{"id":"059d54b7-84f7-4458-81da-5f7db87f1174","arxiv_id":"2501.14594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A multidimensional elephant random walk with stops converges, after normalization, to a Mittag-Leffler distribution times a fixed matrix, with iterated-logarithm laws and Gaussian limits in all regimes.","lead":"This paper proves how a multidimensional elephant random walk, in which the walker sometimes stays put, behaves in the long run. It shows the walk's spread is governed by a Mittag-Leffler random matrix and establishes laws of large numbers, iterated logarithms, and Gaussian fluctuations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Norm LIL in Theorem 3.2 is off by factor d: the proof sums coordinate-wise limsups after (A.20), but the compactness argument gives limsup ||S_n||^2/(2σ_n^2 loglog σ_n^2) = v^2, not d v^2.","rationale":"The most load-bearing concern is not the A.13 fourth-moment bound flagged by the reader; that bound appears repairable and likely true, since with s=bσ_n^2/n and m=(a||S_n||/n)^2 one has m ≤ s^2 ≤ s by Jensen's inequality, making the right side of (A.12) bounded by 7s. The real problem is Theorem 3.2: the proof's step from (A.20) to (A.21) uses limsup of a sum as a sum of limsups. For an isotropic Gaussian limit with covariance v^2 I_d, the norm LIL constant equals the per-coordinate constant v^2, not d v^2. This is internally inconsistent with the paper's own CLT in Theorem 3.3, which gives covariance v^2 I_d for S_n/√σ_n^2. A compactness/epsilon-net argument applied to (A.18) gives the norm LIL constant b/[d(b-2a)] = v^2, so the current theorem statement is false as written. The Gram-matrix Lemma 2.1 and the CLT parts may survive a correction, but the paper's claimed full characterization of the diffusive regime includes a false headline result, so the current version should not be accepted without substantial correction.","tokens_in":18336,"tokens_out":28565,"duration_ms":283532,"concrete_test":"Re-derive the norm LIL from (A.18) without summing coordinates: for any eps>0, choose a finite eps-net {u_j} of the unit sphere; then (1-eps)||M_n|| ≤ max_j M_n(u_j), so limsup ||M_n||^2/(2w_n loglog w_n) = b/d. Converting this to S_n yields limsup ||S_n||^2/(2σ_n^2 loglog σ_n^2) = b/[d(b-2a)] = v^2. If this calculation gives v^2 rather than d v^2, then (3.4)-(3.5) are wrong by a factor d. A numerical check in d=2, r=0.1, p below p_{d,r}, comparing the empirical limsup with v^2 and d v^2, would corroborate the analytic re-derivation but is less decisive.","verdict_should_be":"REJECT","load_bearing_attack":"Section A.2 proves, for every fixed u, limsup M_n(u)^2/(2w_n loglog w_n) = (b/d)||u||^2 a.s. (A.18). For the Euclidean norm this implies limsup ||M_n||^2/(2w_n loglog w_n) = b/d a.s. by the standard epsilon-net/compactness argument: for any eps>0 choose a finite eps-net of the unit sphere; then (1-eps)||M_n|| ≤ max_j M_n(u_j), so the scalar LIL transfers to the norm with the same constant. Converting with M_n=a_nS_n, w_n~Γ^2(a+1)/(b-2a) Σ n^{b-2a}, and σ_n^2~Σ n^b gives limsup ||S_n||^2/(2σ_n^2 loglog σ_n^2) = b/[d(b-2a)] = v^2. Instead, the paper sums the d coordinate inequalities (A.20) and obtains b/(b-2a)=d v^2. This step is invalid: for nonnegative sequences, the limsup of a sum is not the sum of the limsups, and for an isotropic d-dimensional Gaussian limit with covariance v^2 I_d the norm LIL constant is v^2, not d v^2 (d-dimensional Brownian motion has norm LIL constant 1, not d). Hence Theorems 3.2 (3.4)-(3.5) are false as stated; the correct constants should be v^2 and v^2 Σ. The same unjustified coordinate-summation pattern may affect the less explicit critical-regime proof.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the multidimensional elephant random walk with stops (MERWS), a d-dimensional process in which, at each step, a uniformly chosen past increment is repeated with probability p, changed to one of the other 2d-1 directions with probability q, or set to zero with probability r. The main result (Lemma 2.1) states that the Gram matrix of the increments, normalized by n^{1-r}, converges almost surely to (1/d) Σ I_d, where Σ has an ML(1-r) distribution. Building on this, the author proves almost sure convergence of the position in the diffusive and critical regimes, a law of the iterated logarithm, self-normalized central limit theorems, and in the superdiffusive regime almost sure convergence to a nondegenerate limit with Gaussian fluctuations around it. The proofs use multidimensional martingale techniques and generalize the author's one-dimensional results.","tokens_in":18713,"tokens_out":15634,"duration_ms":122298,"significance":"The paper is well motivated and its martingale approach is natural; Lemma 2.1 is a genuine extension of the one-dimensional trace result, and the almost sure convergence and CLT statements are plausible and largely checkable. The derivations are detailed, and the paper relies on a published, parameter-free lemma from [3] for the one-dimensional martingale convergence, which is independent of the multidimensional target and not circular. However, the laws of the iterated logarithm as stated in Theorems 3.2 and 3.5 are incorrect: the norm constants are off by a factor d. This is a substantial error in the main results, and the paper needs revision.","major_comments":[{"comment":"The norm law of the iterated logarithm is stated with the wrong constant. The proof passes from the directional LIL (A.20), which gives limsup_n ⟨u,S_n⟩^2/(2σ_n^2 log log σ_n^2) = v^2||u||^2 for each fixed u, to the norm statement (A.21) by summing the d coordinate inequalities. This step is invalid because the limsup of a sum of nonnegative sequences is not the sum of the limsups; for an isotropic vector with asymptotic covariance v^2 I_d, the correct norm LIL constant is v^2, not d v^2. The standard epsilon-net/compactness argument applied to (A.20) yields limsup_n ||S_n||^2/(2σ_n^2 log log σ_n^2) = v^2 a.s. Consequently, (3.4) should read v^2, and (3.5) should read v^2 Σ, not d v^2 Σ. The equality in (A.22) identifies b/(b-2a) with d v^2, which is the sum of the d per-coordinate constants rather than the norm constant.","section":"Theorem 3.2 and Appendix A.2 (Eqs. (A.20)-(A.21))"},{"comment":"The same factor-d error appears in the critical regime. The directional LIL in the proof gives, for every unit vector u, limsup_n ⟨u,S_n⟩^2/(2σ_n^2 log n log log n) = 2a/d a.s. (since b=2a). The compactness argument then gives the same constant for the norm: limsup_n ||S_n||^2/(2σ_n^2 log n log log n) = 2a/d a.s. Because log σ_n^2 = 2a log n (1+o(1)), this implies limsup_n ||S_n||^2/(2σ_n^2 log σ_n^2 log log log σ_n^2) = 1/d, not 1 as stated in (3.10). Similarly, the constant in (3.11) should be ((1-r)/d)Σ, not (1-r)Σ. The proof omits the conversion step and directly asserts (3.10), which is where the factor d enters.","section":"Theorem 3.5 and Appendix B.2"}],"minor_comments":[{"comment":"The fourth-moment bound (A.13) is asserted after a 'tedious but straightforward calculation' without the intermediate algebra. Since this bound is used to verify Heyde's conditions for the LIL and CLT, the authors should provide a more detailed derivation, at least in an appendix.","section":"Appendix A.2, Eq. (A.13)"},{"comment":"There are several typographical issues: the title in the header reads 'W ALK'; a reference to a monograph on Mittag-Leffler functions appears as '[ ?]' without an entry; and equation (A.12) contains a suspicious '+ -4' sign. These should be corrected.","section":"Various"},{"comment":"In the proof of Theorem 3.6, the derivation of (3.13) is left to the reader; providing the details would increase the paper's verifyability, especially given the delicate conversions used elsewhere.","section":"Appendix B.3, Eq. (3.13)"}],"recommendation":"major_revision","confidential_remarks":"The factor-d error in the LIL statements is the central technical problem. It affects two of the main theorems (Theorems 3.2 and 3.5) and is fixable by replacing the coordinate-summation step with the epsilon-net argument. The rest of the paper, including Lemma 2.1 and the CLT statements, appears sound. I recommend major revision. The importation of Lemma 4.1 from [3] is legitimate: that lemma concerns the one-dimensional trace and is a published result independent of the multidimensional target."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the central claim about the Gram matrix limit is new and probably right, but the LIL statements are wrong by a factor d, and the error is in a step that is easy to fix. The stress-test note holds up on reading.\n\nWhat's genuinely new: this is the first treatment of the multidimensional elephant random walk with stops. Lemma 2.1, the convergence of the normalized Gram matrix to (1/d)Σ I_d with Σ ~ ML(1-r), is a real contribution. The martingale framework is coherent, and the almost sure convergences and CLTs (Theorems 3.1, 3.3, 3.4, 3.6-3.8) seem sound modulo minor repairs. The proof of that Gram matrix limit is the meat of the paper, and it's convincing.\n\nThe soft spots:\n\n1. The norm LIL. Section A.2 proves for each fixed u that limsup M_n(u)^2/(2w_n loglog w_n) = (b/d)||u||^2 a.s. Then the proof sums the d coordinate inequalities (A.20) to get b/(b-2a) = d v^2 in Theorem 3.2. That is invalid: the limsup of a sum is not the sum of limsups, and for an isotropic limit the correct norm constant is v^2, not d v^2. The standard epsilon-net argument gives the right constant. The same coordinate-summation pattern infects Theorem 3.5: the constant in (3.10) should be 1/d rather than 1, and (3.11) should have (1-r)Σ/d. This is a substantial correction, not a typo.\n\n2. The inequality p_n(i) ≤ ((p+q)/n)σ_n^2 in Section 4 fails when a<0 (i.e. p<q). It is repairable by using |a|+2q, so this is a minor flaw in the written bounds, though it does appear in the run-up to the LIL/CLT proofs.\n\n3. There are unresolved citation placeholders (\"[?]\") in the Mittag-Leffler references, and the fourth-moment bound (A.13) is asserted after \"tedious calculations\" with the algebra omitted. That bound is load-bearing for the Heyde-type theorems, so the authors should display the full derivation.\n\nWho this is for: anyone working on reinforced random walks or Mittag-Leffler limit laws. The core Gram matrix lemma will be useful; the LIL statements as written are not. I would send this to a serious referee, but the current version is not acceptable as is. A referee should ask for the LIL constants to be corrected and the supporting algebra displayed. After that, it could be a solid paper.","headline":"The Gram matrix limit with the Mittag-Leffler distribution is new and likely correct, but the LIL theorems are off by a factor d from an invalid coordinate-wise sum; the a.s. and CLT results look sound, and the paper is worth refereeing after corrections.","tokens_in":19235,"tokens_out":12904,"would_cite":false,"duration_ms":100872,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G50","60G42","60F05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Pausing elephant walk obeys a Mittag-Leffler limit in any dimension.","keywords":["elephant random walk","random walk with stops","multidimensional random walk","Mittag-Leffler distribution","martingale","strong law of large numbers","law of the iterated logarithm","asymptotic normality"],"falsifier":"For $d=2$ and a diffusive choice of $p$, compute $E[\\|\\varepsilon_{n+1}\\|^4 \\mid \\mathcal{F}_n]$ directly from the displayed formulas (A.11)-(A.12) for a sequence of $n$, and test the inequality against $7b\\sigma_n^2/n$; a single violation would falsify the bound, and a Monte Carlo estimate of the ratio would provide numerical evidence either way. An algebraic derivation from (A.12) yielding a larger prefactor would also disprove the stated bound.","tokens_in":18119,"feed_emoji":"🐘","tokens_out":10861,"duration_ms":86405,"temperature":0.7,"pith_summary":"The paper extends the one-dimensional elephant random walk with stops to every dimension $d \\ge 1$, allowing the walker to stay put with probability $r$. Its central claim is that the Gram matrix of the walk's step vectors, normalized by $n^{1-r}$, converges almost surely to $\\frac{1}{d}\\Sigma I_d$, where $\\Sigma$ is a Mittag-Leffler random variable with parameter $1-r$. From that single limit, the paper derives a complete asymptotic picture: strong laws in the diffusive and critical regimes, a law of the iterated logarithm, self-normalized central limit theorems, and a superdiffusive limit with Gaussian fluctuations. A curious reader would care because the elephant random walk is a simple non-Markovian model of long-range memory, and allowing stops is the natural way to let the walker hesitate without changing the model's spirit.","feed_headline":"Pausing elephant walk obeys a Mittag-Leffler limit","feed_subtitle":"Scaling the walk's Gram matrix by n^{1-r} yields strong laws, LILs, and CLTs in every regime.","key_machinery":"The central object is the Gram matrix $\\Sigma_n = \\sum_{k=1}^n X_k X_k^{\\mathsf T}$ and its trace $\\sigma_n^2 = \\operatorname{Tr}(\\Sigma_n)$; the walk's increments are the vectors $X_k$. Two martingales carry the proof: $N_n = b_n \\sigma_n^2$, whose conditional dynamics $E[\\sigma_{n+1}^2 \\mid \\mathcal{F}_n] = (1+b/n)\\sigma_n^2$ with $b=1-r$ produce the Mittag-Leffler limit, and $M_n = a_n S_n$ with $a_n = \\Gamma(n)\\Gamma(a+1)/\\Gamma(n+a)$, $a=p-q$, whose predictable quadratic variation splits as $\\langle M\\rangle_n = \\frac{1}{d}I_d + V_n - a^2 W_n$. The Mittag-Leffler distribution $\\mathrm{ML}(\\alpha)$ is the positive random variable with Laplace transform $E_\\alpha(t)=\\sum_{n\\ge0} t^n/\\Gamma(1+n\\alpha)$; it is identified here by its moments $E[X^m]=m!/\\Gamma(1+m\\alpha)$, a characterization that the paper invokes to ensure the distribution is determined by its moments. The fourth-moment bound on the martingale increments is what lets the martingale limit theorems be applied.","core_discovery":"On the paper's own terms, the discovery is Lemma 2.1: for any $p,q \\in [0,1]$ and $r \\in (0,1)$, $\\lim_{n\\to\\infty} n^{-(1-r)} \\Sigma_n = \\frac{1}{d}\\Sigma I_d$ almost surely and in every $L^m$, where $\\Sigma$ has the Mittag-Leffler distribution with parameter $1-r$. The trace version gives $\\sigma_n^2 / n^{1-r} \\to \\Sigma$ almost surely. The paper then shows that in the diffusive regime $p < p_{d,r} = \\frac{2d+1}{4d}(1-r)$, the position satisfies $S_n/n \\to 0$, the law of the iterated logarithm $\\limsup \\|S_n\\|^2/(2\\sigma_n^2 \\log\\log\\sigma_n^2) = d v^2$, and $S_n/\\sqrt{\\sigma_n^2} \\to N(0, v^2 I_d)$; in the critical regime $p = p_{d,r}$, the same results hold with extra $\\log$ factors and covariance $\\frac{1}{d}I_d$; in the superdiffusive regime $p > p_{d,r}$, $S_n/n^{p-q} \\to L$ almost surely with explicit covariance, and $(S_n - n^{p-q}L)/\\sqrt{\\sigma_n^2} \\to N(0, \\vartheta^2 I_d)$.","pith_inferences":["Extension not in the paper: because the limiting Gram matrix is isotropic, the same limit theorems would likely hold for any initial step distribution with covariance proportional to $I_d$, not only for the uniform choice over the $2d$ directions.","Extension not in the paper: other additive functionals of the walk, such as the center of mass, should show the same Mittag-Leffler mixing under the $n^{1-r}$ normalization; checking this is a natural test of the mechanism.","Extension not in the paper: the fourth-moment bound (A.13) is asserted after 'tedious but straightforward calculations'; a reader seeking weaker hypotheses would need a different path than the martingale limit theorems used here.","Extension not in the paper: if the stay probability or the direction probabilities were made coordinate-dependent, the deterministic factor $\\frac{1}{d}I_d$ would become some anisotropic positive semidefinite matrix; the present theorem shows only the isotropic case."],"forward_implications":["The critical value $p_{d,r}=\\frac{2d+1}{4d}(1-r)$ fully separates the regimes; below it the walk is diffusive, at it the normalization picks up logarithmic factors, and above it a nondegenerate random limit $L$ exists.","Self-normalization by $\\sigma_n^2$ removes the Mittag-Leffler randomness, giving pure Gaussian limits in the diffusive and critical regimes; normalization by $n^{1-r}$ instead yields Gaussian mixtures with Mittag-Leffler variance.","The universal almost sure limit $\\sigma_n^2/n^{1-r}\\to\\Sigma$ holds for all parameters and all dimensions, making the squared displacement's random scaling law explicit.","In the superdiffusive regime, the fluctuation of $S_n$ around its random limit $L$, normalized by $\\sqrt{\\sigma_n^2}$, is asymptotically Gaussian with covariance $\\vartheta^2 I_d$."],"supporting_citations":[{"why":"Establishes the one-dimensional walk with stops that this paper extends; supplies the martingale framework and the Mittag-Leffler normalization.","marker":"[3]"},{"why":"Treats the multidimensional walk without stops, the case $r=0$ that this paper completes.","marker":"[4]"},{"why":"Provides the martingale central limit theorem and iterated-logarithm supplements used in the proofs.","marker":"[17]"},{"why":"Gives the strong law of large numbers for martingales used for the almost sure rates.","marker":"[12]"},{"why":"Supplies the martingale convergence theorems used for the $L^2$-boundedness and convergence arguments.","marker":"[16]"},{"why":"Introduces the martingale approach to the one-dimensional elephant random walk on which the multidimensional method is built.","marker":"[2]"}],"fun_headline_variants":["Elephant walk with stops: Gram matrix hits Mittag-Leffler","Multidimensional elephant walk: stops yield Mittag-Leffler law","Stopping elephant walk: almost sure limits across regimes","Elephant walk pauses, Mittag-Leffler emerges","In every regime, elephant walk converges à la Mittag-Leffler"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the conditional fourth-moment bound $E[\\|\\varepsilon_{n+1}\\|^4 \\mid \\mathcal{F}_n] \\le 7b\\sigma_n^2/n$ in Eq. (A.13), which the paper states as the outcome of tedious but straightforward calculations without displaying the full algebra; if that bound fails, the Lindeberg-type conditions and the convergence of the auxiliary martingales used in the limit theorems would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Elephant walk with stops: Gram matrix hits Mittag-Leffler","Multidimensional elephant walk: stops yield Mittag-Leffler law","Stopping elephant walk: almost sure limits across regimes","Elephant walk pauses, Mittag-Leffler emerges","In every regime, elephant walk converges à la Mittag-Leffler"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1407,"prompt_tokens":1014,"completion_tokens":393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":305}},"tokens_in":630,"tokens_out":393,"duration_ms":4575,"temperature":1.0,"reasoning_tokens":305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:00:55.899599+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For $d=2$ and a diffusive choice of $p$, compute $E[\\|\\varepsilon_{n+1}\\|^4 \\mid \\mathcal{F}_n]$ directly from the displayed formulas (A.11)-(A.12) for a sequence of $n$, and test the inequality against $7b\\sigma_n^2/n$; a single violation would falsify the bound, and a Monte Carlo estimate of the ratio would provide numerical evidence either way. An algebraic derivation from (A.12) yielding a larger prefactor would also disprove the stated bound.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the one-dimensional walk with stops that this paper extends; supplies the martingale framework and the Mittag-Leffler normalization."},{"cited_title":"Bercu and L","cited_arxiv_id":null,"evidence_quote":"Treats the multidimensional walk without stops, the case $r=0$ that this paper completes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the martingale central limit theorem and iterated-logarithm supplements used in the proofs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the strong law of large numbers for martingales used for the almost sure rates."},{"cited_title":"Hall and C","cited_arxiv_id":null,"evidence_quote":"Supplies the martingale convergence theorems used for the $L^2$-boundedness and convergence arguments."}],"review_version":1}