{"id":"8e14b296-eff1-4efc-bb7d-4a7b40967a39","arxiv_id":"2507.00572","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The convergence rate of the moment-SOS hierarchy on a compact semi-algebraic set is O(1/r^L), where L is the Lojasiewicz exponent of the set's defining polynomials.","lead":"The paper proves that the convergence speed of the moment-SOS hierarchy for polynomial optimization over a compact domain is governed by the domain's Lojasiewicz exponent, giving explicit rates for polytopes, strongly convex sets, and spheres. This is a significant theoretical advance because it improves the best-known general rate from O(1/r^1/10) to O(1/r) and connects the hierarchy's behavior to a classical geometric invariant.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"For equality-constrained compact sets, the reduced SOS hierarchy (2.9) is not the dual of the moment SDP (2.8), and Corollary 4.8 is false: the τ_i ≥ 0 restriction creates a constant duality gap (e.g., X={0}, f=x).","rationale":"The reader identified the moment-to-dual transfer as the weakest assumption, and that is the right area. However, the load-bearing defect is sharper than 'unproven strong duality for the preordering': the paper's reduced SOS cone R(X)_2r, as defined in (2.3), is not the valid dual of the moment SDP (2.8) because the equality constraint ℓ_y(h_i^2)=0 requires a free multiplier, while (2.3) restricts it to nonnegative values. This is not a missing citation but a concrete mismatch, and it produces a counterexample to Corollary 4.8 within the stated hypotheses. The Hausdorff-distance theorems for the primal moment side (Theorems 4.3, 4.7, and the inequality-only special cases) may well be sound and are valuable; the issue is specifically the claim that the same O(1/r^L) rate holds for the reduced SOS hierarchy (2.9) on general compact semi-algebraic sets with equality constraints. Because the error is localized and repairable (allow free τ_i, or state the dual claims only for the primal hierarchy and for equality-free sets), a conditional verdict rather than outright rejection is appropriate.","tokens_in":40300,"tokens_out":35317,"duration_ms":424784,"concrete_test":"Re-derive the Lagrangian dual of (2.8) for one equality h(x)=x: the constraint ℓ(x^2)=0 has a free multiplier, so the certificate cone must contain both +x^2 and -x^2. Then solve the pair (2.8)-(2.9) on X={0}, f(x)=x, with g_0(x)=1-x^2 ≥ 0 at any level r ≥ 2, either analytically or with an SDP solver. The test lands if the primal value is 0 and the (2.9) value is -1 for all r, contradicting the claimed fmin - lb(f,R(X))_r → 0. A minimal algebraic check: x = ½(x+1)^2 - ½x^2 shows the free-multiplier dual attains 0, while the τ ≥ 0 dual cannot exceed -1.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The transfer of the Hausdorff-distance rate to the SOS side rests on a duality statement that is not merely unproven but false for the reduced hierarchy. The primal (2.8) enforces the equality ℓ_y(h_i^2)=0; the Lagrange multiplier of an equality constraint is free, so the true dual cone is generated by the g_J and ±h_i^2 (equivalently, by h_i^2 with arbitrary real coefficient). Definition (2.3) instead restricts τ_i ∈ R_{\\ge 0}, so (2.9) is a strict subset of the actual dual and strong duality between (2.8) and (2.9) fails. Concrete counterexample: take n=1, X={0}, h_1(x)=x, add g_0(x)=1-x^2 ≥ 0, and f(x)=x. For every r ≥ 1, M_1(y) ⪰ 0 together with y_2 = ℓ_y(h^2) = 0 forces y_1 = 0, so mlb(f,R(X))_r = 0. In the dual (2.9), x - c = σ_0 + τ x^2 + σ_1(1-x^2) with σ_0, σ_1 SOS and τ ≥ 0; evaluating at x = -1 gives -1-c = σ_0(-1) + τ ≥ 0, hence c ≤ -1, and c = -1 is attainable (x+1 = ½(x+1)^2 + ½(1-x^2)). Thus lb(f,R(X))_r = -1 for all r, so fmin - lb = 1 while the claimed O(1/r^L) rate would tend to 0. The defect disappears if τ_i is allowed to be any real number (then x = ½(x+1)^2 - ½x^2 gives c = 0), so the fix is local, but as written Corollary 4.8 is false for equality-constrained sets.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the convergence rate of Lasserre's moment-SOS hierarchy for polynomial optimization over a compact basic semi-algebraic set X defined by polynomial inequalities and equalities. The authors introduce a reduced Schmüdgen-type preordering R(X) in Section 2.1 and analyze the Hausdorff distance between the set of r-truncated pseudo-moment sequences satisfying the associated localizing constraints and the set Mk(X) of true k-truncated moment sequences. The main theorems (4.3, 4.7) bound this distance by O(1/r^L), where L is the Lojasiewicz exponent of X. Corollaries yield O(1/r) for polytopes and sets satisfying CQC, O(1/sqrt(r)) for sets satisfying the Polyak-Lojasiewicz or strong convexity conditions, and O(1/r^2) for products of simple sets and for the sphere. A separate result extends the CD-kernel rate O(1/r^2) to upper-bound hierarchies on products of unit balls and simplices. The paper is self-contained and contains detailed proofs of the CD-kernel estimates and of the Lojasiewicz exponent computations.","tokens_in":40731,"tokens_out":16381,"duration_ms":179532,"significance":"The geometric approach via Hausdorff distance is a genuine novelty: it separates the objective from the domain and ties the rate to a single geometric invariant, the Lojasiewicz exponent. The CD-kernel extension to products and the explicit rates for PL/strongly convex sets are useful contributions, and the appendices provide explicit constants. If the dual-side transfer were valid, the paper would substantially improve the state of the art for general compact semi-algebraic sets, going beyond the O(1/r^{1/10}) of [BMP25]. However, the paper's headline claim that the SOS hierarchy (the dual) converges at the same rate is not established and, for the reduced hierarchy as defined, is false. This is a load-bearing defect in the current version, though it is local to the sign restriction in the definition of R(X).","major_comments":[{"comment":"The dual problem (2.9) is not the dual of the moment problem (2.8). In (2.3), the multipliers τ_i are restricted to R_{\\ge0}, but the constraint ℓ_y(h_i^2)=0 in (2.8) is an equality constraint whose Lagrange multiplier must be free. Consequently (2.9) is a strict subset of the true dual and a constant duality gap can appear. For example, take n=1, X={0} defined by h_1(x)=x and g_0(x)=1-x^2\\ge0, and f(x)=x. For every r\\ge1, the primal (2.8) forces y_2=0 and M_1(y)\\succeq0, hence y_1=0 and mlb(f,R(X))_r=0. In the dual (2.9), any certificate x-c=\\sigma_0+\\tau x^2+\\sigma_1(1-x^2) with \\sigma_0,\\sigma_1 SOS and \\tau\\ge0, evaluated at x=-1, gives -1-c=\\sigma_0(-1)+\\tau\\ge0, so c\\le-1, and c=-1 is attained by x+1=\\tfrac12(x+1)^2+\\tfrac12(1-x^2). Thus lb(f,R(X))_r=-1 for all r, while the right-hand side of Corollary 4.8 tends to 0. This makes Corollary 4.8 false as stated, and it also invalidates the claims about the dual hierarchy (2.9) in Theorem 4.14(2) and Corollary 4.18. The defect is local: allowing τ_i\\in\\mathbb{R} in (2.3) restores the true dual; in the example the supremum in (2.9) then equals 0 for all r, matching mlb. Since the primal feasible set M(R(X)_{2r}) is unchanged by this modification, the Hausdorff bound in Theorem 4.7 is not affected.","section":"§2.1–2.2, Eqs. (2.3), (2.8)–(2.9)"},{"comment":"Lemma 2.4 bounds only fmin-mlb(f,\\cdot)_r, i.e., the error of the moment (primal) relaxation. The paragraph after the lemma asserts that the convergence rates of the pairs (2.4)-(2.5), (2.8)-(2.9), and (2.6)-(2.7) are the same as the rates of the Hausdorff distances. This requires strong duality between each primal and its stated dual, but strong duality is not proved for the Schmüdgen preordering pair (2.4)-(2.5) nor for the reduced pair (2.8)-(2.9). The citation [JH16] supports strong duality for the Putinar-type quadratic module under the Archimedean condition, not for the preordering or for R(X). Even for (2.4)-(2.5), no constraint qualification or Slater condition is verified, so the equality of fmin-lb and fmin-mlb remains an assumption. This gap is load-bearing because Corollary 4.8, Theorem 4.14(2), and Corollary 4.18 all state rates for the dual SOS hierarchies.","section":"§2.3, Lemma 2.4 and the paragraph following it"},{"comment":"The statements that the 'reduced moment-SOS hierarchy (2.8), (2.9)' and the Schmüdgen-type hierarchy (2.4), (2.5) both converge at O(1/r) for polytopes and under CQC inherit both problems above. For the reduced hierarchy, the counterexample in the first comment shows the statement about (2.9) is not merely unproven but false when equalities are present. Since the definition of a polytope in this paper can include equality constraints (see (1.1)), the result as stated covers that case. The authors should restrict the dual-side claims to inequality-only sets, or fix R(X) and then prove strong duality for the corrected pair, before these corollaries can stand.","section":"§4.4, Theorem 4.14 and Corollary 4.18"}],"minor_comments":[{"comment":"The statement begins with '4. If If y ∈ Mk(φ(X))' containing a duplicated 'If'.","section":"Lemma 4.6"},{"comment":"The remark writes the ball constraint as R−∥x∥^2≥0, while the rest of the paper uses R^2−∥x∥^2≥0; the two conventions are used interchangeably and should be harmonized.","section":"Remark 2.3"},{"comment":"The statement says 'for any integer r > ...' and then bounds ub(f,T(X))_{mr}; since T(X)_{2mr} is the certificate cone used in the proof, it would be clearer to denote the level by 2mr or define ub with the degree parameter explicitly.","section":"Theorem 3.9"},{"comment":"In the displayed definition of φ(X), the inequality constraints p_0,...,p_{m+1} are listed, but the equality constraints include q_j(z)=u_j-g_j(x)=0; the notation q_j is not introduced before its use in Lemma 4.5, which may confuse readers.","section":"Section 4.2, Eq. (4.9)"},{"comment":"The remark asserts that Theorem A.1 can be proved by push-forward of the CD kernel, but the actual proof provided uses a general linear transformation and does not construct the CD kernel on A(X); the remark should either be removed or reconciled with the proof.","section":"Appendix A, Remark A.2"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a valuable primal-side geometric analysis, but the dual-side claims are currently not supported. The specific counterexample in my report is small and the fix is local (allow real multipliers for h_i^2 in R(X)). I would encourage the authors to make that correction and to add a rigorous strong-duality statement, with either a proof or a precise reference that covers the preordering (or a Slater-type condition) for the pair they use. Once that is done, the paper would likely be publishable in math.OC."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea—tying moment-SOS convergence rates to the Lojasiewicz exponent—is genuinely new and the primal moment-side bounds are mostly solid. But the claimed transfer to the SOS side in the general compact case is not just unproven; it is false as stated for the reduced hierarchy. Corollary 4.8 fails on a one-point set.\n\nWhat's good: The connection between Hausdorff distance of pseudo-moment sequences and the Lojasiewicz exponent is a real organizing principle. The proofs of Theorems 3.5, 4.3, and 4.7 are detailed and plausible for the primal moment hierarchy. The O(1/r) polytope rate, O(1/sqrt(r)) for PL/strongly convex, and O(1/r^2) for non-homogeneous polynomials on the sphere are genuine improvements over the earlier O(1/r^1/10) under CQC. The CD-kernel extension to products of simple sets is competently done.\n\nThe problem: Definition (2.3) restricts the multipliers on h_i^2 to be nonnegative. The actual dual cone for the equality constraint ℓ_y(h_i^2)=0 should allow free signs. So (2.9) is not the dual of (2.8), and the asserted equality lb = mlb can have a constant gap. Concrete: X={0}, h_1=x, g_0=1-x^2≥0, f=x. For every r≥1, M_1(y)⪰0 and y_2=0 force y_1=0, so mlb=0. But in the dual, x-c = σ_0 + τ x^2 + σ_1(1-x^2) with τ≥0; evaluation at x=-1 gives c≤-1, and c=-1 is attainable. Thus fmin-lb=1 for all r, not O(1/r^L). This is a load-bearing flaw: the SOS rates in Corollary 4.8 and the corollaries for polytopes, CQC, and PL that rely on the dual side collapse unless you either prove a different duality statement or restrict the claims to the primal moment hierarchy. The fix is local—allow τ_i to be any real—but as written the theorem is false.\n\nMinor issues: the Tchakaloff point count in Theorem 3.3's proof is off (N should be the dimension of the polynomial space, not k+deg(K)), and several \"similarly\" proofs are omitted. These are fixable.\n\nWho this is for: researchers working on moment-SOS convergence theory. The geometric insight is worth engaging, but the current version overclaims. I'd send it to a referee, but with a clear instruction to check the duality claim and to require the authors to either prove it or restrict the statements.","headline":"Good idea, false dual transfer: the reduced SOS hierarchy has a constant duality gap, so Corollary 4.8 is false as stated.","tokens_in":41312,"tokens_out":3578,"would_cite":false,"duration_ms":37494,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C22","90C26","14P10","44A60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Moment-SOS hierarchy convergence rate tied to the Lojasiewicz exponent of the domain.","keywords":["moment-SOS hierarchy","polynomial optimization","Lojasiewicz exponent","truncated moment sequences","Hausdorff distance","Christoffel-Darboux kernel","sum-of-squares","semi-algebraic set"],"falsifier":"Compute the Hausdorff distance $d_k(R(X)_{2r})$ for a compact semi-algebraic set with known Lojasiewicz exponent $L$, such as $X = \\{x : g(x) \\ge 0\\}$ with $g$ having a zero of order $L$ at a boundary point, for increasing $r$; if the distance decays strictly slower than $O(1/r^L)$, the theorem fails. A second targeted check is to search for a level $r$ where the Schmüdgen-type moment SDP (2.8) and its dual (2.9) have a nonzero duality gap, since strong duality is asserted but not proved for the preordering, and one gap would sever the claimed transfer from pseudo-moment sequences to the SOS hierarchy.","tokens_in":40054,"feed_emoji":"🧮","tokens_out":6335,"duration_ms":65438,"temperature":0.7,"pith_summary":"This paper establishes that the convergence rate of the moment–sum-of-squares hierarchy for polynomial optimization over a compact basic semi-algebraic set is controlled by the Lojasiewicz exponent of the feasible set: the Hausdorff distance between $r$-level truncated pseudo-moment sequences and true truncated moment sequences is $O(1/r^L)$, giving the same rate for the hierarchy. This matters because previous $O(1/r^2)$ rates were known only for special sets such as balls, simplices, and hypercubes, and relied on explicit Christoffel–Darboux kernel formulas. The paper's geometric route works on general compact semi-algebraic sets after a lifting, and it automatically yields $O(1/r)$ for polytopes and sets satisfying the constraint qualification condition, $O(1/\\sqrt{r})$ for domains satisfying Polyak–Lojasiewicz or strong convexity, and $O(1/r^2)$ for general polynomials on a sphere. It also introduces a reduced Schmüdgen-type certificate that keeps the same theoretical rates.","feed_headline":"Moment-SOS hierarchy rate tied to a domain's Lojasiewicz exponent","feed_subtitle":"For polytopes the error is O(1/r), for strongly convex sets O(1/sqrt(r)), for spheres O(1/r^2); a geometric proof unifies them.","key_machinery":"The central object is the Hausdorff distance $d_k(R(X)_{2r})$ between the projection of the $r$-level pseudo-moment spectrahedron and the set $M_k(X)$ of $k$-truncated moment sequences of probability measures on $X$; Lemma 2.4 converts a bound on this distance into a bound on the optimal-value error of the hierarchy. The arguments are carried by three tools: the perturbed Christoffel–Darboux kernel on products of balls and simplices (Theorem 3.3), which supplies $O(1/r^2)$ approximation of nonnegative polynomials by elements of the preordering; the Lojasiewicz inequality, which bounds the distance to $X$ by a power $L$ of a defining polynomial; and a polynomial lifting $x \\mapsto (x, g(x))$ that embeds a general compact semi-algebraic set into a ball–simplex product so the simple-set estimates apply. The reduced certificate $R(X)$ requires only $h_i^2 = 0$ and the inequalities $g_J \\ge 0$, which is cheaper than the full Schmüdgen preordering yet still carries the same rate.","core_discovery":"On the paper's own terms, the central discovery is that the error of the Schmüdgen-type moment hierarchy is a geometric quantity: the distance from the spectrahedron of $r$-level truncated pseudo-moment sequences to the set of true truncated moment sequences on $X$. Theorem 4.7 bounds this Hausdorff distance by $O(1/r^L)$ whenever the Archimedean condition holds, where $L$ is the Lojasiewicz exponent of the semi-algebraic description of $X$; Corollary 4.8 transfers that bound to the lower-bound hierarchy. The proof lifts $X$ into a product of a ball and a simplex using the map $x \\mapsto (x, g_1(x), \\dots, g_m(x))$, applies a perturbed Christoffel–Darboux kernel bound on that product, then projects back and uses the Lojasiewicz inequality to control how far points in the ambient product are from $X$. The paper also proves sharper rates in special cases and shows the same rate holds for a reduced certificate $R(X)$ instead of the full Schmüdgen preordering.","pith_inferences":["Editorial: because the proof works by lifting any compact semi-algebraic set into a ball–simplex product, the same method could in principle attach rates to hierarchies for sets that are images of simple sets under polynomial maps, as long as the violation functions remain controlled by a Lojasiewicz inequality.","Editorial: the result suggests a practical way to certify convergence rates for a given instance: estimate the Lojasiewicz exponent of the defining inequalities, then predict the decay of the hierarchy; this can be tested numerically on small random instances.","Editorial: if the rate is governed by $L$, then improving a domain's description (for instance, replacing a high-degree inequality by a lower-degree equivalent with exponent $1$) could improve the hierarchy's speed, a design principle not stated in the paper."],"forward_implications":["If correct, for any compact basic semi-algebraic set with Lojasiewicz exponent $L$, both the lower-bound moment hierarchy and its SOS dual converge with error $O(1/r^L)$, matching the geometry of the feasible set.","For polytopes and sets satisfying the constraint qualification condition, the error becomes $O(1/r)$, a concrete improvement over general bounds.","For domains satisfying the Polyak–Lojasiewicz condition or defined by locally strongly convex polynomials, the error is $O(1/\\sqrt{r})$.","For polynomial optimization over a sphere, the hierarchy converges at $O(1/r^2)$ even for non-homogeneous objectives, extending earlier results that only handled homogeneous polynomials.","The reduced certificate $R(X)$ can replace the Schmüdgen preordering without changing the asymptotic rate, with potential computational savings."],"supporting_citations":[{"why":"Supplies the perturbed Christoffel–Darboux kernel bounds for the unit ball and standard simplex that the paper extends to products.","marker":"[Slo21]"},{"why":"Establishes the $O(1/r^2)$ rate on the sphere for homogeneous objectives, which the paper generalizes to non-homogeneous polynomials.","marker":"[FF21]"},{"why":"Cited as the source for strong duality between moment and SOS hierarchies under the Archimedean condition, used to transfer primal rates to the dual hierarchy.","marker":"[JH16]"},{"why":"Provides the Lojasiewicz inequality for semi-algebraic functions, the core estimate linking distance to $X$ with the defining polynomials.","marker":"[BCR98]"},{"why":"Gives the distance-to-a-polyhedron bound that yields Lojasiewicz exponent $1$ for polytopes.","marker":"[BS92]"},{"why":"Establishes that the constraint qualification condition implies Lojasiewicz exponent $1$, which the paper uses to obtain $O(1/r)$.","marker":"[BMP25]"},{"why":"Supplies the Tchakaloff-type discrete representation of truncated moment sequences used throughout the projection arguments.","marker":"[Put97]"},{"why":"Supplies the proof of Tchakaloff's theorem used to represent moment sequences as convex combinations of evaluation points.","marker":"[BT06]"}],"fun_headline_variants":["Moment-SOS rate linked to Lojasiewicz exponent","Geometric bound unifies moment-SOS convergence rates","Error O(1/r^L) for moment-SOS on any compact set","Polytopes O(1/r), spheres O(1/r^2): one proof"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the moment-side semidefinite program and its sum-of-squares dual give exactly the same value at every level; if they ever differ, the claimed rate for the sum-of-squares hierarchy does not follow from the proven moment-side bound.","fun_headline_variants_meta":{"raw":{"variants":["Moment-SOS rate linked to Lojasiewicz exponent","Geometric bound unifies moment-SOS convergence rates","Error O(1/r^L) for moment-SOS on any compact set","Polytopes O(1/r), spheres O(1/r^2): one proof"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1703,"prompt_tokens":1020,"completion_tokens":683,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":605}},"tokens_in":636,"tokens_out":683,"duration_ms":7317,"temperature":1.0,"reasoning_tokens":605,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:13:36.817549+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the Hausdorff distance $d_k(R(X)_{2r})$ for a compact semi-algebraic set with known Lojasiewicz exponent $L$, such as $X = \\{x : g(x) \\ge 0\\}$ with $g$ having a zero of order $L$ at a boundary point, for increasing $r$; if the distance decays strictly slower than $O(1/r^L)$, the theorem fails. A second targeted check is to search for a level $r$ where the Schmüdgen-type moment SDP (2.8) and its dual (2.9) have a nonzero duality gap, since strong duality is asserted but not proved for the preordering, and one gap would sever the claimed transfer from pseudo-moment sequences to the SOS hierarchy.","supporting_citations":[],"review_version":1}