E Evidence Press

Press release · 8 September 2026 · version 0.1.0-candidate

Biased random walks on small-world networks

Candidate proof of the stated AIM mixing bounds for four fixed-weight clock conventions: logarithmic in the denser regime, square-root up to polylogarithmic factors in the sparser regime.

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

Summary

Adding shortcuts to a ring makes distant places closer, but a random walker can still take a long time to forget its starting point. This candidate asks whether a fixed preference for moving in one direction changes that.

The written proof gives logarithmic mixing in the denser shortcut regime and square-root mixing up to polylogarithmic factors in the sparser one. It covers four explicitly specified ways of measuring time, including a nonlazy discrete walk. These are fixed-weight results, not a theorem about every way of adding drift. The release is an unrefereed candidate.

Summary for specialists

Fix $\epsilon>0$ and $u>v>0$ independently of $n$. Add each undirected noncycle pair independently with probability $p$, give forward and reverse cycle edges weights $u,v$, and give shortcuts unit weight. The candidate claims, with high probability over the graph, worst-start total-variation mixing at tolerance $1/4$:

$$p=\epsilon/n:\quad t_{\rm mix}=\Theta(\log n),$$
$$p=\epsilon n^{-3/2}:\quad c\sqrt n\log n\le t_{\rm mix}\le C\sqrt n(\log n)^{18}.$$

The four clocks are the fixed-edge-rate generator $Q$, the rate-one generator $D^{-1}Q$, the lazy kernel $(I+J)/2$, and the nonlazy kernel $J=I+D^{-1}Q$, where $D_{ii}=u+v+d_i$. The invariant law is uniform for $Q$ and proportional to $D_{ii}$ for the other clocks. “Dense” means the denser regime, still with bounded average shortcut degree. The sparse logarithmic exponent is deliberately loose.

Technical account

The proof separates four obstacles that a shortcut-endpoint calculation alone cannot resolve.

  1. Random spacing: evenly spaced endpoints can retain a slow phase mode. In the sparse graph, geometric-gap conditioning and adaptive phase bounds supply coercivity across the relevant frequency windows.
  2. Physical time: crossing duration and destination are dependent. The dense proof controls their joint characteristic kernel, removes the zero-duration atom before inversion, and retains only a stationary second-moment assumption.
  3. Worst starts: a small average distance does not control exceptional vertices. A simultaneous bound for connected endpoint sets and a clock-potential estimate provide the separate upgrade.
  4. Parity: continuous-time mixing does not automatically de-lazify a walk. Signed crossing kernels and negative-side resolvents address the nonlazy obstruction.

Trace elimination, resolvent factorisation, Fourier inversion and configuration-model simplicity are established ingredients. The candidate contribution is their model-specific synthesis into the displayed physical, worst-start bounds, not a claim that those elementary identities are new.

Evidence, assurance and limitations

The seven-page entry-point paper identifies the complete dependency chain; the archive contains the detailed proof modules. Exact rational programs check finite traces, crossing moments, flux identities and resolvent factorizations. Normal, optimised and fresh-extraction replay is supplemented by three deliberately corrupted implementations that the checker rejects.

Those tests do not certify high-probability graph estimates or the asymptotic theorem. The supplied review and five-role producer-coordinated editorial assessment are documented at their actual scope. Unaffiliated specialist validation, formal verification and historical priority remain unestablished. Fixed weights, nonvanishing bias and the sparse polylogarithmic gap are substantive limitations. No performance benefit in deployed networks is measured.

Relationship to earlier work

The reversible Newman–Watts benchmark has logarithmic-squared mixing. Earlier sparse-cycle bounds, directed-cycle spectral estimates, fixed-shortcut-count theorems and average-start results answer related but different questions. The package compares their models, clocks and starting-state quantifiers explicitly. In particular, a theorem with a fixed number of shortcuts cannot simply be evaluated at a shortcut count growing with the network.

Who should care, and why

AudiencePotential useRequired caution
Probability researchersInspect physical-time and worst-start bridges for a nonreversible random graphThe complete asymptotic proof still needs unaffiliated scrutiny
Markov-chain method developersTest the reusable renewal and parity arguments in other settingsVerify every contraction, moment and clock hypothesis anew
Research software reviewersReproduce finite identities and challenge checker semanticsFinite replay is not theorem certification

Why the problem matters

Graph distance and endpoint expansion are tempting proxies for mixing, but they can miss timing, dependence and parity. This candidate makes those missing steps explicit. Its potential value lies in a checkable route from a random network to the behaviour of the actual walk, rather than a spectral calculation alone.

How to inspect or reproduce the recorded checks

Download the exact versioned ZIP and its separate replay companion. From a fresh extraction, run python3 package.py --verify, then verify_trace.py, verify_nonlazy.py and semantic_controls.py, both normally and with python3 -O. The exact checks require only the Python standard library; the producer used Python 3.14.7 and hosted CI checks Python 3.12. The optional NumPy diagnostic is floating-point exploration, not part of the exact evidence.

The most valuable next projects

First reconstruct the dimension-uniform renewal estimate, adaptive phase bound and simultaneous connected-set bound independently. Then investigate whether the sparse logarithmic gap can be reduced. Extensions to vanishing drift or different weights are new problems with new hypotheses, not already established applications.

What is in the evidence package

The release includes the formatted paper and source, complete Markdown proof modules, a cross-file notation guide, a theorem-level literature comparison, exact checkers and mutation controls, claim and assurance records, internal editorial reports and responses, licences, a manifest and a separate hash-bound fresh-replay receipt. GitHub and Zenodo carry the same declared release bytes.

Media

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

Open directions for follow-up research

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

  1. Obtain unaffiliated scrutiny of the dimension-uniform renewal, adaptive phase and connected-set estimates.
  2. Reduce the sparse logarithmic gap without changing the fixed-weight model.
  3. Investigate vanishing drift or other weights as new problems, not consequences of this theorem.
  4. Determine the reach of the renewal and clock-potential lemmas in other graph models.

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:biased-small-world-mixing
Attempt and metric receipts
  • ep-attempt:biased-small-world-mixing-assurance-publication — published / positive

    Measurement scope
    assurance-through-publication — Prospective remaining assurance and publication only. Completed discovery and initial historical notices, notation guide and consolidated-lemma revisions predate this work-ledger registration and are excluded. No discovery clock is reconstructed.
    Frozen target
    Pass five-role editorial review and deterministic checks, publish exact GitHub/Zenodo assets and complete guarded Evidence Press readback.
    Fermi active-time forecast
    150 minutes; plausible interval 120–210; expected unattended wait 30. Reference class: Evidence Press full-candidate procedural prior (n=0) — Procedural prior, not an empirical calibrated sample..
    • Package and PDF checks: 1 × 35/35/50 minutes (low/central/high) — Existing modular analytic proof and standard-library exact tests; minor review revisions and new PDF packaging.
    • Five-role editorial review: 1 × 25/30/45 minutes (low/central/high) — One bounded round.
    • Archives, page and media: 1 × 35/50/65 minutes (low/central/high) — Established scripts and authenticated services.
    • Guarded public readback: 1 × 25/35/50 minutes (low/central/high) — Composite gates and CI.
    Tractability forecast
    Within 300 active minutes: positive signal 0.8; target closure 0.8. Stop rule: Fail closed on proof, rights, CI or byte-integrity defects. Publication timing forecasts are not stopping caps.
    Observed clocks
    20 active-agent; unknown active-human; unknown substantive-compute; 0 unattended-wait; 0 blocked; 0 rework minutes. Calendar elapsed: 35 minutes.
    Research search
    Cycles: 0 positive, 0 negative, 0 inconclusive. Falsification gates: 3. Candidate architectures: 1 tested, 0 rejected.
    Agent and review load
    5 agent runs; maximum parallelism 4; 135 model turns; 16297067 deduplicated model tokens; 1 substantive review rounds; P0/P1 findings 0/0; pre-publication claim corrections 0.
    Result and calibration
    target-closed — Review clarification, five-role internal editorial acceptance, finite exact replay, byte-identical public archives and first canonical publication completed. The prior research is outside this measured attempt. No external whole-proof validation, priority or impact is claimed. Positive signal: true; target reached: true. Active-time error -130 minutes; actual/forecast 0.13333333333333333; inside interval: false. Brier score: positive signal 0.04; target closure 0.04. Variance: Active minutes are the floored union of observed model-response output spans, a strict lower bound excluding initial latency and tool execution, not a comparable total-effort or acceleration estimate. Tokens deduplicate per-response receipts across root and four editorial agents; cached context is included. Zero wait, blocked and rework minutes means no separately metered interval, not absence of CI/build waiting or repairs. Transport fallback, PDF tooling, historical text/locator clarification, forecast decomposition and audio inventory repairs are documented in the publication record. Three finite semantic mutant families were rejected. One existing proof architecture was assessed; research discovery and initial review revisions predate registration.
    Missing telemetry
    activeHumanMinutes — Human effort was not instrumented.; computeMinutes — Substantive compute time was not separately metered.
    Measurement corrections
    • measurement.agentRuns -> metrics.outcome.agentRuns — Opening 1 retained; terminal 5. Reason: Preserve opening root-only snapshot; terminal total includes four completed internal editorial agents.
Prospective work ledger · metrics policy
Intended aims
science
Artifact roles
research-output, evidence-assessment, communication
Decision object
bound — Two-regime worst-start mixing bounds under four fixed-weight clocks. Scope: Biased cycle with independent noncycle unit shortcuts and fixed positive directional bias.
Reusable methods
Structural compression (structural-compression); Exact regime stitching (regime-stitching); Adversarial scientific controls (adversarial-controls); Assurance as a vector (assurance-vector); Agent-readable research objects (agent-readable-research-object) · registry
Targeted clocks
assurance, publication
Semantic bridge
explicit — Trace expansion is bridged to physical mixing using random gaps, joint durations, a worst-start potential and parity control. Remaining risks: Written asymptotic argument awaits external scrutiny.; Finite matrix replay cannot certify graph-uniform estimates.; Priority remains bounded uncertainty..
Human judgement gates
  • Audit dimension-uniform estimates and graph conditioning.
  • Keep physical clocks and starting-state quantifiers separate.
  • Assess historical priority and significance separately.
  • Preserve rights, creator and assurance boundaries.
Next assurance action
Obtain unaffiliated proof reconstruction and separately implemented finite checks. Claim ceiling: Unrefereed candidate proof for four fixed-weight clocks. Sparse bounds differ by polylogarithmic factors. No vanishing drift, arbitrary kernel, historical priority, external specialist validation, formal verification or impact claim.
Aim-scoped impact evidence
  • science: NO_IMPACT_EVIDENCE — Inspectable small-world mixing candidate and renewal statement in Producer-coordinated mathematical publication. Design: none; comparator: No matched comparator.; estimand: No speed or impact effect estimated.. No real-world effect evidence is asserted.

Verification status

Unrefereed candidate proof for four fixed-weight clocks. Sparse bounds differ by polylogarithmic factors. No vanishing drift, arbitrary kernel, historical priority, external specialist validation, formal verification or impact claim.

Cite

Anonymous. (2026). Biased random walks on small-world networks (Version 0.1.0-candidate) [Unrefereed candidate]. Evidence Press. https://doi.org/10.5281/zenodo.22663470
BibTeX
@misc{biasedsmallworldmixing2026,
  title        = {Biased random walks on small-world networks},
  author       = {Anonymous},
  year         = {2026},
  doi          = {10.5281/zenodo.22663470},
  url          = {https://doi.org/10.5281/zenodo.22663470},
  version      = {0.1.0-candidate},
  howpublished = {Zenodo},
  note         = {Unrefereed; internally replayed evidence package. Press page: https://evidencepress.org/releases/biased-small-world-mixing/}
}

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