C H A P T E R 2 0

Additional Network
and LP Algorithms

Chapter Guide. This chapter includes 3 sections. The minimum cost capacitated flow
problem in Section 20.1 has been excised from Chapter 6. It deals with network flow in
which the capacity of the arcs may be limited and external input/output flow occurs at
different nodes. The objective is to determine the flow schedule that minimizes the as-
sociated cost while satisfying the capacity and external flow restrictions. The model can
be specialized to represent the shortest-route and the maximal-flow problems present-
ed in Chapter 6. The remaining 2 sections have been excised from Chapter 7. They deal
with the solution of large-scale LPs. The decomposition algorithm, a product of the
1960s technology, calls for breaking down LPs with special structures into smaller com-
putationally-manageable subproblems. This classic algorithm is interesting theoretically
but no longer viable computationally, because of the present tremendous power of the
computer. On the other hand, Karmarkar’s interior point algorithm, developed in the
1980s, is viable both theoretically and computationally. Unlike the simplex algorithm,
its iterations cut across the interior of the solution space. From the computational
standpoint, the result is a polynomial-time algorithm, as compared with the simplex
exponential-time algorithm. The new algorithm is designed for extremely large LPs.
This chapter includes 7 solved examples, 34 end of section problems, 2 AMPL
models, 1 Solver model, and 2 cases. The cases are in Appendix E on the CD. The
AMPL/Excel/Solver/TORA programs are in folder ch20Files.

20.1 MINIMUM-COST CAPACITATED FLOW PROBLEM

The minimum-cost capacitated flow problem is based on the following assumptions:

1. All arcs are directional (one way).

2. A (nonnegative) unit flow cost is associated with each arc.
3. Arcs may have positive lower capacity limits.
4. Any node in the network may act as a source or as a sink.

CD-1
CD-2 Chapter 20 Additional Network and LP Algorithms

[ fi] [ fj]
$ cij
i j
FIGURE 20.1 (lij, uij)
xij
Capacitated arc with external flow

The new model determines the flows that minimize the total cost while satisfying
the flow restrictions on the arcs and the supply and demand amounts at the nodes. We
first present the capacitated network flow model and its equivalent linear programming
formulation. The linear programming formulation is then used as the basis for the devel-
opment of a special capacitated simplex algorithm for solving the network flow model.

20.1.1 Network Representation

Consider a capacitated network G = 1N, A2, where N is the set of nodes and A is the
set of arcs and define
xij = amount of flow from node i to node j
uij1lij2 = upper 1lower2 capacity of arc1i, j2
cij = unit flow cost from node i to node j
fi = net flow at node i
Figure 20.1 depicts these definitions on arc (i, j). The label [fi] assumes a positive
(negative) value when a net supply (demand) is associated with node i.

Example 20.1-1
GrainCo supplies corn from three silos to three poultry farms. The supply amounts at the three
silos are 100, 200, and 50 thousand bushels and the demands at the three farms are 150, 80, and
120 thousand bushels. GrainCo mostly uses railroads to transport the corn to the farms, with the
exception of three routes where trucks are used.
Figure 20.2 shows the available route between the silos and the farms. The silos are repre-
sented by nodes 1, 2, and 3 whose supply amounts are [100], [200], and [50], respectively. The
farms are represented by nodes 4, 5, and 6 whose demand amounts are [-150], [-80], and
[-120], respectively. The routes allow transhipping between the silos. Arcs (1, 4), (3, 4), and (4, 6)
are truck routes. These routes have minimum and maximum capacities. For example, the capacity
of route (1, 4) is between 50 and 80 thousand bushels. All other routes use trainloads, whose

FIGURE 20.2 [⫺150]

Capacitated network for Example 20.1-1 $1
[100] 1 4
(50, 80) $2
$1
$4
$3 [50] (70, 120) 6 [⫺120]
$4
3 $2
$5 (100, 120)
$6
[200] 2 5
[⫺80]
 20.1 Minimum-Cost Capacitated Flow Problem CD-3

maximum capacity is practically unlimited. The transportation costs per bushel are indicated on
the respective arcs.

PROBLEM SET 20.1A

*1. A product is manufactured to satisfy demand over a 4-period planning horizon according
to the following data:

Unit production Unit holding

Period Units of demand cost ($) cost ($)

1 100 24 1
2 110 26 2
3 95 21 1
4 125 24 2

Given that no back-ordering is allowed, represent the problem as a network model.

2. In Problem 1, suppose that back-ordering is allowed at a penalty of $1.50 per unit per pe-
riod. Formulate the problem as a network model.
3. In Problem 1, suppose that the production capacities of periods 1 to 4 are 110, 95, 125,
and 100 units, respectively, in which case the given demand cannot be satisfied without
back-ordering. Assuming that the penalty cost for back-ordering is $1.5 per unit per peri-
od, formulate the problem as a network model.
4. Daw Chemical owns two plants that manufacture a basic chemical compound for two
customers at the rate of 660 and 800 tons per month. The monthly production capacity
of plant 1 is between 400 and 800 tons and that of plant 2 is between 450 and 900 tons.
The production costs per ton in plants 1 and 2 are $25 and $28, respectively. Raw material
for the plants is provided by two suppliers, who are contracted to ship at least 500 and
700 tons per month for plants 1 and 2 at the costs of $200 and $210 per ton, respectively.
Daw Chemical also pays for the transportation cost of both the raw material and the final
compound. The costs per ton of transporting the raw material from supplier 1 to plants
1 and 2 are $10 and $12. Similar costs from supplier 2 are $9 and $13, respectively. The
transportation costs per ton from plant 1 to clients 1 and 2 are $3 and $4, and from plant
2 are $5 and $2, respectively. Assuming that 1 ton of raw material produces 1 ton of the
final compound, formulate the problem as a network model.
5. Two nonintegrated public schools are required to change the racial balance of their en-
rollments by accepting minority students. Minority enrollment must be between 30%
and 40% in both schools. Nonminority students live in two communities, and minority
students live in three other communities. Traveled distances, in miles, from the five com-
munities to the two schools are summarized in the following table:

Roundtrip miles from school to

Nonminority
Minority areas areas
Maximum
School enrollment 1 2 3 1 2

1 1500 20 12 10 4 5
2 2000 15 18 8 6 5
Student population 500 450 300 1000 1000
CD-4 Chapter 20 Additional Network and LP Algorithms

Formulate the problem as a network model to determine the number of minority

and nonminority students enrolled in each school.
6. (Charnes and Cooper, 1967) During the next three months, farmers harvest their crops
and sell it to a local silo. The silo in turn can sell the crop to retailers. The following table
provides the purchasing cost and selling price per ton during the harvest season.

Purchasing cost Selling price

Month per ton ($) per ton ($)

1 200 250
2 190 230
3 195 240

The maximum capacity of the silo is 800 tons. The owner has $100,000 cash on hand
which can be used to purchase new crops. Also, initially, at the start of month 1, the silo is
half full. It is estimated that the storage cost per ton per month is $10. Surplus cash earns
1% interest monthly. The objective is to determine the buy/sell policy the owner must fol-
low to maximize the total cash accumulation at the end of the three-month season.
Formulate the problem as a network model.

20.1.2 Linear Programming Formulation

The formulation of the capacitated network model as a linear program provides the
foundation for the development of the capacitated simplex algorithm, which we will
present in the next section. Using the notation introduced in Section 20.1.1, the linear
program for the capacitated network is given as

Minimize z = a a cijxij
1i,j2 HA

subject to

a xjk - a xij = fj, j H N

k i
1j,k2HA 1i,j2HA

lij … xij … uij, 1i, j2 H A

The equation for node j measures the net flow fj in node j as

1Outgoing flow from node j2 - 1Incoming flow into node j2 = fj

Node j acts as a source if fj 7 0 and as a sink if fj 6 0.

We can always remove the lower bound lij from the constraints by using the
substitution
xij = xijœ + lij

The new flow variable, xijœ , has an upper limit of uij - lij. Additionally, the net
flow at node i becomes fi - lij, and that at node j is fj + lij. Figure 20.3 shows the trans-
formation of activity (i, j) after the lower bound is substituted out.
 20.1 Minimum-Cost Capacitated Flow Problem CD-5

[ fi] [ fj] [ fi  lij] [ fj
lij]

$ cij $ cij
i j i j
(lij, uij) (uij  lij) FIGURE 20.3
xij xij
Removal of lower bound in arcs

Example 20.1-2
Write the linear program for the network in Figure 20.2, before and after the lower bounds are
substituted out.
The main constraints of the linear program relate the input and output flow for each node,
which yields the following LP:

x12 x13 x14 x23 x25 x34 x35 x46 x56

Minimize 3 4 1 5 6 1 2 2 4

Node 1 1 1 1 = 100
Node 2 -1 1 1 = 200
Node 3 -1 -1 1 1 = 50
Node 4 -1 -1 1 = -150
Node 5 -1 -1 1 = -80
Node 6 -1 -1 = -120

Lower bounds 0 0 50 0 0 70 0 0 100

Upper bounds q q 80 q q 120 q q 120

Note the arrangement of the constraints coefficients. The column associated with variable
xij has exactly one +1 in row i and one -1 in row j, a typical property of network models. The
rest of the coefficients are all 0. This structure is typical of network flow models.
The optimum solution is z = $1,870,000 with x13 = 20 (thousand bushels), x14 = 80,
x23 = 20, x25 = 180, x34 = 90, x46 = 20, and x56 = 100. All the remaining variables are zero.
The variables with lower bounds are substituted as
x14 = x14
œ
+ 50
x34 = x34
œ
+ 70
x56 = x56
œ
+ 100

The substitution leads to the following LP:

x12 x13 x14

œ
x23 x25 x34
œ
x35 x46 x56
œ

Minimize 3 4 1 5 6 1 2 2 4

Node 1 1 1 1 = 50
Node 2 -1 1 1 = 200
Node 3 -1 -1 1 1 = -20
Node 4 -1 -1 1 = -30
Node 5 -1 -1 1 = -180
Node 6 -1 -1 = -20
Upper bounds q q 30 q q 50 q q 20
CD-6 Chapter 20 Additional Network and LP Algorithms

[⫺30]
$1
[50] 1 4
(30) $2
$1
$4
$3 [⫺20] (50) 6 [⫺20]
3 $2 $4
$5 (20)
FIGURE 20.4 $6
[200] 2 5
Network of Example 20.1-2 after substituting out lower
bounds [⫺180]

The corresponding network after substituting out the lower bounds is shown in Figure 20.4.
Note that the lower-bound substitution can be done directly from Figure 20.2 using the substitution
in Figure 20.3, and without the need to express the problem as a linear program first.
The optimum solution is z¿ = $1350 thousand (or z = 1350 + 50 * 1 + 70 * 1 + 100 *
4 = $1870 thousand) with x13 = 20 (thousand bushels), x14 œ
= 30 (or x14 = 30 + 50 = 80),
x23 = 20, x25 = 180, x34 = 20 (or x34 = 20 + 70 = 90), x46 = 20, x56
œ œ
= 0 (or x56 = 0 +
100 = 100), which, of course, is the same solution given by the before-substitution solution.

Example 20.1-3 (Employment Scheduling)

This example illustrates a network model that initially does not satisfy the “node flow” require-
ment (i.e., node output flow less node input flow equals node net flow), but that can be converted
to this form readily through special manipulation of the constraints of the linear program.
Tempo Employment Agency has a contract to provide workers over the next 4 months
(January to April) according to the following schedule:

Month Jan. Feb. March April

No. of workers 100 120 80 170

Because of change in demand, it may be economical to retain more workers than needed in
a given month. The cost of recruiting and maintaining a worker depends on the length of the
employment period, as the following table shows:

Employment period (months) 1 2 3 4

Cost per worker ($) 100 130 180 220

Let
xij = number of workers hired at the start of month
i and terminated at the start of month j
For example, x12 gives the number of workers hired in January for 1 month only. To formu-
late the problem as a linear program for the 4-month period, we add May as a dummy month
(month 5), so that x45 defines hiring in April for April.
 20.1 Minimum-Cost Capacitated Flow Problem CD-7

The constraints recognize that the demand for period k can be satisfied by all xij such that
i … k 6 j. Letting si Ú 0 be the surplus number of workers in month i, the linear program is
given as

x12 x13 x14 x15 x23 x24 x25 x34 x35 x45 s1 s2 s3 s4

Minimize 100 130 180 220 100 130 180 100 130 100

Jan. 1 1 1 1 -1 = 100
Feb. 1 1 1 1 1 1 -1 = 120
March 1 1 1 1 1 1 -1 = 80
April 1 1 1 1 -1 = 170

The preceding LP does not have the 1-1, +12 special structure of the network flow model
(see Example 20.1-2). Nevertheless, the given linear program can be converted into an equivalent
network flow model by using the following arithmetic manipulations:

1. In an n-equation linear program, create a new equation, n + 1, by multiplying equation n

by -1.
2. Leave equation 1 unchanged.
3. For i = 2, 3, Á , n, replace each equation i with 1equation i2 - 1equation i - 12.

The application of these manipulations to the employment scheduling example yields the
following linear program, whose structure fits the network flow model:

x12 x13 x14 x15 x23 x24 x25 x34 x35 x45 s1 s2 s3 s4

Minimize 100 130 180 220 100 130 180 100 130 100

Jan. 1 1 1 1 -1 = 100
Feb. -1 1 1 1 1 -1 = 20
March -1 -1 1 1 1 -1 = -40
April -1 -1 -1 1 1 -1 = 90
May -1 -1 -1 -1 1 = -170

Using the preceding formulation, the employment scheduling model can be represented
equivalently by the minimum cost flow network shown in Figure 20.5 . Actually, because the arcs
have no upper bounds, the problem can be solved also as a transshipment model (see Section 5.5).
The optimum solution is x15 = 100, x25 = 20, x45 = 50, and all the remaining variables are
zero. The following table summarizes the hiring and firing activities over the 4-month horizon.
The total cost is $30,600.

Number of workers Number of workers Net number

Month hired fired of workers

Jan. 100 0 100

Feb. 20 0 120
March 0 0 120
April 50 0 170
CD-8 Chapter 20 Additional Network and LP Algorithms

x15
x14
x13 x35

x12 x34 x45

s2
1 2 3 4 5
s1 s3 s4
[100] [20] x23 [40] [90] [170]
x24
x25

FIGURE 20.5
Network representation of employment scheduling problem

PROBLEM SET 20.1B

*1. Write the linear program associated with the minimum-cost flow network in Figure 20.6,
before and after the lower bounds are substituted out.
2. Use inspection to find a feasible solution to the minimum-cost network model of the em-
ployment scheduling problem in Example 20.1-3 (Figure 20.5 ). Interpret the solution by
showing the pattern of hiring and firing that satisfies the demand for each month, and
compute the associated total cost.
3. Reformulate the employment scheduling model of Example 20.1-3, assuming that a
worker must be hired for at least 2 months. Write the linear program, and convert it to a
minimum-cost flow network.
4. Develop the linear program and the associated minimum-cost flow network for the em-
ployment scheduling model of Example 20.1-3 using the following 5-month demand data.
The per-worker costs of recruiting and maintaining a worker for periods of 1 to 5 months
are $50, $70, $85, $100, and $130, respectively.
(a)
Month 1 2 3 4 5

No. of workers 300 180 90 170 200

(b)
Month 1 2 3 4 5

No. of workers 200 220 300 50 240

FIGURE 20.6 [20]

Network for Problem 1, Set 20.1b 3
(3 $5

$6 )
)
40

(0
,
0,

$4 (10, )

$1 $3
[50] 1 2 4 [30]
(0, ) (10, )
[40]
 20.1 Minimum-Cost Capacitated Flow Problem CD-9

2
$1 $8
) (50
(60 )
[100] 1 4 [100]
$2 (50)
$4
( $1
) 0)
(9
FIGURE 20.7
3
Network for Problem 5, Set 20.1b

5. Conversion of a capacitated network into an uncapacitated network. Show that an arc

1i : j2 with capacitated flow xij … uij can be replaced with two uncapacitated arcs
1i : k2 and 1j : k2 with a net (output) flow of [-uij] at node k and an additional (input)
flow of [+uij] at node j. The result is that the capacitated network can be converted to an
uncapacitated transportation model (Section 5.1). Apply the resulting transformation to
the network in Figure 20.7 and find the optimum solution to the original network by solv-
ing the uncapacitated transportation model.

20.1.3 Capacitated Network Simplex Algorithm

The algorithm is based on the exact steps of the regular simplex method, but is de-
signed to exploit the special network structure of the minimum-cost flow model.
Given that fi is the net flow at node i as defined in the linear program of Section
20.1.2, the capacitated simplex algorithm stipulates that the network must satisfy
n

a fi = 0
i=1

The condition says that the total supply in the network equals the total demand.
We can always satisfy this requirement by adding a balancing dummy source or desti-
nation, which we connect to all other nodes in the network by arcs with zero unit cost
and infinite capacity. However, the balancing of the network does not guarantee feasi-
bility, as this depends on the capacities of the arcs.
We now present the steps of the capacitated algorithm. Familiarity with the sim-
plex method and duality theory (Chapters 3 and 4) is essential. Also, knowledge of the
upper-bounded simplex method (Section 7.3) is helpful.

Step 0. Determine a starting basic feasible solution (set of arcs) for the network. Go
to step 1.
Step 1. Determine an entering arc (variable) using the simplex method optimality
condition. If the solution is optimal, stop; otherwise, go to step 2.
Step 2. Determine the leaving arc (variable) using the simplex method feasibility
condition. Determine the new solution, then go to step 1.

An n-node network with zero net flow (i.e., f1 + f2 + Á + fn = 0) consists of

n - 1 independent constraint equations. Thus, a basic solution must include n - 1 arcs.
It can be proved that a basic solution always corresponds to a spanning tree of the net-
work (see Section 6.2).
CD-10 Chapter 20 Additional Network and LP Algorithms

The entering arc (step 1) is determined by computing the objective coefficients

for all the current nonbasic arcs (i, j). If all the objective coefficients are …0, the current
basis is optimum. Otherwise, we select the nonbasic arc with the most positive objec-
tive coefficients to enter the basis.
The computation of objective coefficients is based on duality, exactly as we did
with the transportation model (see Section 5.3.2). Using the linear program defined in
Section 20.1.2, let wi be the dual variable associated with the constraint of node i, then
the dual problem (excluding the upper bounds) is given as
n
Maximize z = a fiwi
i=1
subject to
wi - wj … cij, 1i, j2 H A
wi unrestricted in sign, i = 1, 2, Á n
From the theory of linear programming, we have
wi - wj = cij, for basic arc1i, j2
Because the original linear program (Section 20.1.2) has one redundant constraint by
definition, we can assign an arbitrary value to one of the dual variables (compare with
the transportation algorithm, Section 5.3). For convenience, we will set w1 = 0. We
then solve the (basic) equations wi - wj = cij to determine the remaining dual values.
From Section 4.2.4, Formula 2, we know that the objective coefficient of nonbasic xij is
the difference between the left-hand side and the right-hand side of the associated dual
constraint; that is, wi - wj - cij. Letting
zij = wi - wj
the objective coefficient of the variable xij is computed as
zij - cij = wi - wj - cij
The only remaining detail is to show how the leaving variable is determined. The
following numeric example provides this detail.

Example 20.1-4
A network of pipelines connects two water desalination plants to two cities. The daily supply
amounts at the two plants are 40 and 50 million gallons and the daily demand amounts at cities 1
and 2 are 30 and 60 million gallons. Nodes 1 and 2 represent plants 1 and 2, and nodes 4 and 5
represent cities 1 and 2. Node 3 is a booster station between the plants and the cities. The model
is already balanced because the sum of the supply at nodes 1 and 2 equals the sum of the demand
at nodes 4 and 5. Figure 20.8 gives the associated network.

Iteration 0

Step 0. Determination of a starting basic feasible solution: The starting basic solution must be
a spanning tree. The feasible spanning tree in Figure 20.9 (shown with solid arcs) is
 20.1 Minimum-Cost Capacitated Flow Problem CD-11

Unit cost Arc capacity

$5
Plant 1 [40] 1 4 [30] City 1
(35)

$7 )
(1
0
$3 () 3 $4 ()
$8
(

(6 2
$
)

0)
$1 FIGURE 20.8
Plant 2 [50] 2 5 [60] City 2
(30) Network for Example 20.1-4

w1  0 w4  5
$5
[40] 1 4 [30] z12  c12  0  (5) 3  2
30(35)
z25  c25  5  (15) 1  9
z45  c45  5  (15)  4  6
$7 0)
10
(1

w3  7
$3 $4 () Arc (2, 5) reaches upper bound at 30.
3
 $ Substitute x25  30  x52.
60 8
(
50 $2
0)

)
(6

Reduce x23 and x35 each by 30.


$1

[50] 2 5 [60]
(30)
w2  5 w5  15

FIGURE 20.9
Network for Iteration 0

obtained by inspection. Normally, we use an artificial-variable technique to find such

a solution (for details, see Bazaraa et al., 1990, pp. 440–446).
In Figure 20.9, the basic feasible solution consists of (solid) arcs (1, 3), (1,4), (2, 3),
and (3, 5) with the feasible flows of 30, 10, 50, and 60 units, respectively. The remaining
(dashed) arcs (1, 2), (2, 5), and (4, 5) represent the nonbasic variables. The notation
x(c) shown on the arcs indicates that a flow of x units is assigned to an arc with capac-
ity c. The default values for x and c are 0 and q , respectively.

Iteration 1

Step 1. Determination of the entering arc: We obtain the dual values by solving the current
basic equations

w1 = 0
wi - wj = cij, for basic arc1i, j2

We thus get

Arc11, 32: w1 - w3 = 7, hence w3 = -7

Arc11, 42: w1 - w4 = 5, hence w4 = -5

CD-12 Chapter 20 Additional Network and LP Algorithms

Arc12, 32: w2 - w3 = 2, hence w2 = -5

Arc13, 52: w3 - w5 = 8, hence w5 = -15

Now, we compute zij - cij for the nonbasic arcs as

Arc11, 22: w1 - w2 - c12 = 0 - 1-52 - 3 = 2

Arc12, 52: w2 - w5 - c25 = 1-52 - 1-152 - 1 = 9

Arc14, 52: w4 - w5 - c45 = 1-52 - 1-152 - 4 = 6

Thus, arc (2, 5) enters the basic solution.

Step 2. Determination of the leaving arc: From Figure 20.9, arc (2, 5) forms a loop with basic
arcs (2, 3) and (3, 5). From the definition of the spanning tree, no other loop can be
formed. Because the flow in the new arc (2, 5) must be increased, we adjust the flow in
the arcs of the loop by an equal amount to maintain the feasibility of the new solution.
To achieve this, we identify the positive 1+2 flow in the loop by the direction of flow
of the entering arc (i.e., from 2 to 5). We then assign 1+2 or 1-2 to the remaining arcs of
the loop, depending on whether the flow of each arc is with or against the direction of
flow of the entering arc. These sign conventions are shown in Figure 20.9.
Determination of the maximum level of flow in the entering arc (2, 5) is based on
two conditions:

1. New flow in the current basic arcs of the loop cannot be negative.
2. New flow in the entering arc cannot exceed its capacity.

The application of condition 1 shows that the flows in arcs (2, 3) and (3, 5) cannot
be decreased by more than min550, 606 = 50 units. By condition 2, the flow in arc (2, 5)
can be increased to at most the arc capacity 1= 30 units2. Thus, the maximum flow
change in the loop is min530, 506 = 30 units. The new flows in the loop are thus 30
units in arc (2, 5), 50 - 30 = 20 units in arc (2, 3), and 60 - 30 = 30 units in arc (3, 5).
Because none of the current basic arcs leave the basis at zero level, the new arc
(2, 5) must remain nonbasic at upper bound. However, to avoid dealing with nonbasic
arcs that are at capacity (or upper bound) level, we implement the substitution

x25 = 30 - x52, 0 … x52 … 30

This substitution is effected in the flow equations associated with nodes 2 and 5 as fol-
lows. Consider

Current flow equation at node 2: 50 + x12 = x23 + x25

Current flow equation at node 5: x25 + x35 + x45 = 60

Then, the substitution x25 = 30 - x52 gives

New flow equation at node 2: 20 + x12 + x52 = x23

New flow equation at node 5: x35 + x45 = x52 + 30

 20.1 Minimum-Cost Capacitated Flow Problem CD-13

w1  0 w4  5
$5
[40] 1 4 [30] z12  c12  0  (5) 3  2
30(35)

z52  c52  15 (5) (1)  9
z45  c45  5  (15) 4  6

$7 0)
10
 w  7

(1
3
$3 $4 () Arc (4, 5) enters at level 5.
3
$ Arc (1, 4) leaves at upper bound.
30 8
(

20 $2
0)
) 
Substitute x14  35  x41.

(6
$1
[20] 2 5 [30] Reduce x13 and x35 each by 5.
(30)*
w2  5 w5  15

FIGURE 20.10
Network for iteration 1

The results of these changes are shown in Figure 20.10. The direction of flow in arc
(2, 5) is now reversed to 5 : 2 with x52 = 0, as desired. The substitution also requires
changing the unit cost of arc (5, 2) to - $1. We will indicate this direction reversal on
the network by tagging the arc capacity with an asterisk.

Iteration 2. Figure 20.10 summarizes the new values of zij - cij (verify!) and shows that arc (4, 5)
enters the basic solution. It also defines the loop associated with the new entering arc
and assigns the signs to its arcs.
The flow in arc (4, 5) can be increased by the smallest of

1. Maximum allowable increase in entering arc14, 52 = q

2. Maximum allowable increase in arc11, 42 = 35 - 30 = 5 units
3. Maximum allowable decrease in arc11, 32 = 10 units
4. Maximum allowable decrease in arc13, 52 = 30 units

Thus, the flow in arc (4, 5) can be increased to 5 units, which will make (4, 5)
basic and will force basic arc (1, 4) to be nonbasic at its upper bound 1= 352.
Using the substitution x14 = 35 - x41, the network is changed as shown in
Figure 20.11, with arcs (1, 3), (2, 3), (3, 5), and (4, 5) forming the basic (spanning
tree) solution. The reversal of flow in arc (1, 4) changes its unit cost to - $5. Also,
convince yourself that the substitution in the flow equations of nodes 1 and 4 will
net 5 input units at each node.

Iteration 3. The computations of the new zij - cij for the nonbasic arcs (1, 2), (4, 1), and (5, 2)
are summarized in Figure 20.11, which shows that arc (1, 2) enters at level 5, and
arc (1,3) becomes nonbasic at level 0. Figure 20.12 depicts the new solution.

Iteration 4. The new zij - cij in Figure 20.12 shows that the solution is optimum. Back substi-
tution yields the values of the original variables as shown in Figure 20.12.
CD-14 Chapter 20 Additional Network and LP Algorithms

w1  0 w4  11
$5
[5] 1 4 [5] z12  c12  0  (5) 3  2
(35)*
z41  c41  11 0 (5)  6
z52  c52  15  (5) (1)  9

$7 0)
5(
1
w3  7
$3  $4 5() Arc (1, 2) enters at level 5.
3


$ Arc (1, 3) leaves at level 0.
25 8
(

20 $2
0)
)

(6
Increase x23 by 5.
$1
[20] 2 5 [30]
(30)*
w2  5 w5  15

FIGURE 20.11
Network for iteration 2

w1  0 w4  9
$5 z13  c13  0  (5) 7  2
[5] 1 4 [5]
(35)* z41  c41  9 0 (5)  4
z52  c52  13  (3) (1)  9
$7 )
(1

w3  5
0

Optimum solution:
$3 5() $4 5() x12  5, x13  0
3
$ x14  35  0  35
25 8 x23  25
(
25 $2
0)

) x25  30  0  30
(6

$1 x35  25, x45  5

[20] 2 5 [30] Total cost  $490
(30)*
w2  3 w5  13

FIGURE 20.12
Network for iteration 3

Solver Moment
The basic idea of the Excel Solver minimum cost capacitated model is similar to the
one detailed in Example 6.3-6 for the shortest-route model. File solverEx20.1-4.xls
provides the details.

AMPL Moment
File amplEx20.1-4.txt is a general model that can be used to solve the minimum cost
capacitated problem of any size. The idea of the model is similar to that of the shortest-
route AMPL model for Example 6.3-6.

PROBLEM SET 20.1C

*1. Solve Problem 1, Set 20.1a, by the capacitated simplex algorithm, and also show that it
can be solved by the transshipment model.
 20.1 Minimum-Cost Capacitated Flow Problem CD-15

2. Solve Problem 2, Set 20.1a, by the capacitated simplex algorithm, and also show that it
can be solved by the transshipment model.
3. Solve Problem 3, Set 20.1a by the capacitated simplex algorithm.
4. Solve Problem 4, Set 20.1a, by the capacitated simplex algorithm.
*5. Solve Problem 5, Set 20.1a, by the capacitated simplex algorithm.
6. Solve the employment scheduling problem of Example 20.1-3 by the capacitated simplex
algorithm.
7. Wyoming Electric uses existing slurry pipes to transport coal (carried by pumped water)
from three mining areas (1, 2, and 3) to three power plants (4, 5, and 6). Each pipe can
transport at most 10 tons per hour. The transportation costs per ton and the supply and
demand per hour are given in the following table:

4 5 6 Supply

1 $5 $8 $4 8

2 $6 $9 $12 10

3 $3 $1 $5 18

Demand 16 6 14

Determine the optimum shipping schedule.

8. The network in Figure 20.13 gives the distances among seven cities. Use the capacitated
simplex algorithm to find the shortest distance between nodes 1 and 7. (Hint: Assume
that nodes 1 and 7 have net flows of [+1] and [-1], respectively. All the other nodes have
zero net flow.)
9. Show how the capacitated minimum-cost flow model can be specialized to represent the
maximum flow model of Section 6.4. Apply the transformation to the network in Exam-
ple 6.4-2. For convenience, assume that the flow capacity from 4 to 3 is zero. All the re-
maining data are unchanged.
10. Solve the following problems using Excel Solver and AMPL:
(a) Problem 3, Set 20.1a.
(b) Problem 4, Set 20.1a.
(c) Problem 7, Set 20.1c.
(d) Problem 8, Set 20.1c.

FIGURE 20.13
Network for Problem 8, Set 20.1c

5
2 5

2 11 6
8

10
1 4 7

4 3 7 9

1
3 6
CD-16 Chapter 20 Additional Network and LP Algorithms

20.2 DECOMPOSITION ALGORITHM

Consider the situation of developing a master corporate plan for several production fa-
cilities. Although each facility has its own independent capacity and production con-
straints, the different facilities are linked together at the corporate level by budgetary
considerations. The resulting model includes two types of constraints: common repre-
senting the corporate budgetary constraints; and independent, representing the internal
capacity and production restrictions of each facility. Figure 20.14 depicts the layout of
the resulting constraints for n activities (facilities). In the absence of the common con-
straints, all activities operate independently.
The decomposition algorithm improves the computational efficiency of the prob-
lem depicted in Figure 20.14 by breaking it down into n subproblems that can be solved
almost independently. We point out, however, that the need for the decomposition al-
gorithm was more justifiable in the past when the speed and memory of the computer
were modest. Today’s impressive computer capabilities obviate the use of the decom-
position algorithm. Nevertheless, we present the algorithm here because of its interest-
ing theoretical contribution.
The corresponding mathematical model is given as

Maximize z = C 1X 1 + C 2X 2 + Á + C nX n
subject to
A 1X 1 + A 2X 2 + Á + A nX n … b0
D1X 1 … b1
D2X 2 … b2
o o
DnX n … bn
X j Ú 0, j = 1, 2, Á , n

FIGURE 20.14
Layout of a decomposable linear program

Common constraints

Activity
1

Activity
Independent
2
constraints
…

Activity
n
 20.2 Decomposition Algorithm CD-17

The slack and surplus variables are added as necessary to convert all the inequal-
ities into equations.
The decomposition principle is based on representing the entire problem in
terms of the extreme points of the sets 5X ƒ DjX j … bj, X j Ú 06, j = 1, 2, Á , n. To do
so, the solution space described by each 5X ƒ DjX j … bj, X j Ú 06 must be bounded. This
requirement can always be satisfied for any set j by adding the artificial restriction
1X j … M, where M is sufficiently large.
Let XN , k = 1, 2, Á , K define the extreme points of 5X ƒ D X … b , X Ú 06.
jk j j j j j
We then have
Kj
N , j = 1, 2, Á , n
X j = a b jkX jk
k=1

kj
where b jk Ú 0 for all k and a b jk = 1
k=1
We can reformulate the entire problem in terms of the extreme points to obtain
the following master problem:
K1 K2 Kn
N b +
Maximize = a C 1X N N
a C 2X 2k b 2k + a C nX nk b nk
1k 1k
Á +
k=1 k=1 k=1

subject to
K1 K2 Kn
N N N
a A 1X 1k b 1k + a A 2X 2k b 2k + a A nX nk b nk … b0
Á +
k=1 k=1 k=1
Kj

a b 1k = 1
k=1
Kj

a b 2k = 1
k=1

o o

Kj

a b nk = 1
k=1

b jk Ú 0, for all j and k

The new variables in the master problem are b jk. Once their optimal values, b …jk,
are determined, we can find the optimal solution to the original problem by back sub-
stitution:
Kj
N ,
X j = a b …jkX j = 1, 2, Á , n
jk
k=1

It may appear that the solution of the master problem requires prior determina-
N , a difficult task indeed! Fortunately, it is not so.
tion of all the extreme points X jk
CD-18 Chapter 20 Additional Network and LP Algorithms

To solve the master problem by the revised simplex method (Section 7.2), we
need to determine the entering and the leaving variables at each iteration. For the en-
tering variable, given C B and B-1 of the current basis of the master problem, then for
nonbasic b jk we have
zjk - cjk = C BB-1Pjk - cjk
where
A jXN
jk
0
N and P
cjk = C jX = ¶
o
∂
jk jk
1
o
0

Now, to decide which, if any, of the nonbasic variables b jk should enter the solu-
tion, we need to determine
zj*k* - cj*k* = min 5zjk - cjk6
all j and k

If zj*k* - cj*k* 7 0, then, according to the maximization optimality condition, b j*k*

must enter the solution; otherwise, the optimum has been reached.
We still have not shown how zj*k* - cj*k* is computed numerically. The idea lies
in the following identity

min 5zjk - cjk6 = min e min5zjk - cjk6 f

all j and k j k

The reason we are able to establish this identity is that each convex set
5X ƒ DjX j … bj, X j Ú 06 has its independent set of extreme points. In effect, what the
identity says is that we can determine zj*k* - cj*k* in two steps:

Step 1. For each convex set 5X ƒ DjX j … bj, X j Ú 06, determine the extreme point
N that yields the smallest z - c —that is, z - c = min 5z - c 6.
Yjk* jk jk jk* jk* k jk jk

Step 2. Determine zj*k* - cj*k* = min5zjk* - cjk*6

j

From LP theory, we know that the optimum solution, when finite, must be associ-
ated with an extreme point of the solution space. Because each of the sets
5X ƒ DjX j … bj, X j Ú 06 is bounded by definition, step 1 is mathematically equivalent
to solving n linear programs of the form
Minimize wj = 5zj - cj ƒ DjX j … bj, X j Ú 06

The objective function wj is a linear function in X j (see Problem 8, Set 20.2a).

The determination of the entering variable b j*k* in the master problem reduces
to solving n (smaller) linear programs to determine the “entering” extreme point YN
j*k*.
 20.2 Decomposition Algorithm CD-19

This approach precludes the need to determine all the extreme points of all n convex
sets. Once the desired extreme point is located, all the elements of the column vector
Pj*k* are defined. We can then determine the leaving variable, and subsequently com-
pute the next B-1 using the revised simplex method computations.

Example 20.2-1
Solve the following LP by the decomposition algorithm:
Maximize z = 3x1 + 5x2 + x3 + x4
subject to
x1 + x2 + x3 + x4 … 40
5x1 + x2 … 12
x3 + x4 Ú 5
x3 + 5x4 … 50
x1, x2, x3, x4 Ú 0
The problem has two subproblems that correspond to the following sets of variables:
X 1 = 1x1, x22T, X 2 = 1x3, x42T

The associated master problem is thus given as

Starting basic
Subproblem 1 Subproblem 2 solution

b 11 b 12 Á b 1K1 b 21 b 22 Á b 2K2 x5 x6 x7
N
C 1X N
C 1X Á N
C 1X N
C 2X N
C 2X Á N
C 2X 0 -M -M
11 12 1K1 21 22 2K2

A 1XN A 1XN Á N
A 1X A 2XN A 2XN Á N
A 2X 1 0 0 = 40
11 12 1K1 21 22 2K2
1 1 Á 1 0 0 Á 0 0 1 0 = 1
0 0 Á 0 1 1 Á 1 0 0 1 = 1

C 1 = 13, 52 C 2 = 11, 12
A 1 = 11, 12 A 2 = 11, 12
Solution space, D1X 1 … b1: Solution space, D2X 2 … b2:
5x1 + x2 … 12 x3 + x4 Ú 5
x1, x2 Ú 0 x3 + 5x4 … 50
x3, x4 Ú 0

Notice that x5 is the slack variable that converts the common constraint to the following
equation:
x1 + x2 + x3 + x4 + x5 = 40

Recall that subproblems 1 and 2 account for variables x1, x2, x3, and x4 only. This is the reason x5
must appear explicitly in the master problem. The remaining starting basic variables, x6 and x7,
are artificial.
CD-20 Chapter 20 Additional Network and LP Algorithms

Iteration 0

X B = 1x5, x6, x72T = 140, 1, 12T

C B = 10, -M, -M2, B = B-1 = I

Iteration 1
Subproblem 1 1j = 12. We have

A 1X 1
z1 - c1 = C BB-1 £ 1 ≥ - C 1X 1
0

11, 12a b
x1
x2
= 10, -M, -M2• µ - 13, 52a b
x1
1
x2
0

= -3x1 - 5x2 - M

Thus, the corresponding LP is

Minimize w1 = -3x1 - 5x2 - M

subject to
5x1 + x2 … 12
x1, x2 Ú 0

The solution of this problem (by the simplex method) yields

N = 10, 122T, z… - c… = w… = -60 - M

X 11 1 1 1

Subproblem 2 1j = 22. The associated linear program is given as

A 2X 2
Minimize z2 - c2 = C BB-1 £ 0 ≥ - C 2X 2
1

11, 12a b
x3
x4
= 10, -M, -M2• µ - 11, 12a b
x4
0
x5
1

= -x4 - x5 - M
 20.2 Decomposition Algorithm CD-21

subject to

x3 + x4 Ú 5
x3 + 5x4 … 50
x 3, x 4 Ú 0

The optimal solution of the problem yields

N = 150, 02T, z… - c… = -50 - M
X 21 2 2

Because the master problem is of the maximization type and z…1 - c…1 6 z…2 - c…2 and
N enters the solution.To determine
z… - c… 6 0, it follows that b associated with extreme point X
1 1 11 11
the leaving variable,

11, 12a b
0
A 1XN 12 12
11
P11 = £ 1 ≥ = • 1 µ = £ 1≥
0 0 0

Thus, B-1P11 = 112, 1, 02T. Given X B = 1x5, x6, x72T = 140, 1, 12T, it follows that x 6 (an artificial
variable) leaves the basic solution (permanently).
The new basis is determined by replacing the vector associated with x6 with the vector P11,
which gives (verify!)

1 12 0 1 -12 0
B = £0 1 0 ≥ Q B-1 = £ 0 1 0≥
0 0 1 0 0 1

The new basic solution is

X B = 1x5, b 11, x72T = B-1140, 1, 12T = 128, 1, 12T

C B = 10, C 1X
N , -M2 = 10, 60, -M2
11

Iteration 2

Subproblem 1 1j = 12. The associated objective function is

Minimize w1 = -3x1 - 5x2 + 60

(verify!). The optimum solution yields z…1 - c…1 = w1 = 0, which means that none of the remain-
ing extreme points in subproblem 1 can improve the solution to the master problem.
Subproblem 2 1j = 22. The associated objective function is (coincidentally) the same as for
j = 2 in Iteration 1 (verify!). The optimum solution yields
N = 150, 02T, z… - c… = -50 - M
X 22 2 2
CD-22 Chapter 20 Additional Network and LP Algorithms

Note that XN is actually the same extreme point as X N . The subscript 2 is used for notational
22 21
convenience to represent iteration 2.
N
From the results of the two subproblems, z…2 - c…2 6 0 indicates that b 22 associated with X 22
enters the basic solution.
To determine the leaving variable, consider

11, 12a b
50
A 2XN 0 50
22
P22 = £ 0 ≥ = • 0 µ = £ 0≥
1 1 1

Thus, B-1P22 = 150, 0, 12T. Because X B = 1x5, b 11, x72T = 128, 1, 12T, x5 leaves.
The new basis and its inverse are given as (verify!)

1
50 12 0 50 - 12
50 0
B = £ 0 1 0 ≥ Q B-1 = £ 0 1 0≥
1 12
1 0 1 - 50 50 1

X B = 1b 22, b 11, x72T = B-1140, 1, 12T = A 14

25 , 1, 25 B
11 T

C B = 1C 2X 1 11, -M2 = 150, 60, -M2

N ,C X
22
N

Iteration 3
Subproblem 1 1j = 12. You should verify that the associated objective function is

Minimize w1 = AM
50 - 2 B x1 + A 50 - 4 B x2 -
M 12M
50 + 48

The associated optimum solution is

N = 10, 02T, z… - c… = - 12M + 48

X 13 1 1 50

Subproblem 2 1j = 22. The objective function can be shown to equal (verify!)

Minimize w2 = AM
50 B 1x3 + x42 - M

The associated optimum solution is

N = 15, 02T, z… - c… = - 9M
X 23 2 2 10

Nonbasic Variable x 5. From the definition of the master problem, zj - cj of x5 must be comput-
ed and compared separately. Thus,

z5 - c5 = C BB-1P5 - c5
= A1 + M
50 , 48 - 12M
50 , -M B 11, 0, 02T - 0
M
= 1 + 50

Thus, x5 cannot improve the solution.

 20.2 Decomposition Algorithm CD-23

N enters the basic solution. To determine

From the preceding information, b 23 associated with X 23
the leaving variable, consider

11, 12a b
5
A 2XN 0 5
23
P23 = £ 0 ≥ = • 0 µ = £0≥
1 1 1

Thus, B-1P23 = A 10
1 9 T
, 0, 10 B . Given that X B = 1b 22, b 11, x72T = A 14
25 , 1, 25 B , the artificial variable
11 T

x10 leaves the basic solution (permanently).

The new basis and its inverse are thus given as (verify!)
1
50 12 5 45 - 12
45
5
- 45
B = £ 0 1 0 ≥ Q B-1 = £ 0 1 0≥
1 12 50
1 0 1 - 45 45 45

X B = 1b 22, b 11, b 232T = B-1140, 1, 12T = A 23

B
22 T
45 , 1, 45

CB = 1C 2X 22 1 11, C 2X 232 = 150, 60, 52

N ,C XN N

Iteration 4
Subproblem 1 1j = 12. w1 = -2x1 - 4x2 + 48. It yields z…1 - c…1 = w…1 = 0.
Subproblem 2 1j = 22. w2 = 0x4 + 0x5 + 48 = 48.
Nonbasic Variable x5: z5 - c5 = 1. The preceding information shows that Iteration 3 is optimal.
We can compute the optimum solution of the original problem by back-substitution:

X …1 = 1x1, x22T = b 11X

N = 110, 122T = 10, 122T
11

X …2 = 1x4, x52T = b 22X

N + b X
22
N
23 23

= A 23
45 B A 50, 02 + A 45 B 15, 0 B
T 22 T

= 128, 02T

All the remaining variables are zero. The optimum value of the objective function can be
obtained by direct substitution.

PROBLEM SET 20.2A

1. In each of the following cases, determine the feasible extreme points graphically and ex-
press the feasible solution space as a function of these extreme points. If the solution
space is unbounded, add a proper artificial constraint.
(a)
x1 + 2x2 … 6
2x1 + x2 … 8
-x1 + x2 … 1
x2 … 2
x1, x2 … 0
CD-24 Chapter 20 Additional Network and LP Algorithms

(b)
2x1 + x2 … 2
3x1 + 4x2 Ú 12
x1, x2 Ú 0
*(c)
x1 - x2 … 10
2x1 … 40
x1, x2 Ú 0

*2. In Example 20.2-1, the feasible extreme points of subspaces D1X 1 = b1, X 1 Ú 0 and
D2X 2 = b2, X 2 Ú 0 can be determined graphically. Use this information to express the
associated master problem explicitly in terms of the feasible extreme points. Then show
that the application of the simplex method to the master problem produces the same
entering variable b jk as that generated by solving subproblems 1 and 2. Hence, convince
yourself that the determination of the entering variable b jk is exactly equivalent to solv-
ing the two minimization subproblems.
3. Consider the following linear program:
Maximize z = x1 + 3x2 + 5x3 + 2x4
subject to
x1 + 4x2 … 8
2x1 + x2 … 9
5x1 + 3x2 + 4x3 Ú 10
x3 - 5x4 … 4
x3 + x4 … 10
x1, x2, x3, x4 Ú 0

Construct the master problem explicitly by using the extreme points of the subspaces,
and then solve the resulting problem directly by the simplex method.
4. Solve Problem 3 using the decomposition algorithm and compare the two procedures.
5. Apply the decomposition algorithm to the following problem:
Maximize z = 6x1 + 7x2 + 3x3 + 5x4 + x5 + x6

subject to
x1 + x2 + x3 + x4 + x5 + x6 … 50
x1 + x2 … 10
x2 … 8
5x3 + x4 … 12
x5 + x6 Ú 5
x5 + 5x6 … 50
x1, x2, x3, x4, x5, x6 Ú 0
 20.3 Karmarkar Interior-Point Method CD-25

*6. Indicate the necessary changes for applying the decomposition algorithm to minimization
LPs. Then solve the following problem:
Minimize z = 5x1 + 3x2 + 8x3 - 5x4
subject to
x1 + x3 + x3 + x4 Ú 25
5x1 + x2 … 20
5x1 - x2 Ú 5
x3 + x4 = 20
x 1 , x 2, x 3, x 4 Ú 0

7. Solve the following problem by the decomposition algorithm:

Minimize z = 10y1 + 2y2 + 4y3 + 8y4 + y5
subject to
y1 + 4y2 - y3 Ú 8
2y1 + y2 + y3 Ú 2
3y1 + y4 + y5 Ú 4
y1 + 2y4 - y5 Ú 10
y1, y2, y3, y4, y5 Ú 0

(Hint: Solve the dual problem first by decomposition.)

8. In the decomposition algorithm, suppose that the number of common constraints in the
original problem is r. Show that the objective function for subproblem j can be written as
Minimize wj = zj - cj = 1C BRA j - C j2X j + C BVr + j

The vector R represents the first r columns of B-1 and Vr + j is its 1r + j2th column.

20.3 KARMARKAR INTERIOR-POINT METHOD

The simplex method obtains the optimum solution by following a path of adjacent ex-
treme points along the edges of the solution space. Although in practice the simplex
method has served well in solving large problems, theoretically the number of itera-
tions needed to reach the optimum solution can grow exponentially. In fact, researchers
have constructed a class of LPs in which all feasible extreme points are visited before
the optimum is reached.
In 1984, N. Karmarkar developed a polynomial-time algorithm that cuts across
the interior of the solution space. The algorithm is effective for extremely large LPs.
We start by introducing the main idea of the Karmarkar method and then pro-
vide the computational details of the algorithm.

20.3.1 Basic Idea of the Interior-Point Algorithm

Consider the following (trivial) example:
Maximize z = x1
CD-26 Chapter 20 Additional Network and LP Algorithms

subject to
0 … x1 … 2
Using x2 as a slack variable, the problem can be rewritten as
Maximize z = x1
subject to
x1 + x2 = 2
x1, x2 Ú 0

Figure 20.15 depicts the problem. The solution space is given by the line segment
AB. The direction of increase in z is in the positive direction of x1.
Let us start with any arbitrary interior (nonextreme) point C in the feasible space
(line AB). The gradient of the objective function (maximize z = x1) at C is the direc-
tion of fastest increase in z. If we locate an arbitrary point along the gradient and then
project it perpendicularly on the feasible space (line AB), we obtain the new point D
with a better objective value z. Such improvement is obtained by moving in the direc-
tion of the projected gradient CD. If we repeat the procedure at D, we will determine
a new closer-to-optimum point E. Conceivably, if we move (cautiously) in the direction
of the projected gradient, we will reach the optimum at point B. If we are minimizing z
(instead of maximizing), the projected gradient will correctly move us away from point
B toward the minimum at point A1x1 = 02.
The given steps hardly define an algorithm in the normal sense, but the idea is in-
triguing! We need some modifications that will guarantee that (1) the steps generated
along the projected gradient will not “overshoot” the optimum point at B, and (2) in

FIGURE 20.15
Illustration of the general idea of Karmarkar’s algorithm

x2

2
A

Gradient of z Maximize
z = x1

C
1

B x1
0 1 2
 20.3 Karmarkar Interior-Point Method CD-27

the general n-dimensional case, the direction created by the projected gradient will not
cause an “entrapment” of the algorithm at a nonoptimum point. This, basically, is what
Karmarkar’s interior-point algorithm accomplishes.

20.3.2 Interior-Point Algorithm

Several variants of Karmarkar’s algorithm are available in the literature. Our presen-
tation follows the original algorithm. Karmarkar assumes that the LP is given as
Minimize z = CX
subject to
AX = 0
1X = 1
X Ú 0
All the constraints are homogeneous equations except for the constraint 1X =
n
a j = 1x j
= 1, which defines an n-dimensional simplex. The validity of Karmarkar’s algo-
rithm rests on satisfying two conditions:

1. X = A n1 , n1 , Á , n1 B satisfies AX = 0.
2. min z = 0.

Karmarkar provides modifications that allow solving the problem when the second
condition is not satisfied. These modifications will not be presented here.
The following example illustrates how a general LP can be made to satisfy the
two stipulated conditions.

Example 20.3-1
Consider the problem.
Maximize z = y1 + y2
subject to

y1 + 2y2 … 2
y1, y2 Ú 0
We start by defining the primal and dual problems of the LP:

Primal Dual

Maximize y0 = y1 + y2 Minimize w0 = 2w1

subject to subject to

f Q w1 Ú 1
y1 + 2y2 … 2 w1 Ú 1
y1, y2 Ú 0 2w1 Ú 1
w1 Ú 0
CD-28 Chapter 20 Additional Network and LP Algorithms

The primal and dual constraints can be put in equation form as

y1 + 2y2 + y3 = 2, y3 Ú 0
w1 - w2 = 1, w2 Ú 0 (20.1)

At the optimum y0 = w0, which yields

y1 + y2 - 2w1 = 0 (20.2)

Selecting M sufficiently large, we have

y1 + y2 + y3 + w1 + w2 … M (20.3)

Now, converting (3) into an equation, we get

y1 + y2 + y3 + w1 + w2 + s1 = M, s1 Ú 0 (20.4)

Next, define a new variable s2. From (4) the following two equations hold if, and only if, the con-
dition s2 = 1 holds:

y1 + y2 + y3 + w1 + w2 + s1 - Ms2 = 0
y1 + y2 + y3 + w1 + w2 + s1 + s2 = M + 1 (20.5)

Now, given s2 = 1 as stipulated by (5), the primal and dual equations (1) can be written as

y1 + 2y2 + y3 - 2s2 = 0
w1 - w2 - 1s2 = 0 (20.6)

Now, define

yj = 1M + 12xj, j = 1, 2, 3
wj - 3 = 1M + 12xj, j = 4, 5
s1 = 1M + 12x6
s2 = 1M + 12x7

Substitution in equations (2), (5), and (6) will produce the following equations:

x1 + x2 - 2x4 = 0
x1 + x2 + x3 + x4 + x5 + x6 - Mx7 = 0
x1 + x2 + x3 + x4 + x5 + x6 + x7 = 1
x1 + 2x2 + x3 - 2x7 = 0
x4 - x5 - x7 = 0
xj Ú 0, j = 1, 2, Á , 7

The final step calls for augmenting the artificial variable y8 in the left-hand side of each
equation. The new objective function will call for minimizing y8, whose obvious minimum value
 20.3 Karmarkar Interior-Point Method CD-29

must be zero (assuming the primal is feasible). Note, however, that Karmarkar’s algorithm re-
quires the solution
X = A 18, 18, 18, 18, 18, 18, 18, 18 B T
to be feasible for AX = 0. This will be true for the homogeneous equations (with zero right-hand
side) if the associated coefficient of the artificial x8 equals the (algebraic) sum of all the coefficients
on the left-hand side. It thus follows that the transformed LP is given as
Minimize z = x8
subject to

x1 + x2 - 2x4 - 0x8 = 0
x1 + x2 + x3 + x4 + x5 + x6 - Mx7 - 16 - M2x8 = 0
x1 + 2x2 + x3 - 2x7 - 2x8 = 0
x4 - x5 - x7 + x8 = 0
x1 + x2 + x3 + x4 + x5 + x6 + x7 + x8 = 1
xj Ú 0, j = 1, 2, Á , 8

Note that the solution of this problem automatically yields the optimum solutions of the primal
and dual problems through back-substitution.
Figure 20.16 provides a typical illustration of the solution space in three dimensions with the
homogeneous set AX = 0 consisting only of one equation. By definition, the solution space con-
sisting of the line segment AB lies entirely in the two-dimensional simplex 1X = 1 and passes
through the feasible interior point A 13, 13, 13 B .

Steps of the Algorithm. Karmarkar’s algorithm starts from an interior point represented
by the center of the simplex and then advances in the direction of the projected gradient to
determine a new solution point. The new point must be strictly interior, meaning that all
its coordinates must be positive. The validity of the algorithm rests on this condition.
For the new solution point to be strictly interior, it must not lie on the boundaries of
the simplex. (In terms of Figure 20.16, points A and B must be excluded.) To guarantee
this result, a sphere with its center coinciding with that of the simplex is inscribed tightly
inside the simplex. In the n-dimensional case, the radius r of this sphere equals 1n1n1 - 12.
A smaller sphere with radius ar10 6 a 6 12 will be a subset of the sphere and
any point in the intersection of the smaller sphere with the homogeneous system
AX = 0 will be an interior point, with strictly positive coordinates. Thus, we can move as
far as possible in this restricted space (intersection of AX = 0 and the ar-sphere) along
the projected gradient to determine the new (improved) solution point.
The new solution point no longer will be at the center of the simplex. For the pro-
cedure to be iterative, we need to find a way to bring the new solution point to the cen-
ter of a simplex. Karmarkar satisfies this requirement by proposing the following
intriguing idea, called projective transformation. Let
xi
xki
yi = n x , i = 1, 2, Á , n
a
i
xkj
j=1
CD-30 Chapter 20 Additional Network and LP Algorithms

x3

(0, 0, 1)

Simplex 1X  1

B
Intersection of
Center of
AX  0 and 1X  1
simplex

1,1,1
3 3 3
x1
(1, 0, 0)

(0, 1, 0)

x2
(a) Three dimensions

(0, 0, 0, 1)

Simplex 1X  1

Intersection of
AX  0 and 1X  1
C
Center of
simplex

1,1,1,1
A
4 4 4 4
(1, 0, 0, 0)
B
(0, 0, 1, 0)

(0, 1, 0, 0)
(b) Four dimensions

FIGURE 20.16
Illustrations of the simplex 1X
1
 20.3 Karmarkar Interior-Point Method CD-31

where xki is the ith element of the current solution point X k. The transformation is
n
valid, because all xki 7 0 by design. You will also notice that a i = 1yi = 1, or 1Y = 1,
by definition. This transformation is equivalent to
Dk-1X
Y =
1Dk-1X
where Dk is a diagonal matrix whose ith diagonal elements equal xki. The transforma-
tion maps the X-space onto the Y-space uniquely because we can show that the last
equation yields
DkY
X =
1DkY

By definition, min CX = 0. Because 1DkY is always positive, the original linear program
is equivalent to
Minimize z = CDkY
subject to
ADkY = 0
1Y = 1
Y Ú 0

The transformed problem has the same format as the original problem. We can thus
start with the simplex center Y = A n1 , n1 , Á , n1 B and repeat the iterative step. After each
iteration, we can compute the values of the original X variables from the Y solution.
We show now how the new solution point can be determined for the transformed
problem. At any iteration k, the problem is given by

Minimize z = CDkY
subject to
ADkY = 0
Y lies in the ar-sphere

Because the ar-sphere is a subset of the space of the constraints 1X = 1 and X Ú 0,

these two constraints can be dispensed with. As a result, the optimum solution of the
preceding problem lies along the projection of the the gradient cp (minimization) and
is given as
cp
7cp 7
Ynew = Y0 - ar

1 t
where Y0 = a , , Á , b and cp is the projected gradient, which can be shown to be
1 1
n n n

c p = [I - P T1PP T2-1P]1CDk2T
CD-32 Chapter 20 Additional Network and LP Algorithms

where

P = a b
ADk
1

The selection of a is crucial to enhancing the efficiency of the algorithm. Normally,

we would select a as large as possible to acquire large jumps in the solution. However, if
a is too large, we may come too close to the prohibited boundaries of the simplex. There
- 1
is no general answer to this problem, but Karmarkar suggests the use of a = n 3n .
The steps of Karmarkar’s algorithm are

Step 0. Start with the solution point X 0 = A n1 , n1 , Á , n1 B and compute r = 1

and a = 1n 3n
1n1n - 12
- 12

General step k. Define

Dk = diag5xk1, Á , xkn6

P = a b
ADk
1
and compute
1 T cp
Ynew = a , Á , b - ar
1
n n 7cp 7
DkYnew
Xk + 1 =
1DkYnew
where
cp = [I - P T1PP T2-1P]1cDk2T

Example 20.3-2

Minimize z = 2x1 + 2x2 - 3x3

subject to

-x1 - 2x2 + 3x3 = 0

x1 + x2 + x3 = 1
x1, x2, x3 Ú 0

The problem satisfies the two conditions imposed by the interior-point algorithm—namely,
X = 1x1, x2, x32T = A 13, 13, 13 B T

satisfies both constraints and the optimum solution

X … = 1x…1, x…2, x…32T = 10, .6, .42T
yields z = 0.
 20.3 Karmarkar Interior-Point Method CD-33

Iteration 0
c = 12, 2, -32, A = 1-1, -2, 32
X0 = A 13, 13, 13 B T, z0 = 13, r = 1
16 , A = 2
9
1
3 0 0
1
D0 = £ 0 3 0≥
1
0 0 3

Using projective transformation, we get

Y0 = A 13, 13, 13 B T

Iteration 1

1
0 0
A 23, 23, -1 B
3
cD0 = 12, 2, -32 £ 0 1
3 0≥ =
1
0 0 3
1
0 0
A - 13, - 23, 1 B
3
AD0 = 1-1, -2, 32 £ 0 1
3 0≥ =
1
0 0 3

- 13 1 -1
-1 - 23 9
1PP 2 = £a 3 b £ - 23 1 ≥ ≥ = a 14 1b
T -1 1 0
1 1 1 0 3
1 1

1 0 0 - 13 1 9
- 13 - 23
I - P 1PP 2 P = £ 0 1 ≥ a 14 1ba b
0 1
T T -1
1 0 ≥ - £ - 23
0 3 1 1 1
0 0 1 1 1

25 -20 -5
1
= 42 £ -20 16 4≥
-5 4 1

Thus,
2
25 -20 -5 3 25
cp = 1I - P 1PP 2 P21cD02 = 42 £ -20
T T -1 T 1
16 4 ≥ £ 23 ≥ = 1
126 £ -20 ≥
-5 4 1 -1 -5

It then follows that

252 + 1-2022 + 1-522
7cp 7 = = .257172
B 1262
Thus,

Ynew = A 13, 13, 13 B T - 2

9 * 1
16 * 1
.257171 * 126 125,
1
-20, -52T

= 1.263340, .389328, .3473322T

CD-34 Chapter 20 Additional Network and LP Algorithms

Next,

1D0Ynew = 1311, 1, 121.263340, .389328, .3473322T = 1

3

Now,

1
3 Ynew
= Ynew = 1.263340, .389328, .3473322T
D0Ynew
X1 = = 1
1D0Ynew
3

z1 = .26334

Iteration 2

.263340 0 0
cD1 = 12, 2, -32 £ 0 .389328 0 ≥ = 1.526680, .778656, -1.0419962
0 0 .347332

.263340 0 0
AD1 = 1-1, -2, 32 £ 0 .389328 0 ≥
0 0 .347332

= 1-.263340, -.778656, 1.0419962

-.263340 1 -1

1PP 2 = £a b £ -.778656
-.26334 -.778656 1.041996
T -1
1≥ ≥
1 1 1
1.041996 1

= a b
.567727 0
0 .333333

I - P T1PP T2-1P

1 0 0 -.263340 1
1≥ a b
.567727 0
= £0 1 0 ≥ - £ -.778656
0 .333333
0 0 1 1.041996 1

a b
-.263340 -.778656 1.041996
1 1 1

.627296 -.449746 -.177550

= £ -.449746 .322451 .127295 ≥
-.177550 .127295 .050254
 20.3 Karmarkar Interior-Point Method CD-35

Thus,

.627296 -.449746 -.177550

cp = 1I - P T1PP T2-1P21cD12T = £ -.449746 .322451 .127295 ≥
-.177550 .127295 .050254

.526680
£ .778656 ≥
-1.041996

.165193
= £ -.118435 ≥
-.046757

It then follows that

7cp 7 = 2.1651932 + 1-.11843522 + 1-.04675722 = .208571

Thus,

Ynew = A 13, 13, 13 B T - 2

9 * 1
16 * .208571 1.165193,
1
-.118435, -.0467572T
= 1.261479, .384849, .3536712T

Next,

.263340 0 0 .261479 .068858

D1Ynew = £ 0 .389328 0 ≥ £ .384849 ≥ = £ .149832 ≥
0 0 .347332 .353671 .122841

1D1Ynew = .341531

Now,

= 1.201616
D1Ynew
X2 = .438707 .3596772T
1D1Ynew
z2 = .201615

Repeated application of the algorithm will move the solution closer to the optimum point
(0, .6, .4). Karmarkar does provide an additional step for rounding the optimal solution to the op-
timum extreme point.

PROBLEM SET 20.3A

1. Use TORA to show that the solution of the transformed LP given at the end of Example
20.3-1 does yield the optimal primal and dual solutions of the parent problem. (Hint: Use
M = 10 and make sure that TORA’s output gives at least 5 decimal points accuracy.)
CD-36 Chapter 20 Additional Network and LP Algorithms

2. Transform the following LP to Karmarkar’s format.

Maximize z = y1 + 2y2

subject to
y1 - y2 … 2
2y1 + y2 … 4
y1, y2 Ú 0

3. Carry out one additional iteration in Example 20.3-3, and show that the solution is mov-
ing toward the optimum z = 0.
4. Carry out two iterations of Karmarkar’s algorithm for the following linear program:
Maximize z = 2x1 + x2

subject to
x1 + x2 … 4
x1, x2 Ú 0

(Hint: The problem must be converted to Karmarkar format first.)

5. Carry out two iterations of Karmarkar’s algorithm for the following problem:
Maximize z = 4x1 + x3 + x4

subject to
-2x1 + 2x2 + x3 - x4 = 0
x1 + x2 + x3 + x4 = 1
x1, x2, x3, x4 Ú 0

(Hint: The problem must be converted to Karmarkar format first.)

REFERENCES
Ahuja, R., T. Magnati, and J. Orlin, Network Flows: Theory, Algorithms, And Applications, Pren-
tice Hall, Upper Saddle River, NJ, 1993.
Bazaraa, M., J. Jarvis, and H. Sherali, Linear Programming and Network Flow, 2nd ed., Wiley,
New York, 1990.
Charnes, A. and W. Cooper, “Some Network Characterization for Mathematical Programming
and Accounting Applications to Planning and Control,” The Accounting Review, Vol. 42,
No. 3, pp. 24–52, 1967.
Hooker, J., “Karmarkar’s Linear Programming Algorithm,” Interfaces, Vol. 16, No. 4, pp. 75–90,
1986.
Lasdon, L., Optimization for Large Systems, Macmillan, New York, 1970.