prob transport
TRANSCRIPT
Mass Transfer ProblemsOptimal NetworksPath Functionals
Optimization Problemsfor Transportation Networks
PhD Thesis
Alessio BrancoliniAdvisor: Prof. Giuseppe Buttazzo
Scuola Normale Superiore, Pisa
Thursday, October 12th 2006
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Outline
1 Optimal Transportation ProblemsMonge’s ProblemKantorovich’s ProblemComparison of the Problems
2 Optimal Networks for Mass Transportation ProblemsSetting of the ProblemTools and results
3 Path Functionals over Wasserstein SpacesPath Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Monge’s ProblemKantorovich’s ProblemComparison of the Problems
Mass Transfer Problems
Mass Transfer Problems
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Monge’s ProblemKantorovich’s ProblemComparison of the Problems
Statement of Monge’s Problem
Problem (Monge’s Problem, 1781)
X metric space; µ+, µ− finite positive Borel measures;mass balance condition µ+(X ) = µ−(X );
minimize
M(t) :=
∫X
c(x , t(x)) dµ+(x);
among measurable maps t : X → X such thatµ−(B) = µ+(t−1(B)) (transport maps); c : X × X → R,non-negative, l.s.c.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Monge’s ProblemKantorovich’s ProblemComparison of the Problems
Statement of Kantorovich’s Problem
Problem (Kantorovich’s Problem, 1940)
X metric space; µ+, µ− finite positive Borel measures;mass balance condition µ+(X ) = µ−(X );
minimize
K (µ) :=
∫X×X
c(x , y) dµ(x , y);
among positive Borel measures on X × X such thatµ+(A) = µ(A× X ), µ−(B) = µ(X × B) (transport plans);c : X × X → R, non-negative, l.s.c.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Monge’s ProblemKantorovich’s ProblemComparison of the Problems
Comparison of the Problems
if t is transport map, the measure µt given byµt(A× B) := µ+(A ∩ t−1(B)) is a transport plan andM(t) = K (µt).
Disadvantages of Monge’s Formulation:non-linearity of M in the unknown of the problem t ;there may be no transport maps.
Advantages of Kantorovich’s Formulation:linearity of K in the unknown of the problem µ;the set of transport plan is not empty, convex, weakly-∗
compact;the previous items imply the existence of an optimalmeasure.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
Transportation Networks
Transportation Networks
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization I
Ω
spt
spt
µ+
µ−
Σ
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization II
γ
Σ
y
γ
x
Σ
dΣ(x , y) = A(H1(γ \ Σ)) + B(H1(γ ∩ Σ)) + C(H1(Σ))
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization III
γ
Σ
y
γ
x
Σ
dΣ(x , y) = infA(H1(γ \ Σ)) + B(H1(γ ∩ Σ)) + C(H1(Σ)) :
γ closed connected, x , y ⊆ γ ⊆ Ω
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization IV
µ+, µ− are the distributions of working people and workingplaces;
the “pseudo-distance” dΣ takes into account thetransportation network Σ via a function J;
in order to introduce the distance dΣ we consider afunction J : [0,+∞]3 → [0,+∞], e.g.J(a, b, c) = A(a) + B(b) + C(c);
the cost of the transportation on a given path γ, will then be
J(H1(γ \ Σ),H1(γ ∩ Σ),H1(Σ)).
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization V
Ω ⊆ RN open bounded set, Σ ⊆ Ω closed connected set,µ+, µ− probability measures on Ω;
J : [0,∞]3 → R l.s.c., non-negative, non-decreasing,continuous in a; J(a, b, c) ≥ G(c) with G(c) → +∞ asc → +∞;
define
dΣ(x , y) = infJ(H1(γ \ Σ),H1(γ ∩ Σ),H1(Σ)) :
γ closed connected, x , y ⊆ γ ⊆ Ω;
T (Σ) is minimum of Kantorovich functional between themeasures µ+, µ− with cost function dΣ.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization VI
Problem
Minimize the functional
Σ 7→ T (Σ)
among Σ ⊂ Ω close connected sets.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization VII
(X , d) complete metric space.
Theorem (Gołab)
Cn closed connected sets, Cn → C. Then
H1(C) ≤ lim infn→∞
H1(Cn).
Theorem (Gołab, generalized)
Γn,Σn closed connected sets, Γn → Γ, Σn → Σ. Then
H1(Γ \ Σ) ≤ lim infn→∞
H1(Γn \ Σn).
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Setting of the ProblemTools and results
A Model of Urban Optimization VIII
Theorem
J : [0,∞]3 → R satisfying the hypoteses before.
Then, there exists a minimizer of the functional T in the class ofclosed connected sets.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Tree-shaped Structures
Tree-shaped Structures
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Tree-shaped Structures I
Many natural systems show a distinctive tree-shapedstructure: plants, trees, drainage networks, root systems,bronchial and cardiovascular systems.
These systems could be described in terms of masstransportation, but Monge-Kantorovich theory turns out tobe the wrong mathematical model since the mass iscarried from the initial to the final point on a straight line.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Tree-shaped Structures II
x1
x2
y1
x1
x2
y1
Figure: V-shaped versus Y-shaped transport.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path Functionals
Path Functionals
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path Functionals, Setting of the Problem
(X , d) metric space such that bounded closed sets arecompact.
The functional we consider is
J (γ) =
∫ 1
0J(γ(t))|γ′|(t) dt ,
γ ∈ Lip([0, 1], X ) such that γ(0) = x0 and γ(1) = x1;
J : X → [0,+∞] is l.s.c.;
|γ′|(t) is the metric derivative, i.e.
|γ′|(t) = lims→t
d(γ(s), γ(t))|s − t |
.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path Functionals, Existence Result
TheoremAssume that
there exists a curve γ0 such that J (γ0) < +∞,
and ∫ ∞
0
[inf
Br (x0)J]
dr = +∞ or J ≥ c > 0,
then the functional admits a minimum in the class of Lipschitzcurves γ : [0, 1] → X such that γ(0) = x0, γ(1) = x1.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Wasserstein Spaces
Wp(Ω), the p-Wasserstein space, the set of Borelprobability measures with finite momentum of order p, thatis all probability measures µ such that∫
X[d(x0, x)]p dµ(x) < +∞;
The p-Wasserstein distance between µ+, µ− is the givenby Kantorovich functional with cost c(x , y) = d(x , y)p
Wp(µ+, µ−) :=
[min
∫X×X
[d(x , y)]p dµ(x , y)
] 1p
.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path functionals over Wasserstein Spaces
X = Wp(Ω); Ω compact metric space; m positive, finite,non-atomic Borel measure.
the coefficient J is l.s.c. local functional defined onmeasures given by:
J(µ) =
∫Ω
f(
dµ
dm
)dm +
∫Ω\Aµ
f∞(
dµs
d|µs|
)d|µs|+∫
Aµ
g(µ(x)) d#(x).
f : R → [0,+∞] convex, l.s.c. and proper; g : R → [0,+∞]l.s.c., subadditive, g(0) = 0; g0(s) = f∞(s).
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path Functionals, Existence Theorem in Wp(Ω)
Theorem
If f (s) > 0 for s > 0 and g(1) > 0, then J ≥ c > 0;
in particular, the functional
J (γ) =
∫ 1
0J(γ(t))|γ′|(t) dt
achieves a minimum in Lip([0, 1],Wp(Ω)) such thatγ(0) = µ+, γ(1) = µ−, provided there exists a curve offinite value.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path Functionals, Concentration and Diffusion Cases
Ω compact, convex subset of RN .
Concentration case: f = +∞, g(z) = |z|r , (0 ≤ r < 1),
J(µ) = Gr (µ) =
∑k∈N(ak )r µ =
∑k∈N akδxk
+∞ otherwise;
Diffusion case: f (z) = |z|q, g = +∞, (q > 1),
J(µ) = Fq(µ) =
∫Ω |u|
qdx µ = u · LN
+∞ otherwise.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path Functionals, Results on Concentration Case
if µ+, µ− are atomic measures, then there exists anoptimal curve γopt such that the functional is finite;
if r > 1− 1/N, the same is true for all couple of measures;
if r ≤ 1− 1/N, measures µ+, µ− such that the functional isnot finite on every curve connecting them can be found.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Path Functionals, Results on Diffusione Case
if µ+, µ− ∈ Lq(Ω), then there exists an optimal curve γopt
such that the functional is finite;
if q < 1 + 1/N, the same is true for all couple of measures;
if q ≥ 1 + 1/N, measures µ+, µ− such that the functional isnot finite on every curve connecting them can be found.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Transport Paths
Transport Paths
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Transport Paths between Atomic Measures I
Ω ⊂ RN compact, convex;
µ+ =∑m
i=1 aiδxi , µ− =
∑nj=1 bjδyj convex combinations of
Dirac measures;
G weighted directed graph; spt µ+, spt µ− ⊆ V (G);
the mass flows from the initial measure µ+ to the finalmeasure µ− “inside” the edges of the graph G;
y1
y2
y3
x1
x2
x3
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Transport Paths between Atomic Measures II
The point is now to provide to each transport path G asuitable cost that makes keeping the mass togethercheaper. The right cost function is
Mα(G) :=∑
e∈E(G)
[w(e)]αl(e),
l(e) length of edge e, 0 ≤ α < 1 fixed;
this cost takes advantage of the subadditivity of thefunction t 7→ tα in order to make the tree-shaped graphsmore economic.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Transport Paths between General Measures I
e (oriented edge) 7→ µe = (H1 e)e (vector measure),div µe = δe+ − δe− ;
G 7→ TG =∑
e∈E(G) w(e)µe;
div G = µ+ − µ− sums up all conditions;
a general transport path is defined by density and the costas a lower semicontinuous envelope:
Mα(T ) = infTGi
→Tlim infi→+∞
Mα(TGi).
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Transport Paths, Some Results
if 1− 1/N < α ≤ 1, the existence of an optimal transportpath is assured between arbitrary measures;
conditions to link arbitrary measures have been found byDe Villanova, Solimini;
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Irrigation Trees
Irrigation Trees
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Irrigation Trees I
(Ω,M, µ) probability space;
associate to each ω ∈ Ω a 1-Lipschitz curveχω : [0,+∞[→ RN such that χω(0) = S for a.e. ω (set offibres);
[ω]t = ω′ : (χω′)|[0,t] = (χω)|[0,t];Mt(χ) = ω : χω not constant on [t ,+∞].
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Irrigation Trees II
Definition (MMS functional)
Let α ∈ [0, 1]. The MMS functional is defined by
MMS(χ) :=
∫ +∞
0
[∫Mt (χ)
[µ([ω]t)]α−1 dµ(ω)
]dt .
Definition
A sequence of set of fibres χn is said to converge to a set offibres χ if, for a.e. ω, χω,n(t) → χω(t) for all t ∈ [0,+∞[.
Alessio Brancolini Optimization Problems for Transportation Networks
Mass Transfer ProblemsOptimal NetworksPath Functionals
Path Functionals over Wasserstein SpacesXia’s Transport PathsMaddalena, Morel, Solimini’s Irrigation Trees
Irrigation Trees, Some Results
Theorem
In a suitable subset functionals of the type
χ 7→ MMS(χ) + F (µχ),
with F is a l.s.c., admits a minimizer.
A case of interest is that of F finite on a certain measure µ−
and infinity elsewhere. Xia’s model and Maddalena, Morel,Solimini’s model come out to be the same.
Alessio Brancolini Optimization Problems for Transportation Networks
Bibliography
Bibliography I
A. Brancolini and G. Buttazzo.Optimal networks for mass transportation problems.
A. Brancolini, G. Buttazzo, and F. Santambrogio.Path Functionals over Wasserstein Spaces.
G. Buttazzo, E. Oudet, and E. Stepanov.Optimal Transportation Problems with Free DirichletRegions.
F. Maddalena, J.M. Morel, and S. Solimini.A Variational Model of Irrigation Patterns.
Q. Xia.Optimal Paths Related to Transport Problems.
Alessio Brancolini Optimization Problems for Transportation Networks