DAMTP, University of Cambridge and Wolfson College
Klas Modin and Milo Viviani. Lie–Poisson methods for isospectral flows. Foundations of Computational Mathematics, 20:889–921, 2020.
asdasdasd
arxiv:2408.16701
Stochastic rigid body
Stochastic (sine) Euler equations
Stochastic point vortex dynamics
Lie group \(G\) with algebra \(\mathfrak{g}\)
Hamiltonian \(H\colon \mathfrak{g}^* \to \mathbb R\)
Two ingredients
Important geometric structure:
LP Systems evolve on coadjoint orbits
(symplectic manifolds that foliate space!)
+ Several preserved quantities, Hamiltonian, Casimirs etc.
Coadjoint orbits
Dynamics on orbit
\(G \subset \operatorname{GL}(n)\) is compact, simply connected and \(J\)-quadratic, i.e., \(A \in \mathfrak{g} \iff AJ + JA^* = 0 \)
Examples: \(\mathfrak{so}(N), \mathfrak{su}(N), \mathfrak{sp}(N)\).
In this setting, LP flow is isospectral:
Classic example: rigid body (ODE)
How to add noise to equations?
Why Strato?
Why these coefficients?
We need a chain rule
Same bracket-type coefficient: same geometric structure
Stochastic LP Systems evolve on coadjoint orbits and have Casimirs!
Stochastic LP Systems evolve on coadjoint orbits and have Casimirs!
Coadjoint orbits
With transport noise
With non-transport noise
What to assume to prove the existence of solutions?
E.g. Lipschitz is not realistic to assume
But! Some smoothness is sufficient for global existence
System remains on compact level sets of Casimirs
\(\rightarrow\) truncation argument to prove existence
Classic example: rigid body
Off the shelf integrator?
Non-physical behavior!
Structure-preserving integrators?
Wish list:
= Lie–Poisson integrator
Also: Avoid exponential map! (expensive, matrices in our applications are \(\sim2000\times2000\) )
LP systems on \(\mathfrak{g}^*\) are reductions of Hamiltonian systems on \(T^*G \cong G \times \mathfrak{g}^*\)
\(H\colon \mathfrak{g}^* \to \mathbb R\) lifts to left-invariant Hamiltonian!
\(G\) acts on \(T^*G\) by \(g.(Q,P) = (gQ,g^{-*}P)\)
Associated momentum map:
Do for all Hamiltonians!
Note: No a priori guarantee that this system remains on \(T^*G\)!
Embedded into \((R^{n \times n})^2\)
However! Remains on \(T^*G\)
\(X_t = \mu(Q_t,P_t) \in \mathfrak{g}^*\) satisfies
\(\Phi_h \colon T^*G \to T^* G\): \(G\)-equivariant symplectic integrator \(\rightarrow\)
Reduces to a Lie–Poisson integrator by \(\mu\)
Exploit the unreduction!
Do we know such an integrator?
Implicit midpoint!
Stochastic isospectral midpoint:
Take \(Q_n,P_n\) such that \(\mu(Q_n,P_n) = X_n\)
Take \(X_{n+1} = \mu(\Phi_h(Q_n,P_n))\)
Exploit the unreduction!
Check the literature for error analysis for implicit midpoint
1) Show that the implicit midpoint method does not travel too far:
\(\sup_{h \geq 0} \sup_{n \geq 0} \|Q_n,P_n\| \leq R(Q_0,P_0). \)
2) Truncate the CHS system on \(T^*G\)
Upstairs
Downstairs
3) Apply error analysis from literature
Reduction by \(\mu\)
4) Convergence of method for truncated LP system
5) Truncated LP system = LP system
Recipe:
Geometric structure of equations and its preservation is central. Used to prove
Without structure preservation, no convergence guarantees
Strong error
Weak error
Rigid body
500 realizations
10^7 realizations
\(\phi(X) = \sin(2\pi x_1) + \sin(2\pi x_2) + \sin(2\pi x_3)\).
Zeitlin is a convenient computational method for 2D ideal spherical hydrodynamics.
Matrices are very large, makes exponential map/algebra to group mappings too expensive.