{"id":"87455a43-c252-4be3-884c-4a3069b263af","arxiv_id":"2411.18502","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"Isometry pursuit is a convex normalization plus multitask basis pursuit procedure for selecting near-orthonormal column subsets, but its theoretical justifications contain proof gaps and an inconsistent normalization formula.","lead":"Scientists propose a convex method, isometry pursuit, that rescales the columns of a wide data matrix and then uses multitask basis pursuit to pick a small subset that is as close to orthonormal as possible. A two-stage version beats greedy search on three small benchmark datasets, which matters for interpretable coordinate selection and diversification tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 18's q_c is not a symmetric normalization: it grows unbounded with ||v|| and the unit sphere is its global minimum, so the printed algorithm does not implement the advertised length penalty.","rationale":"The central assertion of the paper is that a convex program on a suitably normalized matrix selects unit-norm, mutually orthogonal columns. The only concrete normalization supplied is q_c in Eq. 18, and it violates the defining condition of Definition 3. This is not a subtle sign error: q_c is the ground-truth loss g(·,c) from Eq. 8, which is minimized at unit norm, whereas Definition 3 requires the unit sphere to be the argmax. Because Algorithm 1 multiplies each column by q_c, the printed method actively prefers long vectors—for example, 10e1 is scaled by roughly 4.05e3 while e1 is scaled by 1—so the convex objective's support is driven toward long, non-isometric columns rather than the intended orthonormal submatrix. This single defect explains why Proposition 4 cannot hold as stated for the printed q_c: its premise, a normalization satisfying Definition 3, is not met by the paper's own choice, and the proof's QR step also needs patching. I note that the reader's secondary criticism of Proposition 4—that row norms of R^{-1}Q^T are not row norms of R^{-1}—does not survive contact: β=R^{-1}Q^T has rows (R^{-1})_{p.}Q^T, and right multiplication by the unitary Q^T preserves each row's Euclidean norm, so ∥β∥_{1,2}=∥R^{-1}∥_{1,2}. The QR argument is therefore potentially repairable; the normalization equation is not. The two-stage empirical results may survive a corrected formula, but as submitted the central object is mis-defined, so the reader's REJECT is appropriate.","tokens_in":12531,"tokens_out":14242,"duration_ms":137339,"concrete_test":"Run Algorithm 1 exactly as printed on the synthetic matrix X=[e1, e2, 10e1] in D=2 with c=1. The advertised behavior is support S={1,2}, selecting the two unit columns and pruning the long column. With Eq. 18, however, q_c(10e1)=(e^{10}+e^{0.1})/(2e)≈4.05e3 while q_c(e1)=1, so the third column of w(X,c) is enormous and the basis pursuit solution to I_2=w(X,c)β will put weight on it. If the returned support includes column 3 or excludes either of e1,e2, the printed normalization fails Definition 3. Separately, inspect the repository's implementation: if it uses 2e/(e^{||v||^c}+e^{||v||^{-c}}) rather than Eq. 18, that confirms Eq. 18 is a typo and the published definition must be corrected before the claims can be evaluated.","verdict_should_be":"REJECT","load_bearing_attack":"Definition 3 requires q_c to attain its global maximum exactly on the unit sphere. The concrete choice in Eq. 18 is q_c(v)=(e^{||v||^c}+e^{||v||^{-c}})/(2e). For any unit u, q_c(u)=1, while for t>1, q_c(tu)>1 and q_c(tu)→∞ as t→∞. Hence the unit sphere is the global minimum, not the maximum, and the argmax is empty. Since Algorithm 1 normalizes via n(v)=q_c(v)v (Eq. 14), long columns are amplified rather than shrunk, contradicting the stated goal that long and short candidate basis vectors are penalized. All downstream claims—Proposition 4, the two-stage algorithm, and the experiments—use w(X,c) built from this q_c, so the central object of the paper is mis-specified. A likely intended fix is q_c(v)=2e/(e^{||v||^c}+e^{||v||^{-c}}), but that is not what Eq. 18 says, and the proof of Proposition 4 would still need to be written out correctly; the QR step is repairable because unitary right multiplication preserves row norms, but the text's 'by Proposition 3' line is at best compressed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes isometry pursuit, a method for selecting approximately orthonormal D-column submatrices of a wide D×P matrix. The method first normalizes columns with a radially symmetric function q_c, then applies multitask basis pursuit to the normalized matrix to obtain a sparse set of candidate columns; a two-stage variant prunes with this convex program and then applies brute-force search over the reduced set. The authors state two theoretical results: invariance of the selected support under orthonormal changes of basis (Proposition 3) and a square-case guarantee that the multitask basis pursuit solution on a suitably normalized matrix is orthonormal (Proposition 4). Experiments on Iris, Wine, and Ethanol report lower ground-truth loss than greedy search.","tokens_in":12757,"tokens_out":10641,"duration_ms":98336,"significance":"If the method worked as stated, it would offer a convex alternative to greedy or brute-force selection of orthonormal column submatrices, with plausible applications to interpretable coordinate selection and diversification. The paper has clear strengths: the problem is well motivated, the code is publicly available, and the experimental protocol is reproducible. However, the central normalization in Eq. (18) contradicts its own definition, and the proofs of Propositions 3 and 4 contain invalid algebraic steps. Because the algorithm as printed does not implement the intended length penalty and the main theoretical claim is unsupported as proved, the paper's core contribution is not currently established.","major_comments":[{"comment":"The concrete normalization in Eq. (18) contradicts Definition 3. For q_c(v) = (e^{||v||^c} + e^{||v||^{-c}})/(2e), writing t = ||v||_2, the expression e^{t^c} + e^{t^{-c}} has its minimum at t = 1, so q_c(v) = 1 on the unit sphere and q_c(tu) > 1 for t > 1, with q_c(tu) tending to infinity as t tends to infinity. Thus the unit sphere is a global minimum, not the argmax required by Eq. (10), and the normalization n(v) = q_c(v)v in Eq. (14) amplifies long vectors instead of shrinking them. Since Algorithm 1 and Eq. (22) use w(X, c) built from this q_c, the printed algorithm does not implement the proposed method; all downstream theoretical and experimental statements using this normalization are affected. If a reciprocal formula was intended, the text must be corrected and the experiments re-run.","section":"Sec. 3.2, Eq. (18); Sec. 3.3, Algorithm 1"},{"comment":"The proof's chain 'by Proposition 3, l_c(X) = l_c(R) and so ||β||_{1,2} = ||R^{-1}||_{1,2}' is invalid. Proposition 3 concerns the support of the multitask basis pursuit solution under left multiplication by an orthonormal matrix; it does not equate the ground-truth loss l_c with the basis-pursuit objective. Moreover, for a square normalized design W = QR, the constraint I_D = Wβ gives β = R^{-1}Q^T, whose row norms are not in general the row norms of R^{-1}. Therefore the lower bound ||β_{d.}|| ≥ ||X_{.d}||^{-1} = 1 is not established, and Proposition 4 remains unproved.","section":"Sec. 6.2.2, proof of Proposition 4"},{"comment":"The proof of Proposition 6 contains an invalid equality: from I_D = UXβ, left multiplication by U^{-1} yields U^{-1} = Xβ, not I_D = XβU. A correct argument would introduce the variable γ = βU and use the unitary invariance of the ||·||_{1,2} row norm; as written, the displayed derivation does not prove the claimed invariance. Since Proposition 3 is invoked in the proof of Proposition 4, this error is load-bearing.","section":"Sec. 6.2.1, proof of Proposition 3"},{"comment":"The evaluation is partly circular: the ground-truth loss l_c in Eq. (6) is defined through g, and the normalization q_c in Eq. (18) is built from the same g; Figure 1 explicitly states that the two losses are equivalent when D = 1. Thus the method's prior is directly shaped to the metric used for benchmarking, and the final brute-force stage optimizes that same metric. This does not by itself invalidate the two-stage comparison, because greedy search also minimizes l_c, but it weakens the interpretation. To substantiate the claim of a general advantage, the authors should report results against a ground truth not constructed from g, for example ||X_{.S}^T X_{.S} - I_D||_F or the distance to the nearest orthonormal basis, and should compare the normalization's pruning quality against random or length-only pruning.","section":"Sec. 3.1–3.2 and Sec. 4 (experimental evaluation)"}],"minor_comments":[{"comment":"There are typos: 'helps identity' should be 'helps identify', and 'Jacoban matrices' should be 'Jacobian matrices'.","section":"Abstract and Sec. 4"},{"comment":"The statement 'arg min_{X.S ∈ R^{D×D}} bβ_c(X.S) is orthonormal' does not type-check, since bβ_c(X.S) returns a coefficient matrix rather than a scalar objective; please clarify what is being minimized and in which variable.","section":"Proposition 4 statement"},{"comment":"The table and text use l1 to denote the ground-truth loss l_c with c = 1; define this notation explicitly to avoid confusion with the ℓ1 norm.","section":"Table 1 and Sec. 4"},{"comment":"The caption lists '(P, D) = (4, 15) and (13, 18)', which is inconsistent with the P > D convention used elsewhere; swap the roles or correct the values.","section":"Figure 4 caption"},{"comment":"The statement writes ||βU|| without the ||·||_{1,2} subscript in the statement, although the proof uses it.","section":"Proposition 5"},{"comment":"The QR decomposition is introduced for '\\tilde X.S' but the proof then refers to X.S; clarify whether the decomposition is applied to the normalized matrix w(X, c)_{.S} and how the diagonal entries of R relate to the original column norms.","section":"Sec. 6.2.2"},{"comment":"The sentence 'suggest that a related two stage procedure always succeeds' is stated informally and is not proven or fully tested; please label it as a conjecture or provide supporting evidence.","section":"Sec. 5"}],"recommendation":"reject","confidential_remarks":"This is an early workshop-style draft with an interesting idea, but the sign error in Eq. (18) is not a cosmetic typo: it reverses the intended behavior of the normalization and, if the experiments were run with the printed q_c, they reflect a different algorithm than the one described. The proof of Proposition 4 is also not repairable by a small edit because the required lower bound requires a different argument. I recommend rejection; a corrected version with a repaired proof and re-run experiments could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a novel idea — convex selection of orthonormal column subsets by combining length normalization with multitask basis pursuit — but the published normalization is mathematically wrong, and that undermines the central claim. The good news is that the fix appears to be a one-line reciprocal, and the rest of the framing is sensible.\n\nWhat's actually new: the specific recipe (symmetric normalization plus group basis pursuit for this task), the l_c ground truth loss, and the two-stage trimming idea. The empirical comparison to greedy search on Iris, Wine, and Ethanol is reasonable, and the paper is candid about the two-stage method not always reaching the global optimum (Section 6.4). Code is available, and the citation pattern is fine — earlier dictionary-selection work is appropriately acknowledged.\n\nThe main problem is Eq. (18). The function q_c(v) = (e^{||v||^c} + e^{||v||^{-c}})/(2e) is minimized at ||v||=1 and increases without bound as ||v|| grows. So Algorithm 1 amplifies long columns, the opposite of the intended behavior. This directly contradicts Definition 3, which requires the argmax to be the unit sphere. This is not a typo in a minor lemma; it's the normalization on which the whole method rests. The fix is to take the reciprocal, but that's not what's written.\n\nThe proof of Proposition 4 also needs work. The line 'by Proposition 3, lc(X) = lc(R)' mixes the ground truth loss with the basis pursuit penalty, and the equality ||β||_{1,2} = ||R^{-1}||_{1,2} is actually correct because right multiplication by Q^T preserves row norms, but the argument is compressed and the equality conditions are stated too loosely. The reader's claim about Proposition 3's proof being algebraically invalid is not quite right — it's shorthand but valid.\n\nThere's a mild circularity: l_c and q_c are built from the same function g, and for D=1 they're equivalent, so the evaluation is partly self-fulfilling. This is a concern but not a fatal one.\n\nBottom line: this is a promising direction with a load-bearing error. The core intuition may survive a repair, but as written the algorithm does not do what it claims. A serious referee could usefully point to the normalization fix and the proof cleanup. I'd send it to review if the authors fix Eq. (18) first; as it stands, reject.","headline":"Promising idea, broken normalization: Eq. 18 contradicts Definition 3, so the algorithm amplifies long columns; one-line fix could make this a real contribution.","tokens_in":13309,"tokens_out":7833,"would_cite":false,"duration_ms":67681,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Isometry pursuit claims that picking a near-orthonormal D-column block from a wide matrix can be done by normalization plus multitask basis pursuit, a convex surrogate for greedy or brute-force search.","keywords":["isometry pursuit","multitask basis pursuit","orthonormal column selection","symmetric normalization","group lasso","diversification","interpretable coordinates","convex optimization"],"falsifier":"Evaluate $q_1(t) = (e^t + e^{1/t})/(2e)$ at $t=10$ and at $t=1$: at $t=10$ the value is roughly $1.1 \\times 10^3$, while at $t=1$ it is about $0.568$, so unit vectors are not the global maximizer and Definition 3 is violated as written. A corrected normalization would need to be checked numerically before Proposition 4 can hold.","tokens_in":12274,"feed_emoji":"📐","tokens_out":10138,"duration_ms":83157,"temperature":0.7,"pith_summary":"Isometry pursuit is a proposed convex algorithm for selecting a set of $D$ columns from a wide matrix $X \\in \\mathbb{R}^{D\\times P}$ that is as close to orthonormal as possible. The paper claims that after rescaling each column by a symmetric normalization that makes unit-length vectors longest, multitask basis pursuit with the identity matrix as target identifies orthonormal column-submatrices, with a formal guarantee in the square case. If correct, this offers a convex alternative to greedy and brute-force search for tasks like choosing interpretable coordinates from a dictionary or diversifying a selected subset. On the Iris, Wine, and Ethanol datasets, two-stage isometry pursuit achieves lower isometry loss than greedy search.","feed_headline":"Convex optimization finds orthonormal submatrices in wide matrices","feed_subtitle":"Two-stage isometry pursuit beats greedy search for isometric coordinates on Iris, Wine, and Ethanol.","key_machinery":"The key object is the symmetric normalization $q_c(v) = (e^{\\|v\\|^c} + e^{\\|v\\|^{-c}})/(2e)$, intended to satisfy Definition 3: global maximizers are exactly the unit sphere, rescaling is invariant under inversion of vector length, and $q$ depends only on length. It is paired with multitask basis pursuit, the group-lasso penalty $\\|\\beta\\|_{1,2} = \\sum_p \\|\\beta_{p.}\\|_2$ under the equality constraint $w(X,c)\\beta = I_D$. Longer, non-orthogonal columns force larger row norms in $\\beta$, so the support of the minimizer marks the columns closest to an orthonormal block. The ground-truth loss $l_c(X) = \\sum_d g(\\sigma_d(X),c)$ with $g(t,c) = (e^{tc}+e^{t^{-c}})/(2e)$ is uniquely minimized at orthonormal matrices and serves as the brute-force objective.","core_discovery":"On the paper's own terms, the discovery is that orthonormal column-submatrices of a wide matrix can be recovered by a two-step convex procedure: normalize, then solve a group-sparse regression. Each column $v$ is replaced by $q_c(v)v$ where $q_c(v) = (e^{\\|v\\|^c} + e^{\\|v\\|^{-c}})/(2e)$ is designed to peak at unit length; then the algorithm minimizes the multitask basis pursuit penalty $\\|\\beta\\|_{1,2} = \\sum_p \\|\\beta_{p.}\\|_2$ subject to $w(X,c)\\beta = I_D$. Proposition 4 states that for a square normalized submatrix the minimizer is orthonormal and the minimal loss equals $D$, and Proposition 3 states the selected support is invariant under orthonormal changes of basis of $X$. The practical estimator is two-stage: first use isometry pursuit to prune the candidate columns, then run brute-force search on the pruned set with a ground-truth isometry loss $l_c$.","pith_inferences":["If the square-submatrix guarantee extends to overcomplete dictionaries, the same normalize-then-pursuit recipe could replace the greedy feature-by-feature construction used in local isometry embedding methods.","The same mechanism suggests a generic diversification primitive: given a pool of items, normalize their feature vectors and let group lasso pick a small orthogonal spread, which is directly relevant to retrieval and recommendation.","One testable extension is to random dictionaries with a planted orthonormal $D$-block: if isometry pursuit consistently recovers that block as $P$ grows, the convex relaxation is doing genuine combinatorial work rather than only pruning."],"forward_implications":["Two-stage isometry pursuit achieves lower isometry loss than greedy search on Iris, Wine, and Ethanol, with statistically significant paired differences.","Isometry pursuit prunes the candidate set to a small support before brute-force search, making exhaustive enumeration feasible for moderate dictionaries.","Because the support is invariant under orthonormal transformations of the data matrix, the selected coordinates do not depend on the choice of orthonormal frame.","For square normalized submatrices, the multitask basis pursuit minimizer is exactly orthonormal, giving a formal foundation for the pruning step."],"supporting_citations":[{"why":"Supplies the group/multitask basis pursuit formulation and dictionary-based coordinate selection that isometry pursuit adapts.","marker":"[10]"},{"why":"Provides the dictionary-based manifold learning algorithm whose adaptation produces isometry pursuit.","marker":"[1]"},{"why":"Basis pursuit, the convex relaxation that motivates replacing combinatorial subset search with an $\\ell_1$ objective.","marker":"[16]"},{"why":"Defines the grouped variables penalty underlying the multitask basis pursuit norm.","marker":"[17]"},{"why":"Generalized QR factorization used in the proof of Proposition 4 to bound coefficient row norms.","marker":"[72]"},{"why":"Greedy construction of local isometries, the baseline method that two-stage isometry pursuit is compared against.","marker":"[4]"},{"why":"Iris dataset used in the diversification and coordinate selection experiments.","marker":"[22]"},{"why":"Wine dataset used in the diversification experiments.","marker":"[23]"},{"why":"Ethanol dataset used for interpretable coordinate selection experiments.","marker":"[25]"}],"fun_headline_variants":["Convex method finds orthonormal submatrices in wide matrices","Isometry pursuit: convex trick for orthonormal columns","Two-step convex search beats greedy for isometric coordinates","Normalize then regress: new convex algorithm for orthonormal subsets","Finding orthonormal blocks in wide matrices via convex optimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of Proposition 4 assumes that the stated formula for $q_c$ has its global maximum exactly on the unit sphere and that a row-norm comparison in the QR decomposition argument is valid; if either fails, the orthonormality guarantee does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Convex method finds orthonormal submatrices in wide matrices","Isometry pursuit: convex trick for orthonormal columns","Two-step convex search beats greedy for isometric coordinates","Normalize then regress: new convex algorithm for orthonormal subsets","Finding orthonormal blocks in wide matrices via convex optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000443,"raw_usage":{"total_tokens":2177,"prompt_tokens":813,"completion_tokens":1364,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":1279}},"tokens_in":429,"tokens_out":1364,"duration_ms":9521,"temperature":1.0,"reasoning_tokens":1279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:08:45.499966+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate $q_1(t) = (e^t + e^{1/t})/(2e)$ at $t=10$ and at $t=1$: at $t=10$ the value is roughly $1.1 \\times 10^3$, while at $t=1$ it is about $0.568$, so unit vectors are not the global maximizer and Definition 3 is violated as written. A corrected normalization would need to be checked numerically before Proposition 4 can hold.","supporting_citations":[{"cited_title":"Manifold coordinates with physical meaning","cited_arxiv_id":null,"evidence_quote":"Supplies the group/multitask basis pursuit formulation and dictionary-based coordinate selection that isometry pursuit adapts."},{"cited_title":"Consistency of dictionary-based manifold learning","cited_arxiv_id":null,"evidence_quote":"Provides the dictionary-based manifold learning algorithm whose adaptation produces isometry pursuit."},{"cited_title":"Donoho and Michael A","cited_arxiv_id":null,"evidence_quote":"Basis pursuit, the convex relaxation that motivates replacing combinatorial subset search with an $\\ell_1$ objective."},{"cited_title":"Model selection and estimation in regression with grouped variables","cited_arxiv_id":null,"evidence_quote":"Defines the grouped variables penalty underlying the multitask basis pursuit norm."},{"cited_title":"Generalized qr factorization and its applications","cited_arxiv_id":null,"evidence_quote":"Generalized QR factorization used in the proof of Proposition 4 to bound coefficient row norms."},{"cited_title":"LDLE: Low distortion local eigenmaps","cited_arxiv_id":null,"evidence_quote":"Greedy construction of local isometries, the baseline method that two-stage isometry pursuit is compared against."},{"cited_title":"Towards exact molecular dynamics simulations with machine-learned force fields","cited_arxiv_id":null,"evidence_quote":"Ethanol dataset used for interpretable coordinate selection experiments."}],"review_version":1}