betaFeedback

WorksMathematics

No closed-form formula for Stiefel geodesics

Abstract

We use differential Galois theory to prove that given two arbitrary points XX and YY on the Stiefel manifold, there is no closed-form analytic expression for the Riemannian logarithm log⁡XY\log_X Y, the geodesic between XX and YY, or the geodesic distance between XX and YY. In other words, these cannot be expressed in terms of elementary functions of the coordinates of XX and YY; in fact, our result goes further — even if we allow for repeated integrals of elementary functions, there is still no such expression.

Introduction

The Stiefel manifold [28] is the commonest manifold appearing in Riemannian optimization [6], not least because of its newfound use in the manifold Muon optimizer [3, 27], a manifold variant of the popular Muon optimizer [12], for training deep neural networks and transformers. There has been a long and continued search for a closed-form formula to express a geodesic with given initial point XX and endpoint YY on the Stiefel manifold, or, equivalently, a closed-form formula for the Riemannian logarithm log⁡XY\log_X Y. The effort has persisted over at least the last 13 years [21, 32, 4, 14, 2, 29, 25, 19, 18, 11] (references too numerous to be exhaustive), likely making it the most sought after open problem in manifold optimization. In this article, we show that no such formula exists.

For positive integers k≤nk \le n, the Stiefel manifold

V(k,n)={X∈Rn×k:XTX=Ik}(1)V(k,n) = \{X \in\mathbb{R}^{n \times k} : X^{\mathsf T}X = I_k\} \tag*{(1)}

is a Riemannian manifold equipped with the natural O(n)×O(k)O(n) \times O(k)-invariant metric

gX(H1,H2)=tr⁡(H1T(In−12XXT)H2).(2)g_X(H_1,H_2) = \operatorname{tr}\left(H_1^{\mathsf T}\left(I_n - \frac{1}{2}XX^{\mathsf T}\right)H_2\right). \tag*{(2)}

For a geodesic on V(k,n)V(k,n), one may ask for two different kinds of closed-form formulas with different input data:

Problem 1. Given a point X∈V(k,n)X \in V(k,n) and a tangent vector H∈TXV(k,n)H \in T_XV(k,n), find a closed-form formula for the geodesic γ:[0,1]→V(k,n)\gamma: [0,1] \to V(k,n) with γ(0)=X\gamma(0) = X and γ′(0)=H\gamma'(0) = H.

Problem 2. Given two points XX and Y∈V(k,n)Y \in V(k,n), find a closed-form formula for the geodesic γ:[0,1]→V(k,n)\gamma: [0,1] \to V(k,n) with γ(0)=X\gamma(0) = X and γ(1)=Y\gamma(1) = Y.

For the uninitiated, the term “closed-form formula” has a precise, universally accepted definition in mathematics [22]. It means an expression in terms of arithmetic operations over the complex field, arbitrary algebraic functions (which include surds), scalar exponentials, and scalar logarithms; other functions, including trigonometric and hyperbolic functions and their inverses, may be obtained from these. A function admitting such an expression is called an elementary function [23], i.e., a closed-form formula is an elementary function.

What is perhaps not universally known is that standard matrix functions, including the matrix exponential, logarithm, sign, pth root, sine, and cosine [9], are all elementary functions. In fact, this observation does not seem to have been stated anywhere in existing literature so we will explain it here: The reason is that for any X∈Cn×nX \in\mathbb{C}^{n \times n} and complex-valued univariate function φ\varphi analytic near the spectrum of XX, the Lagrange–Hermite formula [9 Equation 1.8] yields

φ(X)=∑i=1r(∏j=1j≠ir(X−λjI)mj)⋅∑k=0mi−1(X−λiI)kk!dkdzk(φ(z)∏j=1j≠ir(z−λj)mj)∣z=λi,\varphi(X) = \sum_{i=1}^{r} \left(\prod_{\substack{j=1 \\ j \ne i}}^{r}(X-\lambda_j I)^{m_j}\right) \cdot\sum_{k=0}^{m_i-1} \frac{(X-\lambda_i I)^k}{k!} \frac{d^k}{dz^k}\left(\frac{\varphi(z)}{\prod_{\substack{j=1 \\ j \ne i}}^{r}(z-\lambda_j)^{m_j}}\right)\bigg|_{z=\lambda_i},

where λ1,…,λr\lambda_1,\ldots,\lambda_r are the distinct eigenvalues of XX and mim_i is the size of the largest λi\lambda_i-Jordan block, i=1,…,ri=1,\ldots,r. Since λi\lambda_i is an algebraic function in the entries of XX and derivatives of elementary scalar functions are elementary, the matrix function φ(X)\varphi(X) is elementary whenever the scalar function φ(x)\varphi(x) is elementary.

Problem 1 has a well-known closed-form formula given by

γ(t)=exp⁡(t(HXT−XHT−X(XTH)XT))X(3)\gamma(t) = \exp\left(t(HX^{\mathsf{T}}-XH^{\mathsf{T}}-X(X^{\mathsf{T}}H)X^{\mathsf{T}})\right)X \tag*{(3)}

where exp⁡\exp is the matrix exponential function. The main result of this article is that Problem 2 has no solution in general, allowing us to deduce that there is also no closed-form formula for the Riemannian logarithm log⁡XY\log_X Y or the geodesic distance between XX and YY. Indeed, we will prove a stronger result showing that the Riemannian logarithm, the geodesic, and the geodesic distance, cannot be expressed as Liouvillian functions, which allows for the operation of taking antiderivatives in addition to those for elementary functions. For example, the error function or Fresnel integrals are Liouvillian but not elementary.

Theorem 1.1. Let n≥4n \ge4 and 2≤k≤n−22 \le k \le n-2. For arbitrary X,Y∈V(k,n)X,Y \in\mathrm{V}(k,n) and t∈[0,1]t \in[0,1], there is no closed-form formula expressing

  1. the unique minimizing geodesic γX,Y(t)\gamma_{X,Y}(t) in terms of the entries of XX, YY, and tt;

  2. the geodesic distance d(X,Y)d(X,Y) in terms of the entries of XX and YY;

  3. the Riemannian logarithm log⁡XY\log_X Y in terms of the entries of XX and YY.

More precisely, there exist a point X∈V(k,n)X \in\mathrm{V}(k,n) and an elementary real-analytic curve Y(s)∈V(k,n)Y(s) \in\mathrm{V}(k,n) defined near s=0s=0 such that if γs:[0,1]→V(k,n)\gamma_s:[0,1] \to\mathrm{V}(k,n) is the unique minimizing geodesic from XX to Y(s)Y(s), then the functions

[γ˙s(0)]2,1=[log⁡XY(s)]2,1,[γs(12)]3,1,d(X,Y(s))(4)[\dot{\gamma}_s(0)]_{2,1} = [\log_X Y(s)]_{2,1}, \qquad[\gamma_s(\tfrac{1}{2})]_{3,1}, \qquad d(X,Y(s)) \tag*{(4)}

are all non-Liouvillian over C(s)\mathbb{C}(s). Here [Z]i,j[Z]_{i,j} denotes the (i,j)(i,j)th entry of the matrix $Z.

Note that there are three exceptional cases for which Problem 2 does have closed-form solutions: V(1,n)=Sn−1\mathrm{V}(1,n)=\mathbb{S}^{n-1}, V(n−1,n)≅SO(n)\mathrm{V}(n-1,n)\cong\mathrm{SO}(n), and V(n,n)=O(n)\mathrm{V}(n,n)=\mathrm{O}(n); for easy reference, we will provide these solutions in § 3. So the bounds n≥4n \ge4 and 2≤k≤n−22 \le k \le n-2 in Theorem 1.1 are sharp. Aside from (k,n)=(1,n),(n−1,n),(n,n)(k,n)=(1,n),(n-1,n),(n,n), Problem 2 has no closed-form solutions. In § 4, we show that Theorem 1.1 extends to all common, and even some uncommon, models of the Stiefel manifold, noting that (1) is just one among many other possible models.1

For contrast, explicit closed-form formulas for the Grassmannian analogs of both Problems 1 and 2 are known for all k≤nk \le n, and in several different models of Gr(k,n)\mathrm{Gr}(k,n). In particular, closed-form formulas for the geodesic between two points on the Grassmannian first appeared in [31] for the quotient model Gr(k,n)≅V(k,n)/O(k)\mathrm{Gr}(k,n)\cong\mathrm{V}(k,n)/\mathrm{O}(k), in [5] for the projection model Gr(k,n)≅{X∈Rn×n:XT=X=X2,tr⁡(X)=k}\mathrm{Gr}(k,n)\cong\{X \in\mathbb{R}^{n \times n}: X^{\mathsf{T}}=X=X^2,\operatorname{tr}(X)=k\}, and in [15] for the involution model Gr(k,n)≅{X∈O(n):XT=X,tr⁡(X)=2k−n}\mathrm{Gr}(k,n)\cong\{X \in\mathrm{O}(n): X^{\mathsf{T}}=X,\operatorname{tr}(X)=2k-n\}.

Proof of Theorem 1.1

We begin with the following general result on constructing non-Liouvillian functions from algebraic functions. The proof assumes basic knowledge of field theory [20].

Lemma 2.1. Let ff and gg be nonconstant algebraic functions over C(x)\mathbb{C}(x). Let x0≠0x_0 \ne0 be a point at which analytic branches of ff and gg are defined, with f(x0)≠0f(x_0) \ne0. Choose an analytic branch of log⁡x\log x and an analytic branch of log⁡f(x)\log f(x) on a sufficiently small neighborhood of x0x_0. If

t(x):=log⁡x+g(x)log⁡f(x)t(x) := \log x + g(x) \log f(x)

satisfies t′(x0)≠0t'(x_0) \ne0, then the local inverse x=x(t)x = x(t) defined near t0=t(x0)t_0 = t(x_0) is not Liouvillian over C(t)\mathbb{C}(t).

Proof. Since t′(x0)≠0t'(x_0) \ne0, tt has a local analytic inverse x=x(t)x = x(t) near t0=t(x0)t_0 = t(x_0) by the inverse function theorem. Throughout the proof, prime always denotes differentiation with respect to tt. Thus

t=log⁡x(t)+g(x(t))log⁡f(x(t)).(5)t = \log x(t) + g(x(t)) \log f(x(t)). \tag*{(5)}

Suppose for contradiction that x(t)x(t) is Liouvillian over C(t)\mathbb{C}(t). Since f(x)f(x) and g(x)g(x) are algebraic over C(x)\mathbb{C}(x), and logarithms can be adjoined in a Liouvillian extension [30 Definition 1.42], there exists a finite tower of fields

C(t)=K0⊆K1⊆⋯⊆Km\mathbb{C}(t) = K_0 \subseteq K_1 \subseteq\cdots\subseteq K_m

such that S:={x(t),f(x(t)),g(x(t)),log⁡x(t),log⁡f(x(t))}⊆KmS := \{x(t), f(x(t)), g(x(t)), \log x(t), \log f(x(t))\} \subseteq K_m, Ki=Ki−1(θi)K_i = K_{i-1}(\theta_i) such that θi\theta_i is either algebraic over Ki−1K_{i-1} or θi′∈Ki−1\theta_i' \in K_{i-1} or θi≠0\theta_i \ne0 and θi′/θi∈Ki−1\theta_i'/\theta_i \in K_{i-1}, and KiK_i has constant field C\mathbb{C} for i=0,…,mi = 0,\ldots,m.

We prove by descending induction on ii that every element of SS is algebraic over KiK_i. The base case i=mi = m trivially holds since S⊆KmS \subseteq K_m. Fix 1≤i≤m1 \le i \le m and assume that every element of SS is algebraic over KiK_i. If Ki/Ki−1K_i/K_{i-1} is algebraic, the claim follows from transitivity of algebraicity. Otherwise, θi\theta_i is transcendental over Ki−1K_{i-1} and either θi′∈Ki−1\theta_i' \in K_{i-1} or θi′/θi∈Ki−1\theta_i'/\theta_i \in K_{i-1}. Let

E=Ki−1(θi,x,f(x),g(x),log⁡x,log⁡f(x)).E = K_{i-1}(\theta_i, x, f(x), g(x), \log x, \log f(x)).

Since Ki=Ki−1(θi)K_i = K_{i-1}(\theta_i) and every element of SS is algebraic over KiK_i by the induction hypothesis, EE is a finite algebraic extension of Ki−1(θi)K_{i-1}(\theta_i).

Let FF be the subfield of EE consisting of all elements algebraic over Ki−1K_{i-1}. For every finite subextension F~/Ki−1\widetilde{F}/K_{i-1} contained in FF, transcendence of θi\theta_i gives

[F~:Ki−1]=[F~(θi):Ki−1(θi)]≤[E:Ki−1(θi)].[\widetilde{F}:K_{i-1}] = [\widetilde{F}(\theta_i):K_{i-1}(\theta_i)] \le[E:K_{i-1}(\theta_i)].

noting that here [E:F][E:F] denotes the degree of EE over FF, i.e., the dimension of EE as an FF-vector space [20 Example 20.11 and Proposition 1.20]. The degrees of these finite subextensions are bounded. Choose F~\widetilde{F} of maximal degree. For any a∈Fa \in F, the extension F~(a)/Ki−1\widetilde{F}(a)/K_{i-1} is finite, so maximality forces a∈F~a \in\widetilde{F}. Therefore, F=F~F = \widetilde{F} and F/Ki−1F/K_{i-1} is finite.

Since algebraic extensions of differential fields in characteristic zero are stable under differentiation, FF is a differential field. By construction, FF is algebraically closed in EE, EE is finite algebraic over F(θi)F(\theta_i), θi\theta_i is transcendental over FF, and FF and EE have the same field of constants C\mathbb{C}. We may therefore apply [24 Theorem, p. 18] to the identities

x′x−(log⁡x)′=0,f′(x)f(x)−(log⁡f(x))′=0.\frac{x'}{x} - (\log x)' = 0,\qquad\frac{f'(x)}{f(x)} - (\log f(x))' = 0.

We proceed by dividing into the following two cases.

Case 1: θi′∈F\theta_i' \in F. By [24 Theorem, p. 18], x,f(x)∈Fx, f(x) \in F, and

log⁡x=aθi+b,log⁡f(x)=cθi+d,(6)\log x = a\theta_i + b,\qquad\log f(x) = c\theta_i + d, \tag*{(6)}

for some a,c∈Ca,c \in\mathbb{C} and b,d∈Fb,d \in F. Since x∈Fx \in F and g(x)g(x) is algebraic over C(x)\mathbb{C}(x), g(x)∈Fg(x) \in F as FF is algebraically closed in EE. Substituting (2.2) into (2.1) gives

(a+cg(x))θi=t−b−dg(x)∈F.(a+cg(x))\theta_i=t-b-dg(x)\in F.

Since θi\theta_i is transcendental over FF, a+cg(x)=0a+cg(x)=0. As g(x)g(x) is nonconstant and a,c∈Ca,c\in\mathbb{C}, a=c=0a=c=0. Therefore, log⁡x,log⁡f(x)∈F\log x,\log f(x)\in F and hence S⊆FS\subseteq F.

Case 2: θi/θj∈F\theta_i/\theta_j\in F. By [24 Theorem, p. 18], log⁡x,log⁡f(x)∈F\log x,\log f(x)\in F. Since ff is nonconstant, log⁡f(x)\log f(x) is not identically zero. Therefore,

g(x)=t−log⁡xlog⁡f(x)∈F.g(x)=\frac{t-\log x}{\log f(x)}\in F.

Since gg is a nonconstant algebraic function of xx, xx and g(x)g(x) are algebraically dependent over C\mathbb{C}, and xx is algebraic over C(g(x))\mathbb{C}(g(x)). Hence, xx is algebraic over FF and x∈Fx\in F as FF is algebraically closed in EE. Similarly, since f(x)f(x) is algebraic over C(x)\mathbb{C}(x), we obtain f(x)∈Ff(x)\in F and hence S⊆FS\subseteq F.

In both cases, S⊆FS\subseteq F. Since FF is algebraic over Ki−1K_{i-1}, every element of SS is algebraic over Ki−1K_{i-1}, completing the induction. Descending from i=mi=m to i=0i=0, we obtain that every element of SS is algebraic over C(t)\mathbb{C}(t).

Since x(t)x(t) is nonconstant and algebraic over C(t)\mathbb{C}(t), the variable tt is algebraic over C(x)\mathbb{C}(x) and hence log⁡x\log x is algebraic over C(x)\mathbb{C}(x). Let

p(x,y)=yr+ar−1(x)yr−1+⋯+a0(x)p(x,y)=y^r+a_{r-1}(x)y^{r-1}+\cdots+a_0(x)

be the monic minimal polynomial of log⁡x\log x over C(x)\mathbb{C}(x), where aj∈C(x)a_j\in\mathbb{C}(x). Using (log⁡x)′=x′/x(\log x)'=x'/x and differentiating p(x(t),log⁡x(t))=0p(x(t),\log x(t))=0 with respect to tt, we obtain

0=x′(∂p∂x(x,log⁡x)+1x∂p∂y(x,log⁡x)).0=x'\left(\frac{\partial p}{\partial x}(x,\log x)+\frac{1}{x}\frac{\partial p}{\partial y}(x,\log x)\right).

Since x′x' is not identically zero,

∂p∂x(x,log⁡x)+1x∂p∂y(x,log⁡x)=0.\frac{\partial p}{\partial x}(x,\log x)+\frac{1}{x}\frac{\partial p}{\partial y}(x,\log x)=0.

Notice that the polynomial

∂p∂x(x,y)+1x∂p∂y(x,y)\frac{\partial p}{\partial x}(x,y)+\frac{1}{x}\frac{\partial p}{\partial y}(x,y)

has degree at most r−1r-1 in yy and vanishes at y=log⁡xy=\log x. By the minimality of pp, it must therefore vanish identically. Comparing the coefficients of yr−1y^{r-1} gives

dar−1dx=−rx.\frac{d a_{r-1}}{d x}=-\frac{r}{x}.

Since the derivative of a rational function cannot have a simple pole, we arrive at a contradiction.

We are now ready to prove Theorem 1.1. The proof strategy is to construct a point XX and an elementary real-analytic curve of endpoints Y(s)Y(s) for which the Riemannian logarithm, the midpoint of the minimizing geodesic, and the geodesic distance from XX to YY are not Liouvillian over C(s)\mathbb{C}(s), the field of complex rational functions in the variable ss.

Proof of Theorem 1.1. Let ε>0\varepsilon>0 be sufficiently small, J:=[0−110]J:=\begin{bmatrix}0&-1\\1&0\end{bmatrix}, and

X:=[I20000Ik−200],Y(s):=[cos⁡εe−sJ0−sin⁡εe−sJ00Ik−200](7)X:=\begin{bmatrix} I_2 & 0\\ 0 & 0\\ 0 & I_{k-2}\\ 0 & 0 \end{bmatrix},\qquad Y(s):=\begin{bmatrix} \cos\varepsilon e^{-sJ} & 0\\ -\sin\varepsilon e^{-sJ} & 0\\ 0 & I_{k-2}\\ 0 & 0 \end{bmatrix} \tag*{(7)}

where the blocks are of row sizes 22, 22, k−2k-2, and n−k−2n-k-2, respectively. By construction, X,Y(s)∈V(k,n)X,Y(s)\in V(k,n), and every entry of Y(s)Y(s) is an elementary function in ss. Using implicit function theorem, let ff be the algebraic function defined near x=1x=1 by f(x)+f(x)−1=(x+x−1)cos⁡εf(x)+f(x)^{-1}=(x+x^{-1})\cos\varepsilon and f(1)=eiεf(1)=e^{i\varepsilon} and let

g(x)=x−x−1f(x)−f(x)−1cos⁡ε.g(x)=\frac{x-x^{-1}}{f(x)-f(x)^{-1}}\cos\varepsilon.

Clearly, both ff and gg are nonconstant algebraic functions, analytic near x=1x=1, with f(1)≠0f(1)\ne0, g(1)=0g(1)=0, and g′(1)=−icot⁡ε≠0g'(1)=-i\cot\varepsilon\ne0. Choose the logarithm branches satisfying log⁡1=0\log1=0 and log⁡f(1)=iε\log f(1)=i\varepsilon. Then

ddx(log⁡x+g(x)log⁡f(x))∣x=1=1+εcot⁡ε≠0.\left.\frac{d}{dx}\left(\log x+g(x)\log f(x)\right)\right|_{x=1}=1+\varepsilon\cot\varepsilon\ne0.

Therefore, by implicit function theorem,

is=log⁡x+g(x)log⁡f(x)(8)is=\log x+g(x)\log f(x) \tag*{(8)}

uniquely determines an analytic function x=x(s)x=x(s) near s=0s=0, with x(0)=1x(0)=1. By Lemma 2.1, x(s)x(s) is not Liouvillian over C(s)\mathbb{C}(s).

By construction, f(1/x‾)=1/f(x)‾f(1/\overline{x})=1/\overline{f(x)} and g(1/x‾)=g(x)‾g(1/\overline{x})=\overline{g(x)}. With the chosen branches of the logarithmic function, log⁡(1/x‾)=−log⁡x‾\log(1/\overline{x})=-\overline{\log x} and log⁡f(1/x‾)=−log⁡f(x)‾\log f(1/\overline{x})=-\overline{\log f(x)}. Consequently, by (8),

log⁡(1/x(s)‾)+g(1/x(s)‾)log⁡f(1/x(s)‾)=is.\log(1/\overline{x(s)})+g(1/\overline{x(s)})\log f(1/\overline{x(s)})=is.

By the local uniqueness of the solution to (8), we obtain that x(s)=1/x(s)‾x(s)=1/\overline{x(s)}, so ∣x(s)∣=∣f(x(s))∣=1|x(s)|=|f(x(s))|=1 and g(x(s))∈Rg(x(s))\in\mathbb{R}.

With x=x(s)x=x(s), define

H(s)=[2i g(x)log⁡f(x)J0−2sin⁡εlog⁡f(x)f(x)−f(x)−1ei(log⁡x)J00000]∈Rn×k.H(s)=\begin{bmatrix} 2i\,g(x)\log f(x)J & 0\\ -\dfrac{2\sin\varepsilon\log f(x)}{f(x)-f(x)^{-1}}e^{i(\log x)J} & 0\\ 0 & 0\\ 0 & 0 \end{bmatrix}\in\mathbb{R}^{n\times k}.

Since XTH(s)=−H(s)TXX^{T}H(s)=-H(s)^{T}X, H(s)∈TXV(k,n)H(s)\in\mathbb{T}_{X}V(k,n) [6 Section 2.2.1].

We now verify that the geodesic γs(t)\gamma_s(t) with initial point XX along H(s)H(s) reaches Y(s)Y(s) at t=1t=1. By definition of ff, we have

(f(x)−f(x)−1)2=cos⁡2ε(x−x−1)2−4sin⁡2ε.\left(f(x)-f(x)^{-1}\right)^2=\cos^2\varepsilon\left(x-x^{-1}\right)^2-4\sin^2\varepsilon.

Combined with J2=−I2J^2=-I_2, plugging it into (3) yields

γs(t)=[exp⁡(itg(x)log⁡f(x)J)(f(x)t+f(x)−t2I2+ig(x)(f(x)t−f(x)−t)2J)0−(f(x)t−f(x)−tf(x)−f(x)−1)exp⁡(i(log⁡x+tg(x)log⁡f(x))J)sin⁡ε00Ik−200],(9)\gamma_s(t)=\begin{bmatrix} \exp\left(itg(x)\log f(x)J\right)\left(\dfrac{f(x)^t+f(x)^{-t}}{2}I_2+\dfrac{ig(x)(f(x)^t-f(x)^{-t})}{2}J\right) & 0\\ -\left(\dfrac{f(x)^t-f(x)^{-t}}{f(x)-f(x)^{-1}}\right)\exp\left(i(\log x+tg(x)\log f(x))J\right)\sin\varepsilon& 0\\ 0 & I_{k-2}\\ 0 & 0 \end{bmatrix}, \tag*{(9)}

where f(x)t=exp⁡(tlog⁡f(x))f(x)^t=\exp(t\log f(x)). Using the definitions of ff and gg and (8) (1,1)(1,1)-block and (2,1)(2,1)-block reduce to (cos⁡ε)e−sJ(\cos\varepsilon)e^{-sJ} and −(sin⁡ε)e−sJ-(\sin\varepsilon)e^{-sJ}, respectively, at t=1t=1. Therefore, γs(1)=Y(s)\gamma_s(1)=Y(s) and H(s)=log⁡XY(s)H(s)=\log_XY(s).

At s=0s=0, we have

H(0)=[00−εI200000].H(0)=\begin{bmatrix} 0 & 0\\ -\varepsilon I_2 & 0\\ 0 & 0\\ 0 & 0 \end{bmatrix}.

Choosing ε\varepsilon sufficiently small and restricting ss to a neighborhood of zero, H(s)H(s) lies in a normal neighborhood of the Riemannian exponential at XX. Therefore, γs\gamma_s is the unique minimizing geodesic joining XX and Y(s)Y(s).

Finally, by the definition of H(s)H(s) and (8),

[γ˙s(0)]2,1=2i g(x)log⁡f(x)=2i(is−log⁡x).[\dot{\gamma}_s(0)]_{2,1}=2i\,g(x)\log f(x)=2i(is-\log x).

If it is Liouvillian over C(s)\mathbb{C}(s), then log⁡x(s)\log x(s), and consequently, x(s)=exp⁡(log⁡x(s))x(s)=\exp(\log x(s)), are Liouvillian, contradicting Lemma 2.12.1.

We next show that the geodesic itself has a non-Liouvillian entry at t=12t=\frac{1}{2}. Consider [γs(12)]3,1[\gamma_s(\frac{1}{2})]_{3,1} as a function of ss. By (9) and (8),

[γs(12)]3,1=−sin⁡εf(x)1/2+f(x)−1/2cosh⁡(log⁡x+is2).\left[\gamma_s\left(\frac{1}{2}\right)\right]_{3,1} = -\frac{\sin\varepsilon}{f(x)^{1/2}+f(x)^{-1/2}} \cosh\left(\frac{\log x+is}{2}\right).

Using the defining equation for ff, we obtain

[γs(12)]3,12=sin⁡2ε(2+eisx+e−isx−1)4(2+cos⁡ε(x+x−1)).\left[\gamma_s\left(\frac{1}{2}\right)\right]_{3,1}^{2} = \frac{\sin^{2}\varepsilon(2+e^{is}x+e^{-is}x^{-1})} {4(2+\cos\varepsilon(x+x^{-1}))}.

Consequently, x=x(s)x=x(s) satisfies

0=(4(cos⁡ε)[γs(12)]3,12−(sin⁡2ε)eis)x2+(8[γs(12)]3,12−2sin⁡2ε)x+4(cos⁡ε)[γs(12)]3,12−(sin⁡2ε)e−is.0=\left(4(\cos\varepsilon)\left[\gamma_s\left(\frac{1}{2}\right)\right]_{3,1}^{2}-(\sin^{2}\varepsilon)e^{is}\right)x^{2} +\left(8\left[\gamma_s\left(\frac{1}{2}\right)\right]_{3,1}^{2}-2\sin^{2}\varepsilon\right)x +4(\cos\varepsilon)\left[\gamma_s\left(\frac{1}{2}\right)\right]_{3,1}^{2}-(\sin^{2}\varepsilon)e^{-is}.

Since [γ0(12)]3,1=−sin⁡(ε/2)\left[\gamma_0\left(\frac{1}{2}\right)\right]_{3,1}=-\sin(\varepsilon/2), the coefficient of x2x^2 at s=0s=0 is −(1−cos⁡ε)2≠0-(1-\cos\varepsilon)^2\ne0. If [γs(12)]3,1\left[\gamma_s\left(\frac{1}{2}\right)\right]_{3,1} were Liouvillian over C(s)\mathbb{C}(s), then x(s)x(s) would be algebraic over a Liouvillian extension of C(s)\mathbb{C}(s), and hence Liouvillian, contradicting Lemma 2.12.1.

Lastly, let

S:=[−J000]∈Rk×k.S:=\begin{bmatrix}-J&0\\0&0\end{bmatrix}\in\mathbb{R}^{k\times k}.

By construction, Y′(s)=Y(s)SY'(s)=Y(s)S. Since right multiplication by orthogonal matrices is an isometry with respect to (2), Z↦ZSZ\mapsto ZS defines a Killing vector field. By the first variation formula and the fact that the inner product of a Killing field with the geodesic velocity is conserved,

ddsd(X,Y(s))2=2gY(s)(γ˙s(1),Y(s)S)=2gX(H(s),XS)=−2[γ˙s(0)]2,1.\frac{d}{ds}d(X,Y(s))^{2} = 2g_{Y(s)}(\dot{\gamma}_s(1),Y(s)S) = 2g_X(H(s),XS) = -2[\dot{\gamma}_s(0)]_{2,1}.

which is not Liouvillian over C(s)\mathbb{C}(s) by the previous argument. Since the derivative of a Liouvillian function is Liouvillian, d(X,Y(s))2d(X,Y(s))^2 is not Liouvillian over C(s)\mathbb{C}(s), so neither is d(X,Y(s))d(X,Y(s)).

Suppose, for contradiction, that at least one of log⁡XY\log_XY, d(X,Y)d(X,Y), or γX,Y(12)\gamma_{X,Y}(\frac{1}{2}) admits an elementary formula in the entries of XX and YY, for all sufficiently close XX and YY. Since elementary functions are closed under composition, substituting (7) into this formula implies that at least one of (4) is elementary, and hence Liouvillian, over C(s)\mathbb{C}(s), a contradiction. □\square

Exceptional cases

For easy reference, we show that geodesics in V(k,n)V(k,n) do indeed have closed-form expressions in the three exceptional cases k=1k=1, n−1n-1, nn excluded from Theorem 1.11.1. For k=1k=1, V(1,n)=Sn−1V(1,n)=S^{n-1}, the unit sphere. Given u,v∈Sn−1u,v\in S^{n-1} with u≠±vu\ne\pm v, we have [16 Proposition 5.27]

log⁡uv=arccos⁡(uTv)1−(uTv)2(v−(uTv)u),d(u,v)=arccos⁡(uTv),\log_u v=\frac{\arccos(u^{T}v)}{\sqrt{1-(u^{T}v)^2}}\left(v-(u^{T}v)u\right), \qquad d(u,v)=\arccos(u^{T}v),
γ(t)=sin⁡((1−t)arccos⁡(uTv))1−(uTv)2u+sin⁡(tarccos⁡(uTv))1−(uTv)2v.\gamma(t)=\frac{\sin((1-t)\arccos(u^{T}v))}{\sqrt{1-(u^{T}v)^2}}u +\frac{\sin(t\arccos(u^{T}v))}{\sqrt{1-(u^{T}v)^2}}v.

For k=n−1k=n-1, V(n−1,n)≅SO⁡(n)V(n-1,n)\cong\operatorname{SO}(n) [8 Section 3.D], identifying (2) with the bi-invariant metric hX(U,V)=12tr⁡(UTV)h_X(U,V)=\frac{1}{2}\operatorname{tr}(U^{T}V) on SO⁡(n)\operatorname{SO}(n). For X,Y∈SO⁡(n)X,Y\in\operatorname{SO}(n) such that XTYX^{T}Y has no eigenvalue −1-1, we have [1 Theorem 2.27]

log⁡XY=Xlog⁡(XTY),γ(t)=Xexp⁡(tlog⁡(XTY)),d(X,Y)=12∥log⁡(XTY)∥F(10)\log_XY=X\log(X^{T}Y),\qquad\gamma(t)=X\exp\bigl(t\log(X^{T}Y)\bigr),\qquad d(X,Y)=\frac{1}{\sqrt{2}}\lVert\log(X^{T}Y)\rVert_F \tag*{(10)}

where log⁡\log denotes the principal matrix logarithm. For k=nk=n, V(n,n)=O⁡(n)=SO⁡(n)∪SO⁡−(n)V(n,n)=\operatorname{O}(n)=\operatorname{SO}(n)\cup\operatorname{SO}^{-}(n), and (10) applies within each connected component.

Other models of the Stiefel manifold

While the matrix representation of the Stiefel manifold in (1) as orthonormal n×kn\times k matrices is taken to be synonymous with the Stiefel manifold, this is strictly speaking a fallacy, although common practice in applied mathematics. The Stiefel manifold is an abstract topological manifold of orthonormal kk-frames, which is Stiefel’s original definition [28] as well as the standard definition in differential geometry, see [13 Section VII.2] or [10 Section 2] for example.

The Stiefel manifold has many different representations aside from (1). For example, it has infinitely many isomorphic representations as a submanifold of Rn×k\mathbb{R}^{n\times k} [17] and may also be represented as quotient manifolds O⁡(n)/O⁡(n−k)\operatorname{O}(n)/\operatorname{O}(n-k), SO⁡(n)/SO⁡(n−k)\operatorname{SO}(n)/\operatorname{SO}(n-k), Spin⁡(n)/Spin⁡(n−k)\operatorname{Spin}(n)/\operatorname{Spin}(n-k), or a myriad of yet more esoteric possibilities. We will extend Theorem 1.1 to other models to show that no closed-form formula exists whenever the model is isometric to (1) by a Nash diffeomorphism. Readers unfamiliar with Nash manifolds and Nash diffeomorphisms may consult [26]. Here it suffices to say that all aforementioned models of the Stiefel manifold are Nash manifolds — see [17] for the submanifold models and [7] for the quotient models.

Proposition 4.1. Let n≥4n\geq4, 2≤k≤n−22\leq k\leq n-2, and MM be a Nash manifold with Riemannian metric g′g'. If there exists an isometry

φ:(V(k,n),g)→(M,g′)\varphi:(V(k,n),g)\to(M,g')

that is a Nash diffeomorphism, then there exist x∈Mx\in M and an elementary real-analytic curve ξ(s)∈M\xi(s)\in M defined near zero such that for the Riemannian logarithm and the midpoint of the unique minimizing geodesic, least one of their coordinate functions in Nash coordinates is not Liouvillian, and the geodesic distance from xx to ξ(s)\xi(s) are not Liouvillian over C(s)\mathbb{C}(s).

Proof. Let XX, Y(s)Y(s), and γs\gamma_s be as defined in the proof of Theorem 1.1, and let x:=φ(X)x:=\varphi(X), ξ(s):=φ(Y(s))\xi(s):=\varphi(Y(s)), and ηs(t):=φ(γs(t))\eta_s(t):=\varphi(\gamma_s(t)). Since φ\varphi is Nash, ξ(s)\xi(s) is elementary and real-analytic. Since φ\varphi is an isometry, ηs\eta_s is the unique minimizing geodesic from xx to ξ(s)\xi(s), and

log⁡xξ(s)=dφX(log⁡XY(s)),dg′(x,ξ(s))=d(X,Y(s)).\log_x\xi(s)=d\varphi_X(\log_XY(s)),\qquad d_{g'}(x,\xi(s))=d(X,Y(s)).

Since dφX−1d\varphi_X^{-1} is a linear map, if every coordinate function of log⁡xξ(s)\log_x\xi(s) is Liouvillian over C(s)\mathbb{C}(s), then so is every entry of log⁡XY(s)\log_XY(s), a contradiction to Theorem 1.1. Similarly, if every coordinate function of ηs(12)\eta_s(\frac{1}{2}) is Liouvillian over C(s)\mathbb{C}(s), then γs(12)=φ−1(ηs(12))\gamma_s(\frac{1}{2})=\varphi^{-1}(\eta_s(\frac{1}{2})) would be Liouvillian as φ−1\varphi^{-1} is Nash, a contradiction to Theorem 1.1. Lastly, dg′(x,ξ(s))=d(X,Y(s))d_{g'}(x,\xi(s))=d(X,Y(s)), which is already known to be non-Liouvillian over C(s)\mathbb{C}(s) by Theorem 1.1. □\square

Conclusion

A negative result like Theorem 1.1 does not necessarily entail a dead end. For one, the knowledge that there is no general closed-form formula ascertains the value of numerical approaches like those in [32, 4, 29, 19, 21]. We hope our work here would impel developments of even better numerical algorithms for the Riemannian logarithm, endpoint geodesic, and geodesic distance on the Stiefel manifold.

AI disclosure

The following are due entirely to the authors without AI assistance: (i) the leap of faith that Problem 2 has no solution; (ii) the observation that matrix functions are elementary functions; (iii) the insight to use differential Galois theory; the twin strategy of (iv) narrowing down the infinite possibilities to (k,n)=(2,4)(k,n)=(2,4) and then (v) extending it back to include all 2≤k≤n−22 \le k \le n-2 and n≥4n \ge4 through an isometric embedding. ChatGPT-6 Astra had supplied one important step — it was used to search for a (k,n)=(2,4)(k,n)=(2,4) counterexample after we put the scaffolding of the proof in place.

Acknowledgment

The authors are supported by a Vannevar Bush Faculty Fellowship ONR N000142312863.

Computational and Applied Mathematics, University of Chicago, Chicago, IL 60637

Email address: [email protected]

Computational and Applied Mathematics Initiative, University of Chicago, Chicago, IL 60637

Email address: [email protected]

References

  1. [1]M. M. Alexandrino and R. G. Bettiol. Lie groups and geometric aspects of isometric actions. Springer, Cham, 2015.DOI
  2. [2]T. Bendokat and R. Zimmermann. Efficient quasi-geodesics on the Stiefel manifold. In Geometric science of information, volume 12829 of Lecture Notes in Comput. Sci., pages 763–771. Springer, Cham, [2021] ©2021.DOI
  3. [3]J. Bernstein. Modular manifolds. Thinking Machines Lab: Connectionism, https://thinkingmachines.ai/blog/modular-manifolds/, 2025.DOI
  4. [4]D. Berner. Endpoint geodesics on the Stiefel manifold embedded in Euclidean space. SIAM J. Matrix Anal. Appl., 38(4):1139–1159, 2017.
  5. [5]C. Davis. Separation of two linear subspaces. Acta Sci. Math. (Szeged), 19:172–187, 1958.
  6. [6]A. Edelman, T. A. Arias, and S. T. Smith. The geometry of algorithms with orthogonality constraints. SIAM J. Matrix Anal. Appl., 20(2):303–353, 1999.arxiv.org/abs/physics/9806030
  7. [7]Y. Fang and B. Sun. Chevalley’s theorem for affine Nash groups. J. Lie Theory, 26(2):359–369, 2016.arxiv.org/abs/1506.03211
  8. [8]A. Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002.
  9. [9]N. J. Higham. Functions of matrices. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2008. Theory and computation.
  10. [10]I. M. James. The topology of Stiefel manifolds, volume No. 24 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge-New York-Melbourne, 1976.DOI
  11. [11]R. Jensen and R. Zimmermann. An new polar factor retraction on the Stiefel manifold with closed-form inverse. arXiv:2602.19923, 2026.arxiv.org/abs/2602.19923
  12. [12]K. Jordan, Y. Jin, V. Boza, J. You, F. Cesista, L. Newhouse, and J. Bernstein. Muon: An optimizer for hidden layers in neural networks. https://kellerjordan.github.io/posts/muon/, 2024.
  13. [13]S. Kobayashi and K. Nomizu. Foundations of differential geometry. Vol. 1. Wiley Classics Library. John Wiley & Sons, Inc., New York, 1996. Reprint of the 1963 original, A Wiley-Interscience Publication.
  14. [14]K. A. Krakowski, L. Machado, F. Silva Leite, and J. Batista. A modified Casteljau algorithm to solve interpolation problems on Stiefel manifolds. J. Comput. Appl. Math., 311:84–99, 2017.DOI
  15. [15]Z. Lai, L.-H. Lim, and K. Ye. Simpler Grassmannian optimization. arXiv:2009.13502, 2020.arxiv.org/abs/2009.13502
  16. [16]J. M. Lee. Introduction to Riemannian manifolds, volume 176 of Graduate Texts in Mathematics. Springer, Cham, second edition, 2018.DOI
  17. [17]L.-H. Lim and K. Ye. Simpler matrix models for the flag, Grassmann, and Stiefel manifolds. arXiv:2407.13482, 2024.arxiv.org/abs/2407.13482
  18. [18]S. Mataigne, P.-A. Absil, and N. Miolane. Bounds on the geodesic distances on the Stiefel manifold for a family of Riemannian metrics. Linear Algebra Appl., 730:1–34, 2026.arxiv.org/abs/2408.07072
  19. [19]S. Mataigne, R. Zimmermann, and N. Miolane. An efficient algorithm for the Riemannian logarithm on the Stiefel manifold for a family of Riemannian metrics. SIAM J. Matrix Anal. Appl., 46(2):879–905, 2025.arxiv.org/abs/2403.11730
  20. [20]P. Morandi. Field and Galois theory, volume 167 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1996.DOI
  21. [21]Q. Rentmeesters. Algorithms for Data Fitting on Some Common Homogeneous Spaces. PhD thesis, Université catholique de Louvain, Louvain-la-Neuve, Belgium, 2013.
  22. [22]R. H. Risch. The problem of integration in finite terms. Trans. Amer. Math. Soc., 139:167–189, 1969.DOI
  23. [23]J. F. Ritt. Integration in Finite Terms. Liouville’s Theory of Elementary Methods. Columbia University Press, New York, 1948.
  24. [24]M. Rosenlicht. On the explicit solvability of certain transcendental equations. Inst. Hautes Études Sci. Publ. Math., (36):15–22, 1969.DOI
  25. [25]A. Séguin and D. Kressner. Hermite interpolation with retractions on manifolds. BIT, 64(4):Paper No. 42, 42, 2024.arxiv.org/abs/2212.12259
  26. [26]M. Shiota. Nash manifolds, volume 1269 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1987.
  27. [27]M. Solonko, A. Molchanov, and M. Rakhuba. Muon on the Stiefel manifold admits an exact closed-form update. arXiv:2608.06218, 2026.arxiv.org/abs/2608.06218
  28. [28]E. Stiefel. Richtungsfelder und Fernparallelismus in n-dimensionalen Mannigfaltigkeiten. Comment. Math. Helv., 8(1):305–353, 1935.
  29. [29]M. Sutti. A single shooting method with approximate Fréchet derivative for computing geodesics on the Stiefel manifold. Electron. Trans. Numer. Anal., 60:501–519, 2024.DOI
  30. [30]M. van der Put and M. F. Singer. Galois theory of linear differential equations, volume 328 of Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 2003.
  31. [31]K. Ye and L.-H. Lim. Schubert varieties and distances between subspaces of different dimensions. SIAM J. Matrix Anal. Appl., 37(3):1176–1197, 2016.arxiv.org/abs/1407.0900
  32. [32]R. Zimmermann. A matrix-algebraic algorithm for the Riemannian logarithm on the Stiefel manifold under the canonical metric. SIAM J. Matrix Anal. Appl., 38(2):322–342, 2017.arxiv.org/abs/1604.05054

Paper details

vlnc:W_hnuzmj8e
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 2

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

  • Mathematics

ChatGPT-6 Astra

View statement
Looking for related papers…
Contents