{"id":"62c49278-cae1-4e03-9e18-78639e438801","arxiv_id":"2506.16013","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FIR-PCA is an iterative robust PCA method that uses projection depth and incremental PCA to select inliers, claiming better point-outlier robustness than DetMCD and FDB.","lead":"A new robust PCA method called FIR iteratively selects inlier points using projection depth and incremental PCA, then estimates location and covariance from that subset. The authors report it runs faster than DetMCD and handles point outliers better than DetMCD and FDB, though the accuracy improvements are mostly limited to the point-outlier scenario.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 step 3(d) is undefined whenever fewer than m unselected points fall in the 50%-expanded IPCA box; no fallback is specified and Theorem 1 assumes existence, so the central FIR/FIR-PCA claims rest on an unproven well-definedness condition.","rationale":"The central promise of the paper is a robust location/covariance estimator and a PCA built from it. Algorithm 1 is the only mechanism that produces the robust subset H; if step 3(d) can be invoked with fewer than m candidates, the estimator is not a well-defined function of the data. This is not a stylistic issue: Theorem 1's equivariance proof is an induction over the selected index sets, and each induction step assumes the next H^{(k+1)} is obtained by choosing m points inside the box. Existence is never shown, and the 0.5 expansion factor appears to be a heuristic rather than a derived quantity. The reader's weakest_assumption identifies the same algorithmic gap; I partially agree because the more pointed sub-issue is not the equivariance of the 0.5 rule (which is plausible under orthogonal transformations) but the missing guarantee that the box contains enough unselected points. I do not elevate this to a rejection: the paper's simulated point-outlier results are suggestive, the complexity analysis is reasonable, and an added fallback plus explicit assumptions would likely repair the defect. Other weaknesses (approximate projection depth, overstated accuracy claims, no code) are secondary and do not change the recommendation. So the reader's CONDITIONAL verdict stands.","tokens_in":16027,"tokens_out":9241,"duration_ms":122783,"concrete_test":"Instrument Algorithm 1 with a counter c_k = |{i not in H^{(k)} : for j=1,2, min_{l in H^{(k)}} z_{l,j} - delta_j <= z_{i,j} <= max_{l in H^{(k)}} z_{l,j} + delta_j}| at each iteration k. Run the algorithm on the paper's Section 4 configuration (n=1000, p=5 and p=10, 40% Point outliers, m=100, alpha=0.5) and on a degenerate control (n=50, p=2, m=2, data with two identical central points plus distant outliers). If c_k < m at any iteration in either run, Algorithm 1 crashes or must be modified; report the failure rate over 1000 Monte Carlo replications. Independently, attempt an analytical counterexample: if one exists, add a fallback rule (e.g., progressively expand the box) and restate Theorem 1 with the existence condition made explicit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"At each iteration, Algorithm 1's step 3(d) selects the m unselected points with smallest scaled distance (Eq. 2) that lie inside the selection box (Eq. 3). The box is obtained by expanding the bounding box of the already-selected points' projections on the first two IPCA axes by 50% in each direction. Nothing in the algorithm or analysis guarantees that at least m unselected points are in that box. If the initial depth-selected points are tightly clustered or contain duplicates, delta_j can be zero and the box reduces to the selected points themselves; with discrete or deterministic data the candidate set can have size less than m. Step 3(d) then has no valid output. The paper offers no fallback (e.g., enlarging the box, taking nearest points outside, or a stopping rule), and the stopping criterion is specified only as a count h = alpha*n, not as a condition on candidates. Theorem 1's proof in Section 3.1 states that 'the set of m points with smallest scaled distance that lie in the bounding box ... is also identical' for transformed data, which presupposes that such a set exists; it does not prove existence. Since FIR location/covariance and FIR-PCA both call Algorithm 1, a single configuration with fewer than m candidates makes the central claim ('reliable dimensionality reduction ... in the presence of anomalous data points') unsubstantiated for that configuration. This is an internal completeness issue rather than a disagreement with existing robust-PCA practice.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FIR, a fast iterative robust location and covariance estimator, and FIR-PCA, a robust PCA built on it. FIR selects an initial subset of m points with highest projection depth, then iteratively adds the m unselected points with smallest scaled distance in an incremental PCA (IPCA) space, restricted to a selection box formed by expanding the bounding box of the current selected points by a factor of 0.5 along the first two IPCA axes. The final subset is used to estimate location and covariance, from which robust PCA scores and distances are derived. The authors state that FIR is orthogonal equivariant and permutation invariant, analyze its computational complexity as O((κ+τ)np + np^2), and compare it against DetMCD and FDB on simulated and real datasets, reporting better robustness for point outliers at 40% contamination.","tokens_in":16379,"tokens_out":9628,"duration_ms":110369,"significance":"If the algorithmic gaps are closed, the paper would make a useful practical contribution: FIR is much faster than DetMCD, and the simulations in Tables 1–2 and Figures 3–5 show that it is more robust than DetMCD and FDB specifically under point outliers at 40% contamination, while remaining comparable for cluster and radial outliers. The paper is honest about limitations (p > n, partial outliers) and provides an explicit complexity analysis. The main missing elements are a rigorous well-definedness argument for the selection step and a complete equivariance proof; until those are supplied, the theoretical claims are conditional and the empirical claims rest on an incompletely specified algorithm.","major_comments":[{"comment":"Algorithm 1 Step 3(d) selects the m unselected points with smallest scaled distance that lie in the selection box defined by Eq. (3), but nothing guarantees that at least m unselected points fall inside this box. The box is only a 50% expansion of the bounding box of the already-selected points on the first two IPCA axes; for discrete data, tightly clustered initial points, or distant outliers, the candidate set can have size less than m. In that case Step 3(d) has no valid output, and neither the stopping rule nor any fallback (e.g., enlarging the box, taking nearest points outside the box, or terminating early) is specified. Because both FIR location/covariance and FIR-PCA call Algorithm 1, this well-definedness gap is load-bearing for the central claims of the paper. Please add a fallback or a condition under which the candidate set always has size at least m, and prove that condition.","section":"Algorithm 1, Step 3(d), Eq. (3)"},{"comment":"The proof of orthogonal equivariance is incomplete. Equations (14)-(15) assert invariance of IPCA scores and singular values, and the text then asserts that the set of m points with smallest scaled distance inside the bounding box of Eq. (3) is identical for Z and ZA+1v^T. This requires a proof that the selection-box rule itself is orthogonal equivariant, i.e., that the expanded bounding box in the transformed IPCA coordinates contains exactly the transformed members of the original candidate set. The proof also presupposes that this candidate set has at least m elements. These points are not established, and the arbitrary 0.5 expansion factor is not justified. Please supply a lemma for the box rule and address existence before invoking induction for H^(k).","section":"Section 3.1, Theorem 1 proof, Eqs. (14)-(17)"},{"comment":"The stopping rule is underspecified when h = alpha*n is not a multiple of the batch size m. The loop 'for k = 1 to floor(h/m - 1)' adds m points per iteration, so the final selected set has size m*floor(h/m), which can be strictly less than h. The text describes the target as h = alpha*n points, but no rounding rule or partial final batch is defined. This affects the actual subset used for location and covariance, and should be clarified (e.g., by defining h as the integer multiple of m closest to alpha*n, or by allowing a final partial batch).","section":"Algorithm 1, Section 2.2"},{"comment":"The orthogonality proof treats projection depth as exactly affine equivariant via Eq. (1), but the algorithm states that u is randomly sampled from the unit p-sphere. A finite random draw of directions is not exactly orthogonal equivariant under a fixed draw, so the exact equivariance claimed in Theorem 1 does not hold for the implemented Algorithm 1. Please state whether the theorem applies to the population depth with a supremum over all directions, and if the random sampling is retained, clarify the sense in which equivariance holds (e.g., in distribution, or for direction sets transformed by A).","section":"Section 2.2, Eq. (1), Algorithm 1 Step 1"}],"minor_comments":[{"comment":"The term delta_j^(k) is used but never defined; the sentence following the equation should define delta_j^(k) = 0.5 (max - min) and the displayed formula needs correction.","section":"Eq. (3)"},{"comment":"The variance formula 'sigma = diag(L)^2' appears to be a typo; if sigma denotes standard deviations, it should be sqrt(diag(L)), and if it denotes variances, the notation should be consistent with the use in Eq. (8).","section":"Eq. (6)"},{"comment":"The execution time '11.43.03 ms' contains a typo.","section":"Section 5.2"},{"comment":"The captions list 'CPCA, DetMCD-PCA, FIR-PCA, and FIR-PCA'; the second FIR-PCA should be FDB-PCA.","section":"Figures 7, 8, 10 captions"},{"comment":"The complexity derivation would be clearer if it stated kappa = h/m - 1 before the equality O(kappa^2 m p + kappa m p^2) = O(kappa n p + n p^2).","section":"Section 3.3"},{"comment":"The theorem calls Z 'nonsingular' with n observations and p features; for rectangular matrices, the intended condition (e.g., full column rank) should be stated.","section":"Theorem 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope. My main concerns are the well-definedness of Algorithm 1, the incomplete proof of Theorem 1, and the underspecified stopping rule; these are fixable, so I do not recommend rejection. The empirical evidence for point-outlier robustness is reasonably convincing, and the paper does not appear to have problematic citation practices. I would also encourage the authors to include a sensitivity analysis for the 0.5 expansion factor, since it is a free parameter that affects both robustness and the existence of candidates."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new algorithm, and the empirical case for point-outlier robustness is credible. But the paper has a load-bearing well-definedness gap in Algorithm 1 and a proof gap in Theorem 1; both are fixable but need to be addressed before publication.\n\nWhat's new: FIR combines projection depth initial selection with IPCA-driven iterative refinement using scaled distances and a selection box. That combination is new relative to FDB and MCD-based methods. The simulations are reasonably thorough: four data sizes, three outlier types, clear tables. The point-outlier results at 40% contamination are striking: FIR's location error stays low where DetMCD and FDB blow up. Speed is also a real advantage over DetMCD, with only about 2x slowdown vs FDB. I believe the central empirical claim: FIR-PCA is more robust to point outliers than its immediate predecessors, at competitive speed.\n\nSoft spots: the stress-test note is right. Step 3(d) requires m unselected points inside the selection box, and nothing guarantees that. The box is defined as a 50% expansion of the current batch's bounding box. With duplicate points, deterministic data, or tight clusters, delta can be zero and the box can be empty of candidates. No fallback is specified. This is not a nit: Algorithm 1 is the method, and if a configuration stalls, FIR and FIR-PCA are undefined for that configuration. The proof of Theorem 1 glosses the same issue: it asserts the set of m smallest-distance points inside the box is identical under rotation, but does not prove existence. Equivariance of the box rule itself is also asserted, not shown. The 0.5 factor is arbitrary, but that is a tuning choice, not a flaw.\n\nAlso worth noting: no code or data is shipped, which matters for a methods paper where implementation details (projection depth directions, IPCA behavior, batch size) affect results. The broad phrasing 'maintains robustness' and 'reliable dimensionality reduction' overshoots what the tables show; a cleaner claim would be point-outlier robustness. Minor: the conclusion honestly admits limitations for p > n and partial outliers.\n\nCitation pattern looks fine; FDB, DetMCD, and IPCA are all properly cited. No self-citation inflation. I would send this to a serious referee. The core idea is worth scrutiny and the empirical claims are strong enough to justify referee time. The authors will likely need to add a fallback or a condition guaranteeing candidate availability, and tighten the proof. If those are fixed, this is a solid contribution to robust PCA.\n\nRecommendation: engage, conditional on revision.","headline":"Genuinely new robust-PCA algorithm with credible point-outlier results, but a load-bearing well-definedness gap and an incomplete equivariance proof need fixing before publication.","tokens_in":16841,"tokens_out":1639,"would_cite":true,"duration_ms":17419,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H25","62G35"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces FIR, a fast iterative robust location and covariance estimator, and constructs FIR-PCA from it, claiming dependable principal component analysis even when a large fraction of data points are outliers.","keywords":["robust PCA","incremental PCA","projection depth","outlier detection","location estimation","covariance estimation","dimensionality reduction","orthogonal equivariance"],"falsifier":"Take a dataset whose true inliers lie in a narrow elongated band, place 40% point outliers just outside the 0.5-expanded box, run Algorithm 1, and check whether every iteration finds m unselected points inside the box; if any iteration selects fewer than m points or pulls in an outlier because the box is empty, the claim that FIR always builds a reliable inlier set is false.","tokens_in":15848,"feed_emoji":"🎯","tokens_out":11512,"duration_ms":116499,"temperature":0.7,"pith_summary":"This paper introduces FIR, a location-and-covariance estimator built for outlier-contaminated data, and constructs a robust PCA (FIR-PCA) from it. The estimator starts with the m points having the largest projection depth, then grows an inlier set one batch at a time: each iteration projects the remaining points into an incremental PCA space and admits the m points closest to the current inlier subset under a singular-value-scaled distance and within a box tied to the two leading principal directions. The authors claim that this selection keeps the estimated mean and covariance accurate when 40% of the data are point outliers, a regime where the DetMCD and depth-based FDB alternatives degrade, while runtime stays close to the cheaper depth-based methods. If the claim holds, FIR-PCA gives a practical way to do dimensionality reduction and outlier diagnostics on datasets that classical PCA would distort, without paying the concentration-step cost of minimum-covariance-determinant approaches.","feed_headline":"Iterative estimator keeps PCA accurate even with 40% point outliers","feed_subtitle":"A fast iterative scheme keeps PCA stable when 40% of points are outliers, at a fraction of current runtimes.","key_machinery":"The load-bearing mechanism is the iteration in Algorithm 1 that turns a depth-based seed into an inlier set. Projection depth (Eq. 1) supplies affine equivariant starting indices; incremental PCA (Ross et al. 2008) supplies batch updates of the subspace and singular values; the scaled distance $d_i^{(k)}$ (Eq. 2) and the 0.5-expanded selection box (Eq. 3) determine which points are accepted next. The box, computed from the two leading singular vectors of the current inlier subset, is the device that keeps the search close to the data's dominant structure; Theorem 1's induction uses the invariance of the IPCA projections and singular values under rotation and translation to prove that the whole selection chain, and therefore the FIR estimates, is orthogonal equivariant and permutation invariant.","core_discovery":"On its own terms, the paper is establishing that a greedy, depth-seeded, incrementally updated inlier subset is enough to make PCA reliable under heavy point-outlier contamination. The procedure is Algorithm 1: initialize with the highest projection-depth points, fit an IPCA subspace to the selected set, project the unselected points, rank them by the scaled distance $d_i^{(k)} = \\sum_j (\\hat{Z}_{i,j}^{(k)}/s_j^{(k)})^2$, and repeatedly accept the m points with smallest distance that fall inside the selection box defined by the two leading IPCA coordinates expanded by a factor of 0.5. After $h = \\alpha n$ points are collected, the mean and covariance of that subset are used to form robust PCA scores, variances, and outlier distances. The central claim is that the resulting FIR and FIR-PCA are orthogonal equivariant and permutation invariant (Theorem 1), and that at high outlier fractions they preserve the underlying structure better than DetMCD-PCA and FDB-PCA, with typical runtime between the two.","pith_inferences":["The 0.5 expansion factor is a design parameter the paper keeps fixed; a natural extension is to test whether smaller expansion rejects borderline outliers more aggressively and larger expansion prevents the selection box from emptying at high contamination, since Algorithm 1 has no fallback if the box holds fewer than m points.","Because the projection-depth computation samples directions from the unit sphere, FIR carries a stochastic component that the theoretical analysis treats as fixed; a stability check across random-direction seeds would quantify how reproducible the selected inlier set H is.","The selection box uses only the two leading IPCA singular vectors, so the method's guarantees are most natural when the underlying structure is approximately two-dimensional; a testable variant would define the box in the leading r directions and compare robustness as r grows.","The paper states FIR is not designed for partial (cellwise) outliers, so a fair inference is that its target regime is rowwise contamination; handling per-feature corruption would likely need a separate feature-scaling step before the iterative selection."],"forward_implications":["For datasets with 40% point outliers, the paper reports that FIR's location error and KL divergence stay close to the clean-data values, while DetMCD and FDB deteriorate (Tables 1-2, Figures 3-4).","The stated complexity of Algorithm 1 is $O((\\kappa+\\tau)np+np^2)$, which avoids the $O(\\Gamma(p^3+np^2))$ cost of MCD concentration steps; measured runtimes show FIR slightly slower than FDB and far faster than DetMCD (Figure 6).","FIR is orthogonal equivariant and permutation invariant (Theorem 1), so rotating, translating, or shuffling the data does not change the selected inlier set or the resulting estimates.","On the octane, forged-bank-note, TopGear, and Philips datasets, FIR-PCA flags the known outliers with runtimes of roughly 5-44 ms, while DetMCD-PCA takes 56-692 ms in these examples.","The batch size $m$ controls the speed-accuracy trade-off: values near 10% of the data give most of the speed benefit without accuracy loss, and values near $p$ are the conservative choice when speed is less important."],"supporting_citations":[{"why":"Defines projection depth and establishes its affine equivariance, which seeds the initial inlier set and anchors the proof of Theorem 1.","marker":"(Zuo & Serfling 2000)"},{"why":"Provides the incremental PCA update that grows the subspace as batches are added, the core of the iterative selection step.","marker":"(Ross et al. 2008)"},{"why":"Introduces deterministic MCD (DetMCD), the main robust-estimator baseline whose accuracy and runtime are compared throughout.","marker":"(Hubbert et al. 2012)"},{"why":"Introduces the fast depth-based (FDB) estimator, the closest speed baseline and the method FIR most directly extends by replacing its selection step.","marker":"(Maoyu Zhang & Dai 2024)"},{"why":"Defines ROBPCA and the octane-data outlier diagnostic convention used to evaluate FIR-PCA on real spectra.","marker":"(Hubert et al. 2005)"},{"why":"Supplies FastMCD and the concentration-step cost that motivates FIR's cheaper iterative scheme.","marker":"(Rousseeuw & Driessen 1999)"}],"fun_headline_variants":["Fast iterative PCA resilient to heavy outlier contamination","Depth-seeded incremental subset keeps PCA accurate with outliers","FIR-PCA: fast robust PCA that iteratively picks inliers","Outlier-resistant PCA without the computational blowup","Iterative inlier subset makes PCA robust and fast"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The algorithm assumes that at every iteration at least m unselected points fall inside the selection box defined in Equation 3; if the box ever contains fewer than m points, the next step is undefined, and the paper does not guarantee this cannot happen.","fun_headline_variants_meta":{"raw":{"variants":["Fast iterative PCA resilient to heavy outlier contamination","Depth-seeded incremental subset keeps PCA accurate with outliers","FIR-PCA: fast robust PCA that iteratively picks inliers","Outlier-resistant PCA without the computational blowup","Iterative inlier subset makes PCA robust and fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000547,"raw_usage":{"total_tokens":2590,"prompt_tokens":898,"completion_tokens":1692,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":1616}},"tokens_in":514,"tokens_out":1692,"duration_ms":17979,"temperature":1.0,"reasoning_tokens":1616,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:44:11.335268+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset whose true inliers lie in a narrow elongated band, place 40% point outliers just outside the 0.5-expanded box, run Algorithm 1, and check whether every iteration finds m unselected points inside the box; if any iteration selects fewer than m points or pulls in an outlier because the box is empty, the claim that FIR always builds a reliable inlier set is false.","supporting_citations":[],"review_version":1}