Long History of the Monge-Kantorovich

Transportation Problem

A. M. Vershik

The Mathematical Intelligencer

ISSN 0343-6993

Math Intelligencer
DOI 10.1007/s00283-013-9380-x

1 23
Your article is protected by copyright and all
rights are held exclusively by Springer Science
+Business Media New York. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

1 23
 Author's personal copy

Long History of the

Monge-Kantorovich
Transportation Problem
(Marking the centennial of L.V. Kantorovich’s birth!)

A. M. VERSHIK

eonid Vital’evich Kantorovich (1912–1986) was one reception was quite different. The ideas of Kantorovich, and

L of the great mathematicians and economists of the

twentieth century.
In 2012, the centenary of his birth was marked in St.
Kantorovich himself, were not appreciated in Soviet Russia.
For a long time—until the end of the 1950s—his ideas on
mathematical economics were considered in official circles as
Petersburg. Short histories were presented describing some of anti-Marxist; consequently it was prohibited and even dan-
the main parts of his legacy, which continue in importance gerous to study and develop them. Such a story, carried much
today: duality in linear programming, the so-called ‘‘Monge- farther, is well-known in Soviet biology (‘‘lysenkovshchina’’).
Kantorovich transportation problem,’’ and the ‘‘Kantorovich So, between 1947 and the late 1950s Kantorovich never
metric.’’ Note that 2012 was also the 70th anniversary of the mentioned his ideas about mathematical economics and linear
publication of his historic paper on the transport metric. The programming. My colleagues and I heard his lectures on
present article offers a somewhat expanded version of my talk functional analysis (later published as a book with G. P. Aki-
on that occasion. lov) but knew nothing of his economically related work. He
lectured openly on the subject only after the beginning of
L. V. Kantorovich the Person Khrushchev’s ‘‘ottepel’’’ (‘‘thaw’’), the liberalization of
We remember Leonid Vital’evich Kantorovich for his massive 1957–1958.
contributions to foundations of mathematics, computational L. V. Kantorovich (LV hereinafter) has left a rich record of
mathematics, and other areas, and in particular as one of the the defence of scientific truth. A majority of the Soviet econ-
founders of mathematical economics. omists of the generation of the 1960s and 1970s were pupils of
He began as a child prodigy, entering Leningrad University LV, and many mathematicians (including the author) consid-
at the age of 14. His first paper on descriptive set theory, which ered themselves his pupils.
caught the attention of the mathematical community and in This article will concentrate on one page of the brilliant
particular N. N. Lusin and A. N. Kolmogorov, appeared when legacy of LV in mathematics and its applications: the transport
he was 18. He went on to work in theory of functions, func- metric or Kantorovich metric. I single out this circle of ideas
tional analysis, numerical methods, computer science, and the from his extensive activity for several reasons. First of all,
main contribution: linear programming and then mathemati- because by now this area seems to be the most discussed of the
cal economics, for which he was awarded the Nobel Prize in discoveries of LV. Second, it is an example of deep and non-
1975. See the recent biographical article [8]. trivial relationship between fundamental science and
In many respects his activity in mathematics and applica- applications, and this relationship was characteristic of the
tions reminds one of the activity of another giant of the creativity of LV. And furthermore, this discovery could have
twentieth century, John von Neumann; both were leaders in been done only by someone who was at that time (late 1930s
functional analysis, mathematical economics, numerical and 1940s) a leader in one of the central areas of mathematics
methods, and computer science, and both played important at that time, functional analysis, and simultaneously in the area
roles in the atomic projects of their countries. But their of applied and computational problems.

 2013 Springer Science+Business Media New York

DOI 10.1007/s00283-013-9380-x
 Author's personal copy
situations in which such a formulation applies. He mentions
briefly also the transport problem, to which he returned later.
It would not be an exaggeration to compare the role of this
booklet in the further development of mathematical eco-
nomics—and, more generally, of the mathematical analysis of
extremal problems—with the role of The Mathematical
Principles of Natural Philosophy by Isaac Newton in
mechanics. LV anticipated by 10 years not only almost all
avenues of practical application of linear programming, but
the main ideas of solution methods: the idea of duality, current
in functional analysis of that time, and a numerical method that
can be regarded as a variant of the simplex method proposed
by Dantzig in the late 1940s.
The works of LV and his collaborators and students in this
field did not become known in the West until the end of the
1950s, and his priority was then recognized, although not
always reflected in terminology.
The idea of duality developed subsequently into a new
economic theory, proposed by LV, the theory of objectively
determined valuations. (In the original brochure the term was
‘‘resolving multipliers,’’ in analogy to Lagrange multipliers.)
Much more could and should be said about the dramatic later
destiny of this theory. Let me just say here that the naturalness
and beauty of this theory was appreciated in the USSR by
mathematicians and a few brave economists, and in the West,
a little later, by many scientists. In the USSR, orthodox econ-
omists, dogmatic and ill equipped to understand it, remained
Figure 1. L. V. Kantorovich in his youth. hostile. Their condemnation extended to LV’s 1942 book
about economics with limited resources.
Linear programming—that is, the theory of linear extremal The spokesmen of orthodoxy had the power. For them to
problems with linear constraints—was established in 1938 by characterize Kantorovich as a revisionist, undermining the
LV as a response to specific practical needs (the famous Marxist labor theory of value, was a serious threat. The con-
‘‘problem of Plywood trust’’). Later this theory was rediscov- sequences for LV could have been dire indeed if his work on
ered in the United States by George Dantzig and others. LV military research had not been so highly regarded. To give an
presented linear programming in the booklet Mathematical idea of the level of the critique of LV’s writings, I quote just one
Methods of Organizing and Planning of Production [1] (1939 of his opponents (not the worst or the most obtuse): ‘‘As
edition from LSU Leningrad; this booklet was reprinted several ‘mathematical physics’ does not have a subject of its own
times and translated into other languages). See Fig. 2. He distinct from physics in general, so ‘mathematical economics’
showed the scope of the theory by a long list of specific does not have any subject matter distinct from political eco-
nomics, and political economics necessarily includes its
quantitative side.’’ LV protested that his theory does not con-
......................................................................... tradict Marx’s, that it seeks only to optimize use of limited
resources, and that in that context it is just as objective; but for a
A. M. VERSHIK is Head of the Laboratory
AUTHOR

long time his arguments went unheard.

of Representation Theory at the St. Clearly what was at stake was not only the purity of out-of-
Petersburg Department of the Steklov date dogmas, but the conservation of the influence of an élite
Institute of Mathematics, and is also a Pro- claiming monopoly on the truth. From 70 years’ hindsight,
fessor at St. Petersburg State University. He setting aside reservations no longer needed, we can say that
says he enjoys ‘‘everything in mathematics the theory of LV significantly extends, and in some respects
involving motion–dynamics, asymptotics, per- supersedes, Marx’s theory of value. There is no need to argue
turbation, randomness, representations,… this here. Even at the time there were in the West several
but this is almost all of mathematics.’’ competent articles about this, see for one [17].
.
Steklov Institute of Mathematics
27 Fontanka Birth of the Transport Metric
St. Petersburg 191023 In the 1939 booklet and subsequently, LV singled out the
Russia transportation problem from other problems of linear pro-
e-mail: vershik@pdmi.ras.ru gramming. Soon after, he began writing together with his

THE MATHEMATICAL INTELLIGENCER

 Author's personal copy

Figure 2. The first presentation of linear programming, 1939.

disciple M. K. Gavurin on a special method for solving a linear not translated into English then, and there were only a few in
transport problem—the potential method [4]. It is an imple- the library.) The following circumstance, important in itself,
mentation of the general method of duality, and the visual helped to speed up the process.
interpretation leads immediately to the analogy with the the- In 1946, the world marked the bicentenary of the birth of
ory of fluid dynamics and flows in networks, which was later the French mathematician and physicist Gaspard Monge
much developed. The article with Gavurin was addressed to (1746–1818), belonging to the brilliant constellation of foun-
transport engineers and planners, but it was rejected by sev- ders of modern mathematics. He was involved not only in
eral serious journals in the field and remained unpublished for mathematics, but also its many applications; he was the
almost 10 years. Not waiting for its publication, LV wrote his founder and head of the famous E´cole Polytechnique, and he
‘‘On mass transfer’’ [3]. I want to say about this work the same was responsible for many inventions. In particular, he for-
that I said about the booklet. This is a classic in all respects: it mulated in 1781 a mathematical problem of ‘‘Excavation and
contains a profound idea that goes beyond those examples embankments’’ (Les de´blais et les remblais)—how to transport
studied previously, it is brief and self-contained, one feels that the soil during the construction of the building of forts and
there is nothing more to be said, just as there is nothing to be roads with minimal transport expenses. His hypothesis of
added to the second law of Newton, and finally, it includes a normality of optimal traffic paths to some surfaces was later
program of future research, one that was followed at first very proved in a 200-page work by Appel (1884).
slowly but proceeds especially quickly today. In Leningrad in 1947 the Commission on History of Mathe-
Before I go into the content, let me say a few words about matical and Physical Sciences of the USSR Academy, headed by
the intellectual mystery story of how this work, followed by a Academician V. I. Smirnov, held a public session dedicated to G.
number of other works by LV and his school, became known Monge, where the well-known Moscow geometer B. N. Dela-
around the world. This tale is told in more detail in the intro- unay spoke on the geometric works of Monge and drew
ductory article to the new edition of the booklet by the son of attention in particular to this problem. The proceedings of the
LV, V. L. Kantorovich [2]. session were published in 1947 [5]. LV, apparently on seeing this
It is clear from the year of publication that in wartime the book, drew attention to this problem of Monge and saw how its
Doklady note could not be immediately known either in the solution was connected with his work. He gave a talk to the
Soviet Union or in the West. (Doklady and other journals were Moscow Mathematical Society (22 December 1947), details of

 2013 Springer Science+Business Media New York

 Author's personal copy

Figure 4. Presentation of the Nobel Prize in Economics to

L. V. Kantorovich, 1975.

Figure 3. Street signs honoring Gaspard Monge in Paris,

and honoring L. V . Kantorovich in Rishon Le Zion, Israel. measure U1 is the initial one that needs to be transported, and
(Thanks to G. Thouvenot and J. Romanovsky for the the second measure U2 is the final one, that is, the desired
photographs.) There is as yet no Kantorovich Street in distribution after the transport.
Russia. A transportation, carriage, or transport plan as it is called by
LV is also a probability measure W on the Cartesian product of
which were published in Uspekhi [6]. In this note it was asserted the compact space with itself X 9 X, whose projection Pr1 on
(with no details) that the surfaces Monge spoke of were just the the first factor and Pr2 on the second (marginal distributions)
level surface of the potentials defined in the Doklady note. are the given measures U1 and U2 : The transportation price for
This very short article caught the eye of several American a given plan is an integral of the metric as a function of two
mathematicians and economists (Uspekhi also was not trans- variables, and the task is to minimize this value over all
lated at the time, but was more widely known in the West than admissible transport plans: in obvious notation it takes the
Doklady), and they began to look up the publications of LV. form
One of them was Tjalling C. Koopmans, who had indepen- Z Z 
dently studied the transportation problem in classified inf rðx; yÞdWðdx; dyÞ; W : Pr1 W ¼ U1 ; Pr2 W ¼ U2
X X
wartime applied work, in the finite-dimensional case, and
who, for his 1949 Econometrica paper, was awarded the 1975  Kr ðU1 ; U2 Þ:
Nobel Prize together with LV.
About the end of the 1950s, LV’s main work on the subject If the space X is finite, then the measures U1 ; U2 are
(not only economics) became known in the West, more so probability vectors, and the measure W is a matrix (a gen-
than in the USSR. Since that time, some scholars in the West eralized doubly stochastic matrix). Note that in this case, r is
speak of ‘‘the Monge–Kantorovich transport problem,’’ which the matrix of the cost of transporting a unit of cargo from one
seems to me pretty fair terminology. point to another, but it is not necessarily a metric.
I turn now in particular to the article of 1942. Its main However—and here the naturalness of LV’s formulation of
content as compared with the previous cycle of finite prob- the problem for continuous spaces appears—we should
lems of linear programming is in a natural generalization of the assume the function r to be a metric; in the particular Monge
situation—here the transportation is a probabilistic measure case it is the Euclidean metric. This characterizes the fruitful
on a compact metric space X with the metric r, namely a generalizations of finite problems to the continuous case. For

THE MATHEMATICAL INTELLIGENCER

 Author's personal copy
finite problems, it is not so important that the objective func- Those connections are extensive. The duality theorem of
tion be a metric. finite-dimensional linear programming is, in a different for-
It is easy to see that feasible plans always exist, and the mulation, based on the fundamental theorem (von
problem is well posed. The indicated infimum Kr satisfies the Neumann) of the theory of matrix games, and it is also Weyl’s
triangle inequality as a function of the measures U1 ; U2 ; and duality theorem in the theory of convex polyhedra, and a
we obtain a metric on the space of probability measures on the theorem about the solution of systems of linear inequalities,
compact space. etc. In hindsight we see these equivalences fall out in a single
Just this Kr as a function of two probability measures on line, and this, with all of those connections, based on the
the compact (X, r) should be called the Kantorovich metric theory of duality, is how I organized the course ‘‘Extremal
on the product space. problems,’’ which I taught for 20 years at Leningrad State
Many mathematicians, not knowing of LV’s work, used University (1973–1993).
such a metric in particular cases (see details in [15]). However,
LV was the first to give a nontrivial example of an infinite-
the point of LV’s treatment is not the definition, but the main
dimensional duality.
theorem, which the subsequent rediscoverers did not prove,
or even formulate.
TH E O R E M 1 Further Development
Kr ðU1 ; U2 Þ
Z  The Kantorovich–Rubinstein Norm
¼ sup uðxÞðU2  U1 ÞðdxÞ : uðxÞ  uðyÞ 6 rðx; yÞ In his article LV emphasizes the utility of studying the metric
X
introduced in the space of probability measures on a compact
Here the supremum is taken over all Lipschitz functions space. It is obvious that the Kantorovich metric is functorial
with constant 1 relative to the given metric; LV calls these (that is, it is preserved under isometries of the compact space),
Lipschitz functions potentials in the transportation problem. but so are a lot of other metrics on the space of probability
The transport plan W is called a potential plan if for almost all measures. To see what distinguishes it among all such metrics,
point pairs x, y in the sense of the measure W; and for a certain we turn to the important work of LV and his student G. Sh.
Lipschitz function U ðÞ (the optimal potential), we have Rubinstein [7]. This work complements the main result of the
note of LV, and its result follows directly from the fundamental
U ðxÞ  U ðyÞ ¼ rðx; yÞ; theorem there.
in other words, the Lipschitz inequality U(x) -
U(y) 6 r(x, y) reduces to equality on a set of full W TH E O R E M 2 Consider the linear space V0(X) of alternating
measure. measures of bounded variation with zero charge on the
compact space (X, r).
CO R O L L A R Y 1 A transport plan W is optimal if and only Every nonzero measure of this type is presented uniquely as
if it is potential. the difference of two nonnegative and mutually disjoint finite
The corollary follows immediately from the theorem. The measures: l = l+ - l- in the space of all b.v. signed measures
optimal potential provides the maximum of the right-hand V(X); it is obvious that l+ = kl0 +; l- = kl0 - with the same
integral. The theorem is the duality theorem of linear pro- positive factor k, where l0 +, l0 - are probability measures.
Define for any such l
gramming, which states that the value of the inf in the original
problem is the same as the sup in the second, dual, problem. jlj  kKr ðl0þ ; l0 Þ:
All these statements imply the existence of an optimal trans- Then j  j is a well-defined norm in V0(X), under which it is
port plan (for a compact space). a separable normed space, which has as its dual the space
At that time there was no duality theory for infinite- of all Lipschitz functions modulo the constants.
dimensional programming problems; it appeared only later. So
Thus, the Kantorovich metric extends to a norm on the
in the article LV proved the theorem by a direct approximation.
signed measures of bounded variation. This norm on the
Note that the sufficiency (the inequality inf > sup) is trivially space of measures is called the Kantorovich-Rubinstein norm
true, and the approximation is needed to prove the necessity;
that is, the nontrivial part is that sup > inf : This example jljKR :
became a model for the development of the general theory of Curiously, before this paper [7] it was not known whether
infinite-dimensional duality of extremal problems, which the space of Lipschitz functions is conjugate to a Banach
nowadays includes a large number of classical and nonclassical space; convergence of bounded sequences in this norm is
problems, and it also became the basis of computational weak convergence of measures. The space ðV0 ðXÞ; j  jÞ is
methods for finite problems. Curiously, the inventor of the in general incomplete, and the completion includes rather
simplex method, G. Dantzig, wrote in his memoirs that he complicated objects that are not measures.
understood the connection of his method with the theory of In my student years I had the good fortune to be on friendly
duality only after a conversation with von Neumann, to whom terms with G. Sh. Rubinstein, and I was perhaps the first whom
he had come to show his invention, and who immediately GSh told about this norm (1957). He modestly noted that LV
explained all the implications and connections. liked this ‘‘Kantorovich norm,’’ in other words he recognized

 2013 Springer Science+Business Media New York

 Author's personal copy
that it had already been implicit in the article by LV alone; I It is surprising that this simple corollary of that same duality
agree with this. theorem is not stated until [11], where it plays an important
Behind Rubinstein’s remark is the fact that the Kantorovich role. Even for the finite-dimensional case these norms deserve
metric is translation-invariant. From this one can readily infer further study from the geometrical point of view. Consider the
the existence of the extension. Indeed, by homogeneity any finite set of n points, with metric function constant. The unit
metric on the simplex of probability measures extends to the ball in that norm coincides with the convex hull of the roots in
cone generated by the simplex; but each signed measure of the Lie algebra SL(n).
finite variation can be represented as a difference of two Now it may seem at first that, for example, the norms on
nonnegative measures, thus if we have a norm on the cone we V0(X)—|.|KR and |.|HK—are in general drastically differ-
have norm on the linear space of measures. Finally, a metric on ent, as well as the isometric embeddings of X into V0(X),
the cone is a norm iff it is invariant under translations. The last |.|KR and into CðXÞ: But it turns out that there are metric
property of the Kantorovich metric trivially follows from the spaces (the so-called linear rigid spaces) for which there is
linearity of the classical transport problem. only one admissible norm on V0, and thus the Kantorovich–
But this remark was very important because if we have a Rubinstein norm coincides with the Hausdorff–Kuratowski
norm we can apply the techniques of the theory of Banach norm and any admissible norm—see [11])! This means that
spaces, which they did. This property is not enjoyed by the up to linear isometry there is only one Banach space in
wider class of new metrics; see below. which our metric space can be embedded isometrically as a
subset whose linear hull is dense. The main example of such
a space is the Urysohn universal metric space (see [11]).
Characterization of the KR-Norm
We can now formulate an axiomatic definition of the Kant-
orovich metric and the norms of Kantorovich-Rubinstein. In Other Admissible Metrics
this case, we drop the requirement of compactness of the Any admissible norm |.| on the space V0(X), restricted to the
space. Indeed, the definition of the Kantorovich metric and the simplex S(X) of Borel probability measures, gives a metric:
norm of Kantorovich-Rubinstein makes sense for noncompact |l1 - l2|: R|.|(l1, l2), where l1 ; l2 2 SðXÞ: This metric
spaces, but not for Rall measures, only for measures l with a R|.| on S(X) evidently has the following property, which is
finite first moment: rðx; yÞdlðxÞdlðyÞ\1: This includes in invariant of the metric under translation:
particular the class of discrete measures with finite support.
Consider, then, a separable metric space X with metric r; the Rj:j ðkl1 þ ð1  kÞm; kl2 þ ð1  kÞmÞ ¼ kRj:j ðl1 ; l2 Þ;
simplex S(X) of all probability Borel measures with finite first k 2 ð0; 1Þ:
moment on X; and the vector space V0(X) over R consisting of
all finite formal real linear combinations of points of X with The Kantorovich metric of course satisfies this condition.
coefficients summing to zero. One can interpret the space Because translation-invariance is a sufficient condition for the
V0(X) as the set of discrete signed measures on X with zero metric to be given by a norm, the opposite assertion is also true:
charge. By an elementary charge I will mean a difference of two
delta measures: ex;y ¼ dx  dy 2 V0 ðXÞ; the set of all elemen- TH E O R E M 4 The admissible metric R on the simplex
tary charges generates the space V0(X) as a linear space. S(X) can be extended to an admissible norm on the space
V0(X)—that is, R(l1, l2) = |l1 - l2|, when l1 ; l2 2 SðXÞ—
DE F I N I T I O N 1 A seminorm j  j on V0(X) (resp. a metric iff it is invariant under translation.
R on the simplex of Borel probability measures S(X)) is called Thus the previous theorem can be restated as follows.
compatible with the metric r, or admissible, if jex;y j ¼ rðx; yÞ The Kantorovich metric is the largest admissible metric
(resp. R(dx, dy) = r(x, y)). on the simplex of probability measures in the class of all
In general, for a given metric space (X, r) there are many translation-invariant admissible metrics.
norms on V0(X) compatible with the metric r, and conse-
quently, many isometric embeddings of X into Banach spaces. But there are many admissible metrics which are not
For example, the well-known isometric embedding THK :
translation-invariant and consequently do not extend to an

V0 ðXÞ ! CðXÞ; the space of bounded continuous functions
admissible norm. One such metric is the Lévy–Prokhorov
on X with sup norm, was considered by F. Hausdorff and in
metric. Here are some others.
detail by K. Kuratowski: Consider the Kantorovich power metrics on the simplex

THK ðex;y Þ ¼ Fx;y ð:Þ; Fx;y ðzÞ ¼ rðx; zÞ  rðy; zÞ; Fx;y 2 CðXÞ; metric space (wrongly called Vasershtein metrics in [14], see
below), defined as follows:
with linear extension to V0. Denote this norm by |.|HK. We  Z Z 1=p
may examine the relationship between different embed- Kr;p ðl1 ; l2 Þ ¼ inf rðx; yÞp Wðdx; dyÞ ;
X X
dings, as was done in [7] and almost simultaneously in [9]. 
W : Pr1 W ¼ U1 Pr2 W ¼ U2 :
TH E O R E M 3 ([11]) For any metric space, the Kantorovich–
Rubinstein norm in the space V0(X) is the largest admissible When p = 1 this is the Kantorovich metric. These power
norm; that is, for any admissible seminorm j  j and any ele- metrics have been considered only in recent years; they are
ment a of the space V0(X) we have jaj 6 jajKR . described in detail in the book [14] and in references therein.

THE MATHEMATICAL INTELLIGENCER

 Author's personal copy
It is clear that if p [ 1, then this norm is not invariant under 2. The dual posing of the transport problems and corre-
translations. sponding class of admissible metrics.
Perhaps the analysis of the space of all transport problems Define Lipschitz-type admissible metrics as
or even linear problems must be studied more carefully. Z
Recent studies have shown the particular importance of the R L ðl1 ; l2 Þ ¼ sup j uðxÞdðl1  l2 Þj  jjl1  l2 jjL ;
Kantorovich quadratic metric (p = 2) for a range of applica- u2L

tions (see [16]). The main result is that the simplex of where L is an appropriate class of Lipschitz functions on
probability measures on a compact Riemannian manifold the metric space (X, r).
equipped with the quadratic Kantorovich metric is a (possibly
infinite) Riemannian manifold. TH E O R E M 5 ([11]) The class of Lipschitz-type admissible
The interrelation of all admissible metrics, in particular of all metrics on S(X) coincides with the class of translation-
Kantorovich power metrics, is more complicated. Power invariant metrics, and consequently with the class of metrics
metrics, as Lp-norms of difference of measures, increase admitting extension to admissible norms on V0(X).
monotonically with increasing p, so the Kantorovich metric
(p = 1) is the smallest in the class. Correspondingly, in the
If L is the space of all Lipschitz functions, we obtain the
class of all admissible metrics, the Kantorovich metric is a sort
Kantorovich metric, which is the largest metric in this class.
of saddle point (see below).
It seems likely that the Kantorovich metric is the unique
The fact that the power rp (p [ 1) of a metric, is, gen-
metric that belongs to both classes of defined metrics. This
erally speaking, not a metric (the triangle inequality may
gives it a sort of ‘‘saddle point position’’ in the set of all
fail) explains the failure of the extension property. If for a
admissible metrics.
metric r and a number p [ 1 it happens that r p is also a
metric, then the study of the metric Kr,p is reduced to the
study of the Kantorovich metric Kpr for a different metric
The Monge Problem (Paths of Transportation)
Actually the original Monge problem was not only to find a
space, in which extension to the Kantorovich–Rubinstein
plan of mass transportation in the sense we have been
norm is possible.
considering, but to choose a transport routing. This ques-
It is interesting that in the case p ¼ 1, the power Kan-
tion is more delicate: in contrast to LV’s problem, here,
torovich metric coincides with the well-known Lévy-
generally speaking, there need not be a unique solution for
Prokhorov metric–see [10], [19], [20].
best transportation plan, as is shown by trivial examples
already in the finite case.
Two Classes of General Transport Metrics
One of the first accurate statements of the problem, to my
Let us look more closely at two kinds of definition of admis- knowledge, was in my work [15] (see also [16]): to find a
sible metrics that motivated the Kantorovich duality theorem. measurable transformation T of a metric space X into itself,
1. Direct posing of the transport problems (nonlinear in taking the initial probability Borel measure l1 to the final
general) and the corresponding class of admissible metrics. one l2 and minimizing the single integral:
For two Borel probability measures l1, l2 on the metric Z 
space (X, r), define the class Wl1 ;l2 of Borel probability
inf rðx; TxÞdl1 ðxÞ : T l1 ¼ l2 :
measures on the space X 9 X (transport plans) whose mar- X
ginal projections are the given l1, l2. Choose a norm N(.) on
This problem was attacked by V. Sudakov, who solved it
the space of measurable functions with respect to the measure
for Euclidean compact sets and Lebesgue measures—see later
space ðX  X; WÞ for all W 2 Wl1 ;l2 : The transportation
corrections in [12, 18] and references therein.
problem is to find Here there is a connection to variational problems on
inf N ðrð:; :ÞÞ  RN ðl1 ; l2 Þ: infinite groups, in this case the group of measure-preserving
w2Wl1 ;l2
transformations (see [16]). The following variational problem
This infimum RN(., .) clearly defines an admissible metric on is closer in nature to the original formulation of Monge: to find
the simplex S(X). a one-parameter group (or semigroup) Tt ; t 2 RðRþ Þ of
If N is the Lp norm, we have the previous Kantorovich measurable transformations of the metric space X minimizing
power metrics. In particular for p = 1 we obtain the Kant- the (time average) integral:
Z t Z 
orovich metric, which is minimal in this family of metrics.
The general linear problem of transportation is the problem inf rðx; Ts xÞdl1 ðxÞds : Ts l1 ¼ ls ; s 2 ½0; t; t 2 ½0; T  :
0 X
of minimizing the (double) integral of the metric (or minimum
of expectation) such as the L1-norm over a variety of plans, Variational problems on infinite groups (in particular Lie
subject to linear constraints: groups) have still not been studied enough. However,
Z Z  striking achievements in the problem of optimal transpor-
inf rðx; yÞdwðdx; dyÞ : w 2 L ; tation, applications to the Monge-Ampère problem in
X X
partial-differential equations, the theory of the Ricci flow,
where L is the set of plans with given marginals and sat- and others testify to the flourishing of ideas in this theory,
isfying the given constraints. which was first stated in the proper generality 70 years ago.

 2013 Springer Science+Business Media New York

 Author's personal copy

Figure 5. Communication was not completely broken between the East and the West in the mid-30’s! Here are letters to
Kantorovich from Garrett Birkhoff and John von Neumann.

What Name to Give These Things occurs primarily in Western literature. It so happened that my
Previously I used what I regard as the only correct attribution colleague and friend R. L. Dobrushin, head of the laboratory
of concepts and theorems, namely ‘‘Monge–Kantorovich where L. N. Vasershtein worked, with understandable
problem,’’ ‘‘Kantorovich metric,’’ etc. There is no doubt that enthusiasm spread the word mostly among probabilists and
the Kantorovich metric was defined by L. V. Kantorovich, and statistical physicists about the ‘‘new’’ metric and its spectacular
his name should be on it. Similarly for the Kantorovich– applications. I spoke to Dobrushin in 1975 and told him that
Rubinstein norm. what he called the Vasershtein metric in the report is the
Scientific cooperation between the USSR and the Kantorovich metric. After some discussion, he agreed fully
Western allies was inadequate (we could almost say non- and even said so in one of his later works. But it was too late,
existent) in those days; even within a country there was the wrong name stuck.
poor awareness of developments—all of this meant that the Vasershtein’s interesting article [13] was very brief (it seems
metric was repeatedly rediscovered by other authors, but to me that few people who refer to it have looked at it), it does
not in such a fundamental way as in LV’s work, and, most contain in passing a definition of the Kantorovich metric and
importantly, without his main result on duality. In the list of applies it to the behavior of Markov fields. But there is no
those who have used similar ideas, some stand out for their definition of power metrics, although in the literature [14]

importance: the d-metric of D. Ornstein in ergodic theory; those are also called Vasershtein metrics. Undoubtedly, the
the method of coupling in probability, statistics, and sta- work of Vasershtein is worthy of mention in this connection,
tistical physics; the theory of polymorphisms; and much, but I think we should restore the correct terminology out of
much more. The importance and popularity of the topic respect for L. V. Kantorovich, to the teachers and pioneers in
first studied by LV in the late 1930s led to the reappearance our science.
of the metric under new names.
It is especially ironic to find the Kantorovich metric called ACKNOWLEDGMENTS
the Vasershtein metric. Leonid Vasershtein is a famous Supported by the grants RFBR 11-01-12092-ofi-m and RFBR
mathematician specializing in algebraic K-theory and other 11-01-00677-a. The author expresses his gratitude to Joseph
areas of algebra and analysis, and my good friend—and he is Romanovsky for his help in translation of the article and to
absolutely not guilty of this distortion of terminology, which Vl. Kantorovich for photos from the family archive.

THE MATHEMATICAL INTELLIGENCER

 Author's personal copy
REFERENCES [10] Rachev, S., Rüschendorf, L., 1996: Mass Transportation Prob-
[1] Kantorovich, L. V., 1939: Mathematical methods in the organi- lems, Vol. I: Theory 1998, XXV, Springer. 540 pp.; Vol. II.
zation and planning of production. Leningrad Univ., 1939. Applications. 450 pp.
[English translation: Management Science, 6, 4 (1960), 363–422.] [11] Melleray, J., Petrov, F., and Vershik, A., 2008: Linearly rigid metric
[2] Kantorovich, L. V., 2012: Mathematical methods in the organi- spaces and the embedding problem. Fund. Math. 199, 2, 177–194.
zation and planning of production. Reprint edition of the book, [12] Bogachev, V. I., and Kolesnikov, A. V., 2012: The Monge-
published in 1939, with introductory paper of L. V. Kantorovich. Kantorovich problem: achievements, connections, and perspec-
St. Petersburg, Publishing House of St. Petersburg Univ., tives Russian Mathematical Surveys, 67, 5, 3–110.
96 pp. [13] Vasershtein, L. N. 1969: Markov processes over denumerable
[3] Kantorovich, L. V., 1942: On translocation of masses. USSR AS products of spaces describing large system of automata,
Doklady. New Serie. vol. 37, 7–8, 227–229 (in Russian). [English Problems Inform. Transmission, 5, 3, 47–52.
translation: J. Math. Sci., 133, 4 (2006), 1381–1382.] [14] Villani, C., 2006: Optimal Transport, Old and New, Springer, 635 pp.
[4] Kantorovich, L. V., and Gavurin, M. K., 1949: Application of [15] Vershik, A., 1970: Some remarks on infinite dimensional prob-
mathematical methods to problems of analysis of freight flows. lems of linear programming, Uspekhi Mathematicheskih nauk,
Problems of raising the efficiency of transport performance, 25, 2, 117–124 (in Russian). [English translation: Russ. Math.
Moscow-Leningrad, 110–138 (in Russian). Surveys, 25, 5 (1970), 117–124.]
[5] Delaunay, B. N., 1947: Gaspard Monge as a mathematician. [16] Vershik, A., 2004: Kantorovich metric: initial history and little-known
Gaspard Monge. Collection of papers to bicentenary of his applications, Zapiski nauch. sem. POMI, 312, 69,1–7 (in Russian).
birthday, USSR Academy of Science, 7, 1–7 (in Russian). [English translation: J. Math. Sci, 133:4 (2006), 1410–1417.]
[6] Kantorovich, L. V., 1948: On a problem of Monge. Uspekhi Mat. [17] Campbell, Robert W., 1961: Marx, Kantorovich, Novozhilov,
Nauk, 3, 225-226 (in Russian). [English translation: J. Math. Sci., Stoimost versus reality, Slavic Review, 20, No. 3, 402–418.
133, 4 (2006), 1383.] [18] Ambrosio, L., 2003: Lecture Notes on Optimal Transport
[7] Kantorovich L. V., Rubinshtein G. S., 1958: On a space of totally Problems. Lecture Notes in Mathematics ‘‘Mathematical aspects
additive functions. Vestn. Leningrad. Univ., 13, 7 (1958), 52–59 of evolving interfaces’’ (Cime Series, Madeira (PT), 2000) 1812,
(in Russian). P. Colli and J. F. Rodriques, eds., 1–52.
[8] Vershik, A. M., Kutateladze, S. S., and Novikov, S. P. 2012: Leonid [19] Strassen, V., 1965: The existence of probability measures with
Vital’evich Kantorovich (on the 100th anniversary of his birth). given marginals, Annals of Math. Statist. 36, 2, 423–439.
Russian Math. Surveys, 67, 3, 589 (in Russian). [20] Vershik, A., 2013: Two ways of defining compatible metrics on
[9] Arens R., Eells J., 1956: On embedding uniform and topological the simplex of measures, Zapiski nauchn.sem. POMI, 411 (in
spaces. Pacif. J. Math. 6, 397–403. Russian). [English translation: J. Math. Sci.]

 2013 Springer Science+Business Media New York