Learning a model of non-stationary nuisances with non-local correlations for object detection

arXiv:2611.14479 · astro-ph.IM · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "Learning a model of non-stationary nuisances with non-local correlations for object detection".

Jocelyn: Here's the gist of the paper. Detecting faint objects in multivariate data corrupted by a strong and spatially correlated nuisance component in the low-sample regime is tackled by proposing a data-driven approach based on sparse precision matrix modeling.

Vera: This method yields a low-complexity yet expressive statistical model of the nuisance and enables stable estimation in the low-sample regime while capturing longrange (spatial) correlations.

Jocelyn: The approach is evaluated on an astrophysical application: the detection of exoplanets by highcontrast imaging, where it achieves state-of-the-art performance.

Vera: Model Formulation and Estimation The problem considers observations where the n-th data sample yn follows the forward model yn = fn(x) + sn, with sn being a nuisance component that is mutually independent but follows an identical Gaussian distribution with mean µ def and precision matrix Λ.

Title and authors: Vera: So we're looking at this paper by Flasseur et al., titled "Learning a model of non-stationary nuisances with non-local correlations for object detection". Basically, they're tackling the problem of finding faint objects in data that gets messed up by something that isn't just random noise, but something spatially correlated and strong.

Jocelyn: That sounds like a tough problem, Vera. The title makes it sound like they are learning how this nuisance component changes over time or space—non-stationary—and it has these long-range connections. It’s about finding faint things in noisy data where the noise itself is structured in a complex way.

Subrahmanyan: And the core idea seems to be that standard methods for handling this kind of noise often get unstable, especially when you only have a small number of data points available, which is what they call the low-sample regime.

Vera: Exactly. The authors propose using a data-driven approach based on sparse precision matrix modeling to get away from those unstable covariance models and still capture those long-range spatial correlations in the noise component.

Jocelyn: They use something like a pivoted sparse Cholesky factorization, where the non-zero coefficients of that factorization are learned using maximum likelihood from the data itself. That lets them build a model of the nuisance that is both low-complexity and expressive.

Subrahmanyan: I’m interested in how they handle those finite sample sizes. The paper points out that getting a consistent estimate for the covariance matrix requires having more data points than there are parameters, which is not always true in these kinds of surveys.

Vera: To deal with that, they approximate the precision matrix by using a structure S defined by some ordering and then selecting only the nearest preceding nodes based on a distance metric. They define model complexity M based on how many non-zero entries are in this structure S.

Jocelyn: And they use cross-validation to pick the right complexity M, maximizing a leave-one-out cross-validated predictive log-likelihood. That’s smart because it tries to find the simplest model that still predicts well without just overfitting to the noise in our small sample set.

Title and authors: Subrahmanyan: So, when you look at their detection criterion, they use a generalized likelihood ratio test that looks like a standard normal distribution under the null hypothesis for the nuisance statistics. This gives them a natural way to set a threshold based on the signal-to-noise ratio.

Vera: Right. It turns the detection problem into something more interpretable, directly related to how much signal you expect versus how much noise you are dealing with, which is what we always want when we're looking at faint objects.

Jocelyn: The paper shows that by using this test and setting a threshold τ greater than zero, you get a constant false alarm rate detector. That means the detection criterion has a clear physical meaning tied to the signal-to-noise ratio of what you are actually observing.

Subrahmanyan: From a cosmological perspective, if we can reliably detect faint sources in these challenging environments, it helps us constrain models of structure formation where the noise might otherwise hide those signals. It connects back to how well we can measure things at large scales versus local details.

Vera: And the method itself is designed to work even when you have non-stationary nuisances, meaning they aren't just a fixed Gaussian component but something that evolves across the data set or space, which is a big deal for real sky observations.

Jocelyn: It’s really about building this statistical framework so that we can trust our measurements more reliably when the data quality is complicated by these kinds of structured uncertainties.

Subrahmanyan: So, to wrap up on the practical side, they show how this method achieves state-of-the-art performance in an astrophysical application—specifically, detecting exoplanets using high-contrast imaging. That’s a very concrete result for the community.

Vera: It is a very strong result because it shows that this data-driven modeling actually outperforms traditional methods in situations where those nuisance components are complex and correlated.

Jocelyn: It suggests that learning the structure of the noise itself can be as powerful as trying to perfectly model every single physical process separately.

Title and authors: Subrahmanyan: And for future work, they leave open questions about how this framework scales when you move beyond a simple Gaussian nuisance model or how it handles truly extreme non-stationarity across vastly different observational scales.

Vera: It’s definitely an interesting direction to explore, moving from the current Gaussian assumption toward more general models of noise structure.

Jocelyn: Exactly. We need to see if this sparse modeling approach can adapt when the underlying physics is even less predictable than what they assumed in this work.

Subrahmanyan: So, it sounds like the main implication here is that we have a tool to stabilize estimation and get better object detection performance when the data is corrupted by correlated noise, even in low-sample regimes.

Vera: It gives us a concrete way to estimate those nuisance parameters without relying on assumptions about their exact distribution beyond them being Gaussian.

Jocelyn: And the application in exoplanet imaging shows that this isn't just theoretical; it has real impact on how we search for faint worlds orbiting other stars.

Subrahmanyan: It’s a nice bridge between complex statistical modeling and the actual observational challenges faced by instruments like those used in high-contrast imaging.

Vera: So, to summarize, this paper proposes learning a sparse model of non-stationary nuisances using maximum likelihood, which leads to a generalized likelihood ratio test that controls the false alarm rate based on signal-to-noise ratios.

Jocelyn: And they use cross-validated predictive log-likelihood to select the optimal complexity of this sparse model, ensuring it captures longrange correlations without overfitting.

Subrahmanyan: The big picture is that we can now build more robust tools for detecting faint signals in messy data, which should eventually help us better constrain cosmological models when those same kinds of noise structures appear in our observations.

Vera: It’s a solid piece of work because it addresses the stability issue that usually plagues covariance-based methods and gives us a principled way to handle those long-range dependencies.

Jocelyn: We'll keep an eye on how they evolve this sparse modeling technique, especially as we look at more complex astrophysical datasets with even more intricate noise.

The paper's summary: Vera: So, to recap, this paper is about building a data-driven way to model noise that isn't just random, but something spatially correlated and changing over time, especially when you don't have a lot of data.

Jocelyn: Right. It takes the problem of finding faint things in messy sky surveys and tries to solve it by learning the structure of that mess itself using sparse models.

Subrahmanyan: The core idea is that instead of guessing a standard noise model, they use something called a sparse precision matrix model to capture those long-range spatial connections in the noise component.

Vera: That's right, and they do this by approximating the precision matrix with this structure S, which basically describes how nodes are connected spatially.

Jocelyn: They then use maximum likelihood estimation to learn that structure S from the data itself, making it a way to estimate those unknown nuisance parameters without relying on standard covariance assumptions that often break down in low-sample regimes.

Subrahmanyan: And they select the complexity of this model M by maximizing a leave-one-out cross-validated predictive log likelihood, which is smart because it tries to find the right balance between capturing those spatial correlations and not just memorizing the noise in a small sample.

Vera: It’s an unsupervised way to pick how complex the model should be, ensuring it actually learns something useful about the underlying physics rather than just fitting the noise perfectly.

Jocelyn: The detection part uses a generalized likelihood ratio test that gives you a signal-to-noise ratio, which is a very natural way to set your threshold for finding those faint objects.

Subrahmanyan: This moves us from needing perfect theoretical knowledge of the noise to using statistical learning to find the best model structure from limited data.

Vera: It means we can get more stable estimates of faint sources even when the data quality is really poor because we aren't fighting a singularity in our math.

Jocelyn: If this works well, it could significantly improve how we search for exoplanets or other faint astrophysical objects in high-contrast imaging surveys.

Subrahmanyan: That’s the immediate payoff, but it also opens up the possibility of applying this kind of sparse modeling to other complex systems where correlated noise is a major hurdle.

The paper's improvements: Tom: So, to recap that paper, it’s about using sparse precision matrix modeling to handle noise that isn't stationary or local, especially when we only have a small amount of data available.

Vera: Right. The authors are proposing a way to learn the structure of this changing noise component using maximum likelihood from the observations themselves.

Jocelyn: They show that by defining a structure S for this matrix, based on how nodes connect spatially, we can get a low-complexity model that still captures those long-range spatial correlations.

Subrahmanyan: The key improvement they highlight is how they select the complexity of this structure M using leave-one-out cross-validated predictive log likelihood.

Vera: That means instead of us guessing how complex the noise structure should be, we let the data tell us by looking at how well a model trained on most of the data predicts the one sample we left out.

Jocelyn: So it’s an unsupervised method for complexity selection, which is really useful because you don't want to overfit when you only have a few hundred data points.

Subrahmanyan: And this whole approach gives us a detection criterion that directly relates to the signal-to-noise ratio, which simplifies setting that threshold for finding faint objects.

Vera: It means we get more stable estimates of what’s out there because we're not relying on assumptions about the noise distribution being perfectly simple.

Jocelyn: If this method holds up, it suggests that learning the structure of correlated noise can be a very powerful way to stabilize measurements in noisy environments.

Subrahmanyan: It pushes us toward building more robust tools for cosmology where those kinds of complex, structured uncertainties are common when looking at large-scale structure formation.

Conclusion: Tom: So we're wrapping up this look at "Learning a model of non-stationary nuisances with non-local correlations for object detection". It boils down to using sparse precision matrix modeling to tackle complex, correlated noise in low-sample data.

Vera: Yeah, it’s about building a statistically stable framework for finding faint objects when the noise itself is changing across space and time.

Jocelyn: Exactly. They show how learning that structure S using maximum likelihood lets us estimate those nuisance parameters without getting stuck in the kind of instability we see with standard covariance models.

Subrahmanyan: From a theoretical standpoint, this is important because it means we can better understand how faint signals get obscured by large-scale noise structures in the universe.

Vera: It translates to a more reliable error bar on those detections, which is crucial for any real survey data we collect across the sky.

Jocelyn: And that stability extends to our time-domain work; if you can model these nuisances well, your follow-up observations have a much better chance of revealing the actual signal.

Subrahmanyan: The next question is whether this method scales up when we move from a simple Gaussian noise component to more complicated physical noise processes.

Vera: That’s what they leave open—how general can this sparse modeling framework be before it starts needing a totally different approach?

Jocelyn: It’s definitely a path forward for making our astrophysical measurements more robust in the face of real-world data imperfections.

Olivier Flasseur, Eric Thiébaut, Loğ Denis, Samuel Theé

Universite Claude Bernard Lyon 1 · CRAL UMR 5574 · ENS de Lyon · Universite Jean Monnet Saint-Etienne · CNRS/INSU · Haystack Observatory, Massachusetts Institute of Technology

astro-ph.IM

Submitted: 2026-11-10

Updated: 2026-11-10

Journal ref: EUSIPCO 2026

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: Detecting faint objects in multivariate data corrupted by a strong and spatially correlated nuisance component in the low-sample regime is tackled by proposing a data-driven approach based on sparse…

Key concepts

Sparse Precision Matrix Modeling
This technique uses a simplified, sparse representation (like a Cholesky factorization) of the noise's precision matrix. This makes the statistical model computationally efficient and less prone to failure when data is scarce, allowing for stable estimation even when standard covariance models become unstable.
Low-Sample Regime Estimation
When there are very few data points, traditional methods for estimating noise parameters often fail because the covariance matrix becomes singular or ill-conditioned. This approach overcomes that by using a learned sparse structure to approximate the complex noise behavior, ensuring reliable estimates in low-sample scenarios.
Leave-One-Out Cross-Validation (LOOCV)
This is a method used to automatically select the best complexity for the statistical model. The system trains on all but one data point, predicts that point's likelihood using the learned model, and then repeats this process for every single data point. The model complexity that maximizes this predictive accuracy is chosen as optimal.
Generalized Likelihood Ratio Test (GLRT)
This is a statistical test used to determine if a signal exists. By comparing the likelihood of data under a 'signal present' hypothesis versus a 'noise only' hypothesis, it provides a clear measure of the signal-to-noise ratio (SNR). This allows researchers to set thresholds for detection that directly relate to how strong the true object's presence is.

Summary

Detecting faint objects in multivariate data corrupted by a strong and spatially correlated nuisance component in the low-sample regime is tackled by proposing a data-driven approach based on sparse precision matrix modeling. p. 0

Where the paper says it

1Universite Claude Bernard Lyon 1, CRAL UMR 5574, ENS de Lyon, CNRS, Villeurbanne, F-69622, France 2Universite Jean Monnet Saint-Etienne, CNRS, IOGS, Laboratoire Hubert Curien UMR 5516, Saint-Etienne, F-42023, France ´ 3Haystack Observatory, Massachusetts Institute of Technology, Westford, MA Abstract—We tackle the problem of detecting faint objects in multivariate data corrupted by a strong and spatially correlated nuisance component in the low-sample regime. Standard covariance-based models often yield unstable or ill-conditioned estimators of the nuisance statistics, severely degrading detection performance.

This method yields a low-complexity yet expressive statistical model of the nuisance and enables stable estimation in the low-sample regime while capturing longrange (spatial) correlations. p. 0
Where the paper says it

We propose a low-complexity, datadriven detection framework based on sparse precision matrix modeling of the nuisance component. The precision matrix is approximated by a pivoted sparse Cholesky factorization learned from the data by maximum likelihood, enabling stable estimation in the low-sample regime while capturing longrange (spatial) correlations (Sect.

The approach is evaluated on an astrophysical application: the detection of exoplanets by highcontrast imaging, where it achieves state-of-the-art performance. p. 0
Where the paper says it

Building on this model, we derive a generalized likelihood ratio test for faint source detection enabling direct control of the false alarm rate. The approach is evaluated on an astrophysical application: the detection of exoplanets by highcontrast imaging, where it achieves state-of-the-art performance.

Model Formulation and Estimation

The problem considers observations where the n-th data sample yn follows the forward model yn = fn(x) + sn, with sn being a nuisance component that is mutually independent but follows an identical Gaussian distribution with mean µ def and precision matrix Λ. The co-log-likelihood of the data sequence given x and omega writes L(x, omega) = 1/2 X N n=1 yn − fn(x) − µ def / Λ−N log(Λ) up to an irrelevant additive constant. p. 1

Where the paper says it

Under forward model (1), the probability density function of the n-th nuisance term is given by: pdf(sn | Ω) = Λ 2 π 1/2 exp − 1 2 yn − fn(x) − µ 2 Λ , where |Λ| is the determinant of Λ, and ∥v∥ 2 Λ def = v ⊤ · Λ · v is the squared Mahalanobis norm. Under these assumptions, the co-log-likelihood of the data sequence {yn}n∈1:N given x and Ω writes: L(x, Ω) = 1 2 X N n=1 yn − fn(x) − µ 2 Λ − N 2 log(|Λ|) (3) up to an irrelevant additive constant.

Under the models of Sects. p. 1
Where the paper says it

II-CIII how to accurately estimate all the unknown parameters, x and Ω, from the finite set of N data samples. Maximum likelihood nuisance statistics Under the models of Sects.

II-A-II-B, the maximum likelihood estimators of the statistics omega knowing x reduce to the sample estimators µb(x) = argmin L(x, omega) = 1/N X N n=1 (yn − fn(x)) and Λb(x) = argmin L(x, omega) = 1/N X N n=1 rn(x) rn(x)⊤ Cb−1 (5), page2>. p. 1
Where the paper says it

Maximum likelihood nuisance statistics Under the models of Sects. II-A-II-B, the maximum likelihood estimators of the statistics Ω knowing x reduce to the sample estimators: µb(x) = argmin µ L(x, Ω) = 1 N XN n=1 (yn − fn(x)), (4) Λb(x) = argmin Λ L(x, Ω) = 1 N X N n=1 rn(x) rn(x) ⊤ | {z } = Cb(x) −1 (5) with the residuals {rn(x) = yn − fn(x) − µ}n∈1:N .

Since the rank1 of Cb is at most N, a consistent covariance estimate requires N >> P. p. 1
Where the paper says it

II-A-II-B, the maximum likelihood estimators of the statistics Ω knowing x reduce to the sample estimators: µb(x) = argmin µ L(x, Ω) = 1 N XN n=1 (yn − fn(x)), (4) Λb(x) = argmin Λ L(x, Ω) = 1 N X N n=1 rn(x) rn(x) ⊤ | {z } = Cb(x) −1 (5) with the residuals {rn(x) = yn − fn(x) − µ}n∈1:N . Since the rank1 of Cb is at most N, a consistent covariance estimate requires N ≫ P.

To cope with this issue, the approach approximates Λ by Λ ≈ S⊤ · S = P⊤ · R⊤ · R · P where S is the composition of a sparse lower triangular matrix R and of a permutation matrix P, page3. p. 1
Where the paper says it

To cope with this issue, we propose to use a matrix with a lower complexity such as the matrices of A family of Sparse Approximations for whitening and Preconditioning (ASAP, [7]) as an estimator of the precision matrix. We thus approximate Λ by: Λ ≈ S ⊤ · S = P ⊤ · R⊤ · R · P , (6) where S is the composition of a sparse lower triangular matrix R and of a permutation matrix P, with P⊤ = P−1 .

The structure S is defined by (i) the ordering of the nodes and (ii), for each node, the set of connected preceding nodes Si = 1:i j ∈ 1:i(i, j) ∈ S where Si denotes the set of nodes connected to node i. p. 1
Where the paper says it

Construction of the sparse structure: The structure of S is defined by (i) the ordering of the nodes (e. , the pixels) and (ii), for each node, the set of connected preceding nodes.

The model complexity M is defined as the mean number of structural non-zeros per row of S and thus set the sparsity level of the model statistics. p. 1
Where the paper says it

ordering is specified by the permutation matrix P, while the connectivity is defined by: Si = {j ∈ 1:i|(i, j) ∈ S} , (∀i ∈ 1:P), (7) where Si denotes the set of nodes connected to node i, and S is the sparsity pattern of R, that is the set of index pairs (i, j) of the structural non-zeros of R. The model complexity M is defined as the mean number of structural non-zeros per row of S and thus set the sparsity level of the model statistics.

A first solution is to assume P = IP (i. p. 1
Where the paper says it

There is a variety of choices to define an ordered sparse structure S that differ by the sampling considered. A first solution is to assume P = IP (i.

e., no permutation) and to select, in lexicographic order, for each node i, the previous nodes that are closest to the node i: S lex i = 1:j ∈ 1:i j ≤ i, D(i, j) ≤ ρ where D(i, j) is a distance (e. p. 1
Where the paper says it

A first solution is to assume P = IP (i. , no permutation) and to select, in lexicographic order, for each node i, the previous nodes that are closest to the node i: S lex i = {j ∈ 1:i| j ≤ i, D(i, j) ≤ ρ} , (8) where D(i, j) is a distance (e.

g., Euclidean) between nodes i and j. p. 1
Where the paper says it

, no permutation) and to select, in lexicographic order, for each node i, the previous nodes that are closest to the node i: S lex i = {j ∈ 1:i| j ≤ i, D(i, j) ≤ ρ} , (8) where D(i, j) is a distance (e. , Euclidean) between nodes i and j, and ρ ≥ 0 controls the sparsity that is the number Mi = |Si | of nodes connected to the i-th one.

Detection Criterion and Model Selection

Assuming that the nuisance statistics omega are known, detection is performed by the generalized likelihood ratio test (GLRT): Γ(θ, omega) = b(θ, omega) / p a(θ, omega) = αb(θ, omega) σb(θ, omega) H1≷H0 τ. Under the null hypothesis H0 and the assumed multivariate Gaussian nuisance model, Γ(θ, omega) follows a standard normal distribution N (0, 1), independently of θ. p. 2

Where the paper says it

Detection criterion Assuming that the nuisance statistics Ω are known, detection is performed by the generalized likelihood ratio test (GLRT): Γ(θ, Ω) = b(θ, Ω) p a(θ, Ω) = αb(θ, Ω) σb(θ, Ω) H1 ≷ H0 τ . (16) Under the null hypothesis H0 and the assumed multivariate Gaussian nuisance model, Γ(θ, Ω) follows a standard normal distribution N (0, 1), independently of θ.

A single threshold τ > 0 thus ensures a constant false alarm rate (CFAR) detector and yields a natural interpretation of the detection criterion as a signal-to-noise ratio (SNR). p. 2
Where the paper says it

(16) Under the null hypothesis H0 and the assumed multivariate Gaussian nuisance model, Γ(θ, Ω) follows a standard normal distribution N (0, 1), independently of θ. A single threshold τ > 0 thus ensures a constant false alarm rate (CFAR) detector and yields a natural interpretation of the detection criterion as a signal-to-noise ratio (SNR).

The model complexity M is selected by maximizing a leaveone-out cross-validated predictive log-likelihood (LOOCV), defined as the sum of the co-log-likelihoods of each sample yn evaluated under a model learned from all remaining samples <ref:2611. p. 2
Where the paper says it

We thus select M by maximizing a leaveone-out cross-validated predictive log-likelihood (LOOCV), defined as the sum of the co-log-likelihoods of each sample yn evaluated under a model learned from all remaining samples. C(M) = 1 2 X N n=1 yn − µb [n] 2 Λb[n] − log(|Λb[n] |+), (17) where Ωb[n] = {µb [n] , Λb[n]} are the nuisance statistics learned with model complexity M from the training set χ [n] = {yn′ | n ′ ∈ 1:N, n′ ̸= n}, i.

14479B,C(M) = 1/2 X N n=1 yn − µb[n] 2 Λb[n] − log(Λb[n]), page4>. p. 2
Where the paper says it

We thus select M by maximizing a leaveone-out cross-validated predictive log-likelihood (LOOCV), defined as the sum of the co-log-likelihoods of each sample yn evaluated under a model learned from all remaining samples. C(M) = 1 2 X N n=1 yn − µb [n] 2 Λb[n] − log(|Λb[n] |+), (17) where Ωb[n] = {µb [n] , Λb[n]} are the nuisance statistics learned with model complexity M from the training set χ [n] = {yn′ | n ′ ∈ 1:N, n′ ̸= n}, i.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Data-driven nuisance modeling for low-sample regimes: The system can estimate unknown parameters x and nuisance statistics Omega using a sparse Cholesky factorization learned by maximum likelihood, enabling stable estimation where standard covariance models fail due to singularity.

  2. Generalized likelihood ratio test for detection: The improved system can derive a generalized likelihood ratio test that allows direct control of the false alarm rate, providing a natural interpretation of detection as a signal-to-noise ratio (SNR).

  3. Unsupervised model complexity selection: The system can automatically select the optimal model complexity by maximizing a leaveone-out cross-validated predictive log-likelihood, ensuring the model captures longrange correlations without overfitting to finite sample sets.

  4. Capturing complex spatial structures: By utilizing random or multi-scale sparse structures for the precision matrix factor S, the system gains the ability to reproduce large scale structures observed in real data, capturing both local and long-range correlations.

Abstract

We tackle the problem of detecting faint objects in multivariate data corrupted by a strong and spatially correlated nuisance component in the low-sample regime. Standard covariance-based models often yield unstable or ill-conditioned estimators of the nuisance statistics, severely degrading detection performance. We propose a data-driven approach based on sparse precision matrix modeling, where the precision matrix is approximated through a pivoted sparse Cholesky factorization whose non-zeros coefficients are learned by maximum likelihood. This yields a low-complexity yet expressive statistical model of the nuisance. Building on this model, we derive a generalized likelihood ratio test for faint source detection enabling direct control of the false alarm rate. The approach is evaluated on an astrophysical application: the detection of exoplanets by high contrast imaging, where it achieves state-of-the-art performance.

Related papers