Bi-invariant Geodesic Regression with Data from the Osteoarthritis Initiative

Johannes Schade, Christoph von Tycowicz, Martin Hanik · 2025

Read Bi-invariant Geodesic Regression with Data from the Osteoarthritis Initiative by Johannes Schade, Christoph von Tycowicz, Martin Hanik (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-11826.txt.

Introduction

Johannes Schade and co-authors build symmetry true regression for shape change data in a 2025 paper. Many data sets are streams of transforms, like joint shapes in medicine. Such data live on Lie groups, where plain line fits break down.

Metrics on those groups oft clash with the group law itself. Their method skips metrics and works through an affine connection. The fit follows geodesic paths that honor left and right shifts. A fixed point method computes it with simple derivative steps. Tests ran on fake data and on osteoarthritis initiative scans. Symmetry true fits dodge false pulls from frame choice.

Catalog record

CREATORJohannes Schade, Christoph von Tycowicz, Martin Hanik
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-11826

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Bi-invariant Geodesic Regression with Data from the Osteoarthritis Initiative Johannes Schade, Christoph von Tycowicz, Martin Hanik Abstract Many phenomena are naturally characterized by measuring continuous transformations such as shape changes in medicine or articulated systems in robotics. Modeling the variability in such datasets requires performing statistics on Lie groups, that is, manifolds carrying an additional group structure. As the Lie group captures the symmetries in the data, it is essential from a theoretical and practical perspective to ask for statistical methods that respect these symmetries; this way they are insensitive to confounding effects, e.g., due to the choice of reference coordinate systems. In this work, we investigate geodesic regression—a generalization of linear regression originally derived for Riemannian manifolds. While Lie groups can be endowed with Riemannian metrics, these are generally incompatible with the group structure. We develop a non-metric estimator using an affine connection setting. It captures geodesic relationships respecting the symmetries given by left and right translations. For its computation, we propose an efficient fixed point algorithm requiring simple differential expressions that can be calculated through automatic differentiation. We perform experiments on a synthetic example and evaluate our method on an open-access, clinical dataset studying knee joint configurations under the progression of osteoarthritis. 1 Introduction We begin with the essential mathematical background before introducing the subject of this work: regression for geometric data using geodesic models. A discussion of related work and our contributions is deferred to the end of this introduction. Basics of Lie groups. We briefly review Lie group theory; refer to [ 26 ] for more detail. In the following, “smooth” means “infinitely differentiable”. A Lie group G is a smooth manifold that has a compatible group structure; that is, it comes with a smooth, associative (often not commutative) group operation G\times G\ni(g,h)\mapsto gh\in G with identity element e and a smooth inversion map G\ni g\mapsto g^{-1} . The group operation defines two automorphisms for each g in a general Lie group G : the left and right translation \mathrm{L}_{h}:g\mapsto hg and \mathrm{R}_{h}:g\mapsto gh . Their derivatives d_{g}\mathrm{L}_{h} and d_{g}\mathrm{R}_{h} map tangent vectors v\in T_{g}G bijectively to the tangent spaces T_{hg}G and T_{gh}G at hg and gh , respectively. Each v\in T_{e}G defines a left-invariant vector field X by X_{g}=\textnormal{d}_{e}\mathrm{L}_{g}(v) for all g\in G . Right invariant vector fields are constructed analogously. The integral curve \alpha_{v}:\mathbb{R}\to G of a (left or right) invariant vector field X with v=X_{e} determines a 1-parameter subgroup of G through e . The group exponential \exp is then defined by \exp(v):=\alpha_{v}(1) ; being a diffeomorphism around e , it has a local inverse: the group logarithm log. Given two vector fields X,Y on G , a so-called connection \nabla enables differentiation of Y along X , yielding a vector field \nabla_{X}Y . With \gamma^{\prime}:=\frac{\scriptstyle{\textnormal{d}}\gamma}{\scriptstyle{\textnormal{d}}t} , a geodesic \gamma:[0,1]\to G is defined by \nabla_{\gamma^{\prime}}\gamma^{\prime}=0 , creating a curve without acceleration. Every f\in G has a so-called normal convex neighborhood U , in which any two points f,h\in U can be connected by a unique geodesic [0,1]\ni t\mapsto\gamma(t;f,h) that lies within U . With v:=\gamma^{\prime}(0;g,h) , the connection exponential \Exp_{g}:T_{g}G\to G at g is defined by \Exp_{g}(v):=\gamma(1;g,h) ; it is also a local diffeomorphism with the connection logarithm \Log_{g}(h)=\gamma^{\prime}(0;g,h) as its inverse. These maps are called Riemannian exponential and logarithm , respectively, if \nabla is the Levi-Civita connection of a Riemannian metric. Translated group and Riemannian maps coincide if a bi-invariant Riemannian metric (that is, one that remains invariant under both left and right translations) is used. Unfortunately, many Lie groups do not allow for such a metric [ 23 ] . If, on the other hand, \nabla is the canonical Cartan-Schouten (CCS) connection [ 8 , 26 ] , then geodesics and left (or right) translated 1-parameter subgroups always coincide. More precisely, there is a maximal normal convex neighborhood U for every g\in G such that [ 25 , Cor. 5.1] \displaystyle\Exp_{g}(v) \displaystyle=g\exp\left(\textnormal{d}_{g}\mathrm{L}_{g^{-1}}(v)\right)=\exp\left(\textnormal{d}_{g}\mathrm{R}_{g^{-1}}(v)\right)g,\quad v\in T_{g}G, \displaystyle\Log_{g}(f) \displaystyle=\textnormal{d}_{e}\mathrm{L}_{g}\big(\log(g^{-1}f)\big)=\textnormal{d}_{e}\mathrm{R}_{g}\big(\log(fg^{-1})\big),\quad f\in U. (1) Geodesics in U are then of the form t\mapsto\gamma(t;g,f)=\Exp_{g}(t\Log_{g}(f))=g\exp(t\log(g^{-1}f))=\exp(t\log(fg^{-1}))g. (2) This fact makes the CCS connection the canonical connection on G . From the above, we have two equivalent ways to parametrize geodesics in U : We can use boundary values (g,f) or initial values (g,v) , which are connected by (g,f)\leftrightarrow(g,\Log_{g}(f)) . We will use the former convention in this work, as its inherent symmetry simplifies, e.g., the implementation of derivatives of geodesics with respect to their parameters. Particularly, estimators of regression problems will be defined by the start and endpoint of the optimal geodesic. Riemannian geodesic regression. Geodesic regression was proposed for Riemannian manifolds by Fletcher in [ 13 ] . We now summarize the approach for Lie groups, adapting our parametrization of geodesics with initial values for consistency of exposition. For Riemannian geodesic regression, the Lie group G must have a Riemannian metric—a smoothly varying inner product on the tangent spaces. The metric determines the corresponding distance function \dist and Levi-Civita connection of G ; the latter defines the geodesics. For a G -valued random variable Y and a non-random, real-valued variable t , the generalization of the linear regression model to the manifold setting is the geodesic model Y=\Exp_{\gamma(t;g_{0},g_{1})}(\epsilon), (3) where \epsilon is a random variable taking values in the tangent space at \gamma(t;g_{0},g_{1}) . The points g_{0},g_{1}\in G are the model’s parameters that must be estimated. Let U\subset G be a normal convex neighborhood. Without loss of generality, we assume in this work that the variable t takes values in the interval [0,1] . Given data (f_{1},t_{1}),\dots,(f_{N},t_{N})\in U\times[0,1] , the least-squares estimator \vartheta^{LS}:(U\times[0,1])^{N}\to U^{2} is the minimizer of the sum-of-squared error \vartheta^{LS}\left((f_{i},t_{i})_{i=1}^{N}\right)=\argmin_{(g_{0},g_{1})\in U^{2}}E(g_{0},g_{1}):=\argmin_{(g_{0},g_{1})\in U^{2}}\frac{1}{2}\sum_{i=1}^{N}\dist^{2}\big(\gamma(t_{i};g_{0},g_{1}),f_{i}\big). (4) It can be computed using gradient descent. Denoting the adjoint operator with respect to the Riemannian metric by “ * ”, the gradients of E with respect to g_{0} and g_{1} are given by [ 17 , Section 3] \displaystyle\textnormal{grad}_{g_{0}}E(\;\cdot\;,g_{1}) \displaystyle=-\sum_{i=1}^{N}\big(\textnormal{d}_{g_{0}}\gamma(t_{i};\;\cdot\;,g_{1})\big)^{*}\Log_{\gamma(t_{i})}(f_{i})\qquad\textnormal{and} (5) \displaystyle\textnormal{grad}_{g_{1}}E(g_{0},\;\cdot\;) \displaystyle=-\sum_{i=1}^{N}\big(\textnormal{d}_{g_{1}}\gamma(t_{i};g_{0},\;\cdot\;)\big)^{*}\Log_{\gamma(t_{i})}(f_{i}), (6) respectively. 1 1 1 The adjoint operators in ( 5 ) and ( 6 ) are given in terms of so-called Jacobi fields; see [ 15 , Section 5.1] for information on when and how they can be computed explicitly. Thus, the desired optima in ( 4 ) are roots to both ( 5 ) and ( 6 ). The problem with Riemannian geodesic regression in Lie groups. A natural property we can expect from an estimator \vartheta:(U\times[0,1])^{N}\to U^{2} for ( 3 ) is that it is equivariant under joint translations of the data; that is, for every h\in G , \displaystyle\vartheta\left(\big(\mathrm{T}_{h}(f_{i}),t_{i}\big)_{i=1}^{N}\right) \displaystyle=\mathrm{T}_{h}\Big(\vartheta\left((f_{i},t_{i})_{i=1}^{N}\right)\Big),\qquad\mathrm{T}_{h}\in\{\mathrm{L}_{h},\mathrm{R}_{h}\}. Indeed, translations are the symmetries of G , and, intuitively, the solution to a symmetry-transformed version of the regression problem should be the symmetry-transformed original solution. The following example shows that equivariance under translations trivially holds for ordinary linear regression. Example 1 Let M=\mathbb{R}^{n} with the Euclidean metric. It is also a Lie group with vector addition as the group operation. We have \mathrm{L}_{z}(x)=\mathrm{R}_{z}(x)=x+z , so the Euclidean metric is bi-invariant. Consequently, the CCS connection coincides with the Levi-Civita connection, which is just the ordinary directional derivative. With the identification T_{x}\mathbb{R}^{n}\cong\mathbb{R}^{n} , we have \Exp_{x}(v)=x+v , implying that geodesics are straight lines. With this, we see that ( 4 )) is the standard least-squares

Read the complete text