Symmetry-Induced Statistic Recovery

• The paper demonstrates that symmetry constraints can force estimators, such as maximum likelihood and variational minimizers, to exactly recover target statistics.
• It details methodical frameworks across variational inference, Gaussian graphical models, and orbit recovery, ensuring effective recovery even under model misspecification.
• The work highlights practical implementations, limitations, and regularization techniques that exploit symmetry for improved statistical estimation.

Searching arXiv for recent and foundational papers on symmetry-induced statistic recovery. arXiv search query: "symmetry-induced statistic recovery variational inference graphical Gaussian models symmetry recovery statistics" Symmetry-induced statistic recovery denotes a class of inference phenomena in which invariance, equivariance, or quotient structure forces an estimator, variational minimiser, test statistic, or recovered representation to reproduce a statistic of interest even when the full target is misspecified, only partially identifiable, or observed through a scrambled measurement process. In the literature, this theme appears in graphical Gaussian models where the maximum likelihood estimator of the mean equals the least-squares estimator exactly under graph-coloring compatibility (Gehrmann et al., 2011), in variational inference where symmetry plus uniqueness forces both the target and the variational minimiser into the same statistic-fixed set (Marks et al., 20 Apr 2026), and in Bayesian inference problems where the correct recoverable object may be an orbit, an invariant, or a canonical orbit representative rather than the raw signal itself (Semerjian, 14 Jan 2025).

1. General mechanism

A common formalism begins with a group $G$ acting on an object space and thereby inducing an action on distributions, parameters, or statistics. In the variational setting, measurable bijections $g:X\to X$ act by pushforward,

$g_\#\pi(A)=\pi(g^{-1}(A)),$

and a statistic is a partial map $S:\mathcal P \rightharpoonup Y$. If $P$ is $G$-invariant, $\mathcal Q$ is $G$-stable, the statistic is compatible with the action, and the variational problem has a unique minimiser $Q^\star$, then the induced action

$\rho_g(y)\coloneqq S(g_\#\pi)$

is well defined and both $g:X\to X$0 and $g:X\to X$1 lie in the fixed set

$g:X\to X$2

The mechanism is that $g:X\to X$3-divergences satisfy

$g:X\to X$4

so uniqueness upgrades invariance of the minimiser set to invariance of the minimiser itself (Marks et al., 20 Apr 2026).

An algebraically parallel formulation appears in Gaussian mean estimation. For fixed covariance $g:X\to X$5, Kruskal’s theorem states that the maximum likelihood estimator $g:X\to X$6 and the least-squares estimator $g:X\to X$7 coincide if and only if the mean space is invariant under the concentration matrix: $g:X\to X$8 with $g:X\to X$9. Here recovery is an invariant-subspace question: the covariance operator must not mix the constrained mean classes asymmetrically (Gehrmann et al., 2011).

A third formulation is orbit recovery. When a symmetry acts nontrivially on the signal but trivially on the observations, the identifiable object is not $g_\#\pi(A)=\pi(g^{-1}(A)),$0 itself but its orbit $g_\#\pi(A)=\pi(g^{-1}(A)),$1. The appropriate loss is then the quotiented distance

$g_\#\pi(A)=\pi(g^{-1}(A)),$2

so recovery is interpreted modulo symmetry rather than pointwise in the original parameterization (Semerjian, 14 Jan 2025).

2. Exact recovery in graphical Gaussian models with symmetries

In undirected graphical Gaussian models, a graph $g_\#\pi(A)=\pi(g^{-1}(A)),$3 encodes conditional independences through zeros of the concentration matrix $g_\#\pi(A)=\pi(g^{-1}(A)),$4. Symmetry restrictions are represented by a colored graph $g_\#\pi(A)=\pi(g^{-1}(A)),$5, where $g_\#\pi(A)=\pi(g^{-1}(A)),$6 is a partition of $g_\#\pi(A)=\pi(g^{-1}(A)),$7 into vertex color classes and $g_\#\pi(A)=\pi(g^{-1}(A)),$8 is a partition of $g_\#\pi(A)=\pi(g^{-1}(A)),$9 into edge color classes. The paper studies three model types: RCON, with equality constraints on entries of $S:\mathcal P \rightharpoonup Y$0; RCOR, with equality constraints on partial correlations; and RCOP, with permutation-symmetry models from a subgroup of automorphisms (Gehrmann et al., 2011).

Mean constraints are encoded by a partition $S:\mathcal P \rightharpoonup Y$1 through

$S:\mathcal P \rightharpoonup Y$2

The main theorem states that for both RCON and RCOR models, and hence also for RCOP models,

$S:\mathcal P \rightharpoonup Y$3

The condition $S:\mathcal P \rightharpoonup Y$4 requires the mean partition to be finer than the vertex-color partition. Vertex regularity requires that for every edge color class $S:\mathcal P \rightharpoonup Y$5, the partition $S:\mathcal P \rightharpoonup Y$6 is equitable with respect to the subgraph $S:\mathcal P \rightharpoonup Y$7, equivalently

$S:\mathcal P \rightharpoonup Y$8

for all $S:\mathcal P \rightharpoonup Y$9, whenever $P$0 lie in the same class of $P$1. The graph-theoretic content is that vertices in the same mean class must “see” every mean class through each edge color in exactly the same way.

This criterion explains both success and failure. In the Frets’s heads example, the symmetry-compatible partition yields fitted means

$P$2

so the mean is recovered by the corresponding averages. In the mathematics marks RCOP example, imposing

$P$3

gives the simple average

$P$4

By contrast, in the Behrens–Fisher-like case with

$P$5

imposing $P$6 makes the partition too coarse relative to the vertex coloring, and $P$7 in general. The exact recovery criterion is therefore not symmetry of $P$8 alone, but alignment between covariance symmetry and mean equality structure.

3. Variational inference and fixed-set recovery

In misspecified variational inference, the optimization problem is

$P$9

and the objective is not exact recovery of $G$0, but recovery of a statistic $G$1 exactly or partially even when $G$2. The general theorem requires three conditions: target invariance, variational stability, and statistic compatibility. Under uniqueness of the minimiser, symmetry then forces

$G$3

so both target and variational minimiser share the same symmetry-determined statistic constraint (Marks et al., 20 Apr 2026).

Previously known recovery results for location-scale families arise as special cases. For even symmetry about $G$4, using the reflection $G$5, the mean transforms as

$G$6

the induced fixed set is $G$7, and uniqueness yields exact mean recovery: $G$8 For elliptical symmetry about $G$9 with scale matrix $\mathcal Q$0, the covariance fixed set is

$\mathcal Q$1

so covariance is recovered only up to a positive scalar factor, while the correlation statistic is recovered exactly: $\mathcal Q$2

The paper’s new application concerns rotationally symmetric distributions on the sphere. If a non-uniform distribution on $\mathcal Q$3 has density $\mathcal Q$4, the axis is identifiable only up to sign, so the statistic is the line

$\mathcal Q$5

For the von Mises–Fisher family $\mathcal Q$6, if the target is rotationally symmetric about $\mathcal Q$7, the family is $\mathcal Q$8-stable, and the $\mathcal Q$9-divergence objective has a unique minimiser $G$0, then

$G$1

The paper also isolates the central failure mode. For targets with axial profile

$G$2

there is a threshold $G$3 such that for $G$4, the minimiser set becomes a whole latitude circle and no minimiser recovers the true axis. Symmetry alone is therefore insufficient; uniqueness is essential.

4. Orbit recovery, asymptotic invariance, and symmetry testing

When a symmetry acts on the signal while leaving the observations unchanged, the raw signal is not identifiable. The correct object is the orbit $G$5, and the correct estimator minimizes posterior risk under the quotient loss

$G$6

This viewpoint yields explicit recoverable statistics in several cases. For scalar sign symmetry $G$7,

$G$8

and the Bayes-optimal estimator is $G$9. For permutation symmetry on coordinates,

$Q^\star$0

so the estimator recovers the sorted signal $Q^\star$1, that is, an orbit representative rather than the original ordering (Semerjian, 14 Jan 2025).

A distinct asymptotic version arises in randomization theory. For a sequence $Q^\star$2, compact groups $Q^\star$3, and Lipschitz statistics $Q^\star$4, the paper studies when

$Q^\star$5

approximates the relevant conditional law of $Q^\star$6. Under Lipschitz control, moment conditions, and the assumption that $Q^\star$7 forms a normal Lévy family, the discrepancy converges to zero almost surely and in $Q^\star$8. This is the paper’s formal statement of asymptotic symmetry recovery: high-dimensional concentration makes many Lipschitz statistics nearly invariant under large compact groups even when exact invariance assumptions fail (Kashlak, 2022).

Symmetry testing makes the recovered statistic explicit. In stationary time series, ordinal patterns

$Q^\star$9

compress the process to the distribution $\rho_g(y)\coloneqq S(g_\#\pi)$0 of local rank orderings. A partition $\rho_g(y)\coloneqq S(g_\#\pi)$1 of $\rho_g(y)\coloneqq S(g_\#\pi)$2 defines the null

$\rho_g(y)\coloneqq S(g_\#\pi)$3

which covers time reversibility, reflection symmetry, and combined symmetry structures. The resulting statistic is a second-order U-statistic on ordinal-pattern labels with degenerate generalized chi-squared limit under the null and asymptotically Gaussian behavior under the alternative (Betken et al., 20 Jan 2026).

A related center-free test uses cumulative past and residual extropy of record values. For the record-based discrepancy

$\rho_g(y)\coloneqq S(g_\#\pi)$4

symmetry is characterized by

$\rho_g(y)\coloneqq S(g_\#\pi)$5

so $\rho_g(y)\coloneqq S(g_\#\pi)$6 under symmetry. The sample statistic therefore measures deviation from a left-right information balance rather than from an estimated center of symmetry (Chaudhary et al., 2022).

5. Discovery, regularization, and operational exploitation

When symmetry is known in advance, it can be used as a regularizer. For covariance estimation, the symmetry-aware estimator is the group Reynolds average

$\rho_g(y)\coloneqq S(g_\#\pi)$7

where $\rho_g(y)\coloneqq S(g_\#\pi)$8 is the empirical covariance and $\rho_g(y)\coloneqq S(g_\#\pi)$9 projects onto the fixed-point subspace. The theory gives improved spectral, Frobenius, and $g:X\to X$00 convergence rates in terms of representation multiplicities and orbit sizes, and uses the same projection as preprocessing for sparse covariance thresholding and sparse inverse covariance estimation (Shah et al., 2011).

When the symmetry is unknown, recent work treats discovery itself as part of recovery. In blind inverse problems, a shallow group-convolutional lifting network learns a latent action

$g:X\to X$01

and a resolving filter $g:X\to X$02, producing

$g:X\to X$03

Training minimizes

$g:X\to X$04

where stationarity is enforced by a Jensen–Shannon penalty, resolution uses total correlation, and InfoMax minimizes $g:X\to X$05. The reported experiments show strong signal recovery correlations, typically above $g:X\to X$06, and strong symmetry similarity scores, with recovery on stochastic processes, Ising models, shuffled and bit-scrambled images, and neural recordings (Efe et al., 16 Jun 2026).

A quantum variant scores candidate groups by

$g:X\to X$07

using either the weak twirl

$g:X\to X$08

or the strong twirl

$g:X\to X$09

The pipeline scores candidate groups, chooses the largest candidate that passes, and then measures in the symmetry-adapted basis. The seeded demonstrations report weak-mode closed-loop shot-count reductions of $g:X\to X$10, $g:X\to X$11, and $g:X\to X$12, and strong-mode reductions of $g:X\to X$13, $g:X\to X$14, and $g:X\to X$15 (Thornton, 17 Jun 2026).

Symmetry can also be exploited after measure relaxation. In optimal control under a finite group action, symmetry-adapted moment-SOS relaxations recover invariant moments of the unique $g:X\to X$16-invariant relaxed measure. Trajectory recovery then reduces to solving the symmetric parametric polynomial system

$g:X\to X$17

for invariant generators $g:X\to X$18. In the reported examples, the symmetric integrator at $g:X\to X$19 reduces variables from $g:X\to X$20 to $g:X\to X$21 and runtime from $g:X\to X$22s to $g:X\to X$23s, while the qubit example at $g:X\to X$24 replaces an out-of-memory dense SDP with a symmetric SDP having $g:X\to X$25 variables and runtime $g:X\to X$26s (Augier et al., 2023).

A dynamical analogue appears in the mass-quenched Hubbard model. The order parameter

$g:X\to X$27

tracks recovery of $g:X\to X$28 symmetry after quenching from a mass-imbalanced to a mass-balanced Hamiltonian. In the reported weak-to-moderate interaction regime, $g:X\to X$29 decreases monotonically and typically decays approximately exponentially, while the late-time momentum distribution and density of states match those of the post-quench $g:X\to X$30-symmetric Hamiltonian at an effective temperature $g:X\to X$31. In that regime, symmetry recovery and thermalization coincide (Du et al., 2017).

6. Limitations and failure modes

Symmetry does not imply unrestricted recoverability. In graphical Gaussian models, symmetry in the concentration matrix is insufficient unless the mean constraints are aligned with the graph coloring. The paper explicitly notes that hypotheses such as “lengths and breadths separately” are not vertex-regular in the Frets graph and therefore require joint maximization over $g:X\to X$32 and $g:X\to X$33. The Behrens–Fisher-like example shows the same point in minimal form: the mean partition can be too coarse relative to the vertex coloring, and then $g:X\to X$34 in general (Gehrmann et al., 2011).

In variational inference, uniqueness is the crucial assumption. Without it, the orbit of a minimiser is merely contained in the minimiser set, and the statistic need not be pinned down. The reverse-KL example with axial profile

$g:X\to X$35

shows that once $g:X\to X$36, the minimiser set becomes a whole latitude circle and no minimiser recovers the true axis (Marks et al., 20 Apr 2026).

In Bayesian inference, symmetry may be ambivalent rather than purely helpful. When the symmetry acts strongly on observations, restricting to invariant or equivariant estimators can reduce complexity without loss. When it acts strongly on signal only, the “obvious” loss can become wrong because the true representative is unobservable. In such cases the recoverable object is an orbit, a radius, sorted coordinates, Gram structure modulo column permutation, or a similar symmetry-invariant statistic (Semerjian, 14 Jan 2025).

Quantum information recovery provides a stronger obstruction. For scrambling dynamics with Lie group conservation laws, the recovery error $g:X\to X$37 obeys universal lower bounds involving a back-reaction term and the SLD quantum Fisher information. The qualitative conclusion is that if the symmetry causes nontrivial back-action on the conserved quantity, perfect recovery is impossible, and large quantum coherence can only mitigate the limitation up to the symmetry-imposed bound. In the Hayden–Preskill setting with energy conservation, the recovery error remains non-negligible until the black hole completely evaporates when the diary size is comparable to the black hole size, and remains significant until a very large fraction of the black hole has evaporated even when the diary is much smaller (Tajima et al., 2021).

Taken together, these results give a precise view of what symmetry can and cannot do. Symmetry can force equality between estimators, pin statistics to fixed sets, collapse inference to quotient variables, regularize ill-posed estimation, and even reveal latent domains. It can also obstruct identifiability, create families of equivalent optima, and impose quantitative lower bounds on recovery. The recoverable object is therefore determined not by symmetry alone, but by symmetry together with compatibility, stability, and—repeatedly—uniqueness.