Spectra & Decisions

The question. Can the eigen-spectrum of a relation or covariance matrix be read as a decision instrument — with random matrix theory as the null model that separates signal from noise — across business data, L-functions, and statistical mechanics alike?

Reading as

Eigenvalues quietly run high-dimensional data, prime numbers, and physics. Can they run a decision?

Status — Early. An open, exploratory thread — shared here for the questions and the method, not for results. Findings are kept off the public site until they're earned and vetted. Tools: NumPy / SciPy · random matrix theory · LaTeX.

The bet

Eigendecomposition is already the engine under PCA and high-dimensional analysis, under the Montgomery–Dyson link between Riemann zeta zeros and the eigenvalue spacings of random Hermitian matrices, and under spectral methods in physics. The bet is that one lens — the spectrum, read against its random-matrix null — gives a principled, transferable way to tell structure from noise, and to act on it.

The framing is process-relational (it shares a substrate with P9): the objects are relations and actions — operators — not static things. We read the world through what transforms, and the invariants are spectral. Mental formulations and mental objects, with no birth or death; only morphisms.

The setup

Definitions first — the part that has to be precise before any computation is allowed to mean anything.

The object

A relation / covariance matrix built from data (or an operator). Its eigenvalues and eigenvectors are the spectrum.

The null

Random matrix theory as the no-structure baseline: Marchenko–Pastur for covariance bulk, Tracy–Widom for the edge, GUE spacings for correlations.

The decision

Eigen-directions that escape the random-matrix null are the candidate signal — the factors a decision should actually rest on.

The method — what actually gets run

Build covariance / relation matrices from a real high-dimensional dataset.
Compare the empirical spectrum to the Marchenko–Pastur / Tracy–Widom null.
Keep only eigen-directions past the null edge; test whether decisions built on them beat ad-hoc thresholds.
Probe the Montgomery–Dyson bridge: do GUE spacing statistics describe a concrete empirical operator, not just zeta zeros?

Open questions (honestly open)

These aren't rhetorical. Each is a place the project could still be a metaphor instead of a result.

S1: eigenvalues beyond the Marchenko–Pastur edge identify decision-relevant factors more robustly than scree-plot rules of thumb.
S2: the GUE spacing statistics matching zeta zeros also describe some real empirical operator — a bridge, not an analogy.
Q: is “reading the world through morphisms and actions” formalisable as operators whose spectra are the invariants? (shared with P9)

Key formulas

λ_{+} = (1 + p / n)^{2}

Marchenko–Pastur upper edge: eigenvalues above it are signal; below, indistinguishable from noise.

r_{n} = \frac{min ( s _{n} , s _{n + 1} )}{max ( s _{n} , s _{n + 1} )}

Consecutive spacing ratio (no unfolding needed). Mean ⟨r⟩: Poisson 0.386, GOE 0.536, GUE 0.600, rigid → 1.

P_{GUE} (s) \approx \frac{32}{π ^{2}} s^{2} e^{- 4 s^{2} / π}

Wigner surmise (GUE): level repulsion — P(s) vanishes as s → 0. The law the Riemann zeros follow.

Reading the work rests on

Modern Multivariate Statistical TechniquesLinear Algebra Done RightInformation TheoryIntroduction to Mathematical Statistics

This is an in-progress thread, not a finished claim. If you work near it — TDA, phenomenology, phonetics, the foundations of inference — I'd genuinely like to compare notes: research@backporch.studio.