WorksMathematics
A Central Limit Theorem for Intrinsic Fréchet Means on Real Grassmannians
Abstract
Subspace-valued data arise in dimension reduction, image-set and video recognition, and signal processing, and the Fréchet mean is their standard sample mean [12,10,16,24]. We establish a central limit theorem for the intrinsic sample Fréchet mean of i.i.d. observations on the real Grassmann manifold , as a specialization of the general theorem of [19]. Unlike prior results which require the distribution to vanish or place negligible mass near the cut locus of the population mean, we allow a continuous, possibly positive density. To enable our results, we derive two properties of : the first is that the union of the cut loci of all points within distance of the mean has volume . The second is that the Hessian of the squared distance function is integrable. We also give a smooth density on for which the mean is unique but the Fréchet Hessian vanishes.
INTRODUCTION
This paper proves a central limit theorem (CLT) for the intrinsic sample Fréchet mean of i.i.d. observations on the real Grassmann manifold of -dimensional subspaces of , for every . Unlike earlier intrinsic CLTs, we allow the population distribution to have a continuous, possibly positive, density on the cut locus (as defined in (4)) of its mean parameter.
Our contribution is a CLT for intrinsic Fréchet means on every real Grassmannian that permits the population distribution to have a continuous, possibly positive density on the cut locus of its mean (although the cut locus cannot have positive probability; see the recent preprint by [22]). We obtain this result by verifying, for all and , the two geometric hypotheses in the general theorem of [19]. To the best of our knowledge, such a result was not previously available for Grassmannians.
Earlier CLTs for intrinsic and extrinsic Fréchet means were established by [8, 6] and developed further by [5]. Subsequent work treated non-identically distributed observations, general metric spaces, and slower (a.k.a. “smeary”) rates [20, 7, 14, 13]. The earlier intrinsic CLTs either exclude mass near the cut locus or require the mass swept out by nearby cut loci to be . For a distribution with bounded density, the latter condition requires the cut locus to have codimension greater than two [7], but the cut locus of a real Grassmannian has codimension one. Those results therefore exclude our setting. The circle exhibits non-standard limits when the distribution has density at the cut locus [18]; the obstruction changes the asymptotic law. For literature on Grassmannians, we refer to the excellent survey by [4] from which much of our preliminary facts are adopted (see also [10]).
Technical challenges. Our results require two geometric conditions for : (i) bounding the volume of a tube around the cut locus, and (ii) verifying the integrability of the distributions around the cut locus. We verify both and show that nondegeneracy must be assumed. These three make our main contributions and can be found as follows:
Volume near the cut locus. Lemma 3.4 identifies the relevant set with the -tube around the cut locus, and Lemma 3.2 shows that its volume is .
Integrability of the Hessian. Lemmas 3.11 and 3.3 prove the required integrability for distributions with a continuous density near the cut locus. Theorem 2.1 then gives the CLT.
Nondegeneracy. Proposition 2.2 gives a smooth, strictly positive density on with a unique mean but a zero Fréchet Hessian.
SETTING AND MAIN RESULT
Geometry of real Grassmannians. A point is a -dimensional subspace of . Let have orthonormal columns spanning . The manifold is a compact, connected symmetric space of dimension with its canonical -invariant metric [12, 2, 4]. The orthogonal group acts transitively on by isometries. For the principal angles are defined by , where are the singular values; they do not depend on the choice of bases [9]. If and with , then . Minimizing over therefore gives
The singular value decomposition also gives orthonormal bases of and of such that
The geodesic distance is [4 eq. (5.3)]
Let denote the exponential map. For there is a unique minimising geodesic from to [11 Ch. 13]; denotes its initial velocity when parametrised on , so and . The cut locus is [4 eq. (5.1)]
The relation is symmetric in and . Let vol denote Riemannian volume and let denote the normalised invariant measure. For , define
Here is a ball, is the -tube around , and is the union of the cut loci based at points near .
Fréchet means. For a probability measure on , define the Fréchet function by
When the minimiser is unique, write . For i.i.d. observations , let be any measurable selection from . If is unique, almost surely [25, 15]. Let be the self-adjoint operator defined, for , by
We now state our main theorem.
Theorem 2.1. Let be a probability measure on with unique population Fréchet mean . Assume that, for some , the restriction of to has a continuous density with respect to vol and is positive definite. Then, is twice differentiable at , and for every measurable choice of the sample Fréchet mean satisfies
where .
Positive definiteness cannot be dropped. Identify with by . Under the canonical metric, Riemannian volume is . The following density gives a counterexample.
Proposition 2.2. For , let have density with respect to
Then is analytic and strictly positive, is the unique Fréchet mean of , and every assumption of Theorem 2.1 other than positive definiteness holds, but
so $\Psi_{\mu_a}(0)=0.
PROOF OF THEOREM 2.1
The proof has three steps. First, we state the general CLT of [19] and reduce Theorem 2.1 to analysing a volume condition (H4) and an integrability condition (H5). The volume condition follows by identifying with the -tube around and bounding its volume (see Lemmas 3.4 and 3.2). The integrability condition follows from an explicit spectral bound for (in Lemmas 3.11 and 3.3). Finally, we apply the general CLT (Lemma 3.1).
We first fix notation. Throughout this section, . For , write for the principal angles and set . For , let , which is smooth near , with . Define the self-adjoint endomorphism by
The identity identifies with the bilinear Hessian of . Therefore is self-adjoint and has the same operator norm as the Hessian. We adapt Theorems 1 and 2 of [19] as the following combined lemma.
Lemma 3.1 ([19], Theorems 1–2 and eq. (15)). Let be an -dimensional Riemannian manifold and a probability measure on . Define by (2.6), let be the Riemannian cut locus of , define , , by (2.5), and define by (3.1). Assume
(H1) is compact and connected;
(H2) is the unique population Fréchet mean of ;
(H3) for some , the restriction of to has a continuous density with respect to ;
(H4) as ;
(H5) is -integrable.
Then is twice differentiable at . Let be the self-adjoint operator defined, for , by
If, in addition,
(H6) is symmetric positive definite,
then every measurable selection of the sample Fréchet mean satisfies
where .
For Theorem 2.1, conditions H2, H3, and H6 are assumptions, while compactness and connectedness of give H1. Proposition 2.2 shows that H6 does not follow from the other assumptions and cannot be dropped. The following two lemmas verify the remaining conditions. Lemma 3.2 verifies H4; its proof uses the cut-locus identity in Lemma 3.4. Lemma 3.3 verifies H5; its proof uses the spectral bound in Lemma 3.9 and the invariant-volume estimate in Lemma 3.11.
Lemma 3.2. For every , as . Moreover, . Thus H4 holds.
Lemma 3.3. If satisfies H3, then
In particular, exists and H5 holds.
Proof of Theorem 2.1
Compactness and connectedness give H1, while Theorem 2.1 assumes H2, H3, and H6. Lemma 3.2 gives H4, and Lemma 3.3 gives H5. Therefore all conditions of Lemma 3.1 hold, which gives the twice differentiability of and the central limit asserted in Theorem 2.1. □
The Volume Condition H4
We require a preparatory lemma.
Lemma 3.4. For all and ,
Consequently, .
Proof. For convenience of notation, we set . We first show that
(1) gives over unit vectors . For , equation (2.4) gives a unit vector . For every unit ,
Since , equation (2.1), applied to the pair , gives
Thus , and consequently by (2.3), . Notice that the RHS of is independent of , hence
We now show that
By (2.2) there are orthonormal bases of and of with ; recall . We first construct a unit vector with
Case . Since is an orthonormal basis of ,
Define . Then , , and . For we have and , hence
Case . Then , so and hence for all ; thus . Since , ; let be any unit vector in . Then (3.2) holds, the orthogonality because and .
Now set . By (3.2) these vectors are orthonormal, so . Moreover , so by (2.4).
Using (3.2) and , we get , for , and for . Hence the matrix of inner products between the orthonormal bases of and of is . Its singular values are the cosines of the principal angles between and . Since with , these angles are , and (3) gives . As ,
Together with the lower bound, , and the infimum is attained at .
Since and have the same singular values, ; applying the first equality with and interchanged gives . Finally, observe that if and only if for some . By the symmetry of cut-locus membership stated after (4), this means meets , that is, . By the distance identity above, this is equivalent to , that is, . Hence .
Proof of Lemma 3.2. Put , so that has dimension . By (4) and (1), a subspace belongs to precisely when contains a nonzero vector. Equivalently, contains a line in .
We first use these lines to parametrise the cut locus. For a Euclidean subspace , write for the set of its -dimensional subspaces, with consisting of one point. Define
For a fixed line , every containing has the unique orthogonal decomposition
Thus the possible choices of are parametrised by , which is diffeomorphic to and has dimension .
Locally in , Gram–Schmidt gives an orthonormal basis of that varies smoothly with . These bases identify locally with the product of an open subset of and . Hence is a smooth manifold of dimension
The containment condition is closed, since it is equivalent to . Therefore is a closed subset of a compact product and is compact.
The projection
is smooth. Its image is exactly , by the characterisation of the cut locus above.
We next show that can be covered by at most balls of radius , for some constant and every . Write . If , the compact manifold is finite, so the claim is immediate.
Suppose . Compactness gives finitely many inverse coordinate maps
whose images cover . We choose the coordinate cubes inside their chart domains, so that each is smooth on a neighbourhood of . Set . Their images cover .
Since there are finitely many maps and is compact, their derivatives have a common bound $L\geq 1. Applying this bound along the straight segment from to gives
Divide into equal cubes where and . Every point in one of these cubes is within Euclidean distance of its centre. Consequently, its image under is within distance
of the image of that centre. Thus is covered by balls of radius . Moreover,
This proves the covering claim. In either case, we can therefore choose centres such that
This covering also controls the tube around the cut locus. Indeed, if , there is a point with . For some , , so the triangle inequality gives . Hence
Finally, small geodesic balls have volume bounded by a constant times the th power of their radius
where is the Euclidean volume of the unit ball in . Combining the ball-volume bound with the covering above, for we obtain
Lemma 3.4 gives , so and H4 holds. Also, for every . The same estimate therefore gives for every sufficiently small . Letting proves .
The Integrability Condition (H5)
We prove H5 in three steps. First we recall standard facts about the Grassmannian (Lemmas 3.5–3.8); the one we need is that for , the singular values of are the principal angles between and . Lemma 3.9 uses this to compute explicitly via Jacobi fields. It gives a pointwise bound on that blows up only when the two largest principal angles both tend to . Lemma 3.11 shows that this singularity is integrable under the invariant measure , and Lemma 3.3 transfers the conclusion to .
The projector model. Throughout . We identify with the orthogonal projector , which realises as the compact submanifold
of , with tangent spaces [4]. The canonical metric is ; it is invariant under , , and is the normalisation under which (3) holds [12, 4].
Fix , and matrices , with orthonormal columns spanning and ; put .
Lemma 3.5 (Tangent vectors; [4 2.12]). The map is a linear isometry from onto , with inverse .
The identification depends on . Replacing them by with , replaces by , which preserves and singular values; hence nothing below depends on this choice. Henceforth we identify with , and similarly at other base points.
Lemma 3.6 (Geodesics; [12 Thm. 2.3], [4]). Let have compact SVD , where , and , . Set
Then has orthonormal columns and is the geodesic with and . In particular is spanned by , and every geodesic from is of this form.
In the notation of [12], the tangent vector corresponds to the horizontal vector , since ; their formula differs from above by right multiplication by , which does not change the span.
Henceforth we identify with , and similarly at other base points.
Lemma 3.7 (Logarithm and principal angles). (a) If , the singular values of are exactly , and they lie in .
(b) The set is open, and is smooth on .
Proof. (a) The SVD formula for the Grassmannian logarithm shows that its singular values are the principal angles; see [4 Sec. 5.2, Thm. 5.4]. The strict upper bound follows from the characterization of the cut locus in [4 Sec. 5.1].
(b) The result follows from the smooth dependence of the Grassmannian logarithm formula on the pair wherever the overlap matrix is invertible, which is precisely the complement of the cut-locus relation.
Lemma 3.8 (Reduction to ). Let .
(a) is an isometry. It maps to , minimising geodesics to minimising geodesics, and onto .
(b) For , the nonzero principal angles of and of coincide, with multiplicities.
(c) If , then and ; in particular the operator norms agree.
Proof. (a) In the projector model is , with differential , which preserves . The remaining claims hold for any isometry [11 Chs. 2–3].
(b) The nonzero principal angles below agree with those of the orthogonal complements [21], and the multiplicity of is for both pairs.
(c) By (a), maps the minimising geodesic from to to the one from to , so . Isometries preserve the Levi-Civita connection, hence covariant derivatives along curves [11 Ch. 2]; so for a curve with , (3.1) gives
which is the stated conjugation. As is a linear isometry, the operator norms agree.
By Lemma 3.8, every statement below for is equivalent to the corresponding one for . We assume from Lemma 3.9 on.
Lemma 3.9. Assume . Let , write , and set . Then is diagonal in an orthonormal basis of , with eigenvalues (read as when ), where runs over: , with multiplicity ; and , once each for every pair ; and , with multiplicity for each . All these values lie in , and
Proof. We represent by a Jacobi field. On the Grassmannian the Jacobi equation has constant coefficients in a suitable frame, with coefficients explicit in the principal angles, which reduces the lemma to scalar ODEs.
Conventions. We follow [11]: is the covariant derivative along curves, , and . The Jacobi operator along a geodesic is self-adjoint [11 Ch. 4], and Jacobi fields satisfy ; variation fields of variations through geodesics are Jacobi fields [11 Ch. 5].
Step 1: Jacobi-field representation. Let , so that , , is the minimising geodesic from to . Fix and a curve with , . Since is open, Lemma 3.7(b) makes smooth for small , so is a variation through geodesics and is a Jacobi field along . From and we get and ; by the symmetry lemma [11 Ch. 2] and ,
by (8).
Step 2: the Jacobi operator at . Every acts as an isometry, hence preserves minimising geodesics and covariant derivatives, so ; principal angles are -invariant as well. Both sides of (10) are therefore unchanged under , and we may take , .
Write and , with the block-diagonal and the off-diagonal skew-symmetric matrices. For put . Since , the map identifies with .
In the opposite sign convention of [17 Ch. IV] the curvature at is (valid for every -invariant metric, as rescaling the metric by a constant changes neither the connection nor the curvature tensor). As , the Jacobi operator agrees in both conventions, so at it is the lower-left block of . Since ,
Step 3: spectrum of . By Lemma 3.7(a), with . Extend the columns of to an orthonormal basis of , let be the columns of , and put . These form an orthonormal basis of for , hence, by Lemma 3.5, an orthonormal basis of . Substituting (zero for ), and (zero for ) into (12) gives for , for , and for . Normalising by yields an orthonormal eigenbasis with , the being exactly the values listed in the lemma.
Step 4: solving the Jacobi equation. By (4) and Lemma 3.7(a) we have and , so every listed satisfies . Fix and let be its parallel transport along . Since on a symmetric space [17 Ch. IV] and is parallel, the field is parallel; it equals at , so throughout. Put . By self-adjointness of ,
with and . As , the unique solution is for and for ; in both cases . Since , (11) gives for every and every , so is diagonal in with the stated eigenvalues.
Step 5: the bound. On the function decreases from 1 to 0, so there. For put ; using ,
and the right-hand side is at least 1, so it bounds both cases. The operator norm of an operator diagonal in an orthonormal basis is its largest , which proves (3.3).
Remark 3.10 (All ). Bound (3.3) holds for all . For , let be the principal angles of , with . By Lemma 3.8(b) the list consists of zeros followed by the nonzero , so . Applying Lemma 3.9 on and Lemma 3.8(c) gives .
For , the bound (3.3) can diverge only when both of the two largest principal angles approach . When , the bound is uniformly bounded by 2. By Lemma 3.2, , so is defined -almost everywhere. We next prove that its operator norm is integrable with respect to the invariant probability measure.
Lemma 3.11. For every , .
Proof. By Lemma 3.8(a),(c), preserves and , so the integral is unchanged under and we may assume . If , then (3.3) with gives , and the integral is at most 2.
Let . For the ordered principal angles have density
on [1]: the are the eigenvalues of , which has the matrix beta distribution with parameters , and (3.6) is the image of its eigenvalue density [23 Ch. 3] under .
Since , every factor in (3.6) lies in , so dropping all but the factor and using (3.3) gives
Put , so that and . Off the null set , the bound gives . Hence the integrand is bounded by 1 and the integral is at most .
Proof of Lemma 3.3. Since and has a density on , we have , so is defined -a.e. and continuous off . Fix . The closed tube is a compact subset of , so the continuous density of is bounded there, say by . Hence, by Lemma 3.11,
If , Lemma 3.4 gives , so and (3.3) gives . Thus , and the integral of the endomorphism-valued map exists, which is H5.
Proof. Since , . The cosine terms integrate to zero over a period, so . Write and .
Derivatives of . Substituting and using -periodicity of ,
Differentiating under the integral and integrating by parts, the boundary terms cancel by periodicity:
Since and ,
Because is even, . Integrating (14) from 0,
Unique mean. for and for . Hence is strictly decreasing on and strictly increasing on , so is its unique global minimiser.
Degeneracy. By (14), . Near 0, and , so and .
Hypotheses. H1 holds, H2 was just shown, and H3 holds because is continuous everywhere. H4 and H5 hold by Lemmas 3.2 and 3.3. Lemma 3.1 then gives .
Proof of Proposition 2.2
Conclusions
This paper establishes intrinsic central limit theorems of valued random variables. This requires us to calculate the geodesic distance of the points on the cut locus (Lemma 3.4), verifying the integrability condition (Lemma 3.3), and bounding the operator norm of the hessian tensor through eigenvalues (Lemma 3.9). As future work, it remains to be seen whether intrinsic CLT’s can be derived for more complicated manifolds like Hadamard and Stiefel manifolds. Such CLT’s are also relevant for postsampling inference from manifolds [3].
References
References
- [1]P.-A. Absil, Alan Edelman, and Plamen Koev, On the largest principal angle between random subspaces, Linear Algebra and its Applications 414 (2006), 288–294.DOI
- [2]P.-A. Absil, Robert Mahony, and Rodolphe Sepulchre, Riemannian geometry of Grassmann manifolds with a view on algorithmic computation, Acta Applicandae Mathematicae 80 (2004), 199–220.DOI
- [3]Imon Banerjee and Riddhiman Bhattacharya, Low dimensional sampling under reconstructed constraints, 2026.arxiv.org/abs/2609.14837
- [4]Thomas Bendokat, Ralf Zimmermann, and P.-A. Absil, A Grassmann manifold handbook: basic geometry and computational aspects, Advances in Computational Mathematics 50 (2024), no. 1, 6.DOI
- [5]Abhishek Bhattacharya and Rabi N. Bhattacharya, Nonparametric inference on manifolds with applications to shape spaces, Cambridge University Press, 2012.
- [6]Rabi Bhattacharya and Vic Patrangenaru, Large sample theory of intrinsic and extrinsic sample means on manifolds—ii, The Annals of Statistics 33 (2005), no. 3, 1225–1259.DOI
- [7]Rabi N. Bhattacharya and Lizhen Lin, Omnibus CLTs for Fréchet means and nonparametric inference on non-Euclidean spaces, Proceedings of the American Mathematical Society 145 (2017), 413–428.arxiv.org/abs/1306.5806
- [8]Rabi Bhattacharya and Vic Patrangenaru, Large sample theory of intrinsic and extrinsic sample means on manifolds I, The Annals of Statistics 31 (2003), 1–29.DOI
- [9]Åke Björck and Gene H. Golub, Numerical methods for computing angles between linear subspaces, Mathematics of Computation 27 (1973), no. 123, 579–594.DOI
- [10]Yasuko Chikuse, Statistics on special manifolds, Lecture Notes in Statistics, vol. 174, Springer, 2003.
- [11]Manfredo P. do Carmo, Riemannian geometry, Birkhäuser, Boston, 1992.
- [12]Alan Edelman, Tomás A. Arias, and Steven T. Smith, The geometry of algorithms with orthogonality constraints, SIAM Journal on Matrix Analysis and Applications 20 (1998), no. 2, 303–353.arxiv.org/abs/physics/9806030
- [13]Benjamin Eltzner, Fernando Galaz-García, Stephan F. Huckemann, and Wilderich Tuschmann, Stability of the cut locus and a central limit theorem for Fréchet means of Riemannian manifolds, Proceedings of the American Mathematical Society 149 (2021), 3947–3963.arxiv.org/abs/1909.00410
- [14]Benjamin Eltzner and Stephan F. Huckemann, A smeary central limit theorem for manifolds with application to high-dimensional spheres, The Annals of Statistics 47 (2019), 3360–3381.arxiv.org/abs/1801.06581
- [15]Steven N. Evans and Adam Q. Jaffe, Limit theorems for Fréchet mean sets, Bernoulli 30 (2024), no. 1, 419–447.DOI
- [17]Sigurdur Helgason, Differential geometry, Lie groups, and symmetric spaces, Academic Press, New York, 1978.
- [18]Thomas Hotz and Stephan Huckemann, Intrinsic means on the circle: uniqueness, locus and asymptotics, Annals of the Institute of Statistical Mathematics 67 (2015), 177–193.arxiv.org/abs/1108.2141
- [19]Thomas Hotz, Huiling Le, and Andrew T. A. Wood, Central limit theorem for intrinsic Fréchet means in smooth compact Riemannian manifolds, Probability Theory and Related Fields 189 (2024), no. 3–4, 1219–1246.arxiv.org/abs/2210.17191
- [20]Wilfrid S. Kendall and Huiling Le, Limit theorems for empirical Fréchet means of independent and non-identically distributed manifold-valued random variables, Brazilian Journal of Probability and Statistics 25 (2011), 323–352.arxiv.org/abs/1102.0228
- [21]Andrew V. Knyazev and Merico E. Argentati, Principal angles between subspaces in an a-based scalar product: Algorithms and perturbation estimates, SIAM Journal on Scientific Computing 23 (2002), no. 6, 2008–2040.DOI
- [22]Alexander Lytchak and Stephan F. Huckemann, Zero probability of the cut locus of a Fréchet mean on a Riemannian manifold, 2025, arXiv:2508.00747.arxiv.org/abs/2508.00747
- [23]Robb J. Muirhead, Aspects of multivariate statistical theory, Wiley, New York, 1982.
- [25]Herbert Ziezold, On expected figures and a strong law of large numbers for random elements in quasi-metric spaces, Transactions of the Seventh Prague Conference on Information Theory, Statistical Decision Functions, Random Processes, vol. A, Reidel, 1977, pp. 591–602.DOI