Coercivity and Local Convergence of Physical Learning in Linear Circuits

Technical Summary: Coercivity and Local Convergence of Physical Learning in Linear Circuits

Problem Statement
Physical learning methods, such as Equilibrium Propagation (EP) and Coupled Learning (CL), offer a paradigm for training physical networks (e.g., resistor networks) using only local update rules. Unlike backpropagation, which requires global error signal propagation, physical learning exploits the system's own equilibrium dynamics to transfer information. While recent experiments have demonstrated the feasibility of these methods in analog devices, rigorous theoretical guarantees regarding their convergence have been scarce. Existing literature establishes the gradient structure of these methods in the small-nudge limit but often stops short of proving that the resulting dynamics actually drive the system to a trained state, particularly in networks with hidden nodes where convexity is not guaranteed. This paper addresses the question of under what conditions a physical system is guaranteed to learn and what structural features might prevent convergence.

Methodology
The authors provide a local convergence analysis for three physical learning methods—Equilibrium Propagation (EP), Coupled Learning (CL), and a newly introduced method, Adjoint Coupled Learning (AL)—applied to linear circuits. The analysis is conducted in two regimes:

1. Continuous Time: The limit of small learning rates ($\tau \to 0$) and small nudging parameters ($\eta \to 0$).
2. Discrete Time: Finite learning rates and nudging parameters.

The mathematical framework models the physical network as an energy minimization problem where the energy function $G(x; k)$ (dissipated power for linear circuits) is minimized subject to input constraints. The learning process involves comparing a "free state" (relaxed under natural dynamics) and a "nudged state" (perturbed toward a target). The parameter updates are derived from the difference in local quantities (e.g., voltage drops) between these two states.

The core of the methodology involves:

• Deriving the exact loss dynamics for EP, CL, and AL in the small-nudge limit.
• Establishing a coercivity condition, expressed as a rank condition on a matrix constructed from the network's incidence structure and the constrained Laplacian.
• Using a bootstrap argument combining the Implicit Function Theorem (for local estimates) and coercivity estimates (for exponential loss decay) to prove that the system remains within the region where these estimates hold.
• Applying Sard's theorem to analyze the genericity of the coercivity condition on the solution manifold.

Key Contributions

1. Introduction of Adjoint Coupled Learning (AL): The authors propose AL, a method that constrains the output to the target and nudges using the resulting constraint force (adjoint error). They demonstrate that AL preserves the physical simplicity of CL while recovering the exact gradient flow structure of EP.
2. Gradient Flow Characterization:
   • EP performs gradient descent on the natural squared error loss $\Phi = \frac{1}{2}\|r\|^2$.
   • AL performs gradient descent on an adjoint loss $\Phi^* = \frac{1}{2}\|\mu_0\|^2$.
   • CL does not follow a pure gradient flow. Its dynamics include an additional cubic correction term arising from the parameter-dependence of the weighting matrix $D$. The authors show that for CL, the error and nudge must be of the same type (voltage-voltage), whereas EP and AL cross the voltage-current duality to achieve gradient flow.
3. Coercivity Analysis: The paper identifies a necessary and sufficient condition for local exponential convergence: the coercivity condition. This condition requires that the "active-incidence matrix" (built from edges with non-zero voltage drops in the free state) has full column rank when projected through the inverse constrained Laplacian.
   • If coercivity holds, the loss decays exponentially, and parameters converge to the solution manifold.
   • If coercivity fails, the dissipative term in the loss dynamics vanishes, potentially leading to slow convergence or stagnation.
4. Counterexample and Genericity:
   • The authors construct a "kite" circuit counterexample where a specific symmetry (ratio conditions on conductances) causes the free and clamped voltage drops to have disjoint support, leading to a degenerate coercivity constant. This demonstrates that coercivity can fail, creating spurious fixed points or slow convergence zones.
   • However, using Sard's theorem, they prove that such degeneracies are non-generic. For almost every choice of desired output $w$, the solution manifold consists entirely of points where the coercivity condition holds.

Results

• Local Convergence Theorems: The paper proves that if the system starts sufficiently close to a solution and the coercivity condition is satisfied, the training loss decays exponentially, and the parameters converge to the solution manifold (Theorems 12, 13, and 14).
• Discrete Convergence: For discrete updates, the authors show that the loss decays geometrically (linear convergence rate) provided the learning rate and initial loss are sufficiently small. For CL, the smallness condition on the initial loss cannot be relaxed by simply making the nudge $\eta$ small, due to the intrinsic non-gradient cubic term.
• Numerical Validation: Numerical experiments on the kite circuit illustrate the coercivity landscape. They confirm that while the coercivity constant vanishes at specific symmetric points (codimension-2), perturbing the target output $w$ shifts the solution manifold away from these degenerate points, restoring exponential convergence rates.

Significance and Claims
The paper claims to provide the first local convergence analysis for EP, CL, and AL in linear circuits with hidden nodes. Its significance lies in:

• Moving beyond the identification of gradient structures to proving that these dynamics actually drive the system to a trained state.
• Clarifying the structural requirements (coercivity) for learning to succeed and identifying that while failures are possible due to symmetries, they are rare (non-generic) for random target outputs.
• Offering a physical interpretation of learning dynamics, showing that the "dual" nature of the error and nudge (voltage vs. current) is the key to recovering gradient flow.
• Providing a theoretical basis for the stability of physical learning methods, suggesting that for generic tasks, physical systems are guaranteed to learn locally, provided the network structure does not possess specific degeneracies.

The authors remain modest, noting that their results are local (requiring the system to start near a solution) and that global convergence is not guaranteed if the initial condition lies near spurious fixed points off the solution manifold. They do not claim to solve the global convergence problem for all physical learning scenarios but establish the conditions under which local exponential convergence is mathematically assured.