the continuum random tree
Just as Brownian motion is the universal scaling limit of random walks, the continuum random tree (CRT) is the universal scaling limit of large random trees. Introduced by Aldous, it is a random compact metric space — a 'tree' with uncountably many points and no edges of positive length — that arises whenever you take a critical Galton-Watson tree (or a uniform random labelled tree), condition it to be large, and shrink the distances. It is the canonical continuous random fractal tree, and the object many discrete tree results converge to.
The cleanest construction is via the Brownian excursion. Take a normalized Brownian excursion (e_s){0<=s<=1}, a Brownian-motion path conditioned to start and end at 0 and stay positive on (0,1). Define a pseudo-distance on [0,1] by d(s, t) = e_s + e_t - 2 min{ e_r : r between s and t }, and glue together points at distance 0; the resulting quotient metric space is the CRT. Intuitively the excursion is read as a contour: its local minima between two times measure how far back you must go to find the most recent common ancestor of the two corresponding leaves. The CRT is almost surely a real tree (any two points are joined by a unique non-self-intersecting path), it is compact, has Hausdorff dimension 2, and its leaves carry a natural mass measure. Aldous's theorem: a critical Galton-Watson tree with finite-variance offspring, conditioned to have n vertices and with graph distances rescaled by a constant times 1/sqrt(n), converges in distribution (in the Gromov-Hausdorff sense) to the CRT. Many discrete trees — uniform labelled trees, random binary trees, random plane trees — share this same limit, the content of the invariance principle for trees.
Why it matters: the CRT is to random-tree theory what Brownian motion is to random walks — a single universal limit unifying a whole class of discrete models, and the foundation for the scaling limits of random maps and the Brownian sphere/Liouville quantum gravity. It also encodes genealogies: the CRT is the genealogical structure dual to the Feller-diffusion continuous-state branching process, tying the population side and the tree side together. The honest hypotheses: universality holds for finite offspring variance; infinite-variance (alpha-stable) offspring converge instead to the family of stable trees, which are NOT the CRT and have Hausdorff dimension alpha/(alpha - 1). Convergence is in the Gromov-Hausdorff topology on metric spaces, a genuinely metric-geometric notion — the CRT is not a subset of any fixed space but an abstract random metric space, which is the conceptual leap the definition demands.
Pick a uniform random tree on n labelled vertices, view it as a metric space with each edge of length 1, then scale all distances by 1/sqrt(n). As n grows, this random shape converges (in Gromov-Hausdorff distance) to a constant multiple of the CRT — the same limit you get from critical Poisson or binary Galton-Watson trees conditioned on size, because all have finite offspring variance.
The CRT is built from a normalized Brownian excursion read as a contour; it is the universal limit of finite-variance random trees.
The CRT is the universal limit ONLY for finite-variance offspring; alpha-stable (infinite-variance) offspring converge to stable trees of dimension alpha/(alpha-1), not the CRT. Convergence is Gromov-Hausdorff on abstract metric spaces, not within any fixed embedding.