pith. sign in

arxiv: 2603.20904 · v6 · pith:T6NZZYXUnew · submitted 2026-03-21 · 📊 stat.ME · math-ph· math.DS· math.MP· nlin.CD· physics.data-an· stat.ML

Data-Driven Weak-form Discovery of Stochastic Systems

Eshwar R A , Gajanan V. Honnavar This is my paper

Pith reviewed 2026-05-21 10:35 UTC · model grok-4.3

classification 📊 stat.ME math-phmath.DSmath.MPnlin.CDphysics.data-anstat.ML
keywords stochastic differential equationsinfinitesimal generatorweak formGaussian kernelsdrift and diffusiondynamical invariantssparse regressionItô processes
0
0 comments X

The pith

Projecting Itô SDEs onto spatial Gaussian kernels removes endogeneity bias exactly and enables unbiased joint recovery of drift and diffusion.

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

The paper demonstrates that weak projections of stochastic differential equations onto temporal test functions create a bias of order O(T dt^{3/2}) that grows with the observation window and cannot be fixed by collecting more data. Switching to projections onto spatial Gaussian kernels eliminates this bias through F_{t_n}-measurability and the tower property, which ensure each regression row is exactly unbiased. The approach then recovers both the drift b(x) and diffusion a(x) in a single sparse regression step, yielding an explicit symbolic form of the infinitesimal generator. A reader would care because the generator directly encodes dynamical invariants such as spectral gaps, escape rates, and position-dependent relaxation timescales that can be extracted from observed trajectories.

Core claim

Weak projection onto temporal test functions produces an endogeneity bias of order O(T dt^{3/2}) that grows with the observation window and cannot be eliminated by additional data. Projecting instead onto spatial Gaussian kernels removes the bias exactly: F_{t_n}-measurability and the tower property guarantee unbiased regression rows at every step. The resulting framework jointly identifies the drift b(x) and diffusion a(x) from a single sparse regression, producing an explicit symbolic generator amenable to spectral analysis.

What carries the argument

Spatial Gaussian kernel projections that exploit F_{t_n}-measurability and the tower property to produce unbiased weak-form regression rows for the infinitesimal generator of an Itô process.

If this is right

  • Joint identification of drift b(x) and diffusion a(x) from one sparse regression on observed trajectories.
  • Explicit symbolic generator that can be subjected directly to spectral analysis.
  • Recovery of dynamical invariants including spectral gaps, Kramers escape rates, and position-dependent relaxation timescales.
  • Validation errors below 5% in coefficients and total-variation distances below 0.01 in stationary densities on benchmark systems.
  • Autocorrelation functions that reproduce true relaxation timescales.

Where Pith is reading between the lines

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

  • The unbiased generator could support downstream tasks such as control or prediction in systems where only trajectory samples are available.
  • Adaptations of the kernel projection might address partial observations or additive measurement noise common in experimental data.
  • The single-regression structure suggests computational advantages for online or streaming identification of stochastic dynamics.
  • Explicit symbolic output opens the possibility of combining the method with symbolic regression tools to discover unknown functional forms.

Load-bearing premise

The data must be generated by an Itô SDE and spatial Gaussian kernel projections with F_{t_n}-measurability and the tower property must produce exactly unbiased regression rows without further conditions on sampling or the process.

What would settle it

Simulate trajectories from a known Itô SDE, apply the spatial-kernel method, and check whether the recovered generator's eigenvalues and stationary density match the true values within the reported error levels of 5% coefficient error and 0.01 total-variation distance.

Figures

Figures reproduced from arXiv: 2603.20904 by Eshwar R A, Gajanan V. Honnavar.

Figure 1
Figure 1. Figure 1: Recovered vs. true drift and diffusion functions for all three benchmark systems. Blue solid lines show the ground truth; red dashed lines show the weak SINDy estimates from Algorithm 1. Top row (drift functions): OU process, b(x) = −θx, mean relative error 2.0%; double-well Langevin system, b(x) = x − x 3 , mean rel. err. 2.7%; multiplicative diffusion, b(x) = −2x, mean rel. err. 3.9%. Bottom row (diffusi… view at source ↗
Figure 2
Figure 2. Figure 2: LassoCV regularisation paths for all six sub-problems. Each panel plots the mean cross￾validated MSE (averaged over five trajectory folds) as a function of the regularisation strength α, shown with α decreasing from left to right. Red dashed vertical lines mark the selected α ∗ . Top row (drift): OU (α ∗ ≈ 1.6 × 10−3 ), double-well (≈ 4.9 × 10−5 ), multiplicative (≈ 2.1 × 10−4 ). Bottom row (diffusion): OU… view at source ↗
Figure 3
Figure 3. Figure 3: Stationary density: true SDE vs. recovered model. Densities are computed analytically using the Fokker–Planck formula π(x) ∝ a(x) −1 exp 2 R x 0 b(y)/a(y) dy . Blue solid lines show the true SDE density; red dashed lines show the recovered model density. The shaded region between the two curves quantifies the pointwise discrepancy. Left (OU): Gaussian stationary distribution reproduced with total variatio… view at source ↗
Figure 4
Figure 4. Figure 4: Autocorrelation check: true SDE vs. recovered model. Empirical autocorrelation functions computed from 200,000-step simulations of both the true dynamics and the recovered model. Light blue solid lines show the true SDE; red dashed lines show the recovered model. Left (OU): Recovered relaxation rate ˆθ = 0.980 (err. 2.0%) closely matches the analytical e −τ (black dotted). The recovered autocorrelation is … view at source ↗
Figure 5
Figure 5. Figure 5: Theoretical noise scaling: Weak Form vs Kramers–Moyal. All curves are purely analytical; no regression is performed. Left: KM noise magnitude σobs/∆t as a function of ∆t for three SNR levels. The noise diverges as ∆t → 0. Centre: WF effective noise σobs/ √ Nheff for the same SNR levels, where N = T /∆t and heff = p π/2 h. The noise grows only as √ ∆t and remains bounded as ∆t → 0. Right: Ratio of KM noise … view at source ↗
read the original abstract

We present an algorithm for learning the governing equations of a stochastic dynamical system from trajectory data. It recovers interpretable symbolic expressions for both the drift $b(x)$ and the diffusion $a(x)$ in a single pass, yielding a model that can be queried directly for relaxation timescales, metastable escape rates, and stationary distributions. Rather than estimating the dynamics one time step at a time, the algorithm averages each candidate term across the whole trajectory before regressing; a drift-informed correction further removes the finite-sampling bias in the diffusion estimate, cutting it from 4.6% to 0.6% for state-dependent noise. We also show that the trajectory averaging must use a spatial rather than a temporal weighting: temporal weighting, as in existing weak-form methods, is biased for stochastic data with an error that grows with dataset size. On three benchmark systems -- the Ornstein--Uhlenbeck process, a double-well Langevin system, and a multiplicative-noise system -- the algorithm recovers all coefficients to within 5%, stationary densities to within 0.01 in total variation, and escape rates that match the true dynamics.

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 manuscript claims that weak projection of an Itô SDE onto temporal test functions induces an endogeneity bias of order O(T dt^{3/2}), while projection onto spatial Gaussian kernels evaluated at the left endpoint X_{t_n} removes this bias exactly. F_{t_n}-measurability combined with the tower property is invoked to guarantee unbiased regression rows for both the drift b(x) and diffusion a(x), enabling their joint recovery from a single sparse regression that yields an explicit symbolic generator suitable for spectral analysis of invariants such as spectral gaps and relaxation timescales. Validation on three benchmark systems is reported to produce coefficient errors below 5%, stationary-density total-variation distances below 0.01, and autocorrelation functions that match true timescales.

Significance. If the exact unbiasedness for both coefficients holds under the stated Itô-SDE assumption and left-endpoint discretization, the framework would constitute a useful advance in data-driven identification of stochastic generators. The joint recovery of b(x) and a(x) from one regression and the direct production of a symbolic operator for subsequent spectral computations are clear strengths that could facilitate computation of dynamical invariants without separate post-processing steps.

major comments (2)
  1. [§3.2, Eq. (9)] §3.2, Eq. (9): the weak-form integral for the diffusion coefficient a(x) involves an integral of a(X_s) times the second derivative of the test function over [t_n, t_{n+1}]. The manuscript must explicitly specify the quadrature (left-endpoint evaluation at X_{t_n}) and confirm that this choice, together with F_{t_n}-measurability, preserves the tower-property unbiasedness; any symmetric or trapezoidal rule would correlate with the Brownian increment and reintroduce bias.
  2. [§5.1, Table 1] §5.1, Table 1: the reported coefficient errors below 5% and TV distances below 0.01 are given without error bars, data-exclusion criteria, or a direct quantitative comparison against the temporal-projection baseline that is asserted to carry O(T dt^{3/2}) bias; this weakens the empirical support for the exact-bias-removal claim.
minor comments (2)
  1. The abstract contains the typographical error 'energerator' (should be 'generator').
  2. [§2] Notation for the infinitesimal generator is introduced as L_op in the abstract but appears as script-L in §2; a single consistent symbol should be used throughout.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation and explicitness of the method.

read point-by-point responses

  1. Referee: [§3.2, Eq. (9)] the weak-form integral for the diffusion coefficient a(x) involves an integral of a(X_s) times the second derivative of the test function over [t_n, t_{n+1}]. The manuscript must explicitly specify the quadrature (left-endpoint evaluation at X_{t_n}) and confirm that this choice, together with F_{t_n}-measurability, preserves the tower-property unbiasedness; any symmetric or trapezoidal rule would correlate with the Brownian increment and reintroduce bias.

    Authors: We agree with the referee that the quadrature rule requires explicit specification for clarity. The manuscript already employs left-endpoint evaluation at X_{t_n} for all integrals, including the diffusion term, to maintain consistency with the F_{t_n}-measurability of the test-function derivatives. In the revision we will add an explicit statement in §3.2 confirming that this left-endpoint choice, combined with the tower property, ensures the conditional expectation of the Itô integral vanishes exactly. We also concur that symmetric or trapezoidal quadrature would correlate with the Brownian increment and reintroduce bias, which is why the left-endpoint discretization is essential to the unbiasedness claim. revision: yes

  2. Referee: [§5.1, Table 1] the reported coefficient errors below 5% and TV distances below 0.01 are given without error bars, data-exclusion criteria, or a direct quantitative comparison against the temporal-projection baseline that is asserted to carry O(T dt^{3/2}) bias; this weakens the empirical support for the exact-bias-removal claim.

    Authors: We acknowledge that the empirical section would benefit from additional statistical detail and direct comparison. In the revised manuscript we will augment Table 1 (and the accompanying text in §5.1) with error bars obtained from multiple independent trajectories, state the precise data-exclusion criteria used (regression-residual thresholds and minimum trajectory length), and add a side-by-side quantitative comparison against the temporal-projection baseline on the same three benchmark systems. This will furnish explicit numerical evidence for the bias reduction achieved by the spatial-kernel formulation. revision: yes

Circularity Check

0 steps flagged

No circularity: unbiasedness follows from standard tower property and measurability

full rationale

The paper's central derivation applies the tower property of conditional expectation and F_tn-measurability of spatial Gaussian kernels to establish unbiased regression rows for both drift and diffusion terms. These are external facts from stochastic calculus, not derived from the paper's own fitted coefficients, self-citations, or data-dependent definitions. The abstract explicitly contrasts this with the temporal test-function approach that retains O(T dt^{3/2}) bias, showing the spatial projection step is justified independently rather than by construction or renaming. No load-bearing equation reduces a claimed prediction to a fitted input, and the joint identification of b(x) and a(x) via sparse regression remains a distinct algorithmic contribution. The result is therefore self-contained against external probabilistic benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the Itô SDE form and standard properties of stochastic filtrations; no free parameters or new entities are introduced in the abstract.

axioms (2)
  • domain assumption The observed process is generated by an Itô SDE with drift b(x) and diffusion a(x).
    Invoked as the governing equation whose weak projection is analyzed.
  • standard math F_tn-measurability and the tower property hold for the natural filtration of the process.
    Used to guarantee that spatial kernel projections yield unbiased regression rows.

pith-pipeline@v0.9.0 · 5708 in / 1459 out tokens · 93386 ms · 2026-05-21T10:35:34.127864+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.