E Evidence Press

Press release · 5 September 2026 · version 0.1.0-candidate

Linear Jordan-block growth in non-backtracking graphs

A constructive proof candidate shows that non-backtracking Jordan blocks can grow linearly with graph order, even under one fixed maximum-degree bound.

Listen to this briefingNarrated summary · OpenAI API synthetic voice (fable) · MP3 · download

Summary

A walk on a network is non-backtracking if it never immediately reverses its last edge. The matrix counting these walks can have Jordan blocks: chains of generalized eigenvectors that are invisible in a list of eigenvalues alone. How large can those blocks become?

This unrefereed candidate supplies a constructive answer: their largest possible size grows linearly with the number of graph vertices. The result persists under one fixed maximum-degree bound. It concerns extremal counting matrices, not typical networks or normalized random-walk transition matrices.

Summary for specialists

Let $f(n)$ be the maximum complex Jordan-block size of $B_G$ over finite simple undirected $n$-vertex graphs without degree-one vertices, with $(B_G)_{(u,v),(x,y)}=1$ exactly when $v=x$ and $u\ne y$. For every $n\ge3$,

$$\max\left\{1,\left\lfloor\frac{n-1}{131072}\right\rfloor-3\right\} \le f(n)\le2n.$$

Thus $f(n)=\Theta(n)$. If maximum degree is at most $393217$, the lower bound is still $\max\{1,\lfloor n/12714016\rfloor\}$. Long blocks occur at $\lambda=\pm512i$. Lower-bound graphs are connected and have minimum degree two. The constants are not claimed optimal. This addresses AIM Problem 1.3(1), not its separate Alon–Boppana question in part (2).

Technical account

The upper bound uses the quadratic non-backtracking determinant identity and a contraction argument excluding nontrivial unit-modulus Jordan blocks. For the lower bound, a banded nilpotent matrix yields a quadratic pencil whose value at $2i$ has a one-dimensional kernel. A separate determinant-valuation argument forces growing algebraic multiplicity. Nilpotence at one point alone would not justify that conclusion.

A four-dimensional rational representation of $\sqrt3$, followed by a fixed integer scaling, produces integer adjacency weights and prescribed degree action. Signed fibers and zero-valued vertices realize both actions in a simple unweighted graph. An incidence intertwiner preserves the chain away from $\pm1$. Padding on zero-valued vertices reaches every larger order; private zero vertices give the fixed maximum-degree strengthening.

Evidence, assurance and limitations

The arbitrary-parameter theorem rests on the written proof. Exact code checks finite pencils, three fully enumerated small graph embeddings, published defective examples, padding and deliberately corrupted inputs. New normal and optimized replays cover even $r=4$ through $12$; a separate historical pencil receipt reaches $r=20$. These are producer-side controls, not universal proof certificates or independent reproduction.

The first full shared-zero graph already has 34,314,518,528 edges. It is represented by an edge-generating rule, not materialized. Internal AI editorial review and the supplied review are not authenticated external specialist review. Formal verification, exhaustive novelty and historical priority remain open. No application benefit, typical-network frequency or optimal constant is claimed.

Relationship to earlier work

Glover and Kempton supply the standard reduced-matrix framework. Heysse, Lorenzen and Reinhart provide defective examples and chain-preserving graph constructions: preserving a fixed chain is not itself a length amplifier. Takata and colleagues construct high-order exceptional points for linear Hamiltonians. The present graph-constrained quadratic pencil needs its own valuation and realization argument. These distinctions identify the mechanism being offered, not a certified historical first.

Who should care, and why

AudiencePotential useRequired caution
Spectral graph theoristsInspect a proposed resolution of the Jordan-growth question.The argument remains unrefereed and constants are not claimed optimal.
Matrix and pencil researchersReuse the valuation and signed-fiber construction.Preserve dimensions, degree constraints and chain-transfer assumptions.
Research agents and tool buildersReplay exact finite controls and trace dependencies.A passing implementation is not independent theorem validation.

Why the problem matters

Undirected adjacency matrices are symmetric and diagonalizable, but counting non-backtracking matrices need not be. The candidate shows that graph structure does not impose a universal small bound on their Jordan complexity. Even bounded maximum degree does not recover the regular-graph bound of two. This is a structural conclusion, not a measured improvement to network algorithms.

How to inspect or reproduce the recorded checks

Download the versioned evidence ZIP, inspect its manifest, and follow its README. With the pinned SymPy and optional performance backend installed, run python verify_research.py --max-r 12 and python -O verify_research.py --max-r 12. Both should report PASS_PRODUCER_EXACT_CHECKS. Start with the manuscript's dimension table and source note before interpreting the nullity receipts. Do not enumerate the enormous full-family graph.

The most valuable next projects

The first priority is unaffiliated reconstruction of the all-parameter proof, especially the valuation and degree-action embedding. Further work could reduce the constants, determine exact extremal sizes, or find the smallest degree bound supporting linear growth. Wider prior-art reconciliation and formalization are separate assurance projects. None is supplied by publishing this release.

What is in the evidence package

The package contains the eight-page PDF and Markdown proof, exact construction and verification code, structured claims, a dated AIM source bridge, bounded prior-art audit, review response, historical and new receipts, a complete manifest, provenance and component licences. Original prose and data are CC0; original code is MIT. Third-party source documents and supplied review files are referenced or hashed, not silently relicensed. The archive, public repository and replay routes are linked in the standard resource panel.

Media

The audio briefing is provided in the header above. Download the MP3 briefing · read the transcript.

Open directions for follow-up research

Also available in machine-readable form for research agents and follow-up projects.

  1. Obtain unaffiliated reconstruction of the valuation, encoding and graph lift.
  2. Improve constants and determine exact extremal sizes or the smallest useful degree bound.
  3. Complete broader prior-art and priority assessment.
  4. Investigate formal verification of the arbitrary-parameter argument; AIM part (2) remains separate.

Research process, metrics and reusable methods

Prospective process metadata under the Evidence Press operating model and research-metrics policy. It records the intended handoff, measured scope and claim boundary; it is not evidence that the method accelerated this work.

Work ID
ep-work:nonbacktracking-jordan-growth
Attempt and metric receipts
  • ep-attempt:nonbacktracking-jordan-growth-assurance-publication — published / positive

    Measurement scope
    assurance-through-publication — Prospective manuscript and telemetry repairs, internal editorial gate, immutable archiving and canonical publication. Intake and the earlier completed research are excluded; historical research receipts remain separate and are not reconstructed.
    Frozen target
    Publish the review-repaired anonymous unrefereed Jordan-growth candidate with exact replay, internal editorial disposition, immutable GitHub and Zenodo assets and canonical Evidence Press readback, or preserve the exact blocking gate.
    Fermi active-time forecast
    150 minutes; plausible interval 120–210; expected unattended wait 30. Reference class: Full-candidate route prior (n=0) — Procedural benchmark, not a matched empirical estimate..
    • Review repairs and source audit: 1 × 25/30/45 minutes (low/central/high) — Minor revisions without a supplied mathematical gap.
    • Deterministic package and five-role editorial gate: 1 × 35/45/60 minutes (low/central/high) — One frozen submission, existing exact controls.
    • Immutable archive, reader-first page and media: 1 × 35/45/60 minutes (low/central/high) — Existing generators and authenticated GitHub and Zenodo.
    • Two seal, CI, deployment and readback cycles: 1 × 25/30/45 minutes (low/central/high) — Normal new-slug workflow.
    Tractability forecast
    Within 240 active minutes: positive signal 0.95; target closure 0.85. Stop rule: Fail closed on unresolved integrity or authority gates. Consolidate at 75 minutes; pause at the workflow's 240-minute hard stop rather than weakening a gate.
    Observed clocks
    1 active-agent; unknown active-human; unknown substantive-compute; 0 unattended-wait; 0 blocked; 0 rework minutes. Calendar elapsed: 41 minutes.
    Research search
    Cycles: 0 positive, 0 negative, 0 inconclusive. Falsification gates: 6. Candidate architectures: 0 tested, 0 rejected.
    Agent and review load
    6 agent runs; maximum parallelism 4; 6 model turns; unknown deduplicated model tokens; 1 substantive review rounds; P0/P1 findings 0/0; pre-publication claim corrections 0.
    Result and calibration
    target-closed — Supplied review actioned without mathematical claim changes; five internal editorial roles and exact replay passed. Six existing scientific negative-control classes were rerun, not six new research cycles. Immutable GitHub and Zenodo bytes agree and the first canonical release is live. Earlier discovery is excluded. Final paired replay used 18.6973869 wall seconds and 16.338528 CPU seconds; this is partial process telemetry, not total compute. Active-agent time is a deliberately incomplete lower bound: one whole minute of directly supervised coordinator work in the clock-tool-bounded 18:38:52Z to 18:40:44Z recovery/readback window (112 seconds, rounded down). Earlier generation, other agents, compaction and uninstrumented waits are not reconstructed. Zero unattended and blocked minutes refer only to that observed window, not the complete attempt. Forecast error and ratio therefore compare an incomplete lower bound, not total effort; they must not be used to infer acceleration or update a class prior. Positive signal: true; target reached: true. Active-time error -149 minutes; actual/forecast 0.006666666666666667; inside interval: false. Brier score: positive signal 0.0025; target closure 0.0225. Variance: Active-agent time is a deliberately incomplete lower bound: one whole minute of directly supervised coordinator work in the clock-tool-bounded 18:38:52Z to 18:40:44Z recovery/readback window (112 seconds, rounded down). Earlier generation, other agents, compaction and uninstrumented waits are not reconstructed. Zero unattended and blocked minutes refer only to that observed window, not the complete attempt. Forecast error and ratio therefore compare an incomplete lower bound, not total effort; they must not be used to infer acceleration or update a class prior.
    Missing telemetry
    activeHumanMinutes — No task-local human-work timer.; computeMinutes — Paired replay process receipts are retained, but complete substantive-computation elapsed coverage is unavailable; no reconstructed total.; deduplicatedModelTokens — No supported fork-aware task-local counter supplied; rollout token events were not summed.; uncachedInputTokens — No supported task-local uncached-input counter supplied.
Prospective work ledger · metrics policy
Intended aims
science
Artifact roles
research-output, evidence-assessment, method-demonstration, communication
Decision object
bound — Two-sided linear Jordan-block growth bounds with explicit graph generation and proof bridges. Scope: Counting non-backtracking operators of finite simple undirected leafless graphs, n>=3.
Reusable methods
Structural compression (structural-compression); Adversarial scientific controls (adversarial-controls); Assurance as a vector (assurance-vector); Agent-readable research objects (agent-readable-research-object) · registry
Targeted clocks
assurance, publication, translation
Semantic bridge
explicit — Banded valuation forces a long pencil chain; uniform rational encoding preserves it; signed fibers realize the adjacency and degree actions; incidence transfer and zero-valued padding preserve the chain in a simple graph. Remaining risks: The universal proof is not formally verified or externally reconstructed.; Finite checks do not establish exact all-parameter block sizes.; Priority and real-world relevance are not established..
Human judgement gates
  • Audit the source counting convention and all quantifiers.
  • Reconstruct valuation, rational encoding and graph realization.
  • Preserve third-party rights and prior attribution.
  • Keep internal review and publication separate from external assurance.
Next assurance action
Seek unaffiliated mathematical reconstruction and wider source reconciliation before stronger assurance claims. Claim ceiling: Unrefereed constructive theorem candidate for part (1), with producer controls; no part (2), exact extremum, optimal constants, priority, external validation or impact established.
Aim-scoped impact evidence
  • science: NO_IMPACT_EVIDENCE — More inspectable spectral graph constructions in Producer-coordinated research and publication. Design: none; comparator: No matched conventional-workflow comparator.; estimand: No causal acceleration, reliability, uptake or scientific-impact effect estimated.. No real-world effect evidence is asserted.
Parent handoffs
  • reuses-method https://doi.org/10.1016/j.laa.2021.01.022 — inherited claim: Reduced non-backtracking matrix and quadratic-pencil framework.; inherited ceiling: Standard antecedent framework, not independent confirmation of the present growth construction.

Verification status

Anonymous AI-assisted unrefereed theorem candidate. Five producer-coordinated editorial roles completed one round; the sole required Minor exposition repair is closed and the final internal decision is Accept/PASS_WITH_NOTES. No external validation, formal verification, exhaustive novelty or historical priority is asserted.

Cite

Anonymous. (2026). Linear Jordan-block growth in non-backtracking graphs (Version 0.1.0-candidate) [Unrefereed theorem candidate and evidence package]. Evidence Press. https://doi.org/10.5281/zenodo.22375005
BibTeX
@misc{linearnonbacktrackingjordangrowth2026,
  title        = {Linear Jordan-block growth in non-backtracking graphs},
  author       = {Anonymous},
  year         = {2026},
  doi          = {10.5281/zenodo.22375005},
  url          = {https://doi.org/10.5281/zenodo.22375005},
  version      = {0.1.0-candidate},
  howpublished = {Zenodo},
  note         = {Unrefereed; internally replayed evidence package. Press page: https://evidencepress.org/releases/linear-nonbacktracking-jordan-growth/}
}

Also: cite.bib · paper.json · this page as Markdown