OctoInform Beta

OpenAI preprint, September 2026 · §4. The operator square · §4

Lemma 4.2 [Exact lines]: For three or more coordinates, two exact tuples are orthogonal or agree up to phases coordinate by coordinate; nonconstant exact lines have mean zero

Independence of all but one coordinate makes exact lines rigid.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Let II be a finite set of indices and YY a standard Borel space. A family (Xi)i∈I(X_i)_{i \in I} of random variables is (|I|−1)(|I|-1)-wise independent when, for each i∈Ii \in I, the family (Xj)j≠i(X_j)_{j \neq i} is independent. The item concerns a probability measure μ\mu on YIY^I with marginals νi\nu_i whose coordinates yiy_i are (|I|−1)(|I|-1)-wise independent, for at least three indices; for |I|=3|I| = 3 this makes μ\mu a pairwise independent coupling of ν1\nu_1, ν2\nu_2 and ν3\nu_3.

The circle 𝕋={z∈ℂ:|z|=1}\mathbb{T} = \{ z \in \mathbb{C} : |z| = 1 \} carries the Borel structure of ℂ\mathbb{C}✦, and the circle group ℳ(Y)\mathcal{M}(Y) of YY is the group of measurable maps f:Y→𝕋f \colon Y \to \mathbb{T} under pointwise multiplication.✦ An exact tuple for μ\mu is a tuple f=(fi)i∈If = (f_i)_{i \in I} in ℳ(Y)I\mathcal{M}(Y)^I whose product is almost surely constant: for some c∈𝕋c \in \mathbb{T}, ∏i∈Ifi(yi)=cfor μ-almost every y∈YI.\prod_{i \in I} f_i(y_i) = c \quad \text{for $\mu$-almost every } y \in Y^I.(1)✦ The exact tuples form the set E(μ)⊆ℳ(Y)IE(\mu) \subseteq \mathcal{M}(Y)^I.

The item is about how rigid exact tuples are, up to constant phases: how two of them compare at each coordinate, and how far one coordinate determines the rest.

Lines and Projective Tuples

For a measure ν\nu on YY, the phase subgroup Φ(ν)\Phi(\nu) of ℳ(Y)\mathcal{M}(Y) consists of the maps that are ν\nu-almost everywhere equal to a constant✦, and the projective circle group of ν\nu is the quotient P(ν)=ℳ(Y)/Φ(ν).P(\nu) = \mathcal{M}(Y) / \Phi(\nu).(2)✦ The line [f]ν[f]_\nu of f∈ℳ(Y)f \in \mathcal{M}(Y) is its class in P(ν)P(\nu). Two lines [f]ν[f]_\nu and [g]ν[g]_\nu in P(ν)P(\nu) are orthogonal when ∫Yfg¯dν=0\int_Y f \overline{g}\, d\nu = 0, which does not depend on the choice of representatives.

The projectivization of tuples is the map π:ℳ(Y)I→∏i∈IP(νi),π(f)=([fi]νi)i∈I,\pi \colon \mathcal{M}(Y)^I \to \prod_{i \in I} P(\nu_i), \qquad \pi(f) = \big([f_i]_{\nu_i}\big)_{i \in I},(3)✦ and the exact projective tuples are its values on exact tuples, ℰ(μ)=π(E(μ))\mathcal{E}(\mu) = \pi(E(\mu)). The projection σi:ℰ(μ)→P(νi)\sigma_i \colon \mathcal{E}(\mu) \to P(\nu_i) to coordinate ii takes a projective tuple to its iith entry, and the exact lines at coordinate ii form its image ℰi(μ)=σi(ℰ(μ))\mathcal{E}_i(\mu) = \sigma_i(\mathcal{E}(\mu)).✦ The partners of an exact line at coordinate ii are the entries, at the other coordinates, of the exact projective tuple whose iith entry is that line.

Codes

Exact lines are labelled by codes, which fix a vector of L2L^2 by its distances to a countable family of test functions. Fix a sequence of sets generating the σ\sigma-algebra of YY; the test algebra 𝒜\mathcal{A} is the algebra of sets it generates.✦ A test key is a finite list of triples (ar,br,Ar)(a_r, b_r, A_r) with ar,br∈ℚa_r, b_r \in \mathbb{Q} and Ar∈𝒜A_r \in \mathcal{A}✦, and its test function is t=∑r(ar+ibr)𝟏Ar.t = \sum_r (a_r + \mathrm{i}\, b_r)\, \mathbf{1}_{A_r}.(4)✦ The test keys form a countable set; fix an enumeration k0,k1,…k_0, k_1, \dots of it and let tnt_n be the test function of knk_n. For a finite measure ν\nu on YY, the test vectors of ν\nu are the classes of t0,t1,…t_0, t_1, \dots in L2(ν)L^2(\nu), in this order.✦

A distance code for a matrix (dnm)(d_{nm}) of reals is a sequence (zn)(z_n) of nonnegative reals with zn≤dnm+zm,dnm≤zn+zmfor all n,m,infnzn=0.z_n \le d_{nm} + z_m, \qquad d_{nm} \le z_n + z_m \quad \text{for all } n, m, \qquad \inf_n z_n = 0.(5)✦ Let κ=(κu)u∈U\kappa = (\kappa_u)_{u \in U} be a kernel from a measurable space UU to YY, that is, a family of finite measures κu\kappa_u on YY depending measurably on u∈Uu \in U. A code over κ\kappa is a pair p=(u,z)p = (u, z) of a point u∈Uu \in U, the base point of pp, and a distance code zz for the distances dnm=∥tn−tm∥d_{nm} = \|t_n - t_m\| of the test vectors of κu\kappa_u.✦ The vector of pp is a vp∈L2(κu)v_p \in L^2(\kappa_u) with zn=∥tn−vp∥z_n = \|t_n - v_p\| for every nn.✦ The approximation index np(m)n_p(m) of pp is the least nn with zn<2−mz_n < 2^{-m}✦, and the representative of pp is the pointwise limit of its approximants tnp(m)t_{n_p(m)}, ϕp(y)=limm→∞tnp(m)(y),y∈Y.\phi_p(y) = \lim_{m \to \infty} t_{n_p(m)}(y), \qquad y \in Y.(6)✦ The distance of two codes pp and qq, with uu the base point of pp, is dist⁡(p,q)=(∫Y|ϕp−ϕq|2dκu)1/2.\operatorname{dist}(p, q) = \Big( \int_Y |\phi_p - \phi_q|^2 \, d\kappa_u \Big)^{1/2}.(7)✦

A vector vv of a Hilbert space is canonical for a sequence (sn)(s_n) of vectors when the first sns_n not orthogonal to vv has a positive inner product with it: for some nn, ⟨sm,v⟩=0 for m<n,⟨sn,v⟩>0.\langle s_m, v \rangle = 0 \ \text{ for } m < n, \qquad \langle s_n, v \rangle > 0.(8)✦ A code pp over κ\kappa with base point uu is canonical when vpv_p is canonical for the test vectors of κu\kappa_u✦, and unimodular when |ϕp|=1|\phi_p| = 1 holds κu\kappa_u-almost everywhere.✦

The Claim

Theorem 1 (Exact lines). Let II be a finite set and (Ω,P)(\Omega, P) a probability space.

Suppose |I|≥3|I| \ge 3. Let (Xi)i∈I(X_i)_{i \in I} be (|I|−1)(|I|-1)-wise independent random variables on Ω\Omega with values in measurable spaces SiS_i, and let fi,gi:Si→ℂf_i, g_i \colon S_i \to \mathbb{C} be measurable maps with |fi(Xi)|=|gi(Xi)|=1|f_i(X_i)| = |g_i(X_i)| = 1 almost surely, such that ∏ifi(Xi)\prod_i f_i(X_i) and ∏igi(Xi)\prod_i g_i(X_i) are almost surely constant. Then, for some phases ci∈𝕋c_i \in \mathbb{T}, 𝔼[fi(Xi)gi(Xi)¯]=0 for all i∈Iorfi(Xi)=cigi(Xi) a.s. for all i∈I.\mathbb{E}\big[f_i(X_i)\, \overline{g_i(X_i)}\big] = 0 \ \text{ for all } i \in I \qquad \text{or} \qquad f_i(X_i) = c_i\, g_i(X_i) \ \text{ a.s. for all } i \in I.(9)✦

Let (fi)i∈I(f_i)_{i \in I} be (|I|−1)(|I|-1)-wise independent complex random variables on Ω\Omega with |fi|=1|f_i| = 1 almost surely and ∏ifi\prod_i f_i almost surely constant. Then |𝔼fi|=∏j≠i|𝔼fj|for every i∈I.|\mathbb{E} f_i| = \prod_{j \neq i} |\mathbb{E} f_j| \quad \text{for every } i \in I.(10)✦

If |I|≥3|I| \ge 3 and nonnegative reals aia_i satisfy ai=∏j≠iaja_i = \prod_{j \neq i} a_j for every i∈Ii \in I, then ai=0 for every i∈Iorai=1 for every i∈I.a_i = 0 \ \text{ for every } i \in I \qquad \text{or} \qquad a_i = 1 \ \text{ for every } i \in I.(11)✦

Suppose now that |I|≥3|I| \ge 3 and that μ\mu is a probability measure on YIY^I with marginals νi\nu_i and (|I|−1)(|I|-1)-wise independent coordinates. If f∈E(μ)f \in E(\mu) and [fi]νi=1[f_i]_{\nu_i} = 1 for one i∈Ii \in I, then [fj]νj=1for every j∈I.[f_j]_{\nu_j} = 1 \quad \text{for every } j \in I.(12)✦ The projections are injective: σi:ℰ(μ)→P(νi) is injective for every i∈I,\sigma_i \colon \mathcal{E}(\mu) \to P(\nu_i) \ \text{ is injective for every } i \in I,(13)✦ and for all i,j∈Ii, j \in I, σj∘σi−1:ℰi(μ)→ℰj(μ) is an isomorphism of groups.\sigma_j \circ \sigma_i^{-1} \colon \mathcal{E}_i(\mu) \to \mathcal{E}_j(\mu) \ \text{ is an isomorphism of groups.}(14)✦ If f∈E(μ)f \in E(\mu) and [fi]νi≠1[f_i]_{\nu_i} \neq 1, then ∫Yfidνi=0.\int_Y f_i \, d\nu_i = 0.(15)✦ Finally, let (pi)i∈I(p_i)_{i \in I} and (qi)i∈I(q_i)_{i \in I} be tuples of canonical unimodular codes over κ\kappa such that pip_i and qiq_i have the same base point uiu_i and νi=κui\nu_i = \kappa_{u_i} for every ii, and such that |∫∏iϕpi(yi)dμ(y)|=|∫∏iϕqi(yi)dμ(y)|=1\big|\int \prod_i \phi_{p_i}(y_i)\, d\mu(y)\big| = \big|\int \prod_i \phi_{q_i}(y_i)\, d\mu(y)\big| = 1. Then p=qordist⁡(pi,qi)≥1 for every i∈I.p = q \qquad \text{or} \qquad \operatorname{dist}(p_i, q_i) \ge 1 \ \text{ for every } i \in I.(16)✦

In the paper, the result reads: if there are m=|I|≥3m = |I| \ge 3 coordinates and any m−1m - 1 of them are independent, then distinct exact lines at a coordinate are orthogonal, an exact line has a unique tuple of partner lines, the exact tuples of lines form an abelian group whose projection to any coordinate is an isomorphism onto the exact lines there, these groups are countable when L2L^2 is separable, and nonconstant exact lines have mean zero.

The box gives the orthogonality as the first alternative of (9), the unique partners as (13), the isomorphisms as (14) and the zero mean as (15).

What Is Checked

The claims are stated over a finite index set II, with at least three elements except in (10): over (|I|−1)(|I|-1)-wise independent random variables on a probability space in (9) and (10), over nonnegative reals in (11), and otherwise over a probability measure μ\mu on YIY^I with marginals νi\nu_i and (|I|−1)(|I|-1)-wise independent coordinates, the circle group ℳ(Y)\mathcal{M}(Y), the projective circle groups P(νi)P(\nu_i), the exact tuples E(μ)E(\mu), the exact projective tuples ℰ(μ)\mathcal{E}(\mu) with their projections σi\sigma_i, and canonical unimodular codes over κ\kappa with their representatives and distances.

The library proves that the measurable maps Y→𝕋Y \to \mathbb{T} form a subgroup of the group of all maps Y→𝕋Y \to \mathbb{T}✦, that Φ(ν)\Phi(\nu) and E(μ)E(\mu) are subgroups of ℳ(Y)\mathcal{M}(Y) and of ℳ(Y)I\mathcal{M}(Y)^I, and that π\pi and each σi\sigma_i are homomorphisms of groups, which makes ℰ(μ)\mathcal{E}(\mu) a subgroup of ∏i∈IP(νi)\prod_{i \in I} P(\nu_i).