Skip to content

The forward S-matrix reconciliation#

A reconciliation, not a result. It gives this project a vocabulary and a lineage that predate it by decades. It adds no measurement — every number that follows is either a physics identity checked on a synthetic instance, or a repeat of a verdict measured and reported elsewhere on this site.

The sentence#

The S-matrix is the unitary matrix connecting sets of asymptotically free particle states — the in-states and the out-states — with no account of the path between them.

Wheeler introduced the object in 1937 ("On the Mathematical Description of Light Nuclei by the Method of Resonating Group Structure," Phys. Rev. 52, 1107–1122); Heisenberg made it the organizing object of particle theory through the 1940s ("Die beobachtbaren Größen in der Theorie der Elementarteilchen," Z. Phys. 120 (1943), 513–538, with a third installment in 1944 on its analytic structure).

That is the whole claim, in one line: where it ends up, from where it started, with no account of the path. It was written down as a physics object nine decades before this project asked the same question of an attention matrix.

The mathematics is the same object, not a metaphor#

The Lippmann–Schwinger equation defines the scattering operator T from a free Hamiltonian H0, a perturbation V, and an energy E approached from the causal side:

T = V + V G0 T                      Lippmann–Schwinger
G0 = (E − H0 ± iε)⁻¹                the free resolvent

Solved in closed form, T = (I − V G0)⁻¹ V. Expanded in powers of V G0 instead of inverted, the same equation is the Born series:

T = V + V G0 V + V G0 V G0 V + …  =  Σ_{k=0}^{∞} (V G0)^k V

The Born series is the Neumann series of the resolvent — the identical object this project calls the hop expansion, (I − gP)⁻¹ = I + gP + (gP)² + …. Both are geometric series of an operator, and both carry the same convergence law:

Born series converges   iff   ρ(G0 V) < 1        (spectral radius of the free resolvent times the perturbation)
hop expansion converges iff   γ·ρ(P) < 1          (this project's own condition, ceqjepa/operator.py)

Two names for one condition: a bounded operator composed with itself, geometrically, converges exactly when its spectral radius sits under 1. Unitarity of the full S-matrix is Born-series mass conservation — probability is neither created nor destroyed by the expansion, term by term. The ± iε prescription is what selects the causal (retarded) resolvent over the anti-causal one; it is the limiting absorption principle, and it is the same margin this project already carries in its own E − H0 + iε construction of G0 and in the resolvent read's own boundary convention.

Worked instance#

Verified on this box rather than asserted. n = 12, complex128, H0 diagonal with entries drawn uniform on [1, 3], V a symmetric real matrix with entries N(0, 0.1²), E = 4 + 0.05i (the +iε margin folded into E's own imaginary part). Script: born_series_check.py, python born_series_check.py, seed 0.

quantity value
ρ(G0 V) 0.2596
Born-series error, V + (VG0)V + … truncated at 5 applications of VG0, ‖·‖F 1.8×10⁻⁴
truncated at 10 applications 2.1×10⁻⁷
truncated at 20 applications 2.8×10⁻¹³

ρ(G0 V) < 1 holds, and the error against the exact T = (I − VG0)⁻¹V falls geometrically, roughly six orders of magnitude every ten hops — the same rate the spectral radius predicts.

The honesty that comes with it#

The first instance run for this reconciliation put E inside the band spanned by H0 (E = 2 + 0.05i against H0 ∈ [1, 3], same seed). ρ(G0 V) = 1.6688. The series diverges — the truncation error grows, not shrinks, with more hops: 3.5, 25, 4.1×10³ at 5, 10, 20 applications, same seed 0, same script. That run is reported here corrected and unscored, not quietly replaced with a convergent E. It is the same standard the rest of this project holds itself to: a bar whose branch never fires is worth nothing, and a divergent instance that gets moved off the page without a note is the same failure in a physics costume.

What this is, and is not#

This reconciliation gives the project a vocabulary and a lineage — the hop expansion is not a new numerical trick, it is the Born series of a resolvent physics has used since 1937, and its convergence condition is not a heuristic, it is a spectral-radius theorem. That is the entire content of the claim.

It adds no measurement. Nothing here changes any number reported elsewhere on this site — not the operator-arm result, not the chess-prediction result, not the open question of why the gate closes. A reader who comes away believing physics validated this project's results, or that the project derived something new about physics, has been misled, and that costs more credibility than the framing buys.

Sentences that are not permitted on this page or any other page of this site, because none of them is true of what was built:

  • Any sentence describing the read in terms of waves — there is no wave equation here, only a resolvent and a spectral condition.
  • Any sentence about inverse scattering — recovering V from S is a different, harder problem this project does not touch. The read runs forward only: parameters and structure in, an outcome distribution out.
  • Any sentence claiming the operator is an S-matrix, rather than that reading it this way has that shape.
  • Any sentence suggesting physics validates this project's results. It does not, and cannot; the two are checked by entirely different evidence.

The one sentence the page is allowed to make, and no larger:

The read is a forward S-matrix — in-state to out-state through the resolvent; unitarity is the Born mass; the +iε is the causality margin the project already carries.

The verdicts this reconciliation does not change#

These are measured elsewhere on this site (docs/PHASE_H.md, docs/PHASE_I.md, docs/PHASE_I1.md); the physics framing above does not add to or subtract from any of them.

  • The operator represents order. Handed S5's 120-state automaton as bare integer symbol ids — no permutation matrices, no transition table, verified by the source not reaching any arm — the operator arm scores 0.8620 ± 0.0556 against a commuting-diagonal control at 0.2860 ± 0.0150, matched at 404 parameters by numel(). An independent re-run put the operator ahead on 5 of 5 seeds pairwise with no overlap, and the control is saturated within 0.031 of its own multiset ceiling of 0.3110. Scope: one parameter budget, one word length, no length-generalization check.
  • The operator does not predict. On the repaired chess bed, recalibrated resolution is 0.001469 against an oracle ceiling of 0.10117 — 1.45% of what the bed offers — and it is tied on that resolution by an eight-bin histogram of total piece count computed from the operator's own 769-float input, losing on Brier score, 0.6334 against 0.5020.
  • The gate closes and nobody knows why. Two proposed mechanisms — gradient starvation and step-0 zero density — are refuted by the same table built to demonstrate them; a third account (exactness as an eval-only property) and a fourth (a live gradient unfreezing a closed gate) were each pre-registered and killed in the three most recent commits to this repository. The cause remains open.
  • Twelve checks that could not fail have been catalogued (docs/PHASE_I1.md), and each could not fail for a structural reason — a null space, a vacuous control, a grep with no true-positive path — rather than because training had settled anything. The standing practice this catalogue set: audit a table's fixed structure before trusting its trainable part.

Diagrams#

Each diagram below is meant to stand alone: read its caption first, then the picture.

The operator family as a cube softmax attention β=1 · g off · qk on unnormalized kernel β=0 · g off · qk on exact path product — β=0 · qk off · any g β: 0 → 1 (total → mean) qk: off → on g: off → on
Shows: the three switches (β, g, qk) as the axes of a cube; softmax attention and the unnormalized kernel are two corners (green, blue), the exact path product is the marked edge (amber). Does not show: which corner performs better at anything — this is a statement about what one family contains, proved in Lean (three_corners_containment), not a ranking.
Forward only: in-state to out-state through the resolvent in⟩ asymptotically free, before T = (I − V G0)⁻¹ V the resolvent, causal branch (the +iε margin) out⟩ asymptotically free, after no arrow runs right to left — this is not an inverse-scattering read
Shows: the S-matrix picture this project's resolvent read has the shape of — an in-state mapped to an out-state by the operator T, with no path drawn between them. Does not show: a wave, a scattering trajectory, or a reverse (out→in) inference — the read this project ships is forward-only, and inverse scattering is not attempted.
Born series = Neumann series of the resolvent = the hop expansion T ≈ V + VG0·V + VG0·VG0·V + … (k hops = k factors of VG0) converges iff ρ(G0 V) < 1 — same law as γ·ρ(P) < 1 measured, seed 0, n=12: ρ(G0V)=0.2596 (converges) 1.0 (divergence boundary) error at 5 / 10 / 20 hops: 1.8e-4 / 2.1e-7 / 2.8e-13
Shows: the hop expansion written as a finite sum of increasing powers of V·G0 applied to V, with the convergence condition ρ(G0V) < 1 stated on the figure and satisfied by the worked instance (bar at 0.2596, well under the 1.0 boundary). Does not show: the divergent instance (ρ = 1.6688) — that failure mode is reported separately, corrected and unscored, in the text above.
G_ij = ∏ m_k along the path — one zero severs every span that crosses it x₀ x₁ x₂ x₃ x₄ x₅ m₁=0.9 m₂=0 m₃=0.8 m₄=0.95 m₅=0.7 G₅,₀ = m₁·m₂·m₃·m₄·m₅ = 0 exactly — every path crossing gate 2 is severed, not merely small
Shows: the gate path product across six tokens; one factor at exactly zero (gate 2) sends every span crossing it — including the full x₀→x₅ path — to exactly zero, not asymptotically small. Does not show: a logit-space decay gate (the additive kind used elsewhere in the literature); Lean's no_prefix_scan_represents_a_zero_gate proves no such gate can do this, because exp(Cᵢ−Cⱼ) is never zero.
The scoreboard: measured, refuted, open Represents order — S5, bare integers, matched params 0.8620 vs control 0.2860 · control saturated at 0.031/0.3110 · scope: 1 budget, no length-gen Does not predict — repaired chess bed resolution 0.001469 / ceiling 0.10117 (1.45%) · tied by an 8-bin histogram · Brier 0.6334 vs 0.5020 loss Gate closes, cause unknown — two accounts refuted by their own table gradient starvation ✗ · step-0 zero density ✗ · eval-only exactness ✗ · live-gradient unfreeze ✗ Twelve checks catalogued that could not fail, each for a structural reason null space · vacuous control · grep with no true-positive path · standing practice: audit structure first This reconciliation (Wheeler 1937 / Lippmann–Schwinger) changes none of the four rows above.
Shows: the four load-bearing verdicts of the project as they stand today — one measured and holding within its stated scope, one measured and lost, one open with its refuted explanations named, one a catalogue of checks that could not fail. Does not show: a trend or a forecast — each row is a single measurement with the command and file that produced it named in the linked pages, not an average or a projection.

Sources#

  • Wheeler, J. A. "On the Mathematical Description of Light Nuclei by the Method of Resonating Group Structure." Physical Review 52, no. 11 (1937): 1107–1122. DOI: 10.1103/PhysRev.52.1107.
  • Heisenberg, W. "Die beobachtbaren Größen in der Theorie der Elementarteilchen." Zeitschrift für Physik 120 (1943): 513–538; continued 120 (1943): 673–702; analytic-structure installment, Zeitschrift für Physik 123 (1944): 93–112.
  • Lippmann, B. A., and J. Schwinger. "Variational Principles for Scattering Processes. I." Physical Review 79, no. 3 (1950): 469–480.
  • This project's own resolvent read: ceqjepa/operator.py, convergence condition γ·ρ(P) < 1.
  • Worked-instance script: born_series_check.py (this session's scratchpad; not part of the shipped package — a verification script, not a claim of a new numerical method).