{"id":"f0ef6e4c-82c2-444a-a645-54b8aadbb01e","arxiv_id":"2412.17475","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Random projections of ℓ_p-balls satisfy a large deviations principle whose rate function is finite only on L_q-zonoids and given by a maximum entropy gap, with almost sure convergence to a Euclidean ball.","lead":"This paper proves a large deviations principle for the random shadows, meaning orthogonal projections, of the unit ball of ℓ_p^n onto random subspaces, describing the rare shapes that appear with exponentially small probability. The rate function is expressed through a maximum entropy problem over measures generating L_q-zonoids, and it yields a new Gamma function inequality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem 1's rate function contains an unjustified symmetrization step: the covariance constraint in the Kim–Ramanan LDP does not survive symmetrization, so the stated rate with Z_2(µ) ⊂ B_2 may differ from the contraction rate.","rationale":"The reader's weakest_assumption pointed to the imported Kim–Ramanan LDP as the fragile external input. My stress-test agrees that the derivation from that LDP is the vulnerable spot, but I locate the precise failure inside the paper's own proof: the symmetrization used to eliminate the barycenter term does not preserve the covariance constraint, and equation (2) conflates the covariance matrix with the second-moment matrix. This is more specific than 'the external theorem might be wrong' and it directly threatens the stated form of Ip(K). The paper has substantial independent support: the continuity of Zq (Proposition 15), the duality arguments for sections, and the maximum-entropy computations in Section 5 are well developed and appear internally consistent. The concern is not about the authors' honesty or the novelty of the framework, but about whether the main rate function as written is actually the one delivered by the contraction principle. A conditional verdict is appropriate because the issue may be resolved either by correcting the proof (if the external constraint is the second moment) or by revising the rate function; the concrete test above would settle which. I do not recommend ACCEPT or REJECT without that check, hence CONDITIONAL, consistent with the reader's verdict but for a more pointed reason.","tokens_in":21664,"tokens_out":32433,"duration_ms":276508,"concrete_test":"Check the exact statement of Kim–Ramanan [23, Thm. 2.8]: determine whether the constraint in Hk is Cov(µ) ≤ I or E[XXᵀ] ≤ I. If it is Cov(µ) ≤ I, perform a numerical optimization for a simple case, e.g., k = 2, q = 1 (p = ∞), with K a small Euclidean ball or a segment. Optimize inf{Ent(γ) − Ent(µ) + ½∥bar µ∥² : Zq(µ) = K, Cov(µ) ≤ I} over non-centered absolutely continuous measures, such as two-component Gaussian mixtures, and compare with inf{Ent(γ) − Ent(µ) : Zq(µ) = K, E[XXᵀ] ≤ I}. If the former is strictly smaller for any K, Theorem 1's rate function is incorrect; if they coincide, the symmetrization gap may be repairable. Independently re-derive the contraction from Proposition 22 without symmetrization to verify the equality of the two infima.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is Theorem 1, whose rate function Ip(K) is derived in Section 3 by contraction from Proposition 22. The contraction gives Ip(K) = inf{Hk(µ) : Zq(µ) = K}, where Hk(µ) = Ent(γ) − Ent(µ) + ½∥bar µ∥² if Cov(µ) ≤ I. The proof then symmetrizes µ: 'The symmetrized version µ̃ of µ ... consequently Hk(µ̃) ≤ Hk(µ)'. This inequality is false. If µ has mean m and covariance Σ ≤ I, the symmetrized measure µ̃ = (µ + µ∘(−id))/2 has covariance Σ + mmᵀ, which need not satisfy ≤ I. Example: µ = N(m, σ²I) with ∥m∥ > √(1−σ²) gives Cov(µ) = σ²I ≤ I, but Cov(µ̃) has largest eigenvalue σ² + ∥m∥² > 1, so Hk(µ̃) = ∞ while Hk(µ) < ∞. Moreover, equation (2) states 'Z2(µ) ⊂ B2 is equivalent to Cov(µ) ≤ I', but Z2(µ) ⊂ B2 is equivalent to the second-moment matrix E[XXᵀ] ≤ I, not to Cov(µ) ≤ I; the two agree only for centered µ. Thus the proof either misstates the external LDP's constraint or fails to justify the symmetrization. Since Theorem 1, Corollary 3, Theorem 5, and Theorem 11 all inherit this rate function, this is the single most load-bearing weakness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper establishes a large deviations principle (LDP) for the sequence of random convex bodies Z_{n,p} = n^{1/p-1/2} Π_{n,k}^T B_p^n, the orthogonal projections of ℓ_p^n balls onto random k-dimensional subspaces, for 2 < p ≤ ∞ (and its dual statement for random sections for 1 ≤ q < 2). The LDP takes place in the space of convex bodies with the Hausdorff metric, and the rate function I_p(K) is expressed as a maximum entropy over probability measures µ generating the L_q-zonoid K, subject to the constraint Z_2(µ) ⊂ B_2^k. The paper derives as corollaries the almost sure convergence of the renormalized projections/sections to a Euclidean ball of radius m_q or m_q^{-1}, computes the exponential rate of the small-ball probability P[Z_{n,p} ⊂ β B_2^k], and proves an auxiliary inequality for the Gamma function. The main proof routes through the Kim–Ramanan LDP for the empirical measures of rows of a Haar-distributed Stiefel matrix, the continuity of the L_q-zonoid map, the contraction principle, and a duality argument for sections.","tokens_in":22013,"tokens_out":13584,"duration_ms":113942,"significance":"If the main theorem is correct, the paper makes a substantial contribution to asymptotic convex geometry and large deviations: it gives a full shape-level LDP for random projections of ℓ_p-balls, with a rate function that is explicit enough to compute small-ball asymptotics, and it unifies several earlier functional and volume results. The approach via L_q-zonoids and maximum entropy is novel and well adapted to the problem. The paper also contains a self-contained proof of an interesting Gamma-function inequality (Proposition 13), and its use of the external Kim–Ramanan LDP is clearly signposted. The main concern is whether the proof of the rate function in Theorem 1 is valid as written; the symmetrization step and equation (2) contain a load-bearing gap that affects all subsequent results.","major_comments":[{"comment":"The symmetrization step is invalid. For a measure µ with covariance Σ ≤ I and mean m, the symmetrized measure µ̃ = (µ + µ∘(−id))/2 has covariance Σ + mmᵀ, which need not satisfy Σ + mmᵀ ≤ I. If it does not, then H_k(µ̃) = ∞ by the definition in (13), while H_k(µ) < ∞, so the claimed inequality H_k(µ̃) ≤ H_k(µ) fails. A concrete counterexample is µ = N(m, σ²I) with σ²I ≤ I and ||m||² > 1 − σ². Thus the reduction to centered measures, which is used to drop the term (1/2)||bar µ||² from the contraction rate, is not justified. Consequently, the rate function stated in Theorem 1, with the constraint Z_2(µ) ⊂ B_2^k, is not shown to be equal to the contraction rate (14). This gap propagates to Corollary 3, Theorem 5, and Theorem 11, all of which rely on the explicit form of I_p.","section":"Section 3, proof of Theorem 1 (equations (13)–(14), (2))"},{"comment":"The statement that Z_2(µ) ⊂ B_2^k is equivalent to Cov(µ) ≤ I_{k×k} is false. By definition Z_2(µ) has support function (∫ |⟨x,u⟩|² dµ(x))^{1/2}, so Z_2(µ) ⊂ B_2^k is equivalent to E[XXᵀ] ≤ I_{k×k}, not to Cov(µ) ≤ I_{k×k}. The two conditions differ by the rank-one matrix mmᵀ, where m = E[X]. They agree only for centered µ. This misstatement is not a minor typo: it is used in the proof of Theorem 1 to replace the covariance constraint from Proposition 22 with the second-moment constraint in the theorem, and it changes the set of admissible measures in the rate function.","section":"Equation (2) and Remark 2"},{"comment":"There is an inconsistency between the exact constraint satisfied by the empirical measures L_n and the stated rate function H_k in (13). For every realization of the Stiefel matrix Π_{n,k}, the rows v_i satisfy Σ_{i=1}^n v_i v_iᵀ = ΠᵀΠ = I_k, so the measure L_n = (1/n)Σ δ_{√n v_i} has second moment exactly I_k. The sequence (L_n) therefore lives in the closed set of probability measures with second moment equal to I_k, and any LDP for it must assign infinite rate to measures with second moment different from I_k. However, H_k as defined in (13) is finite for many such measures, e.g. a centered Gaussian with covariance σ²I for σ² < 1. This suggests that either Proposition 22 misstates the Kim–Ramanan result (the correct constraint may be on the second moment rather than the covariance), or the sequence L_n is not the one to which (13) applies. The authors need to reconcile Proposition 22 with the precise statement in [23] and with the identity in Proposition 20.","section":"Proposition 20 and Proposition 22"}],"minor_comments":[{"comment":"There are a number of typos and OCR artifacts: 'entr opy' in the abstract and Section 1, 'satifies' for 'satisfies', 'δk,q as in Theorem 11 satifies δk,q = k/q a_{k,q}' in Section 5. These should be corrected in a revision.","section":"Throughout"},{"comment":"The proof sketch for extending the almost sure convergence to 1 < p < ∞ is quite compressed; since this remark is not used later, it is acceptable, but a few more details on the Borel–Cantelli argument would improve readability.","section":"Remark 4"},{"comment":"The form of the optimizer v*(r) in (24) is stated without a derivation; the paper refers to [11, Thm. 12.1.1], which is appropriate, but a short explanation of the complementary slackness conditions would make the argument easier to follow.","section":"Section 5, equation (24)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong and well-motivated framework, but the proof of the main theorem contains a genuine gap in the symmetrization step, and equation (2) is a factual error. The issue is load-bearing. I would advise a major revision rather than rejection, because the underlying approach is sound and the gap may be fixable: if the correct Kim–Ramanan rate function is stated with a second-moment constraint (or if the contraction is performed directly on the constrained set), the symmetrization step works and the stated theorem follows. However, as written, the proof does not establish the stated rate function, and the small-ball constant in Theorem 11 depends on the rate function. The authors should also clarify the exact statement of Proposition 22 and its compatibility with the identity in Proposition 20."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a substantial paper that deserves a serious referee, but the proof of the main theorem has a real gap, and until that's fixed I'd be reluctant to rely on the rate function as stated.\n\nWhat's new: the LDP for the random convex body Z_{n,p} in the Hausdorff metric, with rate finite only on L_q-zonoids and expressed as a maximum-entropy gap, is not in the literature. The almost sure convergence to a Euclidean ball is new for 2<p<∞. The Gamma inequality is a nice byproduct. The paper is clearly written, the overall strategy—contraction from the Kim-Ramanan empirical-measure LDP plus continuity of the L_q-zonoid map—is sound in spirit, and the duality argument for sections is clean.\n\nSoft spots: the proof of Theorem 1 contains an unjustified symmetrization. Equation (2) states that Z_2(µ) ⊂ B_2 is equivalent to Cov(µ) ≤ I. That's false; the correct condition is on the second-moment matrix. More seriously, after contracting from Proposition 22, the rate is the infimum over µ with Cov(µ) ≤ I. The paper symmetrizes µ to µ̃, claims H_k(µ̃) ≤ H_k(µ). But if µ has mean m and covariance Σ≤I, then µ̃ has covariance Σ+mmᵀ, which need not be ≤I. For example, take µ = N(m, σ²I) with ||m|| > √(1−σ²). Then H_k(µ) is finite, but H_k(µ̃)=∞ because the constraint is violated. So the symmetrization step fails exactly as written. This matters because Theorem 1, Corollary 3, Theorem 5, and Theorem 11 all inherit this rate function. It is likely fixable—maybe the Kim-Ramanan rate actually has the second-moment constraint, in which case Proposition 22 needs correcting, or one needs a different argument to eliminate the mean term—but as it stands, the proof doesn't establish the stated rate.\n\nMinor: the abstract says the paper identifies the asymptotic probability that the projection is contained in a small ball. That's true for β ≤ β_{k,q} and β ≥ m_q, but for β in between the constant is left undetermined, which the paper itself concedes. The abstract oversells slightly.\n\nBottom line: the ideas are strong, the citation pattern is honest, no fitted parameters, no circularity. I'd send this to a thoughtful referee, but I'd want the referee to press on the symmetrization step before accepting.","headline":"Substantial new LDP for random projections of ℓ_p-balls, but the proof of the main rate function has a gap around symmetrization that needs fixing.","tokens_in":22555,"tokens_out":29494,"would_cite":true,"duration_ms":262373,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["52A23","52A21","52A22","60F10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Random projections of $\\ell_p$-balls obey a large deviations principle with an entropy-based rate function over $L_q$-zonoids.","keywords":["large deviations","random projections","random sections","L_q-zonoids","ℓ_p-balls","maximum entropy","Stiefel manifold","almost sure convergence"],"falsifier":"Take $p=\\infty$, $k=2$, and $\\beta\\le\\beta_{2,1}$; Theorem 11 gives an explicit constant $c_{2,1,\\beta}$. Simulate many independent Haar-distributed projections $\\Pi_{n,2}$ for increasing $n$ and compare $n^{-1}\\log P[Z_{n,\\infty}\\subset\\beta B_2^2]$ with that constant: a systematic deviation as $n$ grows would refute the large deviations principle, while agreement would confirm the rate function and the continuity argument.","tokens_in":21465,"feed_emoji":"🎲","tokens_out":16543,"duration_ms":127705,"temperature":0.7,"pith_summary":"The paper proves a large deviations principle for the $k$-dimensional orthogonal projections of the unit ball of $\\ell_p^n$ onto a random subspace, as $n\\to\\infty$, in the case $2<p\\le\\infty$, and for random sections in the conjugate range $1\\le q<2$. The rate function is finite only on $L_q$-zonoids---convex bodies whose support function is a $q$-th moment of a probability measure---and equals the entropy gap between the standard Gaussian and the maximum-entropy measure that generates the shadow, subject to a covariance constraint. Its unique zero is the Euclidean ball of radius $m_q$, so the classical typical-shadow result becomes the almost sure limit of the rescaled projections. The paper also computes the asymptotic probability that the projection is contained in a Euclidean ball of small radius, with an explicit exponential rate for small radii, and proves a Gamma-function inequality that separates the regimes. This provides the full set of rare shapes of $\\ell_p$-ball shadows in terms of their geometry, rather than only rates for individual functionals.","feed_headline":"Rare shadows of ℓ_p-balls get exact exponential probabilities","feed_subtitle":"A large deviations principle pinpoints the unlikely shapes and gives an explicit rate for small-ball containment","key_machinery":"The central object is the $L_q$-zonoid: a convex body $K$ whose support function is $h_K(u)=(\\int_{\\mathbb{R}^k}|\\langle x,u\\rangle|^q\\,d\\mu(x))^{1/q}$ for some probability measure $\\mu$, so the possible finite-rate shadows are exactly these bodies. The engine of the proof is a large deviations principle for the empirical measures $L_n=\\frac1n\\sum_{i=1}^n\\delta_{\\sqrt n v_i}$ formed from the rows $v_i$ of a Haar-distributed Stiefel matrix (Proposition 22); because the map $\\mu\\mapsto Z_q(\\mu)$ is continuous in the appropriate Wasserstein metric, the contraction principle converts that measure-valued LDP into a body-valued LDP with rate $I_p(K)$. The rate function is then re-expressed as a maximum entropy problem, and the small-ball theorem is solved by identifying the maximizer on rotationally invariant classes: for $\\beta\\le\\beta_{k,q}$ the maximizer has density $x\\mapsto \\omega_k^{-1}Z_{k,q,\\beta}^{-1}e^{-\\lambda_{k,q,\\beta}\\|x\\|^q}$, and the gap regime would correspond to an exponential density $r^{k-1}e^{-\\lambda_1 r^q-\\lambda_2 r^2}$ with both parameters positive.","core_discovery":"The central claim is Theorem 1: for fixed $k\\in\\mathbb{N}$ and $2<p\\le\\infty$, the sequence $Z_{n,p}=n^{1/p-1/2}\\Pi_{n,k}^\\top \\mathbb{B}_p^n$ satisfies a large deviations principle in the space of convex bodies equipped with Hausdorff distance, with good rate function $I_p(K)=\\inf\\{\\mathrm{Ent}(\\gamma^{\\otimes k})-\\mathrm{Ent}(\\mu):\\mu\\in\\mathcal{P}_q(\\mathbb{R}^k),\\, Z_q(\\mu)=K,\\, Z_2(\\mu)\\subset B_2^k\\}$, where $Z_q(\\mu)$ is the $L_q$-zonoid generated by $\\mu$ and $1/p+1/q=1$. The rate function is finite exactly on $L_q$-zonoids whose generating measure has covariance bounded by the identity, and its unique zero is $m_q B_2^k$, giving the almost sure convergence of Corollary 3. By polarity, the same result transfers to random sections: $K_{n,q}=\\Pi_{n,k}^\\top(n^{1/q-1/2}\\mathbb{B}_q^n\\cap E_{n,k})$ has rate function $J_q(K)=I_p(K^\\circ)$ (Theorem 5). The proof obtains the body-valued LDP by contracting an LDP for the empirical measure of the rows of the random Stiefel matrix, then rewrites the rate as a maximum entropy problem. A further maximum-entropy computation yields the explicit small-ball exponent in Theorem 11.","pith_inferences":["Beyond the paper: solving the two-moment maximum entropy problem in the gap $\\beta_{k,q}<\\beta<m_q$ would determine the missing constant $c_{k,q,\\beta}$; the expected optimizer is $v(r)=r^{k-1}e^{-\\lambda_1 r^q-\\lambda_2 r^2}$ with both parameters positive, so the constants can be computed numerically and checked by simulation.","Beyond the paper: because every finite-rate shadow is an $L_q$-zonoid with a covariance constraint, extremal questions about volume, mean width, or other functionals of the projection reduce to optimization over a finite-dimensional family of generating measures, which may yield sharper bounds than functional-level large deviations alone.","Beyond the paper: the complementary range $1<p<2$ is expected to need speed $n^{2/q}$; if a measure-valued LDP with that speed exists, the same contraction and maximum-entropy machinery would give the analogous body-valued LDP, making the present result a template for the remaining cases.","Beyond the paper: the conjectural form of the maximum-entropy density connects the rate function to $q$-stable laws---when the quadratic constraint is dropped, the density is the Fourier transform of a symmetric $q$-stable distribution---so the 'strangeness' of a shadow may be interpretable as a distance in spectral-measure space to Gaussianity."],"forward_implications":["For every fixed $k$ and $2<p\\le\\infty$, the rescaled projections $Z_{n,p}$ converge almost surely in Hausdorff distance to $m_q B_2^k$; the random sections in the dual range converge almost surely to $m_q^{-1}B_2^k$.","Any shadow that appears with merely exponential probability must be an $L_q$-zonoid contained in $B_2^k$ with nonempty interior; all other convex bodies have infinite rate and appear with super-exponentially small probability.","The asymptotic probability that $Z_{n,p}\\subset\\beta B_2^k$ is $\\exp(n\\,c_{k,q,\\beta}+o(n))$ for $\\beta\\le\\beta_{k,q}$, with $c_{k,q,\\beta}$ given explicitly in terms of Gamma functions, and is not exponentially small once $\\beta\\ge m_q$.","Continuous functionals such as volume, intrinsic volumes, or diameter inherit a large deviations principle from the rate function by the contraction principle, so their rare deviations can in principle be read off the same entropy functional.","In the intermediate range $\\beta_{k,q}<\\beta<m_q$ the exponential rate exists but is not computed; the paper shows the relevant sets are continuity sets there, so the missing value is well defined."],"supporting_citations":[{"why":"Supplies the large deviations principle for the empirical measures of the rows of the Haar-distributed Stiefel matrix, which is the input to the contraction principle.","marker":"[23]"},{"why":"Provides the large deviations framework and the contraction principle used to pass from measure-valued to body-valued LDPs.","marker":"[12]"},{"why":"Gives the background on L_q-zonoids, support functions, and polar duality that define the rate function's domain and the section-projection duality.","marker":"[48]"},{"why":"Supplies the maximum-entropy techniques (compactness, complementary slackness, explicit optimizer form) used for the small-ball exponent.","marker":"[22]"},{"why":"Provides the high-dimensional typical-shadow theorem in the form used to motivate the almost sure limit.","marker":"[5]"},{"why":"Earlier result for random projections of the cube that the new almost sure limit extends and reproves through the zonoid route.","marker":"[20]"}],"fun_headline_variants":["Exact probabilities for rare shadows of ℓ_p-balls","Large deviations for random projections of ℓ_p-balls","Rare shapes of ℓ_p-ball projections get exact rates","Maximum entropy pins down ℓ_p-ball shadow tails","ℓ_p-ball projections: large deviations with exact rate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The chain of proof imports, without reproof, a large deviations principle for the empirical measures of the rows of a random orthogonal frame; if that imported theorem, or the range of exponents it covers, were not valid, the contraction argument and the subsequent LDP would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Exact probabilities for rare shadows of ℓ_p-balls","Large deviations for random projections of ℓ_p-balls","Rare shapes of ℓ_p-ball projections get exact rates","Maximum entropy pins down ℓ_p-ball shadow tails","ℓ_p-ball projections: large deviations with exact rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1615,"prompt_tokens":1044,"completion_tokens":571,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":490}},"tokens_in":660,"tokens_out":571,"duration_ms":5056,"temperature":1.0,"reasoning_tokens":490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:40:10.001701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $p=\\infty$, $k=2$, and $\\beta\\le\\beta_{2,1}$; Theorem 11 gives an explicit constant $c_{2,1,\\beta}$. Simulate many independent Haar-distributed projections $\\Pi_{n,2}$ for increasing $n$ and compare $n^{-1}\\log P[Z_{n,\\infty}\\subset\\beta B_2^2]$ with that constant: a systematic deviation as $n$ grows would refute the large deviations principle, while agreement would confirm the rate function and the continuity argument.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the large deviations principle for the empirical measures of the rows of the Haar-distributed Stiefel matrix, which is the input to the contraction principle."},{"cited_title":"Schneider","cited_arxiv_id":null,"evidence_quote":"Gives the background on L_q-zonoids, support functions, and polar duality that define the rate function's domain and the section-projection duality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-entropy techniques (compactness, complementary slackness, explicit optimizer form) used for the small-ball exponent."},{"cited_title":"Aubrun and S","cited_arxiv_id":null,"evidence_quote":"Provides the high-dimensional typical-shadow theorem in the form used to motivate the almost sure limit."},{"cited_title":"Kabluchko, J","cited_arxiv_id":null,"evidence_quote":"Earlier result for random projections of the cube that the new almost sure limit extends and reproves through the zonoid route."}],"review_version":1}