betaFeedback

WorksMathematics

Comment on "The polarization hierarchy for polynomial optimization over convex bodies, with applications to nonnegative matrix rank"

Abstract

We revisit the feasibility and convergence claims for the polarization hierarchy of Plávala, Ligthart and Gross [Linear Algebra Appl.\ 723 (2025) 15–32]. Counterexamples show that the stated assumption is insufficient for the feasibility characterization and for convergence of the hierarchy with the polarization constraint. We establish corrected results under a stronger algebraic condition and supply a missing compactness argument.

Introduction

Let KAK_A and KBK_B be compact convex sets, and let f:KA×KB→Vf : K_A \times K_B \to V and p:KA×KB→Rp : K_A \times K_B \to\mathbb{R} be affine in each argument. The polarization hierarchy of [1] is a sequence of convex relaxations of the problem

F={(a,b)∈KA×KB:f(a,b)=0},p∗=min⁡(a,b)∈Fp(a,b),(1)\mathcal{F} = \{(a,b) \in K_A \times K_B : f(a,b) = 0\}, \qquad p^* = \min_{(a,b)\in\mathcal{F}} p(a,b), \tag*{(1)}

formulated over maximal tensor products of copies of KAK_A and KBK_B. If KAK_A and KBK_B are polytopes, each relaxation is a linear program. The constraint f(a,b)=0f(a,b) = 0 enters through a linear map Π:V⊗V→W\Pi: V \otimes V \to W, the polarization map, which is assumed to satisfy

Π(v⊗v+w⊗w)=0⟹v=w=0(v,w∈V).(2)\Pi(v \otimes v + w \otimes w) = 0 \quad\Longrightarrow\quad v = w = 0 \qquad(v,w \in V). \tag*{(2)}

Email address: [email protected] (Julius A. Zeiss)

Under this condition, [1 Thm. 3] characterizes F≠∅\mathcal{F} \ne\varnothing by bicompatible sequences, and [1 Thm. 5] asserts that the relaxations converge to p∗p^*.

The de Finetti theorem [1 Lem. 2] represents a bicompatible sequence by a probability measure μ\mu on KA×KBK_A \times K_B. For a sequence satisfying the polarization constraint, this representation gives

Π(M)=0,M=∫f(a,b)⊗f(a,b) dμ(a,b).\Pi(M) = 0, \qquad M = \int f(a,b) \otimes f(a,b)\,d\mu(a,b).

Here f(a,b)f(a,b) is the constraint residual, which vanishes exactly at feasible pairs. Fixing an inner product on VV, we regard MM as its second-moment operator under μ\mu. Each outer product f(a,b)⊗f(a,b)f(a,b) \otimes f(a,b) is positive semidefinite, since

⟨z,[f(a,b)⊗f(a,b)]z⟩=⟨z,f(a,b)⟩2≥0(z∈V).\left\langle z, [f(a,b) \otimes f(a,b)]z \right\rangle= \langle z, f(a,b)\rangle^2 \ge0 \qquad(z \in V).

Its average MM is therefore positive semidefinite. Its trace is the average squared constraint violation:

tr⁡M=∫∥f(a,b)∥2 dμ(a,b).\operatorname{tr} M = \int\lVert f(a,b)\rVert^2\,d\mu(a,b).

If M=0M = 0, then f(a,b)=0f(a,b) = 0 for μ\mu-almost every pair (a,b)(a,b), so the original constraint has a solution. The proof of [1 Thm. 3] therefore seeks to deduce M=0M = 0 from Π(M)=0\Pi(M) = 0.

The difficulty lies in this last implication. Since MM is positive semidefinite, its spectral decomposition gives

M=∑j=1rvj⊗vj,r=rank⁡M≤dim⁡V.M = \sum_{j=1}^{r} v_j \otimes v_j, \qquad r = \operatorname{rank} M \le\dim V.

By linearity of Π\Pi, the condition Π(M)=0\Pi(M) = 0 becomes

∑j=1rΠ(vj⊗vj)=0.\sum_{j=1}^{r} \Pi(v_j \otimes v_j) = 0.

Condition (H2_2) forces every vjv_j to vanish when r≤2r \le2, but does not exclude cancellation when r≥3r \ge3. Thus (H2_2) does not, in general, justify the implication Π(M)=0⇒M=0\Pi(M) = 0 \Rightarrow M = 0. We show that this gap is genuine and close it:

  • Exs. 3.1 and 3.2 are counterexamples to [1 Thm. 3] and to the first assertion of [1 Thm. 5] for a polarization map satisfying (H2_2).

  • Both statements hold if the kernel of Π\Pi contains no nonzero positive semidefinite operator (Prop. 4.1, Thms. 5.1 and 5.4). Equivalently, Π(∑j=1Dvj⊗vj)=0\Pi\left(\sum_{j=1}^{D} v_j \otimes v_j\right) = 0 must imply v1=⋯=vD=0v_1 = \cdots= v_D = 0, where D=dim⁡VD = \dim V. The number DD cannot be lowered in general (Prop. 4.3), but a smaller number suffices if Π\Pi has small rank (Prop. 4.4). This strengthened condition is necessary for [1 Thm. 3] to hold for all KAK_A, KBK_B and ff (Prop. 4.5).

  • The proof of [1 Thm. 5] takes limits of two-copy marginals only and does not justify that the limit extends to a bicompatible sequence (Ex. 5.3). Thm. 5.4 supplies the missing argument.

  • The three polarization maps proposed in [1] already satisfy this strengthened condition (Rem. 5.2), so convergence holds for these choices. The stronger hierarchy of [1 Thm. 4], which does not use a polarization map, also converges (Cor. 5.6). The numerical results in [1 Fig. 1] are unaffected (Rem. 5.7).

§ 6 verifies Ex. 3.2 by linear programming.

Preliminaries

All vector spaces are real and finite-dimensional. A state space is a nonempty compact convex subset KK of a vector space. We write A(K)A(K) for the space of affine functions on KK and 1K\mathbf{1}_K for the constant function 11, called the unit effect. We identify x∈Kx \in K with the evaluation functional in A(K)∗A(K)^*. For state spaces KAK_A, KBK_B let E=A(KA)∗⊗A(KB)∗E = A(K_A)^* \otimes A(K_B)^*, and let KA⊗bKB⊂EK_A \otimes_{\mathrm b} K_B \subset E be the maximal tensor product [1 Eq. (4)].

As in [1], a map on KA×KBK_A \times K_B that is affine in each argument is identified with the linear map on EE that agrees with it on the product states a⊗ba \otimes b. This applies to ff and pp. We assume F≠∅\mathcal{F} \ne\varnothing whenever p∗p^* is used; then the minimum in (1) is attained, since F\mathcal{F} is compact and pp is continuous.

We fix an arbitrary inner product on the constraint space VV and identify u⊗v∈V⊗Vu \otimes v \in V \otimes V with the operator

x⟼⟨v,x⟩ux \longmapsto\langle v, x\rangle u

Under this identification, the space Sym⁡2(V)\operatorname{Sym}^2(V) of symmetric tensors corresponds to the space of self-adjoint operators on VV, and S+(V)\mathcal{S}_+(V) denotes the cone of positive semidefinite operators. In particular, v⊗vv \otimes v is positive semidefinite for every v∈Vv \in V.

We write tr⁡\operatorname{tr} for the ordinary trace of operators on VV. Under the above identification,

tr⁡(u⊗v)=⟨v,u⟩,tr⁡(v⊗v)=∥v∥2.\operatorname{tr}(u \otimes v) = \langle v,u\rangle,\qquad\operatorname{tr}(v \otimes v) = \lVert v\rVert^{2}.

For n≥1n \ge1 let

Kn=KA⊗^n⊗^KB⊗^n.\mathcal{K}_{n} = K_{A}^{\widehat{\otimes}n} \widehat{\otimes} K_{B}^{\widehat{\otimes}n}.

with the tensor factors ordered as A1⋯AnB1⋯BnA_{1}\cdots A_{n}B_{1}\cdots B_{n}, and let SnS_{n} be the linear map that reorders them as A1B1⋯AnBnA_{1}B_{1}\cdots A_{n}B_{n} [1 Eq. (8)]. For k≤nk \le n, the marginalization map Rn,kR_{n,k}, also called the partial trace, applies the unit effect 1KA\mathbf{1}_{K_{A}} to the discarded factors Ak+1,…,AnA_{k+1},\ldots,A_{n} and 1KB\mathbf{1}_{K_{B}} to Bk+1,…,BnB_{k+1},\ldots,B_{n}; it maps Kn\mathcal{K}_{n} into Kk\mathcal{K}_{k}. For σ,τ\sigma,\tau in the symmetric group Sn\mathfrak{S}_{n}, the maps SσAS_{\sigma}^{A} and SτBS_{\tau}^{B} permute the factors A1,…,AnA_{1},\ldots,A_{n} and B1,…,BnB_{1},\ldots,B_{n}, respectively. A sequence (yn)n≥1(y_{n})_{n\ge1} with yn∈Kny_{n}\in\mathcal{K}_{n} is bicompatible [1 Def. 1] if

(SσA⊗SτB)yn=yn(σ,τ∈Sn)andRn,kyn=yk(k≤n).(3)(S_{\sigma}^{A}\otimes S_{\tau}^{B})y_{n}=y_{n}\quad(\sigma,\tau\in\mathfrak{S}_{n})\qquad\text{and}\qquad R_{n,k}y_{n}=y_{k}\quad(k\le n). \tag*{(3)}

By [1 Lem. 2], every bicompatible sequence has a Borel probability measure μ\mu on KA×KBK_{A}\times K_{B} such that

yn=∫a⊗n⊗b⊗n dμ(a,b)(n≥1).(4)y_{n}=\int a^{\otimes n}\otimes b^{\otimes n}\,\mathrm{d}\mu(a,b)\qquad(n\ge1). \tag*{(4)}

We call μ\mu a representing measure of (yn)(y_{n}).

The hierarchies. Let Π:V⊗V→W\Pi: V\otimes V\to W be linear. For n≥2n\ge2 let ΩnΠ\Omega_{n}^{\Pi} be the set of y∈Kny\in\mathcal{K}_{n} such that

(SσA⊗SτB)y=y(σ,τ∈Sn)andΠ((f⊗f)S2(Rn,2y))=0,(5)(S_{\sigma}^{A}\otimes S_{\tau}^{B})y=y\quad(\sigma,\tau\in\mathfrak{S}_{n})\qquad\text{and}\qquad\Pi\bigl((f\otimes f)S_{2}(R_{n,2}y)\bigr)=0, \tag*{(5)}

and let

pnΠ=inf⁡{p(Rn,1y):y∈ΩnΠ},p_{n}^{\Pi}=\inf\{p(R_{n,1}y):y\in\Omega_{n}^{\Pi}\},

with inf⁡∅=+∞\inf\varnothing=+\infty. This is the relaxation [1 Eq. (28)], with the order of the tensor factors and the partial traces written explicitly. The stronger relaxation [1 Eq. (29)] replaces the second constraint in (5) by

(f⊗id⁡E⊗(n−1))Sn(y)=0in V⊗E⊗(n−1).(6)(f\otimes\operatorname{id}_{E}^{\otimes(n-1)})S_{n}(y)=0\qquad\text{in }V\otimes E^{\otimes(n-1)}. \tag*{(6)}

We denote its feasible set by Ωn\Omega_n and its value by qnq_n.

If (a,b)∈F(a,b) \in\mathcal{F}, then yn=a⊗n⊗b⊗ny_n = a^{\otimes n} \otimes b^{\otimes n} satisfies (5) and (6), because

(f⊗f)S2(y2)=f(a,b)⊗f(a,b),(f \otimes f)S_2(y_2) = f(a,b) \otimes f(a,b),
(f⊗id⁡E⊗(n−1))Sn(yn)=f(a,b)⊗(a⊗b)⊗(n−1).(f \otimes\operatorname{id}_{E}^{\otimes(n-1)})S_n(y_n) = f(a,b) \otimes(a \otimes b)^{\otimes(n-1)}.

Moreover, p(Rn,1yn)=p(a,b)p(R_{n,1}y_n) = p(a,b). Hence, for every linear map Π\Pi,

F≠∅⟹pnΠ≤p∗ and qn≤p∗(n≥2).(7)\mathcal{F} \ne\varnothing\quad\Longrightarrow\quad p_n^\Pi\le p^* \text{ and } q_n \le p^* \qquad(n \ge2). \tag*{(7)}

In particular, ΩnΠ=∅\Omega_n^\Pi= \varnothing for some nn certifies F=∅\mathcal{F} = \varnothing, whatever Π\Pi is.

Thm. 3 of [1] states that if Π\Pi satisfies (H2_2), then F≠∅\mathcal{F} \ne\varnothing if and only if there is a bicompatible sequence (yn)(y_n) with Π((f⊗f)S2(y2))=0\Pi((f \otimes f)S_2(y_2)) = 0. Thm. 5 of [1] states that, for such Π\Pi, lim⁡n→∞pnΠ=p∗\lim_{n \to\infty} p_n^\Pi= p^* and lim⁡n→∞qn=p∗\lim_{n \to\infty} q_n = p^*.

Condition (H2_2) is insufficient

Let (yn)(y_n) be bicompatible with Π((f⊗f)S2(y2))=0\Pi((f \otimes f)S_2(y_2)) = 0, and let μ\mu be a representing measure. Inserting (4) gives

Π(M)=0,M=∫f(a,b)⊗f(a,b) dμ(a,b).(8)\Pi(M) = 0,\qquad M = \int f(a,b) \otimes f(a,b)\,\mathrm{d}\mu(a,b). \tag*{(8)}

The proof of [1 Thm. 3] then considers the set W+={Π(v⊗v):v∈V}W^+ = \{\Pi(v \otimes v) : v \in V\}. Condition (H2_2) implies W+∩(−W+)={0}W^+ \cap(-W^+) = \{0\}, and [1, p. 20, after Eq. (17)] concludes, citing [2 Prop. B.8], that the dual cone (W+)∗={φ∈W∗:φ≥0 on W+}(W^+)^* = \{\varphi\in W^* : \varphi\ge0 \text{ on } W^+\} is generating, that is, spans W∗W^*.

However, (W+)∗(W^+)^* is also the dual cone of the closed convex cone generated by W+W^+. Indeed, a linear functional nonnegative on W+W^+ is also nonnegative on all nonnegative linear combinations of its elements and on their limits. These combinations can involve more than two generators, whose cancellation is not excluded by (H2_2). Thus (H2_2) does not imply that the closed convex cone generated by W+W^+ is pointed. Since a closed convex cone is pointed if and only if its dual cone is generating, this pointedness is exactly what the argument requires; given (H2_2), it is equivalent to the corrected condition of § 4 (Prop. 4.1(v)).

Ex. 3.1 exhibits this obstruction with a map satisfying (H2_2) for which the convex cone generated by W+W^+ is all of WW, so that (W+)∗={0}(W^+)^* = \{0\}. Using this map, we construct an infeasible problem that nevertheless admits a bicompatible sequence satisfying the polarization constraint, contradicting [1 Thm. 3]. Ex. 3.2 adapts this construction to a feasible optimization problem, disproving the first convergence assertion in [1 Thm. 5].

For this construction and further results throughout the paper, we use the following family of projections onto the symmetric traceless matrices:

ΠD(X)=X+XT2−tr⁡XDID(D≥2).(9)\Pi_D(X) = \frac{X + X^{\mathsf T}}{2} - \frac{\operatorname{tr} X}{D} I_D \qquad(D \ge2). \tag*{(9)}

The subscript DD denotes the matrix dimension. An unsubscripted Π\Pi denotes a general polarization map.

Example 3.1 (Counterexample to [1 Thm. 3]). Let V=R3V = \mathbb{R}^3, identify V⊗VV \otimes V with R3×3\mathbb{R}^{3 \times3} via v⊗w↦vwTv \otimes w \mapsto vw^{\mathsf T}, let WW be the space of symmetric traceless 3×33 \times3 matrices, and take Π=Π3\Pi= \Pi_3 from (9). If Π3(vvT+wwT)=0\Pi_3(vv^{\mathsf T} + ww^{\mathsf T}) = 0, then vvT+wwT=cI3vv^{\mathsf T} + ww^{\mathsf T} = cI_3 with c≥0c \ge0; since the left side has rank at most 2, c=0c = 0 and v=w=0v = w = 0. Thus Π3\Pi_3 satisfies (H2)(\mathrm{H}_2). On the other hand,

∑i=13Π3(eieiT)=Π3(I3)=0,Π3(eieiT)≠0.\sum_{i=1}^{3} \Pi_3(e_i e_i^{\mathsf T}) = \Pi_3(I_3) = 0,\qquad\Pi_3(e_i e_i^{\mathsf T}) \ne0.

In fact, the convex cone generated by W+W^+ is WW: for S∈WS \in W choose t≥0t \ge0 with S+tI3S + tI_3 positive semidefinite and write S+tI3=∑iλiuiuiTS + tI_3 = \sum_i \lambda_i u_i u_i^{\mathsf T} with λi≥0\lambda_i \ge0; then

S=Π3(S+tI3)=∑iΠ3(λiuiuiT).S = \Pi_3(S + tI_3) = \sum_i \Pi_3(\lambda_i u_i u_i^{\mathsf T}).

Now let KA=conv⁡{e1,e2,e3}⊂R3K_A = \operatorname{conv}\{e_1,e_2,e_3\} \subset\mathbb{R}^3, let KB={b0}K_B = \{b_0\} be a singleton, and let f(a,b)=af(a,b) = a. Every a∈KAa \in K_A has nonnegative coordinates summing to 11, so f(a,b)=af(a,b) = a cannot vanish and F=∅F = \varnothing. Here A(KA)∗≅R3A(K_A)^* \cong\mathbb{R}^3, with 1KA(x)=∑ixi\mathbf{1}_{K_A}(x) = \sum_i x_i. Define the sequence yn=13∑i=13ei⊗n⊗b0⊗ny_n = \frac{1}{3}\sum_{i=1}^{3} e_i^{\otimes n} \otimes b_0^{\otimes n}. This sequence chooses one of the three vertices uniformly and repeats the chosen vertex in every AA-copy, with b0b_0 in every BB-copy. Independent permutations leave each term unchanged, and marginalization gives the same construction at lower order. Thus (yn)(y_n) is bicompatible, and

(f⊗f)S2(y2)=13∑i=13eieiT=13I3∈ker⁡Π3.(f \otimes f)S_2(y_2) = \frac{1}{3}\sum_{i=1}^{3} e_i e_i^{\mathsf T} = \frac{1}{3}I_3 \in\ker\Pi_3.

Hence [1 Thm. 3] fails. ⋄

We now turn the preceding construction into a feasible optimization problem for which the hierarchy fails to converge.

Example 3.2 (Counterexample to [1 Thm. 5]). We modify Ex. 3.1 by adding the origin to KAK_A and introducing an objective, while retaining the polarization map Π3\Pi_3. Let KBK_B be a singleton, which we suppress, let KA=conv⁡{0,e1,e2,e3}⊂R3K_A = \operatorname{conv}\{0,e_1,e_2,e_3\} \subset\mathbb{R}^3, and let f(x)=xf(x)=x and p(x)=−∑i=13xip(x)=-\sum_{i=1}^{3}x_i. Then F={0}\mathcal{F}=\{0\} and p∗=0p^*=0. We identify A(KA)∗A(K_A)^* with R×R3\mathbb{R} \times\mathbb{R}^3 and x∈KAx \in K_A with (1,x)(1,x); then 1KA(t,x)=t\mathbf{1}_{K_A}(t,x)=t, f(t,x)=xf(t,x)=x, p(t,x)=−∑ixip(t,x)=-\sum_i x_i, and KAK_A is the simplex with vertices σ0=(1,0)\sigma_0=(1,0) and σi=(1,ei)\sigma_i=(1,e_i), i=1,2,3i=1,2,3. The additional coordinate records normalization; in particular, the newly added origin is represented by the normalized state σ0=(1,0)\sigma_0=(1,0).

Adding the origin creates a feasible point but leaves the three original vertices available. The same uniform mixture of their tensor powers continues to satisfy the polarization constraint, while each of these vertices has objective value −1-1. The sequence yn=13∑i=13σi⊗ny_n=\frac{1}{3}\sum_{i=1}^{3}\sigma_i^{\otimes n} is bicompatible with

(f⊗f)S2(y2)=13I3∈ker⁡Π3,(f \otimes f)S_2(y_2)=\frac{1}{3}I_3 \in\ker\Pi_3,

so yn∈ΩnΠ3y_n \in\Omega_n^{\Pi_3}, and p(Rn,1yn)=−1p(R_{n,1}y_n)=-1. Since p≥−1p \ge-1 on KAK_A,

pnΠ3=−1(n≥2),lim⁡n→∞pnΠ3=−1<0=p∗.(10)p_n^{\Pi_3}=-1 \quad(n\ge2), \qquad\lim_{n\to\infty}p_n^{\Pi_3}=-1<0=p^*. \tag*{(10)}

As (5) is [1 Eq. (28)] and Π3\Pi_3 satisfies (H2_2), this contradicts the first assertion of [1 Thm. 5]. For comparison, pnid=−1/n→p∗p_n^{\mathrm{id}}=-1/n \to p^* (§ 6). The example uses only the constraints in (5); adding f(Rn,1y)=0f(R_{n,1}y)=0 or (6) defines a different relaxation.

Note that Ex. 3.1 also illustrates the insufficiency of (H2_2) for feasibility problems: the polarization hierarchy remains feasible at every level, even though the original problem is infeasible. ⋄

Remark 3.3. The choice of a singleton KBK_B simplifies Exs. 3.1 and 3.2, but both constructions extend to two nontrivial systems with coupled constraints.

In Ex. 3.1, take KB=KAK_B=K_A and define

f(a,b)=∑i,j=13aibjei+j,ek+3=ek.f(a,b)=\sum_{i,j=1}^{3}a_i b_j e_{i+j}, \qquad e_{k+3}=e_k.

Since the entries of f(a,b)f(a,b) sum to 1, the original problem is infeasible. The sequence

yn=13∑i=13ei⊗n⊗ei⊗ny_n=\frac{1}{3}\sum_{i=1}^{3}e_i^{\otimes n}\otimes e_i^{\otimes n}

is nevertheless bicompatible and satisfies the polarization constraint, because its second-moment operator is again I3/3∈ker⁡Π3I_3/3 \in\ker\Pi_3.

In Ex. 3.2, the corresponding modification is

KB=KA,f(a,b)=(a1b1,a2b2,a3b3),p(a,b)=−∑i=13aibi.K_B = K_A,\qquad f(a,b) = (a_1b_1,a_2b_2,a_3b_3),\qquad p(a,b) = -\sum_{i=1}^{3}a_i b_i.

Here p∗=0p^* = 0, while

yn=13∑i=13σi⊗n⊗σi⊗ny_n = \frac{1}{3}\sum_{i=1}^{3}\sigma_i^{\otimes n}\otimes\sigma_i^{\otimes n}

remains feasible at every hierarchy level and gives pnΠ3≤−1p_n^{\Pi_3} \le-1.

Thus the failure occurs both for local affine constraints, as in the original examples, and for constraints coupling the two systems. The local case also connects to [3], whose quantitative convergence results use a finite de Finetti theorem for constraints of the form f(a,b)=fA(a)⊕fB(b)f(a,b) = f_A(a) \oplus f_B(b). Exs. 3.1 and 3.2 show that this local structure alone does not make (H2)(\mathrm{H}_2) sufficient. ⋄

The corrected condition

Let D=dim⁡V≥1D = \dim V \ge1 (the case V={0}V = \{0\} is trivial). For k≥1k \ge1 consider the condition

Π(∑j=1kvj⊗vj)=0⟹v1=⋯=vk=0(v1,…,vk∈V).(11)\Pi\left(\sum_{j=1}^{k}v_j \otimes v_j\right) = 0 \quad\Longrightarrow\quad v_1 = \cdots= v_k = 0 \qquad(v_1,\ldots,v_k \in V). \tag*{(11)}

The corrected hypothesis is the case k=D=dim⁡Vk = D = \dim V, denoted by (HD)(\mathrm{H}_D). Since some vjv_j may vanish, (Hk)(\mathrm{H}_k) implies (Hj)(\mathrm{H}_j) for j≤kj \le k. The following proposition identifies the central algebraic condition underlying the corrected feasibility and convergence theorems and gives several equivalent formulations.

Proposition 4.1. For a linear map Π:V⊗V→W\Pi: V \otimes V \to W the following are equivalent:

(i) Π\Pi satisfies (HD)(\mathrm{H}_D);

(ii) Π\Pi satisfies (Hk)(\mathrm{H}_k) for every k≥1k \ge1;

(iii) ker⁡Π∩S+(V)={0}\ker\Pi\cap\mathcal{S}_+(V) = \{0\};

(iv) there are ℓ∈W∗\ell\in W^* and c>0c > 0 such that ℓ(Π(v⊗v))≥c∥v∥2\ell(\Pi(v \otimes v)) \ge c\lVert v\rVert^2 for all v∈Vv \in V;

(v) Π\Pi satisfies (H1)(\mathrm{H}_1), and the closed convex cone generated by W+W^+ is pointed.

Proof. (i)⇒\Rightarrow(iii): Let M∈ker⁡Π∩S+(V)M \in\ker\Pi\cap\mathcal{S}_+(V). By the spectral theorem, write M=∑j=1Dλjuj⊗ujM = \sum_{j=1}^{D} \lambda_j u_j \otimes u_j, where λj≥0\lambda_j \ge0, and set vj=λj ujv_j = \sqrt{\lambda_j}\,u_j. Then

Π(∑j=1Dvj⊗vj)=Π(M)=0,\Pi\left(\sum_{j=1}^{D} v_j \otimes v_j\right) = \Pi(M) = 0,

so (HD)(\mathrm{H}_D) implies v1=⋯=vD=0v_1 = \cdots= v_D = 0, and hence M=0M = 0.

(iii)⇒\Rightarrow(ii): Fix an arbitrary k≥1k \ge1 and vectors v1,…,vk∈Vv_1,\ldots,v_k \in V such that

Π(M)=0,M=∑j=1kvj⊗vj.\Pi(M) = 0, \qquad M = \sum_{j=1}^{k} v_j \otimes v_j.

Each outer product vj⊗vjv_j \otimes v_j is positive semidefinite, so their sum MM is positive semidefinite as well. Since MM also belongs to ker⁡Π\ker\Pi, condition (iii) gives M=0M = 0. To conclude that the individual vectors vanish, take the trace:

0=tr⁡M=∑j=1k∥vj∥2.0 = \operatorname{tr} M = \sum_{j=1}^{k} \lVert v_j\rVert^2.

This is a sum of nonnegative numbers, so every vj=0v_j = 0. Since kk was arbitrary, Π\Pi satisfies (Hk)(\mathrm{H}_k) for every k≥1k \ge1.

(ii)⇒\Rightarrow(i) is trivial.

(iii)⇒\Rightarrow(iv): The positive semidefinite operators of trace one form a compact convex set. Since Π\Pi is linear, its image

Q=Π({M∈S+(V):tr⁡M=1})Q = \Pi\left(\{M \in\mathcal{S}_+(V) : \operatorname{tr} M = 1\}\right)

is also compact and convex. Moreover, 0∉Q0 \notin Q by (iii).

Fix an inner product on WW. Compactness gives a point q0∈Qq_0 \in Q of minimal norm, and q0≠0q_0 \ne0. For every q∈Qq \in Q, convexity gives q0+t(q−q0)∈Qq_0 + t(q-q_0) \in Q for 0≤t≤10 \le t \le1. Minimality of q0q_0 therefore implies

0≤ddt∣t=0+∥q0+t(q−q0)∥2=2⟨q0,q−q0⟩.0 \le\left.\frac{d}{dt}\right|_{t=0^+} \lVert q_0 + t(q-q_0)\rVert^2 = 2\langle q_0,q-q_0\rangle.

Thus the linear functional ℓ(w)=⟨q0,w⟩\ell(w) = \langle q_0,w\rangle satisfies

ℓ(q)≥∥q0∥2=:c>0(q∈Q).\ell(q) \ge\lVert q_0\rVert^2 =: c > 0 \qquad(q \in Q).

For v≠0v \ne0, the operator (v⊗v)/∥v∥2(v \otimes v)/\lVert v\rVert^2 is positive semidefinite with trace one, so

ℓ(Π(v⊗v))≥c∥v∥2.\ell(\Pi(v \otimes v)) \ge c\lVert v\rVert^2.

For v=0v = 0 this is immediate, completing the proof.

(iv)⇒\Rightarrow(ii): Fix k≥1k \ge1 and suppose that

Π(∑j=1kvj⊗vj)=0.\Pi\left(\sum_{j=1}^{k} v_j \otimes v_j\right)=0.

By linearity of ℓ\ell and Π\Pi, condition (iv) gives

0=∑j=1kℓ(Π(vj⊗vj))≥c∑j=1k∥vj∥2.0=\sum_{j=1}^{k}\ell\left(\Pi(v_j\otimes v_j)\right)\ge c\sum_{j=1}^{k}\lVert v_j\rVert^2.

Since c>0c>0 and each squared norm is nonnegative, every vj=0v_j=0. Thus (Hk)(\mathrm{H}_k) holds for every k≥1k\ge1.

(iv)⇒\Rightarrow(v): Condition (iv) immediately implies (H1)(\mathrm{H}_1). Let

C:=cone⁡(W+)=Π(S+(V)).C:=\operatorname{cone}(W^+)=\Pi(\mathcal{S}_+(V)).

We show that ℓ(w)>0\ell(w)>0 for every nonzero w∈Cw\in C. Summing (iv) over a spectral decomposition gives

ℓ(Π(M))≥ctr⁡M(M∈S+(V)),\ell(\Pi(M))\ge c\operatorname{tr} M \qquad(M\in\mathcal{S}_+(V)),

so ℓ≥0\ell\ge0 on CC. If w∈Cw\in C satisfies ℓ(w)=0\ell(w)=0, choose Mk⪰0M_k\succeq0 with Π(Mk)→w\Pi(M_k)\to w. Then

0≤tr⁡Mk≤c−1ℓ(Π(Mk))⟶0.0\le\operatorname{tr} M_k\le c^{-1}\ell(\Pi(M_k))\longrightarrow0.

Positivity gives ∥Mk∥op≤tr⁡Mk\lVert M_k\rVert_{\mathrm{op}}\le\operatorname{tr} M_k, so Mk→0M_k\to0 and w=0w=0. Thus ℓ\ell is strictly positive on C∖{0}C\setminus\{0\}. Since ℓ(−w)=−ℓ(w)\ell(-w)=-\ell(w), the cone cannot contain both ww and −w-w for any nonzero ww. Hence CC is pointed.

(v)⇒\Rightarrow(iii): if 0≠M=∑jvj⊗vj∈ker⁡Π∩S+(V)0\ne M=\sum_j v_j\otimes v_j\in\ker\Pi\cap\mathcal{S}_+(V) with all vj≠0v_j\ne0, then w=Π(v1⊗v1)≠0w=\Pi(v_1\otimes v_1)\ne0 by (H1)(\mathrm{H}_1), while −w=Π(∑j≥2vj⊗vj)-w=\Pi(\sum_{j\ge2}v_j\otimes v_j) also lies in the cone, contradicting pointedness.

Remark 4.2 (Interpretation). The equivalent conditions in Prop. 4.1 ensure that images of nonzero positive semidefinite outer products cannot cancel to zero. Condition (iv) gives a scalar interpretation of this property: there is a single linear functional ℓ\ell whose value on the polarized output controls the size of every positive semidefinite input. Indeed,

ℓ(Π(M))≥c tr⁡Mfor every M∈S+(V).\ell(\Pi(M)) \ge c\ \operatorname{tr} M \qquad\text{for every } M \in\mathcal{S}_{+}(V).

The same functional works for all such MM, with a uniform constant c>0c > 0. For Π=id⁡\Pi= \operatorname{id}, this is simply the ordinary trace, with c=1c = 1. For the second-moment operator arising from a de Finetti representation, tr⁡M\operatorname{tr} M is the average squared constraint violation. The inequality therefore explains why the polarization constraint cannot conceal a nonzero average squared violation. Condition (v) gives a complementary geometric interpretation. A cone is pointed if it contains no nonzero vector together with its negative, as illustrated by the nonnegative quadrant in R2\mathbb{R}^{2}. Here, (H1)(\mathrm{H}_{1}) ensures that no nonzero outer product is mapped to zero, while pointedness excludes cancellation among nonzero elements of the closed convex cone generated by W+W^{+}. This includes elements obtained by positive sums and limits. The functional ℓ\ell in (iv) makes this geometry explicit: it is strictly positive on every nonzero element of this cone, so two opposite nonzero vectors cannot both belong to it. ◇

The next proposition shows that, without further assumptions on Π\Pi, the corrected condition (HD)(\mathrm{H}_{D}) cannot be replaced by (Hk)(\mathrm{H}_{k}) for any k<Dk < D.

Proposition 4.3. Let V=RDV = \mathbb{R}^{D} with D≥2D \ge2. The map ΠD\Pi_{D} defined in (9) satisfies (HD−1)(\mathrm{H}_{D-1}) but not (HD)(\mathrm{H}_{D}).

Proof. Let

M=∑j=1D−1vjvjTM = \sum_{j=1}^{D-1} v_{j}v_{j}^{\mathsf{T}}

and suppose that ΠD(M)=0\Pi_{D}(M) = 0. Since MM is symmetric, the definition of ΠD\Pi_{D} gives M=cIDM = cI_{D}, where c=tr⁡M/D≥0c = \operatorname{tr} M/D \ge0. The range of MM is contained in span⁡{v1,…,vD−1}\operatorname{span}\{v_{1},\ldots,v_{D-1}\}, so rank⁡M≤D−1\operatorname{rank} M \le D - 1. Since cIDcI_{D} has rank DD whenever c>0c > 0, we obtain c=0c = 0. Taking the trace yields

0=tr⁡M=∑j=1D−1∥vj∥2,0 = \operatorname{tr} M = \sum_{j=1}^{D-1} \lVert v_{j}\rVert^{2},

so every vj=0v_j=0. Thus ΠD\Pi_D satisfies (HD−1)(\mathrm{H}_{D-1}). On the other hand, the standard basis vectors satisfy

ΠD(∑j=1DejejT)=ΠD(ID)=0.\Pi_D\left(\sum_{j=1}^{D} e_j e_j^{\mathsf{T}}\right)=\Pi_D(I_D)=0.

Since these vectors are nonzero, (HD)(\mathrm{H}_D) fails.

Thus DD is the smallest kk such that (Hk)(\mathrm{H}_k) implies the conditions of Prop. 4.1 for every Π\Pi. Carathéodory’s theorem for cones in Sym⁡2(V)\operatorname{Sym}^2(V) gives only k=D(D+1)/2k=D(D+1)/2; the value DD is the Carathéodory number of S+(V)S_+(V) [4]. If VV is the space of real d×dd \times d matrices, as in [1 p. 21], or of Hermitian d×dd \times d matrices, then D=d2D=d^2. The vectors vjv_j in (Hk)(\mathrm{H}_k) range over VV and need not lie in the image of ff.

The next proposition shows that the original two-vector condition suffices when the restriction of Π\Pi to symmetric tensors has rank at most four, and gives a more general bound for higher ranks.

Proposition 4.4. Let m=rank⁡(Π∣Sym⁡2(V))m=\operatorname{rank}(\Pi|_{\operatorname{Sym}^2(V)}) and

κ=min⁡{D,⌊8m+9−12⌋}.(12)\kappa=\min\left\{D,\left\lfloor\frac{\sqrt{8m+9}-1}{2}\right\rfloor\right\}. \tag*{(12)}

Then (Hκ)(\mathrm{H}_\kappa) implies the conditions of Prop. 4.1. In particular, (H2)(\mathrm{H}_2) suffices if m≤4m\leq4.

Proof. We prove the contrapositive: if (iii) of Prop. 4.1 fails, then (Hκ)(\mathrm{H}_\kappa) fails. Normalizing a nonzero positive semidefinite operator in ker⁡Π\ker\Pi by its trace shows that

C={M∈S+(V):Π(M)=0, tr⁡M=1}C=\{M\in S_+(V):\Pi(M)=0,\ \operatorname{tr}M=1\}

is nonempty. This set is compact and convex, so it has an extreme point MM. We show that rank⁡M≤κ\operatorname{rank}M\leq\kappa.

Let U=ran⁡MU=\operatorname{ran}M and r=dim⁡Ur=\dim U. The self-adjoint operators HH with ran⁡H⊆U\operatorname{ran}H\subseteq U correspond to real symmetric r×rr\times r matrices and therefore form a space of dimension r(r+1)/2r(r+1)/2. On this space, Π(H)=0\Pi(H)=0 imposes at most mm independent linear constraints, and tr⁡H=0\operatorname{tr}H=0 adds at most one. Consequently, if r(r+1)/2>m+1r(r+1)/2>m+1, the rank–nullity theorem gives a nonzero such HH with Π(H)=0\Pi(H)=0 and tr⁡H=0\operatorname{tr}H=0.

Both MM and HH vanish on U⊥U^\perp, while M∣UM|_U is positive definite. Thus, for sufficiently small ε>0\varepsilon>0, the restrictions (M±εH)∣U(M\pm\varepsilon H)|_U remain positive definite, and M±εHM \pm\varepsilon H remain positive semidefinite on VV. Moreover, Π(M±εH)=0\Pi(M \pm\varepsilon H) = 0 and tr⁡(M±εH)=1\operatorname{tr}(M \pm\varepsilon H) = 1, so both perturbations belong to CC. Since H≠0H \ne0, the identity

M=12(M+εH)+12(M−εH)M = \frac{1}{2}(M+\varepsilon H) + \frac{1}{2}(M-\varepsilon H)

expresses MM as the midpoint of two distinct elements of CC, contradicting extremality. Hence

r(r+1)2≤m+1,that is,r≤8m+9−12.\frac{r(r+1)}{2} \le m+1, \qquad\text{that is,} \qquad r \le\frac{\sqrt{8m+9}-1}{2}.

This is the rank bound of [5]. Since rr is an integer and r≤Dr \le D, we obtain r≤κr \le\kappa. A spectral decomposition M=∑j=1rvj⊗vjM = \sum_{j=1}^{r} v_j \otimes v_j with vj≠0v_j \ne0 violates (Hr)(\mathrm{H}_r). Padding this decomposition with κ−r\kappa-r zero vectors gives a violation of (Hκ)(\mathrm{H}_\kappa), proving the contrapositive.

Finally, if m≤4m \le4, the bound r(r+1)/2≤5r(r+1)/2 \le5 forces r≤2r \le2, so any failure of (iii) already violates (H2)(\mathrm{H}_2).

For Π3\Pi_3 from the family (9), D=3D=3 and m=5m=5, so κ=3\kappa=3 and Prop. 4.4 is consistent with Ex. 3.1.

Whenever Π\Pi fails (HD)(\mathrm{H}_D), the following construction produces an infeasible problem with a bicompatible sequence satisfying the polarization constraint.

Proposition 4.5. If ker⁡Π∩S+(V)≠{0}\ker\Pi\cap\mathcal{S}_+(V) \ne\{0\}, there are state spaces KA,KBK_A, K_B and a map ff, affine in each argument, with F=∅\mathcal{F} = \varnothing and a bicompatible sequence (yn)(y_n) with Π((f⊗f)S2(y2))=0\Pi((f \otimes f)\mathcal{S}_2(y_2)) = 0.

Proof. Let 0≠M∈ker⁡Π∩S+(V)0 \ne M \in\ker\Pi\cap\mathcal{S}_+(V) and write M=∑i=1rλiui⊗uiM = \sum_{i=1}^{r} \lambda_i u_i \otimes u_i with λi>0\lambda_i > 0 and orthonormal uiu_i. Let KAK_A be a simplex with vertices σ1,…,σr\sigma_1,\ldots,\sigma_r, let KB={b0}K_B = \{b_0\} be a singleton, and let f:KA×KB→Vf : K_A \times K_B \to V be the affine map determined by f(σi,b0)=rλi uif(\sigma_i,b_0) = \sqrt{r\lambda_i}\,u_i. The factor r\sqrt{r} compensates for the uniform weights 1/r1/r in the sequence below, so that its second-moment operator equals MM.

Every a∈KAa \in K_A has barycentric coordinates ti≥0t_i \ge0 with ∑iti=1\sum_i t_i = 1, so

f(a,b0)=∑i=1rtif(σi,b0)=∑i=1rtirλi ui.f(a,b_0) = \sum_{i=1}^{r} t_i f(\sigma_i,b_0) = \sum_{i=1}^{r} t_i \sqrt{r\lambda_i}\,u_i.

If f(a,b0)=0f(a,b_0)=0, orthonormality forces every ti=0t_i=0, contradicting ∑iti=1\sum_i t_i=1. Thus ff has no zero on KA×KBK_A \times K_B, so F=∅\mathcal{F}=\varnothing.

Now define

yn=1r∑i=1rσi⊗n⊗b0⊗n.y_n = \frac{1}{r}\sum_{i=1}^{r}\sigma_i^{\otimes n}\otimes b_0^{\otimes n}.

Each summand is invariant under independent permutations of the AA- and BB-copies, and marginalization preserves the uniform mixture, so (yn)(y_n) is bicompatible. Moreover,

(f⊗f)S2(y2)=1r∑i=1rf(σi,b0)⊗f(σi,b0)=∑i=1rλiui⊗ui=M.\begin{aligned} (f\otimes f)S_2(y_2)=\frac{1}{r}\sum_{i=1}^{r}f(\sigma_i,b_0)\otimes f(\sigma_i,b_0) \\ &=\sum_{i=1}^{r}\lambda_i u_i\otimes u_i=M. \end{aligned}

Since M∈ker⁡ΠM\in\ker\Pi, the polarization constraint is satisfied:

Π((f⊗f)S2(y2))=Π(M)=0.\Pi\bigl((f\otimes f)S_2(y_2)\bigr)=\Pi(M)=0.

Hence condition (HD)(\mathrm{H_D}) is necessary for [1 Thm. 3] to hold for all state spaces and all maps ff. It need not be necessary for a fixed problem.

The corrected theorems

We now show that replacing (H2)(\mathrm{H_2}) by (HD)(\mathrm{H_D}) restores the characterization of feasibility in terms of bicompatible sequences satisfying the polarization constraint.

Theorem 5.1 (Correction of [1 Thm. 3]). Let Π\Pi satisfy (HD)(\mathrm{H_D}). Then F≠∅\mathcal{F}\ne\varnothing if and only if there is a bicompatible sequence (yn)(y_n) with

Π((f⊗f)S2(y2))=0.(13)\Pi\bigl((f\otimes f)S_2(y_2)\bigr)=0. \tag*{(13)}

Every representing measure μ\mu of such a sequence satisfies supp⁡μ⊆F\operatorname{supp}\mu\subseteq\mathcal{F}.

Proof. If (a,b)∈F(a,b)\in\mathcal{F}, the sequence yn=a⊗n⊗b⊗ny_n=a^{\otimes n}\otimes b^{\otimes n} is bicompatible and satisfies

(f⊗f)S2(y2)=f(a,b)⊗f(a,b)=0.(f\otimes f)S_2(y_2)=f(a,b)\otimes f(a,b)=0.

Hence the polarization constraint holds.

Conversely, let (yn)(y_n) be a bicompatible sequence satisfying (13), and let μ\mu be any representing measure, whose existence follows from the de Finetti theorem. Define

M:=(f⊗f)S2(y2)=∫f(a,b)⊗f(a,b) dμ(a,b).M:=(f\otimes f)S_2(y_2)=\int f(a,b)\otimes f(a,b)\,d\mu(a,b).

For every z∈Vz \in V,

⟨z,Mz⟩=∫⟨z,f(a,b)⟩2 dμ(a,b)≥0,\langle z, Mz \rangle= \int\langle z, f(a,b) \rangle^{2} \,d\mu(a,b) \ge0,

so MM is positive semidefinite. The polarization constraint gives Π(M)=0\Pi(M)=0, and (HD)(\mathrm{H}_{D}), through Prop. 4.1(iii), therefore implies M=0M=0. Taking the trace yields

0=tr⁡M=∫∥f(a,b)∥2 dμ(a,b).0=\operatorname{tr} M=\int\lVert f(a,b)\rVert^{2}\,d\mu(a,b).

Since the integrand is nonnegative, f=0f=0 μ\mu-almost everywhere, so μ(F)=1\mu(\mathcal{F})=1. Continuity of ff makes F\mathcal{F} closed, hence supp⁡μ⊆F\operatorname{supp}\mu\subseteq\mathcal{F}. In particular, F≠∅\mathcal{F}\ne\varnothing. As μ\mu was arbitrary, the support inclusion holds for every representing measure.

Remark 5.2. The polarization maps proposed in [1 p. 21] already satisfy the corrected hypothesis (HD)(\mathrm{H}_{D}). This can be verified using the scalar criterion in Prop. 4.1(iv), with c=1c=1: for Π=id⁡V⊗V\Pi=\operatorname{id}_{V\otimes V} take ℓ(v⊗w)=⟨v,w⟩\ell(v\otimes w)=\langle v,w\rangle; for V=Rd×dV=\mathbb{R}^{d\times d} with the Frobenius inner product and the maps

ΠHS(A⊗B)=tr⁡(ATB)∈R,\Pi_{\mathrm{HS}}(A\otimes B)=\operatorname{tr}(A^{\mathsf{T}}B)\in\mathbb{R},
ΠM(A⊗B)=ATB∈Rd×d,\Pi_{\mathrm{M}}(A\otimes B)=A^{\mathsf{T}}B\in\mathbb{R}^{d\times d},

take ℓ=id⁡R\ell=\operatorname{id}_{\mathbb{R}} and ℓ=tr⁡\ell=\operatorname{tr}, respectively. In each case ℓ(Π(v⊗v))=∥v∥2\ell(\Pi(v\otimes v))=\lVert v\rVert^{2}. The codomain of ΠM\Pi_{\mathrm{M}} is Rd×d\mathbb{R}^{d\times d}, since ATBA^{\mathsf{T}}B need not be symmetric.

These functionals also give a direct replacement for the dual-cone argument in [1]. Indeed, with MM and μ\mu as in the proof of Thm. 5.1,

0=ℓ(Π(M))=∫ℓ(Π(f(a,b)⊗f(a,b))) dμ(a,b)≥c∫∥f(a,b)∥2 dμ(a,b),0=\ell(\Pi(M))=\int\ell\bigl(\Pi(f(a,b)\otimes f(a,b))\bigr)\,d\mu(a,b)\ge c\int\lVert f(a,b)\rVert^{2}\,d\mu(a,b),

so f=0f=0 μ\mu-almost everywhere.

Thm. 5.1 assumes a full infinite bicompatible sequence. To prove convergence of the hierarchy, we must first construct such a sequence from minimizers at finite levels. This requires a compactness argument in addition to the corrected polarization condition (HD)(\mathrm{H}_{D}).

The proof of [1 Thm. 5], pp. 24–25, takes a limit of the two-copy marginals only.1 The extension property used there guarantees symmetry under permutations of the pairs AiBiA_iB_i. Bicompatibility, however, requires independent permutations of the AA- and BB-copies at every order, together with compatibility under marginalization. Ex. 5.3 shows that pair-symmetric extensions and independent symmetry at order two do not suffice.

Thm. 5.4 closes this gap by extracting limits of all fixed-order marginals along a common subsequence. These limits form a bicompatible sequence, to which de Finetti applies. Although the finite-level operators need not be positive semidefinite, the resulting second-moment operator is. Condition (HD)(\mathrm{H}_D) then forces it to vanish, so the representing measure is supported on feasible pairs. This proves asymptotic convergence of the optimal values without requiring any finite relaxation to be exact.

Example 5.3. Let KA=KBK_A = K_B be the one-dimensional probability simplex on {0,1}\{0,1\}, the state space of a classical bit. Then KA⊗^KBK_A \widehat{\otimes} K_B is the three-dimensional probability simplex on {0,1}2\{0,1\}^2. Let ρ+=12(δ00+δ11)\rho_+ = \frac{1}{2}(\delta_{00}+\delta_{11}) and ρ−=12(δ01+δ10)\rho_- = \frac{1}{2}(\delta_{01}+\delta_{10}), where δab\delta_{ab} is the deterministic distribution assigning probability 11 to (a,b)(a,b). Under ρ+\rho_+, the two bits are uniformly distributed and always agree; under ρ−\rho_-, they are uniformly distributed and always disagree. Define the two-copy state

x2=12ρ+⊗2+12ρ−⊗2.(14)x_2 = \frac{1}{2}\rho_+^{\otimes2}+\frac{1}{2}\rho_-^{\otimes2}. \tag*{(14)}

A bit string has even or odd parity according to whether its number of ones is even or odd. The mixture chooses with equal probability whether both pairs agree or both pairs disagree. In either case, the total number of ones is even, and each of the eight strings with this property has probability 1/81/8. Thus S2−1x2S_2^{-1}x_2 is uniform on the even-parity strings (a1,a2,b1,b2)∈{0,1}4(a_1,a_2,b_1,b_2)\in\{0,1\}^4.2 Swapping the AA-copies or the BB-copies preserves parity, so this distribution is invariant under these independent permutations. Moreover, x2∈Ext⁡∞(KA⊗^KB,2)x_2 \in\operatorname{Ext}_{\infty}(K_A \widehat{\otimes} K_B,2), being the two-copy marginal of 12ρ+⊗n+12ρ−⊗n\frac{1}{2}\rho_+^{\otimes n}+\frac{1}{2}\rho_-^{\otimes n} for every nn.

We now show that, despite these properties, x2x_2 cannot be the two-copy marginal of a full bicompatible sequence. Suppose that S2(y2)=x2S_2(y_2)=x_2 for a bicompatible sequence (yn)(y_n) with representing measure μ\mu, and write a(0)a(0) for the probability of the outcome 00 under a∈KAa\in K_A. The A1A2A_1A_2-marginal of y2y_2 is uniform because each choice of the two AA-bits has exactly two even-parity completions by the BB-bits. Hence ∫a(0) dμ=12\int a(0)\,d\mu=\frac{1}{2} and ∫a(0)2 dμ=14\int a(0)^2\,d\mu=\frac{1}{4}. Consequently,

∫(a(0)−12)2dμ=∫a(0)2dμ−∫a(0)dμ+14=14−12+14=0.\int\left(a(0)-\frac{1}{2}\right)^2d\mu=\int a(0)^2d\mu-\int a(0)d\mu+\frac{1}{4}=\frac{1}{4}-\frac{1}{2}+\frac{1}{4}=0.

This vanishing variance forces the local probability a(0)a(0) to equal 1/21/2 almost everywhere; similarly, b(0)=1/2b(0)=1/2 almost everywhere. Hence the representing measure is concentrated on the single pair of uniform local distributions. The de Finetti representation would therefore make all four bits independent and uniform, assigning probability 1/161/16 to every string. This contradicts the absence of odd-parity strings in x2x_2. § 6 gives a finite certificate. ⋄

Ex. 5.3 shows that the extension properties used in [1] do not suffice to establish bicompatibility. The proof of Thm. 5.4 supplies this missing step by extracting limits of all fixed-order marginals along a common subsequence and verifying that these limits form a bicompatible sequence. The following theorem shows that, under (HD)(H_D), the optimal values of the hierarchy with the polarization constraint converge monotonically to the true optimum whenever the original problem is feasible.

Theorem 5.4 (Correction of the first assertion of [1 Thm. 5]). Let F≠∅\mathcal{F}\ne\varnothing and let Π\Pi satisfy (HD)(H_D). Then the infima pnΠp_n^\Pi are attained, and

pnΠ≤pn+1Π≤p∗(n≥2),lim⁡n→∞pnΠ=p∗.p_n^\Pi\le p_{n+1}^\Pi\le p^* \quad(n\ge2), \qquad\lim_{n\to\infty}p_n^\Pi=p^*.

Proof. The set Kn\mathcal{K}_n is compact, the constraints in (5) are closed, and the objective is linear. By (7), ΩnΠ≠∅\Omega_n^\Pi\ne\varnothing and pnΠ≤p∗p_n^\Pi\le p^*; hence the infimum is attained.

The partial trace Rn+1,nR_{n+1,n} preserves invariance under independent permutations and satisfies Rn,kRn+1,n=Rn+1,kR_{n,k}R_{n+1,n}=R_{n+1,k}; hence it maps Ωn+1Π\Omega_{n+1}^\Pi into ΩnΠ\Omega_n^\Pi and preserves the objective, so pnΠ≤pn+1Πp_n^\Pi\le p_{n+1}^\Pi. Therefore pnΠp_n^\Pi converges to some L≤p∗L\le p^*. It remains to prove L≥p∗L\ge p^*, which we do by constructing a limiting representing measure supported on F\mathcal{F}.

Choose minimizers y(n)∈ΩnΠy^{(n)}\in\Omega_n^\Pi and define the finite-level operators

Mn=(f⊗f)S2(Rn,2y(n)),Π(Mn)=0.M_n=(f\otimes f)S_2(R_{n,2}y^{(n)}), \qquad\Pi(M_n)=0.

For each kk, the marginals Rn,ky(n)R_{n,k}y^{(n)}, n≥kn\ge k, lie in the compact set Kk\mathcal{K}_k. First choose a subsequence along which the one-copy marginals converge, then a further subsequence along which the two-copy marginals converge, and continue. Taking the jj-th term from the jj-th subsequence gives a single sequence of levels nj→∞n_j\to\infty along which every fixed-order marginal converges:

Rnj,ky(nj)⟶zk∈Kk(k≥1).(15)R_{n_j,k}y^{(n_j)}\longrightarrow z_k\in\mathcal{K}_k \qquad(k\ge1). \tag*{(15)}

Using the same sequence of levels for every marginal order allows compatibility to pass to the limit. Each zkz_k is invariant under independent permutations, as a limit of invariant elements, and for k′<kk' < k,

Rk,k′zk=lim⁡j→∞Rk,k′Rnj,ky(nj)=lim⁡j→∞Rnj,k′y(nj)=zk′.R_{k,k'}z_k = \lim_{j\to\infty} R_{k,k'}R_{n_j,k}y^{(n_j)} = \lim_{j\to\infty} R_{n_j,k'}y^{(n_j)} = z_{k'}.

Thus (zk)(z_k) is bicompatible. Let μ\mu be a representing measure. By continuity and the de Finetti representation,

Mnj⟶M∞:=(f⊗f)S2(z2),M_{n_j} \longrightarrow M_\infty:= (f \otimes f)S_2(z_2),
M∞=∫f(a,b)⊗f(a,b) dμ(a,b)∈S+(V).M_\infty= \int f(a,b) \otimes f(a,b)\,d\mu(a,b) \in\mathbb{S}_+(V).

Consequently, Π(M∞)=lim⁡jΠ(Mnj)=0\Pi(M_\infty) = \lim_j \Pi(M_{n_j}) = 0, and Prop. 4.1 gives M∞=0M_\infty= 0. As in the proof of Thm. 5.1, it follows that f=0f = 0 μ\mu-almost everywhere and supp⁡μ⊆F\operatorname{supp}\mu\subseteq\mathcal{F}. Since μ\mu is supported on F\mathcal{F}, we have p(a,b)≥p∗p(a,b) \ge p^* μ\mu-almost everywhere. By continuity of the objective,

L=lim⁡jpnjΠ=p(z1)=∫p(a,b) dμ(a,b)≥p∗.L = \lim_j p^\Pi_{n_j} = p(z_1) = \int p(a,b)\,d\mu(a,b) \ge p^*.

Together with L≤p∗L \le p^*, this proves L=p∗L = p^*.

Remark 5.5. (a) Thm. 5.4 asserts convergence of the optimal values; it does not assert convergence of minimizers or equality pnΠ=p∗p^\Pi_n = p^* at a finite level. For Π=id⁡\Pi= \operatorname{id} in Ex. 3.2, (17) gives

pnid⁡=−1n<0=p∗(n≥2),p^{\operatorname{id}}_n = -\frac{1}{n} < 0 = p^* \qquad(n \ge2),

so the values converge to p∗p^* without attaining it at any finite level.

(b) Let Π\Pi satisfy (HD)(\mathrm{H}_{\mathrm{D}}) and F=∅\mathcal{F} = \varnothing. Then ΩnΠ=∅\Omega^\Pi_n = \varnothing for some nn, hence for all larger nn since Rn+1,n(Ωn+1Π)⊆ΩnΠR_{n+1,n}(\Omega^\Pi_{n+1}) \subseteq\Omega^\Pi_n: otherwise the diagonal argument, applied to arbitrary y(n)∈ΩnΠy^{(n)} \in\Omega^\Pi_n, yields a bicompatible sequence satisfying (13), contradicting Thm. 5.1. The argument gives no bound on nn.

The stronger hierarchy of [1 Eq. (29)] does not use an auxiliary polarization map and requires no modification. Its convergence follows from Thm. 5.4, as the following corollary shows.

Corollary 5.6 (Second assertion of [1 Thm. 5]). Let F≠∅\mathcal{F} \ne\varnothing. Then the infima qnq_n are attained, and

pnid≤qn≤qn+1≤p∗(n≥2),lim⁡n→∞qn=p∗.p_n^{\mathrm{id}} \le q_n \le q_{n+1} \le p^* \qquad(n \ge2), \qquad\lim_{n \to\infty} q_n = p^*.

If F=∅\mathcal{F} = \varnothing, then Ωn=∅\Omega_n = \varnothing for all sufficiently large nn.

Proof. The stronger constraint applies ff to the first ABAB-pair while retaining all remaining pairs. Applying ff to the second pair and tracing out the others gives the second-moment constraint with Π=id⁡\Pi= \operatorname{id}. Explicitly, applying

id⁡V⊗f⊗(1KA⊗1KB)⊗(n−2)\operatorname{id}_V \otimes f \otimes(\mathbf{1}_{K_A} \otimes\mathbf{1}_{K_B})^{\otimes(n-2)}

to (6) gives (f⊗f)S2(Rn,2y)=0(f \otimes f)S_2(R_{n,2y}) = 0, as in [1 Eq. (22)]; hence Ωn⊆Ωnid\Omega_n \subseteq\Omega_n^{\mathrm{id}} and pnid≤qnp_n^{\mathrm{id}} \le q_n.

Applying

id⁡V⊗id⁡E⊗(n−1)⊗(1KA⊗1KB)\operatorname{id}_V \otimes\operatorname{id}_E^{\otimes(n-1)} \otimes(\mathbf{1}_{K_A} \otimes\mathbf{1}_{K_B})

to (6) at level n+1n+1 shows Rn+1,n(Ωn+1)⊆ΩnR_{n+1,n}(\Omega_{n+1}) \subseteq\Omega_n. Since marginalization also preserves the objective, it follows that qn≤qn+1q_n \le q_{n+1}. Attainment and qn≤p∗q_n \le p^* follow as in the proof of Thm. 5.4.

Since id⁡\operatorname{id} satisfies (HD)(\mathrm{H}_D) (Rem. 5.2), Thm. 5.4 gives pnid→p∗p_n^{\mathrm{id}} \to p^*. The bounds pnid≤qn≤p∗p_n^{\mathrm{id}} \le q_n \le p^* therefore imply qn→p∗q_n \to p^*. Rem. 5.5(b), together with Ωn⊆Ωnid\Omega_n \subseteq\Omega_n^{\mathrm{id}}, gives the last assertion.

Remark 5.7 (Application to nonnegative matrix rank). The insufficiency of (H2)(\mathrm{H}_2) does not invalidate the finite-level infeasibility conclusions in [1 Fig. 1]. The blue and red regions record infeasibility at n=3n = 3 for the stronger hierarchy (6) and the hierarchy with the polarization constraint (5), respectively.3 If the required nonnegative factorization exists, every relaxation is feasible; see (7). Hence infeasibility at any level rules out such a factorization. This conclusion does not require asymptotic convergence.

These certificates alone do not establish that every infeasible instance will eventually be detected. For the stronger hierarchy, this guarantee follows from Cor. 5.6; for the hierarchy with the polarization constraint, it follows under (HD)(\mathrm{H}_D) from Rem. 5.5(b). ⋄\diamond

Remark 5.8 (Several state spaces and higher-degree polynomials). The counterexamples and corrections above also apply to the extensions in Sec. 2.1 of [1], which allow several state spaces K1,…,KmK_1,\ldots,K_m and higher-degree polynomial objectives and constraints [1 Eqs. (25), (26)]. Convergence is stated to follow by the same reasoning [1 p. 23]. For two state spaces and ℓ⃗=(1,1)\vec{\ell}=(1,1) in [1 Eq. (23)], [1 Eq. (25)] is [1 Eq. (28)], so Ex. 3.2 applies.

Choose ℓ≥1\ell\geq1 such that ff and pp have degree at most ℓ\ell in each local state. These polynomials can be represented by linear maps acting on ℓ\ell copies of each state space, using unit functionals to accommodate lower-degree terms. Consequently, for n≥2ℓn\geq2\ell, the objective uses ℓ\ell-copy marginals, while the squared-polarization constraint uses 2ℓ2\ell-copy marginals. The stronger hierarchy retains its full nn-copy constraint.

The compactness argument again extracts limits of all fixed-order marginals along a common subsequence. Compatibility and invariance under independent permutations of the copies of each state space pass to these limits, so the multipartite de Finetti representation gives a measure μ\mu on K1×⋯×KmK_1\times\cdots\times K_m. On each product component of this representation, the multilinearized constraint map evaluates to the original polynomial ff. Thus, writing a=(a1,…,am)\mathbf{a}=(a_1,\ldots,a_m), the limiting polarization constraint is again

Π(∫f(a)⊗f(a) dμ(a))=0,\Pi\left(\int f(\mathbf{a})\otimes f(\mathbf{a})\,d\mu(\mathbf{a})\right)=0,

and the limiting objective is ∫p(a) dμ(a)\int p(\mathbf{a})\,d\mu(\mathbf{a}). The operator inside Π\Pi is positive semidefinite. Condition (HD)(H_D) forces it to vanish, so the representing measure is supported on solutions of the original polynomial constraint. Hence the feasibility and convergence arguments of this section carry over. ◇

Linear-programming verification

We first derive exact formulas for the hierarchy values in Ex. 3.2 and compare them with numerical solutions of the corresponding linear programs. We then give a finite certificate for the failure of bicompatible extendibility in Ex. 5.3 and quantify the distance to the set of extendible marginals.

Exact hierarchy values

In Ex. 3.2, KAK_A is a simplex with vertices σ0,…,σ3\sigma_0,\ldots,\sigma_3, so KnK_n is the set of probability distributions on {0,1,2,3}n\{0,1,2,3\}^n, where the label ii stands for σi\sigma_i. A permutation orbit is specified by its occupation vector k=(k0,k1,k2,k3)\mathbf{k}=(k_0,k_1,k_2,k_3), where kik_i counts the positions carrying label ii and ∑iki=n\sum_i k_i=n. For example, the strings (0,1,1)(0,1,1), (1,0,1)(1,0,1), and (1,1,0)(1,1,0) have the same occupation vector k=(1,2,0,0)k=(1,2,0,0). Permutation invariance assigns them equal probabilities, so we need only record their total probability αk\alpha_k. In general, a distribution invariant under Sn\mathfrak{S}_n is determined by these orbit masses αk≥0\alpha_k\geq0, ∑kαk=1\sum_k\alpha_k=1; there are (n+33)\binom{n+3}{3} orbits.

Let s=k1+k2+k3s=k_1+k_2+k_3 count the positions carrying a nonzero label, and let Eα\mathbb{E}_{\alpha} denote expectation with respect to α\alpha. The objective and the matrix Mn=(f⊗f)S2(Rn,2y)M_n=(f\otimes f)S_2(R_{n,2y}) are

p(Rn,1y)=−Eα[s]n,(Mn)ij=Eα[ki(kj−δij)]n(n−1)(1≤i,j≤3).(16)p(R_{n,1y})=-\frac{\mathbb{E}_{\alpha}[s]}{n},\qquad(M_n)_{ij}=\frac{\mathbb{E}_{\alpha}[k_i(k_j-\delta_{ij})]}{n(n-1)}\quad(1\leq i,j\leq3). \tag*{(16)}

For a fixed occupation vector kk, a single position has a nonzero label with probability s/ns/n. The two-copy marginal samples two distinct positions without replacement. There are n(n−1)n(n-1) ordered choices; the number with labels i,ji,j is kikjk_i k_j if i≠ji\ne j, and ki(ki−1)k_i(k_i-1) if i=ji=j. Averaging these expressions over α\alpha gives (16). Thus pnΠp_n^{\Pi} is the value of the linear program that minimizes −Eα[s]/n-\mathbb{E}_{\alpha}[s]/n subject to Π(Mn)=0\Pi(M_n)=0.

At a finite level, MnM_n records correlations between two distinct copies and need not be positive semidefinite. The positivity of the second moment in (8) follows from the de Finetti representation of a full bicompatible sequence.

For Π=Π3\Pi=\Pi_3, the matrix MnM_n is symmetric, so Π3(Mn)=0\Pi_3(M_n)=0 means Mn=13(tr⁡Mn)I3M_n=\frac{1}{3}(\operatorname{tr}M_n)I_3. The off-diagonal constraints Eα[kikj]=0\mathbb{E}_{\alpha}[k_i k_j]=0, i≠ji\ne j, allow at most one distinct nonzero label per orbit in the support of α\alpha. This label may nevertheless occur in all nn positions. The value −1-1 in (10) requires s=ns=n on the support, and the diagonal constraints then force mass 13\frac{1}{3} on each of the orbits k=(0,n,0,0),(0,0,n,0),(0,0,0,n)k=(0,n,0,0),(0,0,n,0),(0,0,0,n). Hence the minimizer is unique; the gap in (10) is not due to multiple minimizers.

For Π=id⁡\Pi=\operatorname{id}, Mn=0M_n=0 gives

∑i,j(Mn)ij=Eα[s(s−1)]n(n−1)=0,\sum_{i,j}(M_n)_{ij}=\frac{\mathbb{E}_{\alpha}[s(s-1)]}{n(n-1)}=0,

so s≤1s\leq1 on the support: at most one position may carry a nonzero label. Its contribution to the one-copy marginal is therefore at most 1/n1/n, giving p(Rn,1y)≥−1/np(R_{n,1y})\geq-1/n, with equality for the orbit k=(n−1,1,0,0)k=(n-1,1,0,0). Hence

pnid⁡=−1n⟶0=p∗.(17)p_n^{\operatorname{id}}=-\frac{1}{n}\longrightarrow0=p^*. \tag*{(17)}

The same lower bound has a dual certificate. After averaging, the following identity expresses the objective minus −1/n-1/n as a nonnegative quantity plus terms that vanish under Mn=0M_n=0:

−sn+1n+(n−1)∑i,j=13ki(kj−δij)n(n−1)=(s−1)2n≥0;-\frac{s}{n}+\frac{1}{n}+(n-1)\sum_{i,j=1}^{3}\frac{k_i(k_j-\delta_{ij})}{n(n-1)}=\frac{(s-1)^2}{n}\geq0;

averaging over α\alpha and imposing Mn=0M_n=0 gives p(Rn,1y)≥−1/np(R_{n,1}y)\geq-1/n. The minimizers are the mixtures of the three orbits with s=1s=1.

Table 1: Values of the linear programs of § 6 for Ex. 3.2, rounded to five decimal places; pn∗=0p_n^*=0. The exact values for all n≥2n\geq2 are given by {(3.3)} and {(6.2)}.

nnOrbitspnΠ3p_n^{\Pi_3}pnidp_n^{\mathrm{id}}
210−1.00000-1.00000−0.50000-0.50000
435−1.00000-1.00000−0.25000-0.25000
8165−1.00000-1.00000−0.12500-0.12500
16969−1.00000-1.00000−0.06250-0.06250
326545−1.00000-1.00000−0.03125-0.03125

Table 1.

A finite extension certificate

We now show that the obstruction in Ex. 5.3 already appears when extending to three copies of each party. Suppose that S2−1x2S_2^{-1}x_2 is the marginal on A1A2B1B2A_1A_2B_1B_2 of a distribution on three A bits and three B bits that is invariant under independent permutations. Let QklQ_{kl} be the probability that kk of the three A bits and ll of the three B bits equal 1. If three bits contain kk ones, the probability of finding rr ones when choosing two positions uniformly is

ηr(k)=(kr)(3−k2−r)(32).\eta_r(k)=\frac{\binom{k}{r}\binom{3-k}{2-r}}{\binom{3}{2}}.

The denominator is the number (32)=3\binom{3}{2}=3 of choices of two positions; binomial coefficients outside their usual range are taken to be zero. By invariance under independent permutations, the probability that rr of the first two A bits and tt of the first two B bits equal 1 is

Trt=∑k,l=03Qklηr(k)ηt(l).(18)T_{rt}=\sum_{k,l=0}^{3}Q_{kl}\eta_r(k)\eta_t(l). \tag*{(18)}

Fix (k,l)(k,l) and put h=η1(k)h=\eta_1(k), g=η1(l)g=\eta_1(l). Since η0+η2=1−η1\eta_0+\eta_2=1-\eta_1, the contribution of (k,l)(k,l) to T01+T10+T12+T21−T11T_{01}+T_{10}+T_{12}+T_{21}-T_{11} is QklQ_{kl} times

g(1−h)+h(1−g)−hg=h+g−3hg,g(1-h)+h(1-g)-hg=h+g-3hg,

which is nonnegative because h,g∈{0,23}h,g\in\{0,\frac{2}{3}\}. Hence

T11≤T01+T10+T12+T21.(19)T_{11}\le T_{01}+T_{10}+T_{12}+T_{21}. \tag*{(19)}

The right-hand side is the probability of odd total parity among the four bits. Thus any such extension requires this probability to be at least the probability that each local pair contains exactly one 1. For x2x_2, T11=12T_{11}=\frac{1}{2}, T00=T02=T20=T22=18T_{00}=T_{02}=T_{20}=T_{22}=\frac{1}{8} and all other TrtT_{rt} vanish, which violates (6.3). Thus the corresponding feasibility linear program in the 16 variables QklQ_{kl} is infeasible.

The obstruction persists under small perturbations: every two-copy marginal admitting such a three-copy extension is at total variation distance at least 14\frac{1}{4} from x2x_2. We first derive this lower bound from (6.3), then give an extension attaining it. Both x2x_2 and every candidate marginal are uniform within each orbit of independent permutations, so their total variation distance equals that between their distributions of occupation counts. Denote these distributions by T0T^0 and TT, respectively, and use the convention

dTV(T0,T)=12∑r,t=02∣Trt0−Trt∣.d_{\mathrm{TV}}(T^0,T)=\frac{1}{2}\sum_{r,t=0}^{2}\left|T^0_{rt}-T_{rt}\right|.

Let φ\varphi take the value 1 at (1,1)(1,1), the value −1-1 at (0,1),(1,0),(1,2),(2,1)(0,1),(1,0),(1,2),(2,1), and zero elsewhere. Then ET0φ=12\mathbb{E}_{T^0}\varphi=\frac{1}{2}, whereas (6.3) gives ETφ≤0\mathbb{E}_{T}\varphi\le0. Since ∣φ∣≤1|\varphi|\le1, the distance is at least 14\frac{1}{4}:

12≤∣ET0φ−ETφ∣≤2dTV(T0,T).\frac{1}{2}\le\left|\mathbb{E}_{T^0}\varphi-\mathbb{E}_{T}\varphi\right|\le2d_{\mathrm{TV}}(T^0,T).

To attain the lower bound, take

Qkl={964,(k,l)∈{1,2}2,764,(k,l)∈{0,3}2,0,otherwise.Q_{kl}= \begin{cases} \frac{9}{64}, & (k,l)\in\{1,2\}^2,\\ \frac{7}{64}, & (k,l)\in\{0,3\}^2,\\ 0, & \text{otherwise}. \end{cases}

The resulting distribution T⋆T^\star retains the four corner probabilities 18\frac{1}{8}, has T11⋆=14T^\star_{11}=\frac{1}{4}, and assigns probability 116\frac{1}{16} to each of the four remaining cells. Hence

dTV(T0,T⋆)=12(14+4⋅116)=14.d_{\mathrm{TV}}(T^0,T^\star)=\frac{1}{2}\left(\frac{1}{4}+4\cdot\frac{1}{16}\right)=\frac{1}{4}.

Acknowledgements

The author thanks Martin Plávala and David Gross for discussions and for supporting this manuscript. The author acknowledges support from the European Research Council (ERC Grant Agreement No. 948139) and the Excellence Cluster – Matter and Light for Quantum Computing (ML4Q-2).

Data availability

The code, the archived output of the linear programs, and a script verifying the certificates of § 6 in exact rational arithmetic are provided as supplementary material.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

Generative AI tools supported manuscript preparation, mathematical analysis, and code development. Any content not developed by the author was reviewed by the author, who takes full responsibility for the published work.

References

  1. [1]M. Plávala, L.T. Lightart, D. Gross, The polarization hierarchy for polynomial optimization over convex bodies, with applications to nonnegative matrix rank, Linear Algebra Appl. 723 (2025) 15–32, https://doi.org/10.1016/j.laa.2025.05.019.
  2. [2]M. Plávala, General probabilistic theories: an introduction, Phys. Rep. 1033 (2023) 1–64, https://doi.org/10.1016/j.physrep.2023.09.001.
  3. [3]J.A. Zeiss, G. Koßmann, R. Schwonnek, M. Plávala, Finite de FInetti for convex bodies and Polynomial Optimization, arXiv:2601.15184 (2026), https://arxiv.org/abs/2601.15184.
  4. [4]M. Ito, B.F. Lourenço, A bound on the Carathéodory number, Linear Algebra Appl. 532 (2017) 347–363, https://doi.org/10.1016/j.laa.2017.06.043.
  5. [5]G. Pataki, On the rank of extreme matrices in semidefinite programs and the multiplicity of optimal eigenvalues, Math. Oper. Res. 23 (2) (1998) 339–358, https://doi.org/10.1287/moor.23.2.339.

Paper details

vlnc:W_dyd9sjg6
Version
v1 · latest
Published
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Generative AI tools

View statement
Looking for related papers…
Contents