pith. sign in

arxiv: 2605.24136 · v1 · pith:NGIGGMCFnew · submitted 2026-05-22 · 📊 stat.ML · cs.LG· stat.CO

Detecting Metastable Basins in High Dimensions via Marginal Trajectory Distribution Discrimination

Taj Jones-McCormick This is my paper

Pith reviewed 2026-06-30 14:56 UTC · model grok-4.3

classification 📊 stat.ML cs.LGstat.CO
keywords metastable basinsMarkov processestrajectory distributionsbasin detectionclassification riskhigh-dimensional dynamicsneural discrimination
0
0 comments X

The pith

Basin membership reduces to whether a Bayes classifier on marginal trajectory distributions achieves risk near 1/2 or near zero.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that metastable basins can be recovered from trajectory samples alone by testing whether pairs of starting states produce distinguishable marginal trajectory distributions. It proves that the Bayes-optimal risk for distinguishing those distributions is near 1/2 when the states share a basin and near zero when they do not. This turns basin identification into an iterative sequence of two-sample discrimination tasks that a neural network can approximate. The method is evaluated on high-dimensional systems with embedded low-dimensional dynamics where spectral and clustering baselines fail to recover the correct partitions.

Core claim

The central claim is a risk-separation result: for a time-homogeneous Markov process, if two initial states lie in the same basin then the Bayes risk of the optimal classifier between their marginal trajectory distributions approaches 1/2, whereas if the states lie in distinct basins the risk approaches zero. Basin detection is thereby reduced to estimating these risks with a neural network that approximates the Bayes classifier and iteratively merging candidate representatives whose estimated risk is near 1/2.

What carries the argument

The risk-separation theorem on marginal trajectory distributions, which converts basin detection into a sequence of two-sample classification problems solved by neural approximation of the Bayes classifier.

If this is right

  • Spectral and spatial-discretization methods become unnecessary once trajectory discrimination is available.
  • The procedure works on reducible processes where inter-basin transitions are arbitrarily rare on the sampling timescale.
  • High-dimensional ambient noise does not destroy recoverability provided the low-dimensional basin dynamics remain embedded in the marginal distributions.
  • Only a modest number of short trajectories per candidate representative is required once the neural classifier is trained.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Trajectory sampling alone may suffice for basin recovery without ever constructing an explicit transition operator.
  • The same risk-separation idea could be tested on non-Markovian processes whose finite-time marginals still separate by basin.
  • The number of trajectories needed for stable risk estimation may grow with ambient dimension in a manner independent of the intrinsic basin geometry.

Load-bearing premise

The neural network approximates the Bayes-optimal classifier closely enough that its estimated risk reliably signals whether two states belong to the same basin.

What would settle it

Apply the iterative merging procedure to a known metastable system with documented basins and observe whether points from different basins are merged or points from the same basin remain separate when the neural risk estimates deviate from the predicted separation.

Figures

Figures reproduced from arXiv: 2605.24136 by Taj Jones-McCormick.

Figure 1
Figure 1. Figure 1: The figure contains trajectories plotted by cluster label for various different methods. The top row corresponds to the MALA process running on a simple ’double ring’ energy function composed of two Gaussian rings. The bottom row corresponds to the same structure embedded in 100 dimensional ambient space. In this case, we plot only the low dimensional subspace in which the separability of the basins can be… view at source ↗
Figure 2
Figure 2. Figure 2: visualizations of samples drawn from the low dimensional stationary distributions of the MALA processes used in experiments. B.2.2. Gaussian Mixtures. The second experiment we consider is simple Gaussian mixtures in 100 dimensional space. We select some arbitrary number of components between 0 and 50 and the locations for these components are selected randomly from the sphere of radius 10 (i.e. √ d). This … view at source ↗
Figure 3
Figure 3. Figure 3: The plots a) and b) display trajecto￾ries colored by their as￾signed basin from the trained NBI. Plot a) cor￾responds to a model trained only on the 2- d projection of the tra￾jectories whereas plot b) corresponds to cluster￾ing according to a model trained on the full 66 di￾mensional system. Plot c) displays contours for the density of the tra￾jectories plotted above it as well as a coloring de￾noting the… view at source ↗
Figure 4
Figure 4. Figure 4: The figure contains trajectories plotted by cluster label for various different methods. The top row corresponds to the MALA process running on a simple 2d Gaussian Mixture energy function composed of two 3 isotropic Gaussians. The bottom row corresponds to the same structure embedded in 100 dimensional ambient space. In this case, we plot only the low dimensional subspace in which the separability of the … view at source ↗
Figure 5
Figure 5. Figure 5: The figure contains trajectories plotted by cluster label for various different methods. The left corresponds to the MALA process running on a 3d ’helix’ energy function. The right column corresponds to the same structure embedded in 100 dimensional ambient space. In this case, we plot only the low dimensional subspace in which the separability of the basins can be visualized [PITH_FULL_IMAGE:figures/full… view at source ↗
read the original abstract

We study the problem of identifying dynamically distinct basins of attraction in high dimensional time-homogeneous Markov processes using only trajectory sampling. This problem is fundamental in the analysis of metastable dynamical systems, where the process rapidly mixes within basins while transitions between basins occur rarely on the timescale of interest, or even when the state space is reducible. Existing approaches typically rely on spatial discretization or spectral analysis of estimated transition operators, which can become unreliable in high dimensional settings or when the underlying basin geometry is highly nonlinear. We propose a discriminative approach to basin identification based on marginal trajectory distribution comparison. We prove a simple risk separation result: if two initial states belong to the same basin, the Bayes-optimal classifier distinguishing their marginal trajectory distributions achieves risk close to 1/2, whereas if they lie in distinct basins, the optimal risk is close to zero. This observation reduces basin detection to a two-sample discrimination problem between marginal trajectory distributions. Motivated by this principle, we develop a neural algorithm that receives a set of candidate basin representatives and iteratively merges them by estimating classification risk with a neural network that approximates the Bayes classifier. We evaluate the method on various metastable systems. These include synthetic systems constructed by embedding low-dimensional dynamics into high dimensional noisy ambient spaces. In these settings, standard spectral and clustering-based methods often fail, while our approach accurately recovers the underlying basin structure. These results display a shortcoming of existing methods and highlight trajectory discrimination as an effective tool for identifying dynamical basins in high dimensional stochastic systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that metastable basins in high-dimensional time-homogeneous Markov processes can be identified from trajectory samples alone by proving a risk-separation result: the Bayes-optimal classifier on marginal trajectory distributions from same-basin initial states has risk near 1/2, while distinct-basin states yield risk near 0. This reduces basin detection to iterative two-sample discrimination, implemented via a neural network that approximates the Bayes classifier to estimate risk and merge candidate representatives. The approach is shown to recover basin structure on synthetic embeddings of low-dimensional dynamics into high-dimensional noisy spaces, where spectral and clustering baselines fail.

Significance. If the separation theorem holds and the neural estimator reliably tracks the true Bayes risk, the work supplies a mathematically grounded alternative to discretization or spectral methods that degrade in high dimensions or with nonlinear basin geometry. Credit is due for the clean risk-separation statement and for the reproducible synthetic experiments that isolate the failure modes of existing techniques. The significance is limited by the absence of approximation-error controls, which directly affects whether the empirical success generalizes beyond the tested synthetics.

major comments (2)
  1. [§3] §3 (Risk separation theorem and algorithm): The central reduction to two-sample discrimination is load-bearing, yet the manuscript provides no quantitative bound on the neural-network approximation error to the Bayes classifier. In high ambient dimension this gap can be large even when the true marginals are separable, rendering the estimated risk unreliable for the iterative merge decisions; the synthetic experiments do not isolate or bound this gap.
  2. [§4] §4 (Experimental evaluation): The reported recovery of basin structure on embedded low-dimensional systems is encouraging, but the controls do not include ablations that vary network capacity or training schedule while holding the true marginal separation fixed; without such controls it is impossible to attribute success to the risk-separation principle rather than to favorable approximation behavior on the chosen synthetics.
minor comments (2)
  1. [§2] Notation for marginal trajectory distributions is introduced without an explicit equation reference; adding a displayed definition would improve traceability to the risk-separation claim.
  2. [§4] Figure captions for the synthetic embeddings should state the ambient dimension and noise level explicitly rather than referring only to 'high-dimensional' settings.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the value of the risk-separation result. We address each major comment below and outline planned revisions.

read point-by-point responses

  1. Referee: [§3] §3 (Risk separation theorem and algorithm): The central reduction to two-sample discrimination is load-bearing, yet the manuscript provides no quantitative bound on the neural-network approximation error to the Bayes classifier. In high ambient dimension this gap can be large even when the true marginals are separable, rendering the estimated risk unreliable for the iterative merge decisions; the synthetic experiments do not isolate or bound this gap.

    Authors: We agree that the manuscript provides no quantitative bounds on the approximation error between the neural network and the Bayes classifier. The risk-separation theorem applies to the true Bayes risk (near 1/2 for same-basin states and near 0 for distinct basins), while the algorithm relies on a neural approximation whose error is uncontrolled in the current text. This is a genuine limitation for claiming reliability of the iterative merge procedure in arbitrary high-dimensional settings. In the revised manuscript we will add an explicit discussion of this gap, including a statement that the empirical success on the tested synthetics does not yet guarantee performance when the approximation error is large, and we will include a brief analysis of how the estimated risk behaves under increasing network capacity on the same embedded systems. revision: yes

  2. Referee: [§4] §4 (Experimental evaluation): The reported recovery of basin structure on embedded low-dimensional systems is encouraging, but the controls do not include ablations that vary network capacity or training schedule while holding the true marginal separation fixed; without such controls it is impossible to attribute success to the risk-separation principle rather than to favorable approximation behavior on the chosen synthetics.

    Authors: We concur that the current experiments lack ablations that vary network capacity or training schedule while keeping the underlying marginal separation fixed. Such controls are necessary to separate the contribution of the theoretical risk-separation principle from favorable approximation properties on the specific synthetics. In the revision we will add these ablations: for each embedded system we will train networks of varying width and depth (and with different training schedules) on the same trajectory data, report the resulting estimated risks and final merge decisions, and compare against the known ground-truth basin structure. This will provide direct evidence on the sensitivity of the procedure to approximation quality. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper's core step is a mathematical risk-separation claim derived from the definitions of basins and marginal trajectory distributions under time-homogeneous Markov dynamics; this is a direct probabilistic argument (same-basin states yield overlapping marginals, distinct basins yield separable ones) that does not reduce to any fitted parameter, self-citation, or ansatz imported from prior work by the same author. The subsequent neural approximation is explicitly described as an estimator of the Bayes risk rather than part of the theorem itself, and no equation or claim equates the output to its inputs by construction. The method is therefore not circular.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on a mathematical risk separation result whose derivation is not visible in the abstract, plus the practical assumption that a neural network can serve as a reliable proxy for the Bayes classifier. The main added value is therefore the algorithmic procedure and its claimed empirical performance on synthetic high-dimensional embeddings. No new physical entities are postulated.

free parameters (1)
  • Neural network architecture, optimizer, and training schedule
    The algorithm relies on a neural network to approximate the Bayes classifier; all architectural and optimization choices are free parameters whose effect on risk estimation is not quantified in the abstract.
axioms (2)
  • domain assumption The underlying process is a time-homogeneous Markov process
    Stated explicitly as the problem setting.
  • domain assumption Rapid mixing occurs inside basins while transitions between basins are rare on the observation timescale
    Described as the defining property of the metastable systems under study.

pith-pipeline@v0.9.1-grok · 5796 in / 1468 out tokens · 69121 ms · 2026-06-30T14:56:24.096629+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Online stochastic gradient descent on non-convex losses from high-dimensional inference.Journal of Machine Learning Research, 22(106):1–51, 2021

    Gerard Ben Arous, Reza Gheissari, and Aukosh Jagannath. Online stochastic gradient descent on non-convex losses from high-dimensional inference.Journal of Machine Learning Research, 22(106):1–51, 2021

    2021

  2. [2]

    Springer Science & Business Media, 2013

    Gregory R Bowman, Vijay S Pande, and Frank Noé.An introduction to Markov state models and their application to long timescale molecular simulation. Springer Science & Business Media, 2013

    2013

  3. [3]

    siamese

    Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a" siamese" time delay neural network.Advances in neural information processing systems, 6, 1993

    1993

  4. [4]

    CRC press, 2011

    Steve Brooks, Andrew Gelman, Galin Jones, and Xiao-Li Meng.Handbook of markov chain monte carlo. CRC press, 2011

    2011

  5. [5]

    Exploration by Random Network Distillation

    Yuri Burda, Harrison Edwards, Amos Storkey, and Oleg Klimov. Exploration by random network distillation.arXiv preprint arXiv:1810.12894, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  6. [6]

    Hierarchical density estimates for data clustering, visualization, and outlier detection.ACM Transactions on Knowledge Discovery from Data (TKDD), 10(1):1–51, 2015

    Ricardo JGB Campello, Davoud Moulavi, Arthur Zimek, and Jörg Sander. Hierarchical density estimates for data clustering, visualization, and outlier detection.ACM Transactions on Knowledge Discovery from Data (TKDD), 10(1):1–51, 2015

    2015

  7. [7]

    Diffusion maps.Applied and computational harmonic analysis, 21(1):5–30, 2006

    Ronald R Coifman and Stéphane Lafon. Diffusion maps.Applied and computational harmonic analysis, 21(1):5–30, 2006

    2006

  8. [8]

    Identification of almost invariant aggregates in reversible nearly uncoupled markov chains.Linear Algebra and its Applications, 315(1-3):39–59, 2000

    Peter Deuflhard, Wilhelm Huisinga, Alexander Fischer, and Ch Schütte. Identification of almost invariant aggregates in reversible nearly uncoupled markov chains.Linear Algebra and its Applications, 315(1-3):39–59, 2000

    2000

  9. [9]

    Robust perron cluster analysis in conformation dynamics

    Peter Deuflhard and Marcus Weber. Robust perron cluster analysis in conformation dynamics. Linear algebra and its applications, 398:161–184, 2005

    2005

  10. [10]

    Openmm 8: molecular dynamics simulation with machine learning potentials.The Journal of Physical Chemistry B, 128(1):109–116, 2023

    Peter Eastman, Raimondas Galvelis, Raúl P Peláez, Charlles RA Abreu, Stephen E Farr, Emilio Gallicchio, Anton Gorenko, Michael M Henry, Frank Hu, Jing Huang, et al. Openmm 8: molecular dynamics simulation with machine learning potentials.The Journal of Physical Chemistry B, 128(1):109–116, 2023

    2023

  11. [11]

    A density-based algorithm for discovering clusters in large spatial databases with noise

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. Inkdd, volume 96, pages 226–231, 1996

    1996

  12. [12]

    Verso Books, 2020

    Emmanuel Farjoun and Moshe Machover.Laws of Chaos: A probabilistic approach to political economy. Verso Books, 2020

    2020

  13. [13]

    An analytic framework for identifying finite-time coherent sets in time-dependent dynamical systems.Physica D: Nonlinear Phenomena, 250:1–19, 2013

    Gary Froyland. An analytic framework for identifying finite-time coherent sets in time-dependent dynamical systems.Physica D: Nonlinear Phenomena, 250:1–19, 2013

    2013

  14. [14]

    Deep sparse rectifier neural networks

    Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315–323. JMLR Workshop and Conference Proceedings, 2011

    2011

  15. [15]

    Dimensionality reduction by learning an invariant mapping

    Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06), volume 2, pages 1735–1742. IEEE, 2006

    2006

  16. [16]

    Comparing partitions.Journal of classification, 2(1):193– 218, 1985

    Lawrence Hubert and Phipps Arabie. Comparing partitions.Journal of classification, 2(1):193– 218, 1985

    1985

  17. [17]

    On gradient descent algorithm for generalized phase retrieval problem

    Li Ji and Zhou Tie. On gradient descent algorithm for generalized phase retrieval problem. In 2016 IEEE 13th International Conference on Signal Processing (ICSP), pages 320–325. IEEE, 2016

    2016

  18. [18]

    Hamiltonian systems and transformation in hilbert space.Proceedings of the National Academy of Sciences, 17(5):315–318, 1931

    Bernard O Koopman. Hamiltonian systems and transformation in hilbert space.Proceedings of the National Academy of Sciences, 17(5):315–318, 1931. 12 TAJ JONES-MCCORMICKTEJONESM@UWATERLOO.CA

    1931

  19. [19]

    Vampnets for deep learning of molecular kinetics.Nature communications, 9(1):5, 2018

    Andreas Mardt, Luca Pasquali, Hao Wu, and Frank Noé. Vampnets for deep learning of molecular kinetics.Nature communications, 9(1):5, 2018

    2018

  20. [20]

    Some methods of classification and analysis of multivariate observations

    James B McQueen. Some methods of classification and analysis of multivariate observations. In Proc. of 5th Berkeley Symposium on Math. Stat. and Prob., pages 281–297, 1967

    1967

  21. [21]

    Stochasticity helps to navigate rough landscapes: comparing gradient-descent-based algorithms in the phase retrieval problem

    Francesca Mignacco, Pierfrancesco Urbani, and Lenka Zdeborová. Stochasticity helps to navigate rough landscapes: comparing gradient-descent-based algorithms in the phase retrieval problem. Machine Learning: Science and Technology, 2(3):035029, 2021

    2021

  22. [22]

    On spectral clustering: Analysis and an algorithm

    Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001

    2001

  23. [23]

    Kinetic distance and kinetic maps from molecular dynamics simulation.Journal of chemical theory and computation, 11(10):5002–5011, 2015

    Frank Noé and Cecilia Clementi. Kinetic distance and kinetic maps from molecular dynamics simulation.Journal of chemical theory and computation, 11(10):5002–5011, 2015

    2015

  24. [24]

    A variational approach to modeling slow processes in stochastic dynamical systems.Multiscale Modeling & Simulation, 11(2):635–655, 2013

    Frank Noé and Feliks Nuske. A variational approach to modeling slow processes in stochastic dynamical systems.Multiscale Modeling & Simulation, 11(2):635–655, 2013

    2013

  25. [25]

    Number 2

    James R Norris.Markov chains. Number 2. Cambridge university press, 1998

    1998

  26. [26]

    Characterizing metastable states with the help of machine learning.Journal of Chemical Theory and Computa- tion, 18(9):5195–5202, 2022

    Pietro Novelli, Luigi Bonati, Massimiliano Pontil, and Michele Parrinello. Characterizing metastable states with the help of machine learning.Journal of Chemical Theory and Computa- tion, 18(9):5195–5202, 2022

    2022

  27. [27]

    Identification of slow molecular order parameters for markov model construction.The Journal of chemical physics, 139(1), 2013

    Guillermo Pérez-Hernández, Fabian Paul, Toni Giorgino, Gianni De Fabritiis, and Frank Noé. Identification of slow molecular order parameters for markov model construction.The Journal of chemical physics, 139(1), 2013

    2013

  28. [28]

    Markov models of molecular kinetics: Generation and validation.The Journal of chemical physics, 134(17), 2011

    Jan-Hendrik Prinz, Hao Wu, Marco Sarich, Bettina Keller, Martin Senne, Martin Held, John D Chodera, Christof Schütte, and Frank Noé. Markov models of molecular kinetics: Generation and validation.The Journal of chemical physics, 134(17), 2011

    2011

  29. [29]

    Springer, 2004

    Christian P Robert, George Casella, and George Casella.Monte Carlo statistical methods, volume 2. Springer, 2004

    2004

  30. [30]

    Exponential convergence of langevin distributions and their discrete approximations

    Gareth O Roberts and Richard L Tweedie. Exponential convergence of langevin distributions and their discrete approximations. 1996

    1996

  31. [31]

    On the approximation quality of markov state models.Multiscale Modeling & Simulation, 8(4):1154–1177, 2010

    Marco Sarich, Frank Noé, and Christof Schütte. On the approximation quality of markov state models.Multiscale Modeling & Simulation, 8(4):1154–1177, 2010

    2010

  32. [32]

    Springer Science & Business Media, 2006

    Eugene Seneta.Non-negative matrices and Markov chains. Springer Science & Business Media, 2006

    2006

  33. [33]

    World Scientific, 2010

    Gabriel Stoltz, Mathias Rousset, et al.Free energy computations: A mathematical perspective. World Scientific, 2010

    2010

  34. [34]

    Cluster ensembles—a knowledge reuse framework for combining multiple partitions.Journal of machine learning research, 3(Dec):583–617, 2002

    Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for combining multiple partitions.Journal of machine learning research, 3(Dec):583–617, 2002

    2002

  35. [35]

    Online stochastic gradient descent with arbitrary initialization solves non-smooth, non-convex phase retrieval.arXiv preprint arXiv:1910.12837, 2019

    Yan Shuo Tan and Roman Vershynin. Online stochastic gradient descent with arbitrary initialization solves non-smooth, non-convex phase retrieval.arXiv preprint arXiv:1910.12837, 2019

    work page arXiv 1910

  36. [36]

    Depth-first search and linear graph algorithms.SIAM journal on computing, 1(2):146–160, 1972

    Robert Tarjan. Depth-first search and linear graph algorithms.SIAM journal on computing, 1(2):146–160, 1972

    1972

  37. [37]

    Cambridge Molecular Science

    David Wales.Energy Landscapes: Applications to Clusters, Biomolecules and Glasses. Cambridge Molecular Science. Cambridge University Press, 2004

    2004

  38. [38]

    Variational koopman models: Slow collective variables and molecular kinetics from short off-equilibrium simulations.The Journal of chemical physics, 146(15), 2017

    Hao Wu, Feliks Nüske, Fabian Paul, Stefan Klus, Péter Koltai, and Frank Noé. Variational koopman models: Slow collective variables and molecular kinetics from short off-equilibrium simulations.The Journal of chemical physics, 146(15), 2017. DETECTING METASTABLE BASINS IN HIGH DIMENSIONS VIA MARGINAL TRAJECTORY DISTRIBUTION DISCRIMINATION 13 AppendixA.Proo...

    2017