Towards sparse optimization over convex loops: Equivalence of Square Root Velocity distance and Wasserstein-Fisher-Rao
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.
Ada: I'm Ada.
Emmy: And I'm Emmy, here with Kolya.
Ada: We're looking at the paper: "Towards sparse optimization over convex loops".
Paper summary: Ada: Thinking about the overall title, "Towards sparse optimization over convex loops: Equivalence of Square Root Velocity distance and Wasserstein-Fisher-Rao," it really captures the progression of the work. It starts with an equivalence result and then immediately moves into using that equivalence to solve a sparsity problem.
Emmy: I think that’s right. The paper’s real contribution is showing that this precise link between SRVT and WFR distance is geometrically meaningful for the circle, which was previously only known in dimension two. It moves beyond just having an analogy and establishes it rigorously in this specific setting of convex curves on the circle.
Kolya: What does that mean when you look at the implications for people who aren't specialists in differential geometry or optimal transport? In simple terms, what’s the big picture here?
Ada: The big picture is that we have a new formulation for elastic Tikhonov regularization. Instead of just tweaking the shape directly, you can regularize it directly at the level of measures describing the length of those curves. This is presented as a way to approach these problems computationally, where the iteration process involves creating convex loops and reconstructing them by adding vertices one by one.
Emmy: It shifts our focus from manipulating tangent vectors on surfaces to manipulating measures on a circle. That measure-theoretic perspective seems like it offers a cleaner, more structured way to handle the underlying geometry of shape matching.
Kolya: And the sparsity result, where every extreme point is just three combinations of others, that’s pretty powerful for computational feasibility. It suggests that even though these optimization spaces can be complicated, you don't need to explore an infinite number of possibilities to find a good solution.
Ada: That’s what they are pointing toward. They are giving us a concrete, measure-theoretic tool and a way to understand the structure of the solutions when we apply this kind of regularization. It connects geometry, transport theory, and optimization in a way that seems very useful for practical shape analysis tasks.
Conclusion: Ada: So we've been looking at this paper, "Towards sparse optimization over convex loops: Equivalence of Square Root Velocity distance and Wasserstein-Fisher-Rao." It really boils down to them proving that for curves on a circle, the square root velocity transformation distance is exactly the same as a Wasserstein-Fisher-Rao distance based on their length measures.
Emmy: That's the big structural point there. They take this known two-dimensional idea and show how it works in one dimension, on the circle, and they use a technique involving optimal transport on the universal cover to make that connection rigorous.
Kolya: What does that mean for us in practice when we're trying to find these shapes? Like, if I'm doing some shape optimization right now with this regularizer, what is it actually telling me about the geometry of my curve?
Ada: It tells you that you can switch gears and think about the problem entirely in terms of measures. Instead of focusing on the tangent vectors directly, you can work with how mass is distributed along the circle. That's a fundamental shift in how we formulate these elastic regularization strategies.
Emmy: Exactly. And they don't just stop at the equivalence; they go deeper into sparsity, looking at what happens when you use these measures in optimization problems. They find some really interesting structural results about how many extreme points you actually need to consider for a solution.
Kolya: So it suggests that even if the optimization problem looks messy, involving these Wasserstein distances, there's a finite bound on how complex the structure of the solutions can be. That helps with computational feasibility when we're trying to find those optimal loops.
Ada: It points toward a new way to think about elastic Tikhonov regularization, one that’s directly at the level of measures and offers a more structured path for building up these solutions piece by piece. But how does this link actually translate into practical algorithms for finding those shapes?
Giacomo Cristinelli, José A. Iglesias
University of Twente
math.OC
Submitted: 2026-03-23
Updated: 2026-11-18
Comments: 32 pages, 4 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
The gist: The paper establishes that for convex absolutely continuous immersions of S1 in the plane, the shape distance induced by the square root velocity transformation (SRVT) is equivalent to the…
Key concepts
- Square Root Velocity Transformation (SRVT)
- This is a transformation applied to curves that relates the shape distance to an optimal reparametrization. It involves mapping the curve's velocity information in a specific way, which allows the problem to be reformulated in terms of one-dimensional transport problems on the circle.
- Length Measure ($\mu_c$)
- This measure captures geometric information about a loop by tracking how the tangent direction of points on the curve is distributed across all directions on a circle. It is defined as the pushforward of arc-length measure via a Gauss map, providing a way to quantify the shape based on directional density.
- Wasserstein-Fisher-Rao (WFR) Distance ($U(\mu_0, \mu_1)$)
- This distance measures the optimal transport cost between two probability measures ($\mu_0$ and $\mu_1$) using a specific loss function involving the square root of the distance between points on a circle. It is used here to define an energy functional that corresponds to the SRVT shape distance.
- Sparsity in Optimization
- This investigates how optimization problems involving WFR distances behave when the reference measure is discrete. The paper finds structural results, showing that extreme points of certain constrained sets are supported by at most three additional points beyond the support of the initial measure, suggesting a sparsity property.
Terminology
Summary
The paper establishes that for convex absolutely continuous immersions of S1 in the plane, the shape distance induced by the square root velocity transformation (SRVT) is equivalent to the Wasserstein–Fisher–Rao distance acting on their associated length measures The first main contribution of this paper is to show that a precise equivalence between the SRVT distance and Wasserstein-Fisher-Rao can be recovered in a natural and geometrically meaningful setting
Equivalence of Distances
The paper proves that for convex absolutely continuous immersions of S1 in the plane, the shape distance induced by the SRVT coincides with the Wasserstein–Fisher–Rao distance acting on the associated length measures on the circle This result provides a one-dimensional analogue of the previously known two-dimensional correspondence The proof involves exploiting a monotonicity principle for optimal transport on the universal cover of the circle, which guarantees the existence of an optimal reparametrization achieving the SRVT infimum and enables a one-dimensional unbalanced optimal transport reformulation
Sparsity in Optimization
The second part investigates sparsity phenomena in optimization problems involving the Wasserstein–Fisher–Rao distance, specifically analyzing linear optimization over WFR balls and proving a finiteness result for extremal solutions when the reference measure is discrete The authors propose a homogenized
version Sν(·) of the WFR energy with respect to a fixed reference measure ν, which compares measures rescaled by their total mass
Structural Results on Extreme Points
The paper explores the structure of extreme points arising from Sν with finitely supported reference measures, suggesting that measures supported in up to three points appear Corollary 4.8 states that every extreme point of B1(µ1)∩M0(S1) is a convex combination of at most three extreme points of B1(µ1), hence it support contains at most three additional points compared to the support of µ1
Optimization Frameworks
The work introduces several optimization concepts, including linear optimization over WFR balls and the proposal for a convex, positively one-homogeneous regularizer suitable for conditional gradient algorithms The paper also discusses distributionally robust optimization and its connection to the Wasserstein distance
Conclusion and Outlook
The results suggest a complementary and measure-theoretic viewpoint on elastic regularization strategies by identifying the SRVT elastic distance with the Wasserstein–Fisher–Rao distance on length measures This provides an alternative formulation of elastic Tikhonov regularization directly at the level of measures The framework is presented as a first step towards effective computational methods, where iterates are convex loops and reconstruction is performed by progressively adding vertices
Technical Details
The paper details technical arguments such as the lift to the universal cover of S1 to replace transport problems on the circle with one-dimensional problems on the real line It establishes that optimal WFR couplings induce optimal SRVT reparametrizations, and conversely The equivalence theorem is stated as dSRVT(c0, c1) = U(µ0, µ1) The analysis relies on properties of admissible measures and the lifting procedure to construct explicit Monge maps related to optimal reparametrization for the SRVT distance The proof involves constructing a coupling γφ from a given reparametrization φ as the pushforward of ds through the map Φ: S1 → C × C defined by Φ:= c′p0c′0, (c1◦φ)′p(c1◦φ)′ The final equivalence is established through the properties of the cost function cT on R × R, which satisfies a Monge condition The paper concludes by showing that for degenerate orthogonal loops, the distance equals µ0(S1) + µ1(S1) The analysis also includes an extension to extended-valued costs and local modification theory to justify the use of certain lemmas in the proof The paper references extensive literature on optimal transport on S1, such as [3, 15, 22] and results concerning elastic energies as Tikhonov regularizers and distributionally robust optimization with Wasserstein distance The paper's structure is organized into sections covering the SRVT framework, the core theoretical result, the optimization perspective, and technical arguments The organization of the paper is detailed in Section 4. The notation for measures on Polish metric spaces is defined in Notation and standing assumptions The notation for length measure is defined in Definition 2. p. 33 The definition of the Wasserstein-Fisher-Rao distance involves homogeneous couplings and a loss function involving log(cos2) The paper's references span various fields including differential geometry, optimal transport, and nonsmooth optimization
--- Page 1 ---
Towards sparse optimization over convex loops: Equivalence of Square Root Velocity distance and Wasserstein-Fisher-Rao Abstract The Wasserstein–Fisher–Rao (WFR) distance on S2 has recently been shown to coincide with a classical elastic distance between S2-immersions in the theory of Riemannian shape analysis While this correspondence holds in dimension 2, the analogous statement fails in general on S1 and, in the case of convex curves, it cannot be derived from existing twodimensional arguments In this paper, we establish that for convex absolutely continuous immersions of S1 in the plane, the shape distance induced by the square root velocity transformation (SRVT) is indeed equivalent to the WFR distance acting on their associated length measures The proof exploits a monotonicity principle for optimal transport on the universal cover of the circle, which in turn guarantees the existence of an optimal reparametrization achieving the SRVT infimum and enables a one-dimensional unbalanced optimal transport reformulation Motivated by this equivalence, we further investigate the role of sparsity in shape optimization problems formulated in terms of length measures and regularized by the WFR distance Keywords unbalanced optimal transport, elastic shape analysis, nonsmooth optimization sparsity 2020 Mathematics Subject Classification: 49Q22, 49Q10, 52A10, 52A38.
--- Page 6 ---
Definition 2.1 The square root velocity transformation is the function Φ: I → L 2(S1, C-0) given by Φ(c):= c′pc′. p. 5 Note that this is well defined since it’s invariant with respect to translations of c and∥Φ(c)∥ 2 L2 = Z S1 1c' c'squared ds = Z S1 c' ds < ∞ We now define a parametrization-invariant distance on I through the action by precomposition of absolutely continuous reparametrizations: Π(S1):= n φ ∈ W 1,1(S1, S1): φ−1 ∈ W 1,1(S1), φ′(s) > 0 for a.e. s ∈ S1 / ∼. This action defines the quotient space Q:= I/Π(S2). p. 5
--- Page 7 ---
Definition 2.3 Given a loop c ∈ I, we define its length measure µc in M+(S1) as the pushforward of the arc-length measure c' ds via the (tangent) Gauss map Tc: S1 → S1 which is defined by Tc(s):= c′(s)c' ds. Explicitly, µc:= (Tc)-1 c (·)c' ds = Z T−1 c (·)c' ds. The measure µc is called length measure, as it returns the length of the subset of points in Im(c) whose tangent direction lies inside any Borel set U ⊂ S1 However, it is not injective: each measure µ admits infinitely many preimages under this map, see Figure 2 for an example. p. 5 As a result, the measure-theoretic formulation alone is insufficient to capture the full geometric structure of the loops in I. p. 5
--- Page 8 ---
Definition 2.4 We say that c ∈ I is a convex loop if its image is the boundary of some convex set in C. p. 6 We denote by □I the space of such curves, and we assume that they are all positively oriented, meaning that the interior of the convex set lies on the left-hand side of c′ almost everywhere. p. 6 We denote by □Q = □I/Π(S1). p. 3
--- Page 9 ---
Definition 2.7 Given two Borel measures µ0, µ1 ∈ M(S1)⊂ M(C) (inclusion lift), we define the set of homogeneous couplings as Γ(µ0, µ1):= n γ ∈ M 2(C × C): (θi)-(r 2 i γ) = µi, i = 0, 1, where r(x):= x and θ(x) = xx on C - 0 and θ(0) = x for some fixed x ∈ S1. p. 7 We define the Fisher-Rao energy of a coupling γ ∈ Γ(µ0, µ1) as J(γ):= Z C×C x − y squared dγ, and finally the Wasserstein-Fisher-Rao distance as U(µ0, µ1):= inf γ∈Γ(µ0,µ1) p J(γ). p. 7
--- Page 10 ---
Theorem 2.8 The squared distance U squared is the closed convex envelope of the functional S: M(S1) × M(S1) → R ∪ O where S(µ0, µ1):= (√r0x0 − √r1x12 if µi = r iδx i + ∞ otherwise
--- Page 12 ---
Lemma 3.5 Let c0, c1 ∈ □I be two convex loops, and assume that cj is degenerate of the type µj = rδxj + r-xj. p. 2 There exists a sequence x k j → xj for k → ∞ defining µk j:= rδx k j + rdelta-x k j such that µk j, µi are weakly admissible for all k
--- Page 13 ---
Lemma 3.6 Let c ∈ I and φ ∈ Π(S1), and choose lifts c ∈ P and φ in Π(R). p. 5 Then: 1. c◦φ is a lift of c◦φ; 2. p. 7 φ◦ q = q◦ φ
--- Page 14 ---
Lemma 3.7 Let c ∈ □I be a convex loop, and let Tc = c'/c' denote its Gauss map. Then there exists a unique left continuous and non-decreasing function T c: R → R such that q◦ T c = Tc◦ q, T c(0) ∈ [0, 2π), and T c has jumps of size at most π. p. 6 Moreover, for any α ∈ R we have q n µ αc = µc, where µ αc:= (T c)-1 c' dt[α, α + 2π)
--- Page 15 ---
Lemma 3.10 Let c0, c1 ∈ □I be convex loops with length measures µ0, µ1, and σi ∈ L 1+(S1, µi) be two Borel functions defining probability measures σiµi. p. 8 For any coupling η ∈ Γb(σ0µ0,σ1µ1), there exists a coupling ν ∈ Γb(ψ0c′ 0ds, ψ1c'1ds) such that η = (Tc0 × Tc1◦φ)n
--- Page 16 ---
Theorem 3.15 Let c0, c1 ∈ □I, and let µi:= µci be the corresponding length measures. p. 9 Then it holds dSRVT(c0, c1) = U(µ0, µ1)
--- Page 27 ---
References [1] Martin Bauer, Nicolas Charon, Eric Klassen, Sebastian Kurtek, Tom Needham, and Thomas Pierron. Elastic metrics on spaces of euclidean curves: Theory and algorithms. p. 0 p. 0 p. 0 Journal of Nonlinear Science, 34(3):56, 2024 [2] Martin Bauer, Emmanuel Hartman, and Eric Klassen. The square root normal field distance and unbalanced optimal transport. p. 1 p. 1 p. 1 Appl. Math. Optim., 85(3):Paper No. p. 0 p. 0 p. 0 22, 40, 2022 [3] Cl´ement Bonet, Paul Berg, Nicolas Courty, Fran¸cois Septier, Lucas Drumetz, and Minh Tan Pham. Spherical sliced-Wasserstein. p. 0 p. 0 p. 0 In The Eleventh International Conference on Learning Representations, 2023 [4] Kristian Bredies, Marcello Carioni, and Silvio Fanzon. A superposition principle for the inhomogeneous continuity equation with Hellinger-Kantorovich-regular coefficients. p. 1 p. 1 p. 1 Commun. Partial Differ. Equations, 47(10):2023–2069, 2023 [5] Kristian Bredies, Marcello Carioni, and Silvio Fanzon, and Francisco Romero. On the extremal points of the ball of the Benamou–Brenier energy. p. 1 p. 1 p. 1 Bull. Lond. Math. Soc., 53(5):1436–1452, 2021 [6] Kristian Bredies, Marcello Carioni, Silvio Fanzon, and Daniel Walter. Asymptotic linear convergence of fully-corrective generalized conditional gradient methods for composite optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [8] Martins Bruveris. p. 1 p. 1 p. 1 Optimal reparametrizations in the square root velocity framework. p. 3 p. 3 p. 3 SIAM J. Math. Anal., 48(6):4335–4354, 2016 [9] Marcello Carioni and Leonardo Del Grande. A general theory for exact sparse representation recovery in convex optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [10] Marcello Carioni, Jos´e A. p. 1 p. 1 p. 1 Iglesias, and Daniel Walter. p. 0 p. 0 p. 0 Extremal Points and Sparse Optimization for Generalized Kantorovich–Rubinstein Norms
--- Page 26 ---
[1] Martin Bauer, Nicolas Charon, Eric Klassen, Sebastian Kurtek, Tom Needham, and Thomas Pierron. Elastic metrics on spaces of euclidean curves: Theory and algorithms. Journal of Nonlinear Science, 34(3):56, 2024 [2] Martin Bauer, Emmanuel Hartman, and Eric Klassen. The square root normal field distance and unbalanced optimal transport. Appl. Math. Optim., 85(3):Paper No. 22, 40, 2022 [3] Cl´ement Bonet, Paul Berg, Nicolas Courty, Fran¸cois Septier, Lucas Drumetz, and Minh Tan Pham. Spherical sliced-Wasserstein. In The Eleventh International Conference on Learning Representations, 2023 [4] Kristian Bredies, Marcello Carioni, and Silvio Fanzon. A superposition principle for the inhomogeneous continuity equation with Hellinger-Kantorovich-regular coefficients. Commun. Partial Differ. Equations, 47(10):2023–2069, 2023 [5] Kristian Bredies, Marcello Carioni, and Silvio Fanzon, and Francisco Romero. On the extremal points of the ball of the Benamou–Brenier energy. Bull. Lond. Math. Soc., 53(5):1436–1452, 2021 [6] Kristian Bredies, Marcello Carioni, Silvio Fanzon, and Daniel Walter. Asymptotic linear convergence of fully-corrective generalized conditional gradient methods for composite optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [7] Kristian Bredies, Marcello Carioni, Silvio Fanzon, and Daniel Walter. p. 1 p. 1 Asymptotic linear convergence of fully-corrective generalized conditional gradient methods for composite optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [8] Martins Bruveris. Optimal reparametrizations in the square root velocity framework. SIAM J. Math. Anal., 48(6):4335–4354 [9] Marcello Carioni and Leonardo Del Grande. A general theory for exact sparse representation recovery in convex optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [10] Marcello Carioni, Jos´e A. Iglesias, and Daniel Walter. Extremal Points and Sparse Optimization for Generalized Kantorovich–Rubinstein Norms
--- Page 28 ---
[1] Martin Bauer, Nicolas Charon, Eric Klassen, Sebastian Kurtek, Tom Needham, and Thomas Pierron. Elastic metrics on spaces of euclidean curves: Theory and algorithms. Journal of Nonlinear Science, 34(3):56, 2024 [2] Martin Bauer, Emmanuel Hartman, and Eric Klassen. The square root normal field distance and unbalanced optimal transport. Appl. Math. Optim., 85(3):Paper No. 22, 40, 2022 [3] Cl´ement Bonet, Paul Berg, Nicolas Courty, Fran¸cois Septier, Lucas Drumetz, and Minh Tan Pham. Spherical sliced-Wasserstein. In The Eleventh International Conference on Learning Representations, 2023 [4] Kristian Bredies, Marcello Carioni, and Silvio Fanzon. A superposition principle for the inhomogeneous continuity equation with Hellinger-Kantorovich-regular coefficients. Commun. Partial Differ. Equations, 47(10):2023–2069, 2023 [5] Kristian Bredies, Marcello Carioni, and Silvio Fanzon, and Francisco Romero. On the extremal points of the ball of the Benamou–Brenier energy. Bull. Lond. Math. Soc., 53(5):1436–1452, 2021 [6] Kristian Bredies, Marcello Carioni, Silvio Fanzon, and Daniel Walter. Asymptotic linear convergence of fully-corrective generalized conditional gradient methods for composite optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [7] Kristian Bredies, Marcello Carioni, Silvio Fanzon, and Daniel Walter. Asymptotic linear convergence of fully-corrective generalized conditional gradient methods for composite optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [8] Martins Bruveris. Optimal reparametrizations in the square root velocity framework. SIAM J. Math. Anal., 48(6):4335–4354 [9] Marcello Carioni and Leonardo Del Grande. A general theory for exact sparse representation recovery in convex optimization with the nonsmooth term of the cost to be one-homogeneous (often a norm) [10] Marcello Carioni, Jos´e A. Iglesias, and Daniel Walter.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Evaluation of shape matching algorithms: The system can now perform shape matching using a square root velocity transformation distance which is equivalent to Wasserstein-Fisher-Rao distance on length measures for convex curves of S1 in the plane. This allows for more efficient computation and provides a rigorous framework for comparing shapes modulo reparametrization.
-
Sparse optimization for shape reconstruction: The system can solve linear optimization problems over Wasserstein-Fisher-Rao balls with finitely supported reference measures, leading to finiteness results and the proposal of a convex, positively one-homogeneous regularizer suitable for conditional gradient algorithms. This enables the reconstruction of convex loops by progressively adding vertices.
-
Inverse problem regularization: The system can utilize the equivalence between SRVT distance and WFR distance on length measures to provide an alternative formulation of elastic Tikhonov regularization directly at the level of measures, promoting sparsity-promoting methods and convex optimization techniques difficult to access in classical manifold-based frameworks.
Sources
- A General Theory for Exact Sparse Representation Recovery in Convex Optimization
- Linear convergence of a one-cut conditional gradient method for total variation regularization
- An algorithm to reconstruct convex polyhedra from their face normals and areas
Related papers
- An Accelerated Proximal Gradient Method for Bures-Wasserstein Space
- FOMTools: Development and AI-Assisted Code Generation for First-Order Methods
- Lions and Muons: Optimization via Stochastic Frank-Wolfe under Heavy-Tailed Noise
- Dynamic Service Recommendation with Congestion-Dependent Joining: Near-Optimal and Constant-Factor Approximation Algorithms
- Integrating Learning and Optimization for Open Transport Capacity Pooling under Disruption
- Zeroth-Order Gradient Estimators under Decision-Dependent Noise