Professional Documents
Culture Documents
mathrm (C) /mathrm (I) /mathrm (O) /mathrm (M) /mathrm (P) /mathrm (U) /mathrm (T)
mathrm (C) /mathrm (I) /mathrm (O) /mathrm (M) /mathrm (P) /mathrm (U) /mathrm (T)
Abstract. In this work, we construct a novel numerical method for solving the multimarginal
optimal transport problems with Coulomb cost. This type of optimal transport problem arises in
quantum physics and plays an important role in understanding the strongly correlated quantum sys-
tems. With a Monge-like ansatz, we transfer the original high-dimensional problems into mathemati-
cal programmings with generalized complementarity constraints, and thus the curse of dimensionality
is surmounted. However, the latter ones are themselves hard to deal with from both theoretical and
practical perspectives. Moreover, in the presence of nonconvexity, brute-force searching for global
solutions becomes prohibitive as the problem size grows large. To this end, we propose a global
optimization approach for solving the nonconvex optimization problems, by exploiting an efficient
proximal block coordinate descent local solver and an initialization subroutine based on hierarchical
grid refinements. We conduct numerical simulations on some typical physical systems to show the
efficiency of our approach. The results match well with both theoretical predictions and physical
intuitions and provide indications for Monge solutions in two-dimensional contexts. In addition, we
give the first visualization of approximate optimal transport maps for some two-dimensional systems.
Key words. multimarginal optimal transport, Coulomb cost, Monge-like ansatz, mathematical
programming with generalized complementarity constraints, global optimization, grid refinement,
optimal transport maps
DOI. 10.1137/21M1455164
where \scrP (\Omega N ) denotes the space of all N -point probability measures over \Omega N , the cost
function c(r1 , . . . , rN ) is determined by the electron-electron Coulomb interaction
* Submitted
to the journal's Methods and Algorithms for Scientific Computing section October
25, 2021; accepted for publication (in revised form) December 20, 2022; published electronically June
12, 2023.
https://doi.org/10.1137/21M1455164
Funding: The work of the second author was supported by the National Natural Science Foun-
dation of China (11971066). The work of the third author was supported in part by the National
Natural Science Foundation of China (1212500491, 11971466, 11991021) and the Key Research Pro-
gram of Frontier Sciences, Chinese Academy of Sciences (ZDBS-LY-7022).
\dagger
State Key Laboratory of Scientific and Engineering Computing, Academy of Mathematics and
Systems Science, Chinese Academy of Sciences, and University of Chinese Academy of Sciences,
Beijing, China (ykhu@lsec.cc.ac.cn).
\ddagger
School of Mathematical Sciences, Beijing Normal University, Beijing, China (chen.huajie@
bnu.edu.cn).
\S
Corresponding author. State Key Laboratory of Scientific and Engineering Computing, Acad-
emy of Mathematics and Systems Science, Chinese Academy of Sciences, and University of Chinese
Academy of Sciences, Beijing, China (liuxin@lsec.cc.ac.cn).
A1214
\int
\rho \in L1 (\Omega ) refers to the single-electron density satisfying \Omega \rho = N , and \Gamma \mapsto \rightarrow \rho repre-
sents the marginal constraints: for i = 1, . . . , N and any open set \scrA i \subseteq \Omega ,
\int \int
\bigl( \bigr) 1
(1.3) d\Gamma r1 , . . . , rN = \rho (r) dr.
\Omega i - 1 \times \scrA i \times \Omega N - i \scrA i N
Note that the Coulomb interaction 1/| ri - rj | between the electrons in (1.2) can be
approximated or regularized, especially in the simulations of systems with d < 3 [4,
20]. Nevertheless, the approach constructed in this paper will make no difference
as long as the interaction between the electrons is repulsive (i.e., the cost decreases
with respect to | ri - rj | ). Without loss of generality, we will focus on the Coulomb
interaction of the form (1.2).
The MMOT problem (1.1) arises as the strictly correlated electrons (SCE) limit
in the density functional theory (DFT). The DFT has been most widely used for
electronic structure calculations in physics, chemistry, and materials science (see [3]
for a review). It depends on choosing an ansatz for an exact yet unknown density
functional. The SCE limit was first introduced in [45]. Later in [7, 12], it was
recognized that the limit is an MMOT problem. The SCE limit provides an alternative
route to derive the DFT energy functionals and has been exploited to extend the
capability of the DFT to treat strongly correlated quantum systems [9, 10, 23, 35, 37].
Direct discretization of the MMOT problem (1.1) leads to a linear programming,
with the size increasing exponentially fast with respect to N (the number of elec-
trons/marginals). There are several works devoted to the reformulations and nu-
merical methods for the MMOT problem (1.1). In [5], the Sinkhorn scaling algo-
rithm based on iterative Bregman projections was applied to an entropy-regularized
discretized MMOT problem of one-dimensional (1D) systems. In [36], the authors
proposed numerical methods based on the Kantorovich dual of the MMOT problem,
penalizing the nonsmooth reformulation of the original inequality constraints and
utilizing derivative-free methods. In [31, 32], a convex relaxation for the so-called N -
representability formulation was proposed by imposing certain necessary constraints
satisfied by the two-marginal, and the relaxed problem was then solved as a semi-
definite programming to obtain tight lower bounds for the optimal cost. In [1, 2],
the existence of sparse global solutions to semidiscrete formulation was established
and a constrained overdamped Langevin process was proposed to solve the moment
constrained relaxations. In [20, 21], the sparse extremal representations for global
solutions were rigorously justified, and an efficient numerical method was proposed
based on column generation and machine learning.
The starting point of this work is to approximate the N -point measure \Gamma by the
ansatz
\rho (r1 )
(1.4) d\Gamma (r1 , . . . , rN ) = \gamma 2 (r1 , r2 ) \cdot \cdot \cdot \gamma N (r1 , rN ) dr1 \cdot \cdot \cdot drN ,
N
where, for any n \in \{ 2, . . . , N \} , \gamma n \in L1 (\Omega 2 ) is a transport plan fulfilling
\int \int
(1.5) \gamma n (r, r\prime ) \geq 0, \gamma n (r, r\prime ) dr\prime = 1, and \rho (r)\gamma n (r, r\prime ) dr = \rho (r\prime ) \forall r, r\prime \in \Omega .
\Omega \Omega
The condition (1.5) is derived from the marginal constraints (1.3). We sometimes call
the last equality in (1.5) the mass-preserving constraint. From a physical point of
view, \gamma n (r, r\prime ) represents the correlation between the first and the nth electron, which
gives the probability density of finding the nth electron at r\prime while the first electron
is located at r. Under the ansatz (1.4), the MMOT problem (1.1) (with N > 2) can
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
be rewritten as
\left\{
\sum \int \int \int
\rho (r)\gamma m (r, r\prime )\gamma n (r, r\prime \prime )
(1.6) min dr dr\prime dr\prime \prime
\gamma 2 ,...,\gamma N \Omega \Omega \Omega | r\prime - r\prime \prime |
2\leq m<n\leq N
\right\}
\sum \int \int \rho (r)\gamma n (r, r\prime )
+ \prime dr dr\prime : \gamma 2 , . . . , \gamma N satisfy (1.5) .
\Omega \Omega | r - r |
2\leq n\leq N
We mention that in the case of N = 2, the first term in the objective of (1.6) vanishes;
(1.6) then reduces to a linear programming and can be solved by standard algorithms
[10]. In this work, we focus our attention on the N \geq 3 settings. The formulation
(1.6) amounts to a spectacular dimension reduction, in that the unknowns are N - 1
transport plans on \Omega 2 instead of the N -point measure \Gamma on \Omega N . Therefore, the
degrees of freedom now scale linearly with respect to N rather than exponentially
fast. Moreover, the ansatz (1.4) is related to the Monge state [38, 47] by taking
\gamma n (r, r\prime ) = \delta (r\prime - Tn (r)) with \delta being the Dirac measure and Tn (n \in \{ 2, . . . , N \} )
being the so-called optimal transport map. The Monge formulation gives significant
information on the MMOT problem and enjoys physical interpretations; see more
discussions in subsection 1.3.
In practical calculations, we need to discretize (1.6) into some finite-dimensional
problem. The discretization consists of three steps. First, we employ a finite elements--
like mesh \scrT = \{ ek \} K k=1 to partition the domain \Omega into K nonoverlapping elements,
i.e., \cup K k=1 k e = \Omega and ek \cap ek\prime = \emptyset when k \not = k \prime . Let e := [| e1 | , . . . , | eK | ]\top \in \BbbR K + de-
note the volumes of elements. Second, we approximate the \int marginal \rho by a vector
1
\bfitvarrho := [\varrho 1 , . . . , \varrho K ]\top \in \BbbR K
+ , where the kth entry \varrho k := | ek | ek \rho (r) dr gives the mar-
ginal/electron mass on the kth element ek . Finally, the Coulomb interactions and
transport plans \gamma n (n = 2, . . . , N ) can be approximated by the effective interactions
and transports between elements, i.e., for any i, j \in \{ 1, . . . , K\} ,
(1.7) \int \int \int \int
1 1 \prime 1
cij := dr dr and x n,ij := \gamma n (r, r\prime ) dr dr\prime ,
| ei | \cdot | ej | ej ei | r - r\prime | | ei | \cdot | ej | ej ei
respectively, leading to K \times K matrices C := ((1 - aij )cij )ij and Xn := (xn,ij )ij
(n = 2, . . . , N ). Here, aij equals 1 if i = j and 0 otherwise. With a slight abuse of
terminology, we also call Xn (n = 2, . . . , N ) transport plans in what follows. After
discretization, we can approximate (1.6) using the following optimization problem
with unknowns \{ Xn \} N n=2 :
(1.8) \sum \sum
min f (X2 , . . . , XN ) := \langle Xn , \Lambda E CE\rangle + \langle Xn , E\Lambda Xm ECE\rangle
X2 ,...,XN
2\leq n\leq N 2\leq m<n\leq N
s.t. Xn e = 1, Xn\top E\bfitvarrho = \bfitvarrho , Tr(Xn ) = 0, Xn \geq 0, n = 2, . . . , N,
\langle Xm , Xn \rangle = 0, \forall m \not = n,
where 1 is the all-ones vector in \BbbR K , \Lambda := Diag(\bfitvarrho ) and E := Diag(e) are K \times K
diagonal matrices formed by the entries in \bfitvarrho and e, respectively. More detailed
derivation of (1.8) is given in Appendix A. Note that the diagonal elements in matrix
C are removed due to the integral divergence in (1.7). The extra constraints
are hence accordingly added. From a physical point of view, these constraints keep
the electrons spatially away from each other in the case of Coulomb repulsion so that
unfavorable particle clustering can be avoided.
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
quickly as the meshes become finer, state-of-the-art global optimization solvers cannot
be our last resort.
1.1. Optimization background. Although little is known about MPGCC,
there exists a rich literature on MPCC. To overcome the intrinsic difficulties men-
tioned above, several tailored constraint qualifications have been provided for MPCC.
Under these constraint qualifications, points satisfying certain stationary systems are
shown to be the proper candidates of local minimizers. The related notions and
theoretical results are gathered in [41, 51] and the references within.
With these in place, researchers have proposed various numerical approaches,
wherein those based on the original MPCC formulation rank top choices; they em-
ploy modified nonlinear programming solvers. For example, the authors in [17] solved
MPCCs using sequential quadratic programming algorithms with filter techniques
[16]. The software introduced in [8, 49] incorporates a suite of nonlinear program-
ming algorithms to tackle MPCCs, including interior-point methods and sequential
quadratic programming algorithms, together with globalization techniques such as
line search and trust region.
Owing to the troubles when coping with complementarity constraints, methods
based on penalty functions also have gained popularity. Among others, we confine our
attention to the \ell 1 (complementarity) penalty function, which favors direct extension
to MPGCC (1.8) as
\sum
(1.9) f (X2 , . . . , XN ) + \beta \langle Xm , Xn \rangle ,
m<n
namely, penalizing merely the complementarity violation in \ell 1 form. Here, f is the
repulsive energy defined in (1.8), and \beta > 0 is the penalty parameter. It can be verified
under certain conditions that the global solutions of (1.8) coincide with those globally
minimizing (1.9) over \scrS N - 1 , where
(1.10) \scrS := \{ W \in \BbbR K\times K : W e = 1, W \top E\bfitvarrho = \bfitvarrho , Tr(W ) = 0, W \geq 0\} .
A direct consequence is that, if the global solutions of (1.8) are required, one can in
turn minimize (1.9) over \scrS N - 1 starting with proper initializations. However, we are
not aware of any existing method that fully exploits the special structure of (1.9). A
customized algorithm is thus needed, particularly in the large-scale contexts.
pared with methods using modified nonlinear programming solvers or penalty func-
tions, these approaches require solving a sequence of subproblems in the same size to
stationarity or even optimality [30]. This weakness excludes them from our choices,
particularly when the number of grid points K is tremendously large.
\rho (r1 )
(1.11) d\Gamma (r1 , . . . , rN ) = \delta (r2 - T2 (r1 )) \cdot \cdot \cdot \delta (rN - TN (r1 )) dr1 \cdot \cdot \cdot drN ,
N
where the transport map Tn : \Omega \rightarrow \Omega (n \in \{ 2, . . . , N \} ) (we can prescribe T1 (r) = r
for the completeness of notation) preserves the single-electron density \rho . The Monge
solution has a simple physical interpretation: the many-electron repulsive energy is
minimized at a state such that one electron at position r can determine the positions
of all other N - 1 electrons via \{ Tn \} N n=2 . It is known that for 1D systems or systems
with N = 2 electrons, the Monge ansatz (1.11) accommodates the global solutions
of the MMOT problems [11, 12]. But in the general d > 1 and N > 2 cases, it
is unknown whether there exists a minimizer of (1.1) in the form (1.11). Thus far,
a counterexample in real physical scenarios has not been put forward; see [19] and
the references within. It is therefore of scientific interest to search for the Monge
solutions. The ansatz (1.4) is more flexible than (1.11), in that the transport maps
are replaced with transport plans. Unlike (1.11), (1.4) guides one to the problem (1.6)
that can be easily discretized and always admits a minimizer, with which we are able
to approximate the Monge solutions. More precisely, let aj \in \Omega be the barycenter
\sum
(1.12) TnK (aj ) := ak xn,jk | ek | , j = 1, . . . , K, n = 2, . . . , N.
1\leq k\leq K
Although we do not have this symmetric restriction in the MMOT problem (1.1) and
the ansatz (1.11) is in general not symmetric, dropping the restriction does not alter
the minimum value. This is because we have a symmetric cost function c in (1.2) and
equal marginal for any i \in \{ 1, . . . , N \} in (1.3). Hence each nonsymmetric \Gamma can give a
symmetric one with the same energy value by symmetrization. Consequently, we do
not have to impose the symmetric constraints in the optimization formulation (1.8).
Discretization. Most of the existing works discretize the MMOT problems
with real space methods [5, 10]. Particularly, this paper discretizes (1.6) into (1.8)
by representing the marginal \rho with piecewise finite elements and using effective cost
coefficients obtained by integrating the continuous cost functions with respect to these
elements. To further reduce the computational cost (i.e., use fewer grid points where
the marginal is small), we choose the elements adaptively such that each element
carries approximately the same marginal mass.
1.4. Outline. The rest of this paper is organized as follows. We introduce the
global optimization approach in section 2, where the initialization subroutine (subsec-
tion 2.1) and local solver (subsection 2.2) are detailed in order. Section 3 is dedicated
to the rough statements of the convergence properties of PBCD. We corroborate the
proposed approach with numerical simulations on several typical systems in section 4.
Finally, conclusions and discussions appear in section 5.
1.5. Notation. The adjoint and image of a linear operator A are denoted by
A \ast and Im(A ), respectively. The notation \| X\| p gives the p-norm of matrix X, while
\| X\| F yields its Frobenius norm. The components of matrices or vectors are indicated
by subscripts, e.g., xij . The inequality X \geq 0 means xij \geq 0 for any i, j.
For the multiblock objective functions in this work (such as (1.8)), we occasionally
adopt abbreviations in parentheses. For example, f (X<n , Xn , X>n ) means
f (X2 , . . . , Xn - 1 , Xn , Xn+1 , . . . , XN );
abbreviations like X<n , X(m,n) , and X>n represent the aggregation of blocks with
certain subscripts (clearly, X<0 , X(n,n) , and X>n are null variable blocks, which may
be used for notational ease). For any n, we use \nabla n f to refer to the gradient of f with
respect to the nth block variable.
Regarding the algorithm, we use double superscripts within brackets for iterates
(\ell ,k)
in the inner loop; for instance, Xn is the iterate in the kth inner iteration of the
\ell th outer iteration.
(0) (0)
solver; GR subroutine; initial mesh with K (0) elements \{ ek \} K k=1 .
1: Set \ell := 0.
(0, \star )
2: GGR Init: use the global solver for (2.1) with size K (0) and get (Xn )N n=2 .
3: while certain stopping criteria are not satisfied do
(\ell ) (\ell ) (\ell +1) K (\ell +1)
4: Refine the last mesh \{ ek \} K k=1 to \{ ek \} k=1 with K (\ell +1) elements.
(\ell , \star ) N (\ell +1,0) N
5: Modify (Xn )n=2 using the GR subroutine to obtain (Xn )n=2 .
(\ell +1,0) N
6: GGR LS(\ell + 1): start the local solver from (Xn )n=2 for (2.1) with size
(\ell +1, \star ) N
K (\ell +1) and get (Xn )n=2 .
7: Set \ell := \ell + 1.
8: end while
(\ell , \star ) K (\ell ) \times K (\ell ) N - 1
9: return (Xn )N n=2 \in (\BbbR ) .
where f is the repulsive energy defined in (1.8) and \scrS is defined in (1.10). The
problem (2.1) is a nonconvex quadratic programming problem, still NP-hard [39]. In
what follows, when we talk about (2.1) and its solution in space (\BbbR K\times K )N - 1 , we
simply say (2.1) and its solution with size K.
For practical purposes, a global solution of (2.1) is always demanded. Meanwhile,
we notice that the degrees of freedom in (2.1), (N - 1)K 2 , grow fast with respect to K.
This prevents us from brute-force solving (2.1) by state-of-the-art global optimization
methods (e.g., branch-and-bound and cutting plane algorithms) due to exponentially
increasing running time.
Motivated by [5], we propose a global optimization approach, GGR; see Frame-
work 2.1. Here, ``G"" and ``GR"" stand for global optimization and initialization based
on hierarchical grid refinements, respectively. GGR Init refers to the initial step
invoking a global solver and GGR LS to the subsequent step invoking a local solver.
Framework 2.1 progresses step by step along with the process of mesh refinements.
Let us first justify the usage of a global solver in the initial step (line 2 in Frame-
work 2.1). From the point of applicability, given initial size K (0) of moderate magni-
tude, globally solving (2.1) is amenable to state-of-the-art global optimization meth-
ods. Considering the necessity, the qualities of the constructed initial points largely
depend on the solutions in the previous step. Hence it is a natural choice for us to
invoke a global solver in the initial step. For our choices in implementation, please
refer to subsection 4.1.
Without specification, the mesh refinements (line 4 in Framework 2.1) are done
such that the coarse meshes are always embedded into the refined meshes. For more
remarks, see subsection 1.3. Although the refinements are uniform in the numerical
simulations of present work (subsections 4.2 and 4.3), practical implementations focus
on the region where marginals vary violently. Nevertheless, in the latter contexts, our
GGR approach still works.
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
3 4 doubly refine 5 6 7 8
Mesh: −−−−−−−−→
m m
4 7 8
5
Plan: 3
6
Fig. 2.1. 1D case. The red block means there is mass transported from 3 to 4. Then in a doubly
refined mesh, there is mass transported from 5 and 6 to 7 and 8, as marked out by 4 blue blocks.
(1,9)(1,10)
(1,5)
(2,9)(2,10)
doubly refine
Mesh: −−−−−−−−→
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
(3,7)(3,8)
(2,4)
(4,7)(4,8)
m m
35 36 49 50
11
9 ···
10 ···
5
Plan: .. .. .. ..
. . . .
23 ···
24 ···
Fig. 2.2. 2D case (7\times 7 rectangular mesh). The red block means there is mass transported from
(1,5) to (2,4). Then in a doubly refined mesh, there is mass transported from (1,9), (1,10), (2,9) and
(2,10) to (3,7), (3,8), (4,7) and (4,8), as marked out by 16 blue blocks.
\~
After mesh refinement, the original \{ ek \} K k=1 becomes \{ \~ ek \} K
k=1 ; for each k, the original
sk
element ek is divided into sk parts: ek = \cup t=1 e\~kt and e\~kt1 \cap e\~kt2 = \emptyset when kt1 \not = kt2 .
It is reasonable to speculate that there also exists certain mass transported from e\~ju
to e\~jv\prime , where u \in \{ 1, . . . , sj \} , v \in \{ 1, . . . , sj \prime \} . Accordingly in X \~ n , there should be
\~n,ju jv > 0, sj \times sj positive entries in total. We illustrate the 2D case in Figure 2.2.
x \prime \prime
Note that the coordinates in the transport plan are rearranged from the 2D coordi-
nates in mesh.
Based upon the above arguments, we derive the GR subroutine for initializations;
see Algorithm 2.2.
We shall mention that our strategy is completely different from the so-called
shielding neighborhood in the context of standard optimal transport problems [5, 22,
42]. The shielding neighborhood strategy adjusts the supports of transport plans
adaptively by using the strong duality of linear programming and restricts the re-
finement of plans on the adjusted supports. The strategy ensures the optimality of
the refined solutions without increasing too much computational and storage com-
plexities. Compared with the standard optimal transport problems, however, (2.1) is
nonconvex, which renders finding appropriate supports impracticable. Therefore, we
instead keep all the degrees of freedom and concentrate on constructing high-quality
initial points for the local solver.
2.2. Local solver. The first-step global optimization and GR subroutine waive
the need of brute-force globally solving large-scale (2.1). Instead, we only need to
devise a local solver (see line 6 in Framework 2.1). We assume that the procedure is in
the \ell th iteration of Framework 2.1. This is the same in what follows whenever talking
(\ell ) (\ell ) (\ell )
about the local solver. We define the linear operator B (\ell ) : \BbbR K \times K \rightarrow \BbbR 2K +1 as
(\ell )
\times K (\ell )
B (\ell ) (W ) := [e(\ell )\top W \top \bfitvarrho (\ell )\top E (\ell ) W Tr(W )]\top \forall W \in \BbbR K ,
\~
K N
\{ \~
ek \} k=1 ; approximate solution from the previous step (Xn )n=2 ; scaling factor
r > 0.
1: for n = 2, . . . , N do
2: for j = 1, . . . , K do
3: for j \prime = 1, . . . , K do
4: if xn,jj \prime > 0 then
sj
5: Find e\~ju , u = 1, . . . , sj , such that ej = \cup u=1 e\~ju .
sj \prime
6: Find e\~jv , v = 1, . . . , sj , such that ej = \cup v=1 e\~jv\prime .
\prime \prime \prime
7: Set x \~n,ju jv\prime = r \cdot xn,jj \prime for u \in \{ 1, . . . , sj \} and v \in \{ 1, . . . , sj \prime \} .
8: end if
9: end for
10: end for
11: end for
\~ n )N \in (\BbbR K\times \~ K \~ N - 1 \~ n = (\~
12: return (X n=2 ) , where X xn,jj \prime )jj \prime (n = 2, . . . , N ).
4: Set k := k + 1.
5: end while
(\ell , \star ) (\ell ,k) N (\ell ) (\ell )
6: return (Xn )N n=2 := (Xn )n=2 \in (\BbbR K \times K )N - 1 .
(\ell )
and b(\ell ) := [1\top \bfitvarrho (\ell ) 0]\top \in \BbbR 2K +1 . The feasible set in (1.10) can then be rewritten
as \scrS (\ell ) = \{ W : B (\ell ) (W ) = b(\ell ) , W \geq 0\} .
The block structure of (2.1) reminds us of using splitting-type methods. One
natural choice is an (N - 1)-block cyclic PBCD method; see Algorithm 2.3. In PBCD,
the nth block problem merely depends on the nth block variable Xn , while keeping
(\ell ,k)
other block variables their latest values. Moreover, the proximal term \| Xn - Xn \| 2F
is added to the objective function such that the block problem admits a unique global
solution, with \sigma (\ell ) > 0 being the proximal parameter. PBCD invokes certain subsolver
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
for the block problem (2.2) until the inexact criterion (2.3) is met. In (2.3), r(\ell ) acts as
a residual function measuring the violation of the KKT conditions of (2.2), defined as1
r(\ell ) (Xn ,\bfitlambda n , \Phi n , X \~ n ) := \| Xn - X\~ n - B (\ell )\ast (\bfitlambda n ) - \Phi n \| \infty
\Bigl\{ \Bigl\langle \Bigr\rangle \Bigr\}
+ max Xn , Xn - X \~ n - B (\ell )\ast (\bfitlambda n ) - \Phi n , 0
\Bigl\{ \Bigl\langle \Bigr\rangle \Bigr\}
+ max \bfitlambda n , B (\ell ) (Xn ) - b(\ell ) , 0 + max \{ \langle \Phi n , Xn \rangle , 0\}
+ \| B (\ell ) (Xn ) - b(\ell ) \| \infty + \| max \{ - Xn , 0\} \| \infty ;
Remark 3.3. Theorem 3.2 itself is of particular theoretical interest, in that the
iterates are allowed to be infeasible. This has not been covered by existing works
on PBCD (e.g., [50]) and should be credited to the inexact criterion (2.3). This also
provides a theoretical guarantee for the usage of efficient infeasible subsolvers for (2.2),
which is of significant importance in our context because the number of variables far
exceeds that of equality constraints.
4. Numerical experiments. In this section, we validate the proposed GGR
approach via numerical simulations on several typical systems, including both 1D and
2D systems. During the experiments, we mainly monitor the repulsive energy f in
(1.8). We also calculate the approximate transport maps \{ TnK \} N
n=2 as in (1.12), and
evaluate the qualities of solutions through the average error (denoted by err)
K N
1 \sum \sum \bigm| \bigm| \bigm|
err(K, \Omega ) := Tn (ak ) - TnK (ak )\bigm|
K | \Omega | n=2 k=1
if the optimal transport maps \{ Tn \} N n=2 in (1.11) are already available. Moreover,
we approximate the SCE potentials, which is crucial in the applications of electronic
\~ K := \bfitlambda K - minK \{ \lambda K \} \cdot 1 \in \BbbR K . Here,
structure calculations [46, 47], with \bfitlambda j=1 j
N
1 \sum K
(4.1) \bfitlambda K := \bfitlambda \in \BbbR K
N - 1 n=2 n,2
Table 4.1
Values of \beta for different K.
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
K (0, 10) [10, 36) [36, 80) [80, 160) [160, 320)
\beta 22 21 20 2 - 2 2 - 3
K [320, 640) [640, 1280) [1280, 2560) [2560, 5120) [5120, \infty )
\beta 2 - 4 2 - 5 2 - 6 2 - 7 2 - 8
and then performing line searches along d for a sufficient reduction on the dual ob-
jective. Here, the third superscript (j) indicates the iteration of SSNCG, V (\ell ,k,j) \in
(\ell ) (\ell ) (\ell )
\BbbR (2K +1)\times (2K +1) is a positive semidefinite matrix, s(\ell ,k,j) \in \BbbR 2K +1 is the resid-
ual vector, and \tau (\ell ,k,j) > 0. In our context, the linear system can be solved quickly
to desired accuracy by the preconditioned conjugate gradient method equipped with
block Jacobi preconditioner.
Parameter settings. In the GR subroutine, we set the scaling factor r = 1/2d.
For any \ell , we fix \sigma (\ell ) \equiv 10 - 3 in PBCD. For different K, we choose \beta according to Table
4.1. We start SSNCG from origin in the first call; after that, we perform warm start
for acceleration.
Stopping criteria. In SSNCG, we fix \varepsilon (\ell ,k) \equiv 10 - 9 in (2.3) and terminate the
4
algorithm once the subiteration number \surd reaches 10 . We stop PBCD when the scaled
difference of two consecutive iterates \sigma \| X (\ell ) (\ell ,k+1)
- X (\ell ,k) \| F is less than 10 - 4 , or
when the absolute value of the difference between two consecutive energies is less than
10 - 8 , or once the iteration number reaches 106 .
6 3.5 3
6 3.5 3
3
5 2.5
3
5 2.5 2.5
4 2
2.5
2 2
4
3 1.5
2
1.5
3 1.5
2 1.5 1
1
2 1
1 1 0.5
0.5
1 0.5
0.5
0 0 0
-1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1 1.5 -5 0 5
0 0 0
-1 -0.5 0 0.5 1 -1.5 -1 -0.5 0 0.5 1 1.5 -5 0 5
Fig. 4.1. Approximate SCE potentials (blue lines) and ground truths (red lines) for 1D systems
with NFig.
= 3.4.1. Approximate SCE potentials (blue lines) and ground truths (red lines) for 1D systems
with N = 3.
2.5
3.0
1.5
2.5 2.0
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
2.0 1.5
1.0
1.5
1.0
1.0
0.5
0.5
0.5
Fig. 4.2. Marginals (the first row) and approximate transport maps (the remaining four rows)
in 1D systems with N = 3; left to right: systems with \rho 1 , \rho 2 , \rho 3 . The maps in the last four rows
correspond to the rows in Table 4.2(a), where K = 12, 48, 192, 768. The blue and red dots are the
images of T2K and T3K over grid barycenters, respectively.
Starting from the initial meshes, we have performed uniform mesh refinements and
invoked PBCD six times. The output energies and calculated average errors (the ``err e""
\~ K (4.1) as
columns) are collected in Table 4.2(b). The approximate SCE potentials \bfitlambda
well as the ground truths are depicted in Figure 4.3. We show the single-electron
densities (marginals) and approximate transport maps \{ TnK \} N
n=2 (1.12) in Figure 4.4.
\~ K given by our
We observe from the numerical results that the iterates as well as \bfitlambda
GGR approach converge well to the correct solutions and the true SCE potentials.
12 14 6
12
10 5
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
10
8 4
8
6 3
6
4 2
4
2 1
2
0 0 0
-3 -2 -1 0 1 2 3 -4 -3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2 3
Fig. 4.3. Approximate SCE potentials (blue lines) and ground truths (red lines) for 1D systems
with N = 7.
3.5 2.5
4.0
3.0
2.0
2.5 3.0
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
2.0 1.5
2.0
1.5 1.0
1.0
1.0
0.5
0.5
Fig. 4.4. Marginals (the first row) and approximate transport maps (the remaining four rows)
in 1D systems with N = 7; left to right: systems with \rho 4 , \rho 5 , \rho 6 . The maps in the last four rows
correspond to the rows in Table 4.2(b) where K = 14, 56, 224, 896. The blue, red, black, green, brown,
and purple dots are the images of TnK , n = 2, . . . , 7, over grid barycenters, respectively.
We gather the output energies of the GGR approach in Table 4.4, where errors
are absent because no explicit solutions of the original MMOT problems are known in
these contexts. The evolution of the approximate SCE potentials \bfitlambda \~ K (4.1) is shown
in Figure 4.6. The convergence of our GGR approach can be well observed from these
results. Moreover, to show the transport maps T2K , T3K (1.12) approximated by the
obtained solutions, we plot in the remaining three rows of Figure 4.5 the images of
the barycenters of triangular elements within some subregions \omega \subseteq \Omega under T2K and
trials.
The results given by random initializations are the average of 10 trials.
ρ = ρ1 ρ = ρ2 ρ = ρ3
Initialization
Energy Time Energy Time Energy Time
GR 19.022 197.20 11.373 209.30 6.361 98.55
Random 19.037 823.98 11.379 716.02 6.396 538.23
For any n \in \{ 2, . . . , N \} and i \in \{ 1, . . . , K\} , the cardinality nnzn,i corresponds to the
number of sites to which the ith piece of mass is transported by the nth plan and
dn,i measures the total distance between each two of these sites. We draw the fre-
quency percentage distributions of \{ nnzn,i \} n,i and \{ dn,i \} n,i using the built-in function
``hist3"" of matlab; see Figure 4.7. One can observe that there are dominant values
in a few columns for most rows and the distances are short. By the definition of nnzn,i
and dn,i , Figure 4.7 shows the localization of mass transportation. As K grows, the
mesh becomes more refined and the sparsity of solutions becomes more evident.
4.4. Scaling of the GGR approach. We finally investigate the scaling of GGR
numerically on the eight systems in the previous two subsections; see Figures 4.8 and
4.9.3 Leveraging the sparsity of iterates, PBCD entails O(K) expenditure for solving
(2.1) in each step of GGR, which leads to O(K 2 ) cost in total for GGR.
5. Conclusions. In the present work, we consider the MMOT problem with
Coulomb cost arising in quantum physics. The Monge-like ansatz tides us over the
curse of dimensionality, in that the number of unknowns scales linearly with respect to
the number of electrons, however resulting in MPGCC. In quest for global solutions,
3 The time consumed for the initial global optimization is not included.
Fig. 4.5. Contours of marginals (the first row), initial meshes (the second row), and slices of
approximate transport maps (the third to fifth rows) in 2D systems; left to right: systems with \rho 7 , \rho 8 .
In system 7--8, we calculated K to 15360 and 10880, respectively. The gray, blue, and green circles
are preimages \omega \subseteq \Omega , T2K (\omega ), and T3K (\omega ), respectively.
we propose a global optimization approach GGR for dealing with the derived MPGCC.
The GGR approach solves the problem step by step along with the process of mesh
refinements and is equipped with an initialization subroutine such that global solutions
are amenable to the proposed local solver PBCD. The convergence properties of PBCD
are established in the presence of iterate infeasibility. We corroborate the merits of
the GGR approach with numerical simulations on several typical 1D and 2D physical
systems. Notably, we obtain solutions with high resolution in the 1D cases, provide
indications for the Monge solutions, and visualize the approximate optimal transport
System 7 System 8
Step
K Energy K Energy
GGR Init 240 9.503 170 9.491
GGR LS(1) 960 9.577 680 9.533
GGR LS(2) 3840 9.598 2720 9.543
GGR LS(3) 15360 9.604 10880 9.546
GGR LS(4) 61440 9.606 43520 9.547
2 2 2 2
1 1 1 1
1 1 1 1
0 0 0 0
2 2 2 2
0 0 2 0 0 2 0 0 2 0 0 2
0 0 0 0
2 -2 2 -2 2 -2 2 -2
-2 -2 -2 -2
0 2 0 2 0 2 0 2
0 0 0 0
= -2 8 , K -2
= 170 = -28 , K =-2680 = -2
8
, K =-2
2720 = 8-2 -2
, K = 10880
1 1 1 1
0.5 0.5 0.5 0.5
maps Weingather the 2Dthe contexts. output energies of the GGR approach in Table 4.4, where errors
are absent because no explicit solutions of the original MMOT problems are known in
Appendix
these contexts. The A.evolution Discretization of the approximate of (1.6). For SCE thepotentials
repulsive λ̃ energy
K
(4.1)inis(1.6),shownwe
have for any n \in \{ 2, . . . , N \} ,
in Figure 4.6. The convergence of our GGR approach can be well observed from these
results. Moreover, \int \int to show the \prime transport maps \sum \int T\int 2K , T K
\rho (r)\gamma n (r, r ) 3 (1.12)
\rho (r)\gamma n (r, r )
\prime approximated by the
\prime
obtained solutions, we plot in\prime the remaining three rows of Figure dr dr = \prime dr\prime the
dr4.5 . images of
\Omega \Omega | r - r | ej ei | r - r |
the barycenters of triangular elements within some subregions ω ⊆ Ω under T2K and i,j
K
TNote
3 . For that thewhen two-Gaussian i = j, the system integralρ7explodes , the pictures and hence show we that: impose if thexfirst = electron is
0 for any
n,kk
around the left Gaussian center,
k \in \{ 1, . . . , K\} as extra constraints to avoid numerical instability. In the subsequent then the third electron will go to the region near
the right Gaussian
derivation, we takecenter, \gamma n (r, rand \prime
) = 0the second electron
whenever r and r\prime will belong lie in to the the left same part (to satisfy
element:
the marginal constraints) but stay away from the first one (ω and T2K (ω) lie in two
\int \int \sum \int \int
different regions (r, r\prime ) the left
\rho (r)\gamma naround \prime
Gaussian center); \rho (r)\gamma n (r, if rone \prime
) electron \prime
is located around the
right \Omega Gaussian center, \prime dr
then dr the = other two electrons \prime will drbe draround the left Gaussian
\Omega | r - r | | r - r |
i\not =j ej ei
center while keeping distance away \sum from each
\int \int other. For the three-Gaussian system
ρ8 , we can see that if one electron = is\varrho ilocated
xn,ij around1one dr of dr the\prime +Gaussian O(h) centers, the
other two electrons go to the other | r - r\prime | respectively. Our simulations
i\not =jtwo Gaussian ej eicenters,
match physical intuitions quite \sum well and can support the reliability of our approach.
(A.1) = \varrho i xn,ij cij | ei | | ej | + O(h) = \langle Xn , \Lambda E CE\rangle + O(h),
Although no existing works have shown the existence of Monge solutions for
i\not =j
these two systems, our results appear to provide some indications; see Figure 4.5. In
Fig. 4.7. Frequency percentage distribution of \{ nnzn,i \} n,i and \{ dn,i \} n,i of solutions (Xn )N n=2 .
Upper: \rho = \rho 7 (from left to right: K = 240, 960, 3840, 15360). Lower: \rho = \rho 8 (from left to right:
K = 170, 680, 2720, 10880).
d = 1, N = 3 d = 1, N = 7
106 106
105 105
104 104
Running time (seconds)
Running time (seconds)
103 103
102 102
101 101
= 1 = 4
= 2 = 5
0 0
10 = 3 10 = 6
K2 K2
10-1 10-1
24 48 96 192 384 768 28 56 112 224 448 896
K K
Fig. 4.8. Scaling of GGR on 1D systems. Left: \rho = \rho 1 , \rho 2 , \rho 3 . Right: \rho = \rho 4 , \rho 5 , \rho 6 .
where h := \| e\| \infty represents the size of the largest element. By similar arguments, we
can write for any m, n \in \{ 2, . . . , N \} : m \not = n,
108 108
Running time (seconds)
104 104
= 7 = 8
K2 K2
2 2
10 10
960 3840 15360 61440 680 2720 10880 43520
K K
Fig. 4.9. Scaling of GGR on 2D systems. Left: \rho = \rho 7 . Right: \rho = \rho 8 .
Note that we have excluded j = k cases and impose \langle Xm , Xn \rangle = 0 as extra com-
plementarity constraints. By (A.1) and (A.2), the repulsive energy in (1.6) can be
approximated by
\sum \sum
\langle Xn , \Lambda E CE\rangle + \langle Xn , E\Lambda Xm ECE\rangle
2\leq n\leq N m<n
REFERENCES
[1] A. Alfonsi, R. Coyaud, and V. Ehrlacher, Constrained overdamped Langevin dynamics for
symmetric multimarginal optimal transportation, Math. Models Methods Appl. Sci., 32
(2022), pp. 403--455, https://doi.org/10.1142/S0218202522500105.
[2] A. Alfonsi, R. Coyaud, V. Ehrlacher, and D. Lombardi, Approximation of optimal trans-
port problems with marginal moments constraints, Math. Comp., 90 (2021), pp. 689--737,
https://doi.org/10.1090/mcom/3568.
[3] A. D. Becke, Perspective: Fifty years of density-functional theory in chemical physics, J.
Chem. Phys, 140 (2014), 18A301, https://doi.org/10.1063/1.4869598 (18 pages).
[4] S. Bednarek, B. Szafran, T. Chwiej, and J. Adamowski, Effective interaction for charge
carriers confined in quasi-one-dimensional nanostructures, Phys. Rev. B, 68 (2003),
045328, https://doi.org/10.1103/PhysRevB.68.045328, (9 pages).
[5] J.-D. Benamou, G. Carlier, and L. Nenna, A numerical method to solve multi-marginal
optimal transport problems with Coulomb cost, in Splitting Methods in Communication,
Imaging, Science, and Engineering, R. Glowinski, S. J. Osher, and W. Yin, eds., Springer,
Cham, 2016, pp. 577--601, https://doi.org/10.1007/978-3-319-41589-5 17.
[6] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, Distributed optimization and
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
statistical learning via the alternating direction method of multipliers, Found. Trends Mach.
Learn., 3 (2011), pp. 1--122, https://doi.org/10.1561/2200000016.
[7] G. Buttazzo, L. De Pascale, and P. Gori-Giorgi, Optimal-transport formulation of
electronic density-functional theory, Phys. Rev. A, 85 (2012), 062502, https://doi.org/
10.1103/PhysRevA.85.062502.
[8] R. H. Byrd, J. Nocedal, and R. A. Waltz, Knitro: An integrated package for nonlinear
optimization, in Large-Scale Nonlinear Optimization, G. Di Pillo and M. Roma, eds.,
Springer, Boston, 2006, pp. 35--59, https://doi.org/10.1007/0-387-30065-1 4.
[9] H. Chen and G. Friesecke, Pair densities in density functional theory, Multiscale Model.
Simul., 13 (2015), pp. 1259--1289, https://doi.org/10.1137/15M1014024.
[10] H. Chen, G. Friesecke, and C. B. Mendl, Numerical methods for a Kohn-Sham den-
sity functional model based on optimal transport, J. Chem. Theory Comput., 10 (2014),
pp. 4360--4368, https://doi.org/10.1021/ct500586q.
[11] M. Colombo, L. De Pascale, and S. Di Marino, Multimarginal optimal transport
maps for one-dimensional repulsive costs, Canad. J. Math., 67 (2015), pp. 350--368,
https://doi.org/10.4153/CJM-2014-011-x.
[12] C. Cotar, G. Friesecke, and C. Kluppelberg, \" Density functional theory and optimal
transportation with Coulomb cost, Comm. Pure Appl. Math., 66 (2013), pp. 548--599,
https://doi.org/10.1002/cpa.21437.
[13] C. Cotar, G. Friesecke, and B. Pass, Infinite-body optimal transport with Coulomb
cost, Calc. Var. Partial Differential Equations, 54 (2015), pp. 717--742, https://doi.org/
10.1007/s00526-014-0803-0.
[14] F. Facchinei, H. Jiang, and L. Qi, A smoothing method for mathematical programs with
equilibrium constraints, Math. Program., 85 (1999), pp. 107--134, https://doi.org/10.1007/
s10107990015a.
[15] M. L. Flegel and C. Kanzow, On the Guignard constraint qualification for mathematical pro-
grams with equilibrium constraints, Optimization, 54 (2005), pp. 517--534, https://doi.org/
10.1080/02331930500342591.
[16] R. Fletcher and S. Leyffer, Nonlinear programming without a penalty function, Math.
Program., 91 (2002), pp. 239--269, https://doi.org/10.1007/s101070100244.
[17] R. Fletcher and S. Leyffer, Solving mathematical programs with complementarity
constraints as nonlinear programs, Optim. Methods Softw., 19 (2004), pp. 15--40,
https://doi.org/10.1080/10556780410001654241.
[18] R. Fourer, D. M. Gay, and B. W. Kernighan, A modeling language for mathemat-
ical programming, Management Sci., 36 (1990), pp. 519--641, https://doi.org/10.1287/
mnsc.36.5.519.
[19] G. Friesecke, A. Gerolin, and P. Gori-Giorgi, The strong-interaction limit of density
functional theory, https://arxiv.org/abs/2202.09760, 2022.
[20] G. Friesecke, A. S. Schulz, and D. Vogler, \" Genetic column generation: Fast computation
of high-dimensional multimarginal optimal transport problems, SIAM J. Sci. Comput., 44
(2022), pp. A1632--A1654, https://doi.org/10.1137/21M140732X.
[21] \"
G. Friesecke and D. Vogler, Breaking the curse of dimension in multi-marginal Kantorovich
optimal transport on finite state spaces, SIAM J. Math. Anal., 50 (2018), pp. 3996--4019,
https://doi.org/10.1137/17M1150025.
[22] S. Gerber and M. Maggioni, Multiscale strategies for computing optimal transport, J. Mach.
Learn. Res., 18 (2017), pp. 2440--2471, http://jmlr.org/papers/v18/16-108.html.
[23] J. Grossi, D. P. Kooi, K. J. H. Giesbertz, M. Seidl, A. J. Cohen, P. Mori-Sanchez, \' and
P. Gori-Giorgi, Fermionic statistics in the strongly correlated limit of density functional
theory, J. Chem. Theory Comput., 13 (2017), pp. 6089--6100, https://doi.org/10.1021/
acs.jctc.7b00998.
[24] L. Guo and X.-J. Chen, Mathematical programs with complementarity constraints and a non-
Lipschitz objective: Optimality and approximation, Math. Program., 185 (2021), pp. 455--
485, https://doi.org/10.1007/s10107-019-01435-7.
[25] F. Hecht, New development in freefem++, J. Numer. Math., 20 (2012), pp. 251--265,
https://doi.org/10.1515/jnum-2012-0013.
[26] F. J. Hickernell and Y. Yuan, A simple multistart algorithm for global optimization,
Oper. Res. Trans., 1 (1997), pp. 1--12, http://citeseerx.ist.psu.edu/viewdoc/summary?doi=
10.1.1.46.1346.
[27] T. Hoheisel, C. Kanzow, and A. Schwartz, Theoretical and numerical comparison of relax-
ation methods for mathematical programs with complementarity constraints, Math. Pro-
gram., 137 (2013), pp. 257--288, https://doi.org/10.1007/s10107-011-0488-5.
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
[50] Y. Xu and W. Yin, A block coordinate descent method for regularized multiconvex optimization
with applications to nonnegative tensor factorization and completion, SIAM J. Imaging
Sci., 6 (2013), pp. 1758--1789, https://doi.org/10.1137/120887795.
Downloaded 06/12/23 to 219.142.99.17 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
[51] J. Ye, Necessary and sufficient optimality conditions for mathematical programs with equi-
librium constraints, J. Math. Anal. Appl., 307 (2005), pp. 350--369, https://doi.org/
10.1016/j.jmaa.2004.10.032.