Source-linked AI summary
Spectral Uncertainty Decomposition for Sparse Bayesian Learning
Ismail Huseynov, Stefan Haufe
TL;DR
M/EEG source estimation is ill-posed, and usual posterior source covariance does not explicitly distinguish uncertainty in measurement-visible directions from uncertainty in the nullspace. The paper derives complementary spectral and prior-minus-reduction representations of posterior covariance, separating these contributions and accounting for variance removed by data.
Problem
Usual posterior source covariance does not explicitly distinguish likelihood-visible from likelihood-null uncertainty, leaving principled source uncertainty quantification unclear.
Source: §1 · p. 1, §2.1 · p. 3Method
Results
Takeaways & Limitations
The additive form supports visualization of reducible and irreducible uncertainty, while the subtractive form supports accounting for data-driven uncertainty reduction.
Source: §7 · p. 6Empirical-Bayes sparsity can prune directions and assign zero posterior variances to some source coordinates.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Estimating distributed neural sources from M/EEG measurements is an ill-posed inverse problem where principled uncertainty quantification remains challenging. The usual conditional Gaussian posterior covariance does not explicitly distinguish uncertainty along likelihood-visible source directions from uncertainty that is invisible to the measurements. We propose a spectral framework based on the singular value structure of the prior-preconditioned forward operator, which provides an exact, interpretable decomposition of the posterior covariance into data-informed and indeterminacy epistemic components. This formulation identifies uncertainty contributions that are reducible by better measurements from those that are fundamentally irreducible by sensing geometry. We further derive an equivalent subtractive decomposition that expresses the prior covariance as the sum of posterior and data-driven reduction terms, enabling precise accounting of uncertainty reduction through Bayesian learning. Together, these complementary representations offer both theoretical insight and practical tools for interpretable uncertainty quantification in M/EEG source imaging for linear-Gaussian Bayesian inverse solvers where the posterior covariance is available in closed form.
1 Introduction
M/EEG source estimation is ill-posed, and standard posterior source covariance does not explicitly separate likelihood-visible from likelihood-null uncertainty. The paper introduces a spectral decomposition that distinguishes these components and an equivalent accounting of data-driven uncertainty reduction.
1 Introduction
- Standard posterior source covariance does not make likelihood-visible and likelihood-null contributions explicit, despite the need for principled uncertainty quantification.The paper frames this as a source-space gap, distinct from the additive epistemic–aleatoric split available for sensor covariance.Source: §1 · p. 2
- The exact posterior covariance decomposition separates data-informed uncertainty from indeterminacy uncertainty using the prior-preconditioned forward operator’s singular structure.The method uses visible and null singular-vector subspaces; the decomposition is presented for linear-Gaussian inverse solvers with closed-form posterior covariance and fixed or learned diagonal priors.Source: §1 · p. 2
- The equivalent prior-minus-reduction identity attributes posterior covariance to prior covariance minus the variance removed by data.For each visible mode, the removed fraction is determined by the singular value and noise variance; a singular prior’s mapped null component includes no conditional variance along pruned coordinates.Source: §1 · p. 2
- Empirical-Bayes sparsity can yield zero posterior variances in pruned directions, while aleatoric noise enters through shrinkage factors and predictive sensor covariance rather than source posterior covariance.The posterior covariance is characterized as epistemic uncertainty.Source: §1 · p. 2
2 Problem Description
The problem description sets up Type-II Bayesian learning for a linear-Gaussian source model and identifies the source-space uncertainty split the paper seeks. It also distinguishes sensor-level epistemic and aleatoric variability from uncertainty about fitted hyperparameters.
2.1 Type-II Bayesian Learning Framework
- Type-II learning uses a zero-mean Gaussian source prior with diagonal covariance, and estimates its hyperparameters by minimizing negative marginal log-likelihood.The posterior source distribution is Gaussian conditional on the fitted hyperparameters.Source: §2.1 · pp. 2–3
- Sensor covariance separates projected prior-induced epistemic variability from aleatoric measurement noise, but uncertainty about fitted prior hyperparameters is not propagated in the conditional posterior.The sensor-space noise covariance is σ2I.Source: §2.1 · p. 3
- The source posterior covariance is epistemic, while standard formulations do not canonically split it into data-visible reducible and nullspace irreducible components.The paper derives a spectral source-space attribution analogous in spirit to the additive sensor-space covariance split.Source: §2.1 · p. 3
3 Spectral Uncertainty Decomposition
Prior preconditioning makes the uncertainty components orthogonalizable through spectral analysis of a modified Hessian operator. The resulting reformulation remains valid for singular, positive-semidefinite priors, enabling the spectral decomposition without eliminating zero conditional variances for pruned coordinates.
3.1 Prior-Preconditioned Data-Misfit Hessian Operator
- Prior preconditioning orthogonalizes uncertainty components through spectral analysis of a modified Hessian operator.The passage identifies this as the section’s key theoretical insight.Source: §3.1 · p. 3
- In sparse Bayesian learning, empirical-Bayes optimization can drive prior variances to zero, making the prior covariance singular.The prior covariance is typically positive semidefinite under this regime.Source: §3.1 · p. 3
- A positive-semidefinite reformulation avoids prior inverses and supports spectral analysis even when the prior is rank-deficient.The spectrum is analyzed in a whitened parameter space where the prior has been normalized to identity; zero conditional variances for pruned coordinates remain.Source: §3.1 · pp. 3–4
4 Additive Spectral Decomposition
The prior-preconditioned forward operator’s singular-value structure partitions posterior uncertainty into visible and nullspace components. The visible component contracts with data-informed shrinkage, while the nullspace component remains irreducible under the fixed sensing geometry.
4 Additive Spectral Decomposition
- The singular vectors partition prior-preconditioned source coordinates into a visible subspace and a nullspace.The nullspace dimension is N − r, where r is the rank of the operator.Source: §4 · p. 4
- The posterior covariance has an exact additive split into data-informed uncertainty and indeterminacy uncertainty.The split is expressed through the prior-preconditioned forward operator’s spectral structure.Source: §4 · p. 4
- The nullspace term represents epistemic uncertainty from directions that the data cannot identify.The spectral construction uses the prior-preconditioned forward operator.Source: §4 · p. 4
- Visible-mode uncertainty is reduced by larger singular values or smaller noise variance, while noise affects posterior covariance only through shrinkage.The noise variance does not add a separate source-posterior covariance component.Source: §4 · p. 4
1. Data-Informed Uncertainty (reducible, epistemic)
Data-informed uncertainty is the posterior uncertainty in the observable subspace, with each visible mode scaled by a shrinkage factor. Its magnitude depends on observability, effective noise, and the prior allocation along visible modes.
1. Data-Informed Uncertainty (reducible, epistemic)
- Each visible mode’s posterior uncertainty is scaled by a shrinkage factor that measures how strongly the data constrain that direction.The factor approaches 1 when information gain is limited and approaches 0 when the squared singular value is much larger than the noise variance.Source: §1 · p. 5
- Reducible uncertainty decreases with larger singular values or smaller noise variance.The singular values reflect the prior-preconditioned forward operator; effective noise can be reduced by averaging or whitening.Source: §1 · p. 5
- Aleatoric noise affects source-posterior uncertainty parametrically through shrinkage rather than as a separate covariance component.Measurement-noise randomness belongs to the observation model.Source: §1 · p. 5
2. Indeterminacy Uncertainty (irreducible under fixed geometry, epistemic - geometric)
The paper distinguishes irreducible nullspace uncertainty from data-driven variance reduction and illustrates both in simulated EEG. In the reported geometry, top-ranked locations can retain comparable visible and nullspace contributions even as the data remove prior variance.
2. Indeterminacy Uncertainty (irreducible under fixed geometry, epistemic - geometric)
- Nullspace uncertainty is irreducible under fixed geometry and cannot be reduced by lowering noise or averaging trials.It can be reduced only by changing the forward model or constraints, or by altering prior support.Source: §2 · p. 5
5 Prior-Minus-Reduction Decomposition
- The subtractive identity expresses posterior covariance as prior covariance minus data-driven reduction, with each observable mode weighted by the fraction of prior variance removed.The remaining posterior fraction in mode i is r_i.Source: §5 · p. 5
6 Numerical Results
- With M=64 sensors and N=100 sources, many top-k sites show comparable data-informed and nullspace contributions, while Fig. 3 shows substantial prior-variance removal.The plotted variances are marginal per-source quantities at the same top-k locations.Source: §6 · p. 6
7 Discussion and Outlook
The paper presents additive and subtractive views of posterior uncertainty: one separates data-informed uncertainty from nullspace indeterminacy, while the other accounts for data-driven reduction from the prior. The nullspace component is irreducible under fixed sensing geometry, and empirical-Bayes pruning may yield zero posterior variances and overconfidence.
7 Discussion and Outlook
- The additive decomposition separates visible, data-informed posterior uncertainty from nullspace indeterminacy, while the subtractive identity expresses posterior covariance as prior covariance minus data-driven reduction.The data-driven reduction has mode-wise weights 1 − r_i, and Σpost + Rdata = Σprior.Source: §7 · p. 6
- The nullspace term is structurally irreducible under fixed geometry, and empirical-Bayes pruning can yield zero posterior variances that may lead to overconfidence.Calibration was not evaluated in this experiment; mitigating degeneracy is identified as future work.Source: §7 · p. 6
- Bayesian learning reduces uncertainty only in sensor-observable directions, while posterior uncertainty diminishes relative to the prior when data provide stronger informative constraints.Source: §7 · p. 6
A Derivation of Spectral Uncertainty Decomposition
The derivation uses a rank-aware singular-value decomposition of the prior-preconditioned operator, including when the prior is singular. A resolvent identity then relates posterior uncertainty in the visible subspace to the data-driven reduction, while the resulting framework is compared with prior-preconditioned methods and a Bayesian deep-learning decomposition.
A Derivation of Spectral Uncertainty Decomposition
- The derivation starts from an equation valid for singular priors and uses a rank-r decomposition with visible and nullspace bases; it does not require inverting the prior covariance.Source: A Derivation of Spectral Uncertainty… · p. 7
- The derivation extends the same decomposition to a singular prior by replacing M with its rank r.Source: A Derivation of Spectral Uncertainty… · p. 7
- The resolvent identity partitions prior covariance in the visible subspace into the visible posterior component and data-driven uncertainty reduction.Rdata is the prior–posterior difference and acts only in the observable subspace.Source: A Derivation of Spectral Uncertainty… · pp. 7–8
- The subtractive form aligns with prior-preconditioned Hessian methods [1, 2], while the paper's additive form makes the nullspace term explicit; a related Laplace-posterior split has a physical sensing-indeterminacy interpretation here [5].For M/EEG, the nullspace corresponds to physical sensing indeterminacy induced by the forward operator L.Source: A Derivation of Spectral Uncertainty… · p. 8
B Auxiliary lemma
The auxiliary lemma gives an identity for the diagonal matrix Λvis and its resolvent. Its equalities follow from rearrangement and commutation in the Vvis basis.
B Auxiliary lemma
- The lemma states I − (I + Λvis)^−1 = Λvis(I + Λvis)^−1 = (I + Λvis)^−1Λvis.Source: B Auxiliary lemma · p. 8
- Rearranging the difference (I + Λvis) − I gives Λvis, yielding the first equality after right-multiplication by (I + Λvis)^−1.Source: B Auxiliary lemma · p. 8
- Because Λvis is diagonal in the Vvis basis, it commutes with (I + Λvis)^−1, giving the second equality.Source: B Auxiliary lemma · p. 8
C Figures
The figures show the posterior mean on the left hemisphere and variance summaries at selected sites. The third figure compares posterior variance and data-driven reduction against prior variance.
C Figures
- Figure 1 maps time-averaged posterior-mean amplitude |x| in lateral and medial views, with red dots marking the ten sites used in the bar plots.Source: Figure 1 · p. 9
- Figure 2 presents stacked bars of posterior variance at the top ten sites.Source: Figure 2 · p. 9
- Figure 3 stacks posterior variance in green and data-driven reduction in purple, whose sum equals prior variance.Source: Figure 3 · p. 9