We study random relational structures that are relatively exchangeable-that is, whose distributions are invariant under the automorphisms of a reference structure M. When M is ultrahomogeneous and has trivial definable closure, all random structures relatively exchangeable with respect to M satisfy a general Aldous-Hoover-type representation. If M also satisfies the n-disjoint amalgamation property (n-DAP) for all n ≥ 1, then relatively exchangeable structures have a more precise description whereby each component depends locally on M.Date: September 18, 2018. 1991 Mathematics Subject Classification. 03C07 (Basic properties of first-order languages and structures); 03C98 (Applications of model theory); 60G09 (exchangeability). 1 2 HARRY CRANE AND HENRY TOWSNER L-structure on M . We call a pair X and Y of random L-structures on M equal in distribution, written X = D Y, if P(X| S = S) = P(Y| S = S) for every S ∈ L S , for all finite S ⊆ M .1.2. Relative exchangeability. Special cases of L-structures include binary relations, set partitions, undirected graphs, triangle-free graphs, as well as composite objects, e.g., a set together with a binary relation, a pair of graphs, etc. We are particularly interested in random L-structures that satisfy natural invariance properties with respect to the symmetries of another structure, of which exchangeability is a special case.We also call a probability measure µ exchangeable whenever X ∼ µ is an exchangeable L-structure.Given a large structure U = (Ω, R 1 , . . . , R r ) and a probability measure µ on Ω, we can obtain an exchangeable random L-structure X = (N, X 1 , . . . , X r ) by sampling elements φ(1), φ(2), . . . independently and identically distributed (i.i.d.) from µ and then defining X = U φ . Explicit representations of exchangeable structures are detailed in the work of de Finetti [15], Aldous [5], Hoover [17], and Kallenberg [18]. As a special case, the Aldous-Hoover theorem [5,17] characterizes the exchangeable random k-ary hypergraphs X = (N, X )-that is, the exchangeable random structures with a single symmetric k-ary relation-through the decomposition (1)x ∈ X ⇐⇒ f ((ξ s ) s⊆rng x ) = 1, where f is Borel measurable, the random variables ξ s are i.i.d. Uniform[0, 1], and rng x is the set of distinct elements in x. For instance, an exchangeable random graph can be generated by specifying a function f : [0, 1] 4 → {0, 1} with f (·, b, c, ·) = f (·, c, b, ·), selecting independent Uniform[0, 1] parameters ξ ∅ , ξ {i} for each i ∈ N, and ξ {i,j} for each pair i < j, and including the edge {i, j} exactly when f (ξ ∅ , ξ {i} , ξ {j} , ξ {i,j} ) = 1.Exchangeable structures not only play a fundamental role in probability theory [6,18], Bayesian inference [15], and applications in population genetics [19] but also have a natural place in the study of homogeneous structures in combinatorics [23] and mathematical logic [2][3][4]. In many applications, e.g., spin-glass models in statistical physics [8] and combinatorial stochastic processes [9,24], a random structure X is...