{"id":"29dd4cfb-8f00-4093-8395-d2f966f2a1aa","arxiv_id":"2412.05283","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A directed-cycle compartmental model is generically locally identifiable if and only if its leaks are interlaced with inputs and outputs, and catenary models get an explicit coefficient-map formula.","lead":"This paper proves exactly when a directed-cycle compartment model, a circular chain of pools with material flowing one way around the loop, has all rate parameters recoverable from noiseless experiments: the model is identifiable precisely when its leaks are separated from one another by an input or an output.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3.25's D-matrix construction is undefined when an input segment contains no output (j_s=0), leaving the sufficiency proof of Theorem 3.6 incomplete for valid leak-interlacing models.","rationale":"The reader's weakest_assumption focused on reliance on the cited rank criterion (Proposition 2.16) and the imported coefficient-map formula (Lemma 3.10). Those are standard results, and the paper's own spot-checks reproduce them; they are not where I found the sharpest problem. Instead, the most load-bearing concern is an internal gap in Proposition 3.25, the proof of the sufficiency direction of the central theorem. The paper explicitly allows j_s=0 in minimally leak-interlacing models, but the D-matrix construction and the 'at least one coefficient exists' assertion require ν(p_{s,j_s}), which is undefined when j_s=0. This is not a sign typo or an off-by-one issue; it is a missing case in the main proof. I verified concretely that a valid minimally leak-interlacing model with j_1=0 exists (n=6, In={1,3}, Out={5}, Leak={2,4,6}), and a hand computation shows the intended block structure can be made to work if ν is interpreted as the next output in the following segment. Therefore the theorem is likely true, but the proof as written is incomplete for this family of cases. The proposed SIAN/Jacobian rank check would settle whether the gap is merely expository or whether the classification actually fails. Since no counterexample is known and the gap appears repairable, the appropriate verdict remains CONDITIONAL, not REJECT; this stress-test does not change the reader's verdict.","tokens_in":44931,"tokens_out":45943,"duration_ms":388147,"concrete_test":"Compute the Jacobian rank for the directed-cycle model with n=6, In={1,3}, Out={5}, Leak={2,4,6} using the coefficient-map formulas of Proposition 3.14 (or SIAN). The model has 9 parameters (6 edge rates + 3 leak rates); verify that the Jacobian has generic rank 9. If rank 9, Theorem 3.6 holds for this j_s=0 case and the proof gap is a repairable omission. If rank <9, the leak-interlacing characterization is false for this model. As a secondary check, build the 3x3 D matrix of Proposition 3.25 with rows κ(3,5), κ(1,5), e6−∏k_{i+1,i} and columns for leaks 4,2,6 given by J_{k0ℓ}−J_{k_{ℓ+1,ℓ}}, and confirm it is lower triangular with nonzero diagonal entries −k43, −k21k43k54, k21k32k43k54k65.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The sufficiency direction of Theorem 3.6 rests on Proposition 3.25. Its D-matrix row list includes κ(i_s, ν(p_{s,j_s})) for every block s=1,...,x−1, and the proof asserts that 'at least one coefficient exists' in each list, specifically κ(i_s, ν(p_{s,j_s})). But Definition 3.22 and the discussion after inequality (27) explicitly allow j_s=0, meaning an input segment with no output. When j_s=0, p_{s,j_s} does not exist, so ν(p_{s,j_s}) is undefined and the listed coefficient cannot be formed. A concrete minimally leak-interlacing model realizing this is n=6, In={1,3}, Out={5}, Leak={2,4,6}; here x=2, y=1, z=3, j_1=0, and the second block of D would need the undefined κ(1,ν(p_{1,0})). The subsequent 'at least one coefficient' claim also lists κ(i_1,ν(p_{1,j_1})) with j_1=0, which is undefined. Thus, as written, the block-triangularization argument does not cover a nonempty family of leak-interlacing models. This is an internal gap in the proof of the main theorem, independent of the cited rank criterion and coefficient-map formula. The gap appears repairable—interpreting ν(p_{s,0}) as the first output in the next segment makes the n=6 D-matrix lower triangular with nonzero diagonal—but the written proof is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies generic local identifiability of linear compartmental models whose underlying graph is a directed cycle or a bidirected path (catenary model). The main result, Theorem 3.6, states that a directed-cycle model is generically locally identifiable if and only if it is leak-interlacing: outside a small exceptional family, every directed path between consecutive leaks contains at least one input or output. The proof uses the rank of the Jacobian of the input-output coefficient map, the coefficient-map formula of Gerberding-Obatake-Shiu for one-input/one-output cycles, and a reduction (Proposition 2.14) from multi-input/multi-output models to one-input/one-output submodels. Section 4 identifies hyperplanes contained in the singular locus of identifiable directed-cycle models and gives the full singular locus for the case In=Out=Leak={1}. Section 5 states and proves a formula for the input-output equations of catenary models with one input and one output. Appendices contain SIAN-generated identifiability databases for small directed-cycle and catenary models.","tokens_in":45004,"tokens_out":26588,"duration_ms":221202,"significance":"If Theorem 3.6 is correct, it gives a complete and very simply checkable combinatorial characterization of generic local identifiability for all directed-cycle compartmental models, with arbitrary numbers of inputs, outputs, and leaks. This goes well beyond the earlier one-input/one-output results and is the first such classification for a family of models allowing any number of inputs and outputs, as the authors note. The catenary coefficient-map formula in Theorem 5.1 is a potentially useful new tool for future identifiability and singular-locus analyses, and the singular-locus results add explicit information about the measure-zero non-identifiable set. Strengths of the manuscript include worked examples with explicit Jacobians, the clear statements of the combinatorial conditions, and the reproducible SIAN databases in the appendices. The necessity direction and the catenary formula appear sound; however, the sufficiency proof of the main theorem has a load-bearing gap in the construction of the block matrix in Proposition 3.25, as detailed below.","major_comments":[{"comment":"The construction of the D block is undefined when an input segment contains no output, i.e. when j_s=0 for some s. The D rows are specified as κ(i_s,p_{s,1}),...,κ(i_s,p_{s,j_s}), κ(i_s,ν(p_{s,j_s})) for s<x, and the proof states that \"at least one coefficient exists in each list, namely ... κ(i_s,ν(p_{s,j_s}))\". But the discussion immediately after (27) explicitly allows j_s=0 and says that in that case p_{s,1} does not exist, so ν(p_{s,0}) is undefined. A concrete minimally leak-interlacing model realizing this is n=6, In={1,3}, Out={5}, Leak={2,4,6}; here x=2, y=1, j_1=0, and the s=1 block would require the undefined coefficient κ(1,ν(p_{1,0})), while the corresponding column list is also not well-formed. Since Proposition 3.25 is the sufficiency half of Theorem 3.6, the written proof does not cover a nonempty family of leak-interlacing models. The gap appears repairable—for example, by defining ν(p_{s,0}) to be the first output in the next segment, the n=6 example does satisfy the intended lower-triangular form—but the proof as written is incomplete.","section":"§3.3, Proposition 3.25 and inequality (27)"}],"minor_comments":[{"comment":"In the discussion of the upper-right zero block, the last row is said to correspond to the coefficient \"e*_1(ix,pβ)\", but no pβ has been defined; this should be e*_1(ix,px1).","section":"§3.3, Proposition 3.25, Case 1"},{"comment":"The displayed 2×2 matrix D appears to contain sign and index errors relative to the column order defined immediately before it. With columns J_{k0ℓ}-J_{kℓ+1,ℓ} and J_{k0n}-J_{k1n} and the row order κ(1,n), en-∏k_{i+1,i}, the first diagonal entry should be -κ/k_{ℓ+1,ℓ} and the second should be κ, rather than the displayed values; also the subscript k_{ℓ+2,ℓ} should presumably be k_{ℓ+2,ℓ+1}.","section":"§3.3, Proposition 3.25, Case 2"},{"comment":"Theorem 5.1 states n≥1, but the definition of OutM(ℓ) in (39)-(40) refers to k21 for ℓ=1, which does not exist when n=1. The n=1 case should be treated separately or the theorem should assume n≥2.","section":"§5, Theorem 5.1 and equation (39)"},{"comment":"Several rows of the database tables in Appendix A appear garbled or truncated in the typeset version, with parameter lists cut off mid-sentence. The authors should check the final typeset tables for completeness.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The main obstruction is the incomplete sufficiency proof in Proposition 3.25; the gap is localized and I expect it can be fixed without changing the statement or the overall proof strategy. The paper otherwise appears sound and within the scope of the journal. I would be willing to re-review a revision that completes the j_s=0 case."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does what it says: it completes the classification of generic local identifiability for directed-cycle linear compartmental models, resolving the questions raised by Gerberding-Obatake-Shiu, and it goes beyond prior work by allowing arbitrary numbers of inputs, outputs, and leaks. The leak-interlacing condition is natural, and the non-leak-interlacing direction is proven by explicit column dependencies that I spot-checked; the catenary coefficient-map formula (Theorem 5.1) is a genuinely useful new tool, and the appendix databases are a nice resource.\n\nThe main problem is in the sufficiency proof, Proposition 3.25. The D-matrix construction assumes every input segment contains at least one output: the row list uses κ(i_s, ν(p_{s,j_s})) even when j_s=0, which the paper explicitly allows in the discussion after (27). In that case p_{s,j_s} does not exist, so ν(p_{s,j_s}) is undefined, and the claimed 'at least one coefficient exists' fails. This is not a hypothetical edge case: the model n=6, In={1,3}, Out={5}, Leak={2,4,6} is minimally leak-interlacing with j_1=0. The gap is repairable—interpreting the missing term as the path from i_s to the first output after i_s makes the block matrix lower triangular with nonzero diagonal (I checked)—but the written proof is incomplete. This is an internal gap independent of the cited rank criterion.\n\nOther issues are minor: a sign discrepancy between Lemma 3.16(II) and the D-matrix display in Proposition 3.25; an invalid edge label k_{l+2,l} in the Case-2 display; the Theorem 4.5 proof asserts a minor-structure claim without derivation; and the index range in Theorem 5.1 is slightly loose. None of these threaten the main classification once the j_s=0 gap is fixed.\n\nThe paper is honest—conjectures are labeled, the singular-locus results are explicitly partial, and Remark 5.6 corrects an error in a cited result. The reliance on prior coefficient-map formulas from the same group is justified; those are the right tools.\n\nWho should read it: anyone working on structural identifiability of linear compartmental models, especially the algebraic/combinatorial approach. The classification result is the kind of clean statement that will be cited.\n\nRecommendation: send to peer review. The main theorem is important and likely correct, but the sufficiency proof needs a careful revision for the j_s=0 case before it is publishable.","headline":"Completes the directed-cycle identifiability classification with a genuinely new leak-interlacing criterion, but the sufficiency proof has a fixable gap when an input segment contains no output.","tokens_in":45823,"tokens_out":7282,"would_cite":true,"duration_ms":57022,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B30","37N25","92C45","34A30","34A55","05C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"A directed-cycle compartment model is generically locally identifiable exactly when its leaks interlace with its inputs and outputs, with one exceptional configuration set aside.","keywords":["linear compartmental model","structural identifiability","directed-cycle model","catenary model","leak-interlacing","input-output equations","coefficient map","singular locus"],"falsifier":"For the 4-compartment directed-cycle model with In={1}, Out={3}, and Leak={1,3} — a leak-interlacing model that Theorem 3.6 declares identifiable — evaluate the Jacobian matrix of the coefficient map (four edge parameters k21,k32,k43,k14 plus two leak parameters k01,k03) at a generic parameter point. If its determinant is the zero polynomial, or a computer algebra system yields two distinct parameter vectors with identical input-output coefficients, then the classification is false.","tokens_in":44498,"feed_emoji":"🔄","tokens_out":8622,"duration_ms":72588,"temperature":0.7,"pith_summary":"The paper resolves a question about when the rate constants in a cyclic compartment model can be recovered from noiseless input-output data. Its main theorem states that a directed-cycle model with at least one input and one output is generically locally identifiable exactly when it is leak-interlacing: between any two leak compartments along the cycle there is an input or an output, with one explicit exceptional family set aside. The proof works through the coefficients of input-output equations, reducing the question to whether the Jacobian of the coefficient map has full rank, and then showing this happens precisely for leak-interlacing models. For the companion class of catenary models, the paper gives an explicit formula for all coefficients of the input-output equations, which is expected to support future identifiability analysis of those models. The result gives a simple graph-theoretic test that decides identifiability without solving any differential equations.","feed_headline":"Leak placement decides which cycle models are identifiable","feed_subtitle":"A single interlacing condition now predicts generic local identifiability for every directed-cycle compartment model.","key_machinery":"The engine of the proof is the coefficient map c of the model, the vector of coefficients of the input-output equations, together with the rank criterion that a strongly connected linear compartmental model is generically locally identifiable exactly when the Jacobian of c has full column rank at a generic parameter point. For directed cycles the coefficient map has an explicit form (Lemma 3.10): in the one-input, one-output case it is built from elementary symmetric polynomials e_j of edge-plus-leak sums, the path product κ from input to output, and products e*_j κ from the output back to the input; Proposition 2.14 extends these coefficients to models with multiple inputs and outputs by restricting to one-input, one-output submodels. The identifiability proof separates into a failure direction, where two leak columns of the Jacobian are shown to be linearly dependent, and a success direction, where a block-triangular submatrix with nonzero diagonal blocks is exhibited for leak-interlacing models. For catenary models the corresponding coefficient-map formula is derived from the spanning-incoming-forest expansion of input-output equation coefficients, using elementary symmetric polynomials on sets of outgoing sums.","core_discovery":"On the paper's terms, the central discovery is the equivalence in Theorem 3.6: an n-compartment directed-cycle model is generically locally identifiable if and only if it is leak-interlacing, meaning it is not exceptional (the only obstruction is the configuration In={i}, Out={i-1}, |Leak|=2, with i-1 in Leak) and every arc of the cycle from one leak to the next contains at least one input or output compartment, counting endpoints. This covers any number of inputs, outputs, and leaks, extending earlier one-input/one-output results. A direct corollary is that any directed-cycle model with |Leak| at least |In|+|Out|+1 is unidentifiable. The paper also identifies several hyperplanes that always lie in the singular locus of an identifiable directed-cycle model (Theorems 4.3 and 4.5), and, for catenary models, proves a closed coefficient-map formula (Theorem 5.1) expressed through elementary symmetric polynomials and products of edge parameters along paths.","pith_inferences":["Beyond the paper: the same interlacing idea is a natural conjecture for other strongly connected compartmental graphs, replacing the directed cycle by requiring that between any two leaks along every directed path there is an input or output.","Beyond the paper: combining the catenary coefficient formula with the paper's reduction from multiple inputs and outputs to one-input/one-output submodels yields an explicit coefficient map for all catenary models, so a full leak-interlacing-style classification for bidirected paths is a testable next step.","Beyond the paper: the block-triangular structure used to prove the identifiable direction suggests that for leak-interlacing models the Jacobian of the coefficient map is sparse and triangular after column operations; if true, this would give a direct way to compute identifiable parameter combinations, not just the yes/no answer.","Beyond the paper: the singular-locus hyperplanes found for small leak-interlacing cycles suggest the singular locus in the full leak-interlacing family is a hyperplane arrangement; checking the n=4 and n=5 cases in the paper's database would test this."],"forward_implications":["For any directed-cycle model, the leak-interlacing condition gives a purely combinatorial check of generic local identifiability: list the leaks around the cycle and confirm each gap contains an input or an output, after excluding the single exceptional family.","Directed-cycle models with at least |In|+|Out|+1 leaks are always unidentifiable, so adding leaks beyond the number of inputs plus outputs cannot preserve identifiability.","The previous characterization for one-input, one-output directed-cycle models becomes a special case of Theorem 3.6, and the same theorem covers models with several inputs and outputs for the first time.","For identifiable directed-cycle models, Theorems 4.3 and 4.5 show the singular locus contains explicit hyperplanes; in the single-leak case with input and output at compartment 1, the full singular locus is the product of edge parameters times a Vandermonde factor in the leak-shifted edge parameters.","For catenary models with one input and one output, Theorem 5.1 gives an explicit coefficient-map formula usable in later identifiability computations."],"supporting_citations":[{"why":"It supplies the coefficient-map formula (Lemma 3.10) and the earlier one-input/one-output directed-cycle classification (Proposition 2.22) that Theorem 3.6 extends.","marker":"[12]"},{"why":"It supplies the input-output equation formula (Proposition 2.10) and the Jacobian rank criterion (Proposition 2.16) that turn identifiability into a linear-algebra question.","marker":"[19]"},{"why":"It establishes the equivalence between generic local identifiability and the coefficient-map formulation used in Definition 2.15.","marker":"[21]"},{"why":"It provides the spanning-incoming-forest coefficient formula (Lemma 5.5) on which the catenary coefficient-map theorem is built, and the bidirected-tree classification that motivates the paper.","marker":"[2]"},{"why":"It supplies Proposition 2.20, that adding inputs or outputs preserves identifiability, used to reduce leak-interlacing models to minimally leak-interlacing submodels.","marker":"[14]"},{"why":"It defines the singular locus and supplies the single-leak singular-locus result (their Theorem 5.3) that Theorem 4.5 generalizes.","marker":"[15]"}],"fun_headline_variants":["Leak-interlacing rule identifies all cycle models","Cycle model identifiability: leak placement matters","When leaks and inputs interlace, cycle models are identifiable","Interlacing leaks decide cycle model identifiability","New condition: leak-interlacing determines cycle model identifiability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that generic local identifiability is exactly the full-rank condition on the Jacobian of the coefficient map of the input-output equations, and that the cited coefficient-map formula for directed cycles is correct for every leak configuration; if either fails for some model, the leak-interlacing criterion would not decide identifiability.","fun_headline_variants_meta":{"raw":{"variants":["Leak-interlacing rule identifies all cycle models","Cycle model identifiability: leak placement matters","When leaks and inputs interlace, cycle models are identifiable","Interlacing leaks decide cycle model identifiability","New condition: leak-interlacing determines cycle model identifiability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000783,"raw_usage":{"total_tokens":3425,"prompt_tokens":883,"completion_tokens":2542,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":2463}},"tokens_in":499,"tokens_out":2542,"duration_ms":18295,"temperature":1.0,"reasoning_tokens":2463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:19:55.838878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the 4-compartment directed-cycle model with In={1}, Out={3}, and Leak={1,3} — a leak-interlacing model that Theorem 3.6 declares identifiable — evaluate the Jacobian matrix of the coefficient map (four edge parameters k21,k32,k43,k14 plus two leak parameters k01,k03) at a generic parameter point. If its determinant is the zero polynomial, or a computer algebra system yields two distinct parameter vectors with identical input-output coefficients, then the classification is false.","supporting_citations":[{"cited_title":"Identifiability of linear compartmental models: the effect of moving inputs, outputs, and leaks","cited_arxiv_id":null,"evidence_quote":"It supplies the coefficient-map formula (Lemma 3.10) and the earlier one-input/one-output directed-cycle classification (Proposition 2.22) that Theorem 3.6 extends."},{"cited_title":"Identifiability results for several classes of linear compartment models","cited_arxiv_id":null,"evidence_quote":"It supplies the input-output equation formula (Proposition 2.10) and the Jacobian rank criterion (Proposition 2.16) that turn identifiability into a linear-algebra question."},{"cited_title":"Input-output equations and identifiability of linear ode models","cited_arxiv_id":null,"evidence_quote":"It establishes the equivalence between generic local identifiability and the coefficient-map formulation used in Definition 2.15."},{"cited_title":"Linear compartmental models: input-output equations and operations that preserve identifiability","cited_arxiv_id":null,"evidence_quote":"It supplies Proposition 2.20, that adding inputs or outputs preserves identifiability, used to reduce leak-interlacing models to minimally leak-interlacing submodels."}],"review_version":1}