{"id":"8e3e7afb-25f7-4d5c-9744-2f53f55f31b5","arxiv_id":"2511.20037","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"There exist stable switched linear systems whose optimal and worst-case value functions are Lipschitz but non-differentiable on dense subsets of the state space.","lead":"This paper proves that the optimal and worst-case value functions of stable switched linear systems can be non-differentiable on dense sets of states, even with smooth quadratic costs. The result tells control and reinforcement-learning researchers that exact value functions may be too irregular for algorithms that search over piecewise smooth functions.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"n>2 proof uses a cost function inconsistent with Lemma 13, so Theorem 3 is not established beyond n=2 as written.","rationale":"The n=2 construction is careful and largely convincing: the interval-exchange argument, the density of the backward orbit, and the induction in Lemma 11 are coherent. The Lipschitz theorem is standard. But the paper's own Lemma 13 is asserted to be 'straightforward (omitted)' and is in fact inconsistent with the written n>2 cost function. Since the high-dimensional claim is part of the headline theorem, the proof as submitted does not fully support Theorem 3. This is not a question of disagreeing with consensus; it is an internal inconsistency at a load-bearing step. The fix is explicit and small, so conditional acceptance is the right posture rather than rejection. Two smaller slips reinforce this posture: Corollary 3's dense set includes the origin, where J* is differentiable, and Corollary 4's last displayed equation writes 'min' where the worst-case Bellman equation (3) requires 'max'. The reader's rationale already mentions the n>2 cost inconsistency, but the reader's stated weakest assumption was the Bellman uniqueness in Corollary 4, so agreement is partial rather than full.","tokens_in":11479,"tokens_out":16662,"duration_ms":155165,"concrete_test":"Set n=3, \\alpha=\\arctan(3/4), \\rho=0.01, \\theta=\\pi/4, and x=(\\cos\\theta,\\sin\\theta,0). Compute A=\\hat A_1 x and compare \\hat c(A)=\\cos^2(A_1)+2\\sin^2(A_2) with \\rho^2(\\cos^2(\\theta+\\alpha)+2\\sin^2(\\theta+\\alpha)). A mismatch at this single point confirms that Lemma 13 is false for the printed cost; if instead the cost is corrected to x_1^2+2x_2^2, the Bellman recursions (2)–(3) reduce exactly to the n=2 equations on the first two coordinates, which would repair the n>2 part of Theorem 3.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In §IV.A.2 the n>2 construction takes \\hat A_i as block-diagonal scaled rotations in the first two coordinates and zero elsewhere, and defines the cost as \\hat c(x)=cos^2(x_1)+2 sin^2(x_2). Lemma 13 then asserts \\hat J^*(x)=J^*([x_1,x_2]^\\top) and the same for \\hat J^\\circ, where J^*,J^\\circ are the n=2 value functions for the quadratic cost c_2(u)=u_1^2+2u_2^2. That equality cannot hold. For an initial state x(\\theta)=(\\cos\\theta,\\sin\\theta,0,\\dots,0), one step under \\hat A_1 gives \\rho(\\cos(\\theta+\\alpha),\\sin(\\theta+\\alpha),0,\\dots,0). With the printed \\hat c, \\hat c(\\hat A_1 x(\\theta)) = \\cos^2(\\rho\\cos(\\theta+\\alpha))+2\\sin^2(\\rho\\sin(\\theta+\\alpha)), whereas the n=2 recursion needs \\rho^2(\\cos^2(\\theta+\\alpha)+2\\sin^2(\\theta+\\alpha)). These are not equal; take \\theta=\\pi/4, \\alpha=\\arctan(3/4), \\rho=0.01. Thus Lemma 13 is false as stated, and the proof of Corollary 5 and hence Theorem 3 for n>2 fails. The intended repair—using c(x)=x_1^2+2x_2^2 for all n—would restore the argument, but that is not the cost function written in the manuscript. Because Theorem 3 explicitly claims every n\\ge 2, this is a load-bearing gap in the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the value functions of stable discrete-time switched linear systems under controlled and arbitrary switching. It first proves that if the cost is Lipschitz on a neighborhood of the origin, then both the optimal value function J* and the worst-case value function J^∘ are Lipschitz (Theorem 2). The main contribution is Theorem 3, which claims that for every n≥2 there exist rational, stable switched linear systems with C^∞, Lipschitz costs whose value functions are non-differentiable on dense subsets of R^n. The proof for n=2 uses a two-mode system of scaled rotations; the optimal switching law is related to an interval-exchange map, and the backward orbit of a special angle yields a dense non-differentiability set. The n>2 case is intended to follow by block-diagonal embedding of the 2D construction.","tokens_in":11884,"tokens_out":9551,"duration_ms":80173,"significance":"If the construction were fully correct, this would be a significant negative result for the exact computation of value functions of switched linear systems: it would show that no piecewise-smooth template can represent J* or J^∘ in general, despite the functions being differentiable almost everywhere. The n=2 proof is genuinely constructive, with explicit rational matrices, a transparent interval-exchange-map mechanism, and inequalities that are carefully checked. The Lipschitz theorem is clean and useful. However, the extension to n>2 contains a concrete error in the cost function, and the worst-case half of the n=2 proof has a sign error and an unverified uniqueness step. As submitted, Theorem 3 is not established beyond the 2D optimal half.","major_comments":[{"comment":"The cost function used for n>2 is \\hat c(x)=cos^2(x_1)+2 sin^2(x_2). For x=(cosθ,sinθ,0,...,0), \\hat c(\\hat A_1 x)=cos^2(ρ cos(θ+α))+2 sin^2(ρ sin(θ+α)), whereas the 2D recursion with the quadratic cost c_2(u)=u_1^2+2u_2^2 requires ρ^2(cos^2(θ+α)+2 sin^2(θ+α)). These are not equal; e.g., θ=π/4, α=arctan(3/4), ρ=0.01. Hence Lemma 13 is false as stated, and the proof of Corollary 5 and Theorem 3 for n>2 fails. The intended repair—using c(x)=x_1^2+2x_2^2 in every dimension—would restore the decoupling argument, but that is not the cost written in the manuscript.","section":"§IV.A.2, Lemma 13"},{"comment":"The displayed derivation for \\tilde J^∘ ends with 'min{...}' but the algebra, and the definition of the worst-case value, give 'max{...}'. With the printed 'min', the constructed function does not satisfy the worst-case Bellman equation (3), so the conclusion that it equals J^∘ is unsupported. The sign in the final displayed equation must be corrected.","section":"§IV.A.1, Corollary 4"},{"comment":"The step 'This implies that J^∘ is the worst-case value function, by classical arguments ... [19, Thm 3.1]' needs to be made precise. The paper neither states the theorem nor verifies its hypotheses, including its applicability to a supremum (rather than infimum) Bellman equation, the required growth condition on solutions, and the normalization at the origin. Without this, a continuous solution of (3) need not be the true worst-case value. Please either quote the exact result with a hypothesis check, or prove uniqueness directly for this homogeneous quadratic example.","section":"§IV.A.1, Corollary 4"},{"comment":"Even after the cost function is corrected, the proof of Lemma 13 is only 'Straightforward (omitted)'. Since the lemma is load-bearing for the n>2 claim, a short proof should be included, explicitly using the decoupled dynamics and the additive structure of the cost.","section":"§IV.A.2, Lemma 13"}],"minor_comments":[{"comment":"In the proof, the phrase '˜J ⋆ is not differentiable is not differentiable at η' contains a duplicated clause. Please fix this typo.","section":"§IV.A.1, Lemma 12"},{"comment":"The definition J^∘(x)=r^2 \\tilde J^∘(θ) with r≥0 is ambiguous at r=0. The value at the origin should be stated explicitly (it is 0 by continuity and c(0)=0).","section":"§IV.A.1, Corollary 4"},{"comment":"The paper cites [19, Thm 3.1] for Bellman uniqueness but does not reproduce the theorem. A precise statement in the appendix, or at least a precise reference with the assumptions, would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The n>2 gap is a straightforward repair (replace the trigonometric cost by the homogeneous quadratic cost), and the 2D optimal half appears sound. I recommend major revision rather than rejection. The paper's central claim is defensible, but the submitted version contains a false lemma and a sign error that block the main theorem as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The n=2 construction is the real substance here, and it is good. The paper shows that for a stable switched linear system with a smooth quadratic cost, the optimal and worst-case value functions can be non-differentiable on dense subsets. The interval-exchange-map argument is elegant, the inequalities in the key lemmas check out, and the connection to prior work ([8], [17]) is honest. This is a genuine new result for n=2, and it is worth having on record.\n\nThe n>2 extension, however, has a real gap. The proof defines the cost as c(x)=cos^2(x1)+2 sin^2(x2), but Lemma 13 identifies the value function with the n=2 quadratic-cost value function. That identification requires the Bellman recursion for the projected state to match the n=2 recursion, which forces the cost at each step to be the quadratic form x1^2+2x2^2. The printed cost does not do that: after one step you get cos^2(ρ cos(θ+α))+2 sin^2(ρ sin(θ+α)) rather than ρ^2(cos^2(θ+α)+2 sin^2(θ+α)). The stress-test example with θ=π/4, α=arctan(3/4), ρ=0.01 makes the discrepancy concrete. So Theorem 3 is not established for n>2 as written. The fix is immediate—use c(x)=x1^2+2x2^2 for all n—so the result is almost certainly true, but the manuscript as written does not prove it.\n\nTwo smaller issues. Corollary 3 claims non-differentiability on a dense subset but includes the origin, where J* is actually differentiable because homogeneity degree 2 gives J*(x)=O(||x||^2). Excluding the origin fixes that. The worst-case part relies on uniqueness of continuous solutions to the Bellman max equation, which is standard but not proved or precisely referenced; a one-line justification would remove the last soft spot.\n\nThe central n=2 result is solid, the flaws are localized and fixable, and the paper deserves a serious referee. I would send it to review with a request to repair the n>2 cost function, exclude the origin, and pin down the Bellman uniqueness argument.","headline":"Solid n=2 construction showing dense non-differentiability of value functions for switched linear systems, but the n>2 extension as written has a fixable gap in the cost function.","tokens_in":12347,"tokens_out":5607,"would_cite":true,"duration_ms":52255,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49L20","93C30","37E05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that value functions of stable switched linear systems can be non-differentiable on dense subsets of the state space, even when the running cost is smooth and Lipschitz.","keywords":["switched linear systems","value function","non-differentiability","dense subset","Bellman equation","interval exchange map","joint spectral radius","optimal control"],"falsifier":"For the worst-case half, find two distinct continuous functions that both solve J(x)=c(x)+max_i J(A_i x) on a neighborhood of the origin and vanish at 0 for a stable switched linear system with smooth Lipschitz cost; that would break the uniqueness step used to identify the constructed function as the value function. Alternatively, for the paper's concrete two-rotation example, display a periodic backward orbit of the interval-exchange map T; the density argument rules such an orbit out.","tokens_in":11354,"feed_emoji":"📉","tokens_out":7927,"duration_ms":83820,"temperature":0.7,"pith_summary":"The paper asks how irregular the value function of a stable switched linear system can be. It establishes that although these value functions are always Lipschitz continuous—values cannot vary faster than a fixed multiple of the distance between states—they need not be piecewise smooth: for every dimension at least two, there are examples with a smooth quadratic cost where both the optimal and worst-case value functions are non-differentiable on dense subsets (sets that pass arbitrarily close to every state). These examples are stable, with joint spectral radius below one, and can even use rational transition matrices. This matters because exact computation of such value functions—central to optimal control and reinforcement learning—cannot then be achieved by algorithms that search over piecewise-smooth or smooth templates; any exact representation must admit dense pointwise non-differentiability, even though the functions remain differentiable almost everywhere.","feed_headline":"Dense nondifferentiability hits stable switched-linear value functions","feed_subtitle":"Optimal-control and RL algorithms that search piecewise-smooth templates cannot compute these functions exactly.","key_machinery":"The central mechanism is an interval exchange map: a piecewise translation of an interval, cut at a point ν, with the two pieces swapped. Here it takes the form T(θ)=θ+α for θ<ν and T(θ)=θ+β for θ≥ν on an interval of angles, and it is conjugate to an irrational rotation, so every backward orbit is dense. The scaled-rotation matrices reduce the planar dynamics to angle shifts, the homogeneity of the quadratic cost gives scaling identities, and the Bellman (dynamic programming) equation carries non-differentiability from the distinguished point ν backward along the dense orbit.","core_discovery":"The central result, Theorem 3, states that for every n≥2 there is a stable switched linear system and a smooth Lipschitz cost such that the associated optimal and worst-case value functions are non-differentiable on dense subsets of R^n. The construction uses two scaled rotation matrices with angles α and α−π/2, where α/π is irrational, and the quadratic cost c(x)=x1^2+2x2^2; the joint spectral radius is 0.01, so the system is stable. The proof isolates a point on the unit circle where the optimal value function has a kink; the optimal switching dynamics act on the angle coordinate as an interval exchange map, and the backward orbit of that kink point is dense. The dynamic programming equati","pith_inferences":["A natural testable extension is to vary the rotation angle: when α/π is rational, the interval exchange map becomes periodic, so the same mechanism would likely yield piecewise-smooth or finitely nonsmooth value functions, isolating irrationality as the driver of dense non-smoothness.","Although the non-differentiability set is dense, it has measure zero; smooth function-approximation classes may still approximate the value function in norm, but pointwise exact recovery at the dense points would be impossible within finite-parameter smooth templates.","The proof suggests a broader correspondence between value-function irregularity and the orbit complexity of the optimal switching law: minimal or ergodic switching dynamics should produce dense non-smoothness, while periodic switching dynamics should not."],"forward_implications":["Exact value functions of stable switched linear systems can lie outside the class of piecewise-smooth functions, so algorithms searching that class cannot compute them exactly in finite time.","The non-differentiability occurs even with rational transition matrices, so it is not an artifact of specially chosen irrational data and can occur with nonzero probability for finite-precision matrices.","The value functions remain Lipschitz and hence differentiable almost everywhere; the dense non-differentiability set has Lebesgue measure zero, so sampling-based or generic derivative checks will miss it.","Locating extrema of these value functions cannot rely on gradients, because derivatives fail on a dense set.","The paper leaves open sufficient conditions for piecewise differentiability and identifies that as the announced next step."],"fun_headline_variants":["Smooth costs can't prevent dense nondifferentiability in switched linear value functions","Dense nondifferentiability occurs in value functions of stable switched linear systems","Switched linear systems: value functions can be nondifferentiable on dense subsets","Even smooth Lipschitz costs yield dense nondifferentiability in control value functions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The worst-case half of the result rests on the uniqueness of continuous solutions of the dynamic-programming maximum equation with zero at the origin; if two distinct such solutions existed, the constructed function could satisfy the equation without being the true worst-case value.","fun_headline_variants_meta":{"raw":{"variants":["Smooth costs can't prevent dense nondifferentiability in switched linear value functions","Dense nondifferentiability occurs in value functions of stable switched linear systems","Switched linear systems: value functions can be nondifferentiable on dense subsets","Even smooth Lipschitz costs yield dense nondifferentiability in control value functions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001006,"raw_usage":{"total_tokens":4045,"prompt_tokens":654,"completion_tokens":3391,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":398,"completion_tokens_details":{"reasoning_tokens":3309}},"tokens_in":398,"tokens_out":3391,"duration_ms":23920,"temperature":1.0,"reasoning_tokens":3309,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:24:27.874660+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the worst-case half, find two distinct continuous functions that both solve J(x)=c(x)+max_i J(A_i x) on a neighborhood of the origin and vanish at 0 for a stable switched linear system with smooth Lipschitz cost; that would break the uniqueness step used to identify the constructed function as the value function. Alternatively, for the paper's concrete two-rotation example, display a periodic backward orbit of the interval-exchange map T; the density argument rules such an orbit out.","supporting_citations":[],"review_version":1}