betaFeedback

WorksMathematics

Local Koopman modes of saddle-point diffusions: exact singular functions and a data-driven estimator

Abstract

Coherent sets-regions of phase space that resist mixing-organize transport in stochastic dynamical systems and are encoded in the leading singular functions of the stochastic Koopman operator. These are typically computed numerically and globally; a local, analytic theory near a fixed point has been missing, although fixed points and the transitions between them organize the flow. We supply the local theory for a saddle, the transition state through which probability flows between basins. Naively the construction fails-probability escapes along unstable directions, the invariant measure is non-normalizable, the singular functions leave L2L^2, and singular values exceed one. We show that conditioning on non-escape (a Doob's hh-transform) repairs this; that the required hh is the unique stabilizing solution of an algebraic Riccati equation; and that the resulting slow modes decay at the escape rate along the unstable manifold. We obtain these modes in closed form-products of Hermite polynomials in canonical coordinates times the hh-function-and use their unitary structure to derive a data-driven estimator: canonical correlation analysis on conditioned trajectories. When a single unstable direction dominates, the leading mode is the reaction coordinate, locally recovering the stable manifold and committor of transition-path theory. The construction connects coherent sets to transition-path theory, to the stabilizing Riccati solution of linear control, and to the time-reversed dynamics of generative diffusion models.

Introduction

A nonlinear dynamical system is organized by its fixed points and by the transitions between them: trajectories dwell near stable equilibria and cross between them through the neighborhoods of unstable equilibria [1, 2]. Local behavior around fixed points therefore controls the global dynamics, and characterizing it is a natural route to understanding the system. For deterministic dynamics, this local analysis is standard—the Jacobian’s eigenvectors give normal modes and its eigenvalues give growth rates. For stochastic dynamics one would like the analogous local theory of the slow, collective modes that govern how probability is transported through the neighborhood.

Those modes are the singular functions of the stochastic Koopman operator whose singular values lie close to one, the spectral objects underlying coherent sets [3, 4, 5]: subsets of phase space that remain dynamically coherent rather than dispersing. At present, coherent sets are studied almost exclusively numerically and globally: one discretizes a transfer or Koopman operator over the entire phase space of a nonlinear system—using Ulam-type discretizations [6, 7], finite elements [8], extended dynamic mode decomposition [9], or diffusion maps [10]—and reads off a partition [3, 5]. This yields neither a closed form nor a local law: no statement of what the coherent modes look like near a given fixed point. The present paper supplies that missing local, analytic description.

A stable fixed point (a sink) linearizes to a contracting Ornstein–Uhlenbeck (OU) process: an isolated metastable well with no transport through it. The dynamically interesting fixed points are saddles—hyperbolic points with at least one unstable direction. A saddle is a transition state: the bottleneck through which trajectories pass between basins, where reaction coordinates and committor functions are defined [11, 12, 13]. Its linearization is a continuous OU process whose drift matrix has eigenvalues of both signs, and this is the regime we solve.

The breakdown is easy to see. When the drift matrix has unstable eigenvalues, there is no normalizable stationary measure: probability escapes along the unstable manifold. As a consequence, the naive extension of singular functions from stable fixed points to saddles leads to non-integrable functions and singular values greater than one [14] (Figure 2b). This breaks the singular-value problem at its foundation. The singular value represents the correlation coefficient between the left singular function evaluated at the initial time and the right singular function evaluated at time τ\tau. In the coherent-set picture, where these functions approximate indicators of regions, this correlation is essentially the fraction of trajectories that begin in the coherent set and remain there after time τ\tau—a survival probability, which cannot exceed one. The breakdown is not merely analytic: the data-driven route fails in tandem, because the empirical covariances that any CCA-type estimator [5, 15, 16] relies on diverge along the unstable directions as the trajectories escape, so the correlation matrix has no finite limit and the canonical problem is ill-posed on the raw data (Figure 2b). Any correct theory of the saddle case must first restore a proper Hilbert space.

Linear stochastic dynamical system

Figure 2. Linear stochastic dynamical system. (a) Demonstration of a 2D linear saddle system with A=[−1.0.50.0.7]A = \begin{bmatrix} -1. & 0.5 \\ 0. & 0.7 \end{bmatrix} and D=[1.0.30.31.]D = \begin{bmatrix} 1. & 0.3 \\ 0.3 & 1. \end{bmatrix}. The eigenvectors, raw and stabilized drift, and analytic v1v_1 are plotted. (b) Canonical correlations calculated analytically (Eq. (28)) using raw/stabilized dynamics and computed numerically using naive/conditioned CCA for varying t1t_1 and fixed τ\tau. (c) The cosine similarity between analytic v1v_1 and empirical v^1\hat{v}_1 for increasing sample sizes.

Our resolution, and the conceptual core of the paper, is to condition the dynamics: restricting to trajectories that have not yet escaped is a Doob’s hh-transform [17] that stabilizes the drift and makes the slow modes well defined. We show (i) that the required conditioning is the unique stabilizing solution of an algebraic Riccati equation [18], selected automatically as the attractor of a Riccati flow (Sec. III); (ii) that the conditioned slow modes admit closed form whose decay rate is the escape rate along the unstable manifold, and that multiplication by hh is a unitary map relating them to ordinary Hermite modes (Sec. IV); (iii) that this unitary structure yields a data-driven estimator—canonical correlation analysis on conditioned trajectories—which recovers the analytic modes exactly (§ 1.0.4); (iv) that in the special case with a single slow unstable mode, the leading predictive direction converges to the unstable left eigenvector, justifying the reaction coordinate in classical theories (Sec. VI); and (v) that the construction establishes bridges to transition-path theory, stochastic thermodynamics, linear control, and diffusion-model time reversal (Sec. VII).

Setup and main result

Local linearization of stochastic nonlinear dynamics

Consider a stochastic nonlinear system on Rd\mathbb{R}^{d},

dx(t)dt=b(x)+ξ(t),⟨ξ(t)ξ(t′)T⟩=Dδ(t−t′),(1)\frac{\mathrm{d}x(t)}{\mathrm{d}t}=b(x)+\xi(t),\qquad\langle\xi(t)\xi(t')^{\mathsf{T}}\rangle=D\delta(t-t'), \tag*{(1)}

where b:Rd→Rdb:\mathbb{R}^{d}\to\mathbb{R}^{d} is the drift and D≻0D\succ0 is the constant diffusion matrix. Let x∗x_{*} be a fixed point, b(x∗)=0b(x_{*})=0. Writing A=∇b(x∗)A=\nabla b(x_{*}) for the Jacobian and shifting the origin to x∗x_{*}, the local dynamics is the OU process

dx(t)dt=Ax(t)+ξ(t).(2)\frac{\mathrm{d}x(t)}{\mathrm{d}t}=Ax(t)+\xi(t). \tag*{(2)}

We assume for clarity that the eigenvalues and eigenvectors of AA are real (the complex case is discussed in Appendix S4 D). Denote by λi\lambda_i the eigenvalues of AA, by rir_i the right eigenvectors (Ari=λiri)(Ar_i=\lambda_i r_i), and by ℓi\ell_i the left eigenvectors (ATℓi=λiℓi)(A^{\mathsf{T}}\ell_i=\lambda_i\ell_i), normalized biorthogonally, ℓiTrj=δij\ell_i^{\mathsf{T}}r_j=\delta_{ij}. We split the spectrum into a stable part (eigenvalues λi−<0\lambda_i^{-}<0, eigenvectors ri−,ℓi−r_i^{-},\ell_i^{-}) and an unstable part (eigenvalues λi+>0\lambda_i^{+}>0, eigenvectors ri+,ℓi+r_i^{+},\ell_i^{+}), collect the unstable left eigenvectors into the columns of L+L_{+} and the unstable eigenvalues into Λ+=diag⁡(λi+)\Lambda_{+}=\operatorname{diag}(\lambda_i^{+}), so that ATL+=L+Λ+A^{\mathsf{T}}L_{+}=L_{+}\Lambda_{+}. A saddle is the case where both parts are nonempty.

Fokker–Planck operator and the Koopman generator

The density p(x,t)p(x,t) obeys the Fokker–Planck equation ∂tp=Lp\partial_t p=\mathcal{L}p with the forward generator

Lf=−∇⋅(Axf)+12Dαβ∂α∂βf.(3)\mathcal{L}f=-\nabla\cdot(Ax f)+\frac{1}{2}D^{\alpha\beta}\partial_{\alpha}\partial_{\beta}f. \tag*{(3)}

Its adjoint with respect to the Euclidean inner product ⟨f,g⟩=∫fg dx\langle f,g\rangle=\int fg\,\mathrm{d}x is the backward Kolmogorov operator, which is exactly the generator of the finite-time stochastic Koopman operator:

L†g=(Ax)⋅∇g+12Dαβ∂α∂βg,Kτ=eL†τ,(4)\mathcal{L}^{\dagger}g=(Ax)\cdot\nabla g+\frac{1}{2}D^{\alpha\beta}\partial_{\alpha}\partial_{\beta}g,\qquad\mathcal{K}_{\tau}=e^{\mathcal{L}^{\dagger}\tau}, \tag*{(4)}

where the Koopman operator Kτ\mathcal{K}_{\tau} propagates observables g:Rd→Rg:\mathbb{R}^{d}\to\mathbb{R} by the conditional expectation [19]

(Kτg)(x):=E[g(x(τ))∣x(0)=x].(5)(\mathcal{K}_{\tau}g)(x):=\mathbb{E}\left[g(x(\tau))\mid x(0)=x\right]. \tag*{(5)}

A stationary measure satisfying Lρ0=0\mathcal{L}\rho_0=0 is given by ρ0(x)∝exp⁡(−12xTΣ−1x)\rho_0(x)\propto\exp\left(-\frac{1}{2}x^{\mathsf{T}}\Sigma^{-1}x\right), with Σ\Sigma solving the Lyapunov equation

AΣ+ΣAT=−D.(6)A\Sigma+\Sigma A^{\mathsf{T}}=-D. \tag*{(6)}

Eq. (6) has a unique solution iff no two eigenvalues of AA sum to zero [20], and Σ≻0\Sigma\succ0 is the covariance iff AA is Hurwitz (all λi<0\lambda_i<0). If AA has unstable eigenvalues, Σ\Sigma is indefinite (negative definite if all modes are unstable); then ρ0\rho_0 grows in the unstable directions and is not normalizable. Although such a non-normalizable ρ0\rho_0 does not define a probability distribution on Rd\mathbb{R}^{d}, it remains a well-defined local stationary profile for the linearized dynamics near the fixed point. This is the root of the obstruction described in § 1.

Singular functions of the Koopman operator

Coherent modes are defined variationally. Given a reference measure μ\mu with ⟨f,g⟩μ:=∫fg dμ\langle f,g\rangle_{\mu}:=\int fg\,\mathrm{d}\mu and ∥f∥μ:=⟨f,f⟩μ\lVert f\rVert_{\mu}:=\sqrt{\langle f,f\rangle_{\mu}}, the leading singular value and functions solve

σ1=max⁡f,g∈L2(μ)⟨f,Kτg⟩μ∥f∥μ∥g∥μ,(7)\sigma_1=\max_{f,g\in L^2(\mu)}\frac{\langle f,\mathcal{K}_{\tau}g\rangle_{\mu}}{\lVert f\rVert_{\mu}\lVert g\rVert_{\mu}}, \tag*{(7)}

where the maximizers are the leading left singular function f1f_1 and right singular function g1g_1. Further modes maximize the same ratio subject to μ\mu-orthogonality to the previous ones. Introducing the μ\mu-adjoint Kτ†\mathcal{K}_{\tau}^{\dagger} via ⟨f,Kτg⟩μ=⟨Kτ†f,g⟩μ\langle f,\mathcal{K}_{\tau}g\rangle_{\mu}=\langle\mathcal{K}_{\tau}^{\dagger}f,g\rangle_{\mu}, and the forward-backward and backward-forward operators Fτ=KτKτ†\mathcal{F}_{\tau}=\mathcal{K}_{\tau}\mathcal{K}_{\tau}^{\dagger} and Bτ=Kτ†Kτ\mathcal{B}_{\tau}=\mathcal{K}_{\tau}^{\dagger}\mathcal{K}_{\tau}, the singular functions are equivalently given by the following eigenfunctions [19]:

Fτfi=σi2fi,Bτgi=σi2gi,Kτgi=σifi.(8)\mathcal{F}_{\tau}f_i=\sigma_i^2f_i,\qquad\mathcal{B}_{\tau}g_i=\sigma_i^2g_i,\qquad\mathcal{K}_{\tau}g_i=\sigma_i f_i. \tag*{(8)}

Typically, μ\mu is taken as the invariant measure. However, when AA has unstable modes, the invariant measure is non-normalizable, the naive stable-OU polynomial singular functions are not integrable with respect to this measure, and the ratio in Eq. (7) can exceed one—the σ>1\sigma>1 pathology of § 1.

Summary of the main result

References

  1. [1]S. H. Strogatz, Nonlinear dynamics and chaos: with applications to physics, biology, chemistry, and engineering (Chapman and Hall/CRC, 2024).
  2. [2]M. I. Freidlin and A. D. Wentzell, Random Perturbations of Dynamical Systems, 3rd ed. (Springer, 2012).
  3. [3]G. Froyland, An analytic framework for identifying finite-time coherent sets in time-dependent dynamical systems, Physica D: Nonlinear Phenomena 250, 1 (2013).DOI
  4. [4]G. Froyland, Dynamic isoperimetry and the geometry of Lagrangian coherent structures, Nonlinearity 28, 3587 (2015).
  5. [5]S. Klus, B. E. Husic, M. Mollenhauer, and F. Noé, Kernel methods for detecting coherent structures in dynamical data, Chaos: An Interdisciplinary Journal of Nonlinear Science 29 (2019).DOI
  6. [6]S. M. Ulam, A Collection of Mathematical Problems (Interscience Publishers, New York, 1960).
  7. [7]T.-Y. Li, Finite approximation for the frobenius-perron operator. a solution to ulam’s conjecture, Journal of Approximation theory 17, 177 (1976).DOI
  8. [8]G. Froyland and O. Junge, Robust fem-based extraction of finite-time coherent sets using scattered, sparse, and incomplete trajectories, SIAM Journal on Applied Dynamical Systems 17, 1891 (2018).arxiv.org/abs/1705.03640
  9. [9]M. O. Williams, I. G. Kevrekidis, and C. W. Rowley, A data-driven approximation of the koopman operator: Extending dynamic mode decomposition, Journal of Nonlinear Science 25, 1307 (2015).
  10. [10]R. R. Coifman and S. Lafon, Diffusion maps, Applied and computational harmonic analysis 21, 5 (2006).DOI
  11. [11]P. Hänggi, P. Talkner, and M. Borkovec, Reaction-rate theory: fifty years after kramers, Reviews of modern physics 62, 251 (1990).DOI
  12. [12]W. E and E. Vanden-Eijnden, Towards a theory of transition paths, Journal of statistical physics 123, 503 (2006).
  13. [13]P. G. Bolhuis, D. Chandler, C. Dellago, and P. L. Geissler, Transition path sampling: Throwing ropes over rough mountain passes, in the dark, Annual review of physical chemistry 53, 291 (2002).
  14. [14]J. L. Pughe-Sanford, X. Ding, J. Moore, A. Sengupta, C. Epstein, P. Greengard, and D. Chklovskii, Neurons as detectors of coherent sets in sensory dynamics, Advances in Neural Information Processing Systems 38, 158904 (2026).
  15. [15]H. Hotelling, Relations between two sets of variates, in Breakthroughs in statistics: methodology and distribution (Springer, 1992) pp. 162–190.
  16. [16]H. Wu and F. Noé, Variational approach for learning markov processes from time series data, Journal of Nonlinear Science 30, 23 (2020).arxiv.org/abs/1707.04659
  17. [17]J. L. Doob, Conditional brownian motion and the boundary limits of harmonic functions, Bulletin de la Société mathématique de France 85, 431 (1957).DOI
  18. [18]P. Lancaster and L. Rodman, Algebraic riccati equations (Clarendon press, 1995).DOI
  19. [19]S. Klus and N. Djurdjevac Conrad, Dynamical systems and complex networks: A koopman operator perspective, Journal of Physics: Complexity 5, 041001 (2024).
  20. [20]R. A. Horn and C. R. Johnson, Matrix analysis (Cambridge university press, 2012).
  21. [21]B. D. Anderson, Reverse-time diffusion equation models, Stochastic Processes and their Applications 12, 313 (1982).DOI
  22. [22]Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, Score-based generative modeling through stochastic differential equations, in International Conference on Learning Representations (2021).
  23. [23]H. Akaike, Markovian representation of stochastic processes by canonical variables, SIAM journal on control 13, 162 (1975).DOI
  24. [24]M. Fiedler, Algebraic connectivity of graphs, Czechoslovak mathematical journal 23, 298 (1973).DOI
  25. [25]H. A. Kramers, Brownian motion in a field of force and the diffusion model of chemical reactions, physica 7, 284 (1940).DOI
  26. [26]J. S. Langer, Statistical theory of the decay of metastable states, Annals of Physics 54, 258 (1969).DOI
  27. [27]L. Perko, Differential equations and dynamical systems, Vol. 7 (Springer Science & Business Media, 2013).
  28. [28]A. Berezhkovskii and A. Szabo, One-dimensional reaction diffusion activated rate processes in many dimensions, The Journal of chemical physics 122, 014503 (2005).pubmed.ncbi.nlm.nih.gov/15638670
  29. [29]A. Laio and M. Parrinello, Escaping free-energy minima, Proceedings of the national academy of sciences 99, 12562 (2002).DOI
  30. [30]O. Valsson, P. Tiwary, and M. Parrinello, Enhancing important fluctuations: Rare events and metadynamics from a conceptual viewpoint, Annual review of physical chemistry 67, 159 (2016).DOI
  31. [31]P. Tiwary and B. Berne, Spectral gap optimization of order parameters for sampling complex molecular systems, Proceedings of the National Academy of Sciences 113, 2839 (2016).DOI
  32. [32]M. Pavon, Canonical correlations of past inputs and future outputs for linear stochastic systems, Systems & control letters 4, 209 (1984).DOI
  33. [33]P. Collet, S. Martínez, and J. San Martín, Quasi-stationary distributions, Markov chains, diffusions and dynamical systems (2013).DOI
  34. [34]D. Ruelle, Resonances of chaotic dynamical systems, Physical review letters 56, 405 (1986).
  35. [35]A. Bohm, J. D. Dollard, and M. Gadella, Dirac Kets, Gamow Vectors and Gel’fand Triplets: The Rigged Hilbert Space Formulation of Quantum Mechanics Lectures in Mathematical Physics at the University of Texas at Austin (Springer, 1989).

Paper details

vlnc:W_2hqddzat
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents