Skip to content

Notation Reference — Action Operators (foundation)

A standalone glossary for the Action Operators foundation: From actions to operators, Augmenting JEPA with Action Operators, and A Gallery of Operators. Every symbol used across those notes, grouped by role. Keep it open in a second tab while reading.

The encoder is written \(E\) here (and \(\varphi_\psi\) in the GRL literature — the same object). The downstream Operator World Models series uses the same conventions.


Representation — the JEPA side

Symbol Meaning
\(x\) a raw input data point — an image, a behavioral-signal window, a sequence
\(s\) the underlying system state in state space \(\mathcal{S}\) — the physically real thing an action transforms
\(E\) the encoder: a neural network mapping an input to an embedding (written \(\varphi_\psi\) in GRL)
\(z = E(x)\) the latent embedding of \(x\) — a vector capturing meaning, not surface form. Lives in latent space \(\mathcal{Z}\)
\(E_{\text{target}}\) the target encoder: a slow exponential-moving-average (EMA) copy of \(E\), updated by averaging, not by gradients (stop-gradient). Produces JEPA's prediction targets
\(\mathcal{S},\ \mathcal{Z}\) the state space and the latent space. When actions operate on latents, \(\mathcal{S}=\mathcal{Z}\)
\(\mathrm{Pred}(\cdot)\) JEPA's predictor — maps a context embedding plus a query to the predicted embedding of the target
\(q\) a query / position token — tells the predictor which target to predict (where, or how far ahead)

Actions as operators — the GRL side

Symbol Meaning
\(\hat{O}\) an action operator: a function \(\hat O:\mathcal{S}\to\mathcal{S}\) that turns a state into its successor. The hat marks "operator (a function), not a number"
\(\hat{O}_\theta\) the specific state operator configured by parameters \(\theta\) — the physically meaningful (often inaccessible) transformation acting on \(s\)
\(\theta\) operator / action parameters — the knobs that configure the operator (a complete description of how it transforms the state)
\(\Theta\) the space of all operator parameters \(\theta\)
\(f_\theta\) the latent operator: how \(\hat O_\theta\) acts on the latent \(z = E(s)\) instead of on the raw state. The object you compute with
\(\Phi(\theta, s)\) the operator generator: builds an operator from parameters, \(\hat O_\theta(s)=\Phi(\theta,s)\)
\(E(\hat{O})\ge 0\) the energy functional: how large a transformation the operator is — a least-action / parsimony penalty. (Distinct from the encoder \(E\); the argument disambiguates)

Building the latent operator

Symbol Meaning
\(M_\theta\) the generator — a \(D \times D\) matrix; the "velocity field" the parameter \(\theta\) fills in
\(A_\theta = \exp(M_\theta)\) the operator matrix — the group element from exponentiating the generator. \(\exp\) is the matrix exponential
\(b_\theta\) an optional affine bias in latent space
\(f_\theta(z) = A_\theta z + b_\theta\) the latent operator in its concrete linear form
\(B_i\) a generator basis — matrices spanning the allowed operator directions; \(M_\theta = \sum_i \alpha_i B_i\)
\(\alpha\) the coefficient vector over the basis; \(\alpha = \theta\)

Policy and exploration

Symbol Meaning
\(\pi\) the operator policy: given a state, it chooses operator parameters
\(\pi_\psi:\mathcal{Z}\to\Delta(\Theta)\) the policy (weights \(\psi\)) maps a latent to a distribution over operator parameters
\(\Delta(\Theta)\) the set of probability distributions over \(\Theta\). A stochastic policy emits one of these; a deterministic policy emits a single \(\theta\)
\(\Delta z = E(\hat O_\theta(x)) - E(x)\) the change in latent caused by applying an operator — the operationalized meaning of a perturbation

Symbol operators

Symbol Meaning
\(\lVert v \rVert^2\) squared Euclidean (\(\ell_2\)) norm — measures magnitude / distance
\(\exp(M)\) the matrix exponential \(I + M + \tfrac{1}{2}M^2 + \cdots\) (matrix powers, not elementwise)
\(\circ\) function composition: \((\hat O_2 \circ \hat O_1)(s) = \hat O_2(\hat O_1(s))\) — apply \(\hat O_1\), then \(\hat O_2\)
\(\operatorname{Re}(\lambda)\) the real part of an eigenvalue of \(M_\theta\) — positive in any mode flags locally growing dynamics
\(\approx\) "approximately equals" — holds up to a small learned error
\(\sim\) "is distributed as" / "is sampled from"

One overload to keep straight

The symbol \(E\) does double duty: the encoder \(E\) (and its EMA copy \(E_{\text{target}}\)) and the energy functional \(E(\hat O)\). They never appear in the same role, and the argument disambiguates — \(E(x)\) encodes an observation, \(E(\hat O)\) scores an operator.


Foundation home: From actions to operators. Next: Augmenting JEPA with Action Operators.