{"id":"5be1a535-c591-422b-babc-7c0379021d19","arxiv_id":"2412.12297","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"IMEX-BDF2 and IMEX-BDF3 are shown to be stable for DDE systems with non-commuting matrices under field-of-values conditions, with unconditional stability on disks of radius 1/3 and 1/7.","lead":"This paper derives sufficient stability conditions for implicit-explicit BDF time-stepping methods applied to delay differential equations, including cases where the stiff and delay matrices cannot be diagonalized together. It shows how field-of-values arguments yield step-size bounds, and tests them on delayed diffusion problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorems 15/16 establish monotonicity of σ_z only by numerical derivative plots; Lemma 18 and Theorem 24's h* bound depend on that monotonicity, so the central step-size guarantee is not yet proven.","rationale":"The paper's central contribution is Theorem 24: a sufficient, computable step-size bound from a field-of-values radius. Tracing the proof, the chain is: scalar stability region D_z ⊇ D(0,σ_z) (Cor. 11, via Prop. 14), then for non-commuting matrices z ranges over F(hA) and μ over F_p, then Lemma 18 lets one use only the most negative z_d = −λ_d h, giving h* = |χ(r)|/λ_d. The weakest link is not the field-of-values machinery (that part is standard and correctly reduces to the scalar characteristic equation) but the monotonicity of σ_z. If σ_z were not non-decreasing, D(0,σ_{z_d}) would not be contained in D(0,σ_z) for every z, and the step-size restriction would not follow from a disk radius alone. The current proof of Theorems 15 and 16 is explicitly numerical, so the condition is asserted rather than demonstrated. This is a proof-completeness problem, not a demonstrated counterexample; the numerical experiments are consistent with the theory and the monotonicity is very likely true. I therefore do not change the reader's CONDITIONAL verdict. The Eq. (29) μ-formula inconsistency is real but appears to be a typo because the derivation in Section 5.1 and the examples (p=0, F(A^{-1}B)) use the correct A^{p−1}B form. Section 6.2's own concession that nonlinear stability needs more study is a limitation but does not affect the linear central claim. The missing commit hash for the GitHub repository is a reproducibility concern, not a mathematical one.","tokens_in":24079,"tokens_out":14455,"duration_ms":117437,"concrete_test":"Prove the two monotonicity claims symbolically. For BDF2, substitute ψ2 from Eq. (17), compute dψ2/dz, clear denominators, and verify the resulting polynomial inequality is positive on (1/2(−10−9√2), −1/√2) with Sturm's theorem; ψ3'(z)=4/(3z^2)>0 is immediate. For BDF3, obtain the missing ψ̃2(z) from the repository file '2. IMEX BDF3 theory' and repeat the Sturm-sequence check on (−12.655874, −0.722965). If either check fails, recompute h* in Theorem 24 using the true minimum of ψ on [−λ_d h, −λ_1 h]; a dip invalidates the stated h*. If both pass, Lemma 18 is closed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 24 converts a field-of-values radius r into h* = |χ(r)|/λ_d (BDF2) or h* = |χ̃(r)|/λ_d (BDF3). The conversion is valid only if the scalar stability disk D(0, σ_z) is smallest at the most negative z in F(hA) = [−λ_d h, −λ_1 h]. This is exactly Lemma 18, which is stated as 'directly obtained from Theorems 15 and 16', i.e. from the assertions that ψ(z) = σ_z and ψ̃(z) = σ_z are non-decreasing on (−∞,0). The proofs of the two theorems do not prove monotonicity: they check the derivative signs numerically ('Numerically, it is easy to observe...', Figure 4). For BDF2, ψ2(z) in Eq. (17) is written explicitly but its derivative is not shown to be positive; for BDF3, ψ̃2(z) is not even displayed. If either function has a local dip, Lemma 18 can fail for some z ∈ F(hA), and the claimed h* would no longer be a sufficient step-size bound. The theorem may still be true, but the manuscript does not establish it as stated. Secondary gaps: Prop. 14 is justified by a continuity/Cauchy-integral sketch, and Eq. (29) writes μ with Bv instead of A^{p−1}Bv; both should be cleaned up, but monotonicity is the load-bearing one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the stability of IMEX-BDF2 and IMEX-BDF3 methods applied to the linear delay test equation y'(t) = -A y(t) + B y(t-τ), where A is positive definite and B is arbitrary. It first studies the scalar characteristic equation, derives the functions ψ(z) and ψ̃(z) that give the minimal stability radius σ_z, and obtains sufficient stability conditions for the simultaneously diagonalizable case. The central contribution is Theorem 24 in Section 5, which uses the field of values to prove that if the scaled field of values F_p is contained in a disk of radius r, then both methods are stable for h ≤ |χ(r)|/λ_d (IMEX-BDF2) and h ≤ |χ̃(r)|/λ_d (IMEX-BDF3), with unconditional stability when F_p lies inside D(0,1/3) or D(0,1/7), respectively. Numerical experiments for DDEs and PDDEs illustrate the predicted step-size restrictions and convergence orders.","tokens_in":24274,"tokens_out":9642,"duration_ms":84887,"significance":"If Theorem 24 is fully established, the paper gives a genuinely useful and computable sufficient stability test for IMEX-BDF methods beyond the simultaneous-diagonalization setting; the bounds depend only on the largest eigenvalue of A and the radius of a field of values, and the paper provides reproducible symbolic and numerical codes for the constructions. The numerical experiments in Sections 5 and 6 support the qualitative predictions and show that the conditions are not vacuous. The main reservation is that a load-bearing monotonicity assertion in the proof of Theorem 24 is verified only numerically, which leaves the central step-size guarantee not completely proven as stated.","major_comments":[{"comment":"The monotonicity of ψ(z) and ψ̃(z) on (-∞,0) is asserted in Theorems 15 and 16, but the proofs check the sign of the derivatives only numerically ('Numerically, it is easy to observe...') with reference to Figure 4, and for IMEX-BDF3 the derivative of ψ̃2(z) is not even displayed. This monotonicity is exactly what Lemma 18 uses to conclude that D(0,σ_{z2}) ⊆ D(0,σ_{z1}) for z2 < z1 < 0, and Theorem 24 uses that conclusion to replace the entire interval F(hA) by its most negative endpoint z_d. If either function had a local minimum, the claimed bound h* would no longer be sufficient. Please provide an analytic proof of the derivative sign conditions, or an interval-arithmetic/computer-assisted certificate, or alternatively weaken the statement of Theorem 24 by explicitly assuming the numerically observed monotonicity.","section":"Section 3.1, Theorems 15–16, Lemma 18, and Theorem 24"},{"comment":"The displayed formula μ = ⟨v,Bv⟩/⟨v,A^p v⟩ omits the factor A^{p-1} before B. The derivation leading to Eq. (28) and the definition of F_p in Eq. (26) require μ = ⟨v,A^{p-1}Bv⟩/⟨v,A^p v⟩. This is not merely a typographical issue: the p=0 examples use F(A^{-1}B), which corresponds to ⟨v,A^{-1}Bv⟩/⟨v,v⟩, not ⟨v,Bv⟩/⟨v,v⟩. As written, the proof of Theorem 24 does not connect the field-of-values condition to the scalar characteristic equation. The formula and the surrounding derivation should be corrected and checked.","section":"Section 5.1, Eq. (29)"}],"minor_comments":[{"comment":"The estimates 'O(h)^3' and 'O(h)^4' should read 'O(h^3)' and 'O(h^4)', respectively.","section":"Section 2.1, Theorems 3 and 4"},{"comment":"There are inconsistent cross-references: the proof of Theorem 24 cites 'Proposition 11' where Corollary 11 is meant, and the proof of Corollary 25 cites 'Lemma 11' where Lemma 18 is meant.","section":"Sections 4 and 5.1"},{"comment":"The displayed bound contains a typo: '|χ(|µi)|' should be '|χ(|µi|)|'.","section":"Section 4, Theorem 19"},{"comment":"In the formula for μ_{m,z,θ} of IMEX-BDF3, '3e2itheta' should be '3e^{2iθ}'.","section":"Section 3.1"},{"comment":"The remark after Definition 22 is numbered 'Remark 1', although earlier remarks are numbered 12 and 20; it should be renumbered.","section":"Section 5.1"},{"comment":"In the h=0.05 row for IMEX-BDF2, the second error entry 4.9934·10^2 is inconsistent with the other entries and with the reported convergence rate; please check whether the exponent should be -2.","section":"Section 4.2, Table 1"},{"comment":"The proof of Proposition 14 appeals to Cauchy's integral theorem without presenting the argument in detail; please either expand the proof or cite the exact argument in [19], since the equality D_z = D̃_z is used implicitly in the figures and in the identification of σ_z.","section":"Section 3, Proposition 14"}],"recommendation":"major_revision","confidential_remarks":"The central idea is promising and the numerical evidence is consistent, but the proof of Theorem 24 currently rests on a numerically checked monotonicity claim rather than a rigorous argument. This is a concentrated and, in my view, fixable gap: supplying an analytic derivative sign proof for the two functions (or a rigorous enclosure) would place the main theorem on solid ground. I would not recommend rejection on the basis of the current gap, but the manuscript should not be accepted until the monotonicity issue is resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report — I read the paper and the stress-test note, and the note is on target. The genuinely new thing here is the first field-of-values stability treatment for IMEX-BDF2/3 on delay systems with non-commuting A and B, including the explicit unconditional disks D(0,1/3) and D(0,1/7). The scalar stability analysis is careful: the sigma_z functions are derived explicitly, the chi/chitilde inversions are done with clear breakpoints, and the DDE and PDDE experiments fit the theory, including the observed instability thresholds. That part is solid.\n\nThe soft spot is exactly where the stress-test puts it. Theorems 15 and 16 assert that psi and psitilde are non-decreasing on (-infinity,0), but the proof just says 'numerically it is easy to observe' that the derivatives are nonnegative and points to Figure 4. That monotonicity is load-bearing: Lemma 18 and Theorem 24 depend on the smallest stability disk occurring at the most negative z, which is what converts the field-of-values radius r into h* = |chi(r)|/lambda_d. If psi_2 or psitilde_2 has a dip anywhere, the bound could fail. The numerics suggest it doesn't, and the final bound is probably correct, but as written the theorem isn't proven. This is a fixable gap: for psi_3 the derivative is 4/(3z^2), for psi_1 it's zero, and psi_2 has an explicit formula, so an analytic derivative sign check should be possible with some algebra; psitilde_2's formula is omitted, so the authors need to provide it or a proof.\n\nSecondary issues: Proposition 14's identification of the stability region with the innermost boundary curve is a sketch (continuity plus a Cauchy-integral statement), and in the proof of Theorem 24, equation (29) writes mu = <v,Bv>/<v,A^p v>, but from the earlier derivation it should be <v,A^(p-1)Bv>/<v,A^p v> — likely a typo, but in a key proof it is confusing. Section 6.2's nonlinear example is explicitly heuristic; the authors concede the complete nonlinear analysis is open, so that's a stated limitation rather than a hidden flaw. The GitHub repo is referenced without a commit hash, which is minor.\n\nWho should read this: anyone working on IMEX stability for DDEs/PDDEs. It deserves a serious referee, because the technique is relevant and the gap is repairable. I would send it to review, but my review would be conditional: fix the monotonicity proof, supply psitilde_2's formula, clean up Proposition 14 and the Eq. (29) typo, and the theory would stand.","headline":"First field-of-values stability bounds for IMEX-BDF2/3 on delay systems with non-commuting matrices, but the key monotonicity claim is only numerically verified, so the headline step-size bound is not yet proven.","tokens_in":24944,"tokens_out":4130,"would_cite":true,"duration_ms":33408,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65L20","65M20","65L06"],"pacs":[],"model":"deepseek-v4-flash","headline":"IMEX-BDF2 and BDF3 are stable for non-commuting delay systems when a scaled field of values lies in a disk, with explicit step bounds and unconditional stability inside radii $1/3$ and $1/7$.","keywords":["numerical stability","IMEX BDF methods","delay differential equations","unconditional stability","field of values","partial delay differential equations","method of lines","step-size bounds"],"falsifier":"Directly compute the analytical derivative of $\\psi_2(z)$ or $\\tilde\\psi_2(z)$ on the stated intervals and evaluate it at two points $z_2<z_1<0$; any negative value, or any pair with $\\sigma_{z_2}>\\sigma_{z_1}$ from a direct $\\min_\\theta|\\mu(z,\\theta)|$ calculation, would break Lemma 18 and invalidate the step-size bound in Theorem 24.","tokens_in":23728,"feed_emoji":"📐","tokens_out":11475,"duration_ms":92219,"temperature":0.7,"pith_summary":"This paper studies when IMEX-BDF2 and IMEX-BDF3 remain stable for systems of delay differential equations of the form $y'(t)=-Ay(t)+By(t-\\tau)$, where $A$ is positive definite Hermitian and $B$ is arbitrary. The classical scalar test equation covers only the case where $A$ and $B$ diagonalize simultaneously, so the paper extends the analysis to non-commuting matrices by replacing eigenvalues with the field of values of the scaled matrix $A^{p/2-1}BA^{-p/2}$. The main result is a sufficient condition: if that field of values fits inside a disk of radius $r$, then the methods are stable for $h\\le|\\chi(r)|/\\lambda_d$ (BDF2) or $h\\le|\\tilde\\chi(r)|/\\lambda_d$ (BDF3), where $\\lambda_d$ is the largest eigenvalue of $A$; inside disks of radius $1/3$ (BDF2) or $1/7$ (BDF3), stability holds for every step size. This gives a computable stability test for DDE systems and, after the method of lines, for parabolic partial delay differential equations.","feed_headline":"One disk test sets safe step sizes for delayed IMEX-BDF","feed_subtitle":"For non-commuting matrices, stability follows from the field of values plus the largest eigenvalue of A.","key_machinery":"The central machinery is the field of values $F(X)=\\{\\langle x,Xx\\rangle:\\|x\\|=1\\}$, the set of all Rayleigh quotients of a matrix; for Hermitian $A$ it is a real interval, and the scaled field $F_p=F(A^{p/2-1}BA^{-p/2})$ plays the role of the eigenvalues of $A^{-1}B$ when the two matrices do not commute. The paper couples this with the scalar functions $\\psi(z)$ and $\\tilde\\psi(z)$, which for each negative $z$ give the radius $\\sigma_z$ of the largest disk around the origin contained in the scalar stability region of BDF2 or BDF3, respectively. Their inverses $\\chi(r)$ and $\\tilde\\chi(r)$ convert a field-of-values radius $r$ into the maximal allowed $|z|$; dividing by $\\lambda_d$ turns that into the step-size bound. The load-bearing monotonicity of $\\psi$ and $\\tilde\\psi$ on $(-\\infty,0)$ is what allows Lemma 18 to compare stability disks and Theorem 24 to use only the most negative $z$.","core_discovery":"On its own terms, the paper establishes Theorem 24: for the linear delay system $y'(t)=-Ay(t)+By(t-\\tau)$ with $A$ Hermitian positive definite, let $\\lambda_d$ be the largest eigenvalue of $A$, and suppose that for some real $p$ the set $F_p=F(A^{p/2-1}BA^{-p/2})$ is contained in the disk $D(0,r)$ with $0<r\\le1$. Then IMEX-BDF2 is stable for $h\\le|\\chi(r)|/\\lambda_d$ and IMEX-BDF3 for $h\\le|\\tilde\\chi(r)|/\\lambda_d$, with $\\chi$ and $\\tilde\\chi$ the inverse functions built from the scalar stability radii in Section 3. Since the scalar radius functions tend to $1/3$ and $1/7$ as the scaled step $z$ tends to $-\\infty$, the paper also obtains unconditional stability whenever $F_p\\subseteq D(0,1/3)$ for BDF2 or $F_p\\subseteq D(0,1/7)$ for BDF3. The proof passes through a scalar characteristic equation by multiplying with $A^{p-1}$ and forming Rayleigh quotients, then uses the monotonicity of the scalar radius functions to reduce the entire interval of possible $z$ values to the single most negative value $z_d=-\\lambda_d h$. The simultaneous-diagonalization case is treated first and follows as a special case, and the numerical experiments apply the theorem to DDE systems and to parabolic PDDEs semidiscretized by the method of lines.","pith_inferences":["Replacing the containing disk by the convex hull of $F_p$, or by a minimal enclosing ellipse, is a natural way to sharpen the step-size bound; the paper does not explore this.","The same construction of $\\psi$ and $\\chi$ functions could be repeated for higher-order IMEX-BDF methods, giving analogous sufficient conditions if the scalar stability regions are computed.","For the nonlinear delayed Burgers example, the stability analysis linearizes the delayed term at the initial history; a complete nonlinear stability proof remains open, and the observed stability for $h>h^*$ suggests the linear bound is conservative.","In the method-of-lines setting $\\lambda_d$ grows like $(\\Delta x)^{-2}$, so the proved bound should scale like $(\\Delta x)^2$; testing that scaling would show how sharp the sufficient condition is for parabolic problems."],"forward_implications":["Inside $D(0,1/3)$ for BDF2 or $D(0,1/7)$ for BDF3, no step-size restriction is needed for linear stability.","For non-commuting $A$ and $B$, the stability check becomes a numerical-radius computation for $A^{p/2-1}BA^{-p/2}$ instead of a simultaneous-diagonalization eigenvalue problem.","If $A^{-1}B$ is normal and the delay system is stable, the explicit bounds are $h\\le1/(\\lambda_d\\sqrt2)$ for BDF2 and $h\\le0.722965/\\lambda_d$ for BDF3.","Semidiscretizing a parabolic PDDE by the method of lines produces a DDE system to which Theorem 24 applies directly, with $A$ the diffusion matrix and $B$ the delayed reaction matrix.","All the stability conditions are sufficient but not necessary, and the numerical examples show stability for some step sizes larger than the proved bounds."],"supporting_citations":[{"why":"Supplies the coefficient-construction condition, the BDF2 and BDF3 formulas, and the lemma connecting the scalar A-stability region and $\\sigma_z$ to the P-stability region.","marker":"[15]"},{"why":"Establishes the field-of-values route to unconditional stability of multistep IMEX schemes for ODEs, which the paper adapts to the delayed case.","marker":"[20]"},{"why":"Provides the DDE-specific reduction of the stability region to the innermost curve and the matrix characteristic-equation criterion used in Proposition 23.","marker":"[19]"},{"why":"Provides the field-of-values properties, including convexity, spectral containment, and the normal-matrix identity for the numerical radius.","marker":"[9]"},{"why":"Supplies the linear-system stability condition that yields containment of $F(A^{-1}B)$ in the unit disk when $A^{-1}B$ is normal.","marker":"[10]"},{"why":"Shows how field-of-values stability analysis is applied to systems without delay and how nonlinear terms are linearized at an intermediate point.","marker":"[6]"}],"fun_headline_variants":["Safe IMEX-BDF steps for delays via disk condition","Field of values disk test sets IMEX-BDF stability","Disk radius gives stable step sizes for delayed IMEX-BDF","Non-commuting delays: disk test for safe IMEX-BDF steps","Stable IMEX-BDF for delay systems under disk condition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the scalar functions $\\psi(z)$ and $\\tilde\\psi(z)$ are non-decreasing for all negative $z$; this is checked in the paper by plotting numerical derivatives (Figure 4) rather than by an analytic proof, and the step-size bound would not follow if this monotonicity failed.","fun_headline_variants_meta":{"raw":{"variants":["Safe IMEX-BDF steps for delays via disk condition","Field of values disk test sets IMEX-BDF stability","Disk radius gives stable step sizes for delayed IMEX-BDF","Non-commuting delays: disk test for safe IMEX-BDF steps","Stable IMEX-BDF for delay systems under disk condition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001028,"raw_usage":{"total_tokens":4382,"prompt_tokens":1043,"completion_tokens":3339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":3252}},"tokens_in":659,"tokens_out":3339,"duration_ms":22173,"temperature":1.0,"reasoning_tokens":3252,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:15:10.906539+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly compute the analytical derivative of $\\psi_2(z)$ or $\\tilde\\psi_2(z)$ on the stated intervals and evaluate it at two points $z_2<z_1<0$; any negative value, or any pair with $\\sigma_{z_2}>\\sigma_{z_1}$ from a direct $\\min_\\theta|\\mu(z,\\theta)|$ calculation, would break Lemma 18 and invalidate the step-size bound in Theorem 24.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the coefficient-construction condition, the BDF2 and BDF3 formulas, and the lemma connecting the scalar A-stability region and $\\sigma_z$ to the P-stability region."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the field-of-values route to unconditional stability of multistep IMEX schemes for ODEs, which the paper adapts to the delayed case."},{"cited_title":"Rodr ´ ıguez-Fern´ andez and J","cited_arxiv_id":null,"evidence_quote":"Provides the DDE-specific reduction of the stability region to the innermost curve and the matrix characteristic-equation criterion used in Proposition 23."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the field-of-values properties, including convexity, spectral containment, and the normal-matrix identity for the numerical radius."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the linear-system stability condition that yields containment of $F(A^{-1}B)$ in the unit disk when $A^{-1}B$ is normal."},{"cited_title":"Conte, J","cited_arxiv_id":null,"evidence_quote":"Shows how field-of-values stability analysis is applied to systems without delay and how nonlinear terms are linearized at an intermediate point."}],"review_version":1}