Skip to content

Notation Reference — Operator World Models

A standalone glossary for the Operator World Models series. Every symbol used across the series, grouped by role, with a "read as" column. Keep it open in a second tab while reading.

This series writes the encoder as \(E\), matching the Action Operators foundation. The same encoder is written \(\varphi_\psi\) in the GRL literature — the same object, \(E \leftrightarrow \varphi_\psi\).


Representation — the JEPA side

Symbol Read as Meaning
\(x\) "x" a raw observation — an image, a window of behavioral sensor signal, a protein structure
\(s\) "s" the underlying system state in observation/state space \(\mathcal{S}\) — the physically real thing an action transforms (a person's full physiological state, a protein's atomic coordinates)
\(E\) "E" the encoder: a neural network mapping an observation to a latent. Same object as \(\varphi_\psi\) in the GRL literature
\(z = E(x)\) "z equals E of x" the latent embedding — a vector capturing meaning, not surface form. Lives in latent space \(\mathcal{Z}\)
\(z_t\) "z at t" the latent at time \(t\), \(z_t = E_\xi(x_{\le t})\) — the encoded history up to \(t\)
\(E_\xi\) "E-xi" the online encoder, trainable weights \(\xi\) (Greek xi)
\(E_{\bar\xi}\) "E-xi-bar" the target encoder: a slow exponential-moving-average (EMA) copy of \(E_\xi\), used to produce prediction targets. Stop-gradient — no backprop flows into it
\(\bar\xi \leftarrow \tau\bar\xi + (1-\tau)\xi\) the EMA update of the target weights; \(\tau\) close to 1 makes the target drift slowly
\(\tau\) tau the EMA rate (e.g. \(0.999\)): the fraction of the old target weights kept each update
\(\mathrm{sg}\) "stop-grad" stop-gradient: treat the argument as a constant during backprop

Operators — state space and latent space

Symbol Read as Meaning
\(\hat O\) "O-hat" 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" — the physics convention (as in \(\hat H\)), not the statistics "estimate" hat of \(\hat\theta\)
\(\hat O_\theta\) "O-hat-theta" the specific state operator chosen by parameters \(\theta\) — the physically meaningful transformation acting on the real state \(s\). Often inaccessible: you only observe \(s\) and \(s'\), never the operator itself
\(f_\theta\) "f-theta" the latent operator: how \(\hat O_\theta\) acts on the latent \(z=E(s)\) instead of on \(s\). This is the object you actually compute with
\(\theta\) theta operator parameters — the knobs that configure the operator: a complete description of how it transforms the state (for a continuous \(\theta\) the operator is synthesized from \(\theta\), not looked up from a menu)
\(\Theta\) capital Theta the space of all operator parameters

Building the latent operator

Symbol Read as Meaning
\(B_i\) "B-i" the generator basis — a set of \(m\) matrices (each \(D\times D\), where \(D=\dim\mathcal{Z}\)) that span the allowed operator directions. The basis choice sets what kind of operator is possible
\(\alpha\) alpha the coefficient vector over the basis; \(\alpha = \theta\). Component \(\alpha_i\) says "how much of generator \(B_i\)"
\(M_\theta = \sum_i \alpha_i B_i\) "M-theta" the flow generator — a Lie-algebra element; the "infinitesimal" form of the operator, living in a flat vector space
\(A_\theta = \exp(M_\theta)\) "A-theta" the operator matrix — the group element obtained by exponentiating the generator. \(\exp\) is the matrix exponential
\(b_\theta\) "b-theta" an optional affine bias term in latent space
\(f_\theta(z) = A_\theta z + b_\theta\) the latent operator in its concrete linear form, \(\exp(M_\theta) z + b_\theta\)

Conditioning and policy

Symbol Read as Meaning
\(c_t\) "c at t" the context / intervention covariates at time \(t\) — the known causes of state change (hours slept, medication taken, a stressor; a mutation, a ligand)
\(\pi_\psi\) "pi-psi" the context policy: emits the operator coefficients from the latent and context, \(\theta \sim \pi_\psi(z_t, c_t)\), with weights \(\psi\)
\(\Delta(\Theta)\) "simplex over Theta" the set of probability distributions over \(\Theta\). A stochastic policy emits one of these; a deterministic policy emits a single \(\theta\)
\(g_\phi\) "g-phi" JEPA's predictor with weights \(\phi\) — the query-conditioned operator the action operator generalizes
\(q_{\Delta t}\) "query at delta-t" the predictor's query: in vanilla temporal JEPA it carries only the time offset \(\Delta t\); the action operator replaces it with \(\theta(c_t)\)

Energy, spectrum, loss

Symbol Read as Meaning
\(E(\hat O)\ge 0\) "energy of O-hat" the energy functional: how large a transformation the operator is. Here \(E(\hat O)=\lVert M_\theta\rVert_F^2\) — a least-action / parsimony penalty that keeps the operator near identity unless the data demands otherwise. (Distinct from the encoder \(E\); context disambiguates)
\(\lambda\) lambda the weight on the energy penalty
\(\mathrm{Re}(\lambda_i)\) the real part of an eigenvalue of the generator \(M_\theta\). A positive real part in any mode flags locally growing (destabilizing) dynamics
\(\mathcal{L}\) "script L" the training loss — latent-space squared error between the predicted and target latents
\(\lVert v\rVert^2\) squared Euclidean (\(\ell_2\)) norm
\(\lVert M\rVert_F^2\) squared Frobenius norm — the sum of squared matrix entries

One overload to keep straight

The symbol \(E\) does double duty: the encoder \(E\) (and its variants \(E_\xi\), \(E_{\bar\xi}\)) 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.


See every symbol at work. A table defines a symbol; it does not teach one. For all of this notation threaded through a single concrete story — CGM, insulin, carbs, exercise, and medical codes for one person managing diabetes — see the worked example, where each symbol is earned by something real before it is named.

Series home: Operator World Models. The foundation: Augmenting JEPA with Action Operators.