{"id":"e171de1a-b607-4d24-a123-14404cff45c7","arxiv_id":"2412.05306","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A closed-form CDF for Roy's largest root of a non-central F-matrix with rank-one non-centrality is derived, with SNR-scaling limits for detection power in finite and high-dimensional regimes.","lead":"The paper derives a new exact formula for the distribution of the largest eigenvalue of a non-central F-matrix, the quantity used by Roy's largest-root test for detecting signals in colored noise with unknown covariance. It then shows when this test has power as problem dimensions grow, connecting detection limits to known phase transitions in random matrix theory.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's Ψ argument should be 2/t+1, not 2/(t+1); as printed the CDF fails the m=1 sanity check.","rationale":"The reader's weakest assumption was the 0/0 limiting argument in Appendix D, but that limiting step is standard confluence of parameters and is probably sound. The actual load-bearing defect is later, in the conversion from the proof's internal x-threshold to the λ-threshold of the theorem. This is not a matter of pedantry: the printed Corollary 3 is not a valid CDF, since it tends to 0 as t → ∞ for m = 1, p = 1, n = 2. The same error propagates into Theorem 3 and hence into the ROC and high-dimensional claims as stated. I verified that changing the argument from 2/(t+1) to 2/t+1 fixes the m = 1 exact formulas for both central and noncentral cases, which strongly suggests the derivation is otherwise correct and the error is localized. Because the central theorem as printed is false but appears readily correctable, the appropriate verdict is conditional acceptance pending correction of the Ψ argument and re-verification of the numerical figures.","tokens_in":31002,"tokens_out":45772,"duration_ms":374589,"concrete_test":"Evaluate Theorem 3 exactly as printed for m = 1, p = 1, n = 2, ω = 0, t = 1: it yields 1/4, while the exact scalar central F CDF is 3/4. Then rerun the same evaluation with Ψ_{i,j}(t) changed from P(...)(2/(t+1)) to P(...)(2/t+1): the formula should match 1 − (1+t)^{-2} for all t, and for ω > 0 it should match the direct noncentral Laplace integral. As a second check, regenerate Figures 1–4 with the corrected argument; any visible shift confirms that the simulations used the corrected formula or that the theorem statement is erroneous.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim, Theorem 3, is false as stated because of a variable-conversion error in the Jacobi-polynomial argument. In the proof, the symbol t throughout Appendix A and Appendix D is the threshold for x = λ/(1+λ), not the λ-threshold. Equation (21) converts to λ only at the end. Appendix D therefore produces Jacobi arguments of the form 2/t_x − 1, and substituting t_x = t_λ/(1+t_λ) gives 2/t_λ + 1. The theorem instead prints 2/(t_λ+1) inside every Ψ_{i,j}(t). A decisive counterexample is m = 1, p = 1, n = 2 (so α = 1, β = 0), ω = 0. Then F is the scalar ratio R/S with R ~ Exp(1) and S ~ Gamma(2,1), whose exact CDF is 1 − (1+t)^{-2}. Corollary 3 as printed gives F(t) = 2t^2/(1+t)^3, which equals 1/4 at t = 1 instead of 3/4 and tends to 0 as t → ∞ instead of 1. Replacing 2/(t+1) with 2/t+1 makes Ψ_{2,2} = 2/t+1 and reproduces the exact scalar CDF, including the noncentral case ω > 0, which I verified symbolically via the Laplace transform of the noncentral chi-square density. Thus the printed theorem, and any ROC or high-dimensional conclusions evaluated from it, are invalid as written; the evidence points to a localized typo in the Ψ argument rather than a failure of the orthogonal-polynomial method.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies detection of a non-random signal in colored Gaussian noise with unknown covariance, using Roy's largest root of the whitened sample covariance matrix as the test statistic. The main theoretical contribution is an exact finite-dimensional CDF, Theorem 3 (Eq. 23), for the largest eigenvalue of a complex non-central F-matrix with a rank-one noncentrality parameter, derived via orthogonal polynomials and contour integrals; the determinant size depends on α = n − m, yielding a scalar formula for m = n (Corollary 2) and a reduction to the central F case (Corollary 3). The authors use this CDF to derive ROC expressions, the CFAR property, finite-dimensional monotonicity results, and high-dimensional detection limits, including the sub- and super-critical SNR phase transition.","tokens_in":31307,"tokens_out":16234,"duration_ms":136916,"significance":"If the main CDF formula is corrected, the paper would provide a valuable exact finite-dimensional characterization of Roy's largest root under a rank-one noncentrality, with the determinant dimension controlled by n − m rather than by m, enabling efficient evaluation when n − m is small. The α = 0 closed form and its high-dimensional consequences (an O(p²) SNR requirement when m = n) are interesting, and Proposition 1 supplies a matrix-integral identity of independent utility. However, Theorem 3 as printed contains a variable-conversion error in the Jacobi-polynomial argument and fails a scalar sanity check, so the quantitative ROC curves and corollaries depending on general α are not valid as stated. The underlying method appears sound and the error is localized, but the paper's significance will be realized only after the printed formulas are corrected and re-verified.","major_comments":[{"comment":"The argument of the Jacobi polynomial in Eq. (23) is wrong as printed. Appendix D produces an expression in the x-threshold t_x with Jacobi argument 2/t_x − 1; substituting t_x = t/(1+t) via Eq. (21) gives 2/t + 1, not 2/(t+1). As a decisive check, take m = p = 1, n = 2 (α = 1, β = 0) and ω = 0: the exact CDF is 1 − (1+t)⁻², whereas Corollary 3 with the printed Ψ₂,₂ = 2/(t+1) gives 2t²/(1+t)³, which tends to 0 as t → ∞ and equals 1/4 at t = 1 instead of 3/4. Replacing 2/(t+1) by 2/t + 1 reproduces the exact scalar CDF. The same incorrect argument appears in Eqs. (68), (70), (71), and (95)/Proposition 1; these should be corrected consistently, and the numerical figures and finite-dimensional ROC formulas (32)–(33) should be re-generated or verified with the corrected formula.","section":"Section III.B, Theorem 3 / Corollary 3"},{"comment":"The coalescing limit s₁ → y, s₂, ..., s_{α+1} → x is asserted by reference to [12], [56] rather than proved. Because both the numerator determinant and the Vandermonde denominator vanish in this limit, the resulting derivative structure and prefactor are load-bearing for Theorem 3. Please provide a self-contained derivation or a precise statement of the identity being invoked.","section":"Appendix D, Eqs. (100)–(103)"},{"comment":"The displayed equivalence between the ROC inequalities and the Loewner order is overclaimed. For a fixed vector a, P_D depends on Σ only through the scalar aᴴΣ⁻¹a, so the Loewner order is sufficient but not necessary for the quadratic-form inequalities; many matrices that are not Loewner-comparable yield the same ordering. The direction used for the majorization conclusion (Loewner order ⇒ ROC order) is correct and should be stated as an implication, or the equivalence should be formulated over all vectors a.","section":"Section IV.A, Eq. (36)"}],"minor_comments":[{"comment":"The limiting curves in the two panels appear to be swapped: for ω = 2 (O(1)) the limit should be e^{−5/t}, and for ω = p it should be e^{−6/t}.","section":"Fig. 3 caption"},{"comment":"The sentence 'The ROC profiles ... are depicted in Fig. 8a' should refer to Fig. 8b.","section":"Fig. 8 caption"},{"comment":"The phrase 'we the following bounds are in order' is missing a verb; it should read 'we have the following bounds' or 'the following bounds are in order'.","section":"After Eq. (58)"},{"comment":"The relabeling rule for the singular F-matrix case (p < m) is stated without derivation or a specific reference; please justify it or provide a precise citation beyond the joint density result in [37].","section":"Section III.B, Eq. (30)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central formula is wrong as printed, but the error appears to be a localized typo in the Jacobi argument, and the scalar counterexample confirms the intended correction. The orthogonal-polynomial framework and the special-case reductions are otherwise sound, so I view this as a major-revision issue rather than a rejection. I would ask the authors to state the corrected theorem explicitly, supply the missing Appendix D limiting argument, and correct the Fig. 3 caption. A brief verification of the m = 1 sanity check in the revision would be useful for the readers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a real, useful result—an exact CDF for the largest eigenvalue of a rank-one non-central F-matrix that costs a determinant of size n−m+1, not m—but the printed Theorem 3 contains a variable-conversion typo that makes the formula wrong as written. The stress-test is right: the Jacobi argument should be 2/t+1, not 2/(t+1). I checked the m=1, α=1, β=0, ω=0 case: the printed Corollary 3 gives 2t²/(1+t)³, which fails at t=1 (1/4 vs 3/4) and tends to 0 instead of 1. The corrected argument reproduces the exact scalar CDF. So the core method is sound; the theorem as stated is not.\n\nWhat's genuinely new: the α-dimensional determinantal representation (Theorem 3) is a real improvement over the authors' own ISIT version, and the special cases (n=m, ω=0) reduce to clean formulas. The proof follows a known orthogonal-polynomial/contour-integral route and is detailed enough to be checked. The ROC analysis, especially the m=n case and the O(p) vs O(p²) SNR phase transition, is a nice addition.\n\nSoft spots: besides the typo, Section IV overclaims the ROC monotonicity equivalence in Eqs. (36)–(38)—that's a one-way implication for fixed signal direction, not a full equivalence. The singular F relabeling (30) is asserted without proof; it may be true but needs a reference or a short argument. The limiting procedure in Appendix D that coalesces the s_i parameters is cited to [12],[56] rather than proved; that's acceptable for a paper of this type, but a referee should push for a few more details.\n\nBottom line: this paper deserves a serious referee. The main theorem is valuable and likely correct after the fix, but nobody should rely on Eq. (23) in its current form. I would not cite v1; I'd cite the corrected version. For the reading group: worth bringing because it's a clean example of how a localized typo can masquerade as a failed result.","headline":"A genuinely new CDF for Roy's largest root under rank-one non-centrality, but Theorem 3 as printed has a variable-conversion typo (2/(t+1) should be 2/t+1) that fails an m=1 sanity check; the method is sound and the paper deserves referee time after a fix.","tokens_in":31885,"tokens_out":5337,"would_cite":false,"duration_ms":42406,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H10","60B20","62H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper derives a closed-form cumulative distribution function for Roy's largest root, the leading eigenvalue of a non-central F-matrix with rank-one non-centrality, and uses it to analyze detection performance in colored noise with…","keywords":["non-central F-matrix","Roy's largest root","signal detection","colored noise","random matrix theory","orthogonal polynomials","receiver operating characteristics","high-dimensional asymptotics"],"falsifier":"For the smallest non-trivial case α=1, m=2, numerically evaluate both sides of the identity connecting (101) and (103), first with s1 and s2 distinct and then with s1→y, s2→x; if the limit does not reproduce the determinant ratio on the right of (103), Theorem 3 is unsupported. Alternatively, simulate F = $S^{{-1/2}}$$RS^{{-1/2}}$ with m=2, n=3, p=5, ω=2 and compare the empirical CDF of λmax with the closed form in (23).","tokens_in":30792,"feed_emoji":"📡","tokens_out":6163,"duration_ms":52836,"temperature":0.7,"pith_summary":"The paper targets the problem of detecting a non-random signal in colored noise when the noise covariance is unknown, using the leading eigenvalue of the whitened sample covariance matrix (Roy's largest root) as the test statistic. Its central result is an exact finite-dimensional CDF for that statistic: when the non-centrality has rank one, the CDF of the largest eigenvalue of a non-central F-matrix is a closed-form determinant whose size depends only on n−m, the excess of noise-only samples over the system dimension. This makes the n=m case a simple exponential expression. The CDF yields explicit receiver operating characteristic (ROC) curves and, in the high-dimensional limit, shows that the test has power only above a phase-transition SNR, with the required SNR scaling differing between fixed and growing dimensions. A sympathetic reader would care because it turns an eigenvalue detection statistic previously studied through asymptotics or simulation into an exactly computable quantity.","feed_headline":"Exact CDF derived for Roy's largest root in colored noise","feed_subtitle":"A closed-form determinant gives the ROC of the leading-eigenvalue test and a phase-transition SNR threshold.","key_machinery":"The argument combines the zonal-polynomial joint eigenvalue density of a non-central F-matrix with the orthogonal-polynomial method for random matrix integrals. A change of variables x = λ/(1+λ) converts the eigenvalue density into a Jacobi-type weight, after which the m-fold integral for the CDF is evaluated via the Andréief–Heine (Cauchy–Binet) identity, yielding a determinant of size α+1 = n−m+1. A contour-integral representation of the confluent hypergeometric function re-sums the first column of the determinant into the compact Φ_i factors. This determinant dimension depends only on n−m, so computational cost tracks the quality of the noise-covariance estimate rather than the ambient dimension m, and the simplest case n=m collapses to a scalar exponential.","core_discovery":"The paper claims that for independently distributed R ∼ CW_m(p, Σ, Θ) and S ∼ CW_m(n, Σ) with rank-one non-centrality Θ, the CDF of the largest eigenvalue λmax of F = $S^{{-1/2}}$$RS^{{-1/2}}$ is $F^{{(α)}}$_{λmax}(t;ω) = K(α,β,m) $e^{{-ω/(1+t)}}$ (t/(1+t))^{m(α+β+m)} det[Φ_i(t,ω) Ψ_{i,j}(t)], where α = n−m, β = p−m, the determinant is of size α+1, and the special case α=0 (n=m) degenerates to $e^{{-ω/(1+t)}}$ (t/(1+t))^{m(β+m)}. The same theorem drives the ROC analysis in Section IV, including the high-dimensional phase transition: for m/p→c1 and m/n→c2∈(0,1) with SNR = pγ, the test has asymptotic power only when γ exceeds the critical value γp = (c2+r)/(1−c2) with r = √(c1+c2−c1c2). When m=n, the required scaling changes: power is retained in the high-dimensional limit only for SNR of order O($p^{2}$), whereas for fixed m and n the requirement is SNR of order O(p).","pith_inferences":["The determinant structure, whose size is n−m+1, suggests that the same orthogonal-polynomial machinery could produce exact distributions for other spectral statistics of non-central F-matrices—such as the smallest eigenvalue or the trace—by adapting the determinant entries while keeping the dimension fixed by n−m.","Because the rank-one non-centrality is the building block, higher-rank non-centrality parameters might be treated by writing the CDF as a mixture or by a perturbation expansion around the rank-one case, although the paper explicitly leaves arbitrary rank as an open problem.","The phase-transition finding implies that any eigenvalue-based detector in this colored-noise setting inherits a fundamental detection boundary: below the critical SNR the largest root carries no information asymptotically, which could serve as a design target for radar systems that must work with limited noise-only samples.","A natural testable extension is to differentiate the determinant CDF to obtain the exact density of Roy's largest root, which would allow likelihood-ratio-based threshold calibration and finite-sample corrections to Tracy–Widom approximations."],"forward_implications":["For any finite m, n, p with p,n ≥ m and rank-one non-centrality, the detection and false-alarm probabilities of Roy's largest-root test can be computed exactly from the determinant formula, with no simulation needed.","When n = m, the ROC collapses to the explicit closed form PD = 1 − (1−PF) exp{−ω(1 − [1−PF]^{1/[m(β+m)]})}, which is an achievable lower bound on all ROC profiles for fixed other parameters.","In the high-dimensional regime with m/p→c1 and m/n→c2∈(0,1), the largest-root test has no asymptotic detection power for SNR = pγ below the phase-transition threshold γp, and power tends to 1 above it.","For m = n, the required SNR scaling to retain power is O(p) in the fixed-dimensional setting but O(p^2) in the high-dimensional setting, and for m < n with SNR = O(p) the leading eigenvalue cannot detect weak signals asymptotically.","The test has the constant-false-alarm-rate (CFAR) property under H0, since the null CDF is independent of the noise covariance Σ."],"supporting_citations":[{"why":"Supplies the joint eigenvalue density of the non-central F-matrix that is the starting point of the derivation.","marker":"[37]"},{"why":"Provides the orthogonal-polynomial approach used to evaluate the multiple eigenvalue integral.","marker":"[53]"},{"why":"Supplies the Andréief–Heine identity and the limiting argument used in Appendix D to reduce the determinant.","marker":"[12]"},{"why":"Supports the coalescing-limit step that converts the parameter-degenerate determinant into (103).","marker":"[56]"},{"why":"Motivates the rank-one alternative and provides the stochastic representation of Roy's largest root under rank-one spiking.","marker":"[16]"},{"why":"Supplies the phase-transition threshold for high-dimensional spiked F-ratios used in the asymptotic ROC analysis.","marker":"[20]"},{"why":"Provides the central F-matrix CDF special case with which Corollary 3 is verified to agree.","marker":"[4]"},{"why":"Supplies the extreme-eigenvalue results for spiked Fisher matrices that support the high-dimensional null and alternative limits.","marker":"[18]"}],"fun_headline_variants":["Roy's largest root: exact CDF for colored noise","Closed-form CDF for non-central F-matrix leading eigenvalue","Roy's test: phase transition in high-dimensional SNR","Weak signal detection: exact CDF for Roy's root"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 0/0 limiting argument in Appendix D, which coalesces several distinct parameters s_i into a single value inside an Andréief–Heine determinant, is valid; this step is asserted by reference to two earlier works rather than proved, and the main determinant formula collapses if that limit is not legitimate.","fun_headline_variants_meta":{"raw":{"variants":["Roy's largest root: exact CDF for colored noise","Closed-form CDF for non-central F-matrix leading eigenvalue","Roy's test: phase transition in high-dimensional SNR","Weak signal detection: exact CDF for Roy's root"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000292,"raw_usage":{"total_tokens":1806,"prompt_tokens":1150,"completion_tokens":656,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":766,"completion_tokens_details":{"reasoning_tokens":589}},"tokens_in":766,"tokens_out":656,"duration_ms":5934,"temperature":1.0,"reasoning_tokens":589,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:41:12.828533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the smallest non-trivial case α=1, m=2, numerically evaluate both sides of the identity connecting (101) and (103), first with s1 and s2 distinct and then with s1→y, s2→x; if the limit does not reproduce the determinant ratio on the right of (103), Theorem 3 is unsupported. Alternatively, simulate F = $S^{{-1/2}}$$RS^{{-1/2}}$ with m=2, n=3, p=5, ω=2 and compare the empirical CDF of λmax with the closed form in (23).","supporting_citations":[{"cited_title":"Distributions of matrix variates and late nt roots derived from normal samples,","cited_arxiv_id":null,"evidence_quote":"Supplies the joint eigenvalue density of the non-central F-matrix that is the starting point of the derivation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the orthogonal-polynomial approach used to evaluate the multiple eigenvalue integral."},{"cited_title":"Couillet and M","cited_arxiv_id":null,"evidence_quote":"Supplies the Andréief–Heine identity and the limiting argument used in Appendix D to reduce the determinant."},{"cited_title":"On the moments of traces of two matrices in three situations for complex multivariate normal populati ons,","cited_arxiv_id":null,"evidence_quote":"Supports the coalescing-limit step that converts the parameter-degenerate determinant into (103)."},{"cited_title":"Roy’s largest root test u nder rank-one alternatives,","cited_arxiv_id":null,"evidence_quote":"Motivates the rank-one alternative and provides the stochastic representation of Roy's largest root under rank-one spiking."},{"cited_title":"Local Asymptotic Normality of the spectrum of high-dimensional spiked F-ratios","cited_arxiv_id":"1411.3875","evidence_quote":"Supplies the phase-transition threshold for high-dimensional spiked F-ratios used in the asymptotic ROC analysis."},{"cited_title":"Eigenvalue-based detection of a signal in colored no ise: Finite and asymptotic analyses,","cited_arxiv_id":null,"evidence_quote":"Provides the central F-matrix CDF special case with which Corollary 3 is verified to agree."},{"cited_title":"Extreme eigenvalues of large-dimen sional spiked Fisher matrices with application,","cited_arxiv_id":null,"evidence_quote":"Supplies the extreme-eigenvalue results for spiked Fisher matrices that support the high-dimensional null and alternative limits."}],"review_version":1}