---
title: Neural Expectation Operators
url: https://www.emergentmind.com/topics/neural-expectation-operators
type: topic
---

# Neural Expectation Operators

Neural Expectation Operators are non-linear conditional expectations induced by backward stochastic differential equations whose drivers are parameterized by neural networks. Within the paradigm of Measure Learning, they provide a probabilistic model of ambiguity—uncertainty over probability laws—by encoding attitude toward model misspecification directly in the BSDE driver. In the associated theory of Neural Brownian Motion, the same operator, written $\varepsilon^\theta$, becomes the martingale notion relative to which a canonical stochastic process is defined, linking learned ambiguity to endogenous volatility and learned changes of measure [2507.10607][2507.14499].

## 1. BSDE definition and non-linear expectation structure

In its simplest form, a Neural Expectation Operator is induced by a neural BSDE
$$
Y_t = \xi + \int_t^T f_\theta(s,Y_s,Z_s)\,ds - \int_t^T Z_s\,dW_s,
$$
where the driver $f_\theta$ is computed by a neural network with parameters $\theta$. In the general setting, the driver also depends on a forward state process $X$:
$$
Y_s = \xi + \int_s^T f_\theta(r, X_r, Y_r, Z_r)\,dr - \int_s^T Z_r\,dW_r, \quad s\in[t,T].
$$
The operator acts as a non-linear conditional expectation via $Y_t = E[\xi\mid F_t]$. In the notation used for Neural Brownian Motion, one defines
$$
- dY_s = f_\theta(s, X_s, Y_s, Z_s) ds - Z_s dW_s, \qquad Y_T = \xi,
$$
and sets $\varepsilon^\theta[\xi \mid F_t] := Y_t$ [2507.10607].

The economic and probabilistic content of the construction is that $f_\theta$ compactly represents the structure of ambiguity. Quadratic growth in the martingale component $z$ encodes risk and ambiguity aversion. Under suitable convexity and independence assumptions, the operator admits a dual representation over a family of measures related by Girsanov transforms: for drivers of the form $f_\theta(t,x,z)$, independent of $y$ and convex in $z$,
$$
E[\xi\mid F_t] = \operatorname*{ess\,sup}_{u\in\mathcal U} E_{\mathbb{Q}^u}\!\left[\xi + \int_t^T g_\theta(s,X_s,u_s)\,ds \,\bigg|\, F_t\right],
$$
with $d\mathbb{Q}^u/d\mathbb{P} = \mathcal{E}(\int_0^T u_s\,dW_s)$ and $g_\theta$ the (negative) Fenchel–Legendre transform of $f_\theta$ in $z$. In that sense, the non-linear expectation evaluates $\xi$ under a worst-case change of measure from a learned universe of plausible models [2507.10607].

Classical linear conditional expectation is recovered when $f_\theta \equiv 0$. In that case the BSDE reduces to the martingale representation with $Y_t = E[\xi\mid F_t]$, so the neural construction is a strict generalization rather than a replacement of standard expectation.

## 2. Well-posedness under quadratic growth and local Lipschitz structure

The foundational well-posedness theorem is formulated for drivers satisfying continuity in $(t,x,y,z)$, quadratic growth in $z$, uniform local Lipschitz continuity in $y$, and monotonicity in the sense that $y\mapsto f_\theta(t,x,y,z)$ is non-increasing. The terminal condition is required to satisfy exponential integrability, and the random part of the driver induced by the forward state $X$ must satisfy an exponential integrability condition of the form
$$
E\!\left[\exp\!\Big(\lambda_0\int_0^T \sup_{y,z}\frac{f_\theta(s, X_s, y, z)}{1+|y|+|z|^2}\,ds\Big)\right] < \infty.
$$
The main theorem states that if $\beta_0>\alpha$, then there exists a unique adapted solution
$$
(Y,Z)\in \mathcal{S}^\infty([0,T];\mathbb{R})\times \mathcal{H}^2_{\mathrm{BMO}([0,T];\mathbb{R}^{1\times d})},
$$
so $Y$ is essentially bounded and $\int_0^\cdot Z_s\,dW_s$ is a BMO-martingale [2507.10607].

A notable feature of the result is that it circumvents the classical global Lipschitz assumption in the state variable $y$. The theorem is explicitly described as applicable to common neural network architectures, including architectures with ReLU activations, and the exponential integrability condition on terminal data is described as the sharp condition for this setting. A priori estimates are obtained by comparison with dominating quadratic BSDEs. With auxiliary upper and lower generators, one shows
$$
\underline{Y}\le Y\le \bar Y \quad \text{a.s. on }[0,T],
$$
which yields boundedness of $Y$ and BMO control of $Z$; uniqueness follows by strict comparison in $y$ and identification of $Z$ via martingale representation [2507.10607].

A recurrent misconception is that neural parameterization necessarily weakens analytical control. The constructive objective of the theory is the opposite: it builds a bridge between the abstract assumptions of quadratic BSDE theory and concrete, verifiable neural network designs. In that sense, the learned operator is not an unconstrained black box, but a BSDE object whose solvability is tied to explicit structural hypotheses.

## 3. Architectural constraints and axiomatic properties

The theory includes constructive neural architectures that enforce the analytical assumptions required by quadratic BSDEs. A separable driver
$$
f_\theta(t,x,y,z)=N_1(t,x,z)+N_2(y)
$$
satisfies the uniform local Lipschitz property in $y$ when $N_2$ is locally Lipschitz, and monotonicity when $N_2$ is chosen non-increasing. A bounded-interaction form
$$
f_\theta(t,x,y,z)=N_1(t,x,z)+N_2(t,x,z)\cdot N_3(y)
$$
yields local Lipschitz continuity in $y$ provided $N_2$ is uniformly bounded and $N_3$ is locally Lipschitz [2507.10607].

Monotonicity can also be enforced directly by sign constraints. In a feed-forward network with non-decreasing activations, one imposes non-positive weights from $y$ to the first hidden layer and non-negative weights in subsequent hidden layers and in the output layer. A simple induction then gives $\partial f_\theta/\partial y\le 0$ everywhere, hence monotonicity. Convexity in $(y,z)$ can be imposed with Input-Convex Neural Networks. In that case convexity of the operator $\xi\mapsto E[\xi\mid F_t]$ is guaranteed, and under convexity and monotonicity Jensen’s inequality holds:
$$
E[\phi(\xi)\mid F_t]\;\ge\;\phi(E[\xi\mid F_t])\quad\text{for convex }\phi\in C^2.
$$
The same framework is described as encoding economically meaningful axioms such as monotonicity, time consistency, and, via ICNNs, convexity [2507.10607].

These constructions are important because they make the analytical hypotheses inspectable at the level of weights and activations. The operator’s qualitative properties are therefore not merely postulated; they can be enforced by architectural design.

## 4. Fully coupled forward–backward systems and mean-field asymptotics

The framework extends to fully coupled forward–backward stochastic differential equations of the form
$$
\begin{aligned}
dX_s &= b_\theta(s,X_s,Y_s,Z_s)\,ds + \sigma_\theta(s,X_s,Y_s,Z_s)\,dW_s,\\
-dY_s &= f_\theta(s,X_s,Y_s,Z_s)\,ds - Z_s\,dW_s,\qquad Y_T=g_\theta(X_T).
\end{aligned}
$$
Under continuity and global Lipschitz conditions on $b_\theta$, $\sigma_\theta$, and $g_\theta$, together with the quadratic-growth, monotone-in-$y$, and local-Lipschitz-in-$y$ conditions on $f_\theta$, and a small-time or small-Lipschitz condition $T\cdot C<1$, the system admits a unique adapted solution
$$
(X,Y,Z)\in \mathcal{S}^p([0,T];\mathbb{R}^n)\times \mathcal{S}^\infty([0,T];\mathbb{R})\times \mathcal{H}^2_{\mathrm{BMO}([0,T];\mathbb{R}^{1\times d})},
$$
for any $p\ge 2$ [2507.10607].

The proof follows the four-step scheme. One identifies a quasilinear parabolic PDE for $u(t,x)=Y_t$, with the quadratic $z$-dependence entering through $Z=(\nabla_x u)^\top \sigma_\theta$; one then uses viscosity solution theory and comparison to obtain a unique spatially Lipschitz solution $u$; finally one reconstructs $(X,Y,Z)$ by the Markovian ansatz $Y_s=u(s,X_s)$ and $Z_s=(\nabla_x u)^\top \sigma_\theta(s,X_s,u,Z_s)$ [2507.10607].

The same program is pushed to interacting particle systems. For exchangeable particles with empirical law $\mu_t^N$, the theory establishes a Law of Large Numbers, described as propagation of chaos, and a Central Limit Theorem. The LLN states that, for each particle index $i$,
$$
\lim_{N\to\infty} E\Big[\sup_{t\in[0,T]}|X^{i,N}_t-\bar X^i_t|^2+\sup_{t\in[0,T]}|Y^{i,N}_t-\bar Y^i_t|^2+\int_0^T|Z^{i,N}_t-\bar Z^i_t|^2\,dt\Big]=0.
$$
Under differentiability of the coefficients in $(x,y,z)$ and Lions differentiability in the measure argument, the centered fluctuations converge in law to the unique solution of a linear McKean–Vlasov FBSDE. Because the fluctuation dynamics are linear and driven by Brownian inputs, the limit law is Gaussian [2507.10607].

## 5. Self-referential martingales and canonical Neural-Brownian Motion

A particularly distinctive development is the use of the Neural Expectation Operator to define a new martingale notion. For self-referential processes, the driver is specialized by identifying the value and state variables with the process itself:
$$
g_\theta(t,m,z) := f_\theta(t,m,m,z).
$$
A one-dimensional process $(M_t)$ is then called a Neural-Brownian Motion if it is continuous, $M_0=0$, and is an $\varepsilon^\theta$-martingale it generates:
$$
M_s = \varepsilon^\theta[M_t \mid F_s], \qquad 0 \le s \le t \le T.
$$
It is canonical if it has zero drift under the physical measure $\mathbb{P}$ [2507.14499].

If $M$ is an Itô process
$$
dM_t = b_t\,dt + \sigma_t\,dW_t,
$$
then $M$ is an $\varepsilon^\theta$-martingale if and only if
$$
b_t = - g_\theta(t, M_t, \sigma_t)
$$
for almost every $t\in[0,T]$. The canonical condition $b_t=0$ therefore implies the algebraic constraint
$$
g_\theta(t, M_t, \sigma_t)=0,
$$
so the volatility is characterized as a root of $g_\theta(t,M_t,\cdot)$ rather than postulated a priori. Under the implicit volatility assumption that, for each $(t,x)$, the equation $g_\theta(t,x,z)=0$ has a unique positive root $z=\nu_\theta(t,x)$, that the root is regular, and that $\nu_\theta$ has global linear growth, the canonical Neural-Brownian Motion exists and is the unique strong solution of
$$
dM_t = \nu_\theta(t, M_t)\, dW_t,\qquad M_0=0.
$$
The regular root condition yields $C^1$ regularity of $\nu_\theta$ via the Implicit Function Theorem, and local Lipschitz continuity in $x$ follows immediately [2507.14499].

This produces a learned stochastic calculus in which the infinitesimal generator is
$$
Lu(t,x) = \partial_t u(t,x) + \frac{1}{2} \nu_\theta(t,x)^2 \partial_{xx} u(t,x).
$$
The volatility $\nu_\theta$ is endogenous: it emerges as the unique positive root of $g_\theta$ and is therefore learned from the ambiguity structure encoded by the driver. The process is correspondingly no longer characterized by stationary independent increments; its volatility is state- and driver-dependent, learned from data [2507.14499].

## 6. Learned measure changes, sensitivity, applications, and related notions

In the quadratic specialized case
$$
g_\theta(t,x,z) = (\alpha/2) z^2 - \beta,
$$
with $\alpha\neq 0$ and $\beta/\alpha>0$, the canonical constraint yields a constant volatility
$$
\nu_\theta = \sqrt{2\beta/\alpha}.
$$
Under $\mathbb{P}$,
$$
dM_t = \nu_\theta\, dW_t,
$$
so $M$ is a scaled Brownian motion. Defining a new measure $\mathbb{Q}_\alpha$ by the Radon–Nikodym density with kernel $\gamma_t=\alpha\nu_\theta$,
$$
L_t = \exp\!\left(\int_0^t \gamma_s dW_s - \frac12\int_0^t \gamma_s^2 ds\right),
$$
one obtains a Girsanov-type theorem: $W_t^\alpha := W_t - \alpha\nu_\theta t$ is a $\mathbb{Q}_\alpha$-Brownian motion and
$$
dM_t = \alpha \nu_\theta^2\,dt + \nu_\theta\, dW_t^\alpha.
$$
Thus a canonical Neural-Brownian Motion acquires a drift under the learned measure. The sign of $\alpha$ determines the interpretation: $\alpha>0$ corresponds to a convex driver and is interpreted as ambiguity aversion, whereas $\alpha<0$ corresponds to a concave driver and is interpreted as ambiguity seeking [2507.14499].

From the computational side, the framework is designed for learning from data. Neural Expectation Operators are computed by solving the BSDE for $(Y,Z)$ given $\theta$ and $\xi$, and training $\theta$ is described as well-posed because the theory provides a sensitivity BSDE that yields exact gradients for $\theta$. If $Y_t=E[\xi\mid F_t]$ and the driver is differentiable in $\theta$, then the gradient processes $(Y_s^\theta,Z_s^\theta)$ solve the linear BSDE
$$
-dY_s^\theta = \Big[\nabla_\theta f_\theta(s,X_s,Y_s,Z_s) + \partial_y f_\theta(s,X_s,Y_s,Z_s)\,Y_s^\theta + D_z f_\theta(s,X_s,Y_s,Z_s)\,Z_s^\theta\Big]ds - Z_s^\theta\,dW_s,
$$
with $Y_T^\theta=0$,
which enables principled gradient-based learning with SGD or Adam [2507.10607].

The stated application domain includes stochastic control, robust finance, and macro models of systemic ambiguity. A concrete example is the Merton portfolio problem under the driver $f_\theta(t,x,y,z)=-(\theta/2)|z|^2$, which is described as consistent with the theory and as predicting ambiguity-induced caution; the optimal policy is wealth-dependent and strictly more conservative than the classical Merton strategy, with the ambiguity parameter $\theta$ estimable from data via the sensitivity machinery. In the Neural-Brownian Motion formulation, the same learned-volatility mechanism is presented as opening principled applications such as consistent implicit-volatility models in finance, while also opening a path to mean-field limits when interactions are introduced through measure-dependent drivers [2507.10607][2507.14499].

Neural Expectation Operators should be distinguished from methods that estimate linear expectations through PDE surrogates. The Feynman–Kac Operator Expectation Estimator, for example, estimates $\mathbb{E}_{X\sim P}[f(X)]$ by combining diffusion bridge models with approximation of the Feynman–Kac operator using Physically Informed Neural Networks. That construction is an expectation estimator for a target mathematical expectation, whereas Neural Expectation Operators define a non-linear conditional expectation itself through a BSDE driver and are expressly designed to model ambiguity [2407.02010].

Source: https://www.emergentmind.com/topics/neural-expectation-operators