Where the Curvature Penalty Lives: Intrinsic Estimation Bounds Along Entropy Leaves

When a parameter lives on a curved space, such as the set of covariance matrices, the best achievable estimation error is set by that space's geometry. A recent line of work makes this precise with intrinsic Cramér-Rao bounds. These are error floors that measure mistakes along the shortest curved path rather than by straight-line subtraction, and they show that curvature adds an extra penalty. A separate line of work in geometric thermodynamics treats entropy as a Casimir invariant, meaning a quantity that no reversible (energy-conserving) motion can change. Under this view the parameter space is sliced into "leaves" of constant entropy. Reversible motion slides within a leaf, and only dissipation moves the state from one leaf to another. This note proposes to estimate in those two directions separately. We conjecture, and in the covariance case can already show, that the curvature penalty sits entirely inside the entropy leaves. The across-leaf (entropy) direction is geometrically flat. This tells us which part of a curved estimation problem is truly hard, and it suggests trackers for drifting systems that treat the reversible and dissipative parts of the motion separately.

The payoff is practical. In radar clutter tracking, sensor covariance monitoring, or any system whose covariance is reshaped quickly by reversible "mixing" while its overall volume (log-determinant, hence entropy) creeps slowly upward, existing recursive estimators on curved spaces use one update rule for everything. We will derive the error floor for tracking such a system along its foliation, build an estimator that matches it, and test whether the split actually matters against strong single-rule baselines.

Core Claim

The concrete starting point is the space of positive definite covariance matrices with the affine-invariant metric $ds^2 = \mathrm{tr}((\Sigma^{-1}d\Sigma)^2)$. This metric is the Fisher metric of zero-mean Gaussians. In Souriau's construction it is also the Koszul metric whose associated entropy Barbaresco identifies as a generalized Casimir function on coadjoint orbits (Lie Group Statistics and Lie Group Machine Learning…). Write $\Sigma = e^{h}\bar\Sigma$ with $\det\bar\Sigma = 1$ and $h = \tfrac{1}{p}\log\det\Sigma$, where $p$ is the dimension. Gaussian entropy is $S = \tfrac{p}{2}h + \text{const}$, and it is unchanged by the action $\Sigma \mapsto A\Sigma A^\top$ of determinant-one matrices $A$. The constant-entropy leaves are therefore the fixed-determinant slices. Because $\mathrm{tr}(\bar\Sigma^{-1}d\bar\Sigma)=0$, the metric splits exactly:

$$ds^2 = p\,dh^2 + \mathrm{tr}\big((\bar\Sigma^{-1}d\bar\Sigma)^2\big).$$

So the space is a Riemannian product of a flat line (entropy) and the curved symmetric space of unit-determinant matrices (the leaf). It follows that the curvature term in the intrinsic Bayesian bound of Bouchard et al. acts only on the leaf block. The entropy block is a curvature-free scalar bound.

The proposal turns this observation into three results. The first is a general condition for when a Souriau family's Fisher/Koszul metric splits orthogonally along its Casimir foliation, together with the coupling term that appears in the bound when it does not. Barbaresco's "transverse Riemannian foliation" (Symplectic Foliation Structures of Non-Equilibrium Thermodynamics) is the natural candidate condition. The second is an intrinsic tracking bound, a recursive Bayesian floor, for states that move reversibly within leaves while drifting monotonically across them. The third is a split recursive estimator that attains it.

The static entropy bound on its own is not claimed as new. A Cramér-Rao bound for a scalar function does not depend on how the parameters are coordinatized, so the classical delta-method variance already gives it. The new content is where the curvature sits and what that implies for tracking.

Relation to Prior Work

Two bodies of work supply the pieces, and neither touches the other.

Barbaresco's papers construct Gibbs densities on Lie groups, the Koszul-Souriau-Fisher metric, and entropy as a coadjoint invariant (Lie Group Statistics…). They extend this to affine actions and structure-preserving integrators (Lie Group Cohomology and (Multi)Symplectic Integrators) and to dissipation as motion across leaves via a metric bracket (Symplectic Foliation Structures…). These papers motivate the foliation but contain no finite-sample estimation, no bounds, and no data. We take their geometry and add the statistics.

Bouchard and coauthors supply the method. They derive the intrinsic Bayesian bound with its curvature correction and show that MAP and MMSE covariance estimators are asymptotically efficient intrinsically (Intrinsic Bayesian Cramér-Rao Bound). They derive quotient-manifold bounds for low-rank elliptical models attained by Riemannian gradient descent (A Riemannian Framework for Low-Rank Structured Elliptical Models). They also give a recursive estimator for compound Gaussian clutter on a product of unit-determinant matrices and textures that reaches the intrinsic bound at constant cost per step (Riemannian geometry for compound Gaussian distributions).

That last paper is the closest existing work. It already uses the unit-determinant normalization, because it decouples the Fisher metric. It treats the decoupling as a computational convenience for a static or slowly varying parameter. It does not read the split as the entropy foliation, does not model reversible within-leaf motion separately from monotone across-leaf drift, and does not derive a tracking bound. Two other papers constrain the design. Boumal et al. show that intrinsic bounds on compact groups reduce to graph-Laplacian quantities, a reminder that clean closed forms come from symmetry. Prévost et al. show that bounds conditioned on a random constraint can be beaten by real algorithms, so our entropy-drift constraint must be marginalized in the Bayesian bound rather than conditioned on.

Research Program

The theory first states the splitting criterion for a general Souriau Gibbs family. Since the gradient of $S$ in the mean-momentum coordinates is the temperature vector $\beta$, the transverse direction is the metric dual of $\beta$. Orthogonal splitting holds when the leaf metric does not vary along this dual direction. We will check the criterion on the three families with closed forms in the file: Gaussian covariance, which splits; $SU(1,1)$ on the Poincaré disc; and $SE(2)$. In each case we will write the off-diagonal term of the bound when splitting fails.

This also settles the tension the file raises about dually flat exponential-family connections versus Levi-Civita curvature in the intrinsic bound. Our prediction is that the curvature penalty is real but confined to leaf directions. The entropy direction is flat in both senses.

Next we derive a recursive intrinsic Bayesian bound for the following dynamic model. At each step the leaf state is moved by a random determinant-one transformation, standing in for reversible mixing. The entropy coordinate $h$ increases by a nonnegative random increment, standing in for dissipation consistent with $\dot S \ge 0$. Each step yields $n$ samples. On a product space the information recursion separates into a curved leaf recursion and a scalar recursion for $h$. The nonnegativity of the increments enters through the prior, not as a conditioning constraint.

The estimator is the recursive Riemannian update of the compound Gaussian paper, modified in two ways. Leaf and entropy coordinates get separate step sizes matched to their respective process noise, and the entropy estimate is projected to be non-decreasing. We also test a variant that uses Gaussian entropy's exact finite-sample bias correction in the entropy coordinate.

Baselines

The naive reference is the windowed sample covariance, averaged arithmetically over time. Bouchard et al. (compound Gaussian) report that this kind of estimator hits an error floor that does not improve with more data, so any curved-space method should beat it clearly. It will get the same window-length tuning grid as everything else.

The strongest existing competitor is the joint recursive Riemannian gradient descent of the same paper, which reaches the intrinsic bound for slowly varying clutter. We run it on identical simulated sequences with the same samples per step and the same total tuning budget. Its single step size gets a grid search as large as the two-dimensional grid given to our split estimator. The file reports no number for this baseline: no reported result found on Lacuna for the tracking error of recursive Riemannian descent on the K-distribution benchmark, only that it reaches the intrinsic bound. We therefore reproduce it and treat its error relative to our derived tracking bound as the target.

The second strong competitor is a log-Euclidean extended Kalman filter, which tracks the matrix logarithm of the covariance. Its process-noise covariance will be tuned on held-out sequences.

The ablations isolate the key ingredient. One removes the step-size split but keeps monotone projection. Another keeps the split but removes projection. A third runs the split in a non-orthogonal coordinate pair, the trace-normalized matrix plus trace, which does not respect the foliation. If that wrong split performs as well as the right one, the foliation is doing no work.

Beating the strongest baseline means a statistically significant reduction in intrinsic tracking error (leaf and entropy components reported separately) across repeated sequences, in the regime where leaf and entropy rates differ. It also means staying within a small factor of the derived bound where the joint method does not.

Why not just…?

Why not just use the delta method for the entropy bound? For static estimation of entropy alone, the delta method is correct, and we concede this. The scalar bound depends only on the gradient of the entropy and the inverse Fisher information, not on the coordinates. That is exactly why the proposal does not sell a "Casimir entropy bound." What the delta method cannot say is where the curvature correction goes, or how leaf and entropy errors interact under dynamics. The static experiment checks our prediction: the entropy component should track its flat bound at small sample sizes, while the leaf component shows the curvature-driven gap.

Why not just use a Euclidean constrained bound with a fixed-determinant constraint? This handles conservation at one point by linearizing the leaf. It measures error with straight lines, which Bouchard et al. show misjudges efficiency on this space. When the constraint level drifts randomly, Prévost et al. show conditioned constraint bounds can be violated by real estimators. We will compute it alongside ours and report whether the reproduced estimators fall below it.

Why not just tune a single-step joint Riemannian tracker, or a Kalman filter in log coordinates? This is the real threat. Because the metric already decouples in the compound Gaussian paper's coordinates, a joint tracker with a well-chosen step might nearly match the split one. The experiment settles it by sweeping the ratio of leaf-mixing rate to entropy-drift rate over several orders of magnitude. If the tuned joint tracker stays within noise of the split tracker throughout, the estimator claim is dropped and the contribution reduces to the bound and the splitting criterion. We will report it that way.

Why not just use Shannon entropy? For Gaussians, Souriau's entropy equals differential entropy, so the Casimir view adds no new quantity here. What it adds is a principled split between motion that cannot change entropy and motion that must increase it. That split is what the tracker and the bound exploit.

Proposed Experiments

The static check reuses the inverse-Wishart setup with $p=5$ from the intrinsic Bayesian bound paper. It decomposes the intrinsic error of MAP, MMSE, and sample-covariance estimators into leaf and entropy parts and compares each to its own bound as the sample size shrinks toward $p$. The decisive outcome is that the entropy part attains the flat bound (after bias correction) at sample sizes where the leaf part is visibly above its curved bound. If instead the entropy part also departs early, the splitting picture is wrong in practice.

The second check repeats this on the Student-t low-rank benchmark ($p=16$, ranks 4 and 8, three degrees of freedom) from the low-rank elliptical paper. That family's metric is not expected to split cleanly, so the experiment tests the coupling term.

The dynamic check uses the K-distributed clutter setting with $p=10$, $n \in \{20, 50\}$ and $T=1000$ from the compound Gaussian paper, augmented with the reversible-mixing plus monotone-drift dynamics above, and compares all baselines against the derived tracking bound.

The theory check evaluates the splitting criterion and bounds numerically on $SU(1,1)$ and $SE(2)$ using Barbaresco's closed-form Gibbs densities, with Monte Carlo maximum-likelihood estimates. This is the first time those families would be checked against finite-sample data at all.