Numerical Optimal Transport and Applications

Size: px
Start display at page:

Download "Numerical Optimal Transport and Applications"

Transcription

1 Numerical Optimal Transport and Applications Gabriel Peyré Joint works with: Jean-David Benamou, Guillaume Carlier, Marco Cuturi, Luca Nenna, Justin Solomon

2 Histograms in Imaging and Machine Learning Color histograms: Input image

3 Histograms in Imaging and Machine Learning Color histograms: optimal transport Input image Modified image

4 Histograms in Imaging and Machine Learning Color histograms: optimal transport Bag of words: Input image Modified image

5 Overview Transportation-like Problems Regularized Transport Optimal Transport Barycenters Heat Kernel Approximation

6 Radon Measures and Couplings Positive Radon measures M + (X): On a metric space (X, d). dµ(x) =f(x)dx µ = P i p i x i X = R

7 Radon Measures and Couplings Positive Radon measures M + (X): On a metric space (X, d). dµ(x) =f(x)dx Probability measures. µ = P i p i x i X = R µ(x) =1 R f =1 P i p i =1

8 Radon Measures and Couplings Positive Radon measures M + (X): dµ(x) =f(x)dx Probability measures. Couplings: Marginals: On a metric space (X, d). µ = P i p i x i X = R µ(x) =1 R f =1 P i p i =1 (µ, ) def. = { 2 M + (X X) ; P 1] = µ, P 2] = } P 1] (S) def. = (S, X) P 2] (S) def. = (X, S) µ = P i p i xi p = P j q j y j q

9 Radon Measures and Couplings Positive Radon measures M + (X): dµ(x) =f(x)dx Probability measures. Couplings: Marginals: µ = P i p i xi On a metric space (X, d). R f =1 P i p i =1 (µ, ) def. = { 2 M + (X X) ; P 1] = µ, P 2] = } p P 1] (S) def. = (S, X) P 2] (S) def. = (X, S) = P j q j q µ = P i p i x i X = R µ(x) =1 y j µ

10 Optimal Transport Ground cost c(x, y) on X X. Optimal transport: [Kantorovitch 1942] min hc, i = Z X X c(x, y)d (x, y) ; 2 (µ, )

11 Optimal Transport Ground cost c(x, y) on X X. Optimal transport: [Kantorovitch 1942] W (µ, ) def. = min hc, i = Z X X c(x, y)d (x, y) ; 2 (µ, ) For c(x, y) =d(x, y), -Wasserstein distance W.

12 Optimal Transport Ground cost c(x, y) on X X. Optimal transport: [Kantorovitch 1942] W (µ, ) def. = min hc, i = Z X X c(x, y)d (x, y) ; 2 (µ, ) For c(x, y) =d(x, y), -Wasserstein distance W. Linear programming: µ = P N 1 i=1 p i x i, = P N 2 j=1 p j y i

13 Optimal Transport Ground cost c(x, y) on X X. Optimal transport: [Kantorovitch 1942] W (µ, ) def. = min hc, i = Z X X c(x, y)d (x, y) ; 2 (µ, ) For c(x, y) =d(x, y), -Wasserstein distance W. Linear programming: µ = P N 1 i=1 p i x i, = P N 2 j=1 p j y i Hungarian/Auction: O(N 3 ) µ = 1 N P N i=1 x i, = 1 N P N j=1 y j

14 Optimal Transport Ground cost c(x, y) on X X. Optimal transport: [Kantorovitch 1942] W (µ, ) def. = min hc, i = Z X X c(x, y)d (x, y) ; 2 (µ, ) For c(x, y) =d(x, y), -Wasserstein distance W. Linear programming: µ = P N 1 i=1 p i x i, = P N 2 Hungarian/Auction: also reveal that the network simplex behaves in O(n 2 ) in our context, which is a major gain at the scale at which we typical work, i.e. thousands of particles. This finding is also useful for applica- j=1 p j y i tions that use EMD, where using the network simplex instead of the transport simplex can bring a significant performance increase. Our experiments also show that fixed-point precision O(N 3 further speeds up the computation. We observed that the value ) of the final transport cost is less accurate because of the limited precision, but that the particle pairing that produces the actual interpolation scheme remains unchanged. We used the fixed point method to generate µ = 1 N P N i=1 x i, = 1 N P N j=1 y j the results presented in this paper. The results of the performance study are also of broader interest, as current EMD image retrieval or color transfer techniques rely on slower solvers [Rubner et al. 2000; Kanters et al. 2003; Morovic and Sun 2003]. Monge-Ampère/Benamou-Brenier, d = 2 2. Time in seconds Network Simplex (fixed point) Network Simplex (double prec.) Transport Simplex y = α x 2 y = β x 3 Figure 7: Synthetic 2D examples on a Euclidean domain. Th left and right columns show the input distributions, while the cent columns show interpolations for =1/4, =1/2, and =3/ 10 4 positive. A negative side effect of this choice is that interpolatin

15 Optimal Transport Ground cost c(x, y) on X X. Optimal transport: [Kantorovitch 1942] Z def. W (µ, ) = min hc, i = X X c(x, y)d (x, y) ; 2 (µ, ) For c(x, y) = d(x, y), -Wasserstein distance W. LinearPprogramming: P N1 N2 µ = i=1 pi xi, = j=1 pj yi A NUMERICAL ALGORITHM FOR L2 SEMI-DISCRETE OPTIMAL TRANSPORT IN 3D 21 A NUMERICAL ALGORITHM FOR L2 SEMI-DISCRETE A NUMERICAL ALGORITHM FOR L2 SEMI-DISCRETE OPTIMAL TRANSPORT OPTIMAL IN 3D 21 also reveal that the network simplex behaves in O(n2 ) in our context, which is a major gain at the scale at which we typical work, i.e. thousands of particles. This finding is also useful for applications that use EMD, where using the network simplex instead of the transport simplex can bring a significant performance increase. Our experiments also show that fixed-point precision further speeds 3 up the computation. We observed that the value of the final transport cost is less accurate because of the limited precision, but that the particle pairing that produces the actual interpolation scheme remains unchanged. We used the fixed point method to generate the iresults presented in this paper. The results of the j performance study are also of broader interest, as current EMD image retrieval or color transfer techniques rely on slower solvers [Rubner et al. 2000; Kanters et al. 2003; Morovic and Sun 2003]. Hungarian/Auction: PN 1 µ = N i=1 x, = 1 N O(N ) PN j=1 y Monge-Ampe re/benamou-brenier, d = Network Simplex (fixed point) Network Simplex (double prec.) Transport Simplex y = α x2 Time in seconds Semi-discrete: Laguerre cells, d = y = β x3 Figure 7: Synthetic 2D examples on a Euclidean domain. Th left and right columns show the input distributions, while the cent [Levy, 15] columns show interpolations for = 1/4, = 1/2, and = 3/ Figure 9. More examples of9.semi-discrete optimal transport. optimal Figureof More examples of semi-discrete Figure 9. More examples semi-discrete optimal transport. Note how the solids deform and merge to form the sphere ontothe Note how solids and merge form the sph Note how the solids deform and the merge to deform form the sphere on the first row, and how thefirst branches of the star split and merge on thesplit and m row, and how the branches of the star first row, positive. and how the branches of the star split and merge on the A negative side effect of this choice is that interpolatin second row.

16 Optimal Transport Ground cost c(x, y) on X X. Optimal transport: [Kantorovitch 1942] Z def. W (µ, ) = min hc, i = X X c(x, y)d (x, y) ; 2 (µ, ) For c(x, y) = d(x, y), -Wasserstein distance W. LinearPprogramming: P N1 N2 µ = i=1 pi xi, = j=1 pj yi A NUMERICAL ALGORITHM FOR L2 SEMI-DISCRETE OPTIMAL TRANSPORT IN 3D 21 A NUMERICAL ALGORITHM FOR L2 SEMI-DISCRETE A NUMERICAL ALGORITHM FOR L2 SEMI-DISCRETE OPTIMAL TRANSPORT OPTIMAL IN 3D 21 also reveal that the network simplex behaves in O(n2 ) in our context, which is a major gain at the scale at which we typical work, i.e. thousands of particles. This finding is also useful for applications that use EMD, where using the network simplex instead of the transport simplex can bring a significant performance increase. Our experiments also show that fixed-point precision further speeds 3 up the computation. We observed that the value of the final transport cost is less accurate because of the limited precision, but that the particle pairing that produces the actual interpolation scheme remains unchanged. We used the fixed point method to generate the iresults presented in this paper. The results of the j performance study are also of broader interest, as current EMD image retrieval or color transfer techniques rely on slower solvers [Rubner et al. 2000; Kanters et al. 2003; Morovic and Sun 2003]. Hungarian/Auction: PN 1 µ = N i=1 x, = 1 N O(N ) PN j=1 y Monge-Ampe re/benamou-brenier, d = Network Simplex (fixed point) Network Simplex (double prec.) Transport Simplex y = α x2 Time in seconds Semi-discrete: Laguerre cells, d = y = β x3 Figure 7: Synthetic 2D examples on a Euclidean domain. Th left and right columns show the input distributions, while the cent [Levy, 15] columns show interpolations for = 1/4, = 1/2, and = 3/ Need for fast approximate algorithms for generic c Figure 9. More examples of9.semi-discrete optimal transport. optimal Figureof More examples of semi-discrete Figure 9. More examples semi-discrete optimal transport. Note how the solids deform and merge to form the sphere ontothe Note how solids and merge form the sph Note how the solids deform and the merge to deform form the sphere on the first row, and how thefirst branches of the star split and merge on thesplit and m row, and how the branches of the star first row, positive. and how the branches of the star split and merge on the A negative side effect of this choice is that interpolatin second row.

17 Unbalanced Transport (µ, ) 2 M + (X) KL( µ) def. = R X log d dµ dµ + R (dµ d ) X

18 (µ, ) 2 M + (X) Unbalanced Transport KL( µ) def. = R X log d dµ WF c (µ, ) def. =minhc, i + KL(P 1] µ)+ KL(P 2] ) dµ + R (dµ d ) X

19 (µ, ) 2 M + (X) Unbalanced Transport KL( µ) def. = R X log d dµ WF c (µ, ) def. =minhc, i + KL(P 1] µ)+ KL(P 2] ) dµ + R (dµ d ) X µ

20 (µ, ) 2 M + (X) Unbalanced Transport KL( µ) def. = R X log d dµ WF c (µ, ) def. =minhc, i + KL(P 1] µ)+ KL(P 2] ) dµ + R (dµ d ) X Proposition: If c(x, y) = log(cos(min( d(x,y), 2 )) then WF 1/2 is a distance on M + (X). c [Liereo, Mielke, Savaré 2015] [Chizat, Schmitzer, Peyré, Vialard 2015] µ

21 (µ, ) 2 M + (X) Unbalanced Transport WF c (µ, ) def. =minhc, i + KL(P 1] µ)+ KL(P 2] )! Dynamic Benamou-Brenier formulation. [Liereo, Mielke, Savaré 2015] [Chizat, Schmitzer, Peyré, Vialard 2015] µ [Kondratyev, Monsaingeon, Vorotnikov, 2015] KL( µ) def. = R X log d dµ dµ + R (dµ d ) X Proposition: If c(x, y) = log(cos(min( d(x,y), 2 )) then WF 1/2 is a distance on M + (X). c [Liereo, Mielke, Savaré 2015] [Chizat, Schmitzer, Peyré, Vialard 2015]

22 Wasserstein Gradient Flows Implicit Euler step: [Jordan, Kinderlehrer, Otto 1998] µ t+1 =Prox W f (µ t ) def. = argmin µ2m + (X) W (µ t,µ)+ f(µ)

23 Wasserstein Gradient Flows Implicit Euler step: [Jordan, Kinderlehrer, Otto 1998] µ t+1 =Prox W f (µ t ) def. = argmin µ2m + (X) Formal limit! t µ =div(µr(f 0 (µ))) W (µ t,µ)+ f(µ)

24 Wasserstein Gradient Flows Implicit Euler step: [Jordan, Kinderlehrer, Otto 1998] µ t+1 =Prox W f (µ t ) def. = argmin µ2m + (X) Formal limit! t µ =div(µr(f 0 (µ))) f(µ) = R log( dµ dx t µ = µ W (µ t,µ)+ f(µ) (heat di usion)

25 rw Wasserstein Gradient Flows Evolution µ t rw Evolution µ t Implicit Euler step: [Jordan, Kinderlehrer, Otto 1998] µ t+1 =Prox W f (µ t ) def. = argmin µ2m + (X) W (µ t,µ)+ f(µ) Formal limit! t µ =div(µr(f 0 (µ))) f(µ) = R log( dµ dx t µ = µ f(µ) = R t µ =div(µrw) (heat di usion) (advection)

26 rw Wasserstein Gradient Flows Evolution µ t rw Evolution µ t Implicit Euler step: [Jordan, Kinderlehrer, Otto 1998] µ t+1 =Prox W f (µ t ) def. = argmin µ2m + (X) Formal limit! t µ =div(µr(f 0 (µ))) f(µ) = R log( dµ dx t µ = µ W (µ t,µ)+ f(µ) (heat di usion) f(µ) = R t µ =div(µrw) (advection) R f(µ) = 1 (non-linear di usion) m 1 ( dµ dx )m 1 t µ = µ m

27 Overview Transportation-like Problems Regularized Transport Optimal Transport Barycenters Heat Kernel Approximation

28 Transportation-like Problems def. min E( ) = hc, i + f 1 (P 1] )+f 2 (P 2] ) 2M + (X X)

29 Transportation-like Problems def. min E( ) = hc, i + f 1 (P 1] )+f 2 (P 2] ) 2M + (X X) Optimal transport: f 1 = µ f 2 =

30 Transportation-like Problems def. min E( ) = hc, i + f 1 (P 1] )+f 2 (P 2] ) 2M + (X X) Optimal transport: f 1 = µ f 2 = Unbalanced transport: f 1 = KL( µ) f 2 = KL( )

31 Transportation-like Problems def. min E( ) = hc, i + f 1 (P 1] )+f 2 (P 2] ) 2M + (X X) Optimal transport: f 1 = µ f 2 = Unbalanced transport: f 1 = KL( µ) f 2 = KL( ) Gradient flows: f 1 = µt f 2 = f

32 Transportation-like Problems def. min E( ) = hc, i + f 1 (P 1] )+f 2 (P 2] ) 2M + (X X) Optimal transport: f 1 = µ f 2 = Unbalanced transport: f 1 = KL( µ) f 2 = KL( ) Gradient flows: f 1 = µt f 2 = f Regularization: "KL( 0 ) min E( )+"KL( 0 ) Regularization and positivity barrier.

33 Transportation-like Problems def. min E( ) = hc, i + f 1 (P 1] )+f 2 (P 2] ) 2M + (X X) Optimal transport: f 1 = µ f 2 = Unbalanced transport: f 1 = KL( µ) f 2 = KL( ) Gradient flows: f 1 = µt f 2 = f Regularization: "KL( 0 ) min E( )+"KL( 0 ) Regularization and positivity barrier. Discretization grid (prescribed support).

34 Transportation-like Problems def. min E( ) = hc, i + f 1 (P 1] )+f 2 (P 2] ) 2M + (X X) Optimal transport: f 1 = µ f 2 = Unbalanced transport: f 1 = KL( µ) f 2 = KL( ) Gradient flows: f 1 = µt f 2 = f Regularization: Prox KL E/"( 0 ) def. = arg min E( )+"KL( 0 ) Regularization and positivity barrier. "KL( 0 ) Discretization grid (prescribed support). Implicit KL stepping. 0 1 =Prox 1 " E( 0) 2 =Prox 1 " E( 1) argmin E( )

35 def. " = argmin Entropy Regularized Transport {hc, i + "KL( 0 ); 2 (µ, )} " c "

36 def. " = argmin Entropy Regularized Transport Schrödinger s problem: {hc, i + "KL( 0 ); 2 (µ, )} " = argmin 2 (µ, ) KL( K) Gibbs kernel: K(x, y) def. = e c(x,y) " 0 (x, y) Landmark computational paper: [Cuturi 2013]. " c "

37 def. " = argmin Entropy Regularized Transport Schrödinger s problem: {hc, i + "KL( 0 ); 2 (µ, )} " = argmin 2 (µ, ) KL( K) Gibbs kernel: K(x, y) def. = e c(x,y) " 0 (x, y) Landmark computational paper: [Cuturi 2013]. " c Proposition: " "!0! argmin 2 (µ, ) [Carlier, Duval, Peyré, Schmitzer 2015] hc, i " "!+1! µ(x) (y) " µ " "

38 Dykstra-like Iterations Primal: min hc, i + f 1 (P 1] )+f 2 (P 2] )+"KL( 0 )

39 Dykstra-like Iterations Primal: min hc, i + f 1 (P 1] )+f 2 (P 2] )+"KL( 0 ) Dual: max u,v f 1 (u) f 2 (u) "he u ",Ke v " i

40 Dykstra-like Iterations Primal: min hc, i + f 1 (P 1] )+f 2 (P 2] )+"KL( 0 ) Dual: max u,v f 1 (u) f 2 (u) "he u ",Ke v " i (x, y) =a(x)k(x, y)b(y) (a, b) =(e u ",e v " )

41 Dykstra-like Iterations Primal: min hc, i + f 1 (P 1] )+f 2 (P 2] )+"KL( 0 ) Dual: max u,v Block coordinates relaxation: f 1 (u) f 2 (u) "he u ",Ke v " i (x, y) =a(x)k(x, y)b(y) max u f 1 (u) (a, b) =(e u ",e v " ) "he u ",Ke v " i (I u )

42 Dykstra-like Iterations Primal: min hc, i + f 1 (P 1] )+f 2 (P 2] )+"KL( 0 ) Dual: max u,v Block coordinates relaxation: f 1 (u) f 2 (u) "he u ",Ke v " i (x, y) =a(x)k(x, y)b(y) max u max v f 1 (u) (a, b) =(e u ",e v " ) "he u ",Ke v " i (I u ) f 2 (v) "he v ",K e u " i (I v )

43 Dykstra-like Iterations Primal: min hc, i + f 1 (P 1] )+f 2 (P 2] )+"KL( 0 ) Dual: max u,v Block coordinates relaxation: f 1 (u) f 2 (u) "he u ",Ke v " i (x, y) =a(x)k(x, y)b(y) Proposition: max u max v a = ProxKL f 1 /"(Kb) Kb Prox KL f 1 /"(µ) def. f 1 (u) (a, b) =(e u ",e v " ) "he u ",Ke v " i (I u ) f 2 (v) "he v ",K e u " i (I v ) the solutions of (I u ) and (I v ) read: b = ProxKL f 2 /"(K a) K a = argmin f 1 ( )+"KL( µ)

44 Dykstra-like Iterations Primal: min hc, i + f 1 (P 1] )+f 2 (P 2] )+"KL( 0 ) Dual: max u,v Block coordinates relaxation: f 1 (u) f 2 (u) "he u ",Ke v " i (x, y) =a(x)k(x, y)b(y) Proposition: max u max v a = ProxKL f 1 /"(Kb) Kb f 1 (u) (a, b) =(e u ",e v " ) "he u ",Ke v " i (I u ) f 2 (v) "he v ",K e u " i (I v ) the solutions of (I u ) and (I v ) read: b = ProxKL f 2 /"(K a) K a Prox KL f 1 /"(µ) def. = argmin f 1 ( )+"KL( µ)! Only matrix-vector multiplications.! Highly parallelizable.! On regular grids: only convolutions! Linear time iterations.

45 Optimal transport problem: Sinkhorn s Algorithm f 1 = µ f 2 = Prox KL f 1 /"( µ) =µ Prox KL f 2 /"( ) =

46 Optimal transport problem: Sinkhorn s Algorithm f 1 = µ f 2 = Sinkhorn/IPFP algorithm: (`+1) def. a = µ Kb (`) and (`+1) def. b = Prox KL f 1 /"( µ) =µ Prox KL f 2 /"( ) = [Sinkhorn 1967][Deming,Stephan 1940] K a (`+1) (`) " `

47 Optimal transport problem: Sinkhorn s Algorithm Sinkhorn/IPFP algorithm: (`+1) def. a = µ Kb (`) and (`+1) def. b = Proposition: f 1 = µ f 2 = Prox KL f 1 /"( µ) =µ Prox KL f 2 /"( ) = [Sinkhorn 1967][Deming,Stephan 1940] K a (`+1) log( (`) ) log(? ) 1 = O(1 )`, apple 1/" c (`) def. = diag(a (`) )K diag(b (`) ) [Franklin,Lorenz 1989] Local rate: [Knight 2008] (`) " log( log( (`) ) log(? ) 1 ) 0 ` " " `

48 Gradient Flows: Crowd Motion def. µ t+1 = argmin µ W (µ t,µ)+ f(µ) Congestion-inducing function: f(µ) = [0,apple] (µ)+hw, µi [Maury, Roudne -Chupin, Santambrogio 2010]

49 Gradient Flows: Crowd Motion def. µ t+1 = argmin µ W (µ t,µ)+ f(µ) Congestion-inducing function: f(µ) = [0,apple] (µ)+hw, µi [Maury, Roudne -Chupin, Santambrogio 2010] Proposition: Prox 1 " f (µ) =min(e "w µ, apple)

50 Gradient Flows: Crowd Motion def. µ t+1 = argmin µ W (µ t,µ)+ f(µ) Congestion-inducing function: f(µ) = [0,apple] (µ)+hw, µi [Maury, Roudne -Chupin, Santambrogio 2010] Proposition: Prox 1 " f (µ) =min(e "w µ, apple) rw apple = µ t=0 1 apple =2 µ t=0 1 apple =4 µ t=0 1

51 Multiple-Density Gradient Flows (µ 1,t+1,µ 2,t+1 ) def. = argmin (µ 1,µ 2 ) W (µ 1,t,µ 1 )+W (µ 2,t,µ 2 )+ f(µ 1,µ 2 )

52 Multiple-Density Gradient Flows (µ 1,t+1,µ 2,t+1 ) def. = argmin (µ 1,µ 2 ) W (µ 1,t,µ 1 )+W (µ 2,t,µ 2 )+ f(µ 1,µ 2 ) Wasserstein attraction: f(µ 1,µ 2 )=W (µ 1,µ 2 )+h 1 (µ 1 )+h 2 (µ 2 ) rw Example: h i (µ) =hw, µi.

53 Multiple-Density Gradient Flows (µ 1,t+1,µ 2,t+1 ) def. = argmin (µ 1,µ 2 ) W (µ 1,t,µ 1 )+W (µ 2,t,µ 2 )+ f(µ 1,µ 2 ) Wasserstein attraction: f(µ 1,µ 2 )=W (µ 1,µ 2 )+h 1 (µ 1 )+h 2 (µ 2 ) rw Example: h i (µ) =hw, µi.

54 Overview Transportation-like Problems Regularized Transport Optimal Transport Barycenters Heat Kernel Approximation

55 Wasserstein Barycenters P Barycenters of measures (µ k ) k : µ? P 2 argmin k kw (µ k,µ) µ k k =1

56 Wasserstein Barycenters P Barycenters of measures (µ k ) k : µ? P 2 argmin k kw (µ k,µ) µ Generalizes Euclidean barycenter: If µ k = xk then µ? = P k kx k k k =1 µ 1 µ W 2 (µ 1,µ ) W 2 (µ 2,µ ) W 2 (µ 3,µ ) µ 2 µ 3

57 Wasserstein Barycenters P Barycenters of measures (µ k ) k : µ? P 2 argmin k kw (µ k,µ) µ Generalizes Euclidean barycenter: If µ k = xk then µ? = P k kx k k k =1 µ 1 µ W 2 (µ 1,µ ) W 2 (µ 2,µ ) W 2 (µ 3,µ ) µ 2 µ 3

58 Wasserstein Barycenters µ µ2 Mc Cann s displacement interpolation. ) µ1,µ W 2(µ 1 ) ) µ1 µ,µ (µ 3 W2 Generalizes Euclidean barycenter: If µk = xk then µ? = Pk k xk k µ2 =1 2 (µ 2,µ Barycenters of measures (µk )k : k P µ? 2 argmin k k W (µk, µ) W P µ3

59 Wasserstein Barycenters µ µ1,µ W 2(µ 1 ) ) µ1 µ,µ (µ 3 W2 Generalizes Euclidean barycenter: If µk = xk then µ? = Pk k xk ) k µ2 =1 2 (µ 2,µ Barycenters of measures (µk )k : k P µ? 2 argmin k k W (µk, µ) W P µ3 µ2 µ2 Mc Cann s displacement interpolation. µ3 µ1

60 Wasserstein Barycenters µ µ1,µ W 2(µ 1 ) ) µ1 µ,µ (µ 3 W2 Generalizes Euclidean barycenter: If µk = xk then µ? = Pk k xk ) k µ2 =1 2 (µ 2,µ Barycenters of measures (µk )k : k P µ? 2 argmin k k W (µk, µ) W P µ3 µ2 µ2 Mc Cann s displacement interpolation. [Agueh, Carlier, 2010] Theorem: (for c(x, y) = x y 2 ) if µ1 does not vanish on small sets, µ exists and is unique. µ3 µ1

61 min ( k ) k,µ k Regularized Barycenters ( ) X k (hc, k i + "KL( k 0,k )) ; 8k, k 2 (µ k,µ)

62 min ( k ) k,µ k Regularized Barycenters ( ) X k (hc, k i + "KL( k 0,k )) ; 8k, k 2 (µ k,µ)! Need to fix a discretization grid for µ, i.e. choose ( 0,k ) k

63 min ( k ) k,µ k Regularized Barycenters ( ) X k (hc, k i + "KL( k 0,k )) ; 8k, k 2 (µ k,µ)! Need to fix a discretization grid for µ, i.e. choose ( 0,k ) k! Sinkhorn-like algorithm [Benamou, Carlier, Cuturi, Nenna, Peyré, 2015].

64 min ( k ) k,µ Regularized Barycenters ( ) X k (hc, k i + "KL( k 0,k )) ; 8k, k 2 (µ k,µ) k! Need to fix a discretization grid for µ, i.e. choose ( 0,k ) k! Sinkhorn-like algorithm [Benamou, Carlier, Cuturi, Nenna, Peyré, 2015]. [Solomon et al, SIGGRAPH 2015]

65 Color Transfer Input images: (f,g) (chrominance components) Input measures: µ(a) =U(f 1 (A)), (A) =U(g 1 (A)) f µ g

66 Color Transfer Input images: (f,g) (chrominance components) Input measures: µ(a) =U(f 1 (A)), (A) =U(g 1 (A)) f µ g

67 Color Transfer Input images: (f,g) (chrominance components) Input measures: µ(a) =U(f 1 (A)), (A) =U(g 1 (A)) f T T f µ T g g

68 Color Harmonization Raw image sequence

69 Color Harmonization Raw image sequence Compute Wasserstein barycenter Project on the barycenter

70 Overview Transportation-like Problems Regularized Transport Optimal Transport Barycenters Heat Kernel Approximation

71 Optimal Transport on Surfaces Triangulated mesh M. Geodesic distance d M.

72 Optimal Transport on Surfaces Triangulated mesh M. Geodesic distance d M. Ground cost: c(x, y) =d M (x, y). x i d(x i, ) Level sets

73 Optimal Transport on Surfaces Triangulated mesh M. Geodesic distance d M. Ground cost: c(x, y) =d M (x, y). x i d(x i, ) Level sets Computing c (Fast-Marching): N 2 log(n)! too costly.

74 Entropic Transport on Surfaces Heat equation on t u t (x, ) = M u t (x, ), u t=0 (x, ) = x " t

75 Entropic Transport on Surfaces Heat equation on t u t (x, ) = M u t (x, ), u t=0 (x, ) = x Theorem: [Varadhan] " log(u " ) "!0! d 2 M " t

76 Entropic Transport on Surfaces Heat equation on t u t (x, ) = M u t (x, ), u t=0 (x, ) = x Theorem: [Varadhan] " log(u " ) "!0! d 2 M Sinkhorn kernel: K def. = e d 2 M " " u " Id ` M ` " t

77 Barycenter on a Surface µ2 µ

78 Barycenter on a Surface µ2 µ1 0 1 µ2 µ1 µ4 µ3 µ5 µ6 1

79 Barycenter on a Surface µ2 µ1 0 1 µ2 µ1 µ4 µ3 µ5 µ6 = (1,..., 1)/6 1

80 MRI Data Procesing [with A. Gramfort] Ground cost c = d M : geodesic on cortical surface M. L 2 barycenter W 2 2 barycenter

81 Gradient Flows: Crowd Motion with Obstacles M = sub-domain of R 2. apple = µ t=0 1 apple =2 µ t=0 1 apple =4 µ t=0 1 apple =6 µ t=0 1 Potential cos(w)

82 Gradient Flows: Crowd Motion with Obstacles M = sub-domain of R 2. apple = µ t=0 1 apple =2 µ t=0 1 apple =4 µ t=0 1 apple =6 µ t=0 1 Potential cos(w)

83 Crowd Motion on a Surface M = triangulated mesh. apple = µ t=0 1 apple =6 µ t=0 1 Potential cos(w)

84 Crowd Motion on a Surface M = triangulated mesh. apple = µ t=0 1 apple =6 µ t=0 1 Potential cos(w)

85 Histogram features in imaging and machine learning. Conclusion

86 Conclusion Histogram features in imaging and machine learning. Entropic regularization for optimal transport.

87 Conclusion Histogram features in imaging and machine learning. Entropic regularization for optimal transport. Barycenters, unbalanced OT, gradient flows,...

Generative Models and Optimal Transport

Generative Models and Optimal Transport Generative Models and Optimal Transport Marco Cuturi Joint work / work in progress with G. Peyré, A. Genevay (ENS), F. Bach (INRIA), G. Montavon, K-R Müller (TU Berlin) Statistics 0.1 : Density Fitting

More information

softened probabilistic

softened probabilistic Justin Solomon MIT Understand geometry from a softened probabilistic standpoint. Somewhere over here. Exactly here. One of these two places. Query 1 2 Which is closer, 1 or 2? Query 1 2 Which is closer,

More information

Convex discretization of functionals involving the Monge-Ampère operator

Convex discretization of functionals involving the Monge-Ampère operator Convex discretization of functionals involving the Monge-Ampère operator Quentin Mérigot CNRS / Université Paris-Dauphine Joint work with J.D. Benamou, G. Carlier and É. Oudet Workshop on Optimal Transport

More information

Numerical Optimal Transport. Applications

Numerical Optimal Transport. Applications Numerical Optimal Transport http://optimaltransport.github.io Applications Gabriel Peyré www.numerical-tours.com ÉCOLE NORMALE SUPÉRIEURE Grayscale Image Equalization Figure 2.9: 1-D optimal couplings:

More information

Convex discretization of functionals involving the Monge-Ampère operator

Convex discretization of functionals involving the Monge-Ampère operator Convex discretization of functionals involving the Monge-Ampère operator Quentin Mérigot CNRS / Université Paris-Dauphine Joint work with J.D. Benamou, G. Carlier and É. Oudet GeMeCod Conference October

More information

A new Hellinger-Kantorovich distance between positive measures and optimal Entropy-Transport problems

A new Hellinger-Kantorovich distance between positive measures and optimal Entropy-Transport problems A new Hellinger-Kantorovich distance between positive measures and optimal Entropy-Transport problems Giuseppe Savaré http://www.imati.cnr.it/ savare Dipartimento di Matematica, Università di Pavia Nonlocal

More information

Optimal Transport: A Crash Course

Optimal Transport: A Crash Course Optimal Transport: A Crash Course Soheil Kolouri and Gustavo K. Rohde HRL Laboratories, University of Virginia Introduction What is Optimal Transport? The optimal transport problem seeks the most efficient

More information

Optimal mass transport as a distance measure between images

Optimal mass transport as a distance measure between images Optimal mass transport as a distance measure between images Axel Ringh 1 1 Department of Mathematics, KTH Royal Institute of Technology, Stockholm, Sweden. 21 st of June 2018 INRIA, Sophia-Antipolis, France

More information

Lagrangian discretization of incompressibility, congestion and linear diffusion

Lagrangian discretization of incompressibility, congestion and linear diffusion Lagrangian discretization of incompressibility, congestion and linear diffusion Quentin Mérigot Joint works with (1) Thomas Gallouët (2) Filippo Santambrogio Federico Stra (3) Hugo Leclerc Seminaire du

More information

Second order models for optimal transport and cubic splines on the Wasserstein space

Second order models for optimal transport and cubic splines on the Wasserstein space Second order models for optimal transport and cubic splines on the Wasserstein space Jean-David Benamou, Thomas Gallouët, François-Xavier Vialard To cite this version: Jean-David Benamou, Thomas Gallouët,

More information

arxiv: v1 [cs.lg] 13 Nov 2018

arxiv: v1 [cs.lg] 13 Nov 2018 SEMI-DUAL REGULARIZED OPTIMAL TRANSPORT MARCO CUTURI AND GABRIEL PEYRÉ arxiv:1811.05527v1 [cs.lg] 13 Nov 2018 Abstract. Variational problems that involve Wasserstein distances and more generally optimal

More information

A Closed-form Gradient for the 1D Earth Mover s Distance for Spectral Deep Learning on Biological Data

A Closed-form Gradient for the 1D Earth Mover s Distance for Spectral Deep Learning on Biological Data A Closed-form Gradient for the D Earth Mover s Distance for Spectral Deep Learning on Biological Data Manuel Martinez, Makarand Tapaswi, and Rainer Stiefelhagen Karlsruhe Institute of Technology, Karlsruhe,

More information

Optimal Transport in ML

Optimal Transport in ML Optimal Transport in ML Rémi Gilleron Inria Lille & CRIStAL & Univ. Lille Feb. 2018 Main Source also some figures Computational Optimal transport, G. Peyré and M. Cuturi Rémi Gilleron (Magnet) OT Feb.

More information

Optimal transport for Gaussian mixture models

Optimal transport for Gaussian mixture models 1 Optimal transport for Gaussian mixture models Yongxin Chen, Tryphon T. Georgiou and Allen Tannenbaum Abstract arxiv:1710.07876v2 [math.pr] 30 Jan 2018 We present an optimal mass transport framework on

More information

Numerical Optimal Transport. Gromov Wasserstein

Numerical Optimal Transport. Gromov Wasserstein Numerical Optimal Transport http://optimaltransport.github.io Gromov Wasserstein Gabriel Peyré www.numerical-tours.com ÉCOLE NORMALE SUPÉRIEURE Overview Gromov Wasserstein Entropic Regularization Applications

More information

The dynamics of Schrödinger bridges

The dynamics of Schrödinger bridges Stochastic processes and statistical machine learning February, 15, 2018 Plan of the talk The Schrödinger problem and relations with Monge-Kantorovich problem Newton s law for entropic interpolation The

More information

Mathematical Foundations of Data Sciences

Mathematical Foundations of Data Sciences Mathematical Foundations of Data Sciences Gabriel Peyré CNRS & DMA École Normale Supérieure gabriel.peyre@ens.fr https://mathematical-tours.github.io www.numerical-tours.com January 7, 2018 Chapter 18

More information

Gradient Flows: Qualitative Properties & Numerical Schemes

Gradient Flows: Qualitative Properties & Numerical Schemes Gradient Flows: Qualitative Properties & Numerical Schemes J. A. Carrillo Imperial College London RICAM, December 2014 Outline 1 Gradient Flows Models Gradient flows Evolving diffeomorphisms 2 Numerical

More information

a sinkhorn newton method for entropic optimal transport

a sinkhorn newton method for entropic optimal transport a sinkhorn newton method for entropic optimal transport Christoph Brauer Christian Clason Dirk Lorenz Benedikt Wirth February 2, 2018 arxiv:1710.06635v2 [math.oc] 9 Feb 2018 Abstract We consider the entropic

More information

Optimal transport for machine learning

Optimal transport for machine learning Optimal transport for machine learning Rémi Flamary AG GDR ISIS, Sète, 16 Novembre 2017 2 / 37 Collaborators N. Courty A. Rakotomamonjy D. Tuia A. Habrard M. Cuturi M. Perrot C. Févotte V. Emiya V. Seguy

More information

On dissipation distances for reaction-diffusion equations the Hellinger-Kantorovich distance

On dissipation distances for reaction-diffusion equations the Hellinger-Kantorovich distance Weierstrass Institute for! Applied Analysis and Stochastics On dissipation distances for reaction-diffusion equations the Matthias Liero, joint work with Alexander Mielke, Giuseppe Savaré Mohrenstr. 39

More information

Numerical Approximation of L 1 Monge-Kantorovich Problems

Numerical Approximation of L 1 Monge-Kantorovich Problems Rome March 30, 2007 p. 1 Numerical Approximation of L 1 Monge-Kantorovich Problems Leonid Prigozhin Ben-Gurion University, Israel joint work with J.W. Barrett Imperial College, UK Rome March 30, 2007 p.

More information

Generalized incompressible flows, multi-marginal transport and Sinkhorn algorithm arxiv: v2 [math.na] 5 Mar 2018

Generalized incompressible flows, multi-marginal transport and Sinkhorn algorithm arxiv: v2 [math.na] 5 Mar 2018 Generalized incompressible flows, multi-marginal transport and Sinkhorn algorithm arxiv:1710.08234v2 [math.na] 5 Mar 2018 Jean-David Benamou Guillaume Carlier Luca Nenna March 6, 2018 Abstract Starting

More information

arxiv: v1 [math.oc] 27 May 2016

arxiv: v1 [math.oc] 27 May 2016 Stochastic Optimization for Large-scale Optimal Transport arxiv:1605.08527v1 [math.oc] 27 May 2016 Aude Genevay CEREMADE, Univ. Paris-Dauphine INRIA Mokaplan project-team genevay@ceremade.dauphine.fr Gabriel

More information

A shor t stor y on optimal transpor t and its many applications

A shor t stor y on optimal transpor t and its many applications Snapshots of modern mathematics from Oberwolfach 13/2018 A shor t stor y on optimal transpor t and its many applications Filippo Santambrogio We present some examples of optimal transport problems and

More information

arxiv: v1 [math.ap] 19 Sep 2018

arxiv: v1 [math.ap] 19 Sep 2018 Dynamical Optimal Transport on Discrete Surfaces arxiv:1809.07083v1 [math.ap] 19 Sep 2018 HUGO LAVENANT, Université Paris-Sud SEBASTIAN CLAICI, Massachusetts Institute of Technology EDWARD CHIEN, Massachusetts

More information

Fast Computation of Wasserstein Barycenters

Fast Computation of Wasserstein Barycenters Marco Cuturi Graduate School of Informatics, Kyoto University Arnaud Doucet Department of Statistics, University of Oxford MCUTURI@I.KYOTO-U.AC.JP DOUCET@STAT.OXFORD.AC.UK Abstract We present new algorithms

More information

Optimal Transport for Diffeomorphic Registration

Optimal Transport for Diffeomorphic Registration Optimal Transport for Diffeomorphic Registration Jean Feydy, Benjamin Charlier, François-Xavier Vialard, Gabriel Peyré To cite this version: Jean Feydy, Benjamin Charlier, François-Xavier Vialard, Gabriel

More information

arxiv: v4 [math.oc] 28 Aug 2017

arxiv: v4 [math.oc] 28 Aug 2017 GENERALIZED SINKHORN ITERATIONS FOR REGULARIZING INVERSE PROBLEMS USING OPTIMAL MASS TRANSPORT JOHAN KARLSSON AND AXEL RINGH arxiv:1612.02273v4 [math.oc] 28 Aug 2017 Abstract. The optimal mass transport

More information

Stochastic Optimization for Large-scale Optimal Transport

Stochastic Optimization for Large-scale Optimal Transport Stochastic Optimization for Large-scale Optimal Transport Aude Genevay, Marco Cuturi, Gabriel Peyré, Francis Bach To cite this version: Aude Genevay, Marco Cuturi, Gabriel Peyré, Francis Bach. Stochastic

More information

Approximations of displacement interpolations by entropic interpolations

Approximations of displacement interpolations by entropic interpolations Approximations of displacement interpolations by entropic interpolations Christian Léonard Université Paris Ouest Mokaplan 10 décembre 2015 Interpolations in P(X ) X : Riemannian manifold (state space)

More information

ITERATIVE BREGMAN PROJECTIONS FOR REGULARIZED TRANSPORTATION PROBLEMS

ITERATIVE BREGMAN PROJECTIONS FOR REGULARIZED TRANSPORTATION PROBLEMS ITERATIVE BREGMAN PROJECTIONS FOR REGULARIZED TRANSPORTATION PROBLEMS JEAN-DAVID BENAMOU, GUILLAUME CARLIER, MARCO CUTURI, LUCA NENNA, AND GABRIEL PEYRÉ Abstract. This article details a general numerical

More information

Some geometric consequences of the Schrödinger problem

Some geometric consequences of the Schrödinger problem Some geometric consequences of the Schrödinger problem Christian Léonard To cite this version: Christian Léonard. Some geometric consequences of the Schrödinger problem. Second International Conference

More information

Numerical Methods for Optimal Transportation 13frg167

Numerical Methods for Optimal Transportation 13frg167 Numerical Methods for Optimal Transportation 13frg167 Jean-David Benamou (INRIA) Adam Oberman (McGill University) Guillaume Carlier (Paris Dauphine University) July 2013 1 Overview of the Field Optimal

More information

arxiv: v2 [stat.ml] 24 Aug 2015

arxiv: v2 [stat.ml] 24 Aug 2015 A SMOOTHED DUAL APPROACH FOR VARIATIONAL WASSERSTEIN PROBLEMS MARCO CUTURI AND GABRIEL PEYRÉ arxiv:1503.02533v2 [stat.ml] 24 Aug 2015 Abstract. Variational problems that involve Wasserstein distances have

More information

Wasserstein Training of Boltzmann Machines

Wasserstein Training of Boltzmann Machines Wasserstein Training of Boltzmann Machines Grégoire Montavon, Klaus-Rober Muller, Marco Cuturi Presenter: Shiyu Liang December 1, 2016 Coordinated Science Laboratory Department of Electrical and Computer

More information

Least Squares and Linear Systems

Least Squares and Linear Systems Least Squares and Linear Systems Gabriel Peyré www.numerical-tours.com ÉCOLE NORMALE SUPÉRIEURE s=3 s=6 0.5 0.5 0 0 0.5 0.5 https://mathematical-coffees.github.io 1 10 20 30 40 50 Organized by: Mérouane

More information

Free Energy, Fokker-Planck Equations, and Random walks on a Graph with Finite Vertices

Free Energy, Fokker-Planck Equations, and Random walks on a Graph with Finite Vertices Free Energy, Fokker-Planck Equations, and Random walks on a Graph with Finite Vertices Haomin Zhou Georgia Institute of Technology Jointly with S.-N. Chow (Georgia Tech) Wen Huang (USTC) Yao Li (NYU) Research

More information

Discrete Ricci curvature via convexity of the entropy

Discrete Ricci curvature via convexity of the entropy Discrete Ricci curvature via convexity of the entropy Jan Maas University of Bonn Joint work with Matthias Erbar Simons Institute for the Theory of Computing UC Berkeley 2 October 2013 Starting point McCann

More information

Optimal Transport for Applied Mathematicians

Optimal Transport for Applied Mathematicians Optimal Transport for Applied Mathematicians Calculus of Variations, PDEs and Modelling Filippo Santambrogio 1 1 Laboratoire de Mathématiques d Orsay, Université Paris Sud, 91405 Orsay cedex, France filippo.santambrogio@math.u-psud.fr,

More information

Superhedging and distributionally robust optimization with neural networks

Superhedging and distributionally robust optimization with neural networks Superhedging and distributionally robust optimization with neural networks Stephan Eckstein joint work with Michael Kupper and Mathias Pohl Robust Techniques in Quantitative Finance University of Oxford

More information

A DIFFUSIVE MODEL FOR MACROSCOPIC CROWD MOTION WITH DENSITY CONSTRAINTS

A DIFFUSIVE MODEL FOR MACROSCOPIC CROWD MOTION WITH DENSITY CONSTRAINTS A DIFFUSIVE MODEL FOR MACROSCOPIC CROWD MOTION WITH DENSITY CONSTRAINTS Abstract. In the spirit of the macroscopic crowd motion models with hard congestion (i.e. a strong density constraint ρ ) introduced

More information

Batch, Stochastic and Mirror Gradient Descents

Batch, Stochastic and Mirror Gradient Descents Batch, Stochastic and Mirror Gradient Descents Gabriel Peyré www.numerical-tours.com ÉCOLE NORMALE SUPÉRIEURE s=3 s=6 0.5 0.5 0 0 0.5 0.5 https://mathematical-coffees.github.io 1 10 20 30 40 50 Organized

More information

Computing Kantorovich-Wasserstein Distances on d-dimensional histograms using (d + 1)-partite graphs

Computing Kantorovich-Wasserstein Distances on d-dimensional histograms using (d + 1)-partite graphs Computing Kantorovich-Wasserstein Distances on d-dimensional histograms using (d + 1)-partite graphs Gennaro Auricchio, Stefano Gualandi, Marco Veneroni Università degli Studi di Pavia, Dipartimento di

More information

Optimal Transport and Wasserstein Distance

Optimal Transport and Wasserstein Distance Optimal Transport and Wasserstein Distance The Wasserstein distance which arises from the idea of optimal transport is being used more and more in Statistics and Machine Learning. In these notes we review

More information

Jean-David Benamou 1, Guillaume Carlier 2 and Maxime Laborde 3

Jean-David Benamou 1, Guillaume Carlier 2 and Maxime Laborde 3 ESAIM: PROCEEDINGS AND SURVEYS, June 6, Vol. 54, p. -7 B. Düring, C.-B. Schönlieb and M.-T. Wolfram, Editors AN AUGMENTED LAGRANGIAN APPROACH TO WASSERSTEIN GRADIENT FLOWS AND APPLICATIONS Jean-David Benamou,

More information

Nesterov s Optimal Gradient Methods

Nesterov s Optimal Gradient Methods Yurii Nesterov http://www.core.ucl.ac.be/~nesterov Nesterov s Optimal Gradient Methods Xinhua Zhang Australian National University NICTA 1 Outline The problem from machine learning perspective Preliminaries

More information

Optimal transport for Seismic Imaging

Optimal transport for Seismic Imaging Optimal transport for Seismic Imaging Bjorn Engquist In collaboration with Brittany Froese and Yunan Yang ICERM Workshop - Recent Advances in Seismic Modeling and Inversion: From Analysis to Applications,

More information

Characterization of barycenters in the Wasserstein space by averaging optimal transport maps

Characterization of barycenters in the Wasserstein space by averaging optimal transport maps Characterization of barycenters in the Wasserstein space by averaging optimal transport maps Jérémie Bigot, Thierry Klein To cite this version: Jérémie Bigot, Thierry Klein. Characterization of barycenters

More information

Sharp rates of convergence of empirical measures in Wasserstein distance

Sharp rates of convergence of empirical measures in Wasserstein distance Sharp rates of convergence of empirical measures in Wasserstein distance Francis Bach INRIA - Ecole Normale Supérieure ÉCOLE NORMALE SUPÉRIEURE Joint work with Jonathan Weed (MIT) NIPS Workshop, December

More information

Sliced Wasserstein Barycenter of Multiple Densities

Sliced Wasserstein Barycenter of Multiple Densities Sliced Wasserstein Barycenter of Multiple Densities The Harvard community has made this article openly available. Please share how this access benefits you. Your story matters Citation Bonneel, Nicolas

More information

Logarithmic Sobolev Inequalities

Logarithmic Sobolev Inequalities Logarithmic Sobolev Inequalities M. Ledoux Institut de Mathématiques de Toulouse, France logarithmic Sobolev inequalities what they are, some history analytic, geometric, optimal transportation proofs

More information

New Trends in Optimal Transport

New Trends in Optimal Transport New Trends in Optimal Transport 2 6 March 2015 Abstracts Luigi Ambrosio (SNS Pisa) Extended metric measure spaces with Ricci lower bounds I will report on some work in progress with M.Erbar and G.Savare.

More information

Penalized Barycenters in the Wasserstein space

Penalized Barycenters in the Wasserstein space Penalized Barycenters in the Wasserstein space Elsa Cazelles, joint work with Jérémie Bigot & Nicolas Papadakis Université de Bordeaux & CNRS Journées IOP - Du 5 au 8 Juillet 2017 Bordeaux Elsa Cazelles

More information

Minimal time mean field games

Minimal time mean field games based on joint works with Samer Dweik and Filippo Santambrogio PGMO Days 2017 Session Mean Field Games and applications EDF Lab Paris-Saclay November 14th, 2017 LMO, Université Paris-Sud Université Paris-Saclay

More information

Large Scale Semi-supervised Linear SVMs. University of Chicago

Large Scale Semi-supervised Linear SVMs. University of Chicago Large Scale Semi-supervised Linear SVMs Vikas Sindhwani and Sathiya Keerthi University of Chicago SIGIR 2006 Semi-supervised Learning (SSL) Motivation Setting Categorize x-billion documents into commercial/non-commercial.

More information

Nonparametric regression for topology. applied to brain imaging data

Nonparametric regression for topology. applied to brain imaging data , applied to brain imaging data Cleveland State University October 15, 2010 Motivation from Brain Imaging MRI Data Topology Statistics Application MRI Data Topology Statistics Application Cortical surface

More information

Regularity for the optimal transportation problem with Euclidean distance squared cost on the embedded sphere

Regularity for the optimal transportation problem with Euclidean distance squared cost on the embedded sphere Regularity for the optimal transportation problem with Euclidean distance squared cost on the embedded sphere Jun Kitagawa and Micah Warren January 6, 011 Abstract We give a sufficient condition on initial

More information

arxiv: v1 [math.ap] 10 Oct 2013

arxiv: v1 [math.ap] 10 Oct 2013 The Exponential Formula for the Wasserstein Metric Katy Craig 0/0/3 arxiv:30.292v math.ap] 0 Oct 203 Abstract We adapt Crandall and Liggett s method from the Banach space case to give a new proof of the

More information

A primal-dual approach for a total variation Wasserstein flow

A primal-dual approach for a total variation Wasserstein flow A primal-dual approach for a total variation Wasserstein flow Martin Benning 1, Luca Calatroni 2, Bertram Düring 3, Carola-Bibiane Schönlieb 4 1 Magnetic Resonance Research Centre, University of Cambridge,

More information

arxiv: v1 [cs.cv] 5 Oct 2016

arxiv: v1 [cs.cv] 5 Oct 2016 Convex Histogram-Based Joint Image Segmentation with Regularized Optimal Transport Cost Nicolas Papadakis nicolas.papadakis@math.u-bordeaux.fr IMB, UMR 525, Université de Bordeaux, F-33400 Talence, France

More information

Rigorous approximation of invariant measures for IFS Joint work April 8, with 2016 S. 1 Galat / 21

Rigorous approximation of invariant measures for IFS Joint work April 8, with 2016 S. 1 Galat / 21 Rigorous approximation of invariant measures for IFS Joint work with S. Galatolo e I. Nisoli Maurizio Monge maurizio.monge@im.ufrj.br Universidade Federal do Rio de Janeiro April 8, 2016 Rigorous approximation

More information

NEW FUNCTIONAL INEQUALITIES

NEW FUNCTIONAL INEQUALITIES 1 / 29 NEW FUNCTIONAL INEQUALITIES VIA STEIN S METHOD Giovanni Peccati (Luxembourg University) IMA, Minneapolis: April 28, 2015 2 / 29 INTRODUCTION Based on two joint works: (1) Nourdin, Peccati and Swan

More information

A Semi-Lagrangian discretization of non linear fokker Planck equations

A Semi-Lagrangian discretization of non linear fokker Planck equations A Semi-Lagrangian discretization of non linear fokker Planck equations E. Carlini Università Sapienza di Roma joint works with F.J. Silva RICAM, Linz November 2-25, 206 / 3 Outline A Semi-Lagrangian scheme

More information

Optimal transportation and optimal control in a finite horizon framework

Optimal transportation and optimal control in a finite horizon framework Optimal transportation and optimal control in a finite horizon framework Guillaume Carlier and Aimé Lachapelle Université Paris-Dauphine, CEREMADE July 2008 1 MOTIVATIONS - A commitment problem (1) Micro

More information

A Modest Proposal for MFG with Density Constraints

A Modest Proposal for MFG with Density Constraints A Modest Proposal for MFG with Density Constraints Mean Field Games and related topics, Rome, May 011 Filippo Santambrogio October 31, 011 Abstract We consider a typical problem in Mean Field Games: the

More information

AN ELEMENTARY PROOF OF THE TRIANGLE INEQUALITY FOR THE WASSERSTEIN METRIC

AN ELEMENTARY PROOF OF THE TRIANGLE INEQUALITY FOR THE WASSERSTEIN METRIC PROCEEDINGS OF THE AMERICAN MATHEMATICAL SOCIETY Volume 136, Number 1, January 2008, Pages 333 339 S 0002-9939(07)09020- Article electronically published on September 27, 2007 AN ELEMENTARY PROOF OF THE

More information

Anisotropic congested transport

Anisotropic congested transport Anisotropic congested transport Lorenzo Brasco LATP Aix-Marseille Université lbrasco@gmail.com http://www.latp.univ-mrs.fr/~brasco/ Sankt Peterburg, 04/06/2012 References Part of the results here presented

More information

Approximating Riemannian Voronoi Diagrams and Delaunay triangulations.

Approximating Riemannian Voronoi Diagrams and Delaunay triangulations. Approximating Riemannian Voronoi Diagrams and Delaunay triangulations. Jean-Daniel Boissonnat, Mael Rouxel-Labbé and Mathijs Wintraecken Boissonnat, Rouxel-Labbé, Wintraecken Approximating Voronoi Diagrams

More information

Joint distribution optimal transportation for domain adaptation

Joint distribution optimal transportation for domain adaptation Joint distribution optimal transportation for domain adaptation Changhuang Wan Mechanical and Aerospace Engineering Department The Ohio State University March 8 th, 2018 Joint distribution optimal transportation

More information

CONGESTED AGGREGATION VIA NEWTONIAN INTERACTION

CONGESTED AGGREGATION VIA NEWTONIAN INTERACTION CONGESTED AGGREGATION VIA NEWTONIAN INTERACTION KATY CRAIG, INWON KIM, AND YAO YAO Abstract. We consider a congested aggregation model that describes the evolution of a density through the competing effects

More information

Introduction to Optimal Transport Theory

Introduction to Optimal Transport Theory Introduction to Optimal Transport Theory Filippo Santambrogio Grenoble, June 15th 2009 These very short lecture notes do not want to be an exhaustive presentation of the topic, but only a short list of

More information

Displacement convexity of the relative entropy in the discrete h

Displacement convexity of the relative entropy in the discrete h Displacement convexity of the relative entropy in the discrete hypercube LAMA Université Paris Est Marne-la-Vallée Phenomena in high dimensions in geometric analysis, random matrices, and computational

More information

arxiv: v2 [math.ap] 23 Apr 2014

arxiv: v2 [math.ap] 23 Apr 2014 Multi-marginal Monge-Kantorovich transport problems: A characterization of solutions arxiv:1403.3389v2 [math.ap] 23 Apr 2014 Abbas Moameni Department of Mathematics and Computer Science, University of

More information

Optimal Transport for Data Analysis

Optimal Transport for Data Analysis Optimal Transport for Data Analysis Bernhard Schmitzer 2017-05-30 1 Introduction 1.1 Reminders on Measure Theory Reference: Ambrosio, Fusco, Pallara: Functions of Bounded Variation and Free Discontinuity

More information

EE613 Machine Learning for Engineers. Kernel methods Support Vector Machines. jean-marc odobez 2015

EE613 Machine Learning for Engineers. Kernel methods Support Vector Machines. jean-marc odobez 2015 EE613 Machine Learning for Engineers Kernel methods Support Vector Machines jean-marc odobez 2015 overview Kernel methods introductions and main elements defining kernels Kernelization of k-nn, K-Means,

More information

Integral operators. Jordan Bell Department of Mathematics, University of Toronto. April 22, X F x(y)dµ(y) x N c 1 0 x N 1

Integral operators. Jordan Bell Department of Mathematics, University of Toronto. April 22, X F x(y)dµ(y) x N c 1 0 x N 1 Integral operators Jordan Bell jordan.bell@gmail.com Department of Mathematics, University of Toronto April 22, 2016 1 Product measures Let (, A, µ be a σ-finite measure space. Then with A A the product

More information

Optimal Transport for Data Analysis

Optimal Transport for Data Analysis Optimal Transport for Data Analysis Bernhard Schmitzer 2017-05-16 1 Introduction 1.1 Reminders on Measure Theory Reference: Ambrosio, Fusco, Pallara: Functions of Bounded Variation and Free Discontinuity

More information

Optimal Transport Methods in Operations Research and Statistics

Optimal Transport Methods in Operations Research and Statistics Optimal Transport Methods in Operations Research and Statistics Jose Blanchet (based on work with F. He, Y. Kang, K. Murthy, F. Zhang). Stanford University (Management Science and Engineering), and Columbia

More information

Transport between RGB Images Motivated by Dynamic Optimal Transport

Transport between RGB Images Motivated by Dynamic Optimal Transport Transport between RGB Images Motivated by Dynamic Optimal Transport Jan Henrik Fitschen, Friederike Laus and Gabriele Steidl Department of Mathematics, University of Kaiserslautern, Germany {fitschen,

More information

Lecture 9: PGM Learning

Lecture 9: PGM Learning 13 Oct 2014 Intro. to Stats. Machine Learning COMP SCI 4401/7401 Table of Contents I Learning parameters in MRFs 1 Learning parameters in MRFs Inference and Learning Given parameters (of potentials) and

More information

From slow diffusion to a hard height constraint: characterizing congested aggregation

From slow diffusion to a hard height constraint: characterizing congested aggregation 3 2 1 Time Position -0.5 0.0 0.5 0 From slow iffusion to a har height constraint: characterizing congeste aggregation Katy Craig University of California, Santa Barbara BIRS April 12, 2017 2 collective

More information

arxiv: v1 [cs.sy] 14 Apr 2013

arxiv: v1 [cs.sy] 14 Apr 2013 Matrix-valued Monge-Kantorovich Optimal Mass Transport Lipeng Ning, Tryphon T. Georgiou and Allen Tannenbaum Abstract arxiv:34.393v [cs.sy] 4 Apr 3 We formulate an optimal transport problem for matrix-valued

More information

CS598 Machine Learning in Computational Biology (Lecture 5: Matrix - part 2) Professor Jian Peng Teaching Assistant: Rongda Zhu

CS598 Machine Learning in Computational Biology (Lecture 5: Matrix - part 2) Professor Jian Peng Teaching Assistant: Rongda Zhu CS598 Machine Learning in Computational Biology (Lecture 5: Matrix - part 2) Professor Jian Peng Teaching Assistant: Rongda Zhu Feature engineering is hard 1. Extract informative features from domain knowledge

More information

arxiv: v1 [math.ap] 3 Sep 2014

arxiv: v1 [math.ap] 3 Sep 2014 L 2 A NUMERICAL ALGORITHM FOR SEMI-DISCRETE OPTIMAL TRANSPORT IN 3D BRUNO LÉVY arxiv:1409.1279v1 [math.ap] 3 Sep 2014 Abstract. This paper introduces a numerical algorithm to compute the L 2 optimal transport

More information

Unraveling the mysteries of stochastic gradient descent on deep neural networks

Unraveling the mysteries of stochastic gradient descent on deep neural networks Unraveling the mysteries of stochastic gradient descent on deep neural networks Pratik Chaudhari UCLA VISION LAB 1 The question measures disagreement of predictions with ground truth Cat Dog... x = argmin

More information

Numerical Optimization Techniques

Numerical Optimization Techniques Numerical Optimization Techniques Léon Bottou NEC Labs America COS 424 3/2/2010 Today s Agenda Goals Representation Capacity Control Operational Considerations Computational Considerations Classification,

More information

Information geometry connecting Wasserstein distance and Kullback Leibler divergence via the entropy-relaxed transportation problem

Information geometry connecting Wasserstein distance and Kullback Leibler divergence via the entropy-relaxed transportation problem https://doi.org/10.1007/s41884-018-0002-8 RESEARCH PAPER Information geometry connecting Wasserstein distance and Kullback Leibler divergence via the entropy-relaxed transportation problem Shun-ichi Amari

More information

Introduction to optimal transport

Introduction to optimal transport Introduction to optimal transport Nicola Gigli May 20, 2011 Content Formulation of the transport problem The notions of c-convexity and c-cyclical monotonicity The dual problem Optimal maps: Brenier s

More information

A description of transport cost for signed measures

A description of transport cost for signed measures A description of transport cost for signed measures Edoardo Mainini Abstract In this paper we develop the analysis of [AMS] about the extension of the optimal transport framework to the space of real measures.

More information

Entropic and Displacement Interpolation: Tryphon Georgiou

Entropic and Displacement Interpolation: Tryphon Georgiou Entropic and Displacement Interpolation: Schrödinger bridges, Monge-Kantorovich optimal transport, and a Computational Approach Based on the Hilbert Metric Tryphon Georgiou University of Minnesota IMA

More information

4. Ergodicity and mixing

4. Ergodicity and mixing 4. Ergodicity and mixing 4. Introduction In the previous lecture we defined what is meant by an invariant measure. In this lecture, we define what is meant by an ergodic measure. The primary motivation

More information

On the interior of the simplex, we have the Hessian of d(x), Hd(x) is diagonal with ith. µd(w) + w T c. minimize. subject to w T 1 = 1,

On the interior of the simplex, we have the Hessian of d(x), Hd(x) is diagonal with ith. µd(w) + w T c. minimize. subject to w T 1 = 1, Math 30 Winter 05 Solution to Homework 3. Recognizing the convexity of g(x) := x log x, from Jensen s inequality we get d(x) n x + + x n n log x + + x n n where the equality is attained only at x = (/n,...,

More information

arxiv: v1 [math.ca] 20 Sep 2010

arxiv: v1 [math.ca] 20 Sep 2010 Introduction to Optimal Transport Theory arxiv:1009.3856v1 [math.ca] 20 Sep 2010 Filippo Santambrogio Grenoble, June 15th, 2009 Revised version : March 23rd, 2010 Abstract These notes constitute a sort

More information

A Kernel on Persistence Diagrams for Machine Learning

A Kernel on Persistence Diagrams for Machine Learning A Kernel on Persistence Diagrams for Machine Learning Jan Reininghaus 1 Stefan Huber 1 Roland Kwitt 2 Ulrich Bauer 1 1 Institute of Science and Technology Austria 2 FB Computer Science Universität Salzburg,

More information

Gradient Flow. Chang Liu. April 24, Tsinghua University. Chang Liu (THU) Gradient Flow April 24, / 91

Gradient Flow. Chang Liu. April 24, Tsinghua University. Chang Liu (THU) Gradient Flow April 24, / 91 Gradient Flow Chang Liu Tsinghua University April 24, 2017 Chang Liu (THU) Gradient Flow April 24, 2017 1 / 91 Contents 1 Introduction 2 Gradient flow in the Euclidean space Variants of Gradient Flow in

More information

Transport-based analysis, modeling, and learning from signal and data distributions

Transport-based analysis, modeling, and learning from signal and data distributions 1 Transport-based analysis, modeling, and learning from signal and data distributions Soheil Kolouri, Serim Park, Matthew Thorpe, Dejan Slepčev, Gustavo K. Rohde arxiv:1609.04767v1 [cs.cv] 15 Sep 2016

More information

Downloaded 03/07/17 to Redistribution subject to SIAM license or copyright; see

Downloaded 03/07/17 to Redistribution subject to SIAM license or copyright; see SIAM J. MATH.ANAL. Vol. 48, No. 4, pp. 2869 2911 c 216 Society for Industrial and Applied Mathematics OPTIMAL TRANSPORT IN COMPETITION WITH REACTION: THE HELLINGER KANTOROVICH DISTANCE AND GEODESIC CURVES

More information

Random walks for deformable image registration Dana Cobzas

Random walks for deformable image registration Dana Cobzas Random walks for deformable image registration Dana Cobzas with Abhishek Sen, Martin Jagersand, Neil Birkbeck, Karteek Popuri Computing Science, University of Alberta, Edmonton Deformable image registration?t

More information

CS 468, Spring 2013 Differential Geometry for Computer Science Justin Solomon and Adrian Butscher

CS 468, Spring 2013 Differential Geometry for Computer Science Justin Solomon and Adrian Butscher http://igl.ethz.ch/projects/arap/arap_web.pdf CS 468, Spring 2013 Differential Geometry for Computer Science Justin Solomon and Adrian Butscher Homework 4: June 5 Project: June 6 Scribe notes: One week

More information