---
title: 'LongEU: Long-Term Expected Utility Analysis'
url: https://www.emergentmind.com/topics/long-term-expected-utility-longeu
type: topic
---

# LongEU: Long-Term Expected Utility Analysis

Searching arXiv for recent and foundational papers on long-term expected utility, growth-optimality, and related asymptotic utility criteria.
Searching for: "long-term expected utility growth rate expected utility arXiv"
Long-Term Expected Utility, often abbreviated “longEU,” denotes a family of long-horizon decision criteria in which actions, portfolios, or policies are evaluated by expected utility or by the asymptotic growth rate of expected utility over repeated or extended interaction. Taken together, the cited literature suggests that longEU is not a single canonical formalism. In continuous-time portfolio theory it often refers to large-maturity asymptotics such as \(\lim_{T\to\infty} T^{-1}\log \mathbb E[U(\Pi_T)]\); in repeated-game learning it denotes a non-myopic expected-utility proxy that trades immediate payoff against future informational value; and in ergodicity economics it denotes the restricted class of settings in which expected utility can be reinterpreted as a long-run time-average growth criterion for a transformed wealth process [1906.03690][2104.00911][2508.14705][1801.03680].

## 1. Core meanings and formal objects

In long-horizon portfolio theory, the basic object is the finite-horizon optimal expected utility
\[
\sup_{\Pi\in\mathcal X}\mathbb E^{\mathbb P}[U(\Pi_T)],
\]
followed by a large-maturity limit. In the factor-model formulation with CRRA utility \(U(x)=x^p/p\), \(p<0\), the long-run growth rate is defined by
\[
\lim_{T\to\infty}\frac{1}{T}\ln\left|\sup_{\Pi\in\mathcal X}\mathbb E^{\mathbb P}[U(\Pi_T)]\right|,
\]
and the corresponding dual object \(v(\chi,T)\) satisfies
\[
-\lambda=\lim_{T\to\infty}\frac{1}{T}\ln v(\chi,T)
=\frac{1}{1-p}\lim_{T\to\infty}\frac{1}{T}\ln |\mathcal U(\chi,T)|.
\]
This makes longEU an asymptotic property of the finite-horizon optimal expected utility itself, rather than a separate objective imposed from the outset [1906.03690].

A closely related formulation appears in continuous-time portfolio choice with power utility \(U(x)=x^\nu/\nu\), \(\nu<0\), where the central quantities are the optimal expected utility \(u_T\), its logarithmic asymptotics, and especially the sensitivity objects
\[
\frac{\partial}{\partial \nu}\ln u_T,\qquad
\lim_{T\to\infty}\frac{1}{T}\frac{\partial}{\partial \nu}\ln u_T.
\]
Here longEU is not merely the level of expected utility at large \(T\), but its exponential rate and the asymptotic sensitivity of that rate to the risk-tolerance parameter \(\nu\) [2104.00911].

A more restricted but still canonical long-run formulation arises for static investors with constant stock weight \(\alpha\), where the objective is
\[
\Lambda(\alpha):=\lim_{t\to\infty}\frac1t\log \mathbb E[V_t^\theta],
\]
with \(\theta=1-\gamma\in(0,1)\). The investor chooses \(\alpha\in[0,1]\) to maximize \(\Lambda(\alpha)\). This is again a longEU criterion, but in a static-allocation rather than dynamic-control setting [1311.6179].

In repeated multi-objective Stackelberg games, longEU is defined more explicitly as a non-myopic approximation. The leader compares manipulation triples \((l,f,\mathbf c)\) not only by one-step expected utility, but by an approximate remaining-horizon value that extrapolates the consequences of acceptance or rejection over rounds \(t_0,\dots,T\). This use of the term is finite-horizon and sequential rather than asymptotic, but it is explicitly labeled “longEU” in the paper itself [2508.14705].

## 2. Time averages, ergodicity, and the dynamic reinterpretation of utility

A major reinterpretation of longEU comes from ergodicity economics. The central distinction is between expectation values, which average across parallel realizations, and time averages, which track what one individual experiences along one trajectory. The argument is that these coincide only for ergodic observables, and that wealth dynamics relevant to economics are often non-ergodic. For a transformed process \(u(x(t))\), the time-average growth rate is
\[
\bar r=\lim_{\Delta t\to\infty}\frac{\Delta u}{\Delta t},
\]
while the expectation-rate object is
\[
\langle r\rangle=\frac{\langle \Delta u\rangle}{\Delta t}.
\]
If \(u(t)\) has stationary independent increments, then \(\bar r=\langle r\rangle\), and with continuous paths the required utility dynamics are
\[
du=a_u\,dt+b_u\,dW.
\]
Under these conditions, maximizing expected utility growth is equivalent to maximizing long-run time-average growth [1801.03680].

This equivalence is conditional rather than universal. The wealth process must be modeled dynamically as
\[
dx=a_x(x)\,dt+b_x(x)\,dW,
\]
there must exist a monotone invertible transformation \(u(x)\), and the transformed process must have stationary independent increments. The paper excludes bounded utilities, restricts attention to continuous-path processes, and states that if no \(u(x)\) produces Brownian motion with drift in utility space, then expected utility theory cannot be given this dynamic interpretation. This suggests that longEU, in the ergodicity-economics sense, is a partial foundation for expected utility theory rather than a universal justification of it [1801.03680].

The same expectation-versus-time distinction reappears in timing uncertainty. For time lotteries with additive wealth dynamics,
\[
x(t+\Delta t)=x(t)+\Delta x,
\]
the time-average growth rate under repeated sequential play is
\[
\bar g=\frac{\Delta x}{p t_1+(1-p)t_2},
\]
whereas the ensemble-average growth rate is
\[
\langle g\rangle=p\frac{\Delta x}{t_1}+(1-p)\frac{\Delta x}{t_2}.
\]
The paper states explicitly that \(\bar g\neq \langle g\rangle\), interprets this as non-ergodicity, and argues that no utility function of wealth alone can generally restore ergodicity in this setting. This creates a sharp contrast between pathwise long-run criteria and expectation-based intertemporal criteria such as EDUT [2108.08366].

## 3. Long-horizon portfolio choice and asymptotic behavior

In incomplete diffusion markets with factor processes, longEU is governed by spectral asymptotics. For the dual value \(v(\chi,T)\), the asymptotic representation
\[
v(\chi,T)\simeq e^{-\lambda T}\phi(\chi)
\]
implies
\[
\mathcal U(\chi,T)\simeq \frac1p\,e^{-(1-p)\lambda T}\phi(\chi)^{1-p}
\]
up to a multiplicative term converging to a positive constant. The eigenvalue \(\lambda\) determines the long-run exponential rate, while the eigenfunction \(\phi\) determines dependence on the initial factor \(\chi\). The long-run initial-state sensitivity is
\[
\lim_{T\to\infty}\frac{\partial}{\partial\chi}\ln |\mathcal U(\chi,T)|
=(1-p)\frac{\phi'(\chi)}{\phi(\chi)},
\]
and parameter sensitivity is
\[
\lim_{T\to\infty}\frac{1}{T}\frac{\partial}{\partial\epsilon}\Big|_{\epsilon=0}
\ln |\mathcal U^\epsilon(\chi,T)|
=-(1-p)\frac{\partial\lambda_\epsilon}{\partial\epsilon}\Big|_{\epsilon=0}.
\]
This makes longEU a spectral object in which the principal eigenvalue governs rates and the eigenfunction governs prefactors and initial-condition effects [1906.03690].

A parallel conclusion holds for optimal terminal-wealth portfolios under CRRA utility. There the Hansen–Scheinkman decomposition yields
\[
p_T=\phi(\xi)e^{-\lambda T}f(T,\xi),
\]
with \(f(T,\xi)\to\) positive constant, so that
\[
\lim_{T\to\infty}\frac{1}{T}\log p_T=-\lambda,\qquad
\lim_{T\to\infty}\frac{1}{T}\frac{\partial}{\partial \nu}\log p_T
=-\frac{\partial\lambda}{\partial \nu}.
\]
For the utility itself, the large-time sensitivity of \(\log u_T\) is determined by \(r\), \(\lambda\), and \(\partial_\nu\lambda\). The principal eigenvalue controls the exponential rate, while the eigenfunction contributes only \(O(1)\) terms after logarithmic normalization by \(T\) [2104.00911].

LongEU can also determine asymptotic portfolio behavior rather than only value-function growth. In a Bayesian market with unknown but fixed drift, the long-horizon expected-power-utility maximizer becomes extremal. For \(\alpha\in(0,1)\), the optimal portfolio converges to the full-information Merton fraction associated with the largest possible drift \(\mu_d\); for \(\alpha<0\), it converges to the one associated with the smallest possible drift \(\mu_1\). The limiting policy is independent of current posterior beliefs and prior probabilities. This suggests that longEU can asymptotically select support extrema rather than posterior averages [1703.04423].

For static investors, the same long-run objective can be solved model by model. With power utility and constant allocation \(\alpha\), the paper derives explicit \(\Lambda(\alpha)\) functions for the Heston model, the 3/2 model, jump diffusion, and Black–Scholes with correlated Vasicek short rate, and then optimizes over \(\alpha\in[0,1]\). Here longEU is an asymptotic growth rate of \(\mathbb E[V_t^\theta]\), not a dynamic-programming value function [1311.6179].

## 4. Spectral, ergodic, and benchmark methods

The operator-theoretic structure of longEU is especially clear in ergodic HJB and eigenfunction methods. In the incomplete-market factor model, the dual HJB PDE leads to the ergodic eigenvalue problem
\[
\mathcal L\phi=-\lambda\phi,
\]
and the exact finite-horizon decomposition
\[
v(\chi,T)=e^{-\lambda T}\phi(\chi)\,
\mathbb E^{\mathbb Q}\!\left[
\frac{1}{\phi(X_T)}e^{\int_0^T f(X_s,s;T)\,ds}
\right].
\]
The transient factor measures the gap between finite-horizon and ergodic controls. LongEU emerges by showing that this residual term is asymptotically negligible in \(T^{-1}\log\) scale [1906.03690].

The paper on risk-tolerance sensitivity reaches the same structure through Hansen–Scheinkman decomposition and Malliavin calculus. There the eigenpair \((\lambda,\phi)\) of a killed generator isolates the exponential growth/decay rate, while Malliavin derivative formulas are used to prove that the derivative of the remainder term is \(O(1)\). This is why the long-run sensitivity of expected utility is governed by \(\partial_\nu\lambda\), not by derivatives of the remainder [2104.00911].

In robust utility maximization, the asymptotic criterion becomes
\[
\Lambda(\lambda)
=
\sup_{\pi\in\mathcal A}
\overline{\lim}_{T\to\infty}
\frac1T
\ln
\inf_{Q^\eta\in\mathcal Q}
E_{Q^\eta}[(X_T^\pi)^\lambda].
\]
Convex duality transforms this into an infinite-horizon risk-sensitive control problem, and the optimal growth rate is characterized by an ergodic Bellman equation. The same equation identifies an optimal long-run trading strategy and an asymptotic worst-case model \(Q^{\eta^*}\). This places longEU within robust risk-sensitive control and large-deviations duality [1203.1191].

A distinct but related reduction uses a tradeable stochastic discount factor equal to the inverse of the growth-optimal portfolio:
\[
F_t=\frac{1}{V_t^{GP}}.
\]
With fairness imposed, optimal terminal wealth and optimal consumption satisfy first-order conditions involving only \(V_t^{GP}\), not the full market dynamics. In scalar Markov cases, the optimal strategy reduces to a two-fund separation rule between the GP and the baseline riskless asset. This suggests that, in some long-horizon expected-utility problems, the effective state variable is the benchmark portfolio spanning the tradeable SDF [1705.03929].

## 5. Robustness, leverage, and long-run utility growth

LongEU has been developed extensively for leveraged ETFs. With power utility \(u(w)=w^\alpha\), \(0<\alpha\le 1\), and LETF leverage ratio \(\beta\), the long-run object is
\[
\lim_{t\to\infty}\frac1t\log \mathbb E^\mathbb P[L_t^\alpha].
\]
The analytical method is martingale extraction: one first removes the stochastic integral by a Girsanov transform, then solves an eigenpair problem for the infinitesimal generator of a Markov diffusion, and finally reads off the exponential growth rate from the eigenvalue once the transformed-state expectation is shown to converge. The paper works this out explicitly under GBM, GARCH, inverse GARCH, extended CIR, 3/2, quadratic, Heston, 3/2 stochastic volatility, and stochastic short-rate extensions, and then determines the optimal leverage ratio \(\beta^*\) model by model [1612.01013].

Under parameter uncertainty, the criterion becomes robust:
\[
\lim_{T\to\infty}\frac1T\log
\inf_{\alpha\in[\underline\alpha,\overline\alpha]}
\mathbb E^{\mathbb P^\alpha}[L_T^p].
\]
The method combines a comparison principle, used to identify worst-case parameters on the uncertainty set, with martingale extraction, used to compute the asymptotic rate for the worst-case model. Explicit robust long-run growth rates are obtained for GBM, CIR, 3/2, Heston, 3/2 stochastic volatility, and models with Vasicek or inverse GARCH short rate. The robust optimal leverage ratio
\[
\beta^*\in\arg\max_{\beta\in[\underline\beta,\overline\beta]}\Lambda(\beta)
\]
is explicit in some models and numerical in others. The paper also distinguishes long-term expected utility from long-term expected return: the expected-utility criterion contains a volatility penalty term
\[
-\frac12 p(1-p)\beta^2(\cdot),
\]
which disappears at \(p=1\), making robust longEU more conservative than risk-neutral long-run expected return [2310.02084].

These LETF papers make one recurring point precise: in longEU problems for constant-leverage products, leverage magnifies both drift and volatility, while financing costs and stochastic-state persistence determine whether expected utility grows exponentially, decays, or even becomes infinite. LongEU is therefore simultaneously a utility criterion, a model-selection device through worst-case analysis, and an optimizer of leverage exposure [1612.01013][2310.02084].

## 6. Adjacent frameworks, limits, and recurrent misconceptions

Several neighboring literatures clarify what longEU is not. In repeated multi-objective Stackelberg games with unknown follower preferences, longEU is a non-myopic approximation rather than an asymptotic certainty-equivalent rate. The leader compares manipulations by current acceptance probability and remaining-horizon consequences, and under infinite repeated interactions the longEU-based algorithms are proved to converge to the optimal manipulation cost. Here longEU is explicitly a learning-and-control heuristic that values information acquisition, not just current expected payoff [2508.14705].

Other neighboring work is only partially related. “Expected Utility Networks” provide a modular representation of probability and utility dependencies, define conditional expected utility at the event level, and prove that joint probability and utility separation implies conditional EU independence. This is a partial foundation for scalable expected-utility reasoning, but it does not itself define a temporal, discounted, or infinite-horizon longEU model [1301.6714]. Likewise, fluctuation-theorem results on certainty equivalents derive entropy-dependent bounds on one-shot expected-utility valuations and are best classified as thermodynamic constraints on risky choice rather than a long-run expected-utility theory [2307.12358].

The strongest formal limit result comes from universal AI. In a setting with repeated interaction, a prior over environments that assigns positive probability to every computable environment, and a utility function defined on entire infinite perception sequences, the expected utility of any policy is either undefined or infinite if the utility is computably unbounded above or below; if it is computably unbounded in both directions, the expectation is undefined. The paper states that bounded utility is the simplest way to guarantee convergence. This shows that longEU, when extended to universal hypothesis classes and infinite-horizon utilities, can fail as a well-defined scalar criterion [0907.5598].

Several misconceptions recur across the literature. One is that ordinary expected utility is automatically a long-run criterion; the ergodicity paper rejects this and allows the identification only when utility is an ergodicity transformation satisfying stringent dynamical conditions [1801.03680]. Another is that long-run pathwise criteria and expectation-based criteria must agree; the time-lottery results provide a direct counterexample, with ensemble averages and time averages yielding different rankings [2108.08366]. A third is that asymptotic eigenvalue formulas make finite-horizon terms irrelevant; the sensitivity literature explicitly shows that eigenfunctions and remainder terms still govern initial-state dependence and subleading finite-horizon effects, even when the leading rate is spectral [1906.03690][2104.00911].

Taken together, these works suggest that longEU is best understood as a technical umbrella for long-horizon expected-utility evaluation under several distinct regimes: asymptotic portfolio growth, ergodic reinterpretations of utility, robust risk-sensitive control, benchmark-based asset allocation, and non-myopic repeated-game learning. What unifies them is not a single formula, but the attempt to make expected utility dynamically relevant over long horizons.

Source: https://www.emergentmind.com/topics/long-term-expected-utility-longeu