{"id":"1439fbb9-6b5e-44d4-87bc-a5435270647b","arxiv_id":"1908.08606","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For the switch random walk, exceptional times of transience exist almost surely, the exceptional set has Hausdorff dimension 1/2, and the positivity event is maximally noise sensitive.","lead":"This paper shows that a one-dimensional random walk whose direction is controlled by fair coin flips, the switch walk, has rare exceptional times when it escapes to infinity, even though it is recurrent at every fixed time. Such times form a fractal set of Hausdorff dimension 1/2, and the event that the walk stays positive is maximally sensitive to small random perturbations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's weakest assumption points to the independence of the Poisson clocks, which is explicit in the model definition and is used correctly throughout Sections 3, 4, and 6. I focused on the most potentially fragile step, the FKG inequality in Proposition 15, because it supplies the squared factor that drives the second-moment estimate leading to Proposition 8. Reading the events carefully, the inequality is valid: for a fixed W and period length, the event {B_i in (0,2W)} is the intersection of an increasing and a decreasing event in the increments, and the symmetry about W makes the two factors equal, yielding the squared conditional probability. The transfer step in Theorem 2, Part 3, correctly handles the discrepancy between I_K and n, and the influence computations in Proposition 12, while intricate, are consistent. The only issues I found are notational ambiguities and a small edge-case question about k=0 in the E0 union; neither changes the verdict. The paper's central claims are supported by the arguments as written.","tokens_in":86,"tokens_out":33911,"duration_ms":1260816,"concrete_test":"Independently re-derive the conditional FKG bound in Proposition 15 for one even period: fix W > 0 and J, write B_i = W + S_i with S a simple random walk, and verify directly that P(S_i in (-W,W) for all i <= J) is at most P(S_i > -W for all i <= J)^2. If this inequality fails for some W and J, the lower-bound half of Theorem 1 needs repair.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After a full read, I cannot identify a load-bearing concern. The central claims rest on the explicit rate-1 Poisson clocks: period boundaries are independent of the walk values and across coordinates, and the resulting mirroring of increments on even periods is valid because the switch-walk increments are i.i.d. fair. The most delicate point, the FKG step in Proposition 15, is sound when A'_j is read as the stated event {B_i in (0,2W)}: the two events are increasing and decreasing in the underlying increments, and symmetry about W gives exactly the squared factor. The influence estimates in Proposition 12 and the transfer argument in Theorem 2 also check out. There are minor presentation ambiguities, such as the W+B_i spelling of A'_2j and whether k=0 is included in the union for E0, but these are easily repaired and do not threaten the main results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a dynamical version of the 'switch random walk' in one dimension, in which the increments are products of independent fair bits, and each bit is rerandomized by its own rate-1 Poisson clock. It proves two contrast results relative to the usual 'compass' dynamical random walk. Theorem 1 states that almost surely there exist times t in [0,1] at which the switch walk tends to infinity, that the set E_alpha of times with liminf Z_n(t)/n^alpha > 0 has Hausdorff dimension 1/2 almost surely for alpha in [0,1/2), and that E_alpha is empty almost surely for alpha > 1/2. Theorem 2 states that the events {Z_n > 0} are maximally noise sensitive: for any sequence epsilon_n with n epsilon_n -> infinity, the events {Z_n(0) > 0} and {Z_n(epsilon_n) > 0} decorrelate. The proof of Theorem 2 uses a period decomposition into odd and even blocks, on which the two walks agree or mirror each other. Theorem 1 is reduced to a lower-bound estimate (Proposition 8) proved through the same period decomposition together with FKG-type arguments, and an upper-bound influence estimate (Proposition 12) proved by ballot and random-walk estimates.","tokens_in":26447,"tokens_out":9974,"duration_ms":92269,"significance":"If correct, the paper establishes a clean one-dimensional example where recurrence is dynamically sensitive, contrasting with the dynamical stability of recurrence for the compass walk proved by Benjamini, Haggstrom, Peres and Steif, and with the noise stability of the positivity event for the compass walk. The lower-bound argument in Section 6 is genuinely different from the spectral-sample and randomized-algorithm methods used in related dynamical percolation results, and it is elementary and self-contained. The paper also proves a sharp maximal noise sensitivity statement, explicitly noting that the condition n epsilon_n -> infinity is best possible. The results are pure theorems with no fitted parameters, and the proof infrastructure is well organized; the use of external tools such as the local central limit theorem, the reflection principle, the FKG inequality, and the Schramm-Steif results is clearly identified.","major_comments":[],"minor_comments":[{"comment":"The final term X_{I_K(epsilon_n)} in the definition of U_n is not self-explanatory; a short sentence explaining that only the first increment of the next odd period is included because the comparison stops at I_K(epsilon_n) would improve readability.","section":"Section 4, definition of U_n"},{"comment":"There is a notational inconsistency: A'_j(t) is defined using a walk B^{(j)}(t) started from W_{I_{j-1}(t)-1}(t), but in the proof of Proposition 15 the event A'_{2j}(t) is written as {W_{I_{2j-1}(t)-1}(t) + B^{(2j)}_i(t) in (0,2W_{...})}, which suggests that B^{(2j)} is started at 0. The authors should unify the two spellings, for example by stating explicitly that in the second form B^{(2j)} is a walk started at 0 and W + B is the walk started from W.","section":"Section 6, definition of A'_j(t) and proof of Proposition 15"},{"comment":"In the displayed formula for I_m(P_n), the second maximum is written as max_{i <= m-n+1} Z_i >= 2z; the index should be i <= n-m+1, as in the definition of U in equation (17).","section":"Section 7, equation (16)"},{"comment":"The displayed inequality involving P(L^alpha_n(1)>0) is correct but unnecessarily confusing because the same probability appears on both sides; rewriting it as P(A) <= E[L^alpha_n(2)] / E[L^alpha_n(2) | A] with A = {L^alpha_n(1)>0} would make the argument easier to follow.","section":"Section 5.3, equation (7)"},{"comment":"The Chernoff bound in the proof of Lemma 14 is applied to show decay of P(E^odd_n(t)^c), and the sentence 'P(E^even_n(t)) = P(E^odd_n(t))' should read 'P(E^even_n(t)^c) = P(E^odd_n(t)^c)'; as written, the equality is between probabilities of different events.","section":"Section 6, proof of Lemma 14"},{"comment":"The proof that E_alpha is empty for alpha > 1/2 first treats alpha in (1/2,1) and then extends to alpha >= 1 by monotonicity; it may be worth stating explicitly that the event E_alpha is decreasing in alpha, even though this is immediate from the definition.","section":"Section 5.3"}],"recommendation":"accept","confidential_remarks":"I found the central arguments sound and the paper a strong contribution to the dynamical sensitivity literature. The remaining issues are purely cosmetic and can be addressed during production. I recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core news: this is a legitimate, well-executed proof paper, not a repackaged Warren example. The switch walk (coin-turning/bootstrap walk) was known, and Warren had given a noise-sensitive representation, but the exceptional-times theorem and the sharp Hausdorff dimension 1/2 are new. The main results are clean: the switch walk has almost surely a nonempty set of times of transience, with Hausdorff dimension 1/2 for the natural family E_alpha, and the positivity event is maximally noise sensitive. The paper does the job properly: it proves both theorems from the Poisson clocks, with the period decomposition as the key mechanism, and it isolates the technical work in separate sections.\n\nI checked the load-bearing steps. The period boundaries are independent of the walk values and independent across coordinates, and the mirroring of increments on even periods is valid. The FKG step in Proposition 15 is sound when A'_j is read as {B_i in (0,2W)}: the two events are increasing and decreasing, and symmetry about W gives exactly the squared factor. The influence estimates in Proposition 12 and the transfer argument in Theorem 2 also check out. External tools are standard and used correctly. No fitted parameters, no circularity.\n\nSoft spots are minor. Some estimates in Sections 6 and 7 are intricate, especially the lower-bound influence computation in Proposition 12, and I would want an independent check before signing off. There are small presentation ambiguities: the spelling of A'_2j as W+B_i and whether k=0 is included in the union for E0. Both are easily repaired. The proof of Lemma 9 is a little compressed; the claim that the bad times are contained in the set of rerandomisation times is right, but it could be spelled out more.\n\nWho this is for: anyone working on dynamical random walks, noise sensitivity, or exceptional times. It is a useful benchmark and a sharp contrast to the compass walk, where positivity is noise stable and recurrence is dynamically stable. The citation pattern is solid and the novelty is real. This deserves serious refereeing, not a desk reject. I would accept after minor revision.","headline":"A rigorous and genuinely new proof that the switch walk has exceptional times of transience of Hausdorff dimension 1/2 and is maximally noise sensitive; deserves a serious referee.","tokens_in":26954,"tokens_out":1609,"would_cite":true,"duration_ms":18548,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G50","60J10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a switch random walk, times at which the walk escapes to infinity exist almost surely and form a set of Hausdorff dimension 1/2.","keywords":["dynamical random walk","switch random walk","noise sensitivity","exceptional times","Hausdorff dimension","transience","recurrence","simple symmetric random walk"],"falsifier":"Simulate or compute the coupled switch walk with noise level $\\varepsilon_n=n^{-1/2}$; Theorem 2 predicts $P(Z_n(0)>0\\text{ and }Z_n(\\varepsilon_n)>0)\\to1/4$. If the joint-positivity frequency instead stays near $1/2$ for large $n$, the maximal noise-sensitivity claim is false; similarly, an estimate of the Hausdorff dimension of $E_\\alpha$ different from $1/2$ would falsify Theorem 1.","tokens_in":26131,"feed_emoji":"🎲","tokens_out":10609,"duration_ms":93969,"temperature":0.7,"pith_summary":"The paper compares two random walks built from the same fair coin flips: the compass walk $Y_n=\\sum_{j=1}^n X_j$, which keeps a fixed sense of direction, and the switch walk $Z_n=\\sum_{k=1}^n\\prod_{j=1}^k X_j$, which changes direction whenever the running product flips. Although the two have identical distributions, in the dynamical version where each bit rerandomises at rate 1 they behave very differently. The paper proves that the switch walk almost surely has exceptional times at which $Z_n(t)\\to\\infty$, that the set of such times has Hausdorff dimension $1/2$, and that the event $\\{Z_n>0\\}$ is maximally noise sensitive: it decorrelates from itself under any noise level $\\varepsilon_n$ with $n\\varepsilon_n\\to\\infty$. The compass walk has no such exceptional times and its positivity event is noise stable, so the result shows that dynamical and noise sensitivity are properties of how a walk is encoded in the randomness, not just of its marginal law.","feed_headline":"Switch walk escapes at exceptional times of dimension 1/2","feed_subtitle":"Same law as an ordinary random walk, yet recurrence is fragile and positivity is maximally noise sensitive.","key_machinery":"The load-bearing object is the period decomposition of the two coupled walks relative to the change times $I_k(t)=\\min\\{i>I_{k-1}(t):X_i(t)\\ne X_i(0)\\}$. On odd periods the increments of $Z(0)$ and $Z(t)$ coincide; on even periods they are mirror images, so writing $U_n$ and $V_n$ for the sums over odd and even periods gives $Z_n(0)=U_n+V_n$ and $Z_n(t)=U_n-V_n$. Joint positivity of the two walks becomes the event $U_n>|V_n|$, whose probability tends to $1/4$ once there are many periods by step $n$. For the Hausdorff dimension lower bound the same decomposition is combined with the FKG inequality and the half-sum process $W_i(t)=(Z_i(0)+Z_i(t))/2$, which is constant on even periods, to obtain $P(P_n(0)\\cap P_n(t))\\lesssim 1/(nt^{1/2})$; the upper bound is carried by the influence estimate $I_m(P_n)\\asymp (n-m+1)n^{-3/2}$ for the positivity event $P_n$, converted to a dimension bound through a standard energy criterion for exceptional times.","core_discovery":"For the dynamical switch walk $Z_n(t)=\\sum_{k=1}^n\\prod_{j=1}^k X_j(t)$, with each bit $X_j$ rerandomised by an independent rate-1 Poisson clock, the paper establishes three claims. First, the set $E=\\{t\\in[0,1]:Z_n(t)\\to\\infty\\}$ is almost surely non-empty, and for every $\\alpha\\in[0,1/2)$ the set $E_\\alpha=\\{t:\\liminf_n Z_n(t)/n^\\alpha>0\\}$ has Hausdorff dimension $1/2$ almost surely, while $E_\\alpha$ is empty for $\\alpha>1/2$. Second, the events $\\{\\{Z_n>0\\},n\\ge1\\}$ are maximally noise sensitive: for every $\\varepsilon_n\\in(0,1)$ with $n\\varepsilon_n\\to\\infty$, $P(Z_n(0)>0\\text{ and }Z_n(\\varepsilon_n)>0)-P(Z_n(0)>0)^2\\to0$. Third, these statements are in direct contrast to the compass walk $Y_n(t)$, for which recurrence is dynamically stable and the positivity events are noise stable.","pith_inferences":["Because the compass and switch walks have identical laws as sequences of random variables, the paper implies that noise sensitivity is representation-dependent: the same Boolean-function distribution can be encoded so that positivity is noise stable or maximally noise sensitive.","The optimality of $n\\varepsilon_n\\to\\infty$ suggests a quantitative crossover: when $n\\varepsilon_n$ is bounded, some of the first $n$ bits never rerandomise, so the correlation should remain bounded away from zero; quantifying this crossover could connect to Fourier-weight or influence calculations for the switch walk.","The period-decomposition method is not tied to positivity: the same odd/even mirroring gives explicit covariance asymptotics for other additive functionals such as the maximum or the range of the walk, and those would be testable extensions.","Whether $E_{1/2}$ is empty remains open; the natural next step is the same second-moment argument with sharper estimates near the $t^{-1/2}$ singularity, which is exactly the borderline separating the $\\alpha<1/2$ and $\\alpha>1/2$ regimes."],"forward_implications":["Recurrence of the switch walk is dynamically sensitive: with probability one there are times $t$ at which $Z_n(t)\\to\\infty$, so one-dimensional recurrence is not automatically dynamically stable.","The exceptional times form a genuine fractal: for every $\\alpha<1/2$, the times with $\\liminf_n Z_n(t)/n^\\alpha>0$ have Hausdorff dimension exactly $1/2$, while no such times exist for $\\alpha>1/2$.","The positivity events are maximally noise sensitive: any noise level $\\varepsilon_n$ with $n\\varepsilon_n\\to\\infty$ destroys the correlation between $\\{Z_n(0)>0\\}$ and $\\{Z_n(\\varepsilon_n)>0\\}$, and the condition on $\\varepsilon_n$ is optimal.","The law of the iterated logarithm is dynamically sensitive for the switch walk: there almost surely exist times at which $Z_n(t)$ is negative for all large $n$, a phenomenon that cannot occur for the compass walk."],"supporting_citations":[{"why":"Proves recurrence of the compass walk is dynamically stable, providing the contrast that the switch walk's exceptional times overturn.","marker":"[3]"},{"why":"Establishes that the event $\\{Y_n>0\\}$ for the compass walk is noise stable, the baseline against which Theorem 2's maximal noise sensitivity is measured.","marker":"[4]"},{"why":"Supplies the dynamical-percolation framework and the exceptional-times lemma used to move from positive probability to almost-sure dimension statements.","marker":"[13]"},{"why":"Provides the energy/measure criterion and the influence criterion used for the lower and upper Hausdorff dimension bounds.","marker":"[19]"},{"why":"The FKG inequality is applied to compare events on even periods, producing the squared probabilities in Proposition 15.","marker":"[10]"},{"why":"The ballot theorem gives the starting-point decomposition for the influence $I_m(P_n)$ in Proposition 12.","marker":"[2]"},{"why":"The local central limit theorem supplies the order-$n^{-1/2}$ transition probabilities used throughout Lemmas 3, 13, and Corollary 19.","marker":"[17]"},{"why":"Gives the probability that a walk stays above a sublinear boundary, used to show $P(P_n^\\alpha)$ is of order $n^{-1/2}$.","marker":"[18]"}],"fun_headline_variants":["Exceptional times: 1D switch walk escapes to infinity with dimension 1/2","Maximal noise sensitivity for 1D switch walk positivity events","Exceptional transience: Hausdorff dimension 1/2 for 1D switch walk","1D switch walk: exceptional escapes and maximal noise sensitivity","Exceptional times of transience: Hausdorff dimension 1/2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that each bit $X_j$ rerandomises according to an independent rate-1 Poisson clock, so the times at which bits change are independent of the walk's values and of one another; if updates were synchronised or depended on the walk's position, the mirroring of even periods, and with it both main theorems, would fail.","fun_headline_variants_meta":{"raw":{"variants":["Exceptional times: 1D switch walk escapes to infinity with dimension 1/2","Maximal noise sensitivity for 1D switch walk positivity events","Exceptional transience: Hausdorff dimension 1/2 for 1D switch walk","1D switch walk: exceptional escapes and maximal noise sensitivity","Exceptional times of transience: Hausdorff dimension 1/2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001558,"raw_usage":{"total_tokens":6222,"prompt_tokens":942,"completion_tokens":5280,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":5179}},"tokens_in":558,"tokens_out":5280,"duration_ms":32999,"temperature":1.0,"reasoning_tokens":5179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:35:00.623567+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or compute the coupled switch walk with noise level $\\varepsilon_n=n^{-1/2}$; Theorem 2 predicts $P(Z_n(0)>0\\text{ and }Z_n(\\varepsilon_n)>0)\\to1/4$. If the joint-positivity frequency instead stays near $1/2$ for large $n$, the maximal noise-sensitivity claim is false; similarly, an estimate of the Hausdorff dimension of $E_\\alpha$ different from $1/2$ would falsify Theorem 1.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves recurrence of the compass walk is dynamically stable, providing the contrast that the switch walk's exceptional times overturn."},{"cited_title":"Noise sensitivity of Bo olean functions and applications to percolation","cited_arxiv_id":null,"evidence_quote":"Establishes that the event $\\{Y_n>0\\}$ for the compass walk is noise stable, the baseline against which Theorem 2's maximal noise sensitivity is measured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dynamical-percolation framework and the exceptional-times lemma used to move from positive probability to almost-sure dimension statements."},{"cited_title":"Schramm and J.E","cited_arxiv_id":null,"evidence_quote":"Provides the energy/measure criterion and the influence criterion used for the lower and upper Hausdorff dimension bounds."},{"cited_title":"Fortuin, P.W","cited_arxiv_id":null,"evidence_quote":"The FKG inequality is applied to compare events on even periods, producing the squared probabilities in Proposition 15."},{"cited_title":"Solution directe du probleme r´ esolu par M","cited_arxiv_id":null,"evidence_quote":"The ballot theorem gives the starting-point decomposition for the influence $I_m(P_n)$ in Proposition 12."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the probability that a walk stays above a sublinear boundary, used to show $P(P_n^\\alpha)$ is of order $n^{-1/2}$."}],"review_version":1}