Rough paths and the Gap Between Deterministic and Stochastic Di¤erential Equations
P. K. Friz
TU Berlin and WIAS
December 2009
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 1/22 Outline
I Donsker’sinvariance principle and Brownian motion II Itô integration and SDEs III Doss-Sussman’sODE approach to SDEs IV More on ODEs: Euler estimates V Smart ODE limits: rough di¤erential equations VI Brownian motion as a rough path VII SDEs driven by non-Brownian noise VIII Rough path spaces IX Donsker’sprinciple revisited
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 2/22 Limit is a probability measure on C [0, 1] , called Wiener measure W.
Brownian motion (BM) (Bt ) is a stochastic process with law W . Straight-forward extension to Rd -valued case In particular, a d-dimensional Brownian motion is just an ensemble of d independent Brownian motions, say
1 d Bt = Bt , ... , Bt .
Donsker’sinvariance principle
Let (ξi ) be an IID sequence of zero-mean, unit-variance random variables. [Donsker ’52] shows that the rescaled, piecewise-linearly-connected, random-walk
(n) 1 W = ξ + + ξ + (nt [nt]) ξ t n1/2 1 [tn] [nt]+1 converges weakly in the space of continuous functions on [0, 1].
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 3/22 Brownian motion (BM) (Bt ) is a stochastic process with law W . Straight-forward extension to Rd -valued case In particular, a d-dimensional Brownian motion is just an ensemble of d independent Brownian motions, say
1 d Bt = Bt , ... , Bt .
Donsker’sinvariance principle
Let (ξi ) be an IID sequence of zero-mean, unit-variance random variables. [Donsker ’52] shows that the rescaled, piecewise-linearly-connected, random-walk
(n) 1 W = ξ + + ξ + (nt [nt]) ξ t n1/2 1 [tn] [nt]+1 converges weakly in the space of continuous functions on [0, 1]. Limit is a probability measure on C [0, 1] , called Wiener measure W.
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 3/22 Straight-forward extension to Rd -valued case In particular, a d-dimensional Brownian motion is just an ensemble of d independent Brownian motions, say
1 d Bt = Bt , ... , Bt .
Donsker’sinvariance principle
Let (ξi ) be an IID sequence of zero-mean, unit-variance random variables. [Donsker ’52] shows that the rescaled, piecewise-linearly-connected, random-walk
(n) 1 W = ξ + + ξ + (nt [nt]) ξ t n1/2 1 [tn] [nt]+1 converges weakly in the space of continuous functions on [0, 1]. Limit is a probability measure on C [0, 1] , called Wiener measure W.
Brownian motion (BM) (Bt ) is a stochastic process with law W .
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 3/22 In particular, a d-dimensional Brownian motion is just an ensemble of d independent Brownian motions, say
1 d Bt = Bt , ... , Bt .
Donsker’sinvariance principle
Let (ξi ) be an IID sequence of zero-mean, unit-variance random variables. [Donsker ’52] shows that the rescaled, piecewise-linearly-connected, random-walk
(n) 1 W = ξ + + ξ + (nt [nt]) ξ t n1/2 1 [tn] [nt]+1 converges weakly in the space of continuous functions on [0, 1]. Limit is a probability measure on C [0, 1] , called Wiener measure W.
Brownian motion (BM) (Bt ) is a stochastic process with law W . Straight-forward extension to Rd -valued case
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 3/22 Donsker’sinvariance principle
Let (ξi ) be an IID sequence of zero-mean, unit-variance random variables. [Donsker ’52] shows that the rescaled, piecewise-linearly-connected, random-walk
(n) 1 W = ξ + + ξ + (nt [nt]) ξ t n1/2 1 [tn] [nt]+1 converges weakly in the space of continuous functions on [0, 1]. Limit is a probability measure on C [0, 1] , called Wiener measure W.
Brownian motion (BM) (Bt ) is a stochastic process with law W . Straight-forward extension to Rd -valued case In particular, a d-dimensional Brownian motion is just an ensemble of d independent Brownian motions, say
1 d Bt = Bt , ... , Bt .
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 3/22 (ii): continuous 0-mean Gaussian process with covariance
E (Bs Bt ) = min (s, t) s, t [0, 1] 8 2 2 (iii): Markov process with generator L 1 ∂ in the sense that = 2 ∂x 2
E [f (x + Bt )] x 1 Lf = f 00 f nice t ! 2 8 Again, straight-forward extension to Rd -valued case
Brownian motion: alternative characterizations
(i): continuous martingale such that B2 t is also a martingale t
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 4/22 2 (iii): Markov process with generator L 1 ∂ in the sense that = 2 ∂x 2
E [f (x + Bt )] x 1 Lf = f 00 f nice t ! 2 8 Again, straight-forward extension to Rd -valued case
Brownian motion: alternative characterizations
(i): continuous martingale such that B2 t is also a martingale t (ii): continuous 0-mean Gaussian process with covariance
E (Bs Bt ) = min (s, t) s, t [0, 1] 8 2
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 4/22 Again, straight-forward extension to Rd -valued case
Brownian motion: alternative characterizations
(i): continuous martingale such that B2 t is also a martingale t (ii): continuous 0-mean Gaussian process with covariance
E (Bs Bt ) = min (s, t) s, t [0, 1] 8 2 2 (iii): Markov process with generator L 1 ∂ in the sense that = 2 ∂x 2
E [f (x + Bt )] x 1 Lf = f 00 f nice t ! 2 8
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 4/22 Brownian motion: alternative characterizations
(i): continuous martingale such that B2 t is also a martingale t (ii): continuous 0-mean Gaussian process with covariance
E (Bs Bt ) = min (s, t) s, t [0, 1] 8 2 2 (iii): Markov process with generator L 1 ∂ in the sense that = 2 ∂x 2
E [f (x + Bt )] x 1 Lf = f 00 f nice t ! 2 8 Again, straight-forward extension to Rd -valued case
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 4/22 How to de…ne integration against Brownian motion? Itô’sbrilliant idea: with some help and intuition from martingale theory, 1 f (t, ω) dBt (ω) Z0 can be de…ned for reasonable non-anticipating f : start with simple integrands and complete with isometry
1 2 1 2 E f (t, ω) dBt (ω) = E f (t, ω) dt . "Z0 # Z0 t 1 2 Example: Bs dBs = B t ... 2nd order calculus! 0 2 t Fact: Itô-integrals have left-point Riemann-sum approximations. R De…ne Stratonovich-integration via mid-point Riemann-sum t 1 2 approximations = Bs ∂Bs = B (1st order calculus!) ) 0 2 t R
Itô integration
Fact: Typical sample paths of Brownian motion, t Bt (ω), have in…nite variation on every interval. 7!
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 5/22 t 1 2 Example: Bs dBs = B t ... 2nd order calculus! 0 2 t Fact: Itô-integrals have left-point Riemann-sum approximations. R De…ne Stratonovich-integration via mid-point Riemann-sum t 1 2 approximations = Bs ∂Bs = B (1st order calculus!) ) 0 2 t R
Itô integration
Fact: Typical sample paths of Brownian motion, t Bt (ω), have in…nite variation on every interval. 7! How to de…ne integration against Brownian motion? Itô’sbrilliant idea: with some help and intuition from martingale theory, 1 f (t, ω) dBt (ω) Z0 can be de…ned for reasonable non-anticipating f : start with simple integrands and complete with isometry
1 2 1 2 E f (t, ω) dBt (ω) = E f (t, ω) dt . "Z0 # Z0
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 5/22 Fact: Itô-integrals have left-point Riemann-sum approximations. De…ne Stratonovich-integration via mid-point Riemann-sum t 1 2 approximations = Bs ∂Bs = B (1st order calculus!) ) 0 2 t R
Itô integration
Fact: Typical sample paths of Brownian motion, t Bt (ω), have in…nite variation on every interval. 7! How to de…ne integration against Brownian motion? Itô’sbrilliant idea: with some help and intuition from martingale theory, 1 f (t, ω) dBt (ω) Z0 can be de…ned for reasonable non-anticipating f : start with simple integrands and complete with isometry
1 2 1 2 E f (t, ω) dBt (ω) = E f (t, ω) dt . "Z0 # Z0 t 1 2 Example: Bs dBs = B t ... 2nd order calculus! 0 2 t R
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 5/22 De…ne Stratonovich-integration via mid-point Riemann-sum t 1 2 approximations = Bs ∂Bs = B (1st order calculus!) ) 0 2 t R
Itô integration
Fact: Typical sample paths of Brownian motion, t Bt (ω), have in…nite variation on every interval. 7! How to de…ne integration against Brownian motion? Itô’sbrilliant idea: with some help and intuition from martingale theory, 1 f (t, ω) dBt (ω) Z0 can be de…ned for reasonable non-anticipating f : start with simple integrands and complete with isometry
1 2 1 2 E f (t, ω) dBt (ω) = E f (t, ω) dt . "Z0 # Z0 t 1 2 Example: Bs dBs = B t ... 2nd order calculus! 0 2 t Fact: Itô-integrals have left-point Riemann-sum approximations. R
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 5/22 Itô integration
Fact: Typical sample paths of Brownian motion, t Bt (ω), have in…nite variation on every interval. 7! How to de…ne integration against Brownian motion? Itô’sbrilliant idea: with some help and intuition from martingale theory, 1 f (t, ω) dBt (ω) Z0 can be de…ned for reasonable non-anticipating f : start with simple integrands and complete with isometry
1 2 1 2 E f (t, ω) dBt (ω) = E f (t, ω) dt . "Z0 # Z0 t 1 2 Example: Bs dBs = B t ... 2nd order calculus! 0 2 t Fact: Itô-integrals have left-point Riemann-sum approximations. R De…ne Stratonovich-integration via mid-point Riemann-sum t 1 2 approximations = Bs ∂Bs = B (1st order calculus!) ) 0 2 t P. K. Friz (TU Berlin and WIAS) R rough paths, gap ODE/SDEs December2009 5/22 e Let V0, ... , Vd be a collection of nice vector …elds on R
A solution (process) y = yt (ω) to
d i dy = V0 (y) dt + ∑ Vi (y) ∂B i=1 is, by de…nition, a solution to corresponding integral equation.
At the price of modifying the drift vector …eld V0 we can switch to Itô formulation (∂B dB) ! Existence, uniqueness by …xpoint arguments.
For simplicity only: from here on V0 = 0.
Stochastic di¤erential equations
Let B be a d-dimensional Brownian motion.
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 6/22 A solution (process) y = yt (ω) to
d i dy = V0 (y) dt + ∑ Vi (y) ∂B i=1 is, by de…nition, a solution to corresponding integral equation.
At the price of modifying the drift vector …eld V0 we can switch to Itô formulation (∂B dB) ! Existence, uniqueness by …xpoint arguments.
For simplicity only: from here on V0 = 0.
Stochastic di¤erential equations
Let B be a d-dimensional Brownian motion. e Let V0, ... , Vd be a collection of nice vector …elds on R
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 6/22 At the price of modifying the drift vector …eld V0 we can switch to Itô formulation (∂B dB) ! Existence, uniqueness by …xpoint arguments.
For simplicity only: from here on V0 = 0.
Stochastic di¤erential equations
Let B be a d-dimensional Brownian motion. e Let V0, ... , Vd be a collection of nice vector …elds on R
A solution (process) y = yt (ω) to
d i dy = V0 (y) dt + ∑ Vi (y) ∂B i=1 is, by de…nition, a solution to corresponding integral equation.
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 6/22 Existence, uniqueness by …xpoint arguments.
For simplicity only: from here on V0 = 0.
Stochastic di¤erential equations
Let B be a d-dimensional Brownian motion. e Let V0, ... , Vd be a collection of nice vector …elds on R
A solution (process) y = yt (ω) to
d i dy = V0 (y) dt + ∑ Vi (y) ∂B i=1 is, by de…nition, a solution to corresponding integral equation.
At the price of modifying the drift vector …eld V0 we can switch to Itô formulation (∂B dB) !
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 6/22 For simplicity only: from here on V0 = 0.
Stochastic di¤erential equations
Let B be a d-dimensional Brownian motion. e Let V0, ... , Vd be a collection of nice vector …elds on R
A solution (process) y = yt (ω) to
d i dy = V0 (y) dt + ∑ Vi (y) ∂B i=1 is, by de…nition, a solution to corresponding integral equation.
At the price of modifying the drift vector …eld V0 we can switch to Itô formulation (∂B dB) ! Existence, uniqueness by …xpoint arguments.
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 6/22 Stochastic di¤erential equations
Let B be a d-dimensional Brownian motion. e Let V0, ... , Vd be a collection of nice vector …elds on R
A solution (process) y = yt (ω) to
d i dy = V0 (y) dt + ∑ Vi (y) ∂B i=1 is, by de…nition, a solution to corresponding integral equation.
At the price of modifying the drift vector …eld V0 we can switch to Itô formulation (∂B dB) ! Existence, uniqueness by …xpoint arguments.
For simplicity only: from here on V0 = 0.
P. K. Friz (TU Berlin and WIAS) rough paths, gap ODE/SDEs December2009 6/22 Let V be a nice vector …eld on Re Aim: …nd solution to SDE
dy = V (y) ∂dB.
Let etV be the solution ‡ow to the ODE z˙ = V (z) . Then