Lagrangian-based Equilibrium Propagation: generalisation to arbitrary boundary conditions & equivalence with Hamiltonian Echo Learning
Pith reviewed 2026-05-19 10:22 UTC · model grok-4.3
The pith
Generalized Lagrangian Equilibrium Propagation extends the variational formulation of Equilibrium Propagation to entire trajectories of time-varying inputs, yielding a family of algorithms of which only Hamiltonian Echo Learning preserves a
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
By formulating equilibrium propagation in terms of a Lagrangian that accounts for entire trajectories rather than static equilibria, and by allowing different boundary conditions at the start and end of the trajectory, the method yields a family of learning algorithms for energy-based models. Hamiltonian Echo Learning corresponds to the specific choice of boundary conditions that produces a forward-only procedure with local parameter updates, and it is the only such instance that retains the implementation advantages of the original equilibrium propagation.
What carries the argument
Generalized Lagrangian Equilibrium Propagation, obtained by varying a Lagrangian functional defined over system trajectories subject to chosen boundary conditions, from which the parameter update rule follows directly.
If this is right
- Different boundary conditions in GLEP produce distinct learning algorithms for energy-based models driven by time-varying inputs.
- Most choices of boundary conditions result in algorithms that require non-local or backward information and are therefore impractical.
- Hamiltonian Echo Learning is recovered exactly as the special case that operates forward-only with only two or more passes through the system.
- This special case inherits local learning and scales independently of model size.
- Recurrent HEL and Hamiltonian Echo Backpropagation appear as instances inside the same family.
Where Pith is reading between the lines
- The unification suggests that other boundary conditions might be approximated in software to test whether they can match HEL performance without hardware constraints.
- Physical systems governed by Hamiltonian dynamics could be directly configured to implement the HEL case of GLEP for on-device learning with time series.
- The trajectory-based variational view may extend to continuous-time models or stochastic dynamics beyond the deterministic cases treated here.
Load-bearing premise
The variational description of the energy-based model must extend from fixed points to entire trajectories while preserving the same local-update and forward-only character under at least one choice of boundary conditions.
What would settle it
An explicit calculation of the parameter gradient obtained from the Lagrangian under the boundary conditions claimed to recover Hamiltonian Echo Learning that fails to match the updates required to reduce the loss on a trajectory would disprove the claimed equivalence.
read the original abstract
Equilibrium Propagation (EP) is a learning algorithm for training Energy-based Models (EBMs) on static inputs which leverages the variational description of their fixed points. Extending EP to time-varying inputs is a challenging problem, as the variational description must apply to the entire system trajectory rather than just fixed points, and careful consideration of boundary conditions becomes essential. In this work, we present Generalized Lagrangian Equilibrium Propagation (GLEP), which extends the variational formulation of EP to time-varying inputs. We demonstrate that GLEP yields different learning algorithms depending on the boundary conditions of the system, many of which are impractical for implementation. We then show that Hamiltonian Echo Learning (HEL) -- which includes the recently proposed Recurrent HEL (RHEL) and the earlier known Hamiltonian Echo Backpropagation (HEB) algorithms -- can be derived as a special case of GLEP. Notably, HEL is the only instance of GLEP we found that inherits the properties that make EP a desirable alternative to backpropagation for hardware implementations: it operates in a "forward-only" manner (i.e. using the same system for both inference and learning), it scales efficiently (requiring only two or more passes through the system regardless of model size), and enables local learning.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces Generalized Lagrangian Equilibrium Propagation (GLEP) as an extension of Equilibrium Propagation (EP) to time-varying inputs. It formulates a variational principle over system trajectories rather than fixed points and shows that the resulting learning rules depend on the choice of boundary conditions, with many such choices yielding impractical algorithms. The central result is that Hamiltonian Echo Learning (HEL, including RHEL and HEB) arises as a special case of GLEP and is the only variant found that remains forward-only, requires a constant number of passes, and supports local updates suitable for hardware.
Significance. If the equivalence derivations are correct, the work supplies a single Lagrangian framework that classifies EP-style algorithms by boundary conditions and explains why HEL inherits the hardware-friendly properties of classical EP. This could guide the systematic design of energy-based learning rules for physical substrates and unify several recently proposed echo-based methods under one variational umbrella.
major comments (1)
- [Abstract] Abstract: the assertion that HEL 'is the only instance of GLEP we found' that preserves forward-only, O(1)-pass, and local character is load-bearing for the paper's main contribution yet rests on the authors having examined a limited set of boundary conditions. Because boundary conditions enter the variational principle continuously (via weights on initial/final states or adjoint variables), an exhaustive classification or parameterization is required to substantiate uniqueness; without it the claim remains an empirical observation rather than a derived result.
minor comments (2)
- Provide explicit equations for the trajectory-level Lagrangian and the resulting Euler-Lagrange conditions under at least two contrasting boundary conditions (e.g., fixed initial state versus fixed final state) so that readers can verify how locality and forward-only character are lost or retained.
- Clarify whether the 'two or more passes' complexity for HEL is independent of trajectory length or only of network depth; the current wording leaves this ambiguous for recurrent or continuous-time implementations.
Simulated Author's Rebuttal
We thank the referee for their constructive review and for identifying a key point regarding the strength of the uniqueness claim. We address the major comment below and will revise the manuscript accordingly to clarify the scope of our analysis of boundary conditions.
read point-by-point responses
-
Referee: [Abstract] Abstract: the assertion that HEL 'is the only instance of GLEP we found' that preserves forward-only, O(1)-pass, and local character is load-bearing for the paper's main contribution yet rests on the authors having examined a limited set of boundary conditions. Because boundary conditions enter the variational principle continuously (via weights on initial/final states or adjoint variables), an exhaustive classification or parameterization is required to substantiate uniqueness; without it the claim remains an empirical observation rather than a derived result.
Authors: We appreciate the referee's careful reading and agree that the phrasing in the abstract merits clarification to avoid any implication of a fully derived uniqueness result. The manuscript already qualifies the statement with the phrase 'the only instance of GLEP we found,' which reflects that the claim is based on the concrete boundary conditions we explicitly analyzed. These include fixed initial and final states, free boundary conditions at one or both ends, and conditions that incorporate adjoint variables or weighted penalties on the initial and final states in the action integral. For each of these, we derive the associated learning rules and demonstrate that they generally produce non-local updates, require a number of passes that scales with depth or time horizon, or violate the forward-only constraint. Only the specific boundary-condition choice that recovers Hamiltonian Echo Learning (including its RHEL and HEB variants) simultaneously satisfies forward-only operation, constant-pass scaling, and locality. We acknowledge that boundary conditions enter the variational principle in a continuous manner through weighting parameters, and that a complete parameterization of the entire function space is not supplied. In the revised version we will add a short subsection that introduces a general parameterization of the boundary terms (via scalar weights on the initial/final contributions and possible adjoint constraints) and explains why, among the discrete, physically motivated choices that preserve the underlying Lagrangian structure, only the HEL case yields the desired hardware properties. This revision will make the empirical nature of the observation explicit while keeping the paper's focus on the derivation of HEL as a special case of GLEP. We do not believe a re revision: yes
Circularity Check
GLEP-to-HEL derivation is independent; uniqueness claim is empirical sampling rather than definitional
full rationale
The paper extends the variational fixed-point description of EP to full trajectories via a Lagrangian formulation (GLEP) indexed by boundary conditions. It then derives HEL (including RHEL and HEB) as the special case obtained under particular choices of those boundary conditions. This step is presented as an explicit reduction from the general Lagrangian equations rather than a renaming or self-referential fit. The assertion that HEL is the only practical instance is qualified as 'the only instance we found,' which is an empirical observation over sampled boundary conditions and does not reduce the central derivation to a tautology or to a load-bearing self-citation. No equation is shown to equal its own input by construction, and the work remains self-contained against external benchmarks for the generalization itself.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Energy-based models possess a variational description of their fixed points that can be extended to entire trajectories under suitable boundary conditions.
Lean theorems connected to this paper
-
IndisputableMonolith/Cost/FunctionalEquation.leanwashburn_uniqueness_aczel unclear?
unclearRelation between the paper passage and the cited Recognition theorem.
We introduce Generalized Lagrangian Equilibrium Propagation (GLEP), which extends the variational formulation of EP to time-varying inputs... different boundary condition choices in GLEP yield distinct learning algorithms
What do these tags mean?
- matches
- The paper's claim is directly supported by a theorem in the formal canon.
- supports
- The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
- extends
- The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
- uses
- The paper appears to rely on the theorem as machinery.
- contradicts
- The paper's claim conflicts with a theorem or certificate in the canon.
- unclear
- Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.