{"id":"24a32873-d606-45f8-a0e2-5ae139c3e709","arxiv_id":"1908.05949","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Γ-convex free sets, defined by closure under isometries that intertwine a tuple Γ of free polynomials, admit Effros-Winkler-style separation by monic Γ-pencils.","lead":"The paper introduces 'Γ-convexity', a new notion of convexity for sets of matrices that is tailored to partial convexity and biconvexity problems. It proves that such sets can be separated from outside points by linear matrix inequalities built from the defining polynomials Γ, a noncommutative analogue of the classical Hahn-Banach theorem.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closedness of matco(Γ(K)) in Theorem 2.4 is necessary, not automatic: a natural unbounded y2-convex set violates it and admits no Γ-pencil separation.","rationale":"The reader's weakest assumption identified the closedness of matco(Γ(K)) as the key technical condition, and the operator theorem's boundedness hypothesis as the analogous repair. Our explicit K shows the concern is real: closedness is not automatic even for a natural y2-convex free set containing 0, and when it fails the separation conclusion genuinely fails because every monic Γ-pencil that is positive on K remains positive on the offending point. This sharpens the reader's concern from a technical caveat to a necessary condition. However, the paper's stated theorems are correct as written: Theorem 2.4 includes the closedness hypothesis, Remark 2.5 acknowledges the issue, and Theorem 3.8 supplies the bounded operator-level replacement. The only mismatch is the abstract's unqualified claim that Γ-convex sets are delineated by such pencils. That is a framing issue, not a mathematical error in the main results, so the ACCEPT verdict stands unchanged.","tokens_in":20251,"tokens_out":56897,"duration_ms":560366,"concrete_test":"Verify the explicit counterexample: for Γ=(x,y,y^2) and K(n)={(X,Y): Y≽0, X≽Y, ker(Y)⊆ker(X)}, check (i) K is free and y2-convex by Proposition 4.1; (ii) (1,0)∉K; (iii) Γ(1,0)=(1,0,0) is in the closure of matco(Γ(K))(1) via (1,1/n,1/n^2) but not in matco(Γ(K))(1) because a finite convex combination with y-coordinate 0 must have x-coordinate 0; (iv) every monic y2-pencil positive on K is PSD at (1,0) by letting y→0+ along (1,y). If all four checks pass, Theorem 2.4's closedness hypothesis is genuinely necessary and the abstract's unqualified 'delineated by pencils' should be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The finite-dimensional separation theorem (Theorem 2.4, hence Theorem 1.4) is conditional on matco(Γ(K)) being closed, and this is not a harmless technicality. Take Γ=(x,y,y^2) and K(n)={(X,Y)∈S_n(C)^2: Y≽0, X≽Y, ker(Y)⊆ker(X)}. This K is free (direct sums, unitary similarities, and restrictions preserve the conditions), contains 0, has Γ(0)=0, and each slice {X: (X,Y)∈K(n)} is convex, so K is y2-convex by Proposition 4.1. At level 1, Γ(K)(1)=S∪{(0,0,0)} with S={(x,y,y^2): y>0, x≥y}. The point (1,0) is not in K, but Γ(1,0)=(1,0,0) lies in the closure of matco(Γ(K))(1) because (1,1/n,1/n^2)∈S converges to it. It is not in matco(Γ(K))(1): any finite convex combination yielding third coordinate 0 must use only points with y=0, hence x=0, so the first coordinate cannot be 1. Thus matco(Γ(K)) is not closed. Moreover no monic y2-pencil separates (1,0): if L=I+Ax+By+Cy^2 is positive on K, then L(1,y)≽0 for all small y>0, and taking y→0+ gives L(1,0)=I+A≽0. Therefore the closedness hypothesis in Theorem 2.4 is necessary and fails for a natural unbounded Γ-convex set. The operator theorem (Theorem 3.8) repairs exactly this by imposing boundedness; that boundedness is load-bearing and cannot simply be dropped.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Γ-convexity for free sets: given a tuple Γ=(γ_1,...,γ_r) of symmetric free polynomials that includes the coordinate variables, a free set K⊆S(C)^g is Γ-convex if it is closed under V^*XV whenever V is an isometry satisfying V^*Γ(X)V=Γ(V^*XV). The main finite-dimensional result, Theorem 2.4, shows that if Γ(Y) lies outside the matrix convex hull of Γ(K), then a monic Γ-pencil separates Y from K, and consequently, when matco(Γ(K)) is closed, every Y∉K is separated by a monic Γ-pencil. Section 3 extends the theory to bounded, SOT-closed operator Γ-convex sets, proving an operator-level Effros–Winkler theorem (Theorem 3.8) and representation of K as an intersection of positivity sets of monic Γ-pencils without a separate closedness hypothesis for opco(Γ(K)). Section 4 characterizes y2-convex sets as free sets that are convex in x and gives explicit monic y2-pencils for the sets TV_d.","tokens_in":20580,"tokens_out":12837,"duration_ms":119333,"significance":"If the results hold, the paper gives a substantial extension of Effros–Winkler Hahn–Banach separation to constrained noncommutative polynomial inequalities, with natural applications to partial convexity and bilinear matrix inequalities. The proofs are mostly rigorous and transparent: Theorem 2.4 reduces cleanly to the classical Effros–Winkler theorem; Theorem 3.3 uses a compactness principle for bounded operator tuples (Lemma 3.2) to obtain SOT-closedness of opco(Γ(K)); and Proposition 4.2 provides explicit, checkable monic y2-pencils. The paper is also honest about the hypotheses in the operator setting. The main caveats are a notational inconsistency in the definition of matco, a local but real proof defect in Theorem 3.5, and the fact that the finite-dimensional separation theorem is conditional on a closedness hypothesis that fails for natural unbounded Γ-convex sets; these issues affect the presentation and scope of the main claims more than the validity of the individual theorems.","major_comments":[{"comment":"The notation matco(S) is defined in §2.1 as the levelwise closed matrix convex hull, but Proposition 2.2 and Theorem 2.4 use matco(Γ(K)) as the ordinary (not necessarily closed) matrix convex hull. As written, Proposition 2.2 is false if matco denotes the closure: for the unbounded y2-convex set K(n)={(X,Y): Y≽0, X≽Y, ker(Y)⊆ker(X)} with Γ=(x,y,y^2), the point (1,0)∉K has Γ(1,0) in the closure of matco(Γ(K)) but not in the ordinary matco(Γ(K)). Please denote the ordinary hull by matco and its closure by \\overline{matco} throughout, and restate Theorem 2.4 accordingly.","section":"Section 2.1, Proposition 2.2, Theorem 2.4"},{"comment":"In the proof of Theorem 3.5, the equality M(Y)=M(Y')⊕I is not valid for a general Y, because Y need not be block diagonal with respect to the decomposition H=V(C^N)⊕(V(C^N))^\\perp. The desired conclusion still follows by compression: M(Y')=(I⊗V)^*M(Y)(I⊗V); if M(Y) were positive semidefinite, then M(Y') would be, contradicting Proposition 3.7(d). Please replace the displayed equality with this compression argument.","section":"Theorem 3.5, proof"},{"comment":"The closedness hypothesis in Theorem 2.4 is not a harmless technicality. For Γ=(x,y,y^2) and K(n)={(X,Y): Y≽0, X≽Y, ker(Y)⊆ker(X)}, K is a free y2-convex set containing 0 with Γ(0)=0; however matco(Γ(K)) is not closed, and no monic y2-pencil separates (1,0) from K, since L(1,0)=I+A would be a limit of L(1,y)≽0 as y→0+. This example should be discussed, and the abstract's unqualified claim that Γ-convex sets \"are delineated by linear pencils\" should be qualified to the cases where the stated hypotheses hold, or to the operator theorem.","section":"Theorem 1.4, Remark 2.5, abstract"}],"minor_comments":[{"comment":"In the perturbed pencil after the application of Theorem 1.3, the coefficient should be (1−t)A_j, not tA_j: from L'(z)=I+ΣA_jz_j one gets tI+(1−t)L'(z)=I+Σ(1−t)A_jz_j.","section":"Theorem 2.4, proof"},{"comment":"The sentence defining matco(S) as the closed matrix convex hull conflicts with the later use of matco for the unclosed hull; in addition to fixing the notation, please check that all occurrences in Propositions 2.2 and 2.3 and Theorem 2.4 use the intended meaning.","section":"Section 2.1"},{"comment":"Lemma 3.2 is cited from [Man+] as appearing \"to appear\"; if a published version now exists, please update the citation with full publication data.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically within scope and the main ideas are sound, but the definitional inconsistency around matco and the incorrect equality in the proof of Theorem 3.5 should be corrected before publication. The closedness limitation of the finite-dimensional separation theorem is real and worth surfacing more prominently."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper is a genuine step forward for noncommutative convexity. It introduces Γ-convexity, a family of free-set convexity notions parameterized by a tuple Γ of symmetric free polynomials, and proves Effros–Winkler type separation theorems: monic Γ-pencils separate points from closed Γ-convex sets, in the matrix case under a closedness hypothesis, and in the operator case for bounded SOT-closed sets without extra conditions. That is new. The earlier literature only had scattered partial-convexity results; nothing with this generality.\n\nWhat is good: the framework is natural, the proofs are rigorous and mostly reduce to the classical Effros–Winkler theorem, the operator closure theorem (3.3) is subtle and correct, and Section 4 gives a clean characterization of y2-convexity as slice convexity in x plus an explicit pencil representation for the TV_d sets. No data or code, but none is needed.\n\nSoft spots: the finite-dimensional Theorem 2.4 assumes matco(Γ(K)) is closed, and that is not a harmless technicality. Consider Γ=(x,y,y^2) and K={(X,Y): Y≽0, X≽Y, ker(Y)⊆ker(X)}. This is a natural unbounded y2-convex set, yet matco(Γ(K)) is not closed and no monic y2-pencil separates (1,0). So the matrix theorem simply does not reach some natural examples. The paper flags the issue in Remark 2.5, and the operator theorem fixes it by adding boundedness, so the limitation is visible — but it should be stated more boldly. Also, in the proof of Theorem 2.4, the perturbed pencil coefficient should be (1−t)A_j, not tA_j. Minor typo. The citation pattern is fine; the single external result, Lemma 3.2 from [Man+], is a general compactness lemma and not a circular dependency.\n\nBottom line: this deserves a serious referee. It is a solid contribution with real new content. After fixing the typo and, ideally, adding a remark with the counterexample to clarify the scope of Theorem 2.4, it should be published. I would bring it to the reading group and cite it.","headline":"A genuinely new framework for noncommutative partial convexity with clean separation theorems; the matrix result is conditional in an important way, but the operator result and concrete y2 constructions make it a solid paper.","tokens_in":21199,"tokens_out":4852,"would_cite":true,"duration_ms":45378,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46N10","47L07","52A30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Γ-convex sets are separated by monic Γ-pencils whenever the matrix convex hull of their Γ-image is closed, and unconditionally for bounded operator sets.","keywords":["Γ-convexity","partial convexity","biconvexity","bilinear matrix inequalities","noncommutative matrix polynomials","matrix convexity","free semialgebraic sets","Effros-Winkler theorem"],"falsifier":"In the operator setting, the decisive test is a bounded, SOT-closed, operator $\\Gamma$-convex set $K$ containing $0$ with $\\Gamma(0)=0$ and $0\\in\\operatorname{opco}(\\Gamma(K))$, together with some $Y\\notin K$; if every finite monic $\\Gamma$-pencil that is positive semidefinite on $K$ is also positive semidefinite at $Y$, then Theorem 3.8 is refuted. In the matrix setting, the analogous test is a $\\Gamma$-convex $K$ with $\\operatorname{matco}(\\Gamma(K))$ closed and a point $Y\\notin K$ that no monic $\\Gamma$-pencil of matching size separates.","tokens_in":20007,"feed_emoji":"📐","tokens_out":14699,"duration_ms":122187,"temperature":0.7,"pith_summary":"The paper establishes a Hahn–Banach separation theorem for $\\Gamma$-convex sets, a family of noncommutative sets defined by stability under compressions by isometries that intertwine a fixed tuple of symmetric polynomials $\\Gamma=(\\gamma_1,\\dots,\\gamma_r)$. The certificate is a monic $\\Gamma$-pencil, a matrix-valued expression $I+\\sum_j A_j\\gamma_j(x)$ that is positive semidefinite on the whole set but fails at the point being separated. In the finite-matrix setting the theorem requires the matrix convex hull of $\\Gamma(K)$ to be closed; in the operator setting boundedness plus strong-operator-topology closure makes that hypothesis automatic. This gives a common dual description for spectrahedra, biconvex sets, and domains defined by bilinear matrix inequalities.","feed_headline":"A Hahn–Banach theorem for Γ-convex sets","feed_subtitle":"Outside points of Γ-convex sets are cut off by matrix pencils built from Γ, unifying LMI and BMI domains.","key_machinery":"The engine is the $\\Gamma$-pair condition $V^*\\Gamma(X)V=\\Gamma(V^*XV)$, which specifies exactly which compressions a $\\Gamma$-convex set must survive and makes the $\\Gamma$-convex hull commute with the map $\\Gamma$: $X\\in \\Gamma\\text{-co}(K)$ if and only if $\\Gamma(X)\\in \\operatorname{matco}(\\Gamma(K))$. This identity transfers separation of $Y$ from $K$ to separation of $\\Gamma(Y)$ from the matrix convex hull of $\\Gamma(K)$, where the Effros–Winkler theorem produces monic linear pencils; composing those with $\\Gamma$ yields monic $\\Gamma$-pencils. In the operator setting, a compactness principle for bounded operator tuples shows that $\\operatorname{opco}(\\Gamma(K))$ is SOT-closed whenever $K$ is bounded and SOT-closed, supplying the missing closure. The final section adds a slice characterization: for $\\Gamma=\\{x,y,y^2\\}$, a free set is $\\Gamma$-convex exactly when each slice $\\{X:(X,Y)\\in K(n)\\}$ is convex in the ordinary sense, and explicit pencils show the TV sets $\\{1-x^2-y^{2d}\\succeq 0\\}$ are positivity sets of single monic $y^2$-pencils.","core_discovery":"The paper claims that every sufficiently closed $\\Gamma$-convex set containing the origin is described by monic $\\Gamma$-pencils. Theorem 2.4 states that if $K$ is $\\Gamma$-convex, $0\\in K$, $\\Gamma(0)=0$, and $\\operatorname{matco}(\\Gamma(K))$ is closed, then for each $Y\\notin K$ of size $\\ell$ there is a monic $\\Gamma$-pencil $L$ of size $\\ell$ with $L(K)\\succ 0$ and $L(Y)\\not\\succeq 0$. Theorem 3.8 removes the separate closedness hypothesis in the operator setting: for bounded, SOT-closed operator $\\Gamma$-convex sets, every outside point is separated by a finite monic $\\Gamma$-pencil, so $K$ is the intersection of the positivity sets of the monic $\\Gamma$-pencils containing it. Corollary 1.6 then yields a single monic operator $\\Gamma$-pencil $L$ with $K=\\{X:L(X)\\succeq 0\\}$ when $0$ lies in the interior of the convex hull of $\\Gamma(K)$.","pith_inferences":["A natural next step, not pursued in the paper, is to track how the minimal size of a separating $\\Gamma$-pencil grows with the size of the outlier; the matrix theorem gives size $\\ell$, while the operator theorem only promises some finite size, and the gap may encode geometric complexity of $\\Gamma$.","For bilinear matrix inequality feasibility sets, the $xy$-convex results imply that a point outside the feasibility region admits a symbolic $xy$-pencil certificate; this could turn BMI infeasibility into a finite algebraic witness in control applications, though the paper does not address computational cost.","The $y^2$ slice characterization suggests a broader dictionary: for other $\\Gamma$, $\\Gamma$-convexity may correspond to convexity along noncommutative curved directions determined by $\\Gamma$, and testing this on $\\Gamma=\\{x,y,xy+yx\\}$ would give a free analog of biconvexity with bilinear separating pencils."],"forward_implications":["In the operator setting, every bounded, SOT-closed $\\Gamma$-convex set containing $0$ with $\\Gamma(0)=0$ is exactly the intersection of the positivity sets of the monic $\\Gamma$-pencils that contain it, with no extra closedness hypothesis on the hull.","When $0$ lies in the interior of the convex hull of $\\Gamma(K)$—automatically the case when the real span of the coordinates of $\\Gamma$ contains no polynomial that is positive on $K$, e.g. for multilinear $\\Gamma$—a single monic operator $\\Gamma$-pencil alone cuts out $K$.","For any regular free polynomial $p$ whose positivity set is operator $\\Gamma$-convex, every outside matrix point is separated by a monic $\\Gamma$-pencil of the same size, so $D_p$ is an intersection of $\\Gamma$-pencil positivity sets; in particular this applies to the TV sets $1-x^2-y^{2d}$.","For $\\Gamma=\\{x,y,y^2\\}$, $\\Gamma$-convexity is equivalent to ordinary convexity of each $x$-slice at fixed $Y$, giving a free analog of partial convexity with explicit pencils for the TV screen examples."],"supporting_citations":[{"why":"Supplies the Effros–Winkler Hahn–Banach separation theorem for matrix convex sets, the base result that Theorem 2.4 composes with $\\Gamma$.","marker":"[EW97]"},{"why":"Provides the monic-linear-pencil version of the Effros–Winkler theorem cited in Theorem 1.3 and the LMI-representation context.","marker":"[HM12]"},{"why":"Another cited source for the monic-pencil Effros–Winkler theorem and for matrix convexity facts used in the proof.","marker":"[HKM17]"},{"why":"Supplies the unitary-convergence compactness lemma (Lemma 4.5 and Remark 4.6) used in Theorem 3.3 to make operator convex hulls of $\\Gamma(K)$ SOT-closed.","marker":"[Man+]"},{"why":"Earlier noncommutative partial matrix convexity results that motivate the $y^2$-convex case and are extended by the present framework.","marker":"[HHLM08]"},{"why":"Motivates $xy$-convexity and $\\Gamma$-convexity through bilinear matrix inequalities in controller design.","marker":"[KSVS04]"}],"fun_headline_variants":["Separation theorem for Γ-convex sets","Γ-convexity yields matrix pencil separation","Unifying LMI and BMI domains via Γ-convexity","Pencils cut Γ-convex sets from outside","Noncommutative Hahn-Banach via Γ-pencils"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The finite-matrix theorem rests on the assumption that the collection of all matrix compressions of the image $\\Gamma(K)$ is closed—which the paper notes is not automatic—while the operator version obtains that closure from boundedness plus strong-operator-topology closure, leaving unbounded sets outside the theorem.","fun_headline_variants_meta":{"raw":{"variants":["Separation theorem for Γ-convex sets","Γ-convexity yields matrix pencil separation","Unifying LMI and BMI domains via Γ-convexity","Pencils cut Γ-convex sets from outside","Noncommutative Hahn-Banach via Γ-pencils"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1836,"prompt_tokens":891,"completion_tokens":945,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":868}},"tokens_in":507,"tokens_out":945,"duration_ms":8366,"temperature":1.0,"reasoning_tokens":868,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:00:42.135136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the operator setting, the decisive test is a bounded, SOT-closed, operator $\\Gamma$-convex set $K$ containing $0$ with $\\Gamma(0)=0$ and $0\\in\\operatorname{opco}(\\Gamma(K))$, together with some $Y\\notin K$; if every finite monic $\\Gamma$-pencil that is positive semidefinite on $K$ is also positive semidefinite at $Y$, then Theorem 3.8 is refuted. In the matrix setting, the analogous test is a $\\Gamma$-convex $K$ with $\\operatorname{matco}(\\Gamma(K))$ closed and a point $Y\\notin K$ that no monic $\\Gamma$-pencil of matching size separates.","supporting_citations":[],"review_version":1}