🔢 mathematics

A Conservation Law for Equilibrium Propagation and Coupled Learning

This paper demonstrates that coupled learning and equilibrium propagation conserve a mass-like quantity in trainable parameters within the continuous-time, small-nudging limit, a property that ensures reliable convergence in linear circuits and offers practical implications for physical learning systems.

Original authors: Joshua A. McGinnis, Adam G. Kline, Yoichiro Mori

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joshua A. McGinnis, Adam G. Kline, Yoichiro Mori

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 physical machine (like a network of electrical wires) to solve a puzzle. You want the machine to settle into a specific "resting state" that gives the correct answer. To do this, you use a learning method where the machine compares two states: how it behaves normally, and how it behaves when you gently nudge it to try to get the right answer.

This paper, written by McGinnis, Kline, and Mori, discovers a hidden "law of physics" that governs how these machines learn. Here is the breakdown in simple terms:

1. The Two Learning Methods

The paper looks at two ways to teach these machines: Equilibrium Propagation (EP) and Coupled Learning (CL).

The Analogy: Imagine a ball rolling in a bowl. The bottom of the bowl is the "correct answer."
- Normal State: The ball rolls to the natural bottom of the bowl.
- Nudged State: You slightly tilt the bowl (the "nudge") to see how the ball moves toward a different spot.
- The Lesson: The machine learns by comparing where the ball went in the normal bowl versus the tilted bowl. It adjusts the shape of the bowl (the "parameters") to make the ball land in the right spot next time.

2. The Big Discovery: The "Mass" Conservation Law

The authors found that in many of these physical systems, there is a quantity that never changes during the learning process, provided the "nudge" is very small. They call this quantity "Mass."

The Analogy: Imagine the machine's adjustable parts (like the resistance in a wire) are like weights on a scale.
- As the machine learns, some weights get heavier and some get lighter.
- However, the total weight on the scale stays exactly the same.
- If one wire's resistance doubles, another must shrink to compensate so the "total mass" remains constant.

Why does this matter?
In math and physics, when something is "conserved" (like energy or mass), it acts like a guardrail. It prevents the system from going crazy. The paper proves that because this "Mass" is conserved, the learning process is forced to settle down and find a solution, rather than wandering off forever or crashing.

3. When Does This Law Apply?

The law works best in linear circuits (systems where the relationship between electricity and voltage is straightforward, like simple resistors).

The "Separable" Rule: The math works if the energy of the system can be split into independent parts. Think of it like a choir where every singer's volume is controlled by their own knob, but they all sing the same song. If the knobs and the song are "separable," the conservation law holds.
When it breaks: If the system is too messy or "non-linear" (like a diode that acts weirdly depending on voltage), or if you have fixed inputs that don't follow the rules, the "Mass" might not be conserved.

4. The Danger of "Drifting"

The paper warns about what happens if you break this conservation law (for example, by using a "nudge" that is too big or using a specific type of update rule).

The Analogy: Imagine you are trying to balance a stack of books. If you have a rule that says "the total height must stay the same," you can rearrange the books safely.
The Problem: If you lose that rule, the stack might start to grow infinitely tall or shrink until it disappears.
Real-world consequence: In a physical computer chip, if the "weights" (parameters) grow too big, they might hit the physical limits of the hardware and break. If they shrink to zero, the connection is lost, and the machine stops working. The paper shows that using a specific "symmetric" nudge (pushing and pulling equally) keeps the "Mass" safe and prevents this drift.

5. Proving It Works

The authors didn't just guess this; they proved it mathematically for specific types of networks:

Crossbar Arrays: These are grids of wires (like a spreadsheet) often used in new types of computer chips. The paper proves that if you use these conservation rules, the machine will exponentially fast find the correct answer.
Single Path Circuits: For simpler circuits with one input and one output, they proved the machine will either find the answer or a wire will break (reach zero). They argue that in most practical cases, the machine finds the answer before the wires break.

Summary

This paper is like finding a new rule of nature for how physical computers learn. It shows that if you teach these machines using a gentle, balanced method, there is an invisible "budget" (Mass) that stays constant. This budget acts as a safety net, ensuring the machine learns reliably and doesn't crash or drift off into uselessness. This is a big deal for building future computers that learn like biological brains but use electricity instead of neurons.

Technical Summary: A Conservation Law for Equilibrium Propagation and Coupled Learning

Problem Statement
Physical learning methods, specifically Equilibrium Propagation (EP) and Coupled Learning (CL), offer a promising avenue for low-power machine learning by utilizing physical systems (such as resistive circuits) to perform inference and training. These methods rely on contrastive local learning, where updates are driven by the difference between two system states: a "free" state and a "nudged" or "clamped" state. However, the theoretical analysis of these methods faces significant challenges. The learning dynamics involve a bilevel optimization problem where the steady-state is implicitly determined by the trainable parameters, often leading to non-convex loss landscapes. Furthermore, the continuous-time limit of these dynamics, particularly in the small-nudging regime ( $\eta \to 0$ ), lacks a unified theoretical framework to guarantee convergence or explain the stability of the training process. There is a specific need to understand whether these dynamics possess inherent structural properties that ensure reliable convergence, particularly for linear circuits, and how practical implementation choices (such as the magnitude of the nudge) affect the stability of the trainable parameters.

Methodology
The authors establish a general mathematical framework for EP and CL, defining the system via an energy function $G(x; k)$ , where $x$ represents state variables (e.g., node voltages) and $k$ represents trainable parameters (e.g., conductances or resistances). The learning dynamics are analyzed in the continuous-time limit with a small nudging parameter $\eta$ .

The core methodology involves deriving and proving the existence of a conservation law for the trainable parameters. The authors approach this through three distinct proofs:

Direct Calculation for CL: By defining a "mass" function $K$ and analyzing the time derivative of the energy under the specific constraints of Coupled Learning.
Gradient Flow Argument for EP: Demonstrating that EP performs gradient flow on a scale-invariant loss function, implying that the squared radius of the parameter vector remains constant.
Unified KKT-Based Proof: Utilizing Karush-Kuhn-Tucker (KKT) conditions to provide a single proof that applies to both EP and CL, relying on a specific separability identity of the energy function.

The paper then restricts its analysis to linear circuits (specifically resistive networks) to apply these conservation laws to concrete convergence proofs. They examine the dynamics of crossbar arrays and general bipartite circuits, as well as the effects of batched updates and different clamping strategies (asymmetric vs. symmetric).

Key Contributions and Results

Discovery of a Conservation Law: The primary theoretical contribution is the proof that both EP and CL conserve a mass-like quantity $K = \frac{1}{2} \sum f_i(k_i)$ in the continuous-time, small-nudging limit, provided the energy function has a separable structure $G(x; k) = \sum g_i(k_i) \tilde{G}_i(x)$ .
- For linear circuits trained on conductances ( $\kappa$ ), the conserved quantity is the squared norm: $K = \frac{1}{2} \sum \kappa_i^2$ .
- For circuits trained on resistances ( $\rho$ ), the conserved quantity is similarly $K = \frac{1}{2} \sum \rho_i^2$ .
- This conservation holds for a broad range of physically relevant settings, including cases where different edges use different parameterizations (e.g., some conductances, some resistances) as long as the separability condition is met.
Convergence Proofs for Linear Circuits:
- Single Input/Output: The authors prove that for a circuit with a single input and output, the learning dynamics must either converge to a solution satisfying the target or drive at least one conductance to zero (effectively removing the edge).
- Crossbar Arrays: For bipartite crossbar arrays (where inputs and outputs are fully connected), the authors prove exponential convergence of the loss function to zero. This result holds even if some conductances reach zero and are removed during training, provided the input and output remain connected.
Analysis of Nudging and Drift:
- The paper demonstrates that the choice of nudging parameter $\eta$ is critical. In the limit $\eta \to 0$ (or using symmetric clamping), the conservation law holds, preventing the overall scale of parameters from drifting.
- Conversely, using non-zero asymmetric nudging ( $\eta \neq 0$ ) in batched updates leads to a drift in the total "mass" of the parameters. Specifically, positive $\eta$ causes the total conductance to decay toward zero, while negative $\eta$ causes it to grow unbounded.
- This drift can cause training failure by pushing parameters to physical boundaries (e.g., zero conductance) before a solution is reached, or by causing the network to disconnect.
Physical Implementation Insight: The authors note that for linear circuits, symmetric clamping (comparing a positively nudged state $x_\eta$ with a negatively nudged state $x_{-\eta}$ ) reproduces the exact $\eta \to 0$ dynamics at any finite $\eta$ . This suggests that the conservation law can be achieved in physical hardware without requiring the experimentally difficult task of imposing infinitesimally small perturbations.

Significance and Claims
The paper claims that the identification of this conservation law provides a rigorous theoretical foundation for the reliability of physical learning algorithms. By showing that EP and CL conserve a specific quantity, the authors explain why these methods can achieve reliable convergence in important settings like linear crossbar arrays, despite the non-convex nature of the general optimization problem.

The work highlights a practical trade-off: while non-conserving nudging (standard asymmetric clamping with $\eta \neq 0$ ) might offer benefits like regularization, it introduces a risk of parameter scale drift that can lead to "pathologies" such as network disconnection or failure to converge. The authors suggest that symmetric clamping offers a robust alternative for hardware implementation, ensuring stability without the need for infinitesimal perturbations.

The paper remains modest regarding its scope, noting that the conservation law is proven for the small-nudge limit and linear circuits. It acknowledges that for nonlinear circuits, symmetric clamping only conserves the quantity up to $O(\eta^2)$ , and the conditions for conservation in more complex, general topologies (beyond the separable energy structure) remain an open question. Furthermore, while the authors prove convergence for crossbars, the general case of single-input/single-output circuits still leaves open the possibility of "edge death" (disconnection) in complex topologies, which requires further investigation.