OctoInformBeta

OpenAI preprint, September 2026 · §6. From the block tree to the map · §6

Lemma 6.3 [Metric and mass comparison]: Measure-preserving surjections with metric difference ≤2ε\le 2\varepsilon put compact metric probability spaces within ε\varepsilon; continuous functions give tree bounds

Gromov-Hausdorff-Prokhorov bounds from codings and continuous functions.

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

Take two compact metric probability spaces realized as images of a common measure space under measure-preserving surjections whose pairwise distances differ by at most 2ε2\varepsilon for some ε>0\varepsilon > 0. How close must the two spaces be in the Gromov-Hausdorff-Prokhorov (GHP) topology, and what bounds follow for the excursion tree pseudometrics and excursion real trees coded by two continuous functions on [0,1][0, 1] within uniform distance ε\varepsilon?

Gromov-hausdorff-prokhorov Distance

A metric measure space presentation XX is a quadruple (X,dX,ℱX,μX)(X, d_X, \mathcal{F}_X, \mu_X) consisting of a set XX, a distance function dX:X×X→ℝd_X : X \times X \to \mathbb{R}, a σ\sigma-algebra ℱX\mathcal{F}_X on XX, and a measure μX\mu_X on (X,ℱX)(X, \mathcal{F}_X).✦ A compact metric probability space XX is a metric measure space presentation whose distance function dXd_X is a metric on XX under which XX is compact, whose σ\sigma-algebra ℱX\mathcal{F}_X is the Borel σ\sigma-algebra of dXd_X, and whose measure μX\mu_X satisfies μX(X)=1\mu_X(X) = 1.✦

An isometric amalgamation pseudometric dd for two metric measure space presentations XX and YY with canonical inclusions ιX:X→X⊔Y\iota_X : X \to X \sqcup Y and ιY:Y→X⊔Y\iota_Y : Y \to X \sqcup Y into their disjoint union is a pseudometric on X⊔YX \sqcup Y that satisfies (1) for all x,x′∈Xx, x' \in X and y,y′∈Yy, y' \in Y: d(ιX(x),ιX(x′))=dX(x,x′)andd(ιY(y),ιY(y′))=dY(y,y′).d(\iota_X(x), \iota_X(x')) = d_X(x, x') \quad \text{and} \quad d(\iota_Y(y), \iota_Y(y')) = d_Y(y, y').(1)✦

When X⊔YX \sqcup Y carries a pseudometric dd and its Borel σ\sigma-algebra, let dHd_{\mathrm{H}} and dLPd_{\mathrm{LP}} denote the Hausdorff and Lévy-Prokhorov distances on (X⊔Y,d)(X \sqcup Y, d) with values in [0,∞][0, \infty], and write (ιX)*μX(\iota_X)_* \mu_X and (ιY)*μY(\iota_Y)_* \mu_Y for the pushforward measures of μX\mu_X and μY\mu_Y. The GHP ambient embedding cost C(X,Y;d)C(X, Y; d) between metric measure space presentations XX and YY under a pseudometric dd on X⊔YX \sqcup Y is the extended real number in [0,∞][0, \infty] given by (2): C(X,Y;d)=max⁡{dH(ιX(X),ιY(Y)),dLP((ιX)*μX,(ιY)*μY)}.C(X, Y; d) = \max\bigl\{ d_{\mathrm{H}}(\iota_X(X), \iota_Y(Y)),\, d_{\mathrm{LP}}((\iota_X)_* \mu_X, (\iota_Y)_* \mu_Y) \bigr\}.(2)✦

Write 𝒜(X,Y)\mathcal{A}(X, Y) for the collection of all isometric amalgamation pseudometrics for two metric measure space presentations XX and YY. The GHP distance dGHP(X,Y)d_{\mathrm{GHP}}(X, Y) between two metric measure space presentations XX and YY is the extended real number in [0,∞][0, \infty] defined by (3): dGHP(X,Y)=infd∈𝒜(X,Y)C(X,Y;d).d_{\mathrm{GHP}}(X, Y) = \inf_{d \in \mathcal{A}(X, Y)} C(X, Y; d).(3)✦

Excursion Real Trees

Given a function e:[0,1]→ℝe : [0, 1] \to \mathbb{R}, the excursion tree pseudometric de:[0,1]×[0,1]→ℝd_e : [0, 1] \times [0, 1] \to \mathbb{R} is defined by (4): de(s,t)=e(s)+e(t)−2infr∈[s∧t,s∨t]e(r).d_e(s, t) = e(s) + e(t) - 2 \inf_{r \in [s \wedge t,\, s \vee t]} e(r).(4)✦

The excursion real tree 𝒯e\mathcal{T}_e of a function e:[0,1]→ℝe : [0, 1] \to \mathbb{R} is the quotient of [0,1][0, 1] by the equivalence relation generated by pairs (s,t)(s, t) with de(s,t)=0d_e(s, t) = 0.✦ The canonical excursion tree projection πe:[0,1]→𝒯e\pi_e : [0, 1] \to \mathcal{T}_e of a function e:[0,1]→ℝe : [0, 1] \to \mathbb{R} is the quotient map sending each t∈[0,1]t \in [0, 1] to its equivalence class in 𝒯e\mathcal{T}_e.✦

Let λ\lambda denote Lebesgue measure on [0,1][0, 1]. The excursion tree metric measure space presentation 𝒯e\mathcal{T}_e of a function e:[0,1]→ℝe : [0, 1] \to \mathbb{R} is the metric measure space presentation whose carrier set is the excursion real tree 𝒯e\mathcal{T}_e, with distance between x,y∈𝒯ex, y \in \mathcal{T}_e given by de(s,t)d_e(s, t) at chosen representatives s∈xs \in x and t∈yt \in y, Borel σ\sigma-algebra of the quotient topology induced by πe\pi_e, and measure μe=(πe)*λ\mu_e = (\pi_e)_* \lambda.

The Claim

Theorem 1. Let XX and YY be compact metric probability spaces, let ZZ be a nonempty measurable space with a measure μ\mu, and let f:Z→Xf : Z \to X and g:Z→Yg : Z \to Y be measurable surjections satisfying f*μ=μXf_* \mu = \mu_X and g*μ=μYg_* \mu = \mu_Y. If ε>0\varepsilon > 0 satisfies |dX(f(s),f(t))−dY(g(s),g(t))|≤2ε|d_X(f(s), f(t)) - d_Y(g(s), g(t))| \le 2\varepsilon for all s,t∈Zs, t \in Z, then (5) holds: dGHP(X,Y)≤ε.d_{\mathrm{GHP}}(X, Y) \le \varepsilon.(5)✦

Let e,f:[0,1]→ℝe, f : [0, 1] \to \mathbb{R} be continuous functions and let ε∈ℝ\varepsilon \in \mathbb{R} satisfy |e(r)−f(r)|≤ε|e(r) - f(r)| \le \varepsilon for all r∈[0,1]r \in [0, 1]. Then for all s,t∈[0,1]s, t \in [0, 1], the excursion tree pseudometrics satisfy (6): |de(s,t)−df(s,t)|≤4ε.|d_e(s, t) - d_f(s, t)| \le 4\varepsilon.(6)✦

When e,f:[0,1]→ℝe, f : [0, 1] \to \mathbb{R} are continuous and ε>0\varepsilon > 0 satisfies |e(t)−f(t)|≤ε|e(t) - f(t)| \le \varepsilon for all t∈[0,1]t \in [0, 1], the excursion tree metric measure spaces satisfy (7): dGHP(𝒯e,𝒯f)≤2ε.d_{\mathrm{GHP}}(\mathcal{T}_e, \mathcal{T}_f) \le 2\varepsilon.(7)✦

In the paper, the result reads: if a correspondence between two compact metric probability spaces XX and YY has distortion at most ε>0\varepsilon > 0 and a coupling of their probability measures is supported on that correspondence, then dGHP(X,Y)≤εd_{\mathrm{GHP}}(X, Y) \le \varepsilon, and if one marginal of the coupling differs from the desired measure by total variation at most η\eta, the bound is ε+η\varepsilon + \eta.

What Is Checked

The claims are stated over metric measure space presentations XX and YY that are compact metric probability spaces, with GHP distance defined via isometric amalgamation pseudometrics and the ambient embedding cost on X⊔YX \sqcup Y, together with measurable surjections from a nonempty measure space (Z,μ)(Z, \mu) pushing μ\mu forward to μX\mu_X and μY\mu_Y under pairwise distance differences at most 2ε2\varepsilon for ε>0\varepsilon > 0, as well as over continuous functions e,f:[0,1]→ℝe, f : [0, 1] \to \mathbb{R} within uniform distance ε\varepsilon, with ε>0\varepsilon > 0 for the tree GHP comparison, their excursion tree pseudometrics ded_e and dfd_f, and their excursion tree metric measure space presentations 𝒯e\mathcal{T}_e and 𝒯f\mathcal{T}_f defined via the excursion real trees and canonical projections πe\pi_e and πf\pi_f.