Permutation-Averaged Spectral Estimation

• PASE is a group-theoretic framework that estimates spectral properties from a single data vector by averaging over permutation actions.
• It formalizes a group-averaged covariance estimator that recovers classical transforms such as the DFT, DCT, and Karhunen–Loève transform.
• PASE offers polynomial complexity and practical applications in areas like massive MIMO, waveform classification, graph signal processing, and transformer analysis.

Permutation-Averaged Spectral Estimation (PASE) is a group-theoretic framework for spectral analysis that replaces temporal averaging over multiple observations (snapshots) with algebraic averaging over group actions on a single observation. Developed to enable second-order statistical estimation from a single data vector, PASE formalizes the group-averaged covariance estimator and demonstrates that, under suitable conditions, this estimator yields eigenspace decompositions equivalent to those arising from multi-snapshot sample covariances. The symmetric group $S_n$ is proven to be universally optimal within this framework, recovering the Karhunen–Loève (KL) transform from a single sample and unifying classical spectral transforms such as the Discrete Fourier Transform (DFT) and Discrete Cosine Transform (DCT) as special cases (Thornton, 4 Apr 2026).

1. Definition of the Permutation-Averaged Covariance

Given a single real data vector $x \in \mathbb{R}^n$, Permutation-Averaged Spectral Estimation constructs the covariance estimator by averaging over the action of the symmetric group $S_n$ via permutation matrices $P(\pi)$: $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$ where $x_\pi = P(\pi)x$. This fully symmetrized estimator serves as an algebraic surrogate for time-averaged covariance, providing a principled mechanism to synthesize second-order statistics from a single input.

2. Group-Averaged Covariance: Replacement Theorem

The Group Replacement Theorem establishes the equivalence between temporal and group-wise averaging for subspace recovery. For $x = s + n$, with $s$ lying in a low-dimensional subspace equivariant under a group $G$ and $n \sim \mathcal{N}(0, \sigma^2 I)$ being white noise invariant under $x \in \mathbb{R}^n$0:

• The group-averaged estimator can be decomposed as $x \in \mathbb{R}^n$1, with $x \in \mathbb{R}^n$2.
• The expected group-averaged signal contribution, $x \in \mathbb{R}^n$3, block-diagonalizes into at most $x \in \mathbb{R}^n$4 blocks (by Schur's lemma).
• The noise term satisfies $x \in \mathbb{R}^n$5.
• In the high signal-to-noise regime ($x \in \mathbb{R}^n$6), the principal eigenspace of $x \in \mathbb{R}^n$7 converges to the signal subspace, mirroring the behavior of multi-snapshot sample covariance matrices.

This theorem rigorously justifies the use of algebraic group actions in lieu of temporal or ensemble averaging when suitable group symmetries are present in the signal model.

3. Double-Commutator Eigenproblem and KL Transform

Optimal group selection within the PASE framework reduces to finding a basis $x \in \mathbb{R}^n$8 that jointly diagonalizes an unknown covariance $x \in \mathbb{R}^n$9 and a candidate averaging operator $S_n$0 from the group algebra. This requirement is captured by the double-commutator eigenvalue problem: $S_n$1 where $S_n$2. The eigenmatrices $S_n$3 satisfying this equation are precisely those spectral idempotents commuting with $S_n$4. For group-algebraic choices of $S_n$5 (e.g., from the Cayley-graph adjacency), the solution basis $S_n$6 renders both $S_n$7 and $S_n$8 diagonal, identifying $S_n$9 as the KL (Karhunen–Loève) basis. When $P(\pi)$0, the regular-representation Cayley operator $P(\pi)$1 possesses a commutant algebra containing every possible population covariance $P(\pi)$2, so the KL transform arises as the PASE-optimal diagonalization.

4. Universal Optimality of the Symmetric Group $P(\pi)$3

The symmetric group $P(\pi)$4 is proven to yield universal optimality for group-averaged spectral estimation:

• The regular representation of $P(\pi)$5 decomposes into all irreducible representations with full multiplicity, and its group algebra commutant is the full matrix algebra. Therefore, $P(\pi)$6 commutes with any covariance $P(\pi)$7.
• By classical results, the KL transform is unique in decorrelating the data, concentrating variance, and minimizing truncation mean-squared error. As $P(\pi)$8 shares the KL basis, no other group can outperform $P(\pi)$9 in these respects.
• Whenever $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$0 is circulant (i.e., for periodic signals), it commutes with cyclic shifts from $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$1, so PASE over $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$2 subsumes the DFT as a special instance of KL diagonalization.

5. Algorithmic Procedure and Computational Complexity

PASE proceeds as follows:

1. Permutation orbit: For each $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$3, compute $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$4.
2. Covariance averaging: Form $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$5.
3. Eigen-decomposition: Compute $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$6 and extract the leading eigenvectors $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$7.

The naïve algorithmic costs are $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$8 for covariance accumulator formation and $$\hat R_{S_n} = \frac{1}{|S_n|} \sum_{\pi \in S_n} P(\pi)\,x\,x^T\,P(\pi)^T = \frac{1}{n!} \sum_{\pi \in S_n} x_\pi x_\pi^T,$$9 for eigen-decomposition. However, for groups supporting fast transforms (e.g., $x_\pi = P(\pi)x$0 via FFT, $x_\pi = P(\pi)x$1 via DCT), group Fourier transforms reduce the cost to $x_\pi = P(\pi)x$2 for extracting diagonal entries and eigenpairs in each isotypic block. Thus, PASE is of polynomial complexity and near-linear for classical groups.

6. Special Cases: DFT, DCT, and KLT

PASE recovers classical transforms for specific choices of $x_\pi = P(\pi)x$3:

• For $x_\pi = P(\pi)x$4 (cyclic shifts), the PASE estimator $x_\pi = P(\pi)x$5 is circulant and diagonalized by the DFT, matching the KL basis for periodic signals.
• For $x_\pi = P(\pi)x$6 (dihedral group), the group algebra introduces block-Toeplitz and reversal structure, with the DCT family diagonalizing the resulting covariance, suited for signals with even-symmetry boundary conditions.
• For $x_\pi = P(\pi)x$7 (symmetric group), PASE recovers the full KLT for arbitrary covariance structure.

Illustrative examples include:

• Periodic AR(1) signals: PASE over $x_\pi = P(\pi)x$8 yields the DFT, matching the $x_\pi = P(\pi)x$9-snapshot periodogram from a single snapshot.
• LFM (chirp) signals: Applying a dechirp conjugation then PASE with $x = s + n$0 enables automatic estimation of frequency via maximizing spectral concentration and recovers the dechirp-then-FFT procedure.
• Graph signals: With automorphism group $x = s + n$1, PASE yields the graph Fourier basis. For graphs with non-abelian automorphism groups, such as $x = s + n$2 on $x = s + n$3, there is a 15–25% advantage in spectral concentration compared to conjugated cyclic groups, indicating a genuine non-abelian benefit for relevant topologies.

7. Applications and Unified Framework

Permutation-Averaged Spectral Estimation enables single-snapshot recovery of spectral structure in diverse domains:

• Direction-of-arrival estimation (MUSIC) from a single snapshot
• Massive MIMO channel estimation yielding a 64% throughput gain
• Single-pulse waveform classification at 90% accuracy
• Graph signal processing leveraging non-Abelian symmetries
• Transformer LLM analysis, revealing that current position encodings such as RoPE utilize non-optimal algebraic groups for 70–80% of heads, with spectral-concentration-based pruning improving perplexity in large models

All these diagnostics and analyses require only a single forward pass, without gradient-based optimization or training. PASE is thus a unifying framework, implementable in polynomial time, that generalizes and subsumes classical spectral estimators under group-theoretic principles (Thornton, 4 Apr 2026).