PSI-SINDy: Post-Selection Inference for Sparse Identification of Nonlinear Dynamics

arXiv:2610.11486 · stat.ML, cs.LG · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "PSI-SINDy: Post-Selection Inference for Sparse Identification of Nonlinear Dynamics".

Jane: Detailed Research Summary:

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper today, "PSI-SINDy: Post-Selection Inference for Sparse Identification of Nonlinear Dynamics." It tackles a real headache in using SINDy.

Jane: Exactly. The main problem is that if you use the same noisy data to pick the dynamical terms and then test how good those terms actually are, you introduce selection bias, right?

Lu: That's it. The paper proposes PSI-SINDy as a statistical framework to quantify how reliable those selected dynamical terms actually are through hypothesis tests and confidence intervals.

Meng: So we're talking about making sure the math we do after picking our model isn't just guessing, but has some actual statistical grounding.

Tom: Right. The paper introduces this "four-view data thinning construction" to tackle the measurement error in the candidate terms and shared noise between the response and design matrices.

Jane: It splits one observed trajectory into four independent views: V1 for term selection, V2 for constructing the testing response, V3 for a corrected design matrix, and V4 for an instrument that gives us an independent estimate of the noise-free design.

Lu: That's key because it separates the process of selection from the process of inference, and it isolates the testing response from both the noisy design and shared noise effects.

Meng: So we get a cleaner way to look at things without all that messy data interference.

Tom: And they set up two targets for inference. The primary target is what coefficient each term would have in the best linear approximation if we had the noise-free trajectory available, but since we don't have it, they introduce an auxiliary coefficient using the corrected design and instrument from independent views.

Jane: They show that this auxiliary target admits finite-sample valid conditional tests under some assumptions in Section three of their paper <ref:2610.11486#pg2,this auxiliary target admits finite-sample valid conditional tests>.

Lu: That leads into their bridge idea. They develop an asymptotic bridge connecting this exact inference for the auxiliary target to asymptotically valid inference for the primary best linear approximation coefficient within the selected model.

Meng: It’s like they get a precise answer for a related thing first, and then show it gets close to what we really want as you have more data.

Tom: The theoretical foundation is pretty solid too. Theorem one shows that under their measurement model assumptions and the four-view construction, the standardized estimation error of that auxiliary coefficient follows a standard normal distribution conditionally for any non-zero term in the selected set <ref:2610.11486#pg1>.

Jane: That’s a strong statement about how well they've controlled their errors in this specific setup.

Lu: And Theorem two characterizes that bridge asymptotically, showing that as you increase the number of observations M, the relationship between that primary best linear approximation coefficient and the auxiliary IV coefficient converges to a normal distribution with mean zero and variance one/q where q is a positive constant <ref:2610.11486#pg1>.

Title and authors: Meng: That convergence to a positive constant for the variance is what gives them confidence in their long-term results.

Tom: Now let's talk about how this actually performs in practice. The empirical validation uses simulation studies, and they report rejection rates at the true selected best linear approximation coefficient that are between four point four one percent and five point three nine percent.

Jane: That's significantly better than what naive inference methods get, which were around seven point three six percent to eight point six nine percent, and much better than the Bonferroni correction method which was only about one point six one percent to two point zero two percent.

Lu: The paper notes that PSI-SINDy avoids a failure mode where naive reuse of the selection response tends to inflate rejection rates, which is a problem you have in other methods.

Meng: So they are showing superior control over Type I errors while still maintaining reasonable power against the true model selection.

Tom: They also ran a sensitivity analysis with Table twelve testing across various things like different dynamical systems—Van der Pol, Lotka–Volterra, Lorenz two, and Pendulum two—and different noise scales.

Jane: A big finding from that was that under correlated noise conditions, calibrating the auxiliary IV doesn't guarantee a finite-sample selected-BLA calibration.

Lu: That’s a important caveat. The study reports that under correlated noise, all twenty-five upper endpoints for scale inputs remain below the ninety-five percent confidence level.

Meng: And they found estimated-to-true scale ratios range from zero point five nine to zero point seven two under correlation, which is quite different from the near-unity ratios of around one under independent noise conditions.

Tom: So while it's a solid framework, you have to be careful about correlated noise when relying on that specific auxiliary calibration for the final best linear approximation coefficient.

Jane: It shows they've done a rigorous job separating these different statistical hurdles with their four-view construction in the PSI-SINDy paper.

Lu: The computational side is also addressed. They detail the pipeline which includes thinning, constructing the corrected selection design, term selection, and then if applicable, the inference routine itself.

Meng: The inference time scales with M and Bboot, which is the number of bootstrap iterations you use for assessment.

Tom: And they show that on an Apple M4 Pro processor, wall-clock runtime increased from zero point zero zero eight eight seconds to zero point two seven zero one seconds as complexity parameters went up in their simulations.

Jane: So, it’s definitely a computationally intensive process, especially with the rigorous four-view construction they described in PSI-SINDy.

Lu: What this means is that while the theory is very detailed and mathematically grounded, you have to be mindful of the practical computational cost when you try to apply it to real datasets.

Meng: It’s a tool that offers better control over errors than simpler methods, but you still need enough computing power to run that whole four-view thinning construction and the inference routine efficiently.

Title and authors: Tom: So we've seen how this paper, "PSI-SINDy: Post-Selection Inference for Sparse Identification of Nonlinear Dynamics," uses a data thinning approach to bridge an exact auxiliary target to an asymptotically valid primary best linear approximation coefficient.

Jane: It gives us much better control over Type I errors in the context of selecting terms from noisy trajectories.

Lu: The four views—V1 for selection, V2 for response construction, V3 for a corrected design, and V4 for an instrument—are the core mechanism that separates selection from inference.

Meng: It’s a sophisticated way to handle measurement error in the design matrix and shared noise simultaneously.

Tom: The paper shows that this method successfully improves selection reliability by providing finite-sample exact conditional inference for an auxiliary realized IV coefficient, and then showing an asymptotically valid inference for the primary noise-free target.

Jane: That’s a big improvement because it addresses that selection bias issue head-on in a statistically principled way.

Lu: They also introduce a bridge between these two targets to obtain those asymptotically valid results, even when the corrected dynamical terms still have random measurement-error fluctuations.

Meng: It’s clever how they handle that uncertainty propagation without needing the noise-free trajectory directly.

Tom: For anyone who only listens to this, the main point is that this framework gives us a pathway to quantify uncertainty around our selected nonlinear dynamics coefficients in a way that's much more robust than before.

Jane: It’s about moving past just picking terms and moving toward having statistically sound conclusions about those terms.

Lu: The limitations they state are pretty clear: the method doesn't guarantee finite-sample calibration under highly correlated noise conditions, which means you have to be cautious when your data has that kind of noise structure.

Meng: And they also show the computational cost increases with complexity parameters, so it’s not just a quick plug-and-play solution for every single scenario.

Tom: So we've covered the details of how PSI-SINDy uses four independent views and an asymptotic bridge to address selection bias and measurement error in SINDy inference.

Jane: It’s a detailed piece of work that provides strong statistical guarantees for applying SINDy to noisy time-series data.

Lu: The paper's name, "PSI-SINDy: Post-Selection Inference for Sparse Identification of Nonlinear Dynamics," really captures the essence of what they did.

Meng: It’s about moving from just finding terms to statistically validating those terms in a way that accounts for the shared noise and design errors.

Tom: Yeah, that's the big picture. So we’ve seen how this framework uses four independent views to build a path to valid inference on sparse identification of nonlinear dynamics.

Jane: We’re getting ready for another paper now, but this one really shows how deep the statistical tools can get when applied to complex data science problems.

The paper's summary: Tom: So, to recap, PSI-SINDy is this new statistical framework for SINDy that uses four different views of your data to figure out if the terms you picked are actually reliable.

Jane: Exactly. It tackles that big problem where you pick dynamical terms from noisy data and then try to test how good they are, which usually just introduces selection bias into your results.

Lu: The authors get around that by splitting the single trajectory into four independent views: V1 for term selection, V2 for the testing response, V3 for a corrected design matrix, and V4 for an instrument that acts like a noise-free version of the design.

Meng: That separation is what they call decoupling selection from inference. It lets you do different statistical jobs with different parts of the data so they don't interfere with each other’s errors.

Tom: And then they build this bridge between two targets: an auxiliary coefficient that has a very precise, exact conditional test, and the primary best linear approximation coefficient you actually want.

Jane: It's a smart way to get those asymptotically valid results for your main target even when the dynamical terms themselves still have some random measurement error in them.

Lu: They prove that this bridge works because they establish a finite-sample exact conditional test for that auxiliary IV coefficient, and then show it converges to an asymptotic normal distribution as you add more data.

Meng: That convergence to a positive constant for the variance, one over q, is what gives them confidence in the long run when you use a large number of observations.

Tom: Empirically, their rejection rates at the true selected BLA are really good, sitting between four point four one percent and five point three nine percent.

Jane: That's much better than naive methods, which were doing seven point three six to eight point six nine percent, and it keeps your type one error rates in line with what you expect.

Lu: The paper also mentions that their method successfully avoids that common problem where just reusing the selection response ends up inflating those rejection rates.

Meng: But they do have a caveat: if your noise is actually correlated, they find that the auxiliary calibration doesn't guarantee a finite-sample selected BLA calibration.

Tom: So it's a powerful tool for controlling errors, but you gotta watch out for that correlation issue when your noise isn't independent.

Jane: It shows that if you’re working with noisy time-series data, there is a rigorous way to move past just picking terms and start having statistically sound conclusions about the dynamics.

Lu: This opens up new avenues for how we can use SINDy more reliably in real-world applications where the data is inherently messy.

The paper's improvements: Tom: So we've seen how PSI-SINDy uses four views to bridge an exact auxiliary target to an asymptotically valid primary best linear approximation coefficient, and now we're looking at what they suggest for next steps.

Jane: They suggest a few ways to make this framework even more robust, like using a scale factor calculation for the BLA coefficient.

Lu: They propose calculating the combined standard error as sbj-S equals q e squared j S plus bt squared j S, where that gap scale estimate bt j S is found using a parametric bootstrap around a pilot trajectory.

Meng: That means instead of just one error term, you're combining two different ways to estimate uncertainty into one number for your final BLA calculation.

Tom: It’s trying to give you a more comprehensive error picture, especially when dealing with the noise structure we talked about earlier.

Jane: They also suggest using support-conditioned calibration, which means testing each selected coordinate at its own true BLA and seeing if any null hypothesis gets rejected in each dataset.

Lu: That allows for a kind of simultaneous Bonferroni coverage, where you get a union bound over all the IV intervals instead of just looking at one result.

Meng: That sounds like it would help you assess model performance across different dynamical systems more systematically, which is exactly what we need for practical engineering applications.

Tom: It moves the focus from just getting one number to getting a set of statistically sound conclusions for all your selected terms at once.

Jane: The limitation they point out is that this exact calibration under correlated noise conditions might not hold up in finite samples, so you still have to be careful with that noise structure.

Lu: That’s a fair point, and it means the next big challenge for this framework is figuring out how to handle those highly correlated noise scenarios more reliably in a practical setting.

Conclusion: Tom: So we've covered how PSI-SINDy uses four independent views to bridge an exact auxiliary target to an asymptotically valid primary best linear approximation coefficient, and now we’re wrapping up with the conclusions on this paper titled "PSI-SINDy: Post-Selection Inference for Sparse Identification of Nonlinear Dynamics."

Jane: It really shows how you can tackle the messy statistical hurdles in SINDy when you're dealing with real noisy data.

Lu: The implication is that we can build much more reliable methods for quantifying uncertainty around the dynamical terms we identify, which is crucial because those terms are what drive our AI models.

Meng: For me, it means that when I’m building these complex systems on my side, having a statistically sound way to assess the reliability of my identified dynamics helps me trust the predictions much more.

Lalam: From an AI culture standpoint, this kind of rigorous statistical grounding moves us away from just generating patterns and toward building systems that can actually be trusted in sensitive applications.

Tom: And they did confirm that while correlated noise is a bit tricky for finite sample calibration, the framework itself gives us superior control over those type one errors compared to older methods.

Jane: So for someone listening who isn't deep in the math, what this changes is that we have a more principled way to decide if a term we picked actually matters or if it was just noise.

Lu: It shifts the focus from just finding terms to statistically validating those terms in a way that accounts for the shared noise and design errors you talked about earlier.

Meng: It’s about moving past just picking terms and moving toward having statistically sound conclusions about those terms when we apply SINDy to noisy time-series data.

Tom: Exactly, so this framework gives us a pathway to quantify uncertainty around our selected nonlinear dynamics coefficients in a way that's much more robust than before.

Jane: It’s about moving past just picking terms and moving toward having statistically sound conclusions about those terms.

Lu: This opens up new avenues for how we can use SINDy more reliably in real-world applications where the data is inherently messy.

RIKEN 2Nagoya University

stat.ML, cs.LG

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: 47 pages, 3 figures, 16 tables

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: Detailed Research Summary: PSI-SINDy - Post-Selection Inference for Sparse Identification of Nonlinear Dynamics Overview: PSI-SINDy is a novel post-selection inference framework specifically designed

Key concepts

SINDy: Sparse Identification of Nonlinear Dynamics is a technique that finds a small set of mathematical equations (dynamical terms) that best describe complex time-series data. It works by testing many possible simple functions against the data and picking only the most relevant ones.
Post-Selection Inference: This refers to statistical methods used *after* you have already selected a subset of model terms. The challenge is that standard tests often fail here because the selection process itself introduces bias, making it hard to know if your final results are truly reliable.
Four-View Data Thinning: This is the core technique where a single noisy data trajectory is split into four separate, independent views. Each view has a specific job: one for initial term selection, one for building the test response, one to fix measurement errors in the design terms, and one to create a noise-free estimate of those terms.

Terminology

Calibration — 44× in this paper · explained in 7 episodes
Monte Carlo — 18× in this paper
Consistency — 13× in this paper · explained in 3 episodes
Post-Selection — 12× in this paper · explained in 2 episodes
Residualization — 9× in this paper · explained in 2 episodes
Ablation — 8× in this paper · explained in 1 episode
Accuracy — 6× in this paper · explained in 1 episode
Benchmark — 6× in this paper · explained in 2 episodes
Selection Bias — 5× in this paper · explained in 4 episodes
Trace — 4× in this paper · explained in 2 episodes
Clustering — 2× in this paper · explained in 2 episodes
Generative AI — 2× in this paper · explained in 2 episodes
Robustness — 2× in this paper · explained in 5 episodes
Theoretical Guarantees — 2× in this paper · explained in 2 episodes
Transformer — 2× in this paper · explained in 2 episodes
Active Learning — 1× in this paper · explained in 6 episodes
Adaptation — 1× in this paper · explained in 1 episode
Aggregation — 1× in this paper · explained in 1 episode
Bayesian Inference — 1× in this paper · explained in 5 episodes
Calibration Failure — 1× in this paper · explained in 2 episodes
Conformal Prediction — 1× in this paper · explained in 5 episodes
Core — 1× in this paper · explained in 2 episodes
Empirical Coverage — 1× in this paper · explained in 2 episodes
Feature Engineering — 1× in this paper · explained in 3 episodes
Feature Selection — 1× in this paper · explained in 4 episodes
Regularization — 1× in this paper · explained in 3 episodes
Uncertainty Quantification — 1× in this paper · explained in 8 episodes
Verification — 1× in this paper · explained in 2 episodes

Summary

Detailed Research Summary: PSI-SINDy - Post-Selection Inference for Sparse Identification of Nonlinear Dynamics

Overview:

PSI-SINDy is a novel post-selection inference framework specifically designed to quantify the statistical reliability of dynamical terms identified by Sparse Identification of Nonlinear Dynamics (SINDy) when applied to noisy trajectory data. The core challenge addressed by this framework stems from the inherent difficulties in applying standard post-selection inference techniques to SINDy outputs, primarily due to data-dependent model selection, measurement error introduced during term selection, and shared noise between the response and design matrices.

Core Methodology: The Four-View Data Thinning Construction:

To overcome these challenges—namely selection bias, measurement error in the design matrix, and shared noise—PSI-SINDy employs a sophisticated data thinning strategy (referencing Neufeld et al., 2024). This construction decomposes a single observed trajectory into four mutually independent views, each serving a distinct role:

View V1: Dedicated to the initial term selection process.
View V2: Used for constructing the testing response.
View V3: Utilized for constructing a corrected design matrix, mitigating measurement error in the original design terms (addressing the EIV problem).
View V4: Used to construct an independent instrument—a noise-free estimate of the design matrix.

This four-view decomposition is crucial as it effectively separates the processes of selection from inference, and isolates the testing response from both the noisy design and shared noise effects.

Inference Targets and Bridging Mechanism:

The framework establishes a rigorous inferential pipeline to provide statistically sound conclusions:

Primary Target: The Best Linear Approximation (BLA) coefficient derived from the noise-free weak regression restricted to the selected set of terms (S).
Auxiliary Target: A realized instrumental-variable (IV) coefficient for which exact conditional tests and confidence intervals can be established under specific assumptions (Section 3).

The paper establishes a critical bridge between these two targets: it first derives finite-sample exact conditional inference for the auxiliary IV coefficient using the four-view construction. Subsequently, it develops an asymptotic bridge connecting this exact inference for the auxiliary target to asymptotically valid inference for the primary BLA coefficient. This ensures that while exact results are available for a related quantity, they converge to reliable asymptotic results for the ultimate goal (the selected BLA).

Theoretical Foundations and Key Results:

The framework is supported by rigorous mathematical theorems:

Theorem 1 (Exact conditional IV inference): Under specified measurement model assumptions and the four-view construction, it is shown that for any non-zero term j in S, the standardized estimation error of c j,S - c j,S follows a standard normal distribution N(0, 1) conditionally.
Theorem 2 (Selected-model IV–BLA bridge): Under the assumption of a fixed-dimensional weak-design limit (Assumption 1), the relationship between the primary BLA coefficient and the auxiliary IV coefficient is characterized asymptotically. Specifically, sqrt M (c j,S - xi j,S) so N(0, tau 2 j,S), where tau 2 j,S converges to a positive constant 1/q > 0.

Empirical Validation and Performance:

The framework's performance is validated through extensive simulation studies:

Rejection Rates: The primary PSI-SINDy rejection rates at the true selected BLA are reported in Figure 1, showing significantly better performance than naive methods. Specifically, PSI-SINDy achieves rejection rates between 4.41% and 5.39%, compared to 7.36%–8.69% for Naive inference and 1.61%–2.02% for the Bonferroni correction method, indicating superior control over Type I errors while maintaining reasonable power against the true model selection.
Coverage Assessment: Simulations confirm that PSI-SINDy's Type I error rates and confidence interval coverages are compatible with nominal levels. Crucially, the paper notes that naive reuse of the selection response tends to inflate rejection rates, a failure mode PSI-SINDy successfully avoids.

Sensitivity Analysis and Robustness Testing:

The study incorporates a comprehensive sensitivity analysis using Table 12, which examines performance across numerous tuning parameters:

Model Diversity: Assessment is performed on various dynamical systems, including Van der Pol (VdP), Lotka–Volterra (LV), Lorenz (2), and Pendulum (2).
Parameter Variation: The analysis explores a wide range of settings, including sample sizes (M=1600), noise scales (sigma X = 0.02, 0.06, 0.12), threshold levels (e.g., 13, 400, and 5 for the threshold), number of components (K=8 to 24), radius settings, and various misspecification scenarios (sine drift vs. exponential drift).
Correlated Noise Impact: A significant finding is that under correlated noise conditions, the calibration of the auxiliary IV does not guarantee finite-sample selected-BLA calibration. The study reports that under correlated noise, all 25 upper endpoints for scale inputs remain below the 95% confidence level. Furthermore, estimated-to-true scale ratios range from 0.59 to 0.72 under correlation, contrasting sharply with near-unity ratios (0.999–1.002) achieved under independence.

Computational Considerations:

The paper details the computational pipeline, noting that the total pipeline time includes thinning, construction of the corrected selection design, term selection, and (if applicable) the inference routine. The inference time itself is a significant component of this pipeline and scales with M and Bboot (number of bootstrap iterations). The wall-clock runtime for primary simulations was reported on an Apple M4 Pro processor, showing an increase in inference time from 0.0088 to 0.2701 seconds as complexity parameters increased, highlighting the computational cost associated with the rigorous four-view construction and inference routine.

Conclusion:

PSI-SINDy is a highly detailed and mathematically grounded framework that successfully addresses the complex statistical hurdles of performing reliable post-selection inference on SINDy outputs from noisy data. By ingeniously decomposing the data into four independent views, it provides a pathway to establish exact conditional tests for an auxiliary variable, which then serves as a robust asymptotic bridge to infer the primary BLA coefficient. The empirical results demonstrate that PSI-SINDy offers superior control over Type I errors and coverage compared to naive methods across diverse dynamical systems and noise conditions, although sensitivity analysis reveals limitations regarding finite-sample calibration under highly correlated noise.

Improvements for AI systems

Improve selection reliability using PSI-SINDy to quantify statistical significance, as it addresses the key difficulty is that using the same noisy trajectory for both selecting dynamical terms and assessing their statistical significance can introduce selection bias. The system can now provide finite-sample exact conditional inference for the auxiliary realized IV coefficient and asymptotically valid inference for the primary noise-free target.
Implement a four-view data thinning construction to separate information streams, enabling independent inference tasks. This construction allows one to use V1 for term selection, V2 for constructing the testing response, V3 for constructing a corrected design, and V4 for constructing an instrument, thereby accounting for selection bias, shared response–design noise, and measurement error in the design.
Develop a bridge between two inferential targets to obtain asymptotically valid results. This involves establishing an asymptotic bridge from this auxiliary target to the first, primary selected-BLA target, which provides a method for quantifying uncertainty even when the corrected dynamical terms still contain random measurement-error fluctuations.
Utilize the derived scale factor for robust BLA inference. The combined standard error is calculated as sbj,S = qe2j,S + bt2j,S, where the gap scale estimate btj,S is estimated using a "parametric bootstrap (Davison & Hinkley, 1997) around a pilot trajectory."
Employ support-conditioned calibration to assess performance across different models. The system can perform this by testing each selected coordinate at its own true BLA and recording whether any null hypothesis is rejected in each dataset, allowing for simultaneous Bonferroni coverage which provides a union bound over available IV intervals.