WorksMathematics
No closed-form formula for Stiefel geodesics
Abstract
We use differential Galois theory to prove that given two arbitrary points and on the Stiefel manifold, there is no closed-form analytic expression for the Riemannian logarithm , the geodesic between and , or the geodesic distance between and . In other words, these cannot be expressed in terms of elementary functions of the coordinates of and ; 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 and endpoint on the Stiefel manifold, or, equivalently, a closed-form formula for the Riemannian logarithm . 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 , the Stiefel manifold
is a Riemannian manifold equipped with the natural -invariant metric
For a geodesic on , one may ask for two different kinds of closed-form formulas with different input data:
Problem 1. Given a point and a tangent vector , find a closed-form formula for the geodesic with and .
Problem 2. Given two points and , find a closed-form formula for the geodesic with and .
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 and complex-valued univariate function analytic near the spectrum of , the Lagrange–Hermite formula [9 Equation 1.8] yields
where are the distinct eigenvalues of and is the size of the largest -Jordan block, . Since is an algebraic function in the entries of and derivatives of elementary scalar functions are elementary, the matrix function is elementary whenever the scalar function is elementary.
Problem 1 has a well-known closed-form formula given by
where 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 or the geodesic distance between and . 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 and . For arbitrary and , there is no closed-form formula expressing
the unique minimizing geodesic in terms of the entries of , , and ;
the geodesic distance in terms of the entries of and ;
the Riemannian logarithm in terms of the entries of and .
More precisely, there exist a point and an elementary real-analytic curve defined near such that if is the unique minimizing geodesic from to , then the functions
are all non-Liouvillian over . Here denotes the th entry of the matrix $Z.
Note that there are three exceptional cases for which Problem 2 does have closed-form solutions: , , and ; for easy reference, we will provide these solutions in § 3. So the bounds and in Theorem 1.1 are sharp. Aside from , 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 , and in several different models of . In particular, closed-form formulas for the geodesic between two points on the Grassmannian first appeared in [31] for the quotient model , in [5] for the projection model , and in [15] for the involution model .
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 and be nonconstant algebraic functions over . Let be a point at which analytic branches of and are defined, with . Choose an analytic branch of and an analytic branch of on a sufficiently small neighborhood of . If
satisfies , then the local inverse defined near is not Liouvillian over .
Proof. Since , has a local analytic inverse near by the inverse function theorem. Throughout the proof, prime always denotes differentiation with respect to . Thus
Suppose for contradiction that is Liouvillian over . Since and are algebraic over , and logarithms can be adjoined in a Liouvillian extension [30 Definition 1.42], there exists a finite tower of fields
such that , such that is either algebraic over or or and , and has constant field for .
We prove by descending induction on that every element of is algebraic over . The base case trivially holds since . Fix and assume that every element of is algebraic over . If is algebraic, the claim follows from transitivity of algebraicity. Otherwise, is transcendental over and either or . Let
Since and every element of is algebraic over by the induction hypothesis, is a finite algebraic extension of .
Let be the subfield of consisting of all elements algebraic over . For every finite subextension contained in , transcendence of gives
noting that here denotes the degree of over , i.e., the dimension of as an -vector space [20 Example 20.11 and Proposition 1.20]. The degrees of these finite subextensions are bounded. Choose of maximal degree. For any , the extension is finite, so maximality forces . Therefore, and is finite.
Since algebraic extensions of differential fields in characteristic zero are stable under differentiation, is a differential field. By construction, is algebraically closed in , is finite algebraic over , is transcendental over , and and have the same field of constants . We may therefore apply [24 Theorem, p. 18] to the identities
We proceed by dividing into the following two cases.
Case 1: . By [24 Theorem, p. 18], , and
for some and . Since and is algebraic over , as is algebraically closed in . Substituting (2.2) into (2.1) gives
Since is transcendental over , . As is nonconstant and , . Therefore, and hence .
Case 2: . By [24 Theorem, p. 18], . Since is nonconstant, is not identically zero. Therefore,
Since is a nonconstant algebraic function of , and are algebraically dependent over , and is algebraic over . Hence, is algebraic over and as is algebraically closed in . Similarly, since is algebraic over , we obtain and hence .
In both cases, . Since is algebraic over , every element of is algebraic over , completing the induction. Descending from to , we obtain that every element of is algebraic over .
Since is nonconstant and algebraic over , the variable is algebraic over and hence is algebraic over . Let
be the monic minimal polynomial of over , where . Using and differentiating with respect to , we obtain
Since is not identically zero,
Notice that the polynomial
has degree at most in and vanishes at . By the minimality of , it must therefore vanish identically. Comparing the coefficients of gives
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 and an elementary real-analytic curve of endpoints for which the Riemannian logarithm, the midpoint of the minimizing geodesic, and the geodesic distance from to are not Liouvillian over , the field of complex rational functions in the variable .
Proof of Theorem 1.1. Let be sufficiently small, , and
where the blocks are of row sizes , , , and , respectively. By construction, , and every entry of is an elementary function in . Using implicit function theorem, let be the algebraic function defined near by and and let
Clearly, both and are nonconstant algebraic functions, analytic near , with , , and . Choose the logarithm branches satisfying and . Then
Therefore, by implicit function theorem,
uniquely determines an analytic function near , with . By Lemma 2.1, is not Liouvillian over .
By construction, and . With the chosen branches of the logarithmic function, and . Consequently, by (8),
By the local uniqueness of the solution to (8), we obtain that , so and .
With , define
Since , [6 Section 2.2.1].
We now verify that the geodesic with initial point along reaches at . By definition of , we have
Combined with , plugging it into (3) yields
where . Using the definitions of and and (8) -block and -block reduce to and , respectively, at . Therefore, and .
At , we have
Choosing sufficiently small and restricting to a neighborhood of zero, lies in a normal neighborhood of the Riemannian exponential at . Therefore, is the unique minimizing geodesic joining and .
Finally, by the definition of and (8),
If it is Liouvillian over , then , and consequently, , are Liouvillian, contradicting Lemma .
We next show that the geodesic itself has a non-Liouvillian entry at . Consider as a function of . By (9) and (8),
Using the defining equation for , we obtain
Consequently, satisfies
Since , the coefficient of at is . If were Liouvillian over , then would be algebraic over a Liouvillian extension of , and hence Liouvillian, contradicting Lemma .
Lastly, let
By construction, . Since right multiplication by orthogonal matrices is an isometry with respect to (2), 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,
which is not Liouvillian over by the previous argument. Since the derivative of a Liouvillian function is Liouvillian, is not Liouvillian over , so neither is .
Suppose, for contradiction, that at least one of , , or admits an elementary formula in the entries of and , for all sufficiently close and . 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 , a contradiction.
Exceptional cases
For easy reference, we show that geodesics in do indeed have closed-form expressions in the three exceptional cases , , excluded from Theorem . For , , the unit sphere. Given with , we have [16 Proposition 5.27]
For , [8 Section 3.D], identifying (2) with the bi-invariant metric on . For such that has no eigenvalue , we have [1 Theorem 2.27]
where denotes the principal matrix logarithm. For , , 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 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 -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 [17] and may also be represented as quotient manifolds , , , 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 , , and be a Nash manifold with Riemannian metric . If there exists an isometry
that is a Nash diffeomorphism, then there exist and an elementary real-analytic curve 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 to are not Liouvillian over .
Proof. Let , , and be as defined in the proof of Theorem 1.1, and let , , and . Since is Nash, is elementary and real-analytic. Since is an isometry, is the unique minimizing geodesic from to , and
Since is a linear map, if every coordinate function of is Liouvillian over , then so is every entry of , a contradiction to Theorem 1.1. Similarly, if every coordinate function of is Liouvillian over , then would be Liouvillian as is Nash, a contradiction to Theorem 1.1. Lastly, , which is already known to be non-Liouvillian over by Theorem 1.1.
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 and then (v) extending it back to include all and through an isometric embedding. ChatGPT-6 Astra had supplied one important step — it was used to search for a 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]M. M. Alexandrino and R. G. Bettiol. Lie groups and geometric aspects of isometric actions. Springer, Cham, 2015.DOI
- [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]J. Bernstein. Modular manifolds. Thinking Machines Lab: Connectionism, https://thinkingmachines.ai/blog/modular-manifolds/, 2025.DOI
- [4]D. Berner. Endpoint geodesics on the Stiefel manifold embedded in Euclidean space. SIAM J. Matrix Anal. Appl., 38(4):1139–1159, 2017.
- [5]C. Davis. Separation of two linear subspaces. Acta Sci. Math. (Szeged), 19:172–187, 1958.
- [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]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]A. Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002.
- [9]N. J. Higham. Functions of matrices. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2008. Theory and computation.
- [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]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]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]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]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]Z. Lai, L.-H. Lim, and K. Ye. Simpler Grassmannian optimization. arXiv:2009.13502, 2020.arxiv.org/abs/2009.13502
- [16]J. M. Lee. Introduction to Riemannian manifolds, volume 176 of Graduate Texts in Mathematics. Springer, Cham, second edition, 2018.DOI
- [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]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]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]P. Morandi. Field and Galois theory, volume 167 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1996.DOI
- [21]Q. Rentmeesters. Algorithms for Data Fitting on Some Common Homogeneous Spaces. PhD thesis, Université catholique de Louvain, Louvain-la-Neuve, Belgium, 2013.
- [22]R. H. Risch. The problem of integration in finite terms. Trans. Amer. Math. Soc., 139:167–189, 1969.DOI
- [23]J. F. Ritt. Integration in Finite Terms. Liouville’s Theory of Elementary Methods. Columbia University Press, New York, 1948.
- [24]M. Rosenlicht. On the explicit solvability of certain transcendental equations. Inst. Hautes Études Sci. Publ. Math., (36):15–22, 1969.DOI
- [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]M. Shiota. Nash manifolds, volume 1269 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1987.
- [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]E. Stiefel. Richtungsfelder und Fernparallelismus in n-dimensionalen Mannigfaltigkeiten. Comment. Math. Helv., 8(1):305–353, 1935.
- [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]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]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]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