📊 statistics

Data-driven sparse identification of governing PDEs via knockoff filters and multi-criteria trade-offs

The paper proposes KO-PDE-IDENT, a data-driven framework that combines model-X knockoff filters for false discovery rate control, recursive feature elimination, and multi-criteria decision-making to accurately identify parsimonious partial differential equations from noisy observations while eliminating spurious terms caused by multicollinearity.

Original authors: Pongpisit Thanasutives, Naichang Ke, Yoshinobu Kawahara

Published 2026-05-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pongpisit Thanasutives, Naichang Ke, Yoshinobu Kawahara

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 solve a mystery: What are the hidden laws of physics governing a specific system? Maybe it's how a fluid flows, how a disease spreads, or how a chemical reaction happens. You have a massive notebook full of data (observations), but the data is messy, noisy, and full of distractions.

The paper introduces a new detective tool called KO-PDE-IDENT. Its job is to sift through the noise and find the exact few sentences (equations) that describe the system, while ignoring the thousands of fake clues that look important but aren't.

Here is how it works, broken down into three simple steps using everyday analogies:

The Problem: The "Noisy Library"

Imagine you walk into a library with millions of books. You know the answer to your mystery is written in just three specific books, but the library is filled with:

Noise: Books that are just random scribbles.
Clones: Books that say almost the same thing as the real ones, making it hard to tell which is the original.
Distractions: Books that seem relevant but are actually red herrings.

Traditional methods try to pick the "best" books by looking at them one by one. But because the library is so crowded and the books are so similar (a problem called multicollinearity), these methods often grab the wrong books or miss the real ones.

The Solution: The Three-Stage Detective Process

KO-PDE-IDENT solves this with a three-step strategy:

Step 1: The "Fake Twin" Test (Knockoff Filters)

The Goal: Create a safety net to catch fake clues.

Imagine you have a suspect (a candidate term in your equation). To see if they are truly guilty, you create a "Fake Twin" (a "knockoff").

This Fake Twin looks exactly like the real suspect in every way except one thing: it has zero connection to the actual mystery (the physics).
The detective compares the real suspect against their Fake Twin.
If the real suspect is much more "important" than the Fake Twin, they stay on the list. If they look just as suspicious as the Fake Twin, they are likely innocent and get kicked out.

The Magic: This process uses a statistical rule called FDR Control. Think of this as a "guaranteed error rate." The method promises: "We might make a few mistakes, but we will never let more than X% of the innocent people (fake terms) into our final suspect list." It's like a security checkpoint that guarantees no more than 1 in 10 innocent people get through, even if the crowd is chaotic.

Step 2: The "Trimming" (Recursive Feature Elimination)

The Goal: Cut the fat from the list.

After Step 1, you have a shortlist of suspects. It's much smaller than the original library, but it might still have a few "imposter" terms that slipped through because they were so similar to the real ones.

The detective now uses a SHAP tool (a way to measure how much each word contributes to the story).

They look at the list and ask: "If we remove this specific word, does the story fall apart?"
They use a clever trick: they swap the real word with its "Fake Twin" again. If the story works just as well (or better) with the Fake Twin, the real word was unnecessary.
They keep cutting away the unnecessary words until only the essential ones remain.

Step 3: The "Taste Test" (Multi-Criteria Decision Making)

The Goal: Pick the perfect recipe.

Now you have a very small group of candidates. You might have a few different combinations of terms that all look "okay." Which one is the true law of physics?

Instead of just picking the one that fits the data best (which might be too complicated), the detective uses a Panel of Judges (Multi-Criteria Decision Making). They evaluate every candidate equation based on six different criteria, like:

Accuracy: Does it predict the future correctly?
Simplicity: Is it the simplest explanation? (Occam's Razor).
Confidence: Are we sure about the numbers in the equation?
Reliability: Does it hold up under different conditions?

The five judges (named TOPSIS, VIKOR, COMET, etc.) vote on the best equation. They don't just look at one thing; they balance accuracy against simplicity. The winner is the equation that strikes the perfect balance—simple enough to be elegant, but accurate enough to be true.

The Results: Did it work?

The authors tested this detective on five famous physics puzzles (like how waves move or how chemicals react) that were covered in heavy "noise" (like static on a radio).

The Result: KO-PDE-IDENT found the exact correct equations for all five puzzles.
The Score: It had zero fake terms (false discoveries) and found 100% of the real terms.
The Comparison: Old methods got lost in the noise, picking dozens of wrong terms. KO-PDE-IDENT stayed focused and found the truth.

In a Nutshell

KO-PDE-IDENT is a new way to discover the laws of nature from messy data. Instead of guessing and checking, it uses statistical "Fake Twins" to filter out the noise, smart trimming to remove leftovers, and a panel of judges to pick the most elegant, accurate answer. It turns the chaotic search for physics equations into a rigorous, trustworthy process.

Technical Summary: Data-driven sparse identification of governing PDEs via knockoff filters and multi-criteria trade-offs

Problem Statement
The discovery of parsimonious partial differential equations (PDEs) from noisy observational data is a central challenge in data-driven science. Existing frameworks, such as Sparse Identification of Nonlinear Dynamics (SINDy), typically construct large, overcomplete libraries of candidate terms (polynomials and spatial derivatives) and apply sparse regression to identify the governing equation. However, these candidate libraries inherently suffer from extreme multicollinearity. This property renders standard sparse regression algorithms susceptible to observational noise, often leading to the selection of spurious terms (false discoveries) rather than the true physical mechanisms. Furthermore, conventional methods relying on heuristic thresholding, cross-validation, or information criteria (e.g., AIC, BIC) lack formal statistical guarantees regarding the False Discovery Rate (FDR). Consequently, practitioners cannot mathematically assure that a discovered equation captures genuine physics rather than noise artifacts. While Bayesian approaches and deep symbolic architectures offer uncertainty quantification or flexibility, they do not provide finite-sample control over the FDR.

Methodology: KO-PDE-IDENT
The authors propose KO-PDE-IDENT, a three-stage framework that shifts PDE discovery from heuristic optimization to principled statistical inference. The method integrates knockoff filtering for rigorous screening, recursive feature elimination for refinement, and multi-criteria decision-making (MCDM) for final selection.

FDR-Controlled Screening via Model-X Knockoff Filters:
The first stage addresses the multicollinearity and noise issues by constructing a support set of candidate terms with a guaranteed finite-sample FDR.
- Weak Formulation: To mitigate noise amplification inherent in numerical differentiation, the PDE is formulated in a weak (integral) sense using smooth weight functions, transferring derivatives onto the noiseless test functions.
- Knockoff Construction: For each candidate term in the library, a synthetic "knockoff" copy is generated. These copies are statistically indistinguishable from the originals in their correlations with other terms but are conditionally independent of the response (the time derivative). This allows them to serve as negative controls.
- Feature Statistics (SHAP-DS): The framework introduces a new feature statistic, SHAP-DS, which couples an $\ell_0$ -constrained adaptive best-subset selection (abess) solver with SHapley Additive exPlanations (SHAP). The statistic is defined as the difference between the SHAP importance of the original term and its knockoff ( $W_j = Z_j - \tilde{Z}_j$ ). This approach is computationally efficient and satisfies the "flip-sign" property required for valid knockoff inference.
- Aggregation: To reduce the randomness inherent in single knockoff realizations, the method aggregates multiple realizations using the e-value-based Benjamini–Hochberg (e-BH) procedure. This derandomizes the filter while preserving FDR control. The framework adaptively tunes the target FDR to ensure the support set remains overcomplete (containing all true terms) while minimizing false positives.
Refinement via Recursive Feature Elimination (RFE):
Even with FDR control, highly correlated spurious terms may persist. The second stage refines the support set using RFE in the feature-attribution space.
- SHAP Ranking: Terms are ranked by their mean absolute SHAP values.
- Knockoff-Perturbed Hypothesis Testing: For lower-ranked terms, the method performs a one-sided Wilcoxon signed-rank test. It compares the penalized least-squares loss of the original model against a perturbed model where the candidate term is replaced by its knockoff copy. If the knockoff-substituted model performs significantly better (or equally well), the original term is deemed statistically unnecessary and removed. This stage ensures that only terms with genuine marginal contributions are retained.
Model Selection via Multi-Criteria Decision-Making (MCDM):
The final stage selects the single optimal governing equation from the refined set of candidates. Recognizing that single information criteria are insufficient to balance complexity, accuracy, and uncertainty, the authors formulate this as an MCDM problem.
- Candidate Generation: Best-subset selection enumerates all possible PDE alternatives over the reduced support.
- Evaluation Criteria: Each alternative is scored on six distinct criteria: (1) distributional predictive accuracy (via conformalised quantile regression), (2) structural complexity, (3) robust information complexity (RICOMP-MM), (4) coefficient uncertainty (via recentered influence function regression), (5) miscalibration (via Murphy–Ehm decomposition), and (6) a coverage-width criterion.
- Ranking Aggregation: An ensemble of five MCDM algorithms (TOPSIS, VIKOR, COMET, PROMETHEE-II, and CoCoSo) generates individual rankings. These are aggregated using consensus methods (Borda count, Schulze, etc.) to identify the optimal governing equation that best balances the trade-offs.

Key Contributions

First Application of Knockoff Filters to PDE Discovery: The paper introduces the first framework to apply model-X knockoff filters to data-driven PDE identification, providing the first explicit, finite-sample bound on the proportion of false discoveries.
SHAP-DS Statistic: The authors propose a novel, computationally efficient feature statistic combining abess and SHAP, which outperforms traditional swap-type statistics in speed while maintaining effectiveness against spurious terms.
RFE with Knockoff Perturbations: A new refinement procedure that uses knockoff-perturbed hypothesis testing to validate the statistical necessity of terms, effectively eliminating false positives that survive the initial screening.
MCDM for Model Selection: The framework reformulates the final equation choice as a multi-criteria problem, aggregating diverse metrics (accuracy, complexity, uncertainty, calibration) to robustly select the optimal model, a capability previously unexplored in this domain.

Empirical Results
The framework was validated on five canonical PDE systems under severe noise conditions: 1D Burgers, Korteweg–de Vries (KdV), and Kuramoto–Sivashinsky (KS) equations (50% noise); a 2D Reaction–Diffusion system (10% noise); and a 3D Gray–Scott system (0.1% noise).

Structural Recovery: KO-PDE-IDENT achieved an empirical FDR (eFDR) of zero and an empirical power (ePOWER) of one across all benchmarks. It exactly recovered the true PDE structure, eliminating false discoveries while retaining all true underlying terms.
Coefficient Accuracy: The method achieved low coefficient estimation errors (below 5% for 1D systems and below 1% for multi-dimensional systems).
Comparison: The framework consistently outperformed notable baselines such as Sequential Thresholded Least Squares (STLSQ), SR3, and SSR, which failed to control FDR and selected excessive spurious terms under high noise.

Significance
The paper claims that KO-PDE-IDENT advances data-driven PDE discovery by transforming it from an exercise of loss optimization and hyperparameter tuning into a rigorous statistical inference procedure. By ensuring that the candidate set entering the model selection stage carries an explicit, theoretically guaranteed bound on its false discovery proportion, the framework brings the rigor of high-dimensional inference to a domain where it has been conspicuously absent. The combination of a statistically rigorous screening stage with principled empirical refinement allows for the reliable discovery of governing physical laws in complex, noisy dynamical systems, providing practitioners with mathematically rigorous assurance that discovered equations capture genuine mechanisms rather than noise artifacts.