{"id":"b67604f3-acbe-4620-9f7d-b59a1b6264e8","arxiv_id":"2504.12269","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"SEROAISE estimates larger certified regions of attraction for PWA and ReLU dynamical systems by growing a certified invariant set via the NUGIS procedure and then solving a linear program for a Lyapunov-like function on that set.","lead":"This paper introduces a method to compute larger regions of attraction for control systems whose dynamics are piecewise affine or ReLU neural networks. It works by first finding a certified invariant set, then verifying a Lyapunov-like function only inside that set, yielding estimates the authors report as larger than prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LP certificate in §5.2 is unsound: (35) checks the barrier condition only at vertices of a partition aligned with the previous barrier, so the new h's zero set can cross cell interiors and the required pointwise condition (7) can fail there.","rationale":"The reader correctly identifies Section 5.2 as the load-bearing weak point: the certificate that underpins the claimed larger RoA estimates is not sound as written. However, the reader's specific diagnosis is partly inaccurate: because α_m(h)=α0 h for h≥0 under (32)-(33), a vertex placed inside the set is not checked with a weaker slope just because it falls in category I_UC or I_excl. The real failure is that the partition P*_m is refined at the previous barrier's zero set, while the new h's zero set is free to cross cell interiors. The barrier condition (7) is then piecewise affine with an unchecked kink inside a cell, so vertex satisfaction does not imply global satisfaction. The 1-D example above demonstrates zero-slack solutions violating (7); the same mechanism can occur in the numerical examples. The NUGIS growth idea and the computational pipeline may be salvageable by refining the partition at the new h=0 set and re-solving or rechecking, but the paper as written does not do this, so the central certificate fails. This supports the reader's REJECT verdict, although the precise reason is the zero-set alignment gap rather than the α0/αm branch choice at vertices; hence partial agreement.","tokens_in":17459,"tokens_out":20414,"duration_ms":212122,"concrete_test":"Implement the one-dimensional counterexample with D=[0,1], cells [0,0.5] and [0.5,1] (so the previous barrier's zero set is at 0.5), new h(x)=1.1x-0.1, f(x)=-0.8x+0.04545, α0=1, αm=0.5, and categories {0} in I_excl, {0.5} in I_BIS, {1} in I_int. Run optimization (35) and confirm it returns zero slacks. Then refine the returned partition by the new hyperplane h=0 and evaluate the barrier condition (7) at the newly created vertex x=0.0909. If, as the manual calculation shows, g=-0.03 < 0, then Remark 11 is falsified and the certificate requires a post-solve refinement at the new zero set before it can be trusted.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that zero slack in optimization (35) produces a certified invariant set (Remark 11). This requires barrier condition (7) for every x in D. With the Leaky ReLU K∞ function (32)-(33), g(x)=˙h(x)+α(h(x)) is affine on each subregion where h has constant sign but has a kink on the new zero set {x:h(x)=0}. The LP constrains only vertices of P*_m, and P*_m is refined at the previous barrier's zero set (23)-(24), not at the zero set of the newly solved h. After (35) is solved, the new zero set can cut through the interior of a cell of P*_m, where no constraint is checked. Thus Remark 11's certificate is not implied by zero slacks. A one-dimensional cell [0,1] split at x=0.5, with h(x)=1.1x-0.1, f(x)=-0.8x+0.04545, α0=1, αm=0.5, illustrates the gap: the vertex checks at x=0, 0.5, 1 give g=0, 0.06, 0.17, all nonnegative, yet at the new zero set x=0.0909, g=-0.03, violating (7). The same mechanism can affect the Section 7 examples whenever the new barrier's boundary shifts relative to P*_m. Since every claimed RoA estimate inherits this certificate, the load-bearing formal guarantee is missing.","agreement_with_reader":"partial"},"referee_report":null,"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a serious look, but not for the reason the authors hope. The NUGIS growth idea is genuinely new, and the IISE/SEROAISE pipeline—LP-based invariant set expansion plus a Lyapunov-like search restricted to that set—is a sensible engineering direction. The examples are nontrivial, the code is public, and the monotone growth plots are suggestive.\n\nThe soft spot is load-bearing. The certificate behind IISE does not go through as written. The LP in (35) checks the Leaky ReLU barrier condition only at vertices of a partition refined along the previous barrier function's zero set. But the new h solved at iteration m has its own zero set, which can cut through cell interiors. At those interior points the derivative condition uses the steeper alpha_0 slope when h is nonnegative and the flatter alpha_m slope when h is negative, but the LP assigns slopes by the vertex category from the previous iteration, not by the sign of the solved h. A vertex that was outside before but becomes inside after solving is still constrained with alpha_m in (35g), which is weaker than the required alpha_0. The stress-test note's 1D example is concrete and checks out: all vertex constraints are satisfied, yet the pointwise condition (7) fails at the new zero set. So Remark 11's claim that zero slack implies a certified invariant set is not established. Since every claimed RoA estimate inherits this gap, the central formal guarantee is missing.\n\nThere are smaller issues. Theorem 8's set-valued case assumes no sliding behavior without stating it in the theorem. The paper leans on the authors' own prior work for the initial invariant set and for Leaky ReLU feasibility, which is fine but should be clearly flagged. And the \"systematically larger RoA\" claim is demonstrated only by comparison to two baselines on three examples, not by a formal domination result.\n\nThe good news: the certificate gap looks fixable. A post-solve refinement that reclassifies vertices by the solved h's sign, or a formulation that enforces conditions across the new zero set, would likely close it. The NUGIS construction itself is independent and may be salvageable even if this particular LP needs patching.\n\nWho is this for? Researchers working on LP/SOS-based invariant set computation for PWA/ReLU dynamics. It deserves a serious referee: the idea is new, the writing is mostly clear, and the flaw is a precise technical gap rather than a vague hand-wave. Send it to review, but be prepared to require the authors to fix the certificate or explicitly restrict the claims.","headline":"The NUGIS expansion idea is genuinely new, but the IISE certificate is not sound as written: the LP checks the barrier condition on the wrong partition, so the claimed RoA guarantees do not follow.","tokens_in":18324,"tokens_out":1798,"would_cite":false,"duration_ms":17782,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-16T12:35:30.142786+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}