{"id":"4948d6d1-9907-48a8-946b-115359230bbf","arxiv_id":"1908.03544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors extend the Slepian-Bangs formula to non-circular complex elliptical symmetric distributions and prove that Gaussian assumptions do not always maximize the stochastic Cramér-Rao bound.","lead":"This paper derives the Slepian-Bangs formula for non-circular complex elliptical symmetric distributions, giving Cramér-Rao bounds for parameter estimation in these models. It also shows that the Gaussian distribution does not always yield the largest bound, overturning a common assumption in signal processing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unproved diagonal-form reduction in the stochastic representation theorem is the key gap; the FIM formula rests on it.","rationale":"The reader's weakest assumption correctly identifies the unproved diagonal-form reduction as the principal risk: Section III's FIM derivation depends on the augmented stochastic representation, so any failure of Result 1 would invalidate the central Slepian-Bangs extension. The concern is not that the statement is false—standard factorization arguments indicate it is true—but that the submitted proof omits the necessary justification. I therefore keep the CONDITIONAL verdict: the authors should either supply the Takagi-based construction or cite a prior theorem. I also note that Section V proves only the comparison formula, not the existence of a generator with ξ2<1; that is a separate, smaller gap that does not change the verdict.","tokens_in":8451,"tokens_out":17737,"duration_ms":173363,"concrete_test":"Independently verify the diagonal canonical form: for arbitrary positive definite Σ and complex symmetric Ω such that Γ = [[Σ,Ω],[Ω*,Σ*]] is positive definite, set A = Σ^{1/2}, C = A^{-1}ΩA^{-T}, apply Takagi factorization C = U D U^T with U unitary and D real nonnegative diagonal, define A' = A U^* and Δκ = D. Check that Σ = A'A'^H, Ω = A'ΔκA'^T, and that the matrix [[A',0],[0,A'*]] [[I,Δκ],[Δκ,I]] [[A',0],[0,A'*]]^H equals Γ. If this construction succeeds, the stochastic representation (17) is valid for all NC-CES distributions and the Section II proof gap is closed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the proof of Result 1 (Section II). The paper needs every NC-CES vector to admit the stochastic representation z = μ + R A[Δ1 u + Δ2 u*] with Δ1, Δ2 real diagonal; this representation is then used in Section III to compute the FIM via the augmented representation ~z = ~μ + R ~Γ^{1/2}~u and the moment identities (22)-(28). But the proof simply states that solutions of (15) 'can be looked for' in real diagonal form, without showing that this restriction is without loss of generality. For arbitrary scatter Σ and pseudo-scatter Ω, the existence of such a diagonal canonical form requires a simultaneous congruence argument (e.g., Takagi factorization of A^{-1}ΩA^{-T}); the supplement supplies no such argument. The surrounding text also contains typos that obscure the step: Eq. (13) should have ΨΦ^T + ΦΨ^T, Eq. (15) has missing primes, and the eigenvalue decomposition before (18) lists both diagonal blocks as I+Δκ rather than I±Δκ. None of these is fatal by itself, but together they leave the foundational representation theorem unjustified as written. If the diagonal reduction failed, the FIM formula of Result 2 would not be established for all NC-CES distributions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a detailed-proof supplement to the authors' accepted letter [1]. It gives proofs of six results: a stochastic representation theorem for non-circular complex elliptically symmetric (NC-CES) random vectors (Result 1); the Slepian-Bangs formula, i.e., the Fisher information matrix (FIM), for NC-CES distributions (Result 2); the moment identity E[Q phi(Q)] = -M and the inequality M^2 <= M xi_1 (Eq. (9) of [1]); the claim that the Gaussian distribution is not always the least favorable for the stochastic CRB (Result 4); and closed-form stochastic CRBs for noisy mixture models in circular and non-circular cases (Results 5 and 6). The central derivation in Section III expresses the FIM in terms of the augmented covariance matrix and the density-generator moments xi_1 and xi_2, following the template of Besson and Abramovich for circular CES. The paper is self-contained modulo the statements of results in [1], and the algebraic derivations are presented in considerable detail.","tokens_in":8664,"tokens_out":18123,"duration_ms":170661,"significance":"If the representation theorem is properly justified, the main formula is a genuine and useful unification: it contains the non-circular Gaussian and the circular CES cases as special cases, and it is stated in a form directly usable for CRB computations. The proof of Result 4 is also valuable because it gives a concrete parameter regime (xi_2 < 1) in which the common 'Gaussian is least favorable' intuition fails. The supplement is careful in its FIM derivation: the regularity condition (19), the moment identities (22)-(28), and the use of Lemma 1 for the fourth-order moments are internally consistent and appear correct. The main weakness is not the FIM computation itself but the foundational stochastic representation theorem in Section II, whose proof is incomplete as written.","major_comments":[{"comment":"The proof of Result 1 does not justify the reduction to real diagonal matrices (Delta_1, Delta_2). The text states that solutions of (15) 'can be looked for' in real-valued diagonal form and then proceeds to (16), but no argument is given that this restriction is without loss of generality for arbitrary scatter matrix A A^H and pseudo-scatter matrix A Delta_kappa A^T. This is load-bearing because the FIM derivation in Section III uses the canonical augmented representation ~z = ~mu + R ~Gamma^{1/2} ~u, which is obtained from the diagonal choice. A rigorous proof should show either that any solution of (14) yields the same augmented covariance matrix ~Gamma (so the distribution depends only on ~Gamma) or provide a direct simultaneous congruence argument. It should also show that the scalar equations (16) are solvable, i.e., that |kappa_i| <= 1, which follows from the positive semidefiniteness of [[I, Delta_kappa], [Delta_kappa, I]]. As written, the representation theorem is not established for all NC-CES distributions.","section":"Section II, Eq. (15)"},{"comment":"The algebra in the reduction contains several errors that obscure the step. Eq. (13) should read Psi Phi^T + Phi Psi^T, not Psi Phi^T + Psi Phi^T. The change of variables should be Psi' = A^{-1} Psi and Phi' = A^{-1} Phi, not Psi' = A Psi. Eq. (15) is missing primes and should be I = Psi' Psi'^H + Phi' Phi'^H and Delta_kappa = Psi' Phi'^T + Phi' Psi'^T. Finally, the eigenvalue decomposition before Eq. (18) should have blocks I + Delta_kappa and I - Delta_kappa, not two copies of I + Delta_kappa. These are presentation issues in a load-bearing part of the proof, and they should be corrected together with the missing justification in the previous comment.","section":"Section II, Eqs. (13)-(18)"}],"minor_comments":[{"comment":"In the final trace formula, the matrix H is used without being defined. From the preceding line, H should be R_s A_theta^H Sigma^{-1} A_theta R_s (or an equivalent simplification); please define it explicitly.","section":"Section VI, Eq. (52)"},{"comment":"The proof of Result 6 is mostly an appeal to [7, th. 1] with substitutions, and the 'key form expression' for Pi_V^perp is stated without derivation. For a supplement labelled 'detailed proofs', a fuller derivation of the non-circular noisy-mixture CRB would be helpful, even if it follows the template of Result 5.","section":"Section VII"},{"comment":"There are several typographical slips: 'cheek' should be 'check' (Section III), 'manducation' should be 'manipulation' (Section VI), and 'foll ows' should be 'follows'. These do not affect the mathematics but should be cleaned up.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"The gap in Section II is fixable and does not appear to be fatal: the diagonal representation is very likely correct, and the FIM derivation in Section III is sound conditional on it. However, because the representation theorem is the foundation for the main Slepian-Bangs formula, the authors should be required to supply a complete proof of the diagonal reduction before the supplement is published. The dependence on the authors' own prior work [2] and [7] is transparent and not, in my view, a reason for concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline is that this supplement does something real: it extends the Slepian-Bangs formula from circular to non-circular complex elliptical symmetric (NC-CES) distributions and shows the Gaussian is not always the least favorable distribution for Cramér-Rao bounds. That second point is a genuine correction to a common belief, and the criterion (ξ2 < 1) is concrete.\n\nThe FIM derivation in Section III is the strong part. The augmented-representation moment computations are careful, Lemma 1 is a clean device, and the steps leading to the closed-form FIM are checkable. The noisy-mixture SCRB results in Sections V and VI are useful and reduce properly to the known Gaussian forms. The citation pattern is fine: the authors draw on their own earlier work for a lemma and a proof template, but those are published and reproducible, so no issue.\n\nThe soft spot is Section II, the proof of the stochastic representation theorem. The paper asserts we can look for solutions in real diagonal form without proving it. That is exactly Takagi's factorization for the complex symmetric matrix A^{-1}ΩA^{-T}, and because the augmented covariance matrix is positive semidefinite, the Takagi singular values are at most one, so the diagonal Δ1, Δ2 exist. A one-line citation to Takagi would close the gap. As written, a load-bearing justification is missing. The typos in that section make it worse: the change of variables should be Ψ′=A^{-1}Ψ, not AΨ, and the eigenvalue decomposition lists both diagonal blocks as I+Δκ rather than I±Δκ. These are minor individually, but they obscure the step. Section VI is compressed—the phrase 'after some algebraic manducation' is not enlightening—but the final expressions agree with the Gaussian limit, so I think the algebra is sound.\n\nMy bottom line: the central formula is very likely correct, and the missing piece is presentational, not a hidden fatal assumption. This deserves a serious referee, with a request to add the Takagi argument and fix the typos. I would cite it if working on non-circular estimation.","headline":"This supplement genuinely extends the Slepian-Bangs formula to non-circular CES and corrects the Gaussian-worst-case belief, but the stochastic representation proof skips a Takagi step and has typos that need fixing.","tokens_in":9210,"tokens_out":9106,"would_cite":true,"duration_ms":86540,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper extends the Slepian-Bangs formula to non-circular complex elliptical symmetric distributions, giving a closed-form Fisher information matrix and showing that Gaussian noise is not always the worst case.","keywords":["Slepian-Bangs formula","Cramer-Rao bound","non-circular complex elliptical symmetric distributions","Fisher information matrix","stochastic representation","density generator","array processing"],"falsifier":"Construct a non-circular CES density whose scatter matrix $\\Sigma$ and pseudo-scatter matrix $\\Omega$ are not simultaneously diagonalizable in the required real-diagonal sense, then compare its Fisher information with the formula of Result 2; if the formula fails, the representation theorem is the point of collapse. Alternatively, simulate a heavy-tailed non-circular sample with $\\xi_2<1$ and check the prediction that its stochastic CRB exceeds the Gaussian SCRB.","tokens_in":8220,"feed_emoji":"📊","tokens_out":6704,"duration_ms":60154,"temperature":0.7,"pith_summary":"This paper proves a new Slepian-Bangs formula for non-circular complex elliptical symmetric (NC-CES) distributions: the Fisher information matrix has a closed form determined entirely by the augmented covariance matrix and two scalar moments of the radial density generator. The proof rests on a stochastic representation that writes every NC-CES vector as an affine transformation of a spherical complex vector, so all distribution-specific effects reduce to expectations over a scalar radial variable. The formula covers non-circular Gaussian and circular CES distributions as special cases. A direct consequence is that Gaussian noise is not universally the least favorable: when the generator moment $\\xi_2$ is below 1, the NC-CES Fisher information is lower than the Gaussian information, so the Gaussian CRB can be optimistic. The paper also derives closed-form stochastic CRBs for the noisy mixture model, covering direction-of-arrival and source parameter estimation.","feed_headline":"Gaussian noise is not always the worst case for estimation bounds","feed_subtitle":"A closed-form Fisher information formula for non-circular elliptical noise shows when heavy tails push the CRB higher.","key_machinery":"The load-bearing object is the stochastic representation $z = \\mu + R A[\\Delta_1 u + \\Delta_2 u^*]$, where $u$ is uniformly distributed on the complex unit sphere, $R$ is a nonnegative radial variable, and $\\Delta_1, \\Delta_2$ are real diagonal matrices satisfying $\\Delta_1^2+\\Delta_2^2=I$ and $2\\Delta_1\\Delta_2=\\Delta_\\kappa$. This converts expectations with respect to the full density into expectations over the scalar $Q$, whose distribution is carried by $\\xi_1 = E[Q\\varphi(Q)^2]/M$ and $\\xi_2 = E[Q^2\\varphi(Q)^2]/[M(M+1)]$. Lemma 1, which computes $E[(\\tilde{y}^H\\tilde{A}\\tilde{y})(\\tilde{y}^H\\tilde{B}\\tilde{y})]$ for an augmented complex Gaussian vector, supplies the combinatorial coefficients in the Fisher information matrix. Together these pieces turn the Slepian-Bangs formula into a pair of moment computations.","core_discovery":"The central discovery is Result 2: for NC-CES observations with augmented covariance matrix $\\tilde{\\Gamma}$, the Fisher information matrix is given in closed form as a combination of terms involving $\\tilde{\\Gamma}^{-1}\\tilde{\\Gamma}_k$, $\\tilde{\\Gamma}^{-1}\\tilde{\\Gamma}_l$, and the mean derivatives, with coefficients fixed by the two density-generator moments $\\xi_1$ and $\\xi_2$. This generalizes the classical Slepian-Bangs formula, previously known only for Gaussian and circular CES cases. Result 4 then shows that the difference between the NC-CES and non-circular Gaussian information along covariance parameters is proportional to $\\xi_2-1$ times a positive-definite quadratic form, so whenever $\\xi_2<1$ the non-Gaussian information is strictly smaller and the corresponding stochastic CRB is larger. The proofs in Sections III through VI establish these statements for a general parameterization and then specialize them to the noisy mixture, yielding explicit SCRB expressions for the parameters of interest.","pith_inferences":["Because the FIM depends on the density only through $\\xi_1$ and $\\xi_2$, the same formula could support a plug-in CRB: estimate those two moments from data and compute a bound without assuming a full parametric generator.","The proof strategy for the noisy mixture may transfer to other array-processing models beyond the near-field and far-field DOA examples named in the supplement, such as models with structured or partially correlated sources.","If a real heavy-tailed non-circular generator has $\\xi_2<1$, then designs based on the Gaussian CRB will systematically underestimate achievable variance, suggesting a practical diagnostic: estimate $\\xi_2$ before trusting Gaussian-based bounds."],"forward_implications":["For any NC-CES distribution, the stochastic CRB can be computed from the augmented covariance matrix and the two generator moments $\\xi_1$ and $\\xi_2$, without specifying the full density.","The non-circular complex Gaussian and the circular CES bounds are both special cases of the new formula, so the result unifies previously separate derivations.","When $\\xi_2<1$, the NC-CES Fisher information is strictly smaller than the Gaussian information along covariance-parameter directions, meaning the Gaussian distribution does not always give the largest SCRB.","For the noisy mixture model, the supplement derives closed-form SCRBs for the parameters of interest, covering direction-of-arrival and source parameter estimation for both circular and non-circular observations.","The formulas hold for arbitrary parameterizations of the spatial signature matrix $A_\\theta$, including near-field and far-field DOA models with scalar or vector sensors."],"supporting_citations":[{"why":"Defines the Slepian-Bangs formula and the Results whose detailed proofs this supplement provides.","marker":"[1]"},{"why":"Supplies the augmented Gaussian moment identity used in Lemma 1.","marker":"[2]"},{"why":"Defines generalized complex elliptical distributions, the class underlying the stochastic representation.","marker":"[3]"},{"why":"Provides the stochastic representation and the moment identities $E(uu^H)=I/M$ and $E(Q\\varphi(Q))=-M$.","marker":"[4]"},{"why":"Supplies the proof strategy for the Fisher information matrix of elliptical distributions followed in Section III.","marker":"[5]"},{"why":"Gives the projection-matrix derivation for the circular stochastic CRB that Result 5 extends.","marker":"[6]"},{"why":"Provides the direction-of-arrival CRB argument for rectilinear sources that Result 6 builds on.","marker":"[7]"}],"fun_headline_variants":["Closed-form Fisher info for non-circular elliptical noise","Gaussian isn't always the worst-case noise for estimation","Non-circular noise: exact CRB formulas derived","When heavy tails push the CRB above Gaussian's bound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stochastic representation theorem assumes every NC-CES vector can be written with real diagonal matrices $\\Delta_1$ and $\\Delta_2$, and the proof asserts this diagonal form can be chosen without showing that it holds for arbitrary scatter and pseudo-scatter matrices.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form Fisher info for non-circular elliptical noise","Gaussian isn't always the worst-case noise for estimation","Non-circular noise: exact CRB formulas derived","When heavy tails push the CRB above Gaussian's bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000864,"raw_usage":{"total_tokens":3709,"prompt_tokens":868,"completion_tokens":2841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2776}},"tokens_in":484,"tokens_out":2841,"duration_ms":22750,"temperature":1.0,"reasoning_tokens":2776,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:11:02.190259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a non-circular CES density whose scatter matrix $\\Sigma$ and pseudo-scatter matrix $\\Omega$ are not simultaneously diagonalizable in the required real-diagonal sense, then compare its Fisher information with the formula of Result 2; if the formula fails, the representation theorem is the point of collapse. Alternatively, simulate a heavy-tailed non-circular sample with $\\xi_2<1$ and check the prediction that its stochastic CRB exceeds the Gaussian SCRB.","supporting_citations":[{"cited_title":"Abeida and J.P","cited_arxiv_id":null,"evidence_quote":"Defines the Slepian-Bangs formula and the Results whose detailed proofs this supplement provides."},{"cited_title":"Abeida and J.P","cited_arxiv_id":null,"evidence_quote":"Supplies the augmented Gaussian moment identity used in Lemma 1."},{"cited_title":"Ollila and V","cited_arxiv_id":null,"evidence_quote":"Defines generalized complex elliptical distributions, the class underlying the stochastic representation."},{"cited_title":"Ollila, D","cited_arxiv_id":null,"evidence_quote":"Provides the stochastic representation and the moment identities $E(uu^H)=I/M$ and $E(Q\\varphi(Q))=-M$."},{"cited_title":"Besson and Y","cited_arxiv_id":null,"evidence_quote":"Supplies the proof strategy for the Fisher information matrix of elliptical distributions followed in Section III."},{"cited_title":"The stochas tic CRB for array processing: A textbook derivation,","cited_arxiv_id":null,"evidence_quote":"Gives the projection-matrix derivation for the circular stochastic CRB that Result 5 extends."},{"cited_title":"Abeida and J.P","cited_arxiv_id":null,"evidence_quote":"Provides the direction-of-arrival CRB argument for rectilinear sources that Result 6 builds on."}],"review_version":1}