Spaces:
Running
Running
| """Claim 1 / Lemma 3.3 -- Route B: exact counterexample + the paper's own geometry. | |
| Three independent pieces of evidence, none of which uses symbolic algebra: | |
| (A) An explicit 4-atom landscape in EXACT RATIONAL ARITHMETIC that satisfies every | |
| audited clause of Assumption 2.1 and on which m_t strictly INCREASES, converging | |
| to 1 while a_t -> 0. No rounding is involved anywhere, so the refutation of | |
| "there exists rho_eps in (0,1) with m_{t+1} <= rho_eps m_t for all t" is exact. | |
| (B) A sweep over the counterexample family confirming numerically the boundary | |
| delta = log 2 that Route A derived symbolically, from the other side. | |
| (C) The paper's OWN synthetic reward geometry (Appendix C.4: r_i(x) = -||x-mu_i||^2) | |
| on a fine continuous grid, measuring sup_{x not in S_eps} W_t(x) per iteration -- | |
| i.e. measuring whether appendix Assumption B.4 actually holds in the setting the | |
| authors themselves experiment in. | |
| """ | |
| from __future__ import annotations | |
| import math | |
| from fractions import Fraction as F | |
| import numpy as np | |
| from repro.lib import report | |
| from repro.lib.exact import ExactLandscape | |
| from repro.lib.landscape import quadratic_grid_landscape, uniform_init | |
| from repro.lib.verdict import FALSIFIED, Verdict | |
| def _counterexample() -> tuple[ExactLandscape, dict[str, F], F]: | |
| """The 4-atom exact counterexample. | |
| eps = log(10/9) (so eps_ratio = 9/10 and S_{i,eps} = {w_i >= 0.9 w_i*}). | |
| A: w1 = 1, w2 = 1/100 -> S_1 only | |
| B: w1 = 1/100, w2 = 1 -> S_2 only | |
| C: w1 = w2 = 4/5 -> outside (0.8 < 0.9), but only just | |
| D: w1 = w2 = 1/1000 -> outside, deeply | |
| C is the "compromise" atom: not eps-optimal for either reward, yet good enough for | |
| both that the balanced multiplier lifts it above 1 while the basins hold little mass. | |
| Assumption 2.1(iii) permits this because it only requires SOME delta > 0. | |
| """ | |
| w1 = {"A": F(1), "B": F(1, 100), "C": F(4, 5), "D": F(1, 1000)} | |
| w2 = {"A": F(1, 100), "B": F(1), "C": F(4, 5), "D": F(1, 1000)} | |
| lam = ExactLandscape(w1, w2, eps_ratio=F(9, 10), name="exact-4atom-counterexample") | |
| p0 = {"A": F(1, 1000), "B": F(1, 1000), "C": F(998, 1000), "D": F(0)} | |
| return lam, p0, F(1, 2) | |
| def run(params: dict) -> Verdict: | |
| out = report.artifact_dir("claim1", "exact_and_continuous") | |
| steps = int(params.get("steps", 60)) | |
| v = Verdict( | |
| claim_id="claim1/lemma-3.3-outside-decay", | |
| title="Lemma 3.3: geometric decay of mass outside the eps-optimal basins", | |
| status=FALSIFIED, | |
| statement=( | |
| "Under Assumption 2.1 (i,iii) there exists rho_eps in (0,1) with " | |
| "m_{t+1} <= rho_eps m_t for all t, hence m_t -> 0 and a_t + b_t -> 1." | |
| ), | |
| ) | |
| # ================================================================= (A) == # | |
| report.banner("(A) Exact rational counterexample: audit Assumption 2.1 clause by clause") | |
| lam, p0, q = _counterexample() | |
| audit = lam.audit(p0) | |
| for key in ( | |
| "A1_i_bounded_rewards", "A1_ii_basins_disjoint", "A1_ii_nontrivial_init", | |
| "A1_iii_outside_gap", "A1_iii_cross_gaps", | |
| ): | |
| report.kv(key, f"ok={audit[key]['ok']} " + str({ | |
| k: val for k, val in audit[key].items() if k != "ok" | |
| })[:150]) | |
| all_ok = all(audit[k]["ok"] for k in audit if isinstance(audit[k], dict)) | |
| report.kv("eps", f"{audit['eps']:.6f} (= log(10/9), exact ratio {audit['eps_ratio_exact']})") | |
| report.kv("Delta_1, Delta_2", f"{audit['A1_iii_cross_gaps']['Delta1']:.6f}, " | |
| f"{audit['A1_iii_cross_gaps']['Delta2']:.6f} (= log 100)") | |
| report.kv("outside gap delta", f"{audit['A1_iii_outside_gap']['delta']:.6f} (= log(5/4))") | |
| v.add( | |
| "the counterexample satisfies EVERY audited clause of Assumption 2.1", | |
| all_ok, | |
| f"eps={audit['eps']:.6f}, delta={audit['A1_iii_outside_gap']['delta']:.6f}>0, " | |
| f"Delta_1=Delta_2={audit['A1_iii_cross_gaps']['Delta1']:.6f}>0, basins disjoint, " | |
| f"p_0(S_1)={audit['A1_ii_nontrivial_init']['p0_S1']}>0, " | |
| f"p_0(S_2)={audit['A1_ii_nontrivial_init']['p0_S2']}>0", | |
| audit=audit, | |
| ) | |
| # the counterexample must NOT satisfy the appendix's extra hypothesis B.4 | |
| report.banner("(A) Run the exact dynamics") | |
| rows = lam.run(p0, q, steps) | |
| report.write_csv(out / "exact_counterexample_trace.csv", rows) | |
| for r in rows[:6] + rows[-2:]: | |
| report.kv(f"t={r['t']:<3d}", f"m_t={r['m_t']:.12f} m_(t+1)/m_t={r['m_next_over_m_t']:.10f}" | |
| f" a_t={r['a_t']:.3e} supW_out={r['sup_W_outside']:.6f}" | |
| f" W(S1)={r['W_basin1']:.6f}") | |
| increased_every_step = all(r["m_increased"] for r in rows) | |
| max_ratio = max(r["m_next_over_m_t"] for r in rows) | |
| mass_ok = all(r["mass_conserved_exact"] for r in rows) | |
| report.kv("m_t increased at EVERY step (exact)", increased_every_step) | |
| report.kv("max_t m_(t+1)/m_t (exact arithmetic)", f"{max_ratio:.10f}") | |
| report.kv("exact ratio at t=0", rows[0]["m_ratio_exact"]) | |
| report.kv("m_0 -> m_final", f"{rows[0]['m_t']:.10f} -> {rows[-1]['m_t']:.12f}") | |
| report.kv("a_0 -> a_final", f"{rows[0]['a_t']:.3e} -> {rows[-1]['a_t']:.3e}") | |
| v.add( | |
| "total mass is exactly conserved at every step (the update is a valid density)", | |
| mass_ok, "sum_x p_{t+1}(x) == 1 as an exact rational identity at every step", | |
| ) | |
| v.add( | |
| "NO rho_eps in (0,1) exists: m_{t+1} > m_t at every step, in exact arithmetic", | |
| increased_every_step and max_ratio > 1.0, | |
| f"m_(t+1)/m_t = {rows[0]['m_ratio_exact']} = {max_ratio:.10f} > 1 at t=0 and m_t " | |
| f"rises monotonically {rows[0]['m_t']:.6f} -> {rows[-1]['m_t']:.12f}. Since rho_eps " | |
| "is existentially quantified before the universal over t, one such step refutes " | |
| "the statement; here every step does.", | |
| max_ratio=max_ratio, exact_ratio_t0=rows[0]["m_ratio_exact"], | |
| ) | |
| v.add( | |
| "the conclusion inverts: m_t -> 1 and a_t + b_t -> 0, not the reverse", | |
| rows[-1]["m_t"] > rows[0]["m_t"] and rows[-1]["a_t"] < rows[0]["a_t"], | |
| f"a_t falls {rows[0]['a_t']:.3e} -> {rows[-1]['a_t']:.3e} while m_t rises to " | |
| f"{rows[-1]['m_t']:.12f}: all mass ends up OUTSIDE the eps-optimal basins, on the " | |
| "compromise atom C. The lemma asserts exactly the opposite.", | |
| ) | |
| # ---- negative control: repair the landscape, decay returns ------------- # | |
| report.banner("(A) Negative control: deepen the outside atom and the lemma's conclusion returns") | |
| # Move C from w = 4/5 down to w = 1/50, which makes delta large enough that | |
| # 2 e^{-delta} < 1 + e^{-Delta}, i.e. leaves the counterexample family. | |
| w1c = dict(lam.w1); w2c = dict(lam.w2) | |
| w1c["C"] = F(1, 50); w2c["C"] = F(1, 50) | |
| lam_ctrl = ExactLandscape(w1c, w2c, eps_ratio=F(9, 10), name="control-deep-outside") | |
| rows_ctrl = lam_ctrl.run(p0, q, steps) | |
| report.write_csv(out / "exact_control_trace.csv", rows_ctrl) | |
| ctrl_ratios = [r["m_next_over_m_t"] for r in rows_ctrl] | |
| ctrl_decays = all(r["m_increased"] is False for r in rows_ctrl) | |
| report.kv("control: max m_(t+1)/m_t", f"{max(ctrl_ratios):.8f}") | |
| report.kv("control: m_0 -> m_final", f"{rows_ctrl[0]['m_t']:.8f} -> {rows_ctrl[-1]['m_t']:.3e}") | |
| v.add_control( | |
| "control landscape (deep outside atom) DOES decay geometrically as the lemma says", | |
| ctrl_decays, | |
| f"changing only w(C) from 4/5 to 1/50 -- leaving every other atom and p_0 " | |
| f"untouched -- restores contraction: max_t m_(t+1)/m_t = {max(ctrl_ratios):.8f} < 1 " | |
| f"and m_t falls to {rows_ctrl[-1]['m_t']:.3e}. So the failure is caused by the " | |
| "shallow outside gap that Assumption 2.1 permits, not by a coding error in the " | |
| "update: the same code produces the lemma's behaviour on a conforming landscape.", | |
| max_ratio=max(ctrl_ratios), | |
| ) | |
| # ================================================================= (B) == # | |
| report.banner("(B) Sweep the counterexample family; confirm the delta = log 2 boundary") | |
| rows_fam = [] | |
| for delta in (0.05, 0.1, 0.2, 0.4, 0.6, 0.69, 0.70, 0.8, 1.0, 1.5): | |
| for Delta in (3.0, 5.0, 10.0): | |
| # exact rationals approximating e^{-delta}, e^{-Delta} to 12 digits | |
| wC = F(round(math.exp(-delta) * 10**12), 10**12) | |
| wB = F(round(math.exp(-Delta) * 10**12), 10**12) | |
| if not (0 < wC < F(9, 10)): | |
| continue # C must be strictly outside both basins | |
| l = ExactLandscape({"A": F(1), "B": wB, "C": wC}, | |
| {"A": wB, "B": F(1), "C": wC}, F(9, 10)) | |
| p = {"A": F(1, 1000), "B": F(1, 1000), "C": F(998, 1000)} | |
| r0 = l.run(p, F(1, 2), 1)[0] | |
| predicted = 2 * math.exp(-delta) > 1 + math.exp(-Delta) | |
| rows_fam.append({ | |
| "delta": delta, "Delta": Delta, | |
| "m_increased_observed": r0["m_increased"], | |
| "predicted_by_2exp_condition": predicted, | |
| "agree": r0["m_increased"] == predicted, | |
| "m_ratio": r0["m_next_over_m_t"], | |
| }) | |
| report.write_csv(out / "counterexample_family_sweep.csv", rows_fam) | |
| n_agree = sum(r["agree"] for r in rows_fam) | |
| for r in rows_fam: | |
| if r["delta"] in (0.6, 0.69, 0.70, 0.8): | |
| report.kv(f"delta={r['delta']:<5g} Delta={r['Delta']:<5g}", | |
| f"m increased={r['m_increased_observed']} predicted={r['predicted_by_2exp_condition']}" | |
| f" ratio={r['m_ratio']:.8f}") | |
| v.add( | |
| "the closed-form condition 2e^{-delta} > 1 + e^{-Delta} predicts the observed " | |
| "failures exactly across the swept family", | |
| n_agree == len(rows_fam), | |
| f"{n_agree}/{len(rows_fam)} parameter combinations agree with the condition Route A " | |
| f"derived symbolically. The boundary sits between delta=0.69 and delta=0.70, " | |
| f"bracketing log 2 = {math.log(2):.6f}. Two independent routes, same boundary.", | |
| n_agree=n_agree, n_total=len(rows_fam), | |
| ) | |
| # ================================================================= (C) == # | |
| report.banner("(C) The paper's own geometry (Appendix C.4): does Assumption B.4 hold there?") | |
| rows_cont = [] | |
| for d in (2.0, 4.0, 6.0, 8.0): | |
| for eps in (0.1, 0.5): | |
| lam_c = quadratic_grid_landscape(d=d, eps=eps, n=4001, half_width=12.0) | |
| p = uniform_init(lam_c) # the paper's C.9.2 initialisation | |
| res = lam_c.run(p, q=0.5, steps=50) | |
| tr = res["trace"] | |
| sup_W = [r["rho_outside_sup_W"] for r in tr] | |
| ms = np.array([r["m_t"] for r in tr]) | |
| ratios = ms[1:] / np.maximum(ms[:-1], 1e-300) | |
| rows_cont.append({ | |
| "d": d, "eps": eps, | |
| "max_sup_W_outside": float(max(sup_W)), | |
| "B4_holds_uniformly": bool(max(sup_W) < 1.0), | |
| "max_m_ratio": float(ratios.max()), | |
| "uniform_rho_lt_1_exists": bool(ratios.max() < 1.0), | |
| "m_0": float(ms[0]), "m_50": float(ms[-1]), | |
| "m_decays_asymptotically": bool(ms[-1] < ms[0]), | |
| }) | |
| report.kv(f"d={d:<4g} eps={eps:<4g}", | |
| f"max_t sup_(x outside) W_t = {max(sup_W):.4f} " | |
| f"B.4 holds={max(sup_W) < 1.0} max_t m_(t+1)/m_t={ratios.max():.4f} " | |
| f"m: {ms[0]:.4f} -> {ms[-1]:.3e}") | |
| report.write_csv(out / "paper_geometry_outside_domination.csv", rows_cont) | |
| any_b4_violated = any(not r["B4_holds_uniformly"] for r in rows_cont) | |
| all_decay = all(r["m_decays_asymptotically"] for r in rows_cont) | |
| v.add( | |
| "in the paper's own quadratic reward geometry, appendix Assumption B.4 is " | |
| "violated at some iterations, yet m_t still decays asymptotically", | |
| any_b4_violated and all_decay, | |
| "sup_{x outside} W_t(x) exceeds 1 for early t in the configurations marked above " | |
| "(the points just beyond the basin boundary have r_i barely below r_i*-eps, so the " | |
| "balanced multiplier lifts them), which means B.4 -- the hypothesis the appendix " | |
| "proof needs -- does not hold uniformly in t even in the authors' own setting. " | |
| "The CONCLUSION m_t -> 0 nonetheless holds there, because mass does concentrate at " | |
| "the two optima. So the lemma's conclusion is right in the paper's experiments " | |
| "while the stated route to it is not available.", | |
| rows=rows_cont, | |
| ) | |
| v.numbers = { | |
| "exact_m_ratio_t0": rows[0]["m_ratio_exact"], | |
| "exact_m_ratio_t0_float": max_ratio, | |
| "m_final_counterexample": rows[-1]["m_t"], | |
| "a_final_counterexample": rows[-1]["a_t"], | |
| "control_max_ratio": max(ctrl_ratios), | |
| "family_sweep_agreement": f"{n_agree}/{len(rows_fam)}", | |
| } | |
| v.limitations = [ | |
| "The counterexample is a finite (4-atom) landscape. Assumption 2.1 places no " | |
| "cardinality or continuity restriction on X, and a finite X with a counting base " | |
| "measure is a legitimate instance; part (C) additionally exercises continuous " | |
| "landscapes on a 4001-cell grid.", | |
| "Part (C) measures a hypothesis (B.4) on a discretised continuum; the reported " | |
| "sup is over grid cells, so it is a lower bound on the true supremum -- which " | |
| "only strengthens the finding that it exceeds 1.", | |
| ] | |
| v.deviations = [ | |
| "FALSIFIED refers to the main-text statement quoted in the contract, which cites " | |
| "Assumption 2.1(i,iii). The appendix statement (Lemma B.4, which adds Assumption " | |
| "B.4 as a hypothesis) is sound and is separately verified in Route A.", | |
| ] | |
| v.artifacts = [str(p) for p in sorted(out.rglob("*")) if p.is_file()] | |
| return v | |