🤖 machine learning

Beyond Rigid Geometries: The Spline-Pullback Metric for Universal Diffeomorphic SPD Representation Learning

This paper introduces the Spline-Pullback Metric (SPM), a novel framework that replaces rigid, fixed Riemannian metrics with a learnable, rank-invariant B-spline parameterization to achieve universal diffeomorphic SPD representation learning, thereby overcoming gradient instabilities and geometric limitations while delivering state-of-the-art performance across diverse deep learning architectures.

Original authors: Tushar Das, Subrata Dutta, Sarmistha Neogy, Koushlendra Kumar Singh

Published 2026-05-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tushar Das, Subrata Dutta, Sarmistha Neogy, Koushlendra Kumar Singh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to understand complex shapes, like the patterns in a radar signal or the movement of a human hand. These shapes are represented mathematically by special grids of numbers called SPD matrices.

For a long time, computers struggled to learn from these shapes because they were trying to force them into a rigid, straight-line world (Euclidean geometry). It's like trying to measure the distance between two cities on a flat map when the Earth is actually a sphere; the math gets messy, and the computer gets confused.

To fix this, scientists invented "curved" ways to measure these shapes, called Riemannian metrics. But here's the problem: for decades, researchers had to pick one specific "curved ruler" by hand and stick with it. It was like a tailor making a suit using only one fixed pattern size. If the data didn't fit that pattern perfectly, the computer's performance suffered.

The Problem with the Old Rulers

The paper argues that these old, fixed rulers are too stiff.

The "One-Size-Fits-All" Trap: Some rulers work great for small numbers but break for big ones. Others work well for smooth data but get confused by noise.
The "Glitch" Risk: Some newer attempts to make these rulers flexible used math tricks that caused the computer to "fold" the data onto itself or lose information, leading to calculation errors (like dividing by zero).

The New Solution: The "Spline-Pullback Metric" (SPM)

The authors introduce a new tool called the Spline-Pullback Metric (SPM).

Think of the old metrics as a rigid metal ruler. No matter how much you try to bend it, it stays straight or snaps.
The new SPM is like a smart, stretchy rubber band made of a special material called a "B-spline."

Here is how it works in simple terms:

It Learns to Bend: Instead of a fixed shape, the SPM is a flexible curve that the computer learns to shape itself. It looks at the data and asks, "Where is the noise? Where is the important signal?"
The "Squish and Stretch" Trick: Imagine you have a pile of marbles (data points) mixed with some trash (noise).
- Old rulers treat all marbles the same.
- The SPM learns to squish the trash marbles together so they look like a tiny, flat blob, while stretching the important marbles far apart so they are easy to tell apart. It does this by changing its shape dynamically, just like a master tailor adjusting a suit to fit a specific person perfectly.
No More Glitches: The authors proved mathematically that their rubber band never snaps, never folds over itself, and never gets stuck. It always remains a smooth, continuous path, ensuring the computer can always calculate the next step without crashing.

Two Versions of the Tool

The paper offers two ways to use this flexible ruler, depending on the job:

S-SPM (Spectral): This version looks at the "inner frequencies" of the data. It's like tuning a radio to find the clearest signal. It's great for data where the order of sensors doesn't matter.
C-SPM (Cholesky): This version looks at the data in a specific, hierarchical order (like a family tree). It's faster and more efficient for data that has a clear structure, like radar arrays or skeletal movements.

The Results: Why It Matters

The authors tested this new tool on three different real-world challenges:

Human Movement (HDM05): Recognizing actions from body movements.
Hand Gestures (FPHA): Understanding what a person is doing with their hands.
Radar Signals: Distinguishing between different types of radar echoes.

In every single test, the SPM outperformed the old, rigid rulers and even the newer, slightly flexible ones. It achieved the highest accuracy, proving that a "learnable" geometry is better than a "hand-crafted" one.

The Big Picture

The paper claims that this is a major shift in how we teach computers to handle complex shapes. Instead of forcing the data to fit our mathematical rules, we are now giving the computer a flexible tool (the Spline) to learn the best possible shape for the data itself. It's the difference between trying to force a square peg into a round hole versus having a tool that can reshape the hole to fit the peg perfectly.

In short: The paper introduces a new, flexible mathematical ruler that learns to stretch and squeeze data to make it easier for computers to understand, resulting in smarter and more accurate AI for tasks like motion tracking and radar analysis.

Technical Summary: The Spline-Pullback Metric for Universal Diffeomorphic SPD Representation Learning

1. Problem Statement

The integration of Symmetric Positive Definite (SPD) matrices into deep learning has historically relied on fixed, hand-crafted Riemannian metrics (e.g., Affine-Invariant, Log-Euclidean, Log-Cholesky). These static formulations impose rigid geometric priors that limit network expressivity and adaptability. While recent attempts to parameterize these geometries have sought to improve expressivity, they often suffer from fundamental mathematical flaws:

Topological Failures: Rank-dependent mappings (e.g., ALEM) violate the definition of Primary Matrix Functions, leading to one-to-many mappings at spectral singularities (repeated eigenvalues) and causing gradient instability.
Lack of Surjectivity: Polynomial power mappings (e.g., PCM) fail to map the SPD manifold globally onto unconstrained Euclidean space. To prevent undefined negative pre-images, they rely on non-injective computational clipping (e.g., max(x, ϵ)), which breaks the bijectivity required for a geodesically complete Riemannian metric.
Gradient Instability: Existing methods often lack $C^1$ smoothness or require computationally heavy layer cascades to approximate diffeomorphisms, leading to gradient collapse at spectral singularities.

The central challenge is transitioning from static mathematical priors to dynamic, computationally efficient, and completely learnable topological representations that preserve the fundamental axioms of differential geometry.

2. Methodology: The Spline-Pullback Metric (SPM)

The authors introduce the Spline-Pullback Metric (SPM), a framework that parameterizes the global diffeomorphism between the SPD manifold and a flat Euclidean space using B-splines. Unlike previous approaches, SPM treats the Riemannian metric tensor itself as an end-to-end learnable representation.

Core Mechanism

SPM relies on a parameterized, strictly monotonic scalar mapping $f_\theta: \mathbb{R}_{>0} \to \mathbb{R}$ .

Logarithmic Domain: To ensure global surjectivity and avoid gradient instability, the mapping operates in the logarithmic domain.
B-Spline Construction: The core of the mapping is a B-spline $S(x)$ of degree $k$ with learnable control coefficients $c_i$ .
Monotonicity Constraint: To guarantee the mapping is a diffeomorphism (bijective and smooth), the control coefficients are constrained to be strictly increasing ( $c_i > c_{i-1}$ ). This is achieved via a cumulative softplus reparameterization: $c_i = c_{i-1} + \text{softplus}(w_i) + \epsilon$ .
Boundary Handling: Outside the knot vector, the mapping extends via bounded linear extrapolation using the boundary derivatives, ensuring the function covers the entire real line.

Instantiations

The paper proposes two specific instantiations of SPM:

Spectral-SPM (S-SPM): Applies the scalar diffeomorphism $f_\theta$ directly to the eigenvalues of the SPD matrix via eigendecomposition ( $S = U\Sigma U^\top \to U \text{diag}(f_\theta(\lambda_i)) U^\top$ ). This allows for Localized Spectral Modelling, where the metric can dynamically compress noise bands and expand signal bands.
Cholesky-SPM (C-SPM): Applies the scalar diffeomorphism only to the strictly positive diagonal elements of the Cholesky factor ( $L$ ), leaving the lower triangular elements as identity mappings. This avoids the $O(n^3)$ cost of eigendecomposition and the associated Daleckĭ-Kreĭn structural gradient, offering better numerical stability and speed for hierarchically ordered data.

Theoretical Guarantees

The authors provide rigorous proofs establishing that:

Global Diffeomorphism: SPM is a globally bijective, $C^1$ -smooth diffeomorphism.
Universal Approximation: SPM is a dense universal approximator for strictly increasing $C^1$ diffeomorphisms over any compact spectral interval. It theoretically subsumes existing pullback metrics (like Log-Euclidean and Power-Cholesky) as special cases.
Geometric Properties: The induced manifold has zero sectional curvature (globally flat), admits a closed-form Fréchet mean, and allows for path-independent parallel transport.
Gradient Stability: To handle spectral singularities ( $\lambda_i = \lambda_j$ ) in the Daleckĭ-Kreĭn gradient matrix, the authors introduce an asymmetric spectral perturbation (adding a small, increasing vector to eigenvalues in Float64 precision) to strictly bound the Lipschitz constant of the backward pass.

3. Key Contributions

Mathematical Formalization: The paper mathematically formalizes and proves that SPM (and its instantiations S-SPM and C-SPM) constitutes a fully learnable, geodesically complete global Riemannian diffeomorphism and a universal approximator.
Topological Solution: It identifies and resolves the topological limitations of current metrics (rank-swapping discontinuities, lack of surjectivity, and gradient collapse) by introducing a rank-invariant, monotonically constrained B-spline approach.
Numerical Stability Protocol: The authors introduce an asymmetric Float64 matrix perturbation protocol to bound the Lipschitz constant of the Daleckĭ-Kreĭn structural gradient, solving the numerical instability of Riemannian backpropagation at spectral singularities.
Empirical Performance: Under a rigorous evaluation protocol, SPM achieves state-of-the-art accuracy across 9 tested combinations of datasets and architectures (Linear Probes, SPDNets, and Deep Riemannian ResNets) using 7 different metrics.

4. Experimental Results

The authors evaluated SPM on three datasets: HDM05 (kinematic action recognition), FPHA (first-person hand actions), and Radar (signal processing).

Synthetic Validation: In a 1D synthetic function approximation task, SPM successfully learned complex, non-linear mappings (including "staircase" functions to separate noise and signal) that rigid classical metrics could not. In an adversarial classification task with alternating eigenvalue bands, S-SPM achieved 100% accuracy, whereas rigid baselines (Log-Euclidean, Log-Cholesky) suffered severe underfitting (63-66%).
Real-World Datasets:
- HDM05: S-SPM achieved the best results across all architectures, notably 0.6801 on RResNet, outperforming the best baseline (ALEM at 0.6777).
- FPHA: S-SPM achieved a significant improvement over the best baseline (PCM) in the SPDNet architecture (0.9043 vs. 0.7357) and on Linear Probes (0.8574 vs. 0.6817).
- Radar: C-SPM achieved the highest Linear Probe accuracy (0.9687) and maintained competitive scores in deep networks, demonstrating superior efficiency and stability for hierarchical data.

5. Significance and Claims

The paper positions SPM as a paradigm shift from "hand-crafted feature selection" in metric learning to "universal geometric approximation."

Paradigm Shift: The authors claim SPM is the first framework to introduce the paradigm of universal Riemannian diffeomorphic metric learning parameterized via B-splines. It enables localized spectral modelling without sacrificing the fundamental axioms of differential geometry.
Explainability: The learned B-spline coefficients offer direct insight into the dataset's spectral structure (e.g., identifying which eigenvalue ranges correspond to noise vs. signal), opening a new frontier in Riemannian Metric Explainable AI (RMXAI).
Future Directions: The paper modestly suggests that while SPM eliminates rigid priors, future work is needed in Riemannian Metric Initialization (RMI) (tailoring initialization distributions to the SPM manifold) and Riemannian Metric Regularization (RMR) (to prevent overfitting in data-scarce environments by regularizing the spline's second derivative).

The authors conclude that by treating the metric tensor as a learnable, parameter-efficient universal approximator, SPM inherits the high expressivity of deep learning, achieving state-of-the-art performance while providing a mathematically sound, stable, and interpretable framework for SPD manifold learning.