📊 statistics

Symbolic Weak-form Recovery of 2-D Stochastic Generators

This paper introduces WG-SINDy, a novel estimator that combines covariance-shaped kernels, ridge-stabilized projections, and positive-semidefinite constraints to recover two-dimensional Ito generators from trajectory data, demonstrating successful recovery on 19 out of 29 synthetic systems while acknowledging limitations regarding universal or real-data applicability.

Original authors: Sai Sathvik Gullipalli, Eshwar R A

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sai Sathvik Gullipalli, Eshwar R A

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out the secret recipe of a chaotic, invisible machine just by watching a few marbles bounce around inside it. This machine is a "stochastic generator"—a mathematical engine that drives everything from stock markets to the way molecules wiggle in your body. The recipe has two main ingredients: a drift (the average direction the marbles want to go) and a diffusion (the random, jittery kicks they get from invisible wind).

For a long time, detectives could only solve this puzzle if the machine had just one dimension (like a marble rolling on a single track). But what if the machine is two-dimensional, like a marble rolling on a flat table where it can move left-right and forward-backward, and the wind pushing it might push it diagonally?

That's the challenge this paper tackles. The authors, Sai Sathvik Gullipalli and Eshwar R A, built a new detective tool called WG-SINDy. Here is how it works, what it can do, and where it hits a wall.

The Problem: Why the Old Way Failed

If you tried to solve this 2D puzzle using the old 1D method (the "naive port"), it would be like trying to read a complex map using a magnifying glass meant for a single street. It fails in three specific ways:

The Math Gets Messy: The numbers get so unbalanced (like trying to measure a mountain and a pebble with the same ruler) that the computer's calculations go haywire.
The Noise is Uneven: In some spots, the "wind" is gentle; in others, it's a hurricane. The old method treated all spots as if the wind was the same, leading to bad guesses.
The Shape Breaks: The old method might guess a "diffusion" shape that is physically impossible, like a shadow that doesn't cast a shadow. In math terms, it might produce a matrix that isn't "positive semidefinite" (a fancy way of saying the shape is broken).

The Solution: A Smarter Detective Kit

The authors didn't just fix the old tool; they built a whole new kit with six special gadgets:

Shape-Shifting Lenses: Instead of looking at the data with a round, boring lens, they use "covariance-shaped" lenses that stretch and squeeze to fit the actual shape of the data cloud.
Local Polynomials: They don't just look at the average; they fit a tiny, curved polynomial curve to the local neighborhood of data points to smooth out the rough edges.
The "Feasible GLS" Filter: This gadget acts like a noise-canceling headphone specifically for the drift. It listens to the wind speed at each spot and turns down the volume on the noisy, high-wind areas so the signal doesn't get drowned out.
The "Cholesky" Safety Net: This is the most important safety feature. Even if the math gets messy, this gadget forces the final answer to be a valid, physical shape. It reconstructs the diffusion tensor using a specific mathematical recipe (Cholesky decomposition) that guarantees the result is never broken.
Adaptive LASSO: This is a "pruning shears" that cuts away the tiny, fake ingredients in the recipe, keeping only the big, important ones.
Pooling: They combine data from many different "marble runs" (trajectories) to get a clearer picture.

The Results: What Actually Worked

The authors tested this new tool on 29 different synthetic 2D systems. Think of these as 29 different virtual worlds with different rules for how the marbles move.

The Success Rate: In 19 of these systems, the tool passed the test. It successfully recovered the drift and the diffusion tensor.
- For these 19 winners, the "drift metric" (how close the guess was to the real drift) had a median score of 0.204.
- The "tensor error" (how close the guess was to the real diffusion) had a median score of 0.0397.
- When there was a diagonal "cross-wind" (off-diagonal entry), the tool guessed the direction with a cosine score of 0.997 (almost perfect).
- Crucially, the tool never produced a broken shape; the "positive semidefinite" validity was 1.00 (100%) by design.
The Failures (The "Honest Limits"): The tool did not work for the other 10 systems, and the authors are very honest about why. They didn't hide these failures; they labeled them "named limits" or "scoped reviews."
- It failed when the drift signal was too weak (like trying to hear a whisper in a hurricane).
- It failed when the data didn't cover the whole area (like trying to map a whole country by only looking at one city).
- It failed when the "time step" (how often you check the marble's position) was too large.
- It failed for systems with non-polynomial drifts (where the rules of the game are too weird for the tool's library of shapes).

What This Is NOT

It is important to know what this paper doesn't claim.

It is not a universal fix. The authors explicitly state they do not claim "universal 2D recovery." If the data is bad or the library is wrong, the tool admits it.
It is not a real-world proof. All these results come from synthetic simulations. The authors are careful to say these are "in-sample sampled-region diagnostics." They haven't proven it works on real-world stock market data or real biological molecules yet.
It is not a "black box." The goal was to find the symbolic recipe (the actual math equation), not just a computer program that mimics the behavior.

The Bottom Line

This paper suggests that by combining a few clever mathematical tricks—specifically, using shape-shifting lenses, noise-canceling filters, and a safety-net construction—the authors have built a tool that can successfully reverse-engineer the hidden rules of 2D chaotic systems if the data is good and the rules are within a certain range.

In the 19 successful cases, the tool found the drift and diffusion with high precision. But in the 10 cases where the data was too sparse, the noise was too loud, or the rules were too weird, the tool hit a wall. The authors call these walls "honest limits," acknowledging that no amount of math can guess a recipe if the ingredients are missing from the kitchen.

So, for the 19 systems that passed, we have a working symbolic generator. For the rest, we know exactly why the tool stopped, and we know we need better data or a different library before we can claim victory.

Technical Summary: Symbolic Weak-Form Recovery of 2-D Stochastic Generators

Problem Statement
The paper addresses the inverse problem of reconstructing the infinitesimal generator of a two-dimensional Itô diffusion from discretely sampled trajectory data. The generator $L$ is defined by a drift vector $b(x)$ and a diffusion tensor $a(x) = \sigma(x)\sigma(x)^\top$ . While one-dimensional recovery methods exist, extending them to two dimensions presents three specific structural challenges:

Ill-conditioning: Design matrices built from raw bivariate monomials over scale-disparate state clouds (e.g., log-price vs. variance) are severely ill-conditioned.
Statistical Inefficiency: Drift increments have low signal-to-noise ratios, and unweighted normal equations fail to account for heteroscedastic noise across different regions of the state space.
Positive Semidefiniteness (PSD): Unconstrained regression of tensor entries ( $a_{11}, a_{12}, a_{22}$ ) does not guarantee the recovered matrix is a valid covariance (PSD), which is essential for physical interpretation and leverage calculations.

Methodology: WG-SINDy
The authors propose the WG-SINDy estimator, a weak-form spatial-kernel approach that extends the 1D framework of Eshwar and Honnavar [3] to coupled, non-diagonal 2D systems. The method avoids direct derivative estimation by projecting increments onto spatial test functions. The implementation relies on a data-dependent full-cloud smoother rather than fixed weights, meaning exact finite-sample martingale cancellation is not asserted for the released code, though it motivates the population-level identity.

The algorithm integrates six key mechanisms to address the 2D challenges:

Standardization and Anisotropic Kernels: States are standardized to remove axis-scale disparity. Spatial kernels are constructed using covariance-shaped bandwidths (anisotropic) rather than isotropic ones, adapting to the residual covariance structure of the standardized cloud.
Order-2 Local-Polynomial Projection: To eliminate first-order boundary bias inherent in Nadaraya–Watson estimators, the method employs a local-polynomial fit (quadratic basis) at each kernel center.
Feasible Generalized Least Squares (GLS): A per-component diagonal GLS pass reweights the drift estimation. Rows centered in high-diffusion regions are down-weighted based on a first-pass diffusion estimate to stabilize the solution against heteroscedastic noise.
Adaptive Sparsification: The method uses Adaptive-LASSO with weights derived from an initial ridge estimate, followed by OLS debiasing and Sequentially Thresholded Least Squares (STLSQ) pruning to enforce sparsity in the symbolic library.
PSD Cholesky Read-Out: To ensure the diffusion tensor is physically valid, the method does not regress $a_{11}, a_{12}, a_{22}$ directly. Instead, it fits unconstrained entries, projects them pointwise to the PSD cone via eigenvalue clipping, computes the Cholesky factors ( $\ell_{11}, \ell_{21}, \ell_{22}$ ), and then regresses these factors against the library. The final tensor is reconstructed as $\hat{a} = \hat{L}\hat{L}^\top$ , guaranteeing $\hat{a} \succeq 0$ .
Multi-Trajectory Pooling: Independent trajectories are pooled by stacking weak rows to improve coverage and sample efficiency.

Key Contributions

Algorithmic Extension: The first weak-form estimator capable of recovering the full 2D generator (both drift components and all three independent tensor entries) symbolically, handling non-diagonal diffusion.
Structural Guarantees: The Cholesky read-out mechanism ensures the recovered diffusion tensor is positive semidefinite by construction, a critical requirement often violated by naive regression.
Ablation Analysis: The paper isolates the contributions of pooling, feasible GLS, and local-polynomial smoothing as the primary drivers of in-scope performance, while identifying Cholesky recovery and adaptive sparsification as essential for structural validity and stress-regime protection.
Scope Definition: The authors explicitly define the boundaries of their claims, distinguishing between identifiable systems (where the library spans the true generator and data covers the state space) and "honest limits" where recovery fails due to low signal-to-noise, near-singular tensors, or insufficient coverage.

Results
The estimator was evaluated on 29 synthetic two-dimensional systems, including linear (Ornstein–Uhlenbeck), rotational, bistable, multiplicative, and financial (Heston, SABR) models.

Performance: 19 systems met their predeclared recovery contracts (PASS). Across these, the median central-grid drift metric was 0.204, and the median tensor error was 0.0397. For systems with non-degenerate off-diagonal targets, the median cosine similarity for $a_{12}$ was 0.997.
Validity: The recovered tensor was positive semidefinite at all grid points by construction.
Failure Modes: The paper explicitly reports on 8 "named limit" cases and 2 scoped reviews where recovery failed or was incomplete. These failures correspond to specific physical or data limitations: near-singular tensors, boundary concentration, non-polynomial drifts, poor state-space coverage, or excessively large time steps ( $\Delta t$ ).
Comparisons: The WG-SINDy estimator outperformed four in-repository baseline proxies (naive 1D port, local Kramers–Moyal, temporal Weak-SINDy, and generator-EDMD) in terms of in-scope scores, though the authors note these comparisons use unequal data budgets and are diagnostic rather than a definitive ranking of underlying published methods.

Significance and Claims
The paper claims to provide a robust, symbolic recovery method for 2D Itô generators within a declared scope of identifiability. It emphasizes that the results are synthetic, in-sample diagnostics and do not establish universal recovery or real-data applicability.

The authors explicitly state they do not claim exact finite-sample martingale cancellation or a feasible-GLS efficiency theorem for the released implementation, as the smoother is data-dependent.
The significance lies in demonstrating that a combination of weak-form projection, structural constraints (PSD), and variance stabilization (GLS) can successfully recover coupled stochastic dynamics where naive extensions of 1D methods fail.
The work serves as a benchmark for what is recoverable given specific data quality and library constraints, offering a "falsification boundary" where the method correctly identifies when the data is insufficient for symbolic recovery.