teal-sea / zeta-labstate of record · compiled 14 Aug 2026 · revision 9ebdea0 · source

Library · docs/10-trace-formulas-and-connes.md

Trace Formulas and the Connes Program

3,245 words · 301 lines · source

"The explicit formula looks like modal analysis. So what is the structure — and who gets to tap it?"

The short version

A trace formula is an exact accounting identity between two lists: the resonant frequencies of a system (its spectrum) and its closed round-trip paths (its geometry). You have already verified one — the theta identity of docs/02-theta-heat-and-modularity.md is the trace formula of a circle. The Weil explicit formula is the same kind of identity with the zeros of zeta as the modes and the prime powers as the orbits; zeta/weil.py evaluates both sides independently and measures their agreement at ~1e-31. Weil distilled from the formula a THEOREM: RH is equivalent to a certain quadratic form never going negative — zeta.weil.weil_functional computes that form from primes alone, and its measured margin above zero is controlled by the lowest zero, γ₁ = 14.1347…. In one parallel world the dream is already a theorem — Selberg's trace formula for hyperbolic surfaces, where self-adjointness does the work. Over the integers we have the trace formula and no operator; Connes' noncommutative-geometry PROGRAM is the most developed attempt to build the missing space, on which RH becomes exactly a positivity statement about a flow. A deep reformulation, not a proof. Everything below is labelled THEOREM, CONJECTURE, or PROGRAM.

1. What a trace formula is — a tool you already own

Start from matrices, where the word "trace" lives: for a symmetric matrix, the trace — the sum of the diagonal — equals the sum of the eigenvalues. One number, computable two ways: from the entries (how the system is built) or from the spectrum (how it rings). Any identity of that shape is a trace formula. Engineers meet this daily as modal analysis: tap a structure and record the response — in the frequency domain, peaks at the natural frequencies (the spectral readout); in the time domain, echoes at the round-trip travel times of pulses bouncing through the structure (the geometric readout). Same signal, two descriptions, each determining the other.

The baby case in this repository is the heat kernel of docs/02-theta-heat-and-modularity.md, evaluated at coincident points — the trace of the heat operator on a circle:

    Σ_{n ∈ ℤ}  exp(-4π²n²t)     =     Σ_{m ∈ ℤ}  (4πt)^(-1/2) exp(-m²/(4t))

    [spectral side: one term per mode,      [geometric side: one term per closed
     eigenvalue 4π²n² of -d²/dx²]            path winding m times, length |m|]

Left: every vibration mode of the circle, weighted by how much survives diffusion for time t. Right: every closed walk around the circle, weighted by how likely a random walker is to complete it. The identity is Poisson summation — zeta.core.theta_heat_poisson_defect measures its defect at 0.0 to 40 digits (docs/02 §7). Modes left, closed orbits right, an exact equals sign between: that is the entire concept, and the rest of this document is that concept wearing heavier machinery.

Plain-words recap. A trace formula says the list of frequencies at which a system rings and the list of its closed round-trip paths carry the same information, tied by an exact identity — not an analogy. You verified one in docs/02: heat on a circle, modes left, windings right.

2. The Weil explicit formula is a trace formula

docs/04-explicit-formula.md built von Mangoldt's identity — the prime staircase ψ(x) as a smooth term minus one wave per zero — and its dual: a bare sum of cosines over the zeros develops spikes at exactly u = log p^k (zeta.explicit.prime_spectrum). The rigorous symmetric statement is the Riemann–Weil (Guinand–Weil) explicit formula, phrased with a test function — the shape of the tap: probe both sides with a smooth window h and its Fourier partner g, instead of asserting an identity between spike trains.

THEOREM (the explicit formula; Riemann 1859 in outline, von Mangoldt 1895, Guinand and Weil in this symmetric form, c. 1948–1952). Let h(r) be an even test function, analytic and rapidly decaying on a strip |Im r| ≤ 1/2 + δ (fine print in Section 3), with Fourier partner g(u) = (1/2π) ∫ h(r) e^(-iru) dr. Write every non-trivial zero as ρ = 1/2 + iγ_ρ — so γ_ρ is real exactly when that zero obeys RH — and sum over all zeros with multiplicity, both signs of γ. Then, in the convention implemented by zeta/weil.py:

    Σ_ρ h(γ_ρ)  =  h(i/2) + h(-i/2)                                  [the pole of ζ at s = 1]
                 + (1/2π) ∫ h(r) [Re ψ(1/4 + ir/2) - log π] dr       [ψ = Γ'/Γ, digamma]
                 - 2 Σ_{n≥2} Λ(n) n^(-1/2) g(log n)                  [Λ as in docs/04]

Conventions differ across the literature (factors of 2π, where log π sits, one- versus two-sided sums), so the module trusts nothing unmeasured: the placement above was calibrated by computing the zero side (from cached, independently verified zeros) and the arithmetic side (Γ-factors, primes, quadrature — no zeros anywhere) separately, for test functions from three unrelated families, and requiring agreement; tests/test_weil.py re-runs the calibration. Re-measured while writing this document, for the Gaussian window h(r) = exp(-0.01 r²) at dps = 30:

    zero side  (the first 28 pairs ±γ suffice)    0.299396022507559155789163943544
    ---------------------------------------------------------------------------
    h(i/2) + h(-i/2)   [= 2e^(a/4) exactly]      +2.005006255211590
    archimedean integral                         -1.705593433712344
    prime-power sum    (n = 2,3,4,5,7 only)      -0.000016798991687
    ---------------------------------------------------------------------------
    arithmetic side, total                        0.299396022507559155789163943544

    |zero side - arithmetic side|  ≈  7e-32

Read the table as physics. The zeros are the modes: 90.6% of the spectral side is the first zero pair at γ₁ = 14.1347…, the fundamental; a decay envelope certifies everything past the 28th zero negligible at this precision. The prime powers are the closed orbits: each n = p^k enters at "time" log n = k·log p — the k-th traversal of a primitive orbit of length log p — and here 99.99998% of the prime side is the single shortest orbit, n = 2, at log 2 = 0.693. The pole and digamma lines are, in the geometric programs of docs/09-new-ontologies.md, the contribution of the place at infinity — the one place with no prime. And note the cancellation: terms of size 2 conspire to reproduce a spectral side of size 0.3 to thirty-one digits. The identity has no slack anywhere (Section 3 pushes it through eighteen orders of cancellation, and it still balances).

The instrument, briefly. zeta.weil builds admissible pairs (h, g) in closed form — gaussian_pair(a), fejer_pair(b) (whose triangle-shaped g has compact support, making the prime sum exactly finite: n ≤ 54 for b = 2), autocorrelation_pair(coeffs, …) — and explicit_formula_sides(h, g) returns both sides plus explicit bounds on every truncation; the residual must sit inside the accounted error, and the tests assert exactly that.

Plain-words recap. Probe the zeros with a smooth window and you get a number; the explicit formula computes the same number from the primes plus two "infinity" terms, exactly — measured here to thirty-one digits. Zeros play the modes; prime powers play the closed orbits, with lengths log p, 2 log p, …. It is a trace formula with the trace — the underlying operator and space — missing.

3. Weil's positivity criterion

Ask the engineer's question: what forces every mode frequency real, no growing solutions? For physical systems the answer is always some positivity: positive definite mass and stiffness matrices, a passive network, an energy that cannot go negative. Weil found the exact statement.

Let W(h) = Σ_ρ h(γ_ρ) — the spectral side, hence by Section 2 also computable from the primes with no zeros involved; that arithmetic-side evaluation is exactly zeta.weil.weil_functional(h, g). Say h is of positive type if h = |f̂|² on the real axis for some test function f — equivalently, its Fourier partner g is a self-correlation f ⋆ f~. Such h are automatically ≥ 0 on the real axis, but not off it.

THEOREM (Weil, 1952; refined 1972). RH holds if and only if W(h) ≥ 0 for every admissible h of positive type.

One direction is immediate: under RH every γ_ρ is real, where h ≥ 0. The other is the content: a zero off the line has a genuinely complex γ, so h gets evaluated off the real axis, where |f̂|² no longer means anything positive — and one can engineer an admissible h making W(h) < 0. So W, quadratic in f the way x ↦ xᵀAx is quadratic in x, is positive semidefinite exactly when RH holds: an explicitly computable infinite-dimensional "stiffness matrix", built from primes and Gamma factors, with no negative directions.

The honest fine print: the admissible class. The precise smoothness, decay, and analyticity conditions on the test functions differ between Weil's papers, Bombieri's treatments, and the textbooks (compactly supported smooth f versus analyticity-in-a-strip formulations). The equivalence is robust across the usual choices — and every pair built by zeta.weil is of positive type under any of them — but neither this document nor the module re-derives the minimal hypotheses; check a source before leaning on the exact class. Note also that a window with g supported inside (-log 2, log 2) has no prime terms at all — positivity in such restricted-support regimes has been proven in the literature (commonly cited; scope unverified here). The difficulty is uniform positivity over all admissible windows, where every prime gets a vote.

What does measurement say? positivity_probe scans families of positive-type pairs and certifies the sign of W (escalating precision until the sign clears the noise floor); near_tightness_report asks what controls the margin. Measured, and pinned in tests/test_weil.py:

Be clear what all that observed positivity is worth as evidence (docs/00-orientation.md §6): nothing. Each W(h) ≥ 0 reflects zeros already verified on the line; a true negative value would disprove RH, but no finite list of successes supports it — docs/07-equivalences-and-criteria.md §7 is blunt that the criterion admits no meaningful numerical test of RH. The measurements' value is reconnaissance — they locate where the form is tightest — and the criterion's value is structural: it names the kind of input a proof needs. Not a clever inequality; a source of positivity. In the one parallel universe where this criterion has been proven, the source was geometric: an intersection-theory inequality on a surface (docs/09 §1.3). That is why every program in this document is a hunt for geometry.

Plain-words recap. RH is equivalent to: a specific quadratic form — an infinite "stiffness matrix" written down from primes and Gamma factors — is never negative over a class of smooth test windows (a theorem, with fine print on the window class). The lab computes the form from primes alone and finds it positive but asymptotically tight, the margin controlled by the lowest zero — an observation about verified zeros, not evidence for RH. The criterion matters because it names the missing ingredient: a reason for positivity.

4. Selberg: the world where the dream works

There is a setting where all of this — trace formula, zeta function, positivity, RH — is simply true, with proofs. What exists there is a parts list for what is missing here (docs/06-hilbert-polya-and-gue.md §5 has the fuller story).

Take a compact hyperbolic surface: a finite drumhead of constant negative curvature, closed on itself, no boundary. Its Laplacian — the operator whose circle version gave Section 1's modes — is self-adjoint, the operator property that forces a real spectrum (the three-line argument of docs/06 §1); write the eigenvalues λ_n = 1/4 + r_n². The primes' role is played by the closed geodesics — closed orbits of a free particle on the surface — quantitatively so: about e^L / L primitive closed geodesics have length up to L, which is the Prime Number Theorem π(x) ~ x/log x under x = e^L. Geodesic length corresponds to log p, exactly Section 2's orbit lengths.

THEOREM (Selberg, 1956). The spectral sum over the r_n equals a geometric sum over closed geodesic lengths — structurally identical to the explicit formula. Package the geodesics into a Selberg zeta function Z(s), a product over primitive geodesics the way zeta is a product over primes; its non-trivial zeros sit at s = 1/2 ± i r_n. Self-adjointness makes every r_n with λ_n ≥ 1/4 real — so "RH" for Z is a THEOREM, up to at most finitely many exceptional eigenvalues 0 ≤ λ < 1/4, which give finitely many zeros on the real segment.

What exists there and not here? Everything arrives in the right order. Selberg starts with the space; the operator comes with it, self-adjointness included; the "primes" — the geodesic lengths — are an output of the geometry. Over the rationals the order is inverted: we possess the orbit lengths {k·log p} and the trace formula they satisfy, and no space, no operator, no reason. And one door is provably closed: by Weyl's law a hyperbolic surface has about T² modes below frequency T, while zeta has (T/2π) log T zeros below height T — no hyperbolic surface has the right density of states (docs/06 §5). Whatever is being traced in Section 2 is something stranger than a surface.

Plain-words recap. On a curved closed drumhead the whole story is a theorem: trace formula, a prime number theorem for geodesics, and "RH" — from self-adjointness, with finitely many exceptions. The enabling ingredient: the space comes first; the primes are read off it. For the integers we have the primes and no space — and a counting argument proves no ordinary surface can ever be that space.

5. Connes: noncommutative geometry and the adele class space

Status: PROGRAM. Deep theorems inside its own formalism; not a proof of RH, and not claimed to be by its author.

First, noncommutative geometry, in two paragraphs. Quantum mechanics taught physics that a system need not be a point moving in a space; it can be described entirely by its algebra of observables — position, momentum, energy as operators (infinite matrices), measurement outcomes read off their spectra. When the observables commute, this is ordinary geometry in disguise: a commutative observable algebra is secretly the functions on an honest space, which can be rebuilt from it (a classical THEOREM of operator theory, due to Gelfand and Naimark, stated here qualitatively). When they do not commute — position and momentum being the founding example — there is no underlying point-space at all, yet the algebra remains perfectly good to compute with: it still supports measure, dimension, dynamics, and crucially trace. Connes' noncommutative geometry develops exactly that: define a space to be its algebra of observables, and do geometry with algebras where points fail.

Points fail, in practice, at bad quotients: glue together the points of each orbit of a group action, and if the orbits are dense the classical quotient is mush — no continuous function survives to tell points apart — while the observable-algebra description stays rich. That is the situation with the primes. The adeles of the rationals are, in one sentence, all completions of the rationals bundled into one object: the reals plus one p-adic number system per prime — the p-adics measure a number as small when it is divisible by a high power of p (a pointer, not a course; this repo does not develop them). The adele class space is the quotient of the adeles by multiplication by nonzero rationals — a bad quotient of precisely the hopeless kind, on which every prime acts at once. It carries a natural scaling flow, the noncommutative cousin of the dilation operator from the Berry–Keating story (docs/06 §3).

THEOREM (Connes, 1999, Selecta Mathematica). In a restricted ("semilocal") setting — finitely many places at a time — a trace formula holds for this flow, and its shape is Section 2's formula: the primes appear on the geometric side, one term per place, the archimedean terms arising from the place at infinity. The zeros enter as an absorption spectrum — missing lines cut from a continuum rather than emission peaks — matching the otherwise-puzzling sign of the prime sum (docs/06 §3.3). The full statement, all places at once, is equivalent to RH: in this frame RH becomes exactly the Weil positivity of Section 3, now read as a statement about a trace on a genuinely existing (noncommutative) space.

Honest status. What the program has certainly achieved is a reformulation with structure: the explicit formula is no longer a trace formula in search of a space — a space exists, the trace is honest, and "what is being traced," "why these orbit lengths," and "where does the Gamma factor live" all get answers in one formalism (Gate 4 of docs/09-new-ontologies.md §5 passed by construction, Gate 2's slot for positivity built in). What it has not achieved is the positivity itself — which Section 3 showed is the entire remaining content of RH. Connes and Consani have continued the program (the "scaling site"; work around 2020–2021 proving positivity in restricted settings); as in docs/09 §4.2, I have not verified the current scope of those results, and neither author claims a proof. Finer details — which operator algebra, which measure, how the semilocal formula is proven — are deliberately not stated here, because I cannot vouch for them from memory; Connes' 1999 paper is the primary source.

Plain-words recap. Quantum mechanics showed a system can be its algebra of observables; noncommutative geometry does geometry with such algebras, which keep working on glued-up quotients where ordinary geometry sees mush. Connes built such a space from all completions of the rationals, found a scaling flow on it, and proved — in a restricted setting — a trace formula with the primes on the geometric side: the explicit formula, with a space attached at last. RH becomes "a certain trace is non-negative." That positivity is open: a deep reformulation, not a proof.

Where to go next