{"id":"62cbbce4-4d65-41fa-aa87-cac954750d12","arxiv_id":"2412.04284","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For i.i.d. uniform random directions on the sphere, the greedy sign choice has a unique invariant radial measure with mean sqrt(pi)/2 * Gamma((d+1)/2)/Gamma(d/2) ~ sqrt(pi d/8).","lead":"A greedy walker that always jumps toward the origin stays, on average, at distance about sqrt(pi d / 8) when the jump directions are random in d dimensions. The paper proves this and also describes deterministic jump rules that produce periodic polygon patterns.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 depends on an unproved Harris-recurrence step in §2.2; the drift and block-growth claims are plausible but not established, leaving the invariant measure and mean formula without a rigorous foundation.","rationale":"The reader's weakest-assumption identification is correct in targeting §2.2, but the specific objection that 2·(1/100)·√d exceeds the per-step +1 increase in r² misplaces the denominator: X=|γ1|/√(γ1²+χ²), so under the stated event the negative drift term is O(1/√d), not O(√d). The real problem is that the section is a sketch: the growth event is not quantified and the drift is not proved via a Lyapunov function. Since these gaps are fillable and the exact drift computation substantiates the claim, the verdict should remain CONDITIONAL rather than REJECT. The deterministic Lemma 6 omission is also real, but it is not load-bearing for Theorem 1, the paper's main result.","tokens_in":21429,"tokens_out":19154,"duration_ms":187307,"concrete_test":"Verify the two quantitative conditions that would make §2.2 rigorous: (i) with E[X]=Γ(d/2)/(√π Γ((d+1)/2)), check that sup_{r≥√d+1/2}(1−2rE[X])<0 for d=2,…,100 (simple symbolic/numeric evaluation); (ii) compute a uniform lower bound on the Gaussian event of §2.2 and show that under it r_{n+k}²−r_n² ≥ d+1/4 for k=⌈2d⌉, using r_{n+1}²−r_n² ≥ 1−2·(1/100)·r/√(d−√d) for 0≤r≤√d. If both hold, the Harris-recurrence proof is completable and Theorem 1 stands; if either fails for some d, the existence of π_d with the claimed mean is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, Theorem 1, is conditional on Section 2.2's assertion that the radial chain is a positively recurrent Harris chain with minorization on A=[√d, √d+1/2]. That assertion is not proved. The block-growth step says that under the event |γ1|≤1/100 and |χ²(d−1)−d|≤√d for ~2d steps, the chain grows past √d, citing only the inequality r_{n+1}≤(r_n²+1)^{1/2}. That inequality alone does not imply growth past √d; what is needed is a lower bound on r²-increments, which in fact holds (r_{n+1}²−r_n² = 1−2X r_n ≥ 1−2·(1/100)·√d/√(d−√d) ≈ 0.98), but this computation is absent. More seriously, the drift back from r≥√d+1/2 is handled by asymptotic normal approximations for large d and a vague 'replace √d by 10000√d' remark for small d; this does not cover the compact interval [√d+1/2, 10000√d], and no Foster–Lyapunov function is exhibited. A standard argument with V(r)=r² would give E[r_{n+1}²−r_n²|r]=1−2E[X]r, which is strictly negative on the whole tail for every d≥2, but the paper never makes this computation. Because §2.7's mean formula and the exponential-moment bound both presuppose an existing invariant probability π_d, the incompleteness of §2.2 is load-bearing: as written, Theorem 1 is not verified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a greedy sequence in R^d: given x_{n-1} and a unit vector v_n, one chooses x_n = x_{n-1} ± v_n to minimize ||x_n||. For i.i.d. uniform v_n on S^{d-1}, the main result (Theorem 1) claims that the radial process r_n = ||x_n|| has a unique invariant probability measure π_d on [0,∞), with mean (√π/2) Γ((d+1)/2)/Γ(d/2) ~ √(πd/8), and that π_d has finite exponential moments of every order. The proof proceeds by a Harris chain argument for existence (Section 2.2), a Lyapunov function for exponential moments (Sections 2.3–2.4), and a stationarity computation for the mean (Section 2.7). The paper also analyzes a deterministic planar example using the van der Corput sequence, proving that the process reaches a fixed neighborhood of the origin in O(|z|) steps, that initial points in B(0,1) are periodic starts in base 2, and that periodic starts exist or not depending on the parity of the base.","tokens_in":21736,"tokens_out":6795,"duration_ms":65037,"significance":"If Theorem 1 is established, it provides a clean and surprising answer for a natural random greedy process: the radial chain self-stabilizes at distance ~√(πd/8), with sub-Gaussian tails. The mean computation is self-contained and elegant, and the Lyapunov argument for exponential moments is a nice application of Meyn–Tweedie theory. The deterministic sections contain several original elementary results and are interesting in their own right. However, the existence of the invariant measure is not rigorously proved: the Harris recurrence verification in Section 2.2 is sketched and contains gaps that are load-bearing for the whole probabilistic program. As written, the central probabilistic claim is conditional on an unproved positive-recurrence statement, so the paper requires a substantial revision before the main theorem can be accepted.","major_comments":[{"comment":"The argument that starting from r ≤ √d one reaches A in about 2d steps is not justified. The paper derives the upper bound ||x_{n+1}|| ≤ (||x_n||^2+1)^{1/2} and concludes that after ~2d steps the sequence will surpass √d. An upper bound cannot imply growth; what is needed is a lower bound on the squared increment ||x_{n+1}||^2 − ||x_n||^2 = 1 − 2X||x_n|| under the stated event. The paper never writes this lower bound, and without it the hitting-time argument fails. Note also that the event |γ1|≤1/100 alone does not control X = |γ1|/(γ1^2+χ^2(d−1))^{1/2}; the normalization by the chi-squared denominator is essential and is absent from the text.","section":"§2.2, block-growth into A"},{"comment":"The proof that the chain drifts back from the region ||x_n|| > √d + 1/2 is not complete. The displayed inequality E[||x_{n+1}||^2 − ||x_n||^2] ≤ 1 − 2E[|γ1|/(γ1^2+χ^2)^{1/2}]√d is not a negative-drift bound for the full tail, since for r just above √d+1/2 the term 2E[X]r is close to the critical value. The paper splits into 'large dimensions' using χ²(d−1)=d±O(√d) and 'small dimensions' by replacing √d with 10000√d, but this does not cover the entire interval [√d+1/2, 10000√d] uniformly for all d≥2. No Foster–Lyapunov function is exhibited, and finite expected hitting time of A from every starting point is not established.","section":"§2.2, drift from large radius"},{"comment":"The minorization condition on A = [√d, √d+1/2] is asserted rather than verified. The paper states that the transition density is 'uniformly bounded away from 0 in any strict subinterval' and proposes ρ as a multiple of Lebesgue measure on (√d−1/4, √d). However, it does not prove a uniform lower bound over all x ∈ A for the probability of landing in that interval. Since the proposed ρ is supported outside A, one must show that from every x ∈ A the chain can reach that interval with probability bounded below; this is plausible from the density P_d(x,y) in Lemma 3, but the calculation is not provided.","section":"§2.2, minorization condition"},{"comment":"The mean formula E[||x||] = 1/(2E[X]) and the exponential-moment estimate both presuppose the existence of a unique invariant probability π_d with finite second moment. Because the Harris-chain verification in §2.2 is incomplete, the foundation of Theorem 1 is missing. The mean computation itself is formally correct given such a stationary distribution, but the existence of that distribution is the load-bearing unsupported step.","section":"Theorem 1 and §2.7"}],"minor_comments":[{"comment":"The phrase 'conditioning on ||x_n||^2 ≥ √d' appears to be a typo; it should be '||x_n|| ≥ √d' to match the surrounding inequalities.","section":"§2.2"},{"comment":"In the Beta integral computation, after the substitution u = x^2, the differential 'dx' should be 'du' in the displayed integral; also the line 'cd/2 ∫_0^1 (1−u)^{(d−3)/2} dx' should read 'du'.","section":"§2.7"},{"comment":"The invariant measure is stated to be on [0, ∞]; since the chain lives on [0, ∞), the closed bracket at infinity is likely a typo and should be [0, ∞).","section":"Theorem 1"},{"comment":"The statement says 'Let b = 2 and z−1 ∈ R', but the process is in the plane; this should be z−1 ∈ R^2.","section":"Theorem 2"}],"recommendation":"major_revision","confidential_remarks":"The gap in Section 2.2 is central but appears repairable with standard Markov-chain techniques: a proper lower-bound computation for the growth block, a Foster–Lyapunov function such as V(r)=r^2 (whose drift is negative on the tail for every d≥2), and a direct verification of the minorization condition from the explicit kernel. The deterministic sections are independent and may be publishable even if the probabilistic part needs further work. The paper would benefit from a clear statement that Theorem 1 is conditional on the Harris-recurrence proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the new result is Theorem 1. For i.i.d. uniform directions on S^{d-1}, the radial greedy walk has a unique invariant measure with expectation sqrt(pi)/2 * Gamma((d+1)/2)/Gamma(d/2) ~ sqrt(pi d/8), and faster-than-Gaussian decay. The mean computation is self-contained: the density of |v_1| is derived from the uniform sphere, stationarity gives E[||x||] without any fitted constant, and the Lyapunov argument for exponential moments is clean. That part is solid and genuinely new; prior work covered signed harmonic series and deterministic plane sequences, not i.i.d. random directions.\n\nWhere it gets soft: exactly where the stress-test note says. Section 2.2's Harris-chain construction is not complete. The block-growth claim for ~2d steps with |gamma_1|<=1/100 does not follow from the displayed inequality; for large d, 2*(1/100)*sqrt(d) dominates the +1 increment, so the bound is insufficient. The minorization condition is asserted, not verified, and the drift-back argument leaves the interval [sqrt(d)+1/2, 10000 sqrt(d)] uncovered for small d. As written, Theorem 1 is not fully proved. But the gap is fillable: a standard Foster-Lyapunov function V(r)=r^2 gives E[r_{n+1}^2 - r_n^2 | r] = 1 - 2E[X] r, strictly negative on the tail for every d>=2, and a direct lower bound on the r^2-increment gives the growth step. This looks like a missing computation, not a wrong claim.\n\nThe deterministic results (Theorems 2-5) are also new, but rougher. Lemma 6 omits the non-aligned case, and base 3 is explicitly left open. Fine for an exploratory paper, but the proofs need more care.\n\nOverall: the central idea is convincing, the citation pattern is honest, and there are no fitted parameters. The paper deserves a serious referee; I'd send it out and expect a conditional accept after the Harris argument is repaired and Lemma 6 is completed. I'd bring it to reading group.","headline":"The random-case theorem is likely correct and worth refereeing; the Harris-recurrence proof in §2.2 is a real gap, but a fillable one.","tokens_in":22307,"tokens_out":2383,"would_cite":true,"duration_ms":22161,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J05","60G50","37A30","11K31"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that a greedy random walk on the sphere—always choosing the sign that moves closer to the origin—has, in every dimension $d\\ge2$, a unique invariant radial measure with mean…","keywords":["greedy sign choice","random unit vectors","invariant measure","Harris chain","radial Markov chain","van der Corput sequence","exponential moments","concentration of measure"],"falsifier":"Simulate the radial chain in $d=100$ starting at $r=0$ for many trajectories; Theorem 1 requires a positive fraction (bounded away from 0) to hit $[\\sqrt{d},\\sqrt{d}+1/2]$ within $2d$ steps, and requires the long-run empirical mean of $r_n$ to converge to $\\sqrt{\\pi d/8}$. A run showing the hitting fraction going to 0 as $d$ grows, or the empirical mean deviating from the formula, would refute the central claim.","tokens_in":21207,"feed_emoji":"🎯","tokens_out":6896,"duration_ms":62168,"temperature":0.7,"pith_summary":"This paper studies a simple greedy rule: at each step, given a random unit vector, move by either that vector or its negative, whichever puts you closer to the origin. The central result is that in every dimension $d\\ge2$ the distance to the origin has a unique stationary distribution $\\pi_d$ on $[0,\\infty)$, with mean $\\frac{\\sqrt{\\pi}}{2}\\frac{\\Gamma((d+1)/2)}{\\Gamma(d/2)}$, which grows like $\\sqrt{\\pi d/8}$. The distribution decays faster than Gaussian, so the walker is effectively confined to a thin shell near that radius. The paper also proves sharp statements for a deterministic planar rule using the van der Corput sequence: from any start, the walk reaches a fixed small neighborhood of the origin in $O(|z|)$ steps, and periodic starting positions exist exactly for certain bases, tracing regular $b$-gons. A reader should care because the result turns an apparently unstructured random sign choice into a precise quantitative law with a computable universal constant.","feed_headline":"Greedy random walk hovers at sqrt(pi d/8) from origin","feed_subtitle":"In any dimension, always stepping toward the origin leaves distance ~sqrt(pi d/8) on average.","key_machinery":"The carrying object is the radial representation $r_{n+1} = (r_n^2 - 2X_n r_n +1)^{1/2}$, where $X_n = |\\gamma_1|/(\\gamma_1^2+\\chi^2(d-1))^{1/2}$ is the absolute value of the first coordinate of a uniformly random point on $\\mathbb{S}^{d-1}$; its density is $c_d(1-x^2)^{(d-3)/2}$, and it carries the drift toward and away from the origin. The Harris-chain framework with the set $A=[\\sqrt{d},\\sqrt{d}+1/2]$ establishes existence and uniqueness of the invariant measure; the Lyapunov function $e^{\\alpha x^2}$ converts the one-step bound $\\mathbb{E}e^{\\alpha r_{n+1}^2} \\le e^{\\alpha}\\frac{c_d}{2\\alpha r_n} e^{\\alpha r_n^2}$ into finiteness of exponential moments; the identity $\\mathbb{E}[\\|x\\|]=1/(2\\mathbb{E}[X])$ then gives the exact mean. For $d=3$ the transition kernel simplifies to $P_3(x,y)=y/x$, yielding a curious self-referential integral identity for $\\pi_3$.","core_discovery":"The paper's central claim is Theorem 1: for each $d\\ge2$, the greedy random walk's radial coordinate $\\|x_n\\|$ is a positively recurrent Harris chain on $[0,\\infty)$ with a unique invariant probability $\\pi_d$, whose mean is $\\frac{\\sqrt{\\pi}}{2}\\frac{\\Gamma((d+1)/2)}{\\Gamma(d/2)}$, and for which $\\int_0^\\infty e^{\\alpha x^2}\\,d\\pi_d(x)<\\infty$ for every $\\alpha>0$. In particular, the walker's typical distance from the origin is asymptotically $(\\sqrt{\\pi/8}+o(1))\\sqrt{d}$. The proof decouples angular and radial motion, writes the radial update as $r_{n+1}^2 = r_n^2 - 2X_n r_n +1$ with $X_n$ the absolute first coordinate of a uniform sphere point, verifies Harris recurrence with a minorization condition on the interval $[\\sqrt{d},\\sqrt{d}+1/2]$, constructs a Lyapunov function $V(x)=e^{\\alpha x^2}$ to get exponential tail decay, and then computes the mean exactly from the steady-state balance $0 = 1 - 2\\mathbb{E}[X]\\mathbb{E}[\\|x\\|]$.","pith_inferences":["Beyond the paper: the steady-state identity $\\mathbb{E}[\\|x\\|] = 1/(2\\mathbb{E}[X])$ is derived for uniform sphere steps, but the balance argument only needs isotropy and a finite second moment; a natural extension is that any centered isotropic step distribution with finite second moment yields an invariant mean of the form $1/(2\\mathbb{E}|\\cos\\Theta|)$ in the appropriate scaling.","Beyond the paper: the faster-than-Gaussian decay, combined with the paper's conjecture that $\\pi_d$ approaches $\\exp(-(x-\\mu_d)^2)/\\sqrt{\\pi}$, suggests the radial fluctuations are $O(1)$ while the mean radius grows like $\\sqrt{d}$; if true, the walker is localized in a shell of constant thickness, a sharp concentration effect.","Beyond the paper: the base-2 van der Corput result and the periodic-start classification suggest a threshold phenomenon in $b$: for low base, the greedy planar walk is trapped near the origin, while for even $b\\ge4$ it cannot be trapped at all; testing intermediate bases such as $b=3$, where the paper notes periodic starts exist empirically but are unproved, could locate the exact threshold."],"forward_implications":["In all dimensions $d\\ge2$, the invariant radial law is unique and has expectation exactly $\\frac{\\sqrt{\\pi}}{2}\\frac{\\Gamma((d+1)/2)}{\\Gamma(d/2)}$, so the long-run average distance from the origin is $(\\sqrt{\\pi/8}+o(1))\\sqrt{d}$.","Every exponential moment of the radial law is finite; in particular, excursions far beyond the mean radius are exponentially rare with a rate stronger than Gaussian.","For the base-2 van der Corput sequence in the plane, any real starting value reaches $B(0,\\sqrt{2})$ within $O(|z|)$ steps, provided the sign choice never becomes indeterminate.","In base 2, every nonzero point in the unit disk is a periodic start: the orbit is confined to a point plus a semicircle around it, whereas for base $b\\ge4$, periodic starts exist exactly for odd $b$, in two open triangles, and each periodic orbit is a rotated regular $b$-gon."],"supporting_citations":[{"why":"The one-dimensional greedy signed harmonic series whose analogue this paper generalizes to random vectors in $\\mathbb{R}^d$.","marker":"[1]"},{"why":"Supplies the Harris-chain existence criterion used to obtain the unique invariant measure $\\pi_d$.","marker":"[4]"},{"why":"Provides the concentration-of-measure interpretation for why the walk stabilizes near $\\sqrt{d}$.","marker":"[5]"},{"why":"Supplies Theorem 10.0.1 and the Lyapunov drift lemma 14.3.7 used for unique invariant measure and finite exponential moments.","marker":"[6]"},{"why":"Introduces the two-dimensional greedy unit-vector dynamical system whose random version is the subject of Theorem 1.","marker":"[8]"},{"why":"Defines the van der Corput sequence used in the deterministic constructions of Theorems 2 through 5.","marker":"[9]"}],"fun_headline_variants":["Greedy random walk hovers near sqrt(pi d/8)","Always step to origin, stay at sqrt(pi d/8)","Optimal signs yield steady radius sqrt(pi d/8)","Random greedy steps pin distance at sqrt(pi d/8)","Stepping smart keeps walker at sqrt(pi d/8)"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the radial process, from any starting radius no larger than $\\sqrt{d}$, has a uniformly positive chance of reaching the interval $[\\sqrt{d},\\sqrt{d}+1/2]$ within about $2d$ steps, and drifts back from above that interval; if that fails, the invariant measure may not exist.","fun_headline_variants_meta":{"raw":{"variants":["Greedy random walk hovers near sqrt(pi d/8)","Always step to origin, stay at sqrt(pi d/8)","Optimal signs yield steady radius sqrt(pi d/8)","Random greedy steps pin distance at sqrt(pi d/8)","Stepping smart keeps walker at sqrt(pi d/8)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1487,"prompt_tokens":1002,"completion_tokens":485,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":395}},"tokens_in":618,"tokens_out":485,"duration_ms":5392,"temperature":1.0,"reasoning_tokens":395,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:34:59.625110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the radial chain in $d=100$ starting at $r=0$ for many trajectories; Theorem 1 requires a positive fraction (bounded away from 0) to hit $[\\sqrt{d},\\sqrt{d}+1/2]$ within $2d$ steps, and requires the long-run empirical mean of $r_n$ to converge to $\\sqrt{\\pi d/8}$. A run showing the hitting fraction going to 0 as $d$ grows, or the empirical mean deviating from the formula, would refute the central claim.","supporting_citations":[{"cited_title":"Bettin, G","cited_arxiv_id":null,"evidence_quote":"The one-dimensional greedy signed harmonic series whose analogue this paper generalizes to random vectors in $\\mathbb{R}^d$."},{"cited_title":"Harris (1956) The existence of stationary measures for certain Markov processes","cited_arxiv_id":null,"evidence_quote":"Supplies the Harris-chain existence criterion used to obtain the unique invariant measure $\\pi_d$."},{"cited_title":"Ledoux, The concentration of measure phenomenon (No","cited_arxiv_id":null,"evidence_quote":"Provides the concentration-of-measure interpretation for why the walk stabilizes near $\\sqrt{d}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Theorem 10.0.1 and the Lyapunov drift lemma 14.3.7 used for unique invariant measure and finite exponential moments."},{"cited_title":"A curious dynamical system in the plane","cited_arxiv_id":"2409.08961","evidence_quote":"Introduces the two-dimensional greedy unit-vector dynamical system whose random version is the subject of Theorem 1."},{"cited_title":"van der Corput, Verteilungsfunktionen II, Akad","cited_arxiv_id":null,"evidence_quote":"Defines the van der Corput sequence used in the deterministic constructions of Theorems 2 through 5."}],"review_version":1}