--- title: Multidimensional Mean-Preserving Spreads url: https://www.emergentmind.com/topics/multidimensional-mean-preserving-spreads type: topic --- # Multidimensional Mean-Preserving Spreads Searching arXiv for the cited paper to ground the article in the primary source. arXiv search query: 1905.05157 Multidimensional mean-preserving spreads are naturally situated within convex-order comparisons of probability measures, but the most directly relevant result in the present context is formulated for one-dimensional, purely atomic laws. "Mixtures of Mean-Preserving Contractions" [1905.05157] studies mean-preserving contractions on \(X=\mathbb R\) and establishes a finite-support decomposition theorem: if a purely atomic prior \(P\) has support on \(n\) points, then any mean-preserving contraction of \(P\) can be represented as a mixture of simpler contractions, each with support on at most \(n\) points. Although this is not a multivariate theorem on \(\mathbb R^d\), it isolates a support-compression principle, a stochastic-matrix representation, and a convex-order duality that are conceptually informative for multidimensional and generalized formulations. ## 1. Order-theoretic setting The paper is written in terms of **mean-preserving contractions (mpcs)** rather than mean-preserving spreads. In the standard Rothschild–Stiglitz language, a mean-preserving spread is the opposite order relation: if \(Q\) is an mpc of \(P\), then \(P\) is a mean-preserving spread of \(Q\), or equivalently \(Q\) is less risky than \(P\) in convex order. The contraction interpretation is described as **collapsing portions of mass to their barycenters**, whereas the spread interpretation is the opposite operation, in which mass is **spread out**. In convex-order terms, the relation is \[ Q \preceq_{\mathrm{cx}} P \quad\Longleftrightarrow\quad Q \text{ is an mpc of } P \quad\Longleftrightarrow\quad P \text{ is an mps of } Q. \] In martingale language, if \(Q\) is a contraction of \(P\), there exists a coupling \((X,Y)\) with laws \[ X\sim P,\qquad Y\sim Q,\qquad \mathbb E[X\mid Y]=Y. \] Reversing the roles yields the spread formulation \[ \mathbb E[Y\mid X]=X. \] The paper is not explicitly multidimensional. Its measures belong to \(\mathcal P(\mathbb R)\), with \(P\) finitely supported and \(Q\) purely atomic or weak limits thereof. Its relevance to multidimensional mean-preserving spreads is therefore indirect. It is best understood as a result on finite-support mean-preserving transformations represented by stochastic matrices, together with a decomposition principle that depends on atomic convex structure rather than on calculus specific to \(\mathbb R\). ## 2. Atomic formulation and basic definitions The formal setup is \[ X=\mathbb R, \] with \(\mathcal B\) the Borel \(\sigma\)-algebra and \(\mathcal P\) the set of Borel probability measures on \((X,\mathcal B)\). A purely atomic probability measure \(P\) with support on \(n\) points is written \[ \operatorname{supp} P=\vec a:=\{a_1,\dots,a_n\}, \] with masses \(p_1,\dots,p_n>0\), \(\sum_{i=1}^n p_i=1\), and, without loss of generality, \[ a_10\), \(\sum_{j=1}^m q_j=1\), and \[ \vec q=(q_1,\dots,q_m),\qquad \vec{qb}=(q_1b_1,\dots,q_mb_m). \] A **Simple Mean-Preserving Contraction (SMPC)** of \(P\) is defined by the existence of a non-negative row-stochastic \(n\times m\) matrix \(F\) such that \[ \vec p\,F=\vec q \] and \[ (\vec{pa})\,F=\vec{qb}. \] The set of all SMPCs of \(P\) is denoted \(\mathcal m(P)\) [1905.05157]. If \(F=(f_{ij})\), row \(i\) specifies how the mass \(p_i\) at atom \(a_i\) is distributed among the output atoms \(b_j\). Row-stochasticity means \[ f_{ij}\ge 0,\qquad \sum_{j=1}^m f_{ij}=1\quad \forall i. \] The identity \(\vec p\,F=\vec q\) states that the mass arriving at \(b_j\) is \(q_j\), while \[ (\vec{pa})F=\vec{qb} \] states that the first moment arriving at \(b_j\) is \(q_j b_j\). Whenever \(q_j>0\), \[ b_j=\frac{\sum_{i=1}^n p_i a_i f_{ij}}{\sum_{i=1}^n p_i f_{ij}}, \] so each output atom is the barycenter of the mass assigned to it. This is the paper’s exact “collapse to barycenters” interpretation. A general **Mean-Preserving Contraction (MPC)** is then defined by weak closure: \[ R\in\mathcal P \text{ is an MPC of } P \quad\Longleftrightarrow\quad \exists \{Q_m\}_{m=1}^\infty\subset \mathcal m(P)\text{ such that }Q_m\to_w R. \] Thus \[ \mathcal M(P)=\overline{\mathcal m(P)}^{\,w}. \] The paper does not give a separate formal definition of mean-preserving spread, but it explicitly presents it as the equivalent and opposite notion. In standard notation, \[ P \text{ is an mps of } Q \quad\Longleftrightarrow\quad \int \phi\,dP \ge \int \phi\,dQ \quad \forall \text{ convex }\phi, \] subject to equality of means. ## 3. Mixtures, support size, and the principal decomposition theorem Mixtures are defined at the level of the associated Markov matrices. If \(Q,Q',Q''\) are SMPCs of \(P\), then \[ Q=\alpha Q'+(1-\alpha)Q'' \] if and only if \[ F=\alpha F' + (1-\alpha)F'', \tag{1} \] after inserting zero-columns in \(F'\) and \(F''\) for atoms absent from \(Q'\) or \(Q''\), while preserving within-column ratios. In this formulation, “mixture” is not merely an arbitrary convex combination of measures; it is a convex combination compatible with the SMPC matrix representation. The support count is central. If \(P\) has support size \(n\) and \(Q\) has support size \(m\), the main theorem states that support complexity beyond \(n\) is reducible. The principal result is: \[ \textbf{Theorem.} \] Let \(Q\in \mathcal m(P)\) be any SMPC of \(P\) with support on \(n+1\) points, \(n\ge 2\). Then \(Q\) is the convex combination of two purely atomic probability measures \(Q'\) and \(Q''\), with \[ Q',Q''\in \mathcal m(P), \] each with support on at most \(n\) points. Moreover, \(Q'\) and \(Q''\) are unique. The paper’s main corollary extends this reduction to all mean-preserving contractions: \[ \textbf{Corollary.} \] Any \(Q\in \mathcal M(P)\) is a mixture of SMPCs with support on at most \(n\) points. The significance of the theorem, viewed through the dual spread relation, is structural. An apparently complicated mean-preserving transformation with many output atoms is not irreducible. It belongs to the convex hull of simpler transformations whose support size is bounded by the support size of the original law. This identifies support size of the source distribution, rather than ambient Euclidean dimension, as the controlling parameter in the theorem’s decomposition logic. ## 4. Proof architecture and the support-compression mechanism The proof operates directly on the \(n\times (n+1)\) Markov matrix \(F\) associated with an SMPC \(Q\). Writing the columns of \(F\) as vectors \[ \vec f_1,\dots,\vec f_{n+1}\in\mathbb R^n, \] there are \(n+1\) vectors in \(\mathbb R^n\), hence linear dependence: \[ \sum_{j=1}^{n+1}\hat c_j \vec f_j=0 \] for some nonzero coefficients \(\hat c_j\). Because the columns are nonnegative, the coefficients must include both positive and negative signs. After reindexing, \[ \sum_{j=1}^{k} c_j \vec f_j = \sum_{j=k+1}^{n+1} c_j \vec f_j, \qquad c_j\ge 0. \] Then one selects the maximal coefficients on each side, \[ c_{j^*}=\max_{1\le j\le k} c_j, \qquad c_{j^{**}}=\max_{k+1\le j\le n+1} c_j. \] The proof’s central device is the **zeroing** of one of these maximal-coefficient columns. If \(\vec f_{j^*}\) is zeroed, it is rewritten using the dependence relation, and a new matrix \(F'\) is defined by scaling the remaining columns: \[ \vec f'_j= \begin{cases} \left(1-\frac{c_j}{c_{j^*}}\right)\vec f_j, & j\le k,\ j\neq j^*,\\[4pt] \left(1+\frac{c_j}{c_{j^*}}\right)\vec f_j, & j\ge k+1,\\[4pt] 0, & j=j^*. \end{cases} \] Because \(c_{j^*}\) is maximal on its side, the coefficients \(1-\frac{c_j}{c_{j^*}}\) lie in \([0,1]\). The row sums remain \(1\), all entries remain in \([0,1]\), and \(F'\) is again a valid row-stochastic nonnegative matrix. Hence it defines another SMPC \(Q'\) with one fewer nonzero column. The same construction applied to \(j^{**}\) yields \(F''\) and \(Q''\), and then \[ F=\alpha F'+(1-\alpha)F'' \] for some \(\alpha\in(0,1)\), which implies \[ Q=\alpha Q'+(1-\alpha)Q''. \] The support bound follows immediately: each zeroing removes one nonzero column, so an \(n\times(n+1)\) matrix becomes one with at most \(n\) nonzero columns. The uniqueness of \(Q'\) and \(Q''\) is proved by showing that no third independently zeroable column can appear unless it reproduces one of the same two outcomes [1905.05157]. The proof is presented as linear algebra on supports and convex geometry of stochastic matrices. It is not formulated via Strassen’s theorem, explicit martingale couplings, Choquet theory, or Carathéodory’s theorem, even though the overall reduction has an evident convex-analytic character. ## 5. Interpretation for multidimensional mean-preserving spreads The paper does not furnish a multivariate theorem for distributions on \(\mathbb R^d\). Its state space is one-dimensional, and its proof relies on a matrix representation specialized to scalar atomic supports. Accordingly, it does not establish that any multidimensional mean-preserving spread admits an analogous decomposition with the same support bound \(n\). Its relevance to multidimensional mean-preserving spreads is conceptual rather than direct. The theorem suggests that when a finitely supported prior has \(n\) support points, a contraction producing more than \(n\) posterior means is not fundamentally new: it is representable as a mixture of simpler contractions, each involving at most \(n\) support points. Reversing the order relation gives the corresponding intuition for spreads. This suggests that complicated spread–contraction relations may often be analyzed through mixtures of simpler atomic transformations with controlled support complexity. A plausible implication is that finite-dimensional convex decomposition, rather than ambient-space geometry alone, is central to understanding generalized spread phenomena. The paper’s “dimension” is the number of support points of the source measure, not the Euclidean dimension of the state space. That distinction is especially important in multidimensional discussions, where support complexity and ambient dimension need not coincide. At the same time, the limitations are explicit. New work would be required for a full multidimensional theory, because barycenters become vector-valued, convex-order structure in higher dimensions is subtler, and the present proof depends on a stochastic-matrix argument tailored to the one-dimensional atomic case. ## 6. Example, applications, and connections The paper includes a concrete example with a 3-point prior and a 4-point contraction. The prior is \[ P= \begin{Bmatrix} 0 & \tfrac12 & 1\\ \tfrac{3}{10} & \tfrac{3}{10} & \tfrac25 \end{Bmatrix}. \] A 4-point SMPC is \[ Q= \begin{Bmatrix} \tfrac16 & \tfrac12 & \tfrac34 & \tfrac56\\ \tfrac{3}{10} & \tfrac15 & \tfrac15 & \tfrac{3}{10} \end{Bmatrix}, \] with Markov matrix \[ F= \begin{pmatrix} \frac23 & \frac13 & 0 & 0\\ \frac13 & 0 & \frac13 & \frac13\\ 0 & \frac14 & \frac14 & \frac12 \end{pmatrix}. \] The theorem decomposes this \(Q\) into two 3-point SMPCs \(Q'\) and \(Q''\) such that \[ Q=\alpha Q'+(1-\alpha)Q'',\qquad \alpha=\frac47. \] This example exhibits the support-reduction mechanism in explicit finite form. The principal application is to **linear persuasion**. When sender and receiver utilities depend only on posterior means, the sender’s choice of information structure can be reduced to \[ \max_{Q\in\mathcal M(P)} \mathbb E_Q[u(x)]. \] Because \(\mathcal M(P)\) is convex and compact, Bauer’s Maximum Principle implies that a linear objective attains its maximum at an extreme point of \(\mathcal M(P)\). The decomposition theorem yields a necessary condition for extremality: support size at most \(n\). The paper states the resulting proposition as follows: if the prior \(P\) has support on \(n\) points, an optimal signal requires at most \(n\) messages. The paper also treats **competitive persuasion**. Since each pure strategy corresponds to an mpc of the prior, the decomposition result implies that profitable deviations need only be checked among mpcs with support on at most \(n\) points, and that an equilibrium continuous distribution over posterior means can be implemented as a mixed strategy over support-\(\le n\) mpcs. These applications concern information structures and decision under uncertainty rather than a direct multivariate risk-comparison theorem [1905.05157]. The broader connections identified in the paper include Rothschild–Stiglitz, majorization, statistical experiments, Blackwell comparisons, Strassen, and Hill and Joe’s notion of **fusion**. In the one-dimensional finite-first-moment setting discussed there, fusion is identified with mean-preserving contraction. The set of mpcs of a discrete measure is described as compact and convex, so by Krein–Milman it is the closed convex hull of its extreme points. Within that literature, the paper contributes a finite-support convex-analytic representation result: every contraction of an \(n\)-atom prior is a mixture of simpler contractions supported on at most \(n\) points. Source: https://www.emergentmind.com/topics/multidimensional-mean-preserving-spreads