betaFeedback

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 Gr(p,n)\mathrm{Gr}(p,n), 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 Gr(p,n)\mathrm{Gr}(p,n): the first is that the union of the cut loci of all points within distance δδ of the mean has volume O(δ)O(δ). The second is that the Hessian of the squared distance function is integrable. We also give a smooth density on Gr(1,2)\mathrm{Gr}(1,2) 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 Gr⁡(p,n)\operatorname{Gr}(p,n) of pp-dimensional subspaces of Rn\mathbb{R}^n, for every 1≤p<n1 \le p < n. 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 Gr⁡(p,n)\operatorname{Gr}(p,n) 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 pp and nn, 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 o(δ2)o(\delta^{2}). 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 Gr⁡(p,n)\operatorname{Gr}(p,n): (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:

  1. Volume near the cut locus. Lemma 3.4 identifies the relevant set with the δ\delta-tube around the cut locus, and Lemma 3.2 shows that its volume is O(δ)O(\delta).

  1. 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.

  1. Nondegeneracy. Proposition 2.2 gives a smooth, strictly positive density on Gr⁡(1,2)\operatorname{Gr}(1,2) with a unique mean but a zero Fréchet Hessian.

SETTING AND MAIN RESULT

Geometry of real Grassmannians. A point x∈Gr⁡(p,n)x \in\operatorname{Gr}(p,n) is a pp-dimensional subspace of Rn\mathbb{R}^{n}. Let Ux∈Rn×pU_x \in\mathbb{R}^{n \times p} have orthonormal columns spanning xx. The manifold Gr⁡(p,n)\operatorname{Gr}(p,n) is a compact, connected symmetric space of dimension m=p(n−p)m=p(n-p) with its canonical O(n)O(n)-invariant metric [12, 2, 4]. The orthogonal group O(n)={Q∈Rn×n:QTQ=In}O(n)=\{Q \in\mathbb{R}^{n \times n}:Q^{\mathsf{T}}Q=I_n\} acts transitively on Gr⁡(p,n)\operatorname{Gr}(p,n) by isometries. For x,y∈Gr⁡(p,n)x,y \in\operatorname{Gr}(p,n) the principal angles 0≤θ1(x,y)≤⋯≤θp(x,y)≤π/20 \le\theta_1(x,y) \le\cdots\le\theta_p(x,y) \le\pi/2 are defined by cos⁡θi(x,y)=σi(UxTUy)\cos\theta_i(x,y)=\sigma_i(U_x^{\mathsf{T}}U_y), where σ1≥⋯≥σp\sigma_1 \ge\cdots\ge\sigma_p are the singular values; they do not depend on the choice of bases [9]. If Px=UxUxTP_x=U_xU_x^{\mathsf{T}} and v=Uyav=U_y a with ∥a∥=1\lVert a\rVert=1, then ∥Pxv∥=∥UxTUya∥\lVert P_xv\rVert=\lVert U_x^{\mathsf{T}}U_y a\rVert. Minimizing over aa therefore gives

cos⁡θp(x,y)=σp(UxTUy)=min⁡v∈y∥v∥=1∥Pxv∥.(1)\cos\theta_p(x,y)=\sigma_p(U_x^{\mathsf{T}}U_y)=\min_{\substack{v \in y\\ \lVert v\rVert=1}}\lVert P_xv\rVert. \tag*{(1)}

The singular value decomposition also gives orthonormal bases u1,…,upu_1,\ldots,u_p of xx and v1,…,vpv_1,\ldots,v_p of yy such that

⟨vi,uj⟩=δijcos⁡θi(x,y)(1≤i,j≤p).(2)\langle v_i,u_j\rangle=\delta_{ij}\cos\theta_i(x,y)\qquad(1\le i,j\le p). \tag*{(2)}

The geodesic distance is [4 eq. (5.3)]

ρ(x,y)2=∑i=1pθi(x,y)2.(3)\rho(x,y)^2=\sum_{i=1}^{p}\theta_i(x,y)^2. \tag*{(3)}

Let exp⁡x:TxGr⁡(p,n)→Gr⁡(p,n)\exp_x:T_x\operatorname{Gr}(p,n)\to\operatorname{Gr}(p,n) denote the exponential map. For y∉Cxy\notin C_x there is a unique minimising geodesic from xx to yy [11 Ch. 13]; Log⁡x(y)∈TxGr⁡(p,n)\operatorname{Log}_x(y)\in T_x\operatorname{Gr}(p,n) denotes its initial velocity when parametrised on [0,1][0,1], so exp⁡x(Log⁡x(y))=y\exp_x(\operatorname{Log}_x(y))=y and ∥Log⁡x(y)∥=ρ(x,y)\lVert\operatorname{Log}_x(y)\rVert=\rho(x,y). The cut locus is [4 eq. (5.1)]

Cx={y∈Gr⁡(p,n):θp(x,y)=π/2}.(4)C_x=\{y\in\operatorname{Gr}(p,n):\theta_p(x,y)=\pi/2\}. \tag*{(4)}

The relation y∈Cxy \in C_x is symmetric in xx and yy. Let vol denote Riemannian volume and let λ=vol⁡/vol⁡(Gr⁡(p,n))\lambda= \operatorname{vol}/\operatorname{vol}(\operatorname{Gr}(p,n)) denote the normalised invariant measure. For δ>0\delta> 0, define

Bδ(x)={y:ρ(x,y)<δ},Bδ(x)={y:ρ(y,Cx)<δ},Aδ(x)=⋃y∈Bδ(x)Cy.(5)B_\delta(x)=\{y:\rho(x,y)<\delta\},\quad\mathcal{B}_\delta(x)=\{y:\rho(y,C_x)<\delta\},\quad A_\delta(x)=\bigcup_{y\in B_\delta(x)}C_y. \tag*{(5)}

Here Bδ(x)B_\delta(x) is a ball, Bδ(x)\mathcal{B}_\delta(x) is the δ\delta-tube around CxC_x, and Aδ(x)A_\delta(x) is the union of the cut loci based at points near xx.

Fréchet means. For a probability measure μ\mu on Gr⁡(p,n)\operatorname{Gr}(p,n), define the Fréchet function by

Fμ(x)=∫Gr⁡(p,n)ρ(x,ξ)2 dμ(ξ).(6)F_\mu(x)=\int_{\operatorname{Gr}(p,n)}\rho(x,\xi)^2\,\mathrm{d}\mu(\xi). \tag*{(6)}

When the minimiser is unique, write {x0}=arg min⁡x∈Gr⁡(p,n)Fμ(x)\{x_0\}=\operatorname*{arg\,min}_{x\in\operatorname{Gr}(p,n)}F_\mu(x). For i.i.d. observations ξ1,…,ξN∼μ\xi_1,\ldots,\xi_N\sim\mu, let x^N\hat{x}_N be any measurable selection from arg min⁡x∈Gr⁡(p,n)∑i=1Nρ(x,ξi)2\operatorname*{arg\,min}_{x\in\operatorname{Gr}(p,n)}\sum_{i=1}^N\rho(x,\xi_i)^2. If x0x_0 is unique, x^N→x0\hat{x}_N\to x_0 almost surely [25, 15]. Let Ψμ(x0):Tx0Gr⁡(p,n)→Tx0Gr⁡(p,n)\Psi_\mu(x_0):T_{x_0}\operatorname{Gr}(p,n)\to T_{x_0}\operatorname{Gr}(p,n) be the self-adjoint operator defined, for U,V∈Tx0Gr⁡(p,n)U,V\in T_{x_0}\operatorname{Gr}(p,n), by

Hess⁡x0Fμ(U,V)=2⟨Ψμ(x0)U,V⟩.\operatorname{Hess}_{x_0}F_\mu(U,V)=2\langle\Psi_\mu(x_0)U,V\rangle.

We now state our main theorem.

Theorem 2.1. Let μ\mu be a probability measure on Gr⁡(p,n)\operatorname{Gr}(p,n) with unique population Fréchet mean x0x_0. Assume that, for some δ0>0\delta_0>0, the restriction of μ\mu to Bδ0(x0)B_{\delta_0}(x_0) has a continuous density with respect to vol and Ψμ(x0)\Psi_\mu(x_0) is positive definite. Then, FμF_\mu is twice differentiable at x0x_0, and for every measurable choice x^N\hat{x}_N of the sample Fréchet mean satisfies

NLog⁡x0(x^N)→dNm(0,Ψμ(x0)−1V0Ψμ(x0)−1),\sqrt{N}\operatorname{Log}_{x_0}(\hat{x}_N)\xrightarrow{d}\mathcal{N}_m\left(0,\Psi_\mu(x_0)^{-1}V_0\Psi_\mu(x_0)^{-1}\right),

where V0=Cov⁡μ(Log⁡x0(ξ))V_0=\operatorname{Cov}_\mu(\operatorname{Log}_{x_0}(\xi)).

Positive definiteness cannot be dropped. Identify Gr⁡(1,2)\operatorname{Gr}(1,2) with R/πZ\mathbb{R}/\pi\mathbb{Z} by t↦span⁡{(cos⁡t,sin⁡t)}t\mapsto\operatorname{span}\{(\cos t,\sin t)\}. Under the canonical metric, Riemannian volume is dt\mathrm{d}t. The following density gives a counterexample.

Proposition 2.2. For 0<a<1/20<a<1/2, let μa\mu_a have density with respect to dt\mathrm{d}t

ψa(t)=1π(1+acos⁡2t+acos⁡4t),t∈R/πZ.(7)\psi_a(t)=\frac{1}{\pi}(1+a\cos2t+a\cos4t),\qquad t\in\mathbb{R}/\pi\mathbb{Z}. \tag*{(7)}

Then ψa\psi_a is analytic and strictly positive, t=0t=0 is the unique Fréchet mean of μa\mu_a, and every assumption of Theorem 2.1 other than positive definiteness holds, but

Fμa′′(0)=0,F_{\mu_a}''(0)=0,
Fμa(t)−Fμa(0)=at4+O(t6)(t→0),F_{\mu_a}(t)-F_{\mu_a}(0)=at^4+O(t^6)\qquad(t\to0),

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 Aδ(x0)A_\delta(x_0) with the δ\delta-tube around Cx0C_{x_0} and bounding its volume (see Lemmas 3.4 and 3.2). The integrability condition follows from an explicit spectral bound for H(x0∣ξ)H(x_0\mid\xi) (in Lemmas 3.11 and 3.3). Finally, we apply the general CLT (Lemma 3.1).

We first fix notation. Throughout this section, q=n−pq=n-p. For x,ξ∈Gr⁡(p,n)x,\xi\in\operatorname{Gr}(p,n), write θ1≤⋯≤θp\theta_1\le\cdots\le\theta_p for the principal angles θi(x,ξ)\theta_i(x,\xi) and set θ0:=0\theta_0:=0. For ξ∉Cx\xi\notin C_x, let fξ=12ρ(⋅,ξ)2f_\xi=\frac{1}{2}\rho(\cdot,\xi)^2, which is smooth near xx, with ∇fξ=−Log⁡⋅(ξ)\nabla f_\xi=-\operatorname{Log}_{\cdot}(\xi). Define the self-adjoint endomorphism H(x∣ξ)H(x\mid\xi) by

H(x∣ξ)[v]=∇v(Log⁡⋅(ξ))∣x,v∈TxGr⁡(p,n).(8)H(x\mid\xi)[v]=\nabla_v\left(\operatorname{Log}_{\cdot}(\xi)\right)\big|_x,\qquad v\in T_x\operatorname{Gr}(p,n). \tag*{(8)}

The identity ⟨H(x∣ξ)v,w⟩=−Hess⁡xfξ(v,w)\langle H(x\mid\xi)v,w\rangle=-\operatorname{Hess}_x f_\xi(v,w) identifies −H(x∣ξ)-H(x\mid\xi) with the bilinear Hessian of fξf_\xi. Therefore H(x∣ξ)H(x\mid\xi) 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 MM be an mm-dimensional Riemannian manifold and μ\mu a probability measure on MM. Define FμF_\mu by (2.6), let CxC_x be the Riemannian cut locus of xx, define Bδ(x)B_\delta(x), B˚δ(x)\mathring{B}_\delta(x), Aδ(x)A_\delta(x) by (2.5), and define H(x∣ξ)H(x\mid\xi) by (3.1). Assume

(H1) MM is compact and connected;

(H2) x0x_0 is the unique population Fréchet mean of μ\mu;

(H3) for some δ>0\delta>0, the restriction of μ\mu to Bδ(x0)B_\delta(x_0) has a continuous density ψ\psi with respect to vol⁡\operatorname{vol};

(H4) vol⁡(Aδ(x0))=O(δ)\operatorname{vol}(A_\delta(x_0))=O(\delta) as δ↓0\delta\downarrow0;

(H5) ξ↦H(x0∣ξ)\xi\mapsto H(x_0\mid\xi) is μ\mu-integrable.

Then FμF_\mu is twice differentiable at x0x_0. Let Ψμ(x0):Tx0M→Tx0M\Psi_\mu(x_0):T_{x_0}M\to T_{x_0}M be the self-adjoint operator defined, for U,V∈Tx0MU,V\in T_{x_0}M, by

Hess⁡Fμ(x0)(U,V)=2⟨Ψμ(x0)U,V⟩.\operatorname{Hess}F_\mu(x_0)(U,V)=2\langle\Psi_\mu(x_0)U,V\rangle.

If, in addition,

(H6) Ψμ(x0)\Psi_\mu(x_0) is symmetric positive definite,

then every measurable selection x^N\hat{x}_N of the sample Fréchet mean satisfies

NLog⁡x0(x^N)→dNm(0,Ψμ(x0)−1V0Ψμ(x0)−1),\sqrt{N}\operatorname{Log}_{x_0}(\hat{x}_N)\xrightarrow{d}\mathcal{N}_m\left(0,\Psi_\mu(x_0)^{-1}V_0\Psi_\mu(x_0)^{-1}\right),

where V0=Cov⁡μ(Log⁡x0(ξ1))V_0=\operatorname{Cov}_\mu(\operatorname{Log}_{x_0}(\xi_1)).

For Theorem 2.1, conditions H2, H3, and H6 are assumptions, while compactness and connectedness of Gr⁡(p,n)\operatorname{Gr}(p,n) 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 x0∈Gr⁡(p,n)x_0\in\operatorname{Gr}(p,n), vol⁡(Aδ(x0))=O(δ)\operatorname{vol}(A_\delta(x_0))=O(\delta) as δ↓0\delta\downarrow0. Moreover, vol⁡(Cx0)=0\operatorname{vol}(C_{x_0})=0. Thus H4 holds.

Lemma 3.3. If μ\mu satisfies H3, then

∫Gr⁡(p,n)∥H(x0∣ξ)∥op dμ(ξ)<∞.\int_{\operatorname{Gr}(p,n)}\lVert H(x_0\mid\xi)\rVert_{\mathrm{op}}\,d\mu(\xi)<\infty.

In particular, ∫Gr⁡(p,n)H(x0∣ξ) dμ(ξ)\int_{\operatorname{Gr}(p,n)}H(x_0\mid\xi)\,d\mu(\xi) 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 FμF_\mu and the central limit asserted in Theorem 2.1. □

The Volume Condition H4

We require a preparatory lemma.

Lemma 3.4. For all x,ξ∈Gr⁡(p,n)x,\xi\in\operatorname{Gr}(p,n) and δ>0\delta> 0,

ρ(ξ,Cx)=π2−θp(x,ξ)=ρ(x,Cξ).\rho(\xi,C_x)=\frac{\pi}{2}-\theta_p(x,\xi)=\rho(x,C_\xi).

Consequently, Aδ(x)=Bδ(x)A_\delta(x)=B_\delta(x).

Proof. For convenience of notation, we set θ=θp(x,ξ)\theta=\theta_p(x,\xi). We first show that

ρ(ξ,Cx)≥π2−θ.\rho(\xi,C_x)\ge\frac{\pi}{2}-\theta.

(1) gives cos⁡θ=min⁡∥Pxv∥\cos\theta=\min\|P_xv\| over unit vectors v∈ξv\in\xi. For w∈Cxw\in C_x, equation (2.4) gives a unit vector u∈w∩x⊥u\in w\cap x^\perp. For every unit v∈ξv\in\xi,

∣⟨u,v⟩∣=∣⟨u,Px⊥v⟩∣≤∥Px⊥v∥=(1−∥Pxv∥2)1/2≤sin⁡θ.|\langle u,v\rangle|=|\langle u,P_{x^\perp}v\rangle|\le\|P_{x^\perp}v\|=(1-\|P_xv\|^2)^{1/2}\le\sin\theta.

Since u∈wu\in w, equation (2.1), applied to the pair (ξ,w)(\xi,w), gives

cos⁡θp(w,ξ)≤∥Pξu∥=sup⁡v∈ξ∥v∥=1∣⟨u,v⟩∣≤sin⁡θ.\cos\theta_p(w,\xi)\le\|P_\xi u\|=\sup_{\substack{v\in\xi\\ \|v\|=1}}|\langle u,v\rangle|\le\sin\theta.

Thus θp(w,ξ)≥π/2−θ\theta_p(w,\xi)\ge\pi/2-\theta, and consequently by (2.3), ρ(ξ,w)≥θp(w,ξ)≥π/2−θ\rho(\xi,w)\ge\theta_p(w,\xi)\ge\pi/2-\theta. Notice that the RHS of ρ(ξ,w)≥π/2−θ\rho(\xi,w)\ge\pi/2-\theta is independent of ww, hence

ρ(ξ,Cx)=inf⁡w∈Cxρ(ξ,w)≥π2−θ.\rho(\xi,C_x)=\inf_{w\in C_x}\rho(\xi,w)\ge\frac{\pi}{2}-\theta.

We now show that

ρ(ξ,Cx)≤π2−θ.\rho(\xi,C_x)\le\frac{\pi}{2}-\theta.

By (2.2) there are orthonormal bases z1,…,zpz_1,\ldots,z_p of ξ\xi and x1,…,xpx_1,\ldots,x_p of xx with ⟨zi,xj⟩=δijcos⁡θi\langle z_i,x_j\rangle=\delta_{ij}\cos\theta_i; recall θ=θp\theta=\theta_p. We first construct a unit vector u∈x⊥u\in x^\perp with

zp=cos⁡θ xp+sin⁡θ uandu⊥z1,…,zp−1.(9)z_p=\cos\theta\,x_p+\sin\theta\,u\qquad\text{and}\qquad u\perp z_1,\ldots,z_{p-1}. \tag*{(9)}

Case θ>0\theta>0. Since x1,…,xpx_1,\ldots,x_p is an orthonormal basis of xx,

Pxzp=∑j=1p⟨zp,xj⟩xj=cos⁡θ xp,∥Px⊥zp∥2=1−∥Pxzp∥2=sin⁡2θ>0.P_xz_p=\sum_{j=1}^{p}\langle z_p,x_j\rangle x_j=\cos\theta\,x_p,\qquad\|P_{x^\perp}z_p\|^2=1-\|P_xz_p\|^2=\sin^2\theta>0.

Define u:=Px⊥zp/sin⁡θu:=P_{x^\perp}z_p/\sin\theta. Then u∈x⊥u\in x^\perp, ∥u∥=1\|u\|=1, and zp=Pxzp+Px⊥zp=cos⁡θ xp+sin⁡θ uz_p=P_xz_p+P_{x^\perp}z_p=\cos\theta\,x_p+\sin\theta\,u. For i<pi<p we have ⟨zi,zp⟩=0\langle z_i,z_p\rangle=0 and ⟨zi,xp⟩=0\langle z_i,x_p\rangle=0, hence

⟨u,zi⟩=⟨zp,zi⟩−cos⁡θ⟨xp,zi⟩sin⁡θ=0.\langle u,z_i\rangle=\frac{\langle z_p,z_i\rangle-\cos\theta\langle x_p,z_i\rangle}{\sin\theta}=0.

Case θ=0\theta=0. Then θ1=⋯=θp=0\theta_1=\cdots=\theta_p=0, so ⟨zi,xi⟩=1\langle z_i,x_i\rangle=1 and hence zi=xiz_i=x_i for all ii; thus ξ=x\xi=x. Since p<np<n, x⊥≠{0}x^\perp\ne\{0\}; let uu be any unit vector in x⊥x^\perp. Then (3.2) holds, the orthogonality because zi∈xz_i\in x and u∈x⊥u\in x^\perp.

Now set w:=span⁡{z1,…,zp−1,u}w:=\operatorname{span}\{z_1,\ldots,z_{p-1},u\}. By (3.2) these pp vectors are orthonormal, so w∈Gr⁡(p,n)w\in\operatorname{Gr}(p,n). Moreover 0≠u∈w∩x⊥0\ne u\in w\cap x^\perp, so w∈Cxw\in C_x by (2.4).

Using (3.2) and xp⊥ux_p\perp u, we get ⟨zp,u⟩=sin⁡θ\langle z_p,u\rangle=\sin\theta, ⟨zi,u⟩=0\langle z_i,u\rangle=0 for i<pi<p, and ⟨zi,zj⟩=δij\langle z_i,z_j\rangle=\delta_{ij} for j<pj<p. Hence the matrix of inner products between the orthonormal bases {z1,…,zp}\{z_1,\ldots,z_p\} of ξ\xi and {z1,…,zp−1,u}\{z_1,\ldots,z_{p-1},u\} of ww is diag⁡(1,…,1,sin⁡θ)\operatorname{diag}(1,\ldots,1,\sin\theta). Its singular values 1,…,1,sin⁡θ1,\ldots,1,\sin\theta are the cosines of the principal angles between ξ\xi and ww. Since sin⁡θ=cos⁡(π/2−θ)\sin\theta=\cos(\pi/2-\theta) with π/2−θ∈[0,π/2]\pi/2-\theta\in[0,\pi/2], these angles are 0,…,0,π/2−θ0,\ldots,0,\pi/2-\theta, and (3) gives ρ(ξ,w)=π/2−θ\rho(\xi,w)=\pi/2-\theta. As w∈Cxw\in C_x,

ρ(ξ,Cx)=inf⁡w′∈Cxρ(ξ,w′)≤ρ(ξ,w)=π2−θ.\rho(\xi,C_x)=\inf_{w'\in C_x}\rho(\xi,w')\leq\rho(\xi,w)=\frac{\pi}{2}-\theta.

Together with the lower bound, ρ(ξ,Cx)=π/2−θ\rho(\xi,C_x)=\pi/2-\theta, and the infimum is attained at ww.

Since UxTUξU_x^{\mathsf{T}}U_\xi and UξTUxU_\xi^{\mathsf{T}}U_x have the same singular values, θp(x,ξ)=θp(ξ,x)\theta_p(x,\xi)=\theta_p(\xi,x); applying the first equality with xx and ξ\xi interchanged gives ρ(x,Cξ)=π/2−θp(x,ξ)\rho(x,C_\xi)=\pi/2-\theta_p(x,\xi). Finally, observe that ξ∈Aδ(x)\xi\in A_\delta(x) if and only if ξ∈Cy\xi\in C_y for some y∈Bδ(x)y\in B_\delta(x). By the symmetry of cut-locus membership stated after (4), this means CξC_\xi meets Bδ(x)B_\delta(x), that is, ρ(x,Cξ)<δ\rho(x,C_\xi)<\delta. By the distance identity above, this is equivalent to ρ(ξ,Cx)<δ\rho(\xi,C_x)<\delta, that is, ξ∈Bδ(x)\xi\in B_\delta(x). Hence Aδ(x)=Bδ(x)A_\delta(x)=B_\delta(x).

□\square

Proof of Lemma 3.2. Put q=n−pq=n-p, so that Gr⁡(p,n)\operatorname{Gr}(p,n) has dimension m=pqm=pq. By (4) and (1), a subspace ww belongs to Cx0C_{x_0} precisely when w∩x0⊥w\cap x_0^\perp contains a nonzero vector. Equivalently, ww contains a line in x0⊥x_0^\perp.

We first use these lines to parametrise the cut locus. For a Euclidean subspace VV, write Gr⁡(k,V)\operatorname{Gr}(k,V) for the set of its kk-dimensional subspaces, with Gr⁡(0,V)\operatorname{Gr}(0,V) consisting of one point. Define

I={(ℓ,w):ℓ∈Gr⁡(1,x0⊥), w∈Gr⁡(p,n), ℓ⊂w}.I=\{(\ell,w):\ell\in\operatorname{Gr}(1,x_0^\perp),\ w\in\operatorname{Gr}(p,n),\ \ell\subset w\}.

For a fixed line ℓ\ell, every ww containing ℓ\ell has the unique orthogonal decomposition

w=ℓ⊕(w∩ℓ⊥).w=\ell\oplus(w\cap\ell^\perp).

Thus the possible choices of ww are parametrised by Gr⁡(p−1,ℓ⊥)\operatorname{Gr}(p-1,\ell^\perp), which is diffeomorphic to Gr⁡(p−1,n−1)\operatorname{Gr}(p-1,n-1) and has dimension (p−1)q(p-1)q.

Locally in ℓ\ell, Gram–Schmidt gives an orthonormal basis of ℓ⊥\ell^\perp that varies smoothly with ℓ\ell. These bases identify II locally with the product of an open subset of Gr⁡(1,x0⊥)\operatorname{Gr}(1,x_0^\perp) and Gr⁡(p−1,n−1)\operatorname{Gr}(p-1,n-1). Hence II is a smooth manifold of dimension

dim⁡I=(q−1)+(p−1)q=m−1.\dim I=(q-1)+(p-1)q=m-1.

The containment condition ℓ⊂w\ell\subset w is closed, since it is equivalent to PwPℓ=PℓP_wP_\ell=P_\ell. Therefore II is a closed subset of a compact product and is compact.

The projection

π:I⟶Gr⁡(p,n),π(ℓ,w)=w,\pi:I\longrightarrow\operatorname{Gr}(p,n),\qquad\pi(\ell,w)=w,

is smooth. Its image is exactly Cx0C_{x_0}, by the characterisation of the cut locus above.

We next show that Cx0C_{x_0} can be covered by at most Kr−(m−1)Kr^{-(m-1)} balls of radius rr, for some constant KK and every 0<r≤10<r\leq1. Write d=m−1d=m-1. If d=0d=0, the compact manifold II is finite, so the claim is immediate.

Suppose d≥1d\geq1. Compactness gives finitely many inverse coordinate maps

ψi:Q⟶I,Q=[0,1]d,1≤i≤N,\psi_i:Q\longrightarrow I,\qquad Q=[0,1]^d,\qquad1\leq i\leq N,

whose images cover II. We choose the coordinate cubes inside their chart domains, so that each ψi\psi_i is smooth on a neighbourhood of QQ. Set fi=π∘ψif_i=\pi\circ\psi_i. Their images cover Cx0C_{x_0}.

Since there are finitely many maps and QQ is compact, their derivatives have a common bound $L\geq 1. Applying this bound along the straight segment from ss to tt gives

ρ(fi(s),fi(t))≤L∣s−t∣,s,t∈Q.\rho(f_i(s),f_i(t))\leq L\lvert s-t\rvert,\qquad s,t\in Q.

Divide QQ into kdk^d equal cubes where k=⌈Ld/r⌉k = \left\lceil L\sqrt{d}/r\right\rceil and 0<r≤10 < r \le1. Every point in one of these cubes is within Euclidean distance d/(2k)\sqrt{d}/(2k) of its centre. Consequently, its image under fif_i is within distance

Ld2k≤r2\frac{L\sqrt{d}}{2k} \le\frac{r}{2}

of the image of that centre. Thus fi(Q)f_i(Q) is covered by kdk^d balls of radius rr. Moreover,

Nkd≤N(1+Ld)dr−d.Nk^d \le N(1+L\sqrt{d})^d r^{-d}.

This proves the covering claim. In either case, we can therefore choose centres w1,…,wNr∈Cx0w_1,\ldots,w_{N_r}\in C_{x_0} such that

Cx0⊆⋃j=1NrBr(wj),Nr≤Kr−(m−1).C_{x_0}\subseteq\bigcup_{j=1}^{N_r} B_r(w_j),\qquad N_r\le Kr^{-(m-1)}.

This covering also controls the tube around the cut locus. Indeed, if ξ∈Br(x0)\xi\in B_r(x_0), there is a point w∈Cx0w\in C_{x_0} with ρ(ξ,w)<r\rho(\xi,w)<r. For some jj, ρ(w,wj)<r\rho(w,w_j)<r, so the triangle inequality gives ρ(ξ,wj)<2r\rho(\xi,w_j)<2r. Hence

Br(x0)⊆⋃j=1NrB2r(wj).B_r(x_0)\subseteq\bigcup_{j=1}^{N_r}B_{2r}(w_j).

Finally, small geodesic balls have volume bounded by a constant times the mmth power of their radius

vol⁡(Bs(x0))≤2ωmsm,0<s<s0,\operatorname{vol}(B_s(x_0))\le2\omega_m s^m,\qquad0<s<s_0,

where ωm\omega_m is the Euclidean volume of the unit ball in Rm\mathbb{R}^m. Combining the ball-volume bound with the covering above, for 0<r<min⁡{1,s0/2}0<r<\min\{1,s_0/2\} we obtain

vol⁡(Br(x0))≤∑j=1Nrvol⁡(B2r(wj))≤Kr−(m−1)2ωm(2r)m=2m+1ωmKr.\operatorname{vol}(B_r(x_0))\le\sum_{j=1}^{N_r}\operatorname{vol}(B_{2r}(w_j))\le Kr^{-(m-1)}2\omega_m(2r)^m=2^{m+1}\omega_mKr.

Lemma 3.4 gives Aδ(x0)=Bδ(x0)A_\delta(x_0)=B_\delta(x_0), so vol⁡(Aδ(x0))=O(δ)\operatorname{vol}(A_\delta(x_0))=O(\delta) and H4 holds. Also, Cx0⊆Br(x0)C_{x_0}\subseteq B_r(x_0) for every r>0r>0. The same estimate therefore gives vol⁡(Cx0)≤2m+1ωmKr\operatorname{vol}(C_{x_0})\le2^{m+1}\omega_mKr for every sufficiently small rr. Letting r↓0r\downarrow0 proves vol⁡(Cx0)=0\operatorname{vol}(C_{x_0})=0.

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 ξ∉Cx0\xi\notin C_{x_0}, the singular values of Log⁡x0(ξ)\operatorname{Log}_{x_0}(\xi) are the principal angles between x0x_0 and ξ\xi. Lemma 3.9 uses this to compute H(x0∣ξ)H(x_0\mid\xi) explicitly via Jacobi fields. It gives a pointwise bound on ∥H(x0∣ξ)∥op\lVert H(x_0\mid\xi)\rVert_{\mathrm{op}} that blows up only when the two largest principal angles both tend to π/2\pi/2. Lemma 3.11 shows that this singularity is integrable under the invariant measure λ\lambda, and Lemma 3.3 transfers the conclusion to μ\mu.

The projector model. Throughout q=n−pq=n-p. We identify x∈Gr⁡(p,n)x\in\operatorname{Gr}(p,n) with the orthogonal projector Px=UxUx⊤P_x=U_xU_x^\top, which realises Gr⁡(p,n)\operatorname{Gr}(p,n) as the compact submanifold

P={P∈Sym⁡(n):P2=P,tr⁡P=p}\mathcal{P}=\{P\in\operatorname{Sym}(n):P^2=P,\operatorname{tr}P=p\}

of Sym⁡(n)\operatorname{Sym}(n), with tangent spaces TPP⊂Sym⁡(n)T_P\mathcal{P}\subset\operatorname{Sym}(n) [4]. The canonical metric is ⟨Z1,Z2⟩P=12tr⁡(Z1Z2)\langle Z_1,Z_2\rangle_P=\frac{1}{2}\operatorname{tr}(Z_1Z_2); it is invariant under P↦gPg⊤P\mapsto gPg^\top, g∈O(n)g\in O(n), and is the normalisation under which (3) holds [12, 4].

Fix x0x_0, and matrices U∈Rn×pU\in\mathbb{R}^{n\times p}, U⊥∈Rn×qU_\perp\in\mathbb{R}^{n\times q} with orthonormal columns spanning x0x_0 and x0⊥x_0^\perp; put P0=UU⊤P_0=UU^\top.

Lemma 3.5 (Tangent vectors; [4 2.12]). The map ι(B)=U⊥BU⊤+UB⊤U⊥⊤\iota(B)=U_\perp BU^\top+UB^\top U_\perp^\top is a linear isometry from (Rq×p,tr⁡(B⊤C))(\mathbb{R}^{q\times p},\operatorname{tr}(B^\top C)) onto (TP0P,⟨⋅,⋅⟩P0)(T_{P_0}\mathcal{P},\langle\cdot,\cdot\rangle_{P_0}), with inverse Z↦U⊥⊤ZUZ\mapsto U_\perp^\top ZU.

The identification ι\iota depends on U,U⊥U,U_\perp. Replacing them by UQ,U⊥RUQ,U_\perp R with Q∈O(p)Q\in O(p), R∈O(q)R\in O(q) replaces BB by R⊤BQR^\top BQ, which preserves tr⁡(B⊤C)\operatorname{tr}(B^\top C) and singular values; hence nothing below depends on this choice. Henceforth we identify BB with ι(B)\iota(B), and similarly at other base points.

Lemma 3.6 (Geodesics; [12 Thm. 2.3], [4]). Let B∈Rq×pB \in\mathbb{R}^{q\times p} have compact SVD B=U^ΣVTB = \widehat{U}\Sigma V^{\mathsf{T}}, where U^TU^=Ip\widehat{U}^{\mathsf{T}}\widehat{U} = I_p, V∈O(p)V \in O(p) and Σ=diag⁡(σi)\Sigma= \operatorname{diag}(\sigma_i), σi≥0\sigma_i \ge0. Set

Y(t)=UVcos⁡(tΣ)+U⊥U^sin⁡(tΣ),P(t)=Y(t)Y(t)T.Y(t) = UV\cos(t\Sigma) + U_{\perp}\widehat{U}\sin(t\Sigma), \qquad P(t) = Y(t)Y(t)^{\mathsf{T}}.

Then Y(t)Y(t) has orthonormal columns and PP is the geodesic with P(0)=P0P(0) = P_0 and P˙(0)=ι(B)\dot{P}(0) = \iota(B). In particular exp⁡x0(ι(B))\exp_{x_0}(\iota(B)) is spanned by Y(1)Y(1), and every geodesic from x0x_0 is of this form.

In the notation of [12], the tangent vector ι(B)\iota(B) corresponds to the horizontal vector H=U⊥BH = U_{\perp}B, since ddt(YYT)=HYT+YHT\frac{d}{dt}(YY^{\mathsf{T}}) = HY^{\mathsf{T}} + YH^{\mathsf{T}}; their formula differs from Y(t)Y(t) above by right multiplication by VTV^{\mathsf{T}}, which does not change the span.

Henceforth we identify BB with ι(B)\iota(B), and similarly at other base points.

Lemma 3.7 (Logarithm and principal angles). (a) If ξ∉Cx0\xi\notin C_{x_0}, the singular values of Log⁡x0(ξ)\operatorname{Log}_{x_0}(\xi) are exactly θ1(x0,ξ),…,θp(x0,ξ)\theta_1(x_0,\xi),\ldots,\theta_p(x_0,\xi), and they lie in [0,π/2)[0,\pi/2).

(b) The set Ω={(x,ξ)∈Gr⁡(p,n):ξ∉Cx}\Omega= \{(x,\xi) \in\operatorname{Gr}(p,n) : \xi\notin C_x\} is open, and (x,ξ)↦Log⁡x(ξ)(x,\xi) \mapsto\operatorname{Log}_x(\xi) is smooth on Ω\Omega.

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 (x,ξ)(x,\xi) wherever the overlap matrix is invertible, which is precisely the complement of the cut-locus relation.

Lemma 3.8 (Reduction to p≤qp \le q). Let τ(x)=x⊥\tau(x) = x^\perp.

(a) τ:Gr⁡(p,n)→Gr⁡(q,n)\tau: \operatorname{Gr}(p,n) \to\operatorname{Gr}(q,n) is an isometry. It maps λ\lambda to λ\lambda, minimising geodesics to minimising geodesics, and CxC_x onto CτxC_{\tau x}.

(b) For x,ξ∈Gr⁡(p,n)x,\xi\in\operatorname{Gr}(p,n), the nonzero principal angles of (x,ξ)(x,\xi) and of (τx,τξ)(\tau x,\tau\xi) coincide, with multiplicities.

(c) If ξ∉Cx\xi\notin C_x, then Log⁡τx(τξ)=dτxLog⁡x(ξ)\operatorname{Log}_{\tau x}(\tau\xi) = d\tau_x\operatorname{Log}_x(\xi) and H(τx∣τξ)=dτx∘H(x∣ξ)∘dτx−1H(\tau x \mid\tau\xi) = d\tau_x \circ H(x \mid\xi) \circ d\tau_x^{-1}; in particular the operator norms agree.

Proof. (a) In the projector model τ\tau is P↦I−PP \mapsto I-P, with differential Z↦−ZZ \mapsto-Z, which preserves 12tr⁡(Z1Z2)\frac{1}{2}\operatorname{tr}(Z_1Z_2). The remaining claims hold for any isometry [11 Chs. 2–3].

(b) The nonzero principal angles below π/2\pi/2 agree with those of the orthogonal complements [21], and the multiplicity of π/2\pi/2 is p−rank⁡(UxTUξ)=dim⁡(x⊥∩ξ)p-\operatorname{rank}(U_x^{\mathsf{T}}U_\xi)=\dim(x^\perp\cap\xi) for both pairs.

(c) By (a), τ\tau maps the minimising geodesic from xx to ξ\xi to the one from τx\tau x to τξ\tau\xi, so Log⁡τx(τξ)=dτxLog⁡x(ξ)\operatorname{Log}_{\tau x}(\tau\xi) = d\tau_x\operatorname{Log}_x(\xi). Isometries preserve the Levi-Civita connection, hence covariant derivatives along curves [11 Ch. 2]; so for a curve x(s)x(s) with x˙(0)=V\dot{x}(0) = V, (3.1) gives

H(τx∣τξ)[dτxV]=Ddsdτx(s)Log⁡x(s)(ξ)∣s=0=dτxH(x∣ξ)[V],H(\tau x \mid\tau\xi)[d\tau_xV] = \left.\frac{D}{ds}d\tau_{x(s)}\operatorname{Log}_{x(s)}(\xi)\right|_{s=0} = d\tau_xH(x \mid\xi)[V],

which is the stated conjugation. As dτxd\tau_x is a linear isometry, the operator norms agree.

By Lemma 3.8, every statement below for (p,n)(p,n) is equivalent to the corresponding one for (q,n)(q,n). We assume p≤qp \le q from Lemma 3.9 on.

Lemma 3.9. Assume p≤qp \le q. Let ξ∉Cx0\xi\notin C_{x_0}, write θi=θi(x0,ξ)\theta_i = \theta_i(x_0,\xi), and set θ0:=0\theta_0 := 0. Then H(x0∣ξ)H(x_0 \mid\xi) is diagonal in an orthonormal basis of Tx0Gr⁡(p,n)T_{x_0}\operatorname{Gr}(p,n), with eigenvalues −αcot⁡a-\alpha\cot a (read as −1-1 when a=0a=0), where aa runs over: 00, with multiplicity pp; θi+θj\theta_i+\theta_j and ∣θi−θj∣|\theta_i-\theta_j|, once each for every pair i<ji<j; and θj\theta_j, with multiplicity q−pq-p for each jj. All these values lie in [0,θp−1+θp]⊂[0,π)[0,\theta_{p-1}+\theta_p] \subset[0,\pi), and

∥H(x0∣ξ)∥op≤ππ−θp−1−θp.(10)\lVert H(x_0 \mid\xi)\rVert_{\mathrm{op}} \le\frac{\pi}{\pi-\theta_{p-1}-\theta_p}. \tag*{(10)}

Proof. We represent H(x0∣ξ)H(x_0 \mid\xi) 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]: Ddt\frac{D}{dt} is the covariant derivative along curves, J′=DdtJJ'=\frac{D}{dt}J, and R(X,Y)Z=∇Y∇XZ−∇X∇YZ+∇[X,Y]ZR(X,Y)Z=\nabla_Y\nabla_XZ-\nabla_X\nabla_YZ+\nabla_{[X,Y]}Z. The Jacobi operator Kγ˙(C)=R(γ˙,C)γ˙K_{\dot{\gamma}}(C)=R(\dot{\gamma},C)\dot{\gamma} along a geodesic γ\gamma is self-adjoint [11 Ch. 4], and Jacobi fields satisfy J′′+Kγ˙J=0J''+K_{\dot{\gamma}}J=0; variation fields of variations through geodesics are Jacobi fields [11 Ch. 5].

Step 1: Jacobi-field representation. Let B=Log⁡x0(ξ)B=\operatorname{Log}_{x_0}(\xi), so that γ(t)=exp⁡x0(tB)\gamma(t)=\exp_{x_0}(tB), t∈[0,1]t\in[0,1], is the minimising geodesic from x0x_0 to ξ\xi. Fix V∈Tx0Gr⁡(p,n)V\in T_{x_0}\operatorname{Gr}(p,n) and a curve x(s)x(s) with x(0)=x0x(0)=x_0, x˙(0)=V\dot{x}(0)=V. Since Ω\Omega is open, Lemma 3.7(b) makes u(s)=Log⁡x(s)(ξ)u(s)=\operatorname{Log}_{x(s)}(\xi) smooth for small ∣s∣|s|, so F(s,t)=exp⁡x(s)(tu(s))F(s,t)=\exp_{x(s)}(tu(s)) is a variation through geodesics and J(t)=∂sF(0,t)J(t)=\partial_sF(0,t) is a Jacobi field along γ\gamma. From F(s,0)=x(s)F(s,0)=x(s) and F(s,1)=ξF(s,1)=\xi we get J(0)=VJ(0)=V and J(1)=0J(1)=0; by the symmetry lemma [11 Ch. 2] and ∂tF(s,0)=u(s)\partial_tF(s,0)=u(s),

J′(0)=Ddsu(s)∣s=0=H(x0∣ξ)[V](11)J'(0)=\left.\frac{D}{ds}u(s)\right|_{s=0}=H(x_0\mid\xi)[V] \tag*{(11)}

by (8).

Step 2: the Jacobi operator at x0x_0. Every g∈O(n)g\in O(n) acts as an isometry, hence preserves minimising geodesics and covariant derivatives, so H(gx0∣gξ)=dgx0∘H(x0∣ξ)∘dgx0−1H(gx_0\mid g\xi)=dg_{x_0}\circ H(x_0\mid\xi)\circ dg_{x_0}^{-1}; principal angles are O(n)O(n)-invariant as well. Both sides of (10) are therefore unchanged under O(n)O(n), and we may take U=(Ip0)U=\begin{pmatrix}I_p\\0\end{pmatrix}, U⊥=(0Iq)U_\perp=\begin{pmatrix}0\\I_q\end{pmatrix}.

Write Gr⁡(p,n)=O(n)/(O(p)×O(q))\operatorname{Gr}(p,n)=O(n)/(O(p)\times O(q)) and o(n)=k⊕m\mathfrak{o}(n)=\mathfrak{k}\oplus\mathfrak{m}, with k\mathfrak{k} the block-diagonal and m\mathfrak{m} the off-diagonal skew-symmetric matrices. For B∈Rq×pB\in\mathbb{R}^{q\times p} put B^=(0−B⊤B0)∈m\widehat{B}=\begin{pmatrix}0&-B^\top\\B&0\end{pmatrix}\in\mathfrak{m}. Since ddt∣t=0etB^P0e−tB^=[B^,P0]=ι(B)\left.\frac{d}{dt}\right|_{t=0}e^{t\widehat{B}}P_0e^{-t\widehat{B}}=[\widehat{B},P_0]=\iota(B), the map B↦B^B\mapsto\widehat{B} identifies Tx0Gr⁡(p,n)T_{x_0}\operatorname{Gr}(p,n) with m\mathfrak{m}.

In the opposite sign convention of [17 Ch. IV] the curvature at x0x_0 is RH(X,Y)Z=−[[X^,Y^],Z^]R^{H}(X,Y)Z=-[[\widehat{X},\widehat{Y}],\widehat{Z}] (valid for every O(n)O(n)-invariant metric, as rescaling the metric by a constant changes neither the connection nor the (1,3)(1,3) curvature tensor). As R(γ˙,C)γ˙=RH(C,γ˙)γ˙R(\dot{\gamma},C)\dot{\gamma}=R^{H}(C,\dot{\gamma})\dot{\gamma}, the Jacobi operator agrees in both conventions, so at x0x_0 it is the lower-left block of −[[X^,B^],B^]-[[\widehat{X},\widehat{B}],\widehat{B}]. Since [X^,B^]=diag⁡(B⊤X−X⊤B, BX⊤−XB⊤)[\widehat{X},\widehat{B}]=\operatorname{diag}(B^\top X-X^\top B,\,BX^\top-XB^\top),

KB(X)=BB⊤X+XB⊤B−2BX⊤B,X∈Rq×p.(12)K_B(X)=BB^\top X+XB^\top B-2BX^\top B,\qquad X\in\mathbb{R}^{q\times p}. \tag*{(12)}

Step 3: spectrum of KBK_B. By Lemma 3.7(a), B=U^ΘV⊤B=\widehat{U}\Theta V^\top with Θ=diag⁡(θ1,…,θp)\Theta=\operatorname{diag}(\theta_1,\ldots,\theta_p). Extend the columns of U^\widehat{U} to an orthonormal basis u~1,…,u~q\widetilde{u}_1,\ldots,\widetilde{u}_q of Rq\mathbb{R}^q, let v1,…,vpv_1,\ldots,v_p be the columns of VV, and put Fkl=u~kvl⊤F_{kl}=\widetilde{u}_kv_l^\top. These form an orthonormal basis of Rq×p\mathbb{R}^{q\times p} for tr⁡(X⊤Y)\operatorname{tr}(X^\top Y), hence, by Lemma 3.5, an orthonormal basis of Tx0Gr⁡(p,n)T_{x_0}\operatorname{Gr}(p,n). Substituting BB⊤u~k=θk2u~kBB^\top\widetilde{u}_k=\theta_k^2\widetilde{u}_k (zero for k>pk>p), B⊤Bvl=θl2vlB^\top Bv_l=\theta_l^2v_l and BFkl⊤B=θkθlFlkBF_{kl}^\top B=\theta_k\theta_lF_{lk} (zero for k>pk>p) into (12) gives KBFkk=0K_BF_{kk}=0 for k≤pk\le p, KBFkl=θl2FklK_BF_{kl}=\theta_l^2F_{kl} for k>pk>p, and KB(Fkl±Flk)=(θk∓θl)2(Fkl±Flk)K_B(F_{kl}\pm F_{lk})=(\theta_k\mp\theta_l)^2(F_{kl}\pm F_{lk}) for k≠l≤pk\ne l\le p. Normalising Fkl±FlkF_{kl}\pm F_{lk} by 2−1/22^{-1/2} yields an orthonormal eigenbasis {e}\{e\} with KBe=ae2eK_Be=a_e^2e, the ae≥0a_e\ge0 being exactly the values listed in the lemma.

Step 4: solving the Jacobi equation. By (4) and Lemma 3.7(a) we have θi≤θp<π/2\theta_i\le\theta_p<\pi/2 and ∣θi−θj∣≤θp|\theta_i-\theta_j|\le\theta_p, so every listed aa satisfies a≤θp−1+θp<πa\le\theta_{p-1}+\theta_p<\pi. Fix ee and let E(t)E(t) be its parallel transport along γ\gamma. Since ∇R=0\nabla R=0 on a symmetric space [17 Ch. IV] and γ˙\dot{\gamma} is parallel, the field Kγ˙EK_{\dot{\gamma}}E is parallel; it equals ae2ea_e^2e at t=0t=0, so Kγ˙E=ae2EK_{\dot{\gamma}}E=a_e^2E throughout. Put g(t)=⟨J(t),E(t)⟩g(t)=\langle J(t),E(t)\rangle. By self-adjointness of Kγ˙K_{\dot{\gamma}},

g′′=⟨J′′,E⟩=−⟨J,Kγ˙E⟩=−ae2g,g''=\langle J'',E\rangle=-\langle J,K_{\dot{\gamma}}E\rangle=-a_e^2g,

with g(0)=⟨V,e⟩g(0)=\langle V,e\rangle and g(1)=0g(1)=0. As ae<πa_e<\pi, the unique solution is g(t)=g(0)sin⁡(ae(1−t))/sin⁡aeg(t)=g(0)\sin(a_e(1-t))/\sin a_e for ae>0a_e>0 and g(t)=g(0)(1−t)g(t)=g(0)(1-t) for ae=0a_e=0; in both cases g′(0)=−aecot⁡ae g(0)g'(0)=-a_e\cot a_e\,g(0). Since g′(0)=⟨J′(0),e⟩g'(0)=\langle J'(0),e\rangle, (11) gives ⟨H(x0∣ξ)V,e⟩=−aecot⁡ae⟨V,e⟩\langle H(x_0\mid\xi)V,e\rangle=-a_e\cot a_e\langle V,e\rangle for every VV and every ee, so H(x0∣ξ)H(x_0\mid\xi) is diagonal in {e}\{e\} with the stated eigenvalues.

Step 5: the bound. On [0,π/2][0,\pi/2] the function acot⁡aa\cot a decreases from 1 to 0, so ∣acot⁡a∣≤1|a\cot a|\le1 there. For a∈(π/2,π)a\in(\pi/2,\pi) put s=π−as=\pi-a; using tan⁡s≥s\tan s\ge s,

∣acot⁡a∣=acot⁡s≤as≤ππ−a≤ππ−θp−1−θp,|a\cot a|=a\cot s\le\frac{a}{s}\le\frac{\pi}{\pi-a}\le\frac{\pi}{\pi-\theta_{p-1}-\theta_p},

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 ∣aecot⁡ae∣|a_e\cot a_e|, which proves (3.3). □\square

Remark 3.10 (All pp). Bound (3.3) holds for all 1≤p<n1\le p<n. For p>qp>q, let θ1≤⋯≤θq′\theta_1\le\cdots\le\theta'_q be the principal angles of (x0⊥,ξ⊥)(x_0^\perp,\xi^\perp), with θ0′:=0\theta'_0:=0. By Lemma 3.8(b) the list θ1≤⋯≤θp\theta_1\le\cdots\le\theta_p consists of zeros followed by the nonzero θj′\theta'_j, so θp−1+θp=θq−1′+θq′\theta_{p-1}+\theta_p=\theta'_{q-1}+\theta'_q. Applying Lemma 3.9 on Gr⁡(q,n)\operatorname{Gr}(q,n) and Lemma 3.8(c) gives ∥H(x0∣ξ)∥op=∥H(x0⊥∣ξ⊥)∥op≤π/(π−θq−1′−θq′)=π/(π−θp−1−θp)\|H(x_0\mid\xi)\|_{\mathrm{op}}=\|H(x_0^\perp\mid\xi^\perp)\|_{\mathrm{op}}\le\pi/(\pi-\theta'_{q-1}-\theta'_q)=\pi/(\pi-\theta_{p-1}-\theta_p).

For p≥2p\ge2, the bound (3.3) can diverge only when both of the two largest principal angles approach π/2\pi/2. When p=1p=1, the bound is uniformly bounded by 2. By Lemma 3.2, vol⁡(Cx0)=0\operatorname{vol}(C_{x_0})=0, so H(x0∣⋅)H(x_0\mid\cdot) is defined λ\lambda-almost everywhere. We next prove that its operator norm is integrable with respect to the invariant probability measure.

Lemma 3.11. For every x0∈Gr⁡(p,n)x_0\in\operatorname{Gr}(p,n), ∫Gr⁡(p,n)∥H(x0∣ξ)∥op dλ(ξ)<∞\int_{\operatorname{Gr}(p,n)}\|H(x_0\mid\xi)\|_{\mathrm{op}}\,d\lambda(\xi)<\infty.

Proof. By Lemma 3.8(a),(c), τ\tau preserves λ\lambda and ∥H∥op\|H\|_{\mathrm{op}}, so the integral is unchanged under (p,n)↦(q,n)(p,n)\mapsto(q,n) and we may assume p≤qp\le q. If p=1p=1, then (3.3) with θ0=0\theta_0=0 gives ∥H(x0∣ξ)∥op≤π/(π−θ1)≤2\|H(x_0\mid\xi)\|_{\mathrm{op}}\le\pi/(\pi-\theta_1)\le2, and the integral is at most 2.

Let p≥2p\ge2. For ξ∼λ\xi\sim\lambda the ordered principal angles θ=(θ1,…,θp)\theta=(\theta_1,\ldots,\theta_p) have density

cp,n∏i=1p(sin⁡θi)n−2p∏i<j(cos⁡2θi−cos⁡2θj)(13)c_{p,n}\prod_{i=1}^{p}(\sin\theta_i)^{n-2p}\prod_{i<j}(\cos^2\theta_i-\cos^2\theta_j) \tag*{(13)}

on Δ={0≤θ1≤⋯≤θp≤π/2}\Delta=\{0\le\theta_1\le\cdots\le\theta_p\le\pi/2\} [1]: the cos⁡2θi\cos^2\theta_i are the eigenvalues of Ux0TPξUx0U_{x_0}^{\mathsf T}P_\xi U_{x_0}, which has the matrix beta distribution with parameters (p/2,(n−p)/2)(p/2,(n-p)/2), and (3.6) is the image of its eigenvalue density [23 Ch. 3] under λi=cos⁡2θi\lambda_i=\cos^2\theta_i.

Since n−2p≥0n-2p\ge0, every factor in (3.6) lies in [0,1][0,1], so dropping all but the (p−1,p)(p-1,p) factor and using (3.3) gives

∫Gr⁡(p,n)∥H(x0∣ξ)∥op dλ(ξ)≤cp,nπ∫Δcos⁡2θp−1−cos⁡2θpπ−θp−1−θp dθ.\int_{\operatorname{Gr}(p,n)}\|H(x_0\mid\xi)\|_{\mathrm{op}}\,d\lambda(\xi)\le c_{p,n}\pi\int_{\Delta}\frac{\cos^2\theta_{p-1}-\cos^2\theta_p}{\pi-\theta_{p-1}-\theta_p}\,d\theta.

Put a=π/2−θp−1≥b=π/2−θp≥0a=\pi/2-\theta_{p-1}\ge b=\pi/2-\theta_p\ge0, so that π−θp−1−θp=a+b\pi-\theta_{p-1}-\theta_p=a+b and cos⁡2θp−1−cos⁡2θp=sin⁡2a−sin⁡2b=sin⁡(a+b)sin⁡(a−b)\cos^2\theta_{p-1}-\cos^2\theta_p=\sin^2a-\sin^2b=\sin(a+b)\sin(a-b). Off the null set {a+b=0}\{a+b=0\}, the bound sin⁡(a+b)≤a+b\sin(a+b)\le a+b gives sin⁡(a+b)sin⁡(a−b)/(a+b)≤1\sin(a+b)\sin(a-b)/(a+b)\le1. Hence the integrand is bounded by 1 and the integral is at most cp,nπLeb⁡(Δ)<∞c_{p,n}\pi\operatorname{Leb}(\Delta)<\infty. □\square

Proof of Lemma 3.3. Since vol⁡(Cx0)=0\operatorname{vol}(C_{x_0})=0 and μ\mu has a density on Bδ0(x0)⊃Cx0B_{\delta_0}(x_0)\supset C_{x_0}, we have μ(Cx0)=0\mu(C_{x_0})=0, so H(x0∣⋅)H(x_0\mid\cdot) is defined μ\mu-a.e. and continuous off Cx0C_{x_0}. Fix 0<r<δ00<r<\delta_0. The closed tube Kr={ξ:ρ(ξ,Cx0)≤r}K_r=\{\xi:\rho(\xi,C_{x_0})\le r\} is a compact subset of Bδ0(x0)B_{\delta_0}(x_0), so the continuous density ψ\psi of H3H^3 is bounded there, say by ψˉ\bar\psi. Hence, by Lemma 3.11,

∫Br(x0)∥H(x0∣ξ)∥op dμ(ξ)≤ψˉvol⁡(Gr⁡(p,n))∫Gr⁡(p,n)∥H(x0∣ξ)∥op dλ(ξ)<∞.\int_{B_r(x_0)}\|H(x_0\mid\xi)\|_{\mathrm{op}}\,d\mu(\xi)\le\bar\psi\operatorname{vol}(\operatorname{Gr}(p,n))\int_{\operatorname{Gr}(p,n)}\|H(x_0\mid\xi)\|_{\mathrm{op}}\,d\lambda(\xi)<\infty.

If ξ∉Br(x0)\xi\notin B_r(x_0), Lemma 3.4 gives θp(x0,ξ)≤π/2−r\theta_p(x_0,\xi)\le\pi/2-r, so θp−1+θp≤π−2r\theta_{p-1}+\theta_p\le\pi-2r and (3.3) gives ∥H(x0∣ξ)∥op≤π/(2r)\|H(x_0\mid\xi)\|_{\mathrm{op}}\le\pi/(2r). Thus ∫∥H(x0∣ξ)∥op dμ<∞\int\|H(x_0\mid\xi)\|_{\mathrm{op}}\,d\mu<\infty, and the integral of the endomorphism-valued map H(x0∣⋅)H(x_0\mid\cdot) exists, which is H5. □\square

Proof. Since a<1/2a < 1/2, ψa≥(1−2a)/π>0\psi_a \ge(1 - 2a)/\pi> 0. The cosine terms integrate to zero over a period, so ∫0πψa dt=1\int_0^\pi\psi_a\,dt = 1. Write F=FμaF = F_{\mu_a} and ψ=ψa\psi= \psi_a.

Derivatives of FF. Substituting s=t+us = t + u and using π\pi-periodicity of ψ\psi,

F(t)=∫−π/2π/2u2ψ(t+u) du.F(t) = \int_{-\pi/2}^{\pi/2} u^2\psi(t+u)\,du.

Differentiating under the integral and integrating by parts, the boundary terms cancel by periodicity:

F′(t)=∫−π/2π/2u2ψ′(t+u) du=−2∫−π/2π/2uψ(t+u) du,F'(t) = \int_{-\pi/2}^{\pi/2} u^2\psi'(t+u)\,du = -2\int_{-\pi/2}^{\pi/2} u\psi(t+u)\,du,
F′′(t)=−2∫−π/2π/2uψ′(t+u) du=2(1−πψ(t+π/2)).F''(t) = -2\int_{-\pi/2}^{\pi/2} u\psi'(t+u)\,du = 2(1-\pi\psi(t+\pi/2)).

Since cos⁡(2t+π)=−cos⁡2t\cos(2t+\pi)=-\cos2t and cos⁡(4t+2π)=cos⁡4t\cos(4t+2\pi)=\cos4t,

F′′(t)=2a(cos⁡2t−cos⁡4t).(14)F''(t) = 2a(\cos2t-\cos4t). \tag*{(14)}

Because ψ\psi is even, F′(0)=0F'(0)=0. Integrating (14) from 0,

F′(t)=asin⁡2t−a2sin⁡4t=asin⁡2t(1−cos⁡2t).F'(t) = a\sin2t-\frac{a}{2}\sin4t = a\sin2t(1-\cos2t).

Unique mean. F′(t)>0F'(t)>0 for t∈(0,π/2)t\in(0,\pi/2) and F′(t)<0F'(t)<0 for t∈(−π/2,0)t\in(-\pi/2,0). Hence FF is strictly decreasing on [−π/2,0][-\pi/2,0] and strictly increasing on [0,π/2][0,\pi/2], so t=0t=0 is its unique global minimiser.

Degeneracy. By (14), F′′(0)=0F''(0)=0. Near 0, sin⁡2t=2t+O(t3)\sin2t=2t+O(t^3) and 1−cos⁡2t=2t2+O(t4)1-\cos2t=2t^2+O(t^4), so F′(t)=4at3+O(t5)F'(t)=4at^3+O(t^5) and F(t)−F(0)=at4+O(t6)F(t)-F(0)=at^4+O(t^6).

Hypotheses. H1 holds, H2 was just shown, and H3 holds because ψ\psi is continuous everywhere. H4 and H5 hold by Lemmas 3.2 and 3.3. Lemma 3.1 then gives Ψμa(0)=F′′(0)/2=0\Psi_{\mu_a}(0)=F''(0)/2=0.

Proof of Proposition 2.2

Conclusions

This paper establishes intrinsic central limit theorems of Gr⁡(p,n)\operatorname{Gr}(p,n) 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. [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. [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. [3]Imon Banerjee and Riddhiman Bhattacharya, Low dimensional sampling under reconstructed constraints, 2026.arxiv.org/abs/2609.14837
  4. [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. [5]Abhishek Bhattacharya and Rabi N. Bhattacharya, Nonparametric inference on manifolds with applications to shape spaces, Cambridge University Press, 2012.
  6. [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. [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. [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. [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. [10]Yasuko Chikuse, Statistics on special manifolds, Lecture Notes in Statistics, vol. 174, Springer, 2003.
  11. [11]Manfredo P. do Carmo, Riemannian geometry, Birkhäuser, Boston, 1992.
  12. [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. [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. [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. [15]Steven N. Evans and Adam Q. Jaffe, Limit theorems for Fréchet mean sets, Bernoulli 30 (2024), no. 1, 419–447.DOI
  16. [17]Sigurdur Helgason, Differential geometry, Lie groups, and symmetric spaces, Academic Press, New York, 1978.
  17. [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
  18. [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
  19. [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
  20. [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
  21. [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
  22. [23]Robb J. Muirhead, Aspects of multivariate statistical theory, Wiley, New York, 1982.
  23. [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

Paper details

vlnc:W_q8y2x99a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents