Coupling Reflections Exotica Exotic Couplings of Brownian motion X CLAPEM, Lima Wilfrid Kendall w.s.kendall@warwick.ac.uk Department of Statistics, University of Warwick Monday 26 February 2007 Conclusion Coupling Reflections Exotica Exotic Couplings of Brownian motion X CLAPEM, Lima Wilfrid Kendall w.s.kendall@warwick.ac.uk Department of Statistics, University of Warwick Monday 26 February 2007 Conversations and collaborations: Gerard Ben Arous, Sigurd Assing, Stephen Connor, Mike Cranston, Terry Lyons, Catherine Price, Jon Warren. Conclusion Coupling Reflections Exotica Plan of talk Coupling Coupling: maximal or co-adapted? Coupling and Optimal Transport Coupling and Stochastic Control Conclusion Coupling Reflections Exotica Plan of talk Coupling Coupling: maximal or co-adapted? Coupling and Optimal Transport Coupling and Stochastic Control Reflections Brownian coupling Diffusion coupling Coupling Brownian motion on manifolds Conclusion Coupling Reflections Exotica Plan of talk Coupling Coupling: maximal or co-adapted? Coupling and Optimal Transport Coupling and Stochastic Control Reflections Brownian coupling Diffusion coupling Coupling Brownian motion on manifolds Exotica Coupling real Brownian motion with time-integrals Coupling planar BM and stochastic area Coupling n-dimensional BM and all stochastic areas Conclusion Coupling Reflections Exotica Conclusion Probabilistic Coupling Coupling Coupling: maximal or co-adapted? Coupling and Optimal Transport Coupling and Stochastic Control Reflections Exotica e ) builds them on A coupling of two random processes (X and X the same probability space, interdependent in a useful way: e : Ω × [0, ∞) → Rn X, X (prescribed marginals). Coupling Reflections Exotica Some motivations 1 interacting particle systems 2 randomized algorithms 3 explicit approximation (Stein-Chen) 4 rates of convergence in MCMC 5 harmonic functions: gradient estimates etc 6 simulation, especially perfect simulation 7 de-randomizing probabilistic constructions Conclusion Coupling Reflections Exotica Coupling: maximal or co-adapted? Coupling is used for many purposes, but the objective which organizes much of the theory is, Whether a coupling is successful? Conclusion Coupling Reflections Exotica Conclusion Coupling: maximal or co-adapted? Coupling is used for many purposes, but the objective which organizes much of the theory is, Whether a coupling is successful? Successful coupling: h i et for all large enough t P Xt = X = 1. Coupling Reflections Exotica Conclusion Coupling: maximal or co-adapted? Coupling is used for many purposes, but the objective which organizes much of the theory is, Whether a coupling is successful? Successful coupling: h i et for all large enough t P Xt = X Choices: maximal (eg: Pitman 1976) = 1. Coupling Reflections Exotica Conclusion Coupling: maximal or co-adapted? Coupling is used for many purposes, but the objective which organizes much of the theory is, Whether a coupling is successful? Successful coupling: h i et for all large enough t P Xt = X = Choices: maximal (eg: Pitman 1976): potential theory; 1. Coupling Reflections Exotica Conclusion Coupling: maximal or co-adapted? Coupling is used for many purposes, but the objective which organizes much of the theory is, Whether a coupling is successful? Successful coupling: h i et for all large enough t P Xt = X = 1. Choices: maximal (eg: Pitman 1976): potential theory; or co-adapted (prescriptions of marginals respect a specific filtration) Coupling Reflections Exotica Conclusion Coupling: maximal or co-adapted? Coupling is used for many purposes, but the objective which organizes much of the theory is, Whether a coupling is successful? Successful coupling: h i et for all large enough t P Xt = X = 1. Choices: maximal (eg: Pitman 1976): potential theory; or co-adapted (prescriptions of marginals respect a specific filtration): often easier to work with; Coupling Reflections Exotica Conclusion Coupling: maximal or co-adapted? Coupling is used for many purposes, but the objective which organizes much of the theory is, Whether a coupling is successful? Successful coupling: h i et for all large enough t P Xt = X = 1. Choices: maximal (eg: Pitman 1976): potential theory; or co-adapted (prescriptions of marginals respect a specific filtration): often easier to work with; other possibilities . . . . Coupling Reflections Exotica Doeblin’s coupling for Markov chains (run independently till first meeting) Doeblin (1938)’s construction: Conclusion Coupling Reflections Exotica Doeblin’s coupling for Markov chains (run independently till first meeting) Doeblin (1938)’s construction: if X is a finite-state-space Markov chain for which the convergence theorem applies (irreducible, aperiodic); Conclusion Coupling Reflections Exotica Doeblin’s coupling for Markov chains (run independently till first meeting) Doeblin (1938)’s construction: if X is a finite-state-space Markov chain for which the convergence theorem applies (irreducible, aperiodic); e which move independently till construct two copies X , X they couple, and then move synchronously. Conclusion Coupling Reflections Exotica Doeblin’s coupling for Markov chains (run independently till first meeting) Doeblin (1938)’s construction: if X is a finite-state-space Markov chain for which the convergence theorem applies (irreducible, aperiodic); e which move independently till construct two copies X , X they couple, and then move synchronously. e starts in equilibrium, then If T is the coupling time, and X the coupling inequality uses P [T > t] to estimate total variation distance of L (Xt ) from equilibrium. Conclusion Coupling Reflections Exotica Doeblin’s coupling for Markov chains (run independently till first meeting) Doeblin (1938)’s construction: if X is a finite-state-space Markov chain for which the convergence theorem applies (irreducible, aperiodic); e which move independently till construct two copies X , X they couple, and then move synchronously. e starts in equilibrium, then If T is the coupling time, and X the coupling inequality uses P [T > t] to estimate total variation distance of L (Xt ) from equilibrium. Clearly co-adapted. Conclusion Coupling Reflections Exotica Doeblin’s coupling for Markov chains (run independently till first meeting) Doeblin (1938)’s construction: if X is a finite-state-space Markov chain for which the convergence theorem applies (irreducible, aperiodic); e which move independently till construct two copies X , X they couple, and then move synchronously. e starts in equilibrium, then If T is the coupling time, and X the coupling inequality uses P [T > t] to estimate total variation distance of L (Xt ) from equilibrium. Clearly co-adapted. Clearly inefficient. Conclusion Coupling Reflections Exotica Doeblin’s coupling for Markov chains (run independently till first meeting) Doeblin (1938)’s construction: if X is a finite-state-space Markov chain for which the convergence theorem applies (irreducible, aperiodic); e which move independently till construct two copies X , X they couple, and then move synchronously. e starts in equilibrium, then If T is the coupling time, and X the coupling inequality uses P [T > t] to estimate total variation distance of L (Xt ) from equilibrium. Clearly co-adapted. Clearly inefficient. We need to do much better for: randomized algorithms, smart simulation, gradient estimates . . . . Conclusion Coupling Reflections Exotica Coupling and optimal transport There is a significant relationship with optimal transport: Conclusion Coupling Reflections Exotica Conclusion Coupling and optimal transport There is a significant relationship with optimal transport: successful coupling solves a Monge-Kantorovich problem e MIN I L X , X Z = e) π e c(X , X e ) (d X , d X ) L(X ,X subject to prescribed L (X ) , L (Y ) , Coupling Reflections Exotica Conclusion Coupling and optimal transport There is a significant relationship with optimal transport: successful coupling solves a Monge-Kantorovich problem e MIN I L X , X Z = e) π e c(X , X e ) (d X , d X ) L(X ,X subject to prescribed L (X ) , L (Y ) , where e) c(X , X = h i et for all large t . 1 − P Xt = X Coupling Reflections Exotica Coupling and stochastic control Similarly for stochastic control, especially for co-adapted coupling: Conclusion Coupling Reflections Exotica Coupling and stochastic control Similarly for stochastic control, especially for co-adapted coupling: e is controlled by choice of J1 , J2 in suppose X dX e dX = = σ(X ) d B + µ(X ) d t , e )J1 d B + µ(X e , J2 ) d t ; σ(X Conclusion Coupling Reflections Exotica Coupling and stochastic control Similarly for stochastic control, especially for co-adapted coupling: e is controlled by choice of J1 , J2 in suppose X dX e dX = = σ(X ) d B + µ(X ) d t , e )J1 d B + µ(X e , J2 ) d t ; σ(X then choose J1 , J2 to maximize h i et for all large t . P Xt = X Conclusion Coupling Reflections Exotica Conclusion Reflection couplings Coupling Reflections Brownian coupling Diffusion coupling Coupling Brownian motion on manifolds Exotica Most classical couplings use the idea of reflection in one form or another. Coupling Reflections Exotica Conclusion Lindvall’s reflection coupling for Brownian motion do the opposite till first meeting Lindvall (1982): How to couple two Brownian motions? Coupling Reflections Exotica Conclusion Lindvall’s reflection coupling for Brownian motion do the opposite till first meeting Lindvall (1982): How to couple two Brownian motions? Generate first Brownian motion X beginning at x; Coupling Reflections Exotica Conclusion Lindvall’s reflection coupling for Brownian motion do the opposite till first meeting Lindvall (1982): How to couple two Brownian motions? Generate first Brownian motion X beginning at x; Locate y , initial point for second Brownian motion; Coupling Reflections Exotica Conclusion Lindvall’s reflection coupling for Brownian motion do the opposite till first meeting Lindvall (1982): How to couple two Brownian motions? Generate first Brownian motion X beginning at x; Locate y , initial point for second Brownian motion; Construct line of reflection, hence reflection map H; Coupling Reflections Exotica Conclusion Lindvall’s reflection coupling for Brownian motion do the opposite till first meeting Lindvall (1982): How to couple two Brownian motions? Generate first Brownian motion X beginning at x; Locate y , initial point for second Brownian motion; Construct line of reflection, hence reflection map H; Generate second Brownian motion Y using reflection of first; Coupling Reflections Exotica Conclusion Lindvall’s reflection coupling for Brownian motion do the opposite till first meeting Lindvall (1982): How to couple two Brownian motions? Generate first Brownian motion X beginning at x; Locate y , initial point for second Brownian motion; Construct line of reflection, hence reflection map H; Generate second Brownian motion Y using reflection of first; Couple in higher dimensions by reflecting in hyperplane . . . ; Coupling Reflections Exotica Conclusion Lindvall’s reflection coupling for Brownian motion do the opposite till first meeting Lindvall (1982): How to couple two Brownian motions? Generate first Brownian motion X beginning at x; Locate y , initial point for second Brownian motion; Construct line of reflection, hence reflection map H; Generate second Brownian motion Y using reflection of first; Couple in higher dimensions by reflecting in hyperplane; This coupling is co-adapted (no “cheating”) and maximal. Coupling Reflections Exotica Co-adapted coupling is usually not maximal Consider couplings for Ornstein-Uhlenbeck processes dX = d B − 21 X d t ; Conclusion Coupling Reflections Exotica Co-adapted coupling is usually not maximal Consider couplings for Ornstein-Uhlenbeck processes dX = d B − 21 X d t ; e is begun at −x then reflection If X is begun at x and X coupling is maximal. Conclusion Coupling Reflections Exotica Conclusion Co-adapted coupling is usually not maximal Consider couplings for Ornstein-Uhlenbeck processes dX = d B − 21 X d t ; e is begun at −x then reflection If X is begun at x and X coupling is maximal. e is begun in statistical equilibrium If X is begun at 0 and X then reflection coupling is the best co-adapted coupling, but is not maximal. Coupling Reflections Exotica Diffusion coupling do more or less the opposite till first meeting Lindvall and Rogers (1986), also Chen and Li (1989): How to couple two copies of an elliptic diffusion? (not necessarily maximal coupling, but co-adapted.) Conclusion Coupling Reflections Exotica Diffusion coupling do more or less the opposite till first meeting Lindvall and Rogers (1986), also Chen and Li (1989): How to couple two copies of an elliptic diffusion? (not necessarily maximal coupling, but co-adapted.) Specify the diffusion by a stochastic differential equation: dX = σ(X ) d W + b(X ) d t . Conclusion Coupling Reflections Exotica Diffusion coupling do more or less the opposite till first meeting Lindvall and Rogers (1986), also Chen and Li (1989): How to couple two copies of an elliptic diffusion? (not necessarily maximal coupling, but co-adapted.) Specify the diffusion by a stochastic differential equation: dX = σ(X ) d W + b(X ) d t . Reflection map H = H(X ) now depends on σ(X ): e dX = e )H(X ) d W + b(X e) d t . σ(X Conclusion Coupling Reflections Exotica Conclusion Diffusion coupling do more or less the opposite till first meeting Lindvall and Rogers (1986), also Chen and Li (1989): How to couple two copies of an elliptic diffusion? (not necessarily maximal coupling, but co-adapted.) Specify the diffusion by a stochastic differential equation: dX = σ(X ) d W + b(X ) d t . Reflection map H = H(X ) now depends on σ(X ): e dX = e )H(X ) d W + b(X e) d t . σ(X Lindvall and Rogers (1986) consider when success is sure. Coupling Reflections Exotica Coupling Brownian motion on manifolds be geometric about doing the opposite till first meeting When the diffusion is Brownian motion on a Riemannian manifold, the “mirror-map” H can be chosen accordingly. Conclusion Coupling Reflections Exotica Coupling Brownian motion on manifolds be geometric about doing the opposite till first meeting When the diffusion is Brownian motion on a Riemannian manifold, the “mirror-map” H can be chosen accordingly. Reflect the Brownian noise using Conclusion Coupling Reflections Exotica Coupling Brownian motion on manifolds be geometric about doing the opposite till first meeting When the diffusion is Brownian motion on a Riemannian manifold, the “mirror-map” H can be chosen accordingly. Reflect the Brownian noise using (a) stochastic development, Conclusion Coupling Reflections Exotica Coupling Brownian motion on manifolds be geometric about doing the opposite till first meeting When the diffusion is Brownian motion on a Riemannian manifold, the “mirror-map” H can be chosen accordingly. Reflect the Brownian noise using (a) stochastic development, (b) a geodesic connecting the two diffusions. Conclusion Coupling Reflections Exotica Coupling Brownian motion on manifolds be geometric about doing the opposite till first meeting When the diffusion is Brownian motion on a Riemannian manifold, the “mirror-map” H can be chosen accordingly. Reflect the Brownian noise using (a) stochastic development, (b) a geodesic connecting the two diffusions. Coupling can then be analyzed in terms of geometry, relating success of coupling to curvature bounds. Conclusion Coupling Reflections Exotica Exotic couplings Coupling Reflections Exotica Coupling real Brownian motion with time-integrals Coupling planar BM and stochastic area Coupling n-dimensional BM and all stochastic areas Suppose we are more ambitious: is it also possible to couple functionals of the process at the same time? Conclusion Coupling Reflections Exotica Coupling Brownian motion with time-integrals pause occasionally to let the time-integrals catch up Conclusion Coupling Reflections Exotica Conclusion Coupling Brownian motion with time-integrals pause occasionally to let the time-integrals catch up Required: to couple couple A, B, R A d t, R B d t at the same time as we Coupling Reflections Exotica Conclusion Coupling Brownian motion with time-integrals pause occasionally to let the time-integrals catch up R R Required: to couple A d t, B d t at the same time as we couple A, B, where A and B are co-adapted real Brownian motions. Coupling Reflections Exotica Conclusion Coupling Brownian motion with time-integrals pause occasionally to let the time-integrals catch up R R Required: to couple A d t, B d t at the same time as we couple A, B, where A and B are co-adapted real Brownian motions. R We can set B = J d A, where J = ±1 is an adapted function. Coupling Reflections Exotica Conclusion Coupling Brownian motion with time-integrals pause occasionally to let the time-integrals catch up R R Required: to couple A d t, B d t at the same time as we couple A, B, where A and B are co-adapted real Brownian motions. R We can set B = J d A, where J = ±1 is an adapted function. Hence we have to choose J well so as to control Z Z W , W d t = B − A, (B − A) d t to hit (0, 0) . Coupling Reflections Exotica Conclusion Coupling Brownian motion with time-integrals pause occasionally to let the time-integrals catch up R R Required: to couple A d t, B d t at the same time as we couple A, B, where A and B are co-adapted real Brownian motions. R We can set B = J d A, where J = ±1 is an adapted function. Hence we have to choose J well so as to control Z Z W , W d t = B − A, (B − A) d t to hit (0, 0) . Ben Arous et al. (1995), K. and Price (2004). Coupling Reflections Exotica Coupling the Time-integral by Pausing W alternates between 2BM (if J = −1) and constant (if J = 1). Conclusion Coupling Reflections Exotica Coupling the Time-integral Horizontal axis: WR= B − A; Vertical axis: V = W d t. Conclusion Coupling Reflections Exotica Coupling many iterated time-integrals . . . pause in a systematic encoded manner Use a perturbed Morse-Thué sequence + - - + - + + - - + + - + - - + Conclusion Coupling Reflections Exotica Conclusion Application: Relativistic diffusion small perturbations don’t affect strategy Ismael Bailleul (PhD thesis, Paris-Sud 2006): Dudley (1973)’s relativistic diffusion on Minkowski space (driven at sub-light-speed by relativistic Brownian motion). Coupling Reflections Exotica Conclusion Application: Relativistic diffusion small perturbations don’t affect strategy Ismael Bailleul (PhD thesis, Paris-Sud 2006): Dudley (1973)’s relativistic diffusion on Minkowski space (driven at sub-light-speed by relativistic Brownian motion). limiting velocity direction; asymptotically location lies on parallel geodesic. Coupling Reflections Exotica Conclusion Application: Relativistic diffusion small perturbations don’t affect strategy Ismael Bailleul (PhD thesis, Paris-Sud 2006): Dudley (1973)’s relativistic diffusion on Minkowski space (driven at sub-light-speed by relativistic Brownian motion). limiting velocity direction; asymptotically location lies on parallel geodesic. is this the full set of asymptotics? Coupling Reflections Exotica Conclusion Application: Relativistic diffusion small perturbations don’t affect strategy Ismael Bailleul (PhD thesis, Paris-Sud 2006): Dudley (1973)’s relativistic diffusion on Minkowski space (driven at sub-light-speed by relativistic Brownian motion). limiting velocity direction; asymptotically location lies on parallel geodesic. is this the full set of asymptotics? (compute conditional diffusion, coupling): Coupling Reflections Exotica Conclusion Application: Relativistic diffusion small perturbations don’t affect strategy Ismael Bailleul (PhD thesis, Paris-Sud 2006): Dudley (1973)’s relativistic diffusion on Minkowski space (driven at sub-light-speed by relativistic Brownian motion). limiting velocity direction; asymptotically location lies on parallel geodesic. is this the full set of asymptotics? (compute conditional diffusion, coupling): Can we control (W , V ) to hit (0, 0) by switching K in {0, 1}? dW = K d A + W d B − 32 W d t dV = W dt . Coupling Reflections Exotica Conclusion Application: Relativistic diffusion small perturbations don’t affect strategy Ismael Bailleul (PhD thesis, Paris-Sud 2006): Dudley (1973)’s relativistic diffusion on Minkowski space (driven at sub-light-speed by relativistic Brownian motion). limiting velocity direction; asymptotically location lies on parallel geodesic. is this the full set of asymptotics? (compute conditional diffusion, coupling): Can we control (W , V ) to hit (0, 0) by switching K in {0, 1}? dW = K d A + W d B − 32 W d t dV = W dt . Coupling Reflections Exotica Conclusion Application: Relativistic diffusion small perturbations don’t affect strategy Ismael Bailleul (PhD thesis, Paris-Sud 2006): Dudley (1973)’s relativistic diffusion on Minkowski space (driven at sub-light-speed by relativistic Brownian motion). limiting velocity direction; asymptotically location lies on parallel geodesic. is this the full set of asymptotics? (compute conditional diffusion, coupling): Can we control (W , V ) to hit (0, 0) by switching K in {0, 1}? dW = K d A + W d B − 32 W d t dV = W dt . YES Coupling Reflections Exotica Coupling planar Brownian motion and Lévy stochastic area pause when the areal difference gets too big Lévy stochastic area Z Z B1 d B2 − B2 d B1 , with symmetries given by the Heisenberg group. Conclusion Coupling Reflections Exotica Conclusion Coupling planar Brownian motion and Lévy stochastic area pause when the areal difference gets too big Lévy stochastic area Z Z B1 d B2 − B2 d B1 , with symmetries given by the Heisenberg group. Invariance arguments: control the following to hit zero. Z A12 = Z B1 d B2 − Z Z e e e e B2 d B1 − B1 d B2 − B2 d B1 e 2 B1 . e2 − B + B1 B Coupling Reflections Exotica Conclusion Lévy stochastic area coupling strategy SDEs Coupling Reflections Exotica Coupling n-dimensional Brownian motion and all Lévy stochastic areas Can we do the same for n-dimensional Brownian motion, B = [B1 , . . . , Bn ]T ? Conclusion Coupling Reflections Exotica Coupling n-dimensional Brownian motion and all Lévy stochastic areas Can we do the same for n-dimensional Brownian motion, B = [B1 , . . . , Bn ]T ? 1 We R need toR couple 2 n(n − 1) different stochastic areas Bi d Bj − Bj d Bi . Conclusion Coupling Reflections Exotica Conclusion Coupling n-dimensional Brownian motion and all Lévy stochastic areas Can we do the same for n-dimensional Brownian motion, B = [B1 , . . . , Bn ]T ? 1 We R need toR couple 2 n(n − 1) different stochastic areas Bi d Bj − Bj d Bi . The invariant differences of stochastic area can be viewed (as before) as entries in an n × n skew-symmetric matrix: Z Aij = Z Bi d Bj − Z Z e e e e Bj d Bi − Bi d Bj − Bj d Bi ej − B e j Bi . + Bi B Coupling Reflections Exotica Simple ideas don’t work It is no good trying first to couple Aij , then Apq , et cetera. Conclusion Coupling Reflections Exotica Conclusion Simple ideas don’t work It is no good trying first to couple Aij , then Apq , et cetera. The best that can be done this way seems to be, to couple a single “row” of stochastic areas: A12 , A13 , A14 , . . . , A1n . Coupling Reflections Exotica Conclusion Simple ideas don’t work It is no good trying first to couple Aij , then Apq , et cetera. The best that can be done this way seems to be, to couple a single “row” of stochastic areas: A12 , A13 , A14 , . . . , A1n . Trying to couple more stochastic areas in this sequential way is like herding cats! Coupling Reflections Exotica Conclusion Simple ideas don’t work It is no good trying first to couple Aij , then Apq , et cetera. The best that can be done this way seems to be, to couple a single “row” of stochastic areas: A12 , A13 , A14 , . . . , A1n . Trying to couple more stochastic areas in this sequential way is like herding cats! But a quite different technique allows us to couple the whole set of stochastic areas at once. Coupling Reflections Exotica A general formulation for Brownian coupling e is The general co-adapted coupling between B and B determined by an n × n covariance matrix J: Conclusion Coupling Reflections Exotica Conclusion A general formulation for Brownian coupling e is The general co-adapted coupling between B and B determined by an n × n covariance matrix J: h i dB T eT × d B d B e dB = In J × dt . JT In Coupling Reflections Exotica Conclusion A general formulation for Brownian coupling e is The general co-adapted coupling between B and B determined by an n × n covariance matrix J: h i dB T eT × d B d B e dB = In J × dt . JT In Matrix constraint (since right-hand side must be nonnegative-definite): JT J ≤ In . Coupling Reflections Exotica Conclusion A general formulation for Brownian coupling e is The general co-adapted coupling between B and B determined by an n × n covariance matrix J: h i dB T eT × d B d B e dB = In J × dt . JT In Matrix constraint (since right-hand side must be nonnegative-definite): JT J ≤ In . Control theory arguments suggest we focus on orthogonal matrices J, and be prepared for interfaces . . . . Coupling Reflections Exotica Conclusion A general formulation for Brownian coupling (ctd.) e define matrix of invariant differences of Set X = B − B, stochastic areas A as above. Coupling Reflections Exotica Conclusion A general formulation for Brownian coupling (ctd.) e define matrix of invariant differences of Set X = B − B, stochastic areas A as above. Focus on distance V and areal discrepancy U Coupling Reflections Exotica Conclusion A general formulation for Brownian coupling (ctd.) e define matrix of invariant differences of Set X = B − B, stochastic areas A as above. Focus on distance V and areal discrepancy U defined by X = Vν and A = UZ , where ν is a randomly evolving unit vector and Z is a randomly evolving skew-symmetric matrix of unit Hilbert-Schmidt norm, Coupling Reflections Exotica Conclusion A general formulation for Brownian coupling (ctd.) e define matrix of invariant differences of Set X = B − B, stochastic areas A as above. Focus on distance V and areal discrepancy U defined by X = Vν and A = UZ , where ν is a randomly evolving unit vector and Z is a randomly evolving skew-symmetric matrix of unit Hilbert-Schmidt norm, together measuring the geometric configuration underlying the classical and areal distances. Coupling Reflections Exotica Relevant coupling recipes K. (2007): Use a little bit of mirror coupling: Conclusion Coupling Reflections Exotica Relevant coupling recipes K. (2007): Use a little bit of mirror coupling: the mirror strategy J = In − 2νν T Conclusion Coupling Reflections Exotica Relevant coupling recipes K. (2007): Use a little bit of mirror coupling: the mirror strategy J = In − 2νν T Mix judiciously with a slight rotation coupling: Conclusion Coupling Reflections Exotica Conclusion Relevant coupling recipes K. (2007): Use a little bit of mirror coupling: the mirror strategy J = In − 2νν T Mix judiciously with a slight rotation coupling: J = exp − γ Z U/V 2 SDEs Coupling Reflections Exotica Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, Conclusion Coupling Reflections Exotica Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; Conclusion Coupling Reflections Exotica Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; defining K = log V (spatial) and H = log U (areal), Conclusion Coupling Reflections Exotica Conclusion Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; defining K = log V (spatial) and H = log U (areal), we find that with a positive chance we can make K , H tend to −∞ as τe → ∞ at approximately linear rates, Coupling Reflections Exotica Conclusion Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; defining K = log V (spatial) and H = log U (areal), we find that with a positive chance we can make K , H tend to −∞ as τe → ∞ at approximately linear rates, and these rates are such that exp(K − H) = V /U tends to +∞ at approximately linear rate (so τe tends to +∞ while t remains bounded), Coupling Reflections Exotica Conclusion Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; defining K = log V (spatial) and H = log U (areal), we find that with a positive chance we can make K , H tend to −∞ as τe → ∞ at approximately linear rates, and these rates are such that exp(K − H) = V /U tends to +∞ at approximately linear rate (so τe tends to +∞ while t remains bounded – Lamperti’s trick), Coupling Reflections Exotica Conclusion Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; defining K = log V (spatial) and H = log U (areal), we find that with a positive chance we can make K , H tend to −∞ as τe → ∞ at approximately linear rates, and these rates are such that exp(K − H) = V /U tends to +∞ at approximately linear rate (so τe tends to +∞ while t remains bounded – Lamperti’s trick), and nevertheless exp(H − 2K ) = U/V 2 tends to infinity (required if the estimates are to work!). Coupling Reflections Exotica Conclusion Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; defining K = log V (spatial) and H = log U (areal), we find that with a positive chance we can make K , H tend to −∞ as τe → ∞ at approximately linear rates, and these rates are such that exp(K − H) = V /U tends to +∞ at approximately linear rate (so τe tends to +∞ while t remains bounded – Lamperti’s trick), and nevertheless exp(H − 2K ) = U/V 2 tends to infinity (required if the estimates are to work!). This implies, with positive chance U, V both R ∞hit zero (so coupling occurs) at the finite random time 0 U 2 /V 2 d τe. Coupling Reflections Exotica Conclusion Careful but simple estimates from stochastic calculus shows that, under a mixture of these two strategies, using a time-change d τe = 4V 2 d t/U 2 ; defining K = log V (spatial) and H = log U (areal), we find that with a positive chance we can make K , H tend to −∞ as τe → ∞ at approximately linear rates, and these rates are such that exp(K − H) = V /U tends to +∞ at approximately linear rate (so τe tends to +∞ while t remains bounded – Lamperti’s trick), and nevertheless exp(H − 2K ) = U/V 2 tends to infinity (required if the estimates are to work!). This implies, with positive chance U, V both R ∞hit zero (so coupling occurs) at the finite random time 0 U 2 /V 2 d τe. Scaling arguments then show how to make coupling sure! Coupling Reflections Exotica Conclusion Brownian motion functionals couple much more than I thought was possible! Conclusion Coupling Reflections Exotica Conclusion Brownian motion functionals couple much more than I thought was possible! What about higher-order stochastic areas? Conclusion Coupling Reflections Exotica Conclusion Conclusion Brownian motion functionals couple much more than I thought was possible! What about higher-order stochastic areas? We believe the result generalizes, but the situation is complicated. Coupling Reflections Exotica Conclusion Conclusion Brownian motion functionals couple much more than I thought was possible! What about higher-order stochastic areas? We believe the result generalizes, but the situation is complicated. Carnot-Caratheodory distance may provide an alternative. Coupling Reflections Exotica Conclusion Conclusion Brownian motion functionals couple much more than I thought was possible! What about higher-order stochastic areas? We believe the result generalizes, but the situation is complicated. Carnot-Caratheodory distance may provide an alternative. What about (finite-dimensional) hypoelliptic diffusions? Should be straightforward (coupling strategies are robust). Coupling Reflections Exotica Conclusion Conclusion Brownian motion functionals couple much more than I thought was possible! What about higher-order stochastic areas? We believe the result generalizes, but the situation is complicated. Carnot-Caratheodory distance may provide an alternative. What about (finite-dimensional) hypoelliptic diffusions? Should be straightforward (coupling strategies are robust). Applications . . . . Coupling Reflections Exotica Conclusion Conclusion Brownian motion functionals couple much more than I thought was possible! What about higher-order stochastic areas? We believe the result generalizes, but the situation is complicated. Carnot-Caratheodory distance may provide an alternative. What about (finite-dimensional) hypoelliptic diffusions? Should be straightforward (coupling strategies are robust). Applications . . . . Infinite-dimensional case . . . . Coupling Reflections Exotica Conclusion Conclusion Brownian motion functionals couple much more than I thought was possible! What about higher-order stochastic areas? We believe the result generalizes, but the situation is complicated. Carnot-Caratheodory distance may provide an alternative. What about (finite-dimensional) hypoelliptic diffusions? Should be straightforward (coupling strategies are robust). Applications . . . . Infinite-dimensional case . . . . QUESTIONS? References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Bibliography This is a rich hypertext bibliography. Journals are linked to their homepages, and stable or Project Euclid ) have been URL links (as provided for example by JSTOR added where known. Access to such URLs is not universal: in case of difficulty you should check whether you are registered (directly or indirectly) with the relevant provider. In the case of preprints, icons , , , linking to homepage locations are inserted where available: note that these are probably less stable than journal links!. Ben Arous, G., M. Cranston, and W. S. Kendall (1995). Coupling constructions for hypoelliptic diffusions: Two examples. In M. Cranston and M. Pinsky (Eds.), Stochastic Analysis: Summer Research Institute July 11-30, 1993, Volume 57, Providence, RI Providence, pp. 193–212. American Mathematical Society. Chen, M. F. and S. F. Li (1989). Coupling methods for multidimensional diffusion processes. The Annals of Probability 17 (1), 151–177, . Doeblin, W. (1938). Exposé de la théorie des chaînes simples constants de Markoff á un nombre fini d’états. Revue Math. de l’Union Interbalkanique 2, 77–105. References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Dudley, R. M. (1973). Asymptotics of some relativistic Markov processes. Proc. Nat. Acad. Sci. U.S.A. 70, 3551–3555, . Kendall, W. S. (1986a). Nonnegative Ricci curvature and the Brownian coupling property. Stochastics and Stochastic Reports 19, 111–129. Kendall, W. S. (1986b). Stochastic differential geometry, a coupling property, and harmonic maps. The Journal of the London Mathematical Society (Second Series) 33, 554–566. Kendall, W. S. (2001). Symbolic Itô calculus: an ongoing story. Statistics and Computing 11, 25–35, . Also: University of Warwick Department of Statistics Research Report 327. . Kendall, W. S. (2007). Coupling all the Lévy stochastic areas of multidimensional Brownian motion. The Annals of Probability to appear, . Also available as University of Warwick Department of Statistics Research Report 445 , including animations. References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Kendall, W. S. and C. J. Price (2004). Coupling iterated Kolmogorov diffusions. Electronic Journal of Probability 9, 382–410, . Also available as University of Warwick Department of Statistics Research Report 416 , with figures linked to animations. Lamperti, J. (1972). Semi-stable Markov processes. I. Zeitschrift für Wahrscheinlichkeitstheorie und verve Gebiete 22, 205–225, Lindvall, T. (1982). On coupling of Brownian motions. Technical report 1982:23, Department of Mathematics, Chalmers University of Technology and University of Göteborg. Lindvall, T. and L. C. G. Rogers (1986). Coupling of multidimensional diffusions by reflection. The Annals of Probability 14(3), 860–872, . Pitman, J. W. (1976). On coupling of Markov chains. Z. Wahrscheinlichkeitstheorie und Verw. Gebiete 35(4), 315–322, . . References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for single stochastic area (I) Effect of reflection coupling: (d V )2 = 4 d t , Drift d V = 0 , (d V ) × (d U) = 0 (d U)2 = 8V 2 d t , Drift d U = 0 so that the spatial distance V moves as a scalar Brownian motion at least till it hits 0, and the areal discrepancy U moves as a scalar Brownian motion subject to a V -dependent time-change. References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for single stochastic area (II) Effect of synchronous coupling: (d V )2 = 0 , Drift d V = 0 , (d V ) × (d U) = 0 , (d U)2 = 8V 2 d t , Drift d U = 0 so that the spatial distance V is held constant, while the areal discrepancy U moves as a scalar Brownian motion subject to a V -dependent time-change in the same way as for reflection coupling. References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for single stochastic area (III) Apply Lamperti (1972)’s observation: the random time-change 4dt = V2 dτ . makes K = log(V ) into an (interrupted) Brownian motion with constant negative drift. Writing W = U/V 2 , and N indicating whether process is in reflected coupling mode (N = 0 if W big), (d K )2 = N d τ , (d K ) × (d W ) = 2NW d τ , (d W )2 = 2 1 + 2NW 2 d τ , 1 Drift d K = − N d τ , 2 Drift d W = 3NW d τ . BACK References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for multiple stochastic areas (I) Use (τ -)time-scale defined by 4 d t = V 2 d τ , set W = U/V 2 , K = log(V ) and H = log(U), and S = (J + JT )/2, A = (J − JT )/2; 1 (d K )2 = 1 − νT S ν d τ , 2 1 n − tr S − 2 1 − ν T S ν dτ , Drift d K = 4 1 (d K ) × (d H) = − ν T ZT A ν dτ , W 1 (d H)2 = 2ν T ZT I + S Z ν dτ , W2 1 1 n − 1 + tr S − ν T S ν − 4ν T ZT I + S Z ν Drift d H = dτ 2 W2 1 1 + tr ZT A dτ . 2 W References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for multiple stochastic areas (II) Reflection coupling is defined by Jreflection = I − 2ν ν T , which implies (d K )2 = d τ , 1 Drift d K = − d τ , 2 (d K ) × (d H) = 0 , (d H)2 = 4kZ νk2 dτ dτ 2 , Drift d H = νk . n − 1 − 4kZ W2 W2 References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for multiple stochastic areas (III) Synchronous coupling is defined by Jreflection = I, which implies (d K )2 = 0 , Drift d K = 0 , (d K ) × (d H) = 0 , (d H)2 = 4kZ νk2 dτ dτ 2 n − 1 − 4kZ νk , Drift d H = . W2 W2 So if n ≥ 3 then H is a non-constant submartingale under both controls! References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for multiple stochastic areas (III) Synchronous coupling is defined by Jreflection = I, which implies (d K )2 = 0 , Drift d K = 0 , (d K ) × (d H) = 0 , (d H)2 = 4kZ νk2 dτ dτ 2 n − 1 − 4kZ νk , Drift d H = . W2 W2 So if n ≥ 3 then H is a non-constant submartingale under both controls! We need something more than reflection/synchronous coupling. References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for multiple stochastic areas (IV) Use Taylor series and the adaptive rotation coupling Jrotation (θZ) = exp θZ , θ = −γ/W , γ2 dτ kZ νk2 2 4 W dτ γ2 Drift d K = 1 − 2kZ νk2 8 W2 dτ (d K ) × (d H) = γkZ νk2 2 W d τ (d H)2 = 4kZ νk2 2 W γ d τ Drift d H = − − n − 1 − 4kZ νk2 2 W2 (d K )2 = γ4 O(1) d τ , W4 γ4 + 4 O(1) d τ , W γ3 + 4 O(1) d τ , W γ2 + 4 O(1) d τ , W γ2 + γ3 O(1) d τ . + W4 + References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” SDEs for multiple stochastic areas (IV) Use Taylor series and the adaptive rotation coupling Jrotation (θZ) = exp θZ , θ = −γ/W , γ2 dτ kZ νk2 2 4 W dτ γ2 Drift d K = 1 − 2kZ νk2 8 W2 dτ (d K ) × (d H) = γkZ νk2 2 W d τ (d H)2 = 4kZ νk2 2 W γ d τ Drift d H = − − n − 1 − 4kZ νk2 2 W2 (d K )2 = (so work with d τe = d τ /W 2 ) γ4 O(1) d τ , W4 γ4 + 4 O(1) d τ , W γ3 + 4 O(1) d τ , W γ2 + 4 O(1) d τ , W γ2 + γ3 O(1) d τ . + W4 + BACK References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Possible strategy for higher-order stochastic areas construct traces of nonlinear invariant higher-order stochastic areas; References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Possible strategy for higher-order stochastic areas construct traces of nonlinear invariant higher-order stochastic areas; recursive construction and analysis . . . References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Possible strategy for higher-order stochastic areas construct traces of nonlinear invariant higher-order stochastic areas; recursive construction and analysis . . . compute stochastic differential system; References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Possible strategy for higher-order stochastic areas construct traces of nonlinear invariant higher-order stochastic areas; recursive construction and analysis . . . compute stochastic differential system; identify perturbations in strategy for step-2 stochastic areas (mixed reflection and small-rotation) which control higher-order stochastic areas on slower time-scales References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Possible strategy for higher-order stochastic areas construct traces of nonlinear invariant higher-order stochastic areas; recursive construction and analysis . . . compute stochastic differential system; identify perturbations in strategy for step-2 stochastic areas (mixed reflection and small-rotation) which control higher-order stochastic areas on slower time-scales, as in Kendall and Price (2004) for iterated time-integrals. BACK References SDEs for single stochastic area SDEs for multiple stochastic areas Possible future Define “exotic” Cambridge Advanced Learner’s Dictionary Definition exotic: adjective unusual and often exciting because of coming (or seeming to come) from a distant, especially tropical country BACK