{"id":"cf95bb7d-0b8a-40b6-881f-8f3f863bae0c","arxiv_id":"2411.19920","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Explicit formulas for the codimension and component count of fiber varieties of the matrix multiplication map, with applications to deep linear networks and singular learning theory.","lead":"This paper computes the exact codimension and number of top-dimensional components of the algebraic set of matrix tuples whose product is zero or a fixed matrix. The formulas apply to deep linear neural networks and show their Bayesian singularities are mild.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QIP and explicit formulas in Theorems 6.1 and 7.10 rest on an unproven structural classification of lace diagrams for weakly increasing dimension vectors.","rationale":"The reader's weakest_assumption identifies exactly the same gap: the structural lemma behind the QIP is only sketched. My independent reading confirms that the proof of Theorem 6.1 hinges on the assertion that top-dimensional components have lace diagrams of a very special form; the top-row argument is mostly sound but the lower-row part and Lemma 6.4 itself are not proved. If this structural classification fails for even one weakly increasing dimension vector, the QIP and the explicit formulas of Section 7 would not compute the true C and θ, and the permutation-invariance application of these formulas would also be unsupported. The Poincaré-series route (Theorem 5.5) appears to be rigorous, and the explicit formula of Theorem 7.10 is an algebraic consequence of the QIP, so the core unresolved issue is this combinatorial-geometric lemma. The RLCT result Theorem 8.6 depends on an external result by Aoyagi; while this is a secondary concern, it reinforces the conditional verdict. The proposed computational check on a small weakly increasing dimension vector would directly settle whether the structural claim holds, and I would accept the paper if this check succeeds and the lemma is given a real proof.","tokens_in":26132,"tokens_out":31810,"duration_ms":255343,"concrete_test":"For d'=(2,3,4,5), N=4, enumerate all Kostant partitions, compute orbit codimensions via Equation (4), find the minimal-codimension orbits in Σ^0_{d'}, and test whether every such orbit has a representative lace diagram with exactly one missing gap per top row and all lower rows of type [j,4]. Compare the number of such orbits with the number of optimal solutions to the QIP.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the proof of Theorem 6.1, after invoking Lemma 6.4 (whose proof is only 'Induction on N'), the paper asserts that every lace diagram representing a top-dimensional component of Σ^0_{d'} has exactly one missing [i-1,i] interval in each of the top d'_0 rows, and that every lower row is an interval [j,N] with multiplicity f_j = d'_j - d'_{j-1}. The 'exactly one missing' argument for the top rows is plausible: adding one of two missing intervals merges the two pieces of the row but leaves another gap, so the product stays zero and the orbit closure strictly grows, contradicting component maximality. However, the lower-row structure is only called 'analoguous'; it is not shown that a lower row cannot be [a,b] with b<N or a>1. Extending such a row requires reconnecting dots from other rows, and it is not demonstrated that the resulting Kostant partition is larger in the orbit closure order and still belongs to Σ^0. Because the quadratic integer program (QIP) and hence Theorems 7.5, 7.10, and the corollaries for C and θ depend directly on this structural classification, this is the most load-bearing unproven step. A counterexample to the lower-row claim would invalidate the QIP and all explicit formulas.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the variety of tuples of composable matrices whose product has a fixed rank r, and in particular the zero-product locus Σ^0_d and the fibers mult^{-1}(B) of the multiplication map. The main results are three descriptions of the codimension C of the top-dimensional components and their number θ: a Poincaré-series formula in equivariant cohomology (Theorem 5.5), a quadratic integer program (Theorem 6.1), and an explicit closed formula (Theorem 7.10). The paper also proves that C and θ are invariant under permutations of the dimension vector (Corollary 5.10) and reformulates Aoyagi's computation of the real log-canonical threshold of the squared Frobenius loss of deep linear networks as rlct(K_B^DLN) = codim mult^{-1}(B)/2 (Theorem 8.6). The exposition is generally clear, and the quiver-representation framework is used effectively to organize the orbit structure.","tokens_in":26362,"tokens_out":10789,"duration_ms":96963,"significance":"If fully established, the results solve a natural problem about the geometry of matrix multiplication and give a clean, explicit description of the singularities relevant to deep linear networks. The Poincaré-series method in Section 5 is elegant, and the paper provides several worked examples that make the formulas concrete. The main obstruction is that the proof of Theorem 6.1 rests on a structural classification of lace diagrams that is only sketched; the explicit formulas in Theorem 7.10 inherit this gap. Because the missing step is localized and appears fillable, the paper is promising but needs a completed proof before it can be accepted.","major_comments":[{"comment":"The classification of the lower rows in the lace diagram is asserted by the sentence 'An analoguous argument shows that in the rows below the top d'_0 rows all possible horizontal intervals are laces of L' and is not proved. This assertion is load-bearing: the displayed decomposition M = e_1(I_{00}+I_{1N}) + ... + e_N(I_{0,N-1}+I_{NN}) + f_1(I_{1N}) + ... + f_N(I_{NN}), the quadratic integer program (QIP), and hence Theorems 7.5 and 7.10 all depend on it. Please provide a proof that a lower row with a missing [i-1,i] interval can be extended to a strictly larger orbit closure that still lies in Σ^0_{d'}, for instance using the rank-pattern order of Theorem 3.8, or supply a reference for this structural fact.","section":"Section 6.2, proof of Theorem 6.1"},{"comment":"The proof of Lemma 6.4 is given only as 'Induction on N'. This lemma is used to represent every orbit in the weakly increasing case by a lace diagram with only horizontal laces, and it is not immediate from the definition of Kostant partitions. Please give the full induction argument or an explicit reference; as written, this is a gap in the proof of Theorem 6.1.","section":"Lemma 6.4"},{"comment":"The step in which the projection of s to the hyperplane has 'possibly a few last components negative' and the problem is reduced to the first m coordinates is not justified. This truncation step is the bridge from the quadratic integer program to the explicit formula, and it needs a proof, for example a water-filling or convexity argument showing that the optimal e_i vanish for i > m and that the remaining problem is exactly the closest-vector problem in the smaller simplex.","section":"Theorem 7.5"}],"minor_comments":[{"comment":"The product (1-q^{s-r+1})...(1-q^r) should read (1-q^{s-r+1})...(1-q^s); the same typo appears in the displayed computation after Eq. (12). The subsequent use in Theorem 5.5 is with r = s, so the conclusion is unaffected, but the statement as written is false for r < s.","section":"Lemma 5.9"},{"comment":"In the displayed chain of equalities, the middle term is missing a factor of 2 in front of the first sum; the final identity is correct, but the intermediate display should be corrected.","section":"Proof of Theorem 7.5"},{"comment":"The last sentence says 'since d' is weakly decreasing', but d' is weakly increasing; this should be corrected.","section":"Lemma 7.1"},{"comment":"The word 'analoguous' appears in Section 6.2; it should be 'analogous'.","section":"Throughout"},{"comment":"Remark 8.7 refers to 'the formula of Theorem' without a number; the intended theorem number should be supplied.","section":"Remark 8.7"},{"comment":"The proof of Theorem A.1 is only a sketch. Since this appendix is presented as a bonus and is not used in the main theorems, this is acceptable, but a precise reference to the details in [FR04] and [FR02] would be helpful.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The main gap is localized to Section 6.2: the structural classification of lace diagrams for weakly increasing dimension vectors is asserted rather than proved, and Theorem 6.1, the QIP, and all subsequent explicit formulas depend on it. If the authors can supply the missing proof, the paper is likely to be acceptable. I also note that Theorem 6.1 deliberately uses Corollary 5.10, so the QIP approach does not give an independent proof of permutation invariance; this is acknowledged in the paper and is not by itself a problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this: the paper has a real result and a real gap. The Poincaré series (Thm 5.5) computing C and θ, with permutation invariance as a corollary, is solid and the argument holds up. The QIP (Thm 6.1) and explicit formula (Thm 7.10) are genuinely new and far more usable, but both rest on a structural claim about lace diagrams in Section 6.2 that is asserted, not proven.\n\nThe load-bearing step: for weakly increasing d', every top-dimensional component of Σ^0_d' is supposed to have a lace diagram with exactly one missing [i-1,i] interval in each top row and every lower row running to N. The top-row part is argued correctly. The lower-row part is dismissed with \"analoguous\", and Lemma 6.4 (the existence of horizontal lace diagrams) is \"Induction on N\". I traced the mechanism myself: a product path in a horizontal lace diagram stays within a single row, so extending a lower row cannot create a nonzero product, and the rank-pattern closure argument should complete the proof. So I expect the statement is true. It just is not proven, and the QIP, the closed formulas, and the RLCT simplification all inherit the gap.\n\nThe RLCT section is honest but deserves a caveat. Theorem 8.6 is a translation of Aoyagi's theorem, not an independent derivation; the paper credits this, and the notation comparison checks out. The clean consequence — rlct = codim/2, so deep linear networks are \"mildly singular\" — is genuinely useful for the singular-learning-theory crowd. The paper is also upfront that the general non-weakly-increasing description is deferred to a forthcoming note [KR24], which marks the boundary of what is settled here.\n\nMinor issues: an indexing typo in Lemma 5.9's Pochhammer factor (harmless, since only r=s is used), a missing factor of 2 in one line of Thm 7.5's proof, and a scattered misspelling of \"analogous\". Final identities verify on the worked examples. Citation practice is fine: SB24 and TKB20 get proper credit for the linear-algebra groundwork, and the self-citations (Rim13, FR02) are the appropriate sources for the spectral-sequence and Ext machinery.\n\nVerdict: send it to peer review. The Poincaré-series result alone justifies referee time, the gap is likely a few pages of argument rather than a wrong theorem, and the referee's main assignment is clearly Section 6.2. The real audience is representation theory and singular learning theory; the ML framing is motivation, not substance.","headline":"Solid Poincaré-series computation of C and θ for zero-product matrix tuples, with a real but probably fixable proof gap in the structural lemma behind the explicit formulas.","tokens_in":26919,"tokens_out":29602,"would_cite":true,"duration_ms":231439,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["16G20","14L30","14M12","14B05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper determines, for tuples of composable matrices of fixed shapes whose product is zero (or a fixed matrix), the codimension $C$ and the number $\\theta$ of largest-dimensional irreducible pieces of that variety, and proves both are…","keywords":["deep linear networks","multiplication map","quiver representations","Kostant partitions","lace diagrams","codimension","real log-canonical threshold","irreducible components"],"falsifier":"Compute the top-dimensional components of the zero-product variety $\\Sigma^0_{(2,3,5,7)}$ by direct elimination, and check each component's Kostant partition for a lace diagram with exactly one missing step in each of the top two rows and all horizontal intervals present below; a single component without such a diagram would show the quadratic integer program does not give the true codimension.","tokens_in":25897,"feed_emoji":"🧮","tokens_out":24388,"duration_ms":180036,"temperature":0.7,"pith_summary":"Deep linear networks are just tuples of composable matrices, and the set of weights that realize a fixed input-output map is the fiber of the multiplication map. This paper determines the codimension $C$ of the zero-product locus (the tuples whose product is zero), the number $\\theta$ of its largest irreducible components, and derives the corresponding numbers for every fiber $\\mathrm{mult}^{-1}(B)$. The answer comes in three equivalent forms: a Poincaré series in equivariant cohomology, a quadratic integer program, and closed-form formulas in the layer widths. A corollary is that $C$ and $\\theta$ are invariant under arbitrary permutations of the layer widths, even though the geometry itself changes. Since the real log-canonical threshold of the squared-Frobenius loss equals half the fiber codimension, the paper shows that deep linear networks are 'mildly singular' as statistical models.","feed_headline":"Deep linear networks: zero-product fibers now have explicit formulas","feed_subtitle":"A quadratic integer program and a closed formula give the codimension and component count from layer widths alone.","key_machinery":"The argument runs through the orbit stratification of the representation space: the group $G = \\prod_i \\mathrm{GL}_{d_i}$ acts on tuples by change of basis, and for the equioriented type A quiver the $G$-orbits are classified by Kostant partitions, i.e. multiplicities of interval modules $M_{ij}$ in the decomposition of a representation. Lace diagrams, arrays of dots in $N+1$ columns partitioned into intervals, are the combinatorial pictures of these partitions. Three facts make the machinery work: orbit closure is controlled by entrywise inequality of rank patterns; the codimension of an orbit is a bilinear form in the partition coming from $\\mathrm{Ext}(M,M)$; and the longest interval module $M_{0N}$ is both injective and projective, so adding copies of it leaves normal slices unchanged. This last fact is what ultimately makes $C$ and $\\theta$ independent of the ordering of the layer widths, and it reduces the top-dimensional components to the solutions of a quadratic integer program that can be solved by closest-vector rounding.","core_discovery":"The central discovery is that the zero-product variety $\\Sigma^0_d = \\{ (A_1,\\ldots,A_N) : A_N \\cdots A_1 = 0 \\}$ has codimension $C$ equal to the minimum of a quadratic integer program in the layer widths, and the number $\\theta$ of its top-dimensional irreducible components equals the number of minimizers. The same two numbers for the fiber $\\mathrm{mult}^{-1}(B)$ over a rank-$r$ matrix $B$ follow by reduction (Lemmas 4.5 and 4.6). The paper proves the statement three ways: Theorem 5.5 expresses the relevant generating series as $Q^r_d = P_r \\sum_{s=0}^{\\min d - r} (-1)^s q^{\\binom{s}{2}} P_s P_{d-r-s}$, whose lowest term is $\\theta q^C$; Theorem 6.1 derives the quadratic program from the combinatorics of lace diagrams; Theorem 7.10 solves the program explicitly using the weakly increasing rearrangement of $d$, the closest-vector problem in a type A root lattice, and the quantity $S = \\sum_{i=0}^m d'_i$. Corollary 5.10 records that $C$ and $\\theta$ are permutation-invariant, and Theorem 8.6 identifies the real log-canonical threshold of $K_B^{\\mathrm{DLN}}(A) = \\|\\mathrm{mult}(A)-B\\|_2^2$ with half the codimension of $\\mathrm{mult}^{-1}(B)$.","pith_inferences":["Beyond the paper: the permutation invariance suggests there should be a bijective, purely combinatorial involution on Kostant partitions that preserves codimension and the number of minimal orbits; finding one could give an elementary proof and possibly extend the formulas to other quivers.","Beyond the paper: because the global threshold equals half the fiber codimension, the formulas give a practical route to the local real log-canonical threshold for deep linear networks if the local equality $\\mathrm{rlct}_A(K_B)=\\mathrm{codim}_A\\,\\mathrm{mult}^{-1}(B)/2$, which the paper leaves to future work, holds.","Beyond the paper: the closest-vector reformulation ties the counting problem for $\\theta$ to the geometry of type A root lattices, so $\\theta$ can be read as the number of lattice points at a fixed minimal distance; this may connect component counting to the complexity of integer programming at large depth."],"forward_implications":["For any fixed layer widths, the codimension and the number of top components of the zero-product variety, and of every fiber $\\mathrm{mult}^{-1}(B)$, can be read off from a quadratic integer program or a closed formula instead of from resolving the variety.","Permuting the layer widths leaves $C$ and $\\theta$ unchanged: a network with widths $(d_0,d_1,d_2)$ and its mirror image have the same zero-product codimension and component count, though not the same ambient dimension or total component count.","The real log-canonical threshold of the squared-Frobenius loss equals $C/2$, saturating the general inequality $\\mathrm{rlct}(F) \\le \\mathrm{codim}\\,F^{-1}(0)/2$; for Bayesian inference on deep linear networks this makes the leading asymptotic term explicitly computable from the widths.","For equal widths $d$ and depth $N$, the explicit formula gives codimension $d(d+1)/2$ whenever $N>d$, independent of depth, and for fixed $N$ the codimension grows like $\\frac{N+1}{2N} d^2$ as $d \\to \\infty$, showing that the zero-product locus is controlled mainly by width."],"supporting_citations":[{"why":"Supplies the lemma that a normal slice to a quiver orbit at a module M is isomorphic to Ext(M,M), the foundation for orbit codimension formulas.","marker":"[Voi77]"},{"why":"Gives the bilinear Ext dimension formula for interval modules and the fact that the longest module M_0N is both injective and projective, used for permutation invariance and the Poincaré series.","marker":"[FR02]"},{"why":"Establishes the orbit-closure order via entrywise comparison of rank patterns, which identifies irreducible components of the zero-product locus with minimal orbits.","marker":"[AD80]"},{"why":"Provides the generating-function identity for quiver representation spaces that underlies Theorem 5.7 and hence the Poincaré-series formula.","marker":"[Rei10]"},{"why":"Supplies the degenerate spectral sequence computation in equivariant cohomology that turns orbit codimensions into the Poincaré series P_d.","marker":"[Rim13]"},{"why":"Supplies the filtered equivariant-cohomology spectral sequence that turns orbit codimensions into the generating function used in Theorem 5.5.","marker":"[Kaz97]"},{"why":"Gives the closest-vector rounding solution for type A root lattices used to solve the quadratic integer program explicitly in Theorem 7.10.","marker":"[CS88]"},{"why":"Computes the real log-canonical threshold of deep linear networks; Theorem 8.6 identifies that threshold with half the fiber codimension by comparing formulas.","marker":"[Aoy24]"}],"fun_headline_variants":["Fiber geometry of deep linear nets: codimension and components in closed form","Zero-product fibers: explicit codimension and component count from widths only","Deep linear networks' fibers: permutation-invariant geometry solved","Fiber codimension and components: quadratic program gives exact formula","Singular learning theory meets fibers: explicit geometry of deep linear nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The explicit computation rests on the assertion, argued by analogy and a sketched induction, that for weakly increasing widths every largest zero-product component can be drawn as a lace diagram with exactly one missing adjacent interval in each of the top rows and all horizontal intervals present in the rows below; if that assertion failed, the quadratic program and closed formulas would not describe the true $C$ and $\\theta$.","fun_headline_variants_meta":{"raw":{"variants":["Fiber geometry of deep linear nets: codimension and components in closed form","Zero-product fibers: explicit codimension and component count from widths only","Deep linear networks' fibers: permutation-invariant geometry solved","Fiber codimension and components: quadratic program gives exact formula","Singular learning theory meets fibers: explicit geometry of deep linear nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000962,"raw_usage":{"total_tokens":4138,"prompt_tokens":1028,"completion_tokens":3110,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":3020}},"tokens_in":644,"tokens_out":3110,"duration_ms":19065,"temperature":1.0,"reasoning_tokens":3020,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:42:28.503747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the top-dimensional components of the zero-product variety $\\Sigma^0_{(2,3,5,7)}$ by direct elimination, and check each component's Kostant partition for a lace diagram with exactly one missing step in each of the top two rows and all horizontal intervals present below; a single component without such a diagram would show the quadratic integer program does not give the true codimension.","supporting_citations":[],"review_version":1}