Fréchet Mean & Variance Methods

• Fréchet mean and variance methods extend classical definitions of location and dispersion to general metric spaces by minimizing expected squared or transformed distances.
• They provide a robust framework applicable to diverse geometries such as Riemannian manifolds, Hadamard spaces, and graph spaces, facilitating inference in non-Euclidean settings.
• The methodology encompasses weak convergence, computational strategies, and asymptotic theories, addressing challenges like nonuniqueness and heavy-tailed data.

Fréchet mean and variance methods extend the notions of location and dispersion from Euclidean spaces to random objects valued in a metric space by minimizing expected squared distance, or more generally transformed distance, without requiring linear structure. In their most basic form, they study the Fréchet function $p \mapsto \mathbb{E}[d(X,p)^q]$, the associated mean set of minimizers, and the minimal value as variance; in current research, this framework now covers general metric spaces, infinite-dimensional and non-smooth settings, Hadamard spaces, Riemannian manifolds, Wasserstein spaces, tropical projective tori, persistence-diagram spaces, BHV treespace, graph spaces, and regression problems with non-Euclidean responses (Jaffe, 2024, Dubey et al., 2017, Schötz, 10 Nov 2025).

1. Core definitions and variants

Let $(S,d)$ be a metric space and $X \sim \mu$ a Borel probability measure on $S$. For $q \ge 1$, the Fréchet $q$-function is

$$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$

For $q=2$, the Fréchet mean set and Fréchet variance are

$$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$

If the mean is unique $m \in S$, then $(S,d)$0. For i.i.d. $(S,d)$1, the sample analogues are

$(S,d)$2

(Jaffe, 2024).

A key generalization is the canonical renormalized Fréchet functional

$(S,d)$3

defined for $(S,d)$4 and $(S,d)$5. By construction, $(S,d)$6 is well-defined under only $(S,d)$7-moment assumptions, and the Fréchet $(S,d)$8-mean set can be written equivalently as

$(S,d)$9

which does not depend on $X \sim \mu$0. The corresponding origin-dependent Fréchet $X \sim \mu$1-variance is

$X \sim \mu$2

(Jaffe, 2024).

A second generalization replaces squared distance by a transformation $X \sim \mu$3. In a Hadamard space, for $X \sim \mu$4 nondecreasing, convex, with concave derivative $X \sim \mu$5, the transformed Fréchet functional is

$X \sim \mu$6

and a transformed Fréchet mean $X \sim \mu$7 is any minimizer of $X \sim \mu$8. This class includes the Fréchet median, the Fréchet mean, the Huber loss-induced Fréchet mean, power Fréchet means, the pseudo-Huber mean, and related robust statistics (Schötz, 2023, Schötz, 10 Nov 2025).

On compact Riemannian manifolds, a smooth surrogate is provided by Varadhan functions. With heat kernel $X \sim \mu$9,

$S$0

For an $S$1-valued random variable $S$2 with law $S$3, the population Varadhan function, variance, and mean are

$S$4

which reduce at $S$5 to the Fréchet function, variance, and mean (Cao, 6 Jan 2026).

2. Existence, continuity, and asymptotic theory in general metric spaces

A central development is a weak-convergence framework for Fréchet means in general metric spaces. A convergence $S$6 on $S$7 is called a weak convergence for $S$8 if it satisfies three axioms: weak compactness of $S$9-bounded sets, lower semicontinuity of the distance, and a Kadec–Klee-type strengthening that upgrades weak convergence to metric convergence when distances agree. This $q \ge 1$0 need not be topological in the usual sense and is genuinely weaker than metric convergence (Jaffe, 2024).

For $q \ge 1$1, define

$q \ge 1$2

and let the topology $q \ge 1$3 on $q \ge 1$4 be given by $q \ge 1$5 iff $q \ge 1$6 and the $q \ge 1$7-moments converge. The fundamental continuity theorem states: if $q \ge 1$8 is separable and admits a weak convergence $q \ge 1$9, $q$0, and $q$1 in $q$2, then for any selections $q$3, the sequence $q$4 is relatively $q$5-compact, and every $q$6-cluster point lies in $q$7. Equivalently, the mean sets $q$8-converge, including Painlevé–Kuratowski outer limit, and one obtains one-sided Hausdorff convergence

$q$9

(Jaffe, 2024).

Several consequences are immediate. For every $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$0, $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$1 is nonempty and $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$2-compact, avoiding Heine–Borel assumptions. The set-valued map $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$3 is continuous as a map into compact subsets with one-sided Hausdorff distance. For any fixed origin $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$4, $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$5 whenever $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$6 in $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$7. The theorem works under the provably minimal assumption $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$8; for $$F_q(p) := \mathbb{E}[d(X,p)^q] = \int_S d(x,p)^q \,\mu(dx).$$9, only one finite moment is needed, and for $q=2$0, no moment condition is required (Jaffe, 2024).

This framework removes uniqueness assumptions. The mean object is treated as a closed, bounded, compact set, and convergence is established for the entire mean sets. Thus, when the mean is nonunique, every sample mean is asymptotically close to some population mean; when unique, continuity reduces to standard metric continuity (Jaffe, 2024).

The asymptotic consequences include strong laws, ergodic theorems, and large deviations. For i.i.d. $q=2$1,

$q=2$2

where $q=2$3. Variances are consistent, and under exponential $q=2$4-moment conditions and uniqueness, an LDP holds for empirical Fréchet means via Sanov’s theorem and contraction. By contrast, CLTs are not derived in this abstract framework because of lack of differentiable structure (Jaffe, 2024).

Hadamard-space theory adds curvature-driven lower bounds. For $q=2$5, if $q=2$6 is a transformed Fréchet mean, then for all $q=2$7,

$q=2$8

These variance inequalities describe how the expected transformed distance grows when moving away from a minimizer, imply uniqueness under natural positivity assumptions on $q=2$9, and are useful in the theory of estimation and numerical approximation (Schötz, 2023). Finite-sample rates in expectation for transformed Fréchet means in Hadamard spaces strengthen this picture: the parametric rate of convergence is obtained under fewer than two moments, while a subclass of estimators with bounded positive derivative exhibits a breakdown point of $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$0 (Schötz, 10 Nov 2025).

3. Geometry, robustness, and smooth surrogates

The weak-convergence framework applies in many infinite-dimensional or non-smooth settings. Metric convergence itself is a weak convergence in Heine–Borel spaces. Uniformly convex Banach spaces admit weak convergence through Milman–Pettis weak compactness of closed balls, weak lower semicontinuity of the norm, and the Kadec–Klee or Radon–Riesz property. Hadamard spaces inherit Jost’s weak convergence without local compactness and in infinite dimensions. Wasserstein spaces $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$1 over complete, separable, locally compact bases also fit through tightness, Prokhorov, Fatou in optimal couplings, and uniform integrability criteria. The framework is closed under weak-closed subspaces, products, quotients by compact groups, and semi-quotients by compact connected Lie groups (Jaffe, 2024).

Robustness enters through transformed losses and through explicit robust estimators. In Hadamard spaces, transformed Fréchet means interpolate between median and classical mean; for power means $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$2, $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$3, finite-sample risk bounds attain the parametric $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$4 rate under fewer than two moments. For bounded positive derivative losses such as pseudo-Huber and $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$5, one obtains both parametric rates in expectation under arbitrarily weak polynomial moments and a breakdown point of $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$6 (Schötz, 10 Nov 2025). A complementary non-asymptotic construction is a trimmed Fréchet mean estimator

$$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$7

which is robust to $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$8-contamination up to breakdown at $$M(\mu) := \arg\min_{p \in S} F_2(p), \qquad V(\mu) := \inf_{p \in S} F_2(p).$$9. Its high-probability error bounds decompose into a confidence-independent global variance term and a confidence-dependent local variance term, recovering the optimal Euclidean structure in Hilbert spaces (Bartl et al., 17 Sep 2025).

On compact Riemannian manifolds, Varadhan functions furnish smooth approximations to Fréchet counterparts. Uniform convergence $m \in S$0 yields $m \in S$1 and $m \in S$2 as $m \in S$3. Uniform laws of large numbers hold uniformly over $m \in S$4 for Varadhan functions, variances, and means; for fixed $m \in S$5, there is a CLT for $m \in S$6 in $m \in S$7, a scalar CLT for $m \in S$8, and for fixed $m \in S$9, a tangent-space CLT for Varadhan means. Small-time asymptotics of gradients and Hessians recover the Fréchet mean CLT with the non-standard cut-locus correction term $(S,d)$00, but without assumptions on the geometry of the cut locus (Cao, 6 Jan 2026). This suggests a precise methodological distinction: smooth surrogates can retain Fréchet targets while bypassing the nonsmoothness of $(S,d)$01 across cut loci.

4. Inference beyond location: equivariance, ANOVA, regression, and dependence

Fréchet variance is not merely a descriptive quantity; it supports full inferential procedures. In general bounded metric spaces, under existence, uniqueness, separation, and entropy conditions, the sample Fréchet variance satisfies

$(S,d)$02

with a consistent plug-in estimator of $(S,d)$03. This leads to Fréchet analysis of variance for $(S,d)$04 populations, based on a pooled Fréchet mean, groupwise Fréchet means and variances, and a combined test statistic

$(S,d)$05

which converges under the null to $(S,d)$06 (Dubey et al., 2017).

On homogeneous Riemannian manifolds, symmetry leads to equivariant estimation. If $(S,d)$07 is a homogeneous Riemannian manifold with isometry group $(S,d)$08, the Fréchet mean functional is equivariant in the sense $(S,d)$09. Under transitive group action and factorization over orbits, the minimum risk equivariant estimator is

$(S,d)$10

or equivalently an integral over $(S,d)$11 when the orbit structure allows it. In the von Mises–Fisher model on $(S,d)$12, this yields $(S,d)$13; in the hyperbolic analogue, the MRE coincides with the normalized sum in the hyperboloid model. When the action is non-transitive, an adaptive equivariant estimator computes an MRE on an estimated orbit and remains equivariant (McCormack et al., 2021).

Fréchet regression extends the framework to Euclidean predictors and metric-space responses. A Bayesian version targets the Fréchet Bayes rule

$(S,d)$14

reducing the object-valued problem to scalar regression tasks for $(S,d)$15. Under a Gaussian working model in an RKHS, this yields closed-form posterior means and an explicit shrinkage path between prior-informed and frequentist objectives; under weak conditional expectations, the same formulas remain valid under moment conditions and uncorrelatedness, not a full likelihood model (Fontaine et al., 6 Jun 2026).

Conditioning can also be quantified directly through variance reduction. The Fréchet correlation coefficient is defined by

$(S,d)$16

where $(S,d)$17 is the conditional Fréchet mean and $(S,d)$18 the global one. It is directional, model-free, and unit-scale: it equals one under almost sure functional dependence and zero when conditioning leaves the Fréchet mean unchanged. A partition-based estimator avoids explicit nonparametric estimation of the full conditional Fréchet mean surface and admits consistency and null asymptotic distributions under fixed-partition and growing-partition regimes (He et al., 12 Apr 2026).

5. Computation and representative spaces

Practical computation depends strongly on geometry. In the general weak-convergence framework, sample Fréchet means can be obtained by minimizing $(S,d)$19 or the renormalized empirical functional

$(S,d)$20

In Banach or Hilbert settings one uses gradient, proximal, or stochastic algorithms; in CAT(0) spaces one exploits geodesic convexity; in Wasserstein spaces one uses barycenter solvers such as entropic regularization and iterative Bregman projections. When mean sets are nonunique, the recommended output is the mean set itself, or all minimizing solutions found, rather than a single arbitrary selection (Jaffe, 2024).

On Riemannian manifolds, differentiability through the Fréchet mean enables end-to-end optimization. For weighted points $(S,d)$21,

$(S,d)$22

and the first-order condition is $(S,d)$23. An implicit differentiation formula yields the Jacobian of the mean with respect to inputs, and in hyperbolic space this is accompanied by explicit log, exp, and parallel transport formulas, together with hyperparameter-free solvers in both the Poincaré ball and hyperboloid models. These constructions integrate Fréchet means and variances into hyperbolic neural network pipelines, including hyperbolic graph aggregation and hyperbolic batch normalization (Lou et al., 2020).

Several spaces display distinctive mean-set geometry. In tropical projective tori $(S,d)$24, the tropical Fréchet mean minimizes a sum of squared tropical distances. Means need not be unique; the set of all means is a polytrope, both tropically and classically convex, and equals the intersection of finitely many tropical balls. A positivity certificate for maxima of quadratic polynomials with sums-of-squares homogeneous parts supports an exact symbolic algorithm for the Fréchet mean polytrope; a reduced-gradient method supplies a numerical mean, after which the entire mean set and the tropical Fréchet variance can be recovered exactly (Lin et al., 7 Feb 2025).

In deformable models for curves and images, the relevant distance is deformation-invariant: $(S,d)$25 The Fréchet mean becomes an alignment-plus-averaging estimator. A two-step Procrustean procedure first estimates deformation parameters by minimizing a joint criterion and then averages the aligned signals. The decisive theoretical point is that consistency depends on the number of design points $(S,d)$26 and smoothing, not on increasing the number of curves $(S,d)$27 alone; lower bounds show that with $(S,d)$28 fixed, even ideal alignment-averaging cannot achieve consistency by increasing $(S,d)$29 alone (Bigot et al., 2010, Bigot et al., 2012).

In BHV treespace, a CAT(0) space of phylogenetic trees, the Fréchet mean is unique and geodesic convexity underpins iterative geodesic-averaging algorithms such as SturmMean. The Fréchet variance

$(S,d)$30

is faster and more precise than pairwise sum-of-squares for large samples, and permutation tests based on distances between group means provide mean hypothesis testing in treespace (Brown et al., 2017). In persistence-diagram spaces, deterministic Fréchet means may be nonunique and discontinuous as diagrams vary continuously, so a probabilistic Fréchet mean replaces a single diagram by a probability measure on mean diagrams induced by perturbation-based groupings; the resulting map is Hölder continuous on finite diagrams and is suited to vineyards (Munch et al., 2013).

For SPD matrices, metric choice is especially consequential. Under the affine-invariant metric, the Karcher mean solves

$(S,d)$31

under the log-Euclidean metric, the Fréchet mean is

$(S,d)$32

In high-dimensional low-sample regimes, random matrix theory can replace naive squared Fisher distances by a corrected estimator and define an RMT-improved Fréchet mean as the minimizer of the corrected objective over $(S,d)$33, with Riemannian gradient descent under the Fisher metric (Bouchard et al., 2024). Yet homogeneous complex Wishart matrices provide a counterpoint: under the affine-invariant metric, intrinsic bias and Riemannian risk calculations show that the simple arithmetic mean can be preferable overall to the sample Fréchet mean, because the Fréchet mean carries a curvature-induced intrinsic bias that does not diminish with the number of matrices averaged (Zhuang et al., 2017).

Graph-valued data also exhibit metric-sensitive behavior. For Erdős–Rényi $(S,d)$34, with Frobenius distance on graph Laplacians, the Fréchet mean set consists of quasi-regular graphs, and the squared distance to any Fréchet mean has closed-form mean and variance invariant to the choice of mean. Weak limits range from a Poisson-derived law to asymptotic normality, with a phase transition governed by the growth of $(S,d)$35 (Feng et al., 30 Mar 2026). This suggests that metric selection fundamentally shapes Fréchet mean geometry.

6. Limitations, comparisons, and open directions

A persistent misconception is that Fréchet means are automatically unique, continuous, or Euclidean-like. The current theory shows otherwise. Nonuniqueness is generic in tropical geometry, persistence-diagram spaces, and many quotient or positively curved settings; infinite-dimensional theory therefore treats mean sets rather than single means (Lin et al., 7 Feb 2025, Munch et al., 2013, Jaffe, 2024). Conversely, in CAT(0) or other strongly convex settings, uniqueness can be recovered, sometimes together with quantitative variance inequalities (Schötz, 2023).

Another misconception is that higher moments are always needed. The renormalized functional $(S,d)$36 establishes continuity, existence, compactness, SLLNs, and variance consistency under only $(S,d)$37-moment assumptions, and transformed Fréchet means achieve parametric finite-sample rates under fewer than two moments (Jaffe, 2024, Schötz, 10 Nov 2025). A different misconception is that the empirical Fréchet mean is uniformly satisfactory: under contamination or heavy tails it can be arbitrarily bad, while trimmed estimators and bounded-derivative transformed means retain non-asymptotic robustness and, in the latter case, breakdown point $(S,d)$38 (Bartl et al., 17 Sep 2025, Schötz, 10 Nov 2025).

The most important structural limitation is the absence of a universal CLT theory in abstract metric spaces. General weak-convergence theory currently covers consistency and LDPs but not CLTs, because tangent-space linearization, gradients, Hessians, and influence functions are unavailable without differentiable structure (Jaffe, 2024). Compact-manifold Varadhan theory and manifold-specific Fréchet mean analyses show one route around this obstacle, but only in geometries where heat-kernel or differential tools are available (Cao, 6 Jan 2026).

Open problems are explicit. Verifying the weak-convergence axioms in “super infinite-dimensional” Bures–Wasserstein spaces over infinite-dimensional Hilbert spaces remains challenging. Extending weak convergence to Banach manifolds modeled on uniformly convex spaces and to metrics defined via dynamical formulations such as LDDMM or Wasserstein–Fisher–Rao is promising but technically demanding. General CAT$(S,d)$39 spaces with $(S,d)$40 and bounded diameter may admit Jost-type weak convergence, but systematic verification remains open (Jaffe, 2024). For adaptive equivariant estimation, whether the adaptive MRE asymptotically dominates any consistent equivariant estimator from which its orbit is derived remains open (McCormack et al., 2021). In tropical geometry, predicting the dimension of the Fréchet mean polytrope from data, extending the convexity picture to $(S,d)$41, and analyzing $(S,d)$42-approximate means remain open directions (Lin et al., 7 Feb 2025).

Taken together, these developments establish Fréchet mean and variance methods as a geometry-aware statistical infrastructure rather than a single estimator. Their scope now includes set-valued continuity under weak convergence, strong laws, ergodic theorems, large deviations, smooth manifold surrogates, robust transformed losses, non-asymptotic contamination bounds, regression and dependence measures, and computational schemes tailored to the specific metric geometry of the data space (Jaffe, 2024, Fontaine et al., 6 Jun 2026, He et al., 12 Apr 2026).