optimal mass transport and density flows
play

Optimal mass transport and density flows Tryphon Georgiou - PowerPoint PPT Presentation

Optimal mass transport and density flows Tryphon Georgiou Mechanical & Aerospace Engineering University of California, Irvine Joint work with Yongxin Chen (MSKCC), M. Pavon (Padova) , A. Tannenbaum (Stony Brook), and Wilfrid Gangbo (UCLA)


  1. Optimal mass transport and density flows Tryphon Georgiou Mechanical & Aerospace Engineering University of California, Irvine Joint work with Yongxin Chen (MSKCC), M. Pavon (Padova) , A. Tannenbaum (Stony Brook), and Wilfrid Gangbo (UCLA) LCCC, Lund June 13, 2017 Supported by the NSF & AFOSR

  2. Plan of the talk: – Nexus of ideas: Mass Transport ⇔ Schr¨ odinger bridges ⇔ Stochastic control with a bit on LQG, Riccati, etc. – Discrete-space counterpart: Markov chains and networks – Non-commutative counterpart: Quantum flows & non-commutative geometry

  3. Density flows

  4. Optimal Mass Transport (OMT) Gaspard Monge 1781 Leonid Kantorovich 1976 Work in early 1940’s, Nobel 1975 CIA file on Kantorovich (wikipedia)

  5. Monge’s formulation Le m´ emoire sur les d´ eblais et les remblais Gaspard Monge 1781 � � 2 dµ ( x ) � x − T ( x ) inf T � �� � y where T # µ = ν

  6. Kantorovich’s formulation �� � x − y � 2 dπ ( x, y ) 3 3 inf 2.5 2.5 π ∈ Π( ρ 0 ,ρ 1 ) 2 2 1.5 1.5 1 1 0.5 0.5 0 0 0 1 2 0 1 2 3 where Π( µ, ν ) are “couplings”: 6 5 4 � y π ( dx, dy ) = ρ 0 ( x ) dx = dµ ( x ) 3 2 1 0 � 0 1 2 3 x π ( dx, dy ) = ρ 1 ( y ) dy = dν ( y ) .

  7. B&B’s fluid dynamic formulation Benamou and Brenier (2000): � 1 � � v ( x, t ) � 2 ρ ( x, t ) dtdx inf R n ( ρ,v ) 0 ∂ρ ∂t + ∇ · ( vρ ) = 0 ρ ( x, 0) = ρ 0 ( x ) , ρ ( y, 1) = ρ 1 ( y ) McCann, Gangbo, Otto, Villani, ...

  8. Stochastic control formulation �� 1 � � v ( x, t ) � 2 dt inf v E ρ 0 x ( t ) = v ( x, t ) ˙ x (0) ∼ ρ 0 ( x ) dx x (1) ∼ ρ 1 ( y ) dy

  9. OMT as a control problem – derivation � 1 � x − y � 2 = x � 2 dt, � ˙ inf x ∈X xy 0 X xy = { x ∈ C 1 | x(0) = x, x(1) = y } . Inf attained at constant speed geodesic x ∗ ( t ) = (1 − t ) x + ty

  10. OMT as a control problem – Dirac marginals Also, Inf = any probabilistic average in X xy �� 1 � � x − y � 2 = inf x( t ) � 2 dt � ˙ , E P xy P xy 0 P xy ∈ D ( δ x , δ y ) : prob. measures on C 1 with delta marginals

  11. OMT as a control problem – general marginals �� 1 � � R n × R n � x − y � 2 dπ ( x, y ) = x( t ) � 2 dt � ˙ inf inf . E P π ∈ Π( ρ 0 ,ρ 1 ) P ∈ D ( ρ 0 ,ρ 1 ) 0 ⇒ OMT ≃ stochastic control problem with atypical boundary constraints �� 1 � � v � 2 dt inf v E 0 x (0) ∼ ρ 0 dx, x (1) ∼ ρ 1 dy. x ( t ) = v ( x ( t ) , t ) , ˙ a . s .,

  12. Schr¨ odinger’s Bridges ρ = Ψ ¯ Ψ Ψ t = U ( t )Ψ0 Erwin Schr¨ odinger Work in 1926, Nobel 1935 Bridges 1931/32

  13. Schr¨ odinger’s Bridge Problem (SBP) – Cloud of N independent Brownian particles ( N large) – empirical distr. ρ 0 ( x ) dx and ρ 1 ( y ) dy at t = 0 and t = 1 , resp. – ρ 0 and ρ 1 not compatible with transition mechanism � 1 ρ 1 ( y ) � = p ( t 0 , x, t 1 , y ) ρ 0 ( x ) dx, 0 where � � −| x − y | 2 p ( s, y, t, x ) = [2 π ( t − s )] − n 2 exp , s < t 2( t − s ) Particles have been transported in an unlikely way Schr¨ odinger (1931): Of the many unlikely ways in which this could have happened, which one is the most likely?

  14. Large deviations formulation of SBP � � log dQ Minimize H ( Q, W ) = E Q dW over Q ∈ D ( ρ 0 , ρ 1 ) distributions on paths with marginals ρ ’s H ( · , · ) : relative entropy F¨ ollmer 1988: This is a problem of large deviations of the empirical distribution on path space connected through Sanov’s theorem to a maximum entropy problem .

  15. Relative entropy w.r.t. Wiener measure dX = vdt + dB Girsanov: � � � t � � log dQ 1 � v � 2 ds E Q = E Q dW 2 0 is a quadratic cost!!!

  16. SBP as a stochastic control problem � 1 � � v ( x, t ) � 2 ρ ( x, t ) dtdx, inf R n ( ρ,v ) 0 ∂ρ ∂t + ∇ · ( vρ ) = 1 2∆ ρ ρ ( x, 0) = ρ 0 ( x ) , ρ ( y, 1) = ρ 1 ( y ) . Blaqui` ere, Dai Pra, ... compare with OMT: � 1 � 1 2 � v ( x, t ) � 2 ρ ( x, t ) dtdx inf ( ρ,v ) 0 R n ∂ρ ∂t + ∇ · ( vρ ) = 0 ρ ( x, 0) = ρ 0 ( x ) , ρ ( y, 1) = ρ 1 ( y )

  17. Fluid-dynamic formulation of SBP (time-symmetric) � 1 � � � � v ( x, t ) � 2 + � 1 2 ∇ log ρ ( x, t ) � 2 inf ρ ( x, t ) dtdx, R n ( ρ,v ) 0 ∂ρ ∂t + ∇ · ( vρ ) = 0 , ρ (0 , x ) = ρ 0 ( x ) , ρ (1 , y ) = ρ 1 ( y ) . 2 ∇ log ρ ( x, t ) � 2 : Fisher information , Nelson’s osmotic power � 1 Chen-Georgiou-Pavon, On the relation between optimal transport and Schr¨ odinger bridges: A stochastic control viewpoint, J. Opt. Theory Appl. , 2015 Mikami 2004, Mikami-Thieullen 2006,2008, L´ eonard 2012

  18. Erwin Schr¨ odinger’s insight on SBP the density factors into ρ ( x, t ) = ϕ ( x, t ) ˆ ϕ ( x, t ) where ϕ and ˆ ϕ solve (Schr¨ odinger’s system): � ϕ ( x, t ) = p ( t, x, 1 , y ) ϕ ( y, 1) dy, ϕ ( x, 0) ˆ ϕ ( x, 0) = ρ 0 ( x ) � ϕ ( x, t ) = ˆ p (0 , y, t, x ) ˆ ϕ ( y, 0) dy, ϕ ( x, 1) ˆ ϕ ( x, 1) = ρ 1 ( x ) . compare with Ψ ¯ Ψ = ρ Existence and uniqueness for Schr¨ odinger’s system: Fortet 1940, Beurling 1960, Jamison 1974/75, F¨ ollmer 1988. ∼ Sinkhorn iteration & Quantum version: Georgiou-Pavon 2015

  19. SBP schematic

  20. SBP schematic

  21. SBP schematic

  22. SBP schematic

  23. Schr¨ odinger system − ∂ϕ ∂t ( t, x ) = 1 2 ∆ ϕ ( t, x ) ∂ ˆ ϕ ∂t ( t, x ) = 1 2 ∆ ˆ ϕ ( t, x ) ϕ (0 , x ) ˆ ϕ (0 , x ) = ρ 0 ( x ) ϕ (1 , x ) ˆ ϕ (1 , x ) = ρ 1 ( x )

  24. Existence & uniqueness (Sinkhorn scaling) − ∂ϕ ∂t ( t, x ) = 1 2 ∆ ϕ ( t, x ) ∆ / 2 − − − − − − → ϕ ˆ ϕ ˆ ∂ ˆ ϕ ∂t ( t, x ) = 1 � ↑ 2 ∆ ˆ ϕ ( t, x ) � ρ 0 � ρ 1 � ↓ · · − ∆ / 2 ϕ (0 , x ) ˆ ϕ (0 , x ) = ρ 0 ( x ) ← − − − − − − − ϕ ϕ ϕ (1 , x ) ˆ ϕ (1 , x ) = ρ 1 ( x ) iteration is contractive in the Hilbert metric! d H ( p, q ) = log M ( p, q ) m ( p, q ) inf { λ | p ≤ λq } M ( p, q ) := sup { λ | λq ≤ p } m ( p, q ) := Chen-Georgiou-Pavon, Entropic and displacement interpolation: a computational approach using the Hilbert metric, SIAM J. Appl. Math.

  25. OMT as limit to SBP: numerics in general OMT interpolation: Marginal distributions ρ t + ∇ · ρv = 0 ρ t + ∇ · ρv = ǫ ∆ ρ , varying ǫ

  26. Applications: Image interpolation Interpolation of 2D images to a 3D model:

  27. LQG - covariance control �� T � � u ( t ) � 2 dt min , E u 0 s.t. dX = AXdt + Budt + BdW X (0) ∼ N (0 , Σ 0 ) , X ( T ) ∼ N (0 , Σ 1 ) ⇐ these are the ρ ’s Beghi (1996), Grigoriadis- Skelton (1997) Brockett (2007, 2012), Vladimirov-Petersen (2010, 2015)

  28. Bridges - LQG - covariance control in general �� T � � u ( t ) � 2 dt min , E u 0 s.t. dX = AXdt + Budt + B 1 dW X (0) ∼ N (0 , Σ 0 ) , X ( T ) ∼ N (0 , Σ 1 ) connection with SBP ⇒ φ ( t, x ) = exp( −� x � 2 Q ( t ) − 1 ) & Riccati’s Chen-Georgiou-Pavon (TAC 2016)

  29. SBP Riccati’s – nonlinearly coupled Riccati equations ≡ Schr¨ odinger system Π = − A ′ Π − Π A + Π BB ′ Π ˙ H = − A ′ H − H A − H BB ′ H ˙ � (Π + H) . � BB ′ − B 1 B ′ + (Π + H) 1 Σ − 1 = Π(0) + H(0) 0 Σ − 1 = Π( T ) + H( T ) . T φ ) ⇔ Σ − 1 = Π + H log( ρ ) = log( φ ) + log( ˆ Chen-Georgiou-Pavon, Optimal steering of a linear stochastic system to a final probability distribution, IEEE Trans. Aut. Control , May 2016

  30. stationary SBP When can Σ be a stationary state-covariance for dx ( t ) = ( A − BK ) x ( t ) dt + B 1 dw ( t )? i.e., when is Σ = Exx ′ , for suitable choice of K ? – not all Σ can be realized by state feedback

  31. stationary SBP When can Σ be a stationary state-covariance for dx ( t ) = ( A − BK ) x ( t ) dt + B 1 dw ( t )? This is so iff � 0 � A Σ + Σ A ′ + B 1 B ′ � � 1 B B rank = rank . B 0 B 0 – Chen-Georgiou-Pavon, Optimal steering..., Part II IEEE TAC , May 2016 – Georgiou, Structure of state covariances... TAC 2002 – recent work with Mihailo Jovanovic etal. on inverse problems, etc., 2016, 2017

  32. stationary SBP Assuming � 0 � A Σ + Σ A ′ + B 1 B ′ � � B B 1 rank = rank , B 0 B 0 find K so that for u = − Kx and dx = ( A − BK ) xdt + B 1 dw , we have: Σ = Exx ′ and J power ( u ) := E {� u � 2 } is minimal Via semidefinite programming: – Chen-Georgiou-Pavon, Optimal steering..., Part II IEEE TAC , May 2016.

  33. Application: Cooling Efficient steering from initial condition ρ 0 to ρ 1 at finite time – Efficient stationary state of stochastic oscillators to desired ρ 1 – thermodynamic systems, controlling collective response – magnetization distribution in NMR spectroscopy,.. - Chen-Georgiou-Pavon Fast cooling for a system of stochastic oscillators, J. Math. Phys. Nov. 2015.

  34. Cooling (cont’d) Nyquist-Johnson noise driven oscillator Ldi L ( t ) = v C ( t ) dt − v C ( t ) dt − Ri L ( t ) dt + u ( t ) dt + dw ( t ) RCdv C ( t ) =

  35. Cooling & keeping it cool ! Inertial particles with stochastic excitation trajectories in phase space transparent tube: “ 3 σ region”

  36. Application: OMT with dynamics via SBP Schr¨ odinger bridge with ǫ = 9 Schr¨ odinger bridge with ǫ = 0 . 01 Schr¨ odinger bridge with ǫ = 4 Optimal transport with prior

Recommend


More recommend