berge’s Maximum
Theorem [6, VI.3: Theorem 2]; the basic semicontinuity result of [44, Example 4’
after (5.13)] is simply a special case of Berge’s. The above semicontinuity results
for profit and cost as functions of the fixed quantities (Lemmas 6.2.3–6.2.5) are of
the same kind, being also applications of Berge’s Theorem (after the Krein-Smulian
Theorem has been used to “localize” to bounded sets).
1
This criterion of dual-value continuity is mentioned above in the Proof of Theorem 5.3.1.
For the context of short-run profit or the long-run cost as the primal value,
Slater’s Condition is spelt out in Sect. 6.4. In the case of SRP maximization with
conditionally fixed coefficients—i.e., the case of maximizing the operating profit
of a rigid plant—Slater’s Condition boils down to strict positivity of the fixed-
input bundle k which represents the plant’s capacities; this guarantees continuity
of …SR . p; ; w/ on a neighbourhood of k (Part 2 of Proposition 7.4.2). The
alternative upper semicontinuity result for …SR . p; ; w/ on the fixed-input space
K (Lemma 6.2.3) requires that the given price system lie in the predual of the
commodity space, i.e., that p 2 Y 0 and w 2 V 0 (in addition to assumptions on
the production set that hold whenever Parts 1, 2 and 4 of Lemma 7.3.1 apply).
Either condition—positive capacities or a predual output price system—
precludes a duality gap between the programmes of profit-maximizing operation
and plant valuation (for a c.f.c. technique satisfying the relevant assumptions).
Between them, the two sufficient conditions cover a lot of ground: although the
alternation “p 2 Y 0 or k 0” is not actually necessary for …SR to equal …SR at
. p; k; w/, it comes close to being so with technologies such as pumped storage and
hydroelectric generation. In the case of pumped storage, if the reservoir capacity
kSt is zero and the price system p 2 L1 Œ0; T has a singular a.k.a. purely finitely
additive part pFA ¤ 0 (in addition to a density a.k.a. countably additive part
pCA 2 L1 Œ0; T), then the operating profit is obviously zero, but the unit value of
conversion capacity is positive. Spelt out next, this example shows also that failure
of Slater’s Condition can lead to nonexistence of an exact dual solution. A similar
example of a duality gap with the hydro technology is given in [23].2
Example A.1 (Duality Gap Between Operation and Valuation of Incomplete Plant)
With the pumped-storage technology (5.1.3), take an output price system p 2
L1 Œ0; T with pFA ¤ 0 and pCA 2 BV
L1 Œ0; T, i.e., with a nonzero singular
part and a density part of bounded variation. If additionally kCo > 0 but kSt D 0, i.e.,
the plant has a conversion capacity but no storage capacity, then the optimal output
(the primal solution) is yO D 0 (it is the only feasible output), and so the maximum
operating profit (the primal optimal value) is zero, i.e., …PS
SR . pI 0; kCo / D 0. But the
O
optimal stock price (the dual solution) is D pCA , and so the minimum capacity
3
2
In the case of a hydro plant without a reservoir, i.e., with kSt D 0, take a p 0 with pFA ¤ 0.
If also kTu > Sup .e/, then h p j ei < h pCA j ei C kTu kpFA k. Since the optimal output is obviously
equal to the inflow e, and since the optimal stock price equals pCA as in Example A.1, this inequality
means that the maximum revenue is strictly less than the minimum value of the inputs (turbine and
inflow).
3
For the dual in the form (5.2.27)–(5.2.28) but with p 2 L1 here.
Appendix A 159
Fig. A.1 The total capacity value (…SR ) and the operating profit (…SR ) of a pumped-storage plant
as functions of its storage capacity kSt (for a fixed conversion capacity kCo > 0 and a fixed TOU
price, p 2 L1 nL1 , of the storable good). When kSt > 0, Slater’s Condition is met and so … D …,
but a duality gap opens at kSt D 0, where … is right-continuous but … drops to zero (Example A.1)
The same duality gap is present if pCA 2 L1 nBV (and again kCo > 0 but kSt D 0).
Although the dual (stock-pricing) programme for has then no (exact) solution,
any sequence of ’s in BV that converges to pCA in the L1 -norm is a sequence of
approximate dual optima (and so the infimal capacity value is still kCo kpFA k1 > 0).
4
A finite concave function on a polyhedral set Z Rn is l.s.c. on Z (so if it is u.s.c. on Z then it is
continuous on Z): see [42, 10.2 and 20.5]. This applies to Z D RnC for every n (here, n D 1).
160 Appendix A
That is, CLR is L1 -l.s.c. (and a fortiori L1 -l.s.c.) in y 2 L1 . This can be seen
either by applying Lemma 6.2.4 or directly from the formulae, (5.1.8)–(5.1.9),
for the capacity requirement functions kL St and kL Co . The same l.s. continuity is a
property of CSR as a function of .y; k/, since it is simply the 0-1 indicator of
the closed set Y (there are no variable inputs with this technique, i.e., the SRC
programme is merely a check of capacity sufficiency). So neither permutation
of p, k, y and r can yield a cost-minimizing solution and its dual counterpart.
Indeed, both permutations manifestly fail: (i) the LRC programme’s solution has
kCo D 0, unlike the SRP data for this example; and (ii) every solution to the
OFIV programme (the dual of the SRC programme) has rCo D 0, unlike the
solution to the FIV programme (the dual of the SRP programme), which has
rCo D kpFA k1 > 0. (In detail, the SRP primal-dual solution pair—given a
nonconstant pCA 2 BV, pFA ¤ 0, kSt D 0 and kCo > 0—is: y D 0 together
with r D .rSt ; rCo / D VarC c . p CA / ; kp k
FA 1 0. But, given y D 0 and
r D .rSt ; rCo / 0, an LRC solution pair is: .kSt ; kCo / D .0; 0/ with any LRMC
as p, i.e., with any p 2 rSt @ kL St .0/ C rCo @ kL Co .0/ C const:; recall from (6.2.10)
that @ WD L1 \ @a . And finally, given y D 0, kSt D 0 and kCo > 0, an SRC
dual solution is: rCo D 0 with any rSt 0 and any p 2 rSt @ kL St .0/ C const:)
Appendix B
Convex Conjugacy and Subdifferential Calculus
0
.lsc C/ .y/ WD min C .y/ ; lim
0
inf C y
y !y
or globally by the formula epi lsc C WD cl epi C, where cl means the T -closure, and
A proper convex function is one that takes a finite value (somewhere) and does not
take the value 1 (anywhere). A convex function taking the value 1 is peculiar:
it may take finite values only on the algebraic boundary of its effective domain,1 and
it has no finite value at all if it is lower semicontinuous along each straight line: see,
e.g., [42, 7.2 and 7.2.1], [44, Theorem 4] or [48, 5.12 with Proof].
for p 2 P; it is l.s.c. and either proper convex or an infinite constant (C1 or 1).
Obviously
C# . p/ h p j yi C .y/ (B.2.2)
Furthermore, if C is convex, then C## D lsc C unless lsc C takes the value 1 (and
hence has no finite value).2 In the latter case, C## D 1 (everywhere on Y) and
lsc C D 1 on the convex set cl dom C, but lsc C D C1 on the complement set.
So if a convex C is l.s.c. at y then: (i) C## .y/ and C .y/ can differ only by being
oppositely infinite, and (ii) C## .y/ D C .y/ if and only if either C .y/ < C1 or
both C .y/ D C1 and lsc C > 1 everywhere on Y. Also, C## D C (everywhere
on Y) if and only if C is either l.s.c. proper convex or an infinite constant.3 Applied
1
In precise terms, if C .y/ D 1 for some y 2 Y, and C is convex, then C .y/ D 1 for every y
in the intrinsic core (a.k.a. the relative algebraic interior) of dom C.
2
When additionally Y is finite-dimensional (and C is convex), if lsc C takes the value 1, then so
does C itself. This follows from [42, 7.5]; it is stated in, e.g., [44, Example 1”].
3
In [42] and [44], C is called “closed” when C D C## , and cl C serves as an alternative notation
for C## . This is abandoned in [45], and rightly so: cl C can be misinterpreted as lsc C, especially
since others—e.g., [37]—do use cl C instead of lsc C (to have epi cl C WD cl epi C).
Appendix B 163
C### D C# (B.2.5)
#
(which can be seen also directly from (B.2.1) and (B.2.4): C## C# because
##
C## C, but also C# C# ).
For a bivariate function C, its partial second conjugate (i.e., its second conjugate
taken w.r.t. just one variable, y, with the other variable, k, kept fixed) lies always
between the original function (C) and its total second conjugate (i.e., the second
conjugate w.r.t. both variables). Formally, the partial first conjugate w.r.t., say, the
first variable of a bivariate function C on Y K (where K is another vector space) is
defined as
C .y C y/ C .y/ C h p j yi (B.3.1)
for every y 2 Y. The set of all subgradients (at y) is the subdifferential a.k.a.
subgradient set , @C .y/. In other words,
See, e.g., [5, V.3.3 (b)] or [18, Exercise 2.10 (g)].5 A fortiori , TSLC makes Y a
barrelled space (since this requires (B.3.4) only for closed convex sets).
Directly from the subgradient inequality (B.3.1), if C0 and C00 are functions with
values in R [ fC1g, i.e., not taking the value 1, then
@ C0 C C00 .y/ @C0 .y/ C @C00 .y/ . (B.3.5)
4
In other words, the ˛ in (B.2.9) is allowed to vary with k in any way (subject to the stated
inequality), whereas the corresponding term in (B.2.10) is hr j ki C ˇ, which varies linearly with k.
5
For this reason, TSLC is also known as the convex-core topology . In [5] it is called “topologie
naturelle”.
Appendix B 165
Equality holds for proper convex functions under a continuity assumption: if, in
addition to C0 and C00 being convex with values in R [ fC1g, there exists a point of
Y at which both C0 and C00 are finite and at least one (C0 or C00 ) is continuous, then
@ C0 C C00 .y/ D @C0 .y/ C @C00 .y/ (B.3.6)
for every y 2 Y. See, e.g., [44, Theorem 20 (i) under (a)] or [48, 5.38 (b)]. Applied
to the case of 0-1 indicator functions of convex subsets of Y, (B.3.6) gives the
outward normal cone to the intersection of sets Z 0 and Z 00 as the sum of their normal
cones, i.e.,
N y j Z 0 \ Z 00 WD @ı y j Z 0 \ Z 00 D @ı y j Z 0 C @ı y j Z 00 (B.3.7)
DW N y j Z 0 C N y j Z 00
for every y 2 Y if Z 0 \ int Z 00 ¤ ;. This is stated in, e.g., [32, 4.3: Proposition 1].
Also directly from (B.3.1), for every ˛ > 0,
and this holds also for ˛ D 0 if (and only if) @C .y/ ¤ ;, i.e., if C is subdifferentiable
at y.
For C to be subdifferentiable at y, it is necessary that C be l.s.c. at y and actually
that C## .y/ D C .y/; in this case @C## .y/ D @C .y/. In other words,
p 2 @C .y/ , p 2 @C## .y/ and C## .y/ D C .y/ (B.3.9)
6
Since C# . p/ D C### . p/ h p j yi C## .y/ h p j yi C .y/ always, the first and the last terms
are equal if and only if both inequalities hold actually as equalities—which proves (B.3.9).
166 Appendix B
; ¤ @a C .y/ P (B.3.10)
Any continuous function is bounded from above (by a finite number) on a neigh-
bourhood of any point where its value is either finite or 1. With convex functions,
this obvious necessary condition is also sufficient for continuity. In precise terms, if
CW Y ! R [ f˙1g is convex then the following conditions are equivalent to one
another:
1. C is continuous on int dom C, which is nonempty.
2. C is continuous at some y 2 Y with C .y/ < C1.
3. There exists an open set N Y and a % 2 R such that C .y/ % for every y 2 N
(or, equivalently, the epigraph of C has a nonempty interior in Y R).
See, e.g., [18, 14A], [32, 3.2: Theorem 1], [44, Theorem 8] or [48, 5.20]. In
particular, this shows that, for a convex function, continuity is a property that
“propagates” from any single point to the whole interior of the effective domain
(Condition 2 ) Condition 1, as is stated also in Part 1 of Theorem B.3.1).
The sufficiency of local boundedness for continuity can also be combined with a
Baire category argument to deduce continuity from mere lower semicontinuity, for
a convex function on a Banach space or, more generally, on a barrelled space. The
result comes in two variants (which are very similar but not identical): see, e.g., [44,
Corollary 8B] and [18, p. 84 and Exercise 3.50]. In the particular case of TSLC (which
makes Y barrelled), l.s. continuity need not be assumed (because (B.3.4) holds for
all convex sets, whether closed or not). That is, every convex CW Y ! R [ f˙1g is
TSLC -continuous on cor dom C D int dom C. A finite convex function CW Y ! R is
therefore TSLC -continuous (everywhere) on Y: see, e.g., [5, V.3.3 (d)].
Appendix B 167
Those of the preceding concepts and results that depend on the function’s convexity
can be reoriented to concavity—which, again, will be assumed only when it is
required. In particular, when …W K ! R [ f˙1g is an extended-real function
on a real vector space K that is paired with another space R, the concave conjugate
of … is the concave function on R defined as
0
.usc …/ .k/ WD max … .k/ ; lim sup … k
k0 !k
is the hypograph of …. If … is concave then usc … depends on the dual space R but
not on the choice of a consistent topology on K. Also, with the effective domain of
a concave … defined as the convex set
if … is concave then usc … .k/ and …## .k/ differ in only one case: if k … cl dOom…
and usc … .k00 / D C1 for some k00 , then usc … .k/ D 1 but …## D C1 (on K).
168 Appendix B
So if … is u.s.c. proper concave (i.e., takes a finite value and does not take the value
C1), then …## D … (everywhere). Hence
…### D …# . (B.5.3)
… .k C k/ … .k/ C hr j ki (B.5.4)
O .k/,
for every k 2 K. The set of all supergradients (at k) is the superdifferential , @…
i.e.,
O .k/ , k maximizes … hr j i
r 2 @… (B.5.5)
, …# .r/ D hr j ki … .k/ . (B.5.6)
Also,
O .k/ , r 2 @…
r 2 @… O ## .k/ and …## .k/ D … .k/ . (B.5.7)
7
This is stated in, e.g., [4, 4.4.4], [42, 23.5 (a) and (a*)] and [44, Corollary 12A].
Appendix B 169
For a concave function … (on a space K paired with R), this becomes:
Proof This follows from the Fenchel-Young Inequality (B.2.2) and from the case
of equality therein as a characterization of the subdifferential: apply (B.3.2)–(B.3.3)
twice, to C and to C# (instead of C), to see that the conditions p 2 @C .y/ and
y 2 @C# . p/ are equivalent when C## .y/ D C .y/. It remains to show that this
equality holds when p 2 @C .y/. And this is because, by (B.2.2) and by (B.3.2)–
(B.3.3) applied to C# , C## .y/ h p j yi C# . p/ D C .y/ C## .y/ by (B.2.4).
The Inversion Rule and the First-Order Condition (B.3.2) are next combined in
a derivative property of conjugate functions. In convex programming, this yields
the derivative property of the optimal value (in the same way as is shown here, in
Sect. 3.11, for the cases of profit and cost programmes and their duals).
Corollary B.6.2 (Derivative Property of the Conjugate) Assume that CW Y !
R [ f˙1g is convex (and the space Y is paired with P). Then, for every y 2 Y and
p 2 P,
y maximizes h p j i C , y 2 @C# . p/ and C## .y/ D C .y/ . (B.6.3)
ı # . p j Z/ D sup h p j i (B.6.5)
Z
@ı # . p j Z/ D argmax h p j i (B.6.6)
Z
8
This is stated in, e.g., [4, 4.4.5], [42, 23.5 (b) and (a*)] and [44, Corollary 12B]. It holds formally
also when C is the constant 1 (but not when C is C1 because argmax .1/ WD ; by
convention, whereas @ .1/ . p/ WD Y).
170 Appendix B
if Z is nonempty, convex and closed. This is stated in, e.g., [42, 23.5.3] and [44,
p. 36, lines 1–7]. Similarly, the inf-support function of a set Z R is the concave
conjugate of ı . j Z/, i.e.,
9
This argument assumes that the maximand f is nowhere C1 and that the minimand g is nowhere
1. These sensible conditions are met when the perturbed primal constrained maximand, F, is a
u.s.c. proper concave function on a space X A paired with B Y (where A and B are the spaces
of primal and dual perturbations). This is because it follows that: (i) f .x/ D F .x; 0/ < C1 for
every x, and (ii) the perturbed dual constrained maximand, G .b; y/ WD F# .b; y/, is l.s.c. proper
convex, and so g .y/ WD G .0; y/ > 1 for every y. See, e.g., [44, (4.17)] for the equality of G
and F# (up to the signs).
Appendix B 171
See, e.g., [32, 4.1: Proposition 1], [42, 13.2.1] or [48, 6.22]. By (B.6.6), it follows
that, for every l.s.c. sublinear C,
( )
@C .y/ WD p 2 @C .0/ W h p j yi D sup h j yi D C .y/ (B.6.11)
@C.0/
which is stated in, e.g., [32, 4.2.1: Example 3], [42, 23.5.3] and [44, p. 36, lines
1–7]. This is a variant of Euler’s Theorem on homogeneous functions.
on P R. In other words, if
10
For example, it yields the extension (3.9.1) of the Wong-Viner Theorem.
172 Appendix B
D sup .… . p; k/ hr j ki/
k
as required.
Comment (“Staged” Conjugacy and Alternative Proof of the Inequality
Between Partial and Total Second Conjugates) The second conjugate, too, can
be obtained in stages:
C## D C#1 #1 #2 #2 .
That is, the total second conjugate of C is equal to the partial second conjugate,
w.r.t. one variable, of the partial second conjugate of C w.r.t. the other variable.
This gives another proof of the first inequality in (B.2.8): C## D C#1 #1 #2 #2 C#1 #1
(by (B.2.4) applied to the function C#1 #1 .y; / on K, instead of C). Or, in terms
of the partial second concave conjugate of … WD C#1 w.r.t. the second variable,
#1
C## D C#1 #2 #2 C#1 #1 (because …#2 #2 …).
h p j yi C .y; k/ … . p; k/ (B.7.2)
hr j ki C … . p; k/ C# . p; r/ (B.7.3)
and
… . p; k/ D h p j yi C .y; k/ < C1
… . p; k/ D C# . p; r/ C hr j ki > 1
11
From … D C#1 it follows that for every k 2 K either (i) … .; k/ D 1 (everywhere on P), or
(ii) … .; k/ does not take the value 1 (anywhere on P). The latter is the case for some k (since
C .; k/ ¤ C1 for some k by properness of C). So … . p; / ¤ 1 for every p 2 P.
12
That … . p; k/ > 1 can also be deduced from r 2 @O k … . p; k/, since … . p; / ¤ 1.
174 Appendix B
Corollary B.7.3 (Partial Inversion Rule) Under the assumptions of Lemma B.7.2,
the following conditions are equivalent to each other13 :
1. . p; r/ 2 @C .y; k/.
2. y 2 @p … . p; k/ and r 2 @O k … . p; k/, and C .; k/ is finite and lower semicontinu-
ous at y.
Also, either condition implies that both C .y; k/ and … . p; k/ are finite.
Proof By Lemma B.7.2, if . p; r/ 2 @C .y; k/ then, in addition to r 2 @O k … . p; k/
and C .y; k/ < C1, one has p 2 @y C .y; k/. By the Inversion Rule (B.6.1)
and (B.2.4), this implies that y 2 @p … . p; k/ and that C .; k/ is l.s.c. at y. So
Condition 1 implies Condition 2.
For the converse, since C .y; k/ < C1 and C .; k/ is l.s.c. at y, one has C .y; k/ D
C#1 #1 .y; k/. So if y 2 @p … . p; k/ then p 2 @y C .y; k/ by the Inversion Rule (B.6.1).
And if additionally r 2 @O k … . p; k/, then . p; r/ 2 @C .y; k/ by Lemma B.7.2.
Comments (on the PIR and the SSL)
• Finiteness of C .y; k/ can be dropped from Condition 2 (and the proof of its
equivalence to Condition 1 simplifies) if either (i) C .; k/ is assumed to be l.s.c.
on the whole space Y (and not just at the particular point y), or (ii) Y is finite-
dimensional. This is because, in either case, the assumption (of Lemmas B.7.2
and B.7.3) that C .; k/ > 1 on Y implies that lsc .C .; k// > 1 on Y (when
Y is finite-dimensional, this follows from [42, 7.5]). Therefore lsc .C .; k// D
C#1 #1 .; k/ on Y, and so the Inversion Rule (B.6.1) shows that p 2 @y C .y; k/ if
and only if both y 2 @p … . p; k/ and C .; k/ is l.s.c. at y. Thus Corollary B.7.3
reduces immediately to Lemma B.7.2.
• There is a structural difference between the Subdifferential Sections Lemma and
the Partial Inversion Rule. The SSL turns the condition . p; r/ 2 @y;k C into the
pair of conditions p 2 @y C and r 2 @O k …—which involve two functions but use
partial subdifferentials w.r.t. the same variables as in the joint subdifferential (of
C). The PIR turns the condition . p; r/ 2 @y;k C into the pair of conditions y 2
@p … and r 2 @O k …. These use a single function …, but only one of its arguments
(k) is the same as in the parent function C: the other argument (y) is replaced by
its dual (p) in inverting @y C into @p …. This step requires the semicontinuity of C
w.r.t. y—and this is why the PIR (its Condition 2) is not purely algebraic like the
SSL.
An obvious but remarkable implication of the SSL is that the saddle “child”
function (…) is “smoother” than its bivariate convex parent (C) by having a smaller
subdifferential w.r.t. their common, non-conjugated variable (k). This means that
the child function “has a good chance” of being differentiable even when its convex
parent is not. And, indeed, this heuristic is borne out by the study of peak-load
13
This is in, e.g., [4, 4.4.14], [41, Lemma 4], [42, 37.5] and [45, 11.48].
Appendix B 175
pricing with storage summarized in Sect. 5.2, which shows that the operating profit
of each type of plant is a differentiable function of the capital inputs. For details, see
[21, 23, 27] and [30].
Remark B.7.4 (Partial Subdifferentials of Saddle Function and Its Convex Parent)
Under the assumptions of Lemma B.7.2,
@k C .y; k/ contains the section of @C .y; k/ through any p 2 @y C .y; k/—and this
section equals @O k … . p; k/ by Lemma B.7.2.
Comments
• A simpler proof of (B.7.5) comes straight from the definition (B.7.1):
For a saddle function S with a (bivariate) convex parent I, the following useful
variant of Corollary B.7.3 transposes the saddle differential correspondence @S into
@I # instead of @I (i.e., into the subdifferential correspondence of I’s total conjugate
instead of I itself).
Corollary B.7.5 (Dual Partial Inversion Rule) Assume that IW YV ! R[fC1g
is proper convex and (jointly) lower semicontinuous for the pairing of the space V
with W (and of Y with P), and that SW Y W ! R [ f˙1g is its partial convex
conjugate I #2 , i.e., that
for every y 2 Y and w 2 W. Then the following conditions are equivalent to each
other:
1. .y; u/ 2 @I # . p; w/.
2. p 2 @y S .y; w/ and u 2 @O w S .y; w/.
Also, either condition implies that both I . p; w/ and S .y; w/ are finite.
Proof Since I ## D I by the assumption that I is l.s.c., the Inversion Rule (B.6.1)
shows that Condition 1 is equivalent to: . p; w/ 2 @I .y; u/. And this is equivalent to
Condition 2 by the Partial Inversion Rule (Corollary B.7.3) and the first Comment
thereafter.
# #1
Comment (on Another Derivation of DPIR) By Lemma B.7.1, I D I 2 #
D
S#1 , i.e., the convex function I # is a partial conjugate of the saddle function S.
When this relationship can be inverted to represent S as the partial conjugate I ##1 ,
the required invertibility of @I # into @S will follow from the PIR alone—without
recourse to the Inversion Rule for the joint subdifferentials (of I and I ## ). But this
argument requires S .; w/ to be l.s.c. on Y, and this is a condition that S can fail
at some points (even when I is l.s.c.). The actual Proof of Corollary B.7.5 obviates
the need to ensure that S .y; w/ is l.s.c. in y (by applying the PIR directly to the
assumed partial conjugacy S D I #2 , rather than intending to derive from it the
partial conjugacy S D I ##1 in order to apply the PIR to that).
adjoining just any partial subgradient w.r.t. the other variable. But, as is shown
below, it usually can be so complemented by a suitable choice of the other partial
subgradient (Theorem B.8.3 at interior points, and Corollary B.8.4 at boundary
points). Not only does such a complementary choice exist, but in principle it can also
be calculated—by superdifferentiating the relevant partial conjugate (Lemma B.7.2
on subdifferential sections).
Proposition B.8.1 (Nonfactorable Subdifferentials of Sublinear Function)
Assume that CW Y K ! R [ fC1g is (jointly) positively linearly homogeneous,
convex and lower semicontinuous for the pairing of Y K with another product,
P R, of real vector spaces, and that . p0 ; r0 / and . p00 ; r00 / are two joint
subgradients of C at .y; k/, i.e., . p0 ; r0 / and . p00 ; r00 / are in @y;k C .y; k/ PR.14
If additionally
˝ ˛ ˝ ˛
p0 j y ¤ p00 j y (B.8.1)
then r0 ¤ r00 (and so the partial subdifferential @k C .y; k/ is not a singleton, i.e.,
C .y; / is not Gâteaux differentiable at k). What is more, neither . p0 ; r00 / nor
. p00 ; r0 / is in @y;k C .y; k/, and so
C .y; k/ D h p j yi hr j ki (B.8.3)
for every . p; r/ 2 @y;k C .y; k/. So (B.8.3) holds for both . p0 ; r0 / and . p00 ; r00 /,
but it therefore fails for both . p0 ; r00 / and . p00 ; r0 / because of (B.8.1). So neither
. p0 ; r00 / nor . p00 ; r0 / is in @y;k C .y; k/, which shows that this set is not a Cartesian
product.
Example B.8.2 (Nonfactorable Subdifferential) Take the function cW R2C ! R
defined as in (2.2.3), i.e., c .y; k/ D wy if 0 y k and C1 otherwise (given
a number w 0). With the scalar product hp; r j y; ki WD py rk=T where T > 0
is a given number, the joint subdifferential of c at a point with y D k > 0 is
n r o
@y;k c .y; k/ D .p; r/ 2 RC R W p D w C ; r 0
T
(which, being a half-line not parallel to either coordinate axis of the plane R2 , is not
a Cartesian product).
14
The minus sign in . p; r/ is there to make r nonnegative when C .y; / is nonincreasing on K.
178 Appendix B
15
Then h p0 j yi < h p00 j yi, since p0 < p00 and y 0.
Appendix B 179
mere existence of such a complementary r, but also gives a method of calculating it.
This alternative proof is detailed next.
Alternative Proof of Theorem B.8.3 Take any y, any k at which C .y; / is finite and
continuous, and any p 2 @y C .y; k/. Introduce … WD C#1 , the convex-concave
function on P K defined by (B.7.1). By the SSL (Lemma B.7.2), it suffices to
show that @O k … . p; k/ ¤ ;. Since … . p; / h p j yiC .y; /, which is bounded from
below on a neighbourhood of k, the concave function … . p; / is continuous on all of
intK dOom… . p; /, which contains k (Sect. B.4).16 Also, since p 2 @y C .y; k/, one has
… . p; k/ D h p j yi C .y; k/, which is finite. It follows that … . p; / nowhere takes
the value C1.17 Therefore, … . p; / is superdifferentiable at k, i.e., @O k … . p; k/ ¤ ;
by Theorem B.3.1.
A basic shortcoming of Theorem B.8.3 is its failure to apply at the boundary
points of the function’s effective domain, dom C. And indeed, at a boundary point,
a partial subgradient may have no complement to a joint one—but it is useful
to identify those cases in which such complements do exist. This is because the
boundary points of dom C can be the points of greatest interest: for example, when
C is the short-run cost as a function of the output bundle y and the fixed-input bundle
k, all the efficient combinations of y and k lie on the boundary of dom C. If, however,
C has a finite convex extension CEx , defined on the whole space (or at least on
a neighbourhood of dom C), and dom C can be represented as the sublevel set of
another finite convex function CDm , then Theorem B.8.3 can be applied to both of
these functions (CEx and CDm ). For the original function C, this yields a result that
applies also to the boundary points of dom C.
Corollary B.8.4 (Complementary Subgradient at Boundary Point) Let CW Y
K ! R [ fC1g be a (jointly) convex function, where Y and K are topological
vector spaces paired with their continuous duals (P and R). Assume that:
1. The effective domain of C has the form
˚
dom C D .y; k/ W CDm .y; k/ 0 and k 2 K0 (B.8.4)
16
Continuity of … . p; / can also be proved more succinctly: the maximization programme that
defines … by (B.7.1) satisfies, at k, Slater’s Condition for generalized perturbed CPs as formulated
in [44, Theorem 18 (a)]. So … . p; / is continuous at k by [44, Theorem 18 (a)]; to apply this
formally, the programme must of course be reoriented to minimization.
17
If … . p; / took the value C1 anywhere, then it would have to equal C1 everywhere on
intK dOom… . p; /—see the end of Sect. B.1—but it is finite at k.
180 Appendix B
4. C (or, more precisely, its restriction to dom C) has a finite, continuous and convex
extension CEx W Y K ! R.
Then for every p 2 @y C .y; k/ there exists an r such that . p; r/ 2 @y;k C .y; k/.
Proof Every p 2 @y C .y; k/ has the form p D p0 C ˛p00 for some p0 2 @y CEx .y; k/,
p00 2 @y CDm .y; k/ and a scalar ˛ 0, with ˛ D 0 if CDm .y; k/ < 0. This is because,
since C D CEx C ı . j dom C/,
when CDm .y; k/ D 0. When CDm .y; k/ < 0, the term @y ı—which is the outward
normal cone to the sublevel set of CDm .; k/—equals f0g, in which case the term
denoting the cone generated by @y CDm must be deleted from (B.8.5). For additivity
of @ (with a similar application to a sum of the form C C ı), see, e.g., [42, 23.8
and proof of 28.3.1], [44, Theorem 20] or [48, 5.38 and 7.2]. The relevant formula
for the normal cone to a sublevel set is given in, e.g., [32, 4.3: Proposition 2], [42,
23.7.1] or [48, 7.8].
Since CEx and CDm are finite and continuous (everywhere on Y K), Theo-
rem B.8.3 applies to both; so there exist r0 and r00 with
p0 ; r0 2 @y;k CEx .y; k/ and p00 ; r00 2 @y;k CDm .y; k/ . (B.8.6)
It now suffices to set r WD r0 C ˛r00 . To see this, use again the formula for the normal
cone and additivity of @ (this time for joint subdifferentials) to obtain from (B.8.6)
that
. p; r/ D p0 C ˛p00 ; r0 ˛r00
˚
2 @y;k CEx .y; k/ C @y;k ı y; k j y0 ; k0 W CDm y0 ; k0 0
˚
@y;k CEx .y; k/ C @y;k ı y; k j CDm 0 C @y;k ı .y; k j Y K0 /
D @y;k CEx .y; k/ C @y;k ı .y; k j dom C/ D @y;k C .y; k/ .
The penultimate equality follows from (B.8.4); by the way, on its l.h.s.
@y;k ı .y; k j Y K0 / D f0g @k ı .k j K0 /.
The Alternative Proof of Theorem B.8.3—the proof based on the SSL—has a
counterpart that gives another proof of Corollary B.8.4. This is detailed next.
Alternative Proof of Corollary B.8.4 Let … WD C#1 (a saddle function on P K).
By the SSL (Lemma B.7.2), it suffices to show that @O k … . p; k/ ¤ ; for the given
k and any p 2 @y C .y; k/. For this, note first that … D …0 ı . j P K0 /, where
Appendix B 181
…0 p; kQ denotes, for each kQ 2 K, the supremum of h p j yiCEx y; kQ over y subject
to CDm y; kQ 0.
Next, sinceCDm yS ; k < 0 and CDm˝ is continuous,
˛ on a neighbourhood of k one
has CDm yS ; < 0 and so …0 . p; / p j yS CEx yS ; , which is bounded from
below on a neighbourhood of k. So the concave function …0 . p; / is continuous on
all of intK dOom…0 . p; /, which contains k (Sect. B.4).18 Also, since k 2 K0 and
p 2 @y C .y; k/, one has …0 . p; k/ D … . p; k/ D h p j yi C .y; k/, which is finite. It
follows that … . p; / nowhere takes the value C1.19
Therefore, …0 . p; / is superdifferentiable at k, i.e., @O k …0 . p; k/ ¤ ; by Theo-
rem B.3.1. Finally, since …0 . p; / is continuous,
@O k … . p; k/ D @O k …0 . p; k/ @ı .k j K0 / ¤ ;
since both terms are nonempty sets (@ı is a cone); for additivity of @, see, e.g., [44,
Theorem 20 (i) under (a)] or [48, 5.38 (b)].
18
Continuity of …0 . p; / can also be proved more succinctly: by the third assumption of
Corollary B.8.4, the maximization programme that defines …0 satisfies, at kQ D k, Slater’s
Condition as generalized to infinite-dimensional inequality constraints in [44, (8.12)]. So …0 . p; /
is continuous at k by [44, Theorem 18 (a)].
19
If …0 . p; / took the value C1 anywhere, then it would have to equal C1 everywhere on
intK dOom…0 . p; /—see the end of Sect. B.1—but it is finite at k.
Appendix C
Notation List
Notation is grouped below in several categories. See also Table 5.1 for correspon-
dence between the general duality scheme (Sects. 3.3 and 3.12) and its application
to electricity supply (Sects. 5.2 and 5.3).1
Profit and Cost Optimization and Shadow-Pricing Programmes: Parameters
and Decision Variables, Solutions, Optimal Values and Marginal Values
1
Note two unrelated meanings of the symbols s and : in the general duality scheme of Sects. 3.3
and 3.12, these mean the standard parameters (s) paired with the standard dual variables ( ), but
in the description of energy storage techniques in Sect. 5.1 they mean the energy stock (s) and
spillage ( ). Also, the n, nSt and nTu of Sect. 5.2 mean lower constraint parameters (whose original,
unperturbed values are zeros), whereas in the short-run approach to equilibrium and its application
to electricity pricing, in Sects. 4.2 and 5.3, n means an input of the numeraire.
'O h . p; %I M/ consumer h’s demand for the Industrial User’s product when its
price is %, the differentiated good’s (e.g., electricity) price system is p, and the
consumer’s income is M
zO . p; %/ the Industrial User’s factor demand for the differentiated good (e.g.,
electricity) when its price system is p and the User’s product price is %
nO . p; %/ the Industrial User’s factor demand for the numeraire when the User’s
product price is % and the differentiated good’s (e.g., electricity) price system is p
Short-Run General-Equilibrium Prices and Quantities
p?SR , %?SR prices for the differentiated good (electricity) and for the IU’s product
y?SR output of the differentiated good (electricity) by production technique
?
vSR variable input into production technique
xSR h , z?SR consumer demand and factor demand for the differentiated good
?
(electricity)
m?SR h , n?SR consumer demand and factor demand for the numeraire
' ?SR the Industrial User’s output
Long-Run General-Equilibrium Prices and Quantities
w the given prices of the Supply Industry’s variable inputs
rF the given rental prices of the Supply Industry’s fixed-priced capital inputs
rE rental prices of the Supply Industry’s equilibrium-priced capital inputs—to
be determined in long-run equilibrium
r? the equilibrium prices of the equilibrium-priced inputs (i.e., the equilibrium
value of rE )
k? equilibrium capacities of producer in the Supply Industry
p?LR , y?LR , etc. equilibrium prices and quantities—as above, but for the long-
run equilibrium
Electricity Generation (All Techniques)
p .t/ electricity price at time t (in $/kWh), i.e., p is a time-of-use (TOU) tariff
y .t/ rate of electricity output from a plant, at time t (in kW)
Dt .p/ cross-price independent demand for electricity (in kW) at time t, if the
current price is p (in $/kWh)
Thermal Generation
S .p/ in the short run, the cross-price independent rate of supply (in kW) of
thermally generated electricity, if the current price is p (in $/kWh)
cSR .y/ the instantaneous short-run thermal cost per unit time (in $/kWh), if the
current output rate is y (in kW); the common graph of the correspondences S and
@cSR is the thermal SRMC curve
a type of thermal plant
fuel type used by plant type
v fuel input of a thermal plant (in kWh of heat)
technical efficiency of a thermal plant, i.e., 1= is the heat rate
Appendix C 187
w unit running cost of a thermal plant (in $ per kWh of electricity output), equal
to the price of fuel (in $ per kWh of heat) times the heat rate
k a thermal generating capacity (in kW)
unit value of a thermal generating capacity at time t, per unit time (in $/kWh)
RT
r D 0 .t/ dt unit value of a thermal generating capacity in total for the cycle
(in $/kW) RT
.t/ D .t/ = 0 .t/ dt density, at time t, of the distribution of capacity
charges over the cycle, i.e., a function representing a subgradient of the convex
functional EssSup on L1 Œ0; T (more generally, a subgradient of any capacity
requirement function)
rF the given rental price of a thermal generating capacity (in $/kW)
.t/ unit value of nonnegativity constraint on the output of a thermal plant at
time t, per unit time (in $/kWh)
YO Th . p; k; w/ the set of all the electricity output bundles that maximize the
operating profit of a thermal plant of capacity k with a unit running cost w when
p is the TOU electricity tariff
yO Th . p; k; w/ the electricity output bundle—if unique—that maximizes the
operating profit of a thermal plant of capacity k with a unit running cost w when
p is the TOU electricity tariff
y? .t/ the general-equilibrium rate of electricity output from the thermal plant
of type at time t (in kW)
Pumped Storage
kSt the plant’s storage a.k.a. reservoir capacity (in kWh)
St .dt/ unit value of storage capacity on a time interval of length dt (in $/kWh)
RT
rSt D 0 St .dt/ unit value of storage capacity in total for the cycle (in $/kWh)
?
rSt the (long-run) equilibrium rental price of storage capacity (in $/kWh)
G .kSt / the supply cost of kSt of storage capacity
St .dt/ unit value of nonnegativity constraint on energy stock on an interval of
length dt (in $/kWh)
kCo the plant’s conversion capacity (in kW)
Pu .t/ unit value of converter’s pump capacity at time t, per unit time (in
$/kWh)
Tu .t/ unit value of converter’s turbine capacity at time t, per unit time (in
$/kWh)
Co .t/ D Pu .t/ C Tu .t/ unit value of converter’s capacity at time t, per unit
time (inR $/kWh)
T
rCo D 0 Co .t/ dt unit value of conversion capacity in total for the cycle (in
$/kW)
F
rCo the given rental price of conversion capacity (in $/kW)
O
Y PS . pI kSt ; kCo / the set of all the electricity output bundles that maximize the
operating profit of a pumped-storage plant with capacities kSt and kCo when p is
the TOU electricity tariff
188 Appendix C
m .Y; P/ the Mackey topology on Y for its pairing with P (e.g., with P D Y or
with P D Y 0 when Y is a dual Banach space)
w and m abbreviations for w .P ; P/ and m .P ; P/, the weak* and the
Mackey topologies on the norm-dual of a Banach space P
bw the bounded weak* topology (on a dual Banach space)
clY;T Z the closure of a set Z relative to a (larger) set Y with a topology T
intY;T Z the interior of a set Z relative to a (larger) set Y with a topology T
Y a the algebraic dual of a vector space Y
TSLC D m .Y; Y a / the strongest locally convex topology on a vector space Y
Sets Derived from a Set in a Vector Space
cone Z the cone generated by a subset, Z, of a vector space (i.e., the smallest
cone containing Z)
conv Z the convex hull of a subset, Z, of a vector space (i.e., the smallest convex
set containing Z)
cor Z the core of a subset, Z, of a vector space
ext Z the set of all the extreme points of a subset, Z, of a vector space
span Z the linear span of a subset, Z, of a vector space
N .y j Z/ D @ı .y j Z/ the outward normal cone to a convex set Z at a point
y 2 Z (a cone in the dual space)
Na .y j Z/ D @a ı .y j Z/ the algebraic normal cone to Z at y (a cone in the
algebraic dual space); @a is the algebraic subdifferential
Sets and Functions Derived from Functions or Operations on a Vector Space
1. Afriat SN (1971) Theory of maxima and the method of Lagrange. SIAM J Appl Math 20:343–
357
2. Aliprantis C, Burkinshaw O (1985) Positive operators. Academic Press, New York-London
3. Anderson EJ, Nash P (1987) Linear programming in infinite-dimensional spaces. Wiley,
New York-Chichester-Brisbane-Toronto-Singapore
4. Aubin JP, Ekeland I (1984) Applied nonlinear analysis. Wiley, New York
5. Bair J, Fourneau R (1975) Etude géometrique des espaces vectoriels (Lecture notes in
mathematics, vol 489). Springer, Berlin-Heidelberg-New York
6. Berge C (1963) Topological spaces. Oliver and Boyd, Edinburgh
7. Bewley T (1972) Existence of equilibria in economies with infinitely many commodities. J
Econ Theory 4:514–540
8. Birkhoff G (1967) Lattice theory. American Mathematical Society, Providence
9. Boiteux M (1964) Peak-load pricing. In: Nelson JR (ed) Marginal cost pricing in practice
(Chapter 4). Prentice Hall, Engelwood Cliffs ((1949) A translation of “La tarification des
demandes en pointe: application de la théorie de la vente au coût marginal”, Revue Général
de l’Electricité 58:321–340)
10. Boutacoff D (1989) Emerging strategies for energy storage. Electric Power Res Instit J 14(5):
4–13
11. Chvatal V (1983) Linear programming. Freeman, New York
12. Craven BD (1978) Mathematical programming and control theory. Chapman and Hall, London
13. Diewert WE (1974) Applications of duality theory. In: Intriligator MD, Kendrick DA (eds)
Frontiers of quantitative economics, vol II (Chapter 3). North-Holland, Amsterdam
14. Diewert WE (1982) Duality approaches to microeconomic theory. In: Arrow KJ, Intriliga-
tor MD (eds) Handbook of mathematical economics, vol II (Chapter 12). North-Holland,
Amsterdam
15. Drèze JH (1964) Some postwar contributions of French economists to theory and public policy.
Am Econ Rev 54(supplement):1–64
16. Foran J (1991) Fundamentals of real analysis. Dekker, New York-Basel-Hong Kong
17. Henderson JM, Quandt RE (1971) Microeconomic theory. McGraw-Hill, New York-London
18. Holmes RB (1975) Geometric functional analysis and its applications. Springer, Berlin-
Heidelberg-New York
19. Horsley A, Wrobel AJ (1988) Subdifferentials of convex symmetric functions: An application
of the Inequality of Hardy, Littlewood and Pólya. J Math Anal Appl 135:462–475
20. Horsley A, Wrobel AJ (1996) Uninterruptible consumption, concentrated charges, and equilib-
rium in the commodity space of continuous functions. STICERD Discussion Paper TE/96/300,
LSE
21. Horsley A, Wrobel AJ (1996) Efficiency rents of storage plants in peak-load pricing, I: pumped
storage. STICERD Discussion Paper TE/96/301, LSE (This is a fuller version of Ref. [27])
22. Horsley A, Wrobel AJ (1996) Comparative statics for a partial equilibrium model of investment
with Wicksell-complementary capital inputs. STICERD Discussion Paper TE/96/302, LSE
23. Horsley A, Wrobel AJ (1999) Efficiency rents of storage plants in peak-load pricing, II:
hydroelectricity. STICERD Discussion Paper TE/99/372, LSE (This is a fuller version of
Refs. [24] and [30])
24. Horsley A, Wrobel AJ (1999) Efficiency rents of hydroelectric storage plants in continuous-
time peak-load pricing. In: Dahiya SB (ed) The current state of economic science, vol 1.
Spellbound Publications, Rohtak, pp 453–480
25. Horsley A, Wrobel AJ (2000) Localisation of continuity to bounded sets for nonmetrisable
vector topologies and its applications to economic equilibrium theory. Indag Math (New Series)
11:53–61
26. Horsley A, Wrobel AJ (2002) Boiteux’s solution to the shifting-peak problem
and the equilibrium price density in continuous time. Econ Theory 20:503–537.
DOI:10.1007/s001990100226
27. Horsley A, Wrobel AJ (2002) Efficiency rents of pumped-storage plants and their uses for
operation and investment decisions. J Econ Dyn Control 27:109–142. DOI: 10.1016/S0165-
1889(01)00030-6
28. Horsley A, Wrobel AJ (2005) Continuity of the equilibrium price density and its uses in peak-
load pricing. Econ Theory 26:839–866. DOI: 10.1007/s00199-004-0568-3
29. Horsley A, Wrobel AJ (2006) Demand continuity and equilibrium in Banach commodity
spaces. In: Wieczorek A, Malawski M, Wiszniewska-Matyszkiel A (eds) Game theory and
mathematical economics, vol 71. Banach Center Publications, pp 163–183. (Also avail-
able as CDAM Research Report LSE-CDAM-2005-01, http://www.cdam.lse.ac.uk/Reports/
reports2005.html)
30. Horsley A, Wrobel AJ (2007) Profit-maximizing operation and valuation of hydroelectric
plant: a new solution to the Koopmans problem. J Econ Dyn Control 31:938–970. DOI:
10.1016/j.jedc.2006.03.004
31. Horsley A, Wrobel AJ, Van Zandt T (1998) Berge’s Maximum Theorem with two topologies
on the action set. Econ Lett 61:285–291. DOI:10.1016/S0165-1765(98)00177-3
32. Ioffe AD, Tihomirov VM (1979) Theory of extremal problems. North-Holland, Amsterdam-
New York-Oxford
33. Kantorovich LW, Akilov PG (1982) Functional analysis. Oxford, Pergamon Press
34. Klein E, Thompson AC (1984) Theory of correspondences. Wiley, New York-Chichester-
Brisbane-Toronto
35. Koopmans TC (1957) Water storage policy in a simplified hydroelectric system. In: Proceed-
ings of the First international conference on operational research, pp 193–227. Operations
Research Society of America, Baltimore (Also in: Koopmans TC (1970) Scientific papers of
Tjalling C. Koopmans. Springer, Berlin-Heidelberg, pp 282–316)
36. Laurent P-J (1972) Approximation et optimisation. Hermann, Paris
37. Levin VL (1985) Convex analysis in spaces of measurable functions and its applications to
mathematics and economics (in Russian). Nauka, Moscow
38. Marsh WD (1980) Economics of electric utility power generation. Oxford University Press-
Clarendon Press, Oxford-New York
39. Meyer-Nieberg P (1991) Banach lattices. Springer, Berlin-Heidelberg-New York
40. Moore T (1994) Storing megawatthours with SMES. Electric Power Res Inst J 19(5):24–33
41. Rockafellar RT (1968) A general correspondence between dual minimax problems and convex
programs. Pac J Math 25:597–611
42. Rockafellar RT (1970) Convex analysis. Princeton University Press, Princeton
References 195
43. Rockafellar RT (1970) Conjugate convex functions in optimal control and the calculus of
variations. J Math Anal Appl 32:174–222
44. Rockafellar RT (1974) Conjugate duality and optimization. SIAM, Philadelphia
45. Rockafellar RT, Wets RJB (1997) Variational analysis. Springer, Berlin-Heidelberg-New York
46. Rudin W (1973) Functional analysis. McGraw-Hill, New York
47. Takayama A (1985) Mathematical economics. Cambridge University Press, Cambridge-
London-New York
48. Tiel J van (1984) Convex analysis. Wiley, Chichester-New York-Brisbane
49. Wrobel AJ (2016) Efficiency rents of a hydroelectric storage plant with a variable head.
A manuscript in preparation