{"id":"63ad9815-8a48-4982-985a-dbf72978521a","arxiv_id":"1908.00531","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The worst-case first-step error of cyclic alternating projections equals a maximum of a product of adjacent off-diagonal matrix entries, with exact value f_3(c)=4c^2 for c<=1/4 and f_3(c)=c for c>=1/4, and optimal slope 2(n-1)sin^2(pi/(2n)) near c=1 for all n.","lead":"This paper finds the exact worst-case error after one round of alternating projections onto three subspaces, controlled by the angle parameter c. It recasts the problem as a matrix optimization, solves it for three subspaces, and identifies the optimal first-order behavior for any number of subspaces.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader located the only unproved external ingredient, the path-Laplacian spectral gap, but that fact is standard and can be verified in one line, so it is not a load-bearing risk. I checked the places where a hidden gap could hide: the c=1 endpoint in Proposition 1.2, the cancellation leading to ||u_n|| <= K||u_1|| in the proof of Theorem 2.1, the positivity of D_n-epsilon in Theorem 2.6, the monotonicity and Taylor expansions in Theorem 2.7, and the existence of optimal matrices with nonnegative adjacent entries used in Theorem 2.4. No circularity, unstated compactness failure, or sign error was found. The non-explicit constant b_tilde_n in the lower bound is a stated limitation but does not affect the first-order optimality claim. The verdict ACCEPT with high confidence remains appropriate.","tokens_in":21094,"tokens_out":18523,"duration_ms":187797,"concrete_test":"Verify the spectral input directly: for L the path Laplacian defined in §3.11 and v_j = cos((2j-1)pi/(2n)) for j=1,...,n, compute (Lv)_j = 4 sin^2(pi/(2n)) v_j. This confirms lambda_2(P_n)=4 sin^2(pi/(2n)) and closes the only unproved lemma used in the near-c=1 slope of the bounds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. The central equivalence (Theorem 2.1), the exact evaluation of f_3 (Theorem 2.2), and the small-c exact value (Theorem 2.3) are all supported by complete arguments. The reduction via Proposition 1.2 correctly handles the c=1 case separately, the Gram-matrix argument in §3.3 has no hidden boundedness assumption, and the lower-bound Taylor expansion in §3.12 is internally consistent. The only external input is the standard spectral gap lambda_2(P_n)=4 sin^2(pi/(2n)) of the path Laplacian used in Lemma 3.1 and Theorem 2.7; this is an elementary, easily verified fact and does not threaten the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the worst-case first-step error of the method of cyclic alternating projections for n closed subspaces of a Hilbert space. It defines f_n(c) as the supremum of ||P_n...P_2P_1 - P_0|| over all n-tuples of subspaces whose Friedrichs number is at most c, and establishes that this supremum is exactly the matrix maximum of |a_{12}a_{23}...a_{n-1,n}| over Hermitian positive semidefinite matrices with unit diagonal and upper bound A <= (1+(n-1)c)I (Theorem 2.1). Using this equivalence, the paper obtains the exact value f_3(c) = 4c^2 for c in [0,1/4] and f_3(c)=c for c in [1/4,1] (Theorem 2.2), the exact small-c value f_n(c)=(n-1)^{n-1}c^{n-1} for c in [0,1/(n-1)^2] (Theorem 2.3), concavity of f_n^{1/(n-1)} and a functional equation (Theorems 2.4 and 2.5), and two-sided quadratic bounds near c=1 whose linear coefficient a_n = 2(n-1)sin^2(pi/(2n)) is sharp (Theorems 2.6 and 2.7).","tokens_in":21182,"tokens_out":14102,"duration_ms":140677,"significance":"If the results hold, the paper gives the first exact solution of the first-step worst-case problem for three subspaces and a sharp asymptotic first-order coefficient for all n, improving on the earlier bounds of Badea-Grivaux-Muller and Badea-Seifert. The matrix reformulation in Theorem 2.1 is elegant and likely to be a useful tool for further study of f_n for n>=4. The central proofs are complete and self-contained: the equivalence is proved in both directions, exact optima are exhibited with explicit matrices, and the near-c=1 lower bound is derived from a concrete family of one-dimensional subspaces. External results are used only as comparisons, and the paper ships no unexplained numerical or fitting steps. Overall this is a solid, well-motivated contribution to the theory of alternating projections.","major_comments":[],"minor_comments":[{"comment":"The proof invokes the spectral gap lambda_2(P_n)=4 sin^2(pi/(2n)) of the path Laplacian as 'well-known' without proof or a precise reference; since this fact is exactly what determines the sharp slope a_n, please add a short derivation or a concrete citation.","section":"Section 3.11, Lemma 3.1"},{"comment":"The passage from the principal-minor criterion to the displayed inequalities for x and y is compressed with 'one can easily check'; spelling out the 2x2 and 3x3 determinant conditions would make the proof easier to verify.","section":"Section 3.5, proof of Theorem 2.2"},{"comment":"The manuscript contains numerous typographical and spacing artifacts (for example, 'metho d', 'orthog onal', 'Frie drichs', and 'n /greaterorequalslant4'); a careful proofreading pass is needed before publication.","section":"Throughout"}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper, worth taking seriously. The genuinely new content is the matrix formulation (Theorem 2.1), the exact formula for f_3 (Theorem 2.2), the exact small-c value (Theorem 2.3), and the sharp first-order coefficient as c approaches 1. The earlier literature supplied upper bounds but not exact values or matching lower bounds, so this is a real advance in the subfield.\n\nThe hinge is Theorem 2.1, and the proof is solid in both directions. The direction from matrices to subspaces via Gram matrices is clean, and the reverse direction, using a clever chain of inner products to control the norm of the projection product, works. The reduction to real symmetric persymmetric matrices (Lemma 2.1) is also sound and simplifies the optimization.\n\nThe exact evaluation of f_3 is the centerpiece. The proof reduces to maximizing x^2 over a 3-by-3 matrix with one parameter y, and then uses the principal-minor criterion for positive semidefiniteness. The phrase “one can easily check” is doing a bit of work here; the inequalities are correct, but a referee might ask for a few more lines. This is a minor exposition issue, not a correctness issue.\n\nFor n>=4 the paper does not find f_n, and the author says so explicitly. The n=4 case is reduced to maximizing x^2 y with two free parameters, which is left open. That is an honest limitation, not a flaw.\n\nThe near-c=1 analysis rests on the spectral gap of the path Laplacian, lambda_2 = 4 sin^2(pi/(2n)). This is invoked as well-known and used in Lemma 3.1 and Theorem 2.7. It is an elementary fact and not a burden on the argument. The lower bound in Theorem 2.7 has a nonexplicit constant, so it proves optimality of the linear coefficient only up to quadratic error. That is fine, and the explicit upper bound matches the same linear coefficient.\n\nNo circularity, no fitting, no suspicious citation pattern. The paper is self-contained modulo standard facts, and the external references are used as comparisons, not as load-bearing inputs.\n\nWho is this for? Specialists in alternating projections, operator theory, and numerical analysis who want the exact worst-case behavior of the first step, not the user looking for a new algorithm. It deserves a serious referee and, if the minor computations are verified, publication. I would be happy to see it in print after the author fills in the terse principal-minor verification and maybe says a bit more about the nonexplicit lower-bound constant.","headline":"A clean, genuinely new result on the optimal first-step error in cyclic alternating projections, with a complete proof of the exact f_3 and a sharp first-order expansion near c=1.","tokens_in":21717,"tokens_out":1469,"would_cite":true,"duration_ms":18491,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46C07","47B15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The worst-case first error in cyclic projections equals a matrix product maximum, and the three-subspace case is solved exactly.","keywords":["cyclic alternating projections","Friedrichs number","Dixmier number","orthogonal projections","worst-case error","Hermitian matrices","path graph Laplacian","rate of convergence"],"falsifier":"For n=4, carry out a high-precision semidefinite optimization of |a12 a23 a34| over the reduced class H'_4(1+3c) for c=0.99, 0.999, 0.9999. The theorem predicts f_4(c)=1-a_4(1-c)+O((1-c)^2) with a_4=6 $sin^{2}$(π/8)≈0.87868; a fitted slope differing from a_4 beyond numerical error, or any subspace family achieving first-step error larger than the upper bound, would refute the claim. Alternatively, verify the unproved eigenvalue formula λ_2(P_4)=4 $sin^{2}$(π/8) directly, since Theorem 2.7 chooses the perturbation along that eigenvector.","tokens_in":20884,"feed_emoji":"🔁","tokens_out":9839,"duration_ms":89626,"temperature":0.7,"pith_summary":"This paper asks how large the first step of the method of cyclic alternating projections can be, in the worst case, when a cosine-like measure of the mutual position of n closed subspaces — the Friedrichs number — is bounded by c. It proves that the worst-case error is exactly the maximum of the product |a12 a23 ... a_{n-1,n}| over n×n Hermitian matrices with unit diagonal and 0 ≤ A ≤ (1+(n-1)c)I. From this equivalence the paper solves the three-subspace case completely: f_3(c)=$4c^{2}$ for c≤1/4 and f_3(c)=c for c≥1/4. For general n it obtains two-sided bounds whose linear coefficient near c=1 is 2(n-1) $sin^{2}$(π/(2n)), so the first-order behavior is sharp even though the exact f_n for n≥4 remains open. Knowing f_n matters because it converts directly into a bound on how quickly the alternating projection sequence approaches the intersection.","feed_headline":"Cyclic projections' worst first step equals one matrix product","feed_subtitle":"For three subspaces the bound is exact; for any n the slope near c=1 is pinned down.","key_machinery":"The carrying object is the Hermitian matrix A with unit diagonal, understood as the Gram matrix of unit vectors v_1,...,v_n; the quantity Π(A)=|a_{12}a_{23}...a_{n-1,n}| is exactly the norm of the product of orthogonal projections onto the lines spanned by those vectors. The condition 0≤A≤(1+(n-1)c)I is shown to be equivalent to the Dixmier/Friedrichs number constraint, reducing an infinite-dimensional operator problem to a finite-dimensional semidefinite one. For the near-c=1 estimates the decisive inequality is a comparison of quadratic forms: sum_{i<j}(x_i-x_j)^2 ≤ D_n sum_{i=1}^{n-1}(x_i-x_{i+1})^2, whose sharp constant is controlled by the second eigenvalue λ_2 of the path graph Laplacian, λ_2(P_n)=4 $sin^{2}$(π/(2n)); together with the identity ||x_i-x_{i+1}||^2=||x_i||^2-||x_{i+1}||^2 for projected vectors, this yields the upper bound, and a perturbation along the corresponding eigenvector yields the matching lower bound.","core_discovery":"The paper's central claim is Theorem 2.1: f_n(c) equals max{|a12 a23 ... a_{n-1,n}|} over Hermitian matrices A=(a_{ij}) with a_{ii}=1 and 0≤A≤(1+(n-1)c)I. The proof identifies such matrices with Gram matrices of unit vectors; the product of the first superdiagonal entries is exactly the norm of the product of projections onto the one-dimensional subspaces spanned by those vectors, and the upper/lower constraint on A encodes the condition c_D≤c. The paper then derives f_3(c)=$4c^{2}$ on [0,1/4] and f_3(c)=c on [1/4,1], with explicit optimal matrices, and shows that for c≤(n-1)^{-2} the exact formula f_n(c)=(n-1)^{n-1}$c^{{n-1}}$ holds. For all n it proves 1 - a_n(1-c) - \\tilde b_n(1-c)^2 ≤ f_n(c) ≤ 1 - a_n(1-c) + b_n(1-c)^2, with a_n = 2(n-1) $sin^{2}$(π/(2n)) and b_n = 6(n-1)^2 $sin^{4}$(π/(2n)).","pith_inferences":["The matrix formulation turns f_n into a finite-dimensional semidefinite optimization problem; for n=4 one could run a numerical SDP solver to conjecture a closed form for f_4 and to test the functional equation numerically.","The appearance of the path graph Laplacian suggests that the second-order coefficient of f_n near c=1 may be expressible through other Laplacian eigenvalues, such as λ_3, which would be a natural extension the paper does not pursue.","Question 1 in the paper — whether f_n(c) ≤ 1 - a_n(1-c) for all c — is testable for n=3 using the exact f_3, since f_3(c)=c equals the linear bound; testing n=4 numerically would give evidence about whether the linear bound holds universally.","The functional equation gives a self-similar structure: iterating c ↦ 1/((n-1)^2 c) from the explicitly known small-c interval could generate many further points of f_n, potentially enough to pin down f_4 if combined with concavity and continuity."],"forward_implications":["For three subspaces, the exact worst-case first-step error is known in closed form: 4c^2 for c≤1/4 and c for c≥1/4, so sharp three-iteration bounds become available.","For every n, f_n(c)=(n-1)^{n-1}c^{n-1} on [0,(n-1)^{-2}], giving exact small-Friedrichs-number behavior and an explicit optimal matrix in the reduced class.","For c close to 1, the first-order term of f_n(c) is exactly 2(n-1) sin^2(π/(2n))(1-c), so the previous square-root upper bounds are not first-order optimal; the new upper bound matches the lower bound to first order.","The functions f_n^{1/(n-1)} are concave, hence f_n is continuous, and f_n obeys a scaling relation f_n(1/((n-1)^2 c)) = f_n(c)/((n-1)^{n-1}c^{n-1}); these structural properties constrain any future exact solution for n≥4.","The optimal first diagonal (a12,a23,...,a_{n-1,n}) is unique whenever c>0, so the maximization problem has a well-defined answer even though f_n for n≥4 remains open."],"supporting_citations":[{"why":"Introduces the Friedrichs number of n subspaces, the quantity whose threshold c defines f_n.","marker":"[2]"},{"why":"Supplies the Dixmier-number formula and the convergence-rate estimates for cyclic alternating projections that the paper's worst-case functions are built to sharpen.","marker":"[3]"},{"why":"Gives an earlier upper bound for f_n that the paper's new upper bound improves near c=1.","marker":"[4]"},{"why":"Provides the two-subspace facts used in the introduction, including the equality c_F(H1,H2)=||P1P2-P0|| that fixes f_2(c)=c.","marker":"[6]"},{"why":"Establishes convergence of the cyclic alternating projections scheme whose first-step error f_n measures.","marker":"[7]"},{"why":"Proves the sharp two-subspace error identity that motivates studying worst-case first-step error as a function of the Friedrichs number.","marker":"[8]"}],"fun_headline_variants":["Exact worst-case first step in cyclic alternating projections","One matrix product rules cyclic projections' worst first step","Cyclic projections: worst first step now exact for n=3","Hermitian matrices settle first-step error in cyclic projections"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The near-c=1 analysis stands on the unproved spectral fact that the path graph's second-smallest Laplacian eigenvalue is 4 $sin^{2}$(π/(2n)); if that number were different, the claimed slope a_n in both the upper and lower bounds would change.","fun_headline_variants_meta":{"raw":{"variants":["Exact worst-case first step in cyclic alternating projections","One matrix product rules cyclic projections' worst first step","Cyclic projections: worst first step now exact for n=3","Hermitian matrices settle first-step error in cyclic projections"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000964,"raw_usage":{"total_tokens":4220,"prompt_tokens":1175,"completion_tokens":3045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":791,"completion_tokens_details":{"reasoning_tokens":2979}},"tokens_in":791,"tokens_out":3045,"duration_ms":23682,"temperature":1.0,"reasoning_tokens":2979,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:49:36.348516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For n=4, carry out a high-precision semidefinite optimization of |a12 a23 a34| over the reduced class H'_4(1+3c) for c=0.99, 0.999, 0.9999. The theorem predicts f_4(c)=1-a_4(1-c)+O((1-c)^2) with a_4=6 $sin^{2}$(π/8)≈0.87868; a fitted slope differing from a_4 beyond numerical error, or any subspace family achieving first-step error larger than the upper bound, would refute the claim. Alternatively, verify the unproved eigenvalue formula λ_2(P_4)=4 $sin^{2}$(π/8) directly, since Theorem 2.7 chooses the perturbation along that eigenvector.","supporting_citations":[{"cited_title":"Badea, S","cited_arxiv_id":null,"evidence_quote":"Introduces the Friedrichs number of n subspaces, the quantity whose threshold c defines f_n."},{"cited_title":"Badea, S","cited_arxiv_id":null,"evidence_quote":"Supplies the Dixmier-number formula and the convergence-rate estimates for cyclic alternating projections that the paper's worst-case functions are built to sharpen."},{"cited_title":"Badea, D","cited_arxiv_id":null,"evidence_quote":"Gives an earlier upper bound for f_n that the paper's new upper bound improves near c=1."},{"cited_title":"Deutsch, The angle between subspaces of a Hilbert space","cited_arxiv_id":null,"evidence_quote":"Provides the two-subspace facts used in the introduction, including the equality c_F(H1,H2)=||P1P2-P0|| that fixes f_2(c)=c."},{"cited_title":"Halperin, The product of projection operators , Acta Sci","cited_arxiv_id":null,"evidence_quote":"Establishes convergence of the cyclic alternating projections scheme whose first-step error f_n measures."},{"cited_title":"Kayalar, H","cited_arxiv_id":null,"evidence_quote":"Proves the sharp two-subspace error identity that motivates studying worst-case first-step error as a function of the Friedrichs number."}],"review_version":1}