9  Certainty equivalence in POMDPs

9.1 Classical certainty equivalence

In linear-quadratic-Gaussian (LQG) control with output feedback, the optimal control law has the following structure. The optimal action at each time is obtained by evaluating the optimal state-feedback policy of the fully observed stochastic LQG system at the MMSE state estimate: if \(\hat π^\star\) is optimal for that fully observed problem and \(σ_t(h_t) = \EXP[S_t \mid H_t = h_t]\) is the MMSE estimate, then \[ π_t(h_t) = \hat π^\star_t\bigl(σ_t(h_t)\bigr). \]

The comparison model \(\ALPHABET M\) is fully observed but still stochastic. One does not replace process noise by its mean and then solve a deterministic open-loop problem. (That different idea is sometimes also called certainty equivalence in stochastic control; we do not treat that idea here.)

In LQG this construction is optimal. Outside LQG, composing a fully observed optimal policy with a state estimate is a heuristic. In the rest of this lecture, we make that heuristic quantitative for finite POMDPs.

9.2 Generalized CE for finite POMDPs

Fix a finite-horizon POMDP with finite spaces \(\ALPHABET S\), \(\ALPHABET A\), \(\ALPHABET Y\), history \(H_t = (Y_{1:t}, A_{1:t-1})\), and additive costs \(C_t = c_t(S_t, A_t)\). Write \(V^π_t(h_t)\) and \(V^\star_t(h_t)\) for the history value of a policy \(π\) and for the optimal history value, as in lecture 7.

Let \(\ALPHABET M\) be the associated fully observed MDP \[ \ALPHABET M = \bigl\langle \ALPHABET S, \ALPHABET A, P, c, T \bigr\rangle, \] in which the controller sees \(S_t\) at time \(t\) and \(P_t(\cdot \mid s, a)\) is the state transition kernel of the POMDP. Let \(\hat π^\star = (\hat π^\star_1,\dots,\hat π^\star_T)\) denote the optimal policy of the MDP and \(\{\hat V^\star_t\}_{t=1}^T\) be the corresponding optimal value function.

Consider an arbitrary sequence of estimators \(σ_t \colon \ALPHABET H_t \to \ALPHABET S\) that generate a state estimate from the history. For example, this could be the conditional mean, MAP estimate, the last observation, and so on. Write \(\hat S_t = σ_t(H_t)\) for the resulting estimate. The certainty equivalent policy generated by \(\{σ_t\}\) is \[ π^{\mathrm{CE}}_t(h_t) = \hat π^\star_t\bigl(σ_t(h_t)\bigr). \] In words: pretend the estimate is the true state, and play the optimal fully observed action. In the language of lecture 8 this is exactly the lifted AIS policy \(π_t = \hat π^\star_t \circ σ_t\). We are interested in quantifying the sub-optimality gap of this policy.

9.2.1 Quality of estimation

Let \(\ALPHABET S\) be a metric space with a metric \(d_{\ALPHABET S}\). Define the conditional estimation error \[\begin{equation}\label{eq:ce-eta-local} η_t(h_t) \coloneqq \EXP\bigl[ d_{\ALPHABET S}\bigl(S_t, σ_t(h_t)\bigr) \bigm| H_t = h_t \bigr], \end{equation}\] with the convention \(η_{T+1} \equiv 0\). By the belief lemma of lecture 7 (Lemma 7.2), the conditional law of \(S_t\) given \(h_t\) does not depend on the control policy, so \(η_t(h_t)\) is a property of the observation model and the estimator alone.

9.3 Smoothness assumption

Assume the cost and dynamics of \(\ALPHABET M\) are Lipschitz in the state: there exist constants \(L^c_t, L^P_t \ge 0\) such that for all \(s,\tilde s \in \ALPHABET S\) and \(a \in \ALPHABET A\), \[\begin{align} \ABS{c_t(s,a) - c_t(\tilde s,a)} &\le L^c_t\, d_{\ALPHABET S}(s,\tilde s), \label{eq:ce-Lip-c} \\ W_1\bigl( P_t(\cdot \mid s,a),\, P_t(\cdot \mid \tilde s,a) \bigr) &\le L^P_t\, d_{\ALPHABET S}(s,\tilde s). \label{eq:ce-Lip-P} \end{align}\] As in Theorem 3.3, \[ \operatorname{Lip}(\hat V^\star_t) \le L^c_t + L^P_t\, \operatorname{Lip}(\hat V^\star_{t+1}), \] with \(\operatorname{Lip}(\hat V^\star_{T+1}) = 0\). Thus Lipschitz constants of \(\hat V^\star\) are controlled by \(\{L^c_t, L^P_t\}\) alone.

9.4 Main insight: the CE generator is an AIS

Certainty equivalence is an AIS policy for a concrete generator whose pointwise errors are linear in \(η_t(h_t)\). The approximate MDP is exactly the fully observed model \(\ALPHABET M\), and the lifted AIS policy is \(π^{\mathrm{CE}} = \hat π^\star \circ σ\).

Theorem 9.1 (CE as an AIS) Under \(\eqref{eq:ce-Lip-c}\)\(\eqref{eq:ce-Lip-P}\), the compression \(σ_t\) with cost approximators \(\hat c_t = c_t\) and dynamics approximators \(\hat P_t = P_t\) (the dynamics of \(\ALPHABET M\)) are an AIS generator with tolerances \[ ε_t(h_t) = L^c_t\, η_t(h_t), \qquad δ_t(h_t, a_t) = L^P_t\, η_t(h_t) + \EXP\bigl[ η_{t+1}(H_{t+1}) \bigm| h_t, a_t \bigr]. \] Consequently \(π^{\mathrm{CE}}\) is the lifted AIS policy for this generator, with approximate optimal value functions \(\{\hat V^\star_t\}\).

The theorem is proved by the next two lemmas: Lemma 9.1 is pointwise (AP1) and Lemma 9.2 is pointwise (AP2). The policy gap is deferred to the next section (sample-path envelopes of lecture 8), rather than the uniform form of Theorem 8.1.

9.4.1 (AP1): cost approximation

Lemma 9.1 (Cost approximation for CE) For every history \(h_t\) and action \(a_t\), \[ \ABS{ \EXP\bigl[ c_t(S_t, a_t) \bigm| H_t = h_t,\, A_t = a_t \bigr] - c_t\bigl(σ_t(h_t), a_t\bigr) } \le L^c_t\, η_t(h_t) = ε_t(h_t). \]

NoteProof

Write \(\hat s_t = σ_t(h_t)\). By the triangle inequality and \(\eqref{eq:ce-Lip-c}\), \[\begin{align*} \ABS{ \EXP\bigl[ c_t(S_t, a_t) \bigm| h_t \bigr] - c_t(\hat s_t, a_t) } &\le \EXP\bigl[ \ABS{c_t(S_t, a_t) - c_t(\hat s_t, a_t)} \bigm| h_t \bigr] \\ &\le L^c_t\, \EXP\bigl[ d_{\ALPHABET S}(S_t, \hat s_t) \bigm| h_t \bigr] \\ &= L^c_t\, η_t(h_t), \end{align*}\] where the last step is \(\eqref{eq:ce-eta-local}\).

9.4.2 (AP2): dynamics approximation

Lemma 9.2 (Dynamics approximation for CE) For every history \(h_t\) and action \(a_t\), writing \[ μ_t(\,\cdot\,) = \PR\bigl( \hat S_{t+1} \in \cdot \bigm| H_t = h_t,\, A_t = a_t \bigr), \quad ν_t(\,\cdot\,) = P_t\bigl(\cdot \bigm| σ_t(h_t), a_t\bigr) \] as in Definition 8.1, we have \[ W_1(μ_t, ν_t) \le L^P_t\, η_t(h_t) + \EXP\bigl[ η_{t+1}(H_{t+1}) \bigm| h_t, a_t \bigr] = δ_t(h_t, a_t). \]

NoteProof

Write \(\hat s_t = σ_t(h_t)\) and let \[ \tilde μ_t(\cdot \mid h_t, a_t) = \PR\bigl( S_{t+1} \in \cdot \bigm| h_t, a_t \bigr) = \sum_{s \in \ALPHABET S} P_t(\cdot \mid s, a_t)\, b_t[h_t](s) \] be the predictive law of the true next state. By the triangle inequality for \(W_1\), \[ W_1(μ_t, ν_t) \le W_1(μ_t, \tilde μ_t) + W_1(\tilde μ_t, ν_t). \]

Estimate versus true next state. Coupling \((\hat S_{t+1}, S_{t+1})\) under the joint conditional law given \((h_t, a_t)\) and using the tower property with \(\eqref{eq:ce-eta-local}\), \[\begin{align*} W_1(μ_t, \tilde μ_t) &\le \EXP\bigl[ d_{\ALPHABET S}\bigl(\hat S_{t+1}, S_{t+1}\bigr) \bigm| h_t, a_t \bigr] \\ &= \EXP\Bigl[ \EXP\bigl[ d_{\ALPHABET S}\bigl(σ_{t+1}(H_{t+1}), S_{t+1}\bigr) \bigm| H_{t+1} \bigr] \Bigm| h_t, a_t \Bigr] \\ &= \EXP\bigl[ η_{t+1}(H_{t+1}) \bigm| h_t, a_t \bigr]. \end{align*}\]

Dynamics Lipschitz under current error. Convexity of \(W_1\) and \(\eqref{eq:ce-Lip-P}\) give \[\begin{align*} W_1(\tilde μ_t, ν_t) &\le \sum_{s} b_t[h_t](s)\, W_1\bigl( P_t(\cdot \mid s, a_t),\, P_t(\cdot \mid \hat s_t, a_t) \bigr) \\ &\le L^P_t\, \EXP\bigl[ d_{\ALPHABET S}(S_t, \hat s_t) \bigm| h_t \bigr] \\ &= L^P_t\, η_t(h_t). \end{align*}\] Adding the two estimates yields the claim.

Together, Lemma 9.1 and Lemma 9.2 establish Theorem 9.1: the CE generator meets the stated pointwise \((ε_t,δ_t)\) tolerances, so \(π^{\mathrm{CE}} = \hat π^\star \circ σ\) is the corresponding lifted AIS policy.

9.5 Sample-path sub-optimality of CE

We view the history-based MDP as the true model, \(\ALPHABET M\) as the abstract model, and \(σ_t\) as the abstraction. Absolute-value envelopes (Theorem 6.4) give a history-dependent gap; signed envelopes (Theorem 6.3) are available when mismatches are one-sided.

A convenient explicit upper bound is \[\begin{equation}\label{eq:ce-alpha-local} α_t(h_t) = \sum_{τ=t}^{T} \EXP\Bigl[ ε_τ(H_τ) + δ_τ(H_τ, A_τ)\, \operatorname{Lip}(\hat V^\star_{τ+1}) \Bigm| H_t = h_t \Bigr], \end{equation}\] where \(ε_τ(H_τ) = L^c_τ η_τ(H_τ)\) and \(δ_τ(H_τ, A_τ)\) is as in Theorem 9.1, and we use \(\hat V^\star_{T+1} = 0\) (hence \(\operatorname{Lip}(\hat V^\star_{T+1}) = 0\)), so the \(τ=T\) dynamics term vanishes. (If an action-free upper bound is preferred, replace \(δ_τ(H_τ,A_τ)\) by \[ \bar δ_τ(H_τ) = L^P_τ\, η_τ(H_τ) + \sup_{a \in \ALPHABET A} \EXP\bigl[ η_{τ+1}(H_{τ+1}) \bigm| H_τ, a \bigr] \] before taking the conditional expectation.)

Theorem 9.2 (Sub-optimality of certainty equivalence (sample-path)) Under \(\eqref{eq:ce-Lip-c}\)\(\eqref{eq:ce-Lip-P}\), the certainty equivalent policy \(π^{\mathrm{CE}}\) satisfies \[ V^{π^{\mathrm{CE}}}_t(h_t) - V^\star_t(h_t) \le 2 α_t(h_t) \] for every history \(h_t\), where \(α_t(h_t)\) is given by \(\eqref{eq:ce-alpha-local}\) (or any absolute envelope from Theorem 6.4 driven by the pointwise mismatches of Theorem 9.1).

NoteProof sketch

Lemma 9.1 and Lemma 9.2 bound the pointwise mismatch of the history-based MDP against \(\ALPHABET M\) at \((h_t,a_t)\) by \[ \ABS{\bigl[\hat\MISMATCH_t \hat V^\star_{t+1}\bigr](h_t,a_t)} \le ε_t(h_t) + δ_t(h_t,a_t)\, \operatorname{Lip}(\hat V^\star_{t+1}). \] Theorem 6.4 then yields \(V^{\hat π^\star \circ σ}_t(h_t) - V^\star_t(h_t) \le 2 α_t(h_t)\) for any absolute envelope that upper-bounds those one-step terms; \(π^{\mathrm{CE}} = \hat π^\star \circ σ\) and optimality of \(V^\star\) give the claim. Unfolding the envelope recursion produces \(\eqref{eq:ce-alpha-local}\).

TipReading the bound

Sub-optimality scales with the local estimation error \(\{η_t(h_t)\}\), amplified by the Lipschitz constants of \(\ALPHABET M\). Histories with accurate estimates pay a small gap; histories with poor estimates pay more. Signed envelopes replace the factor two by \(α^+_t(h_t) - α^-_t(h_t)\) when the CE mismatch is one-sided (lecture 8).

9.6 Corollary: uniform / sup-norm bound

Collapsing history dependence by a supremum recovers the classical scalar statement. Define \[\begin{align} η_t &\coloneqq \sup_{h_t \in \ALPHABET H_t} η_t(h_t) = \NORM{η_t(\,\cdot\,)}_∞, \label{eq:ce-eta} \\ ε_t &\coloneqq L^c_t\, η_t, \label{eq:ce-eps-unif} \\ δ_t &\coloneqq L^P_t\, η_t + η_{t+1}, \label{eq:ce-delta-unif} \end{align}\] and the uniform aggregate \[\begin{equation}\label{eq:ce-alpha-unif} α_t^{\mathrm{unif}} = \sum_{τ=t}^{T} \bigl[ ε_τ + δ_τ\, \operatorname{Lip}(\hat V^\star_{τ+1}) \bigr]. \end{equation}\] For every \((h_t,a_t)\) one has \(ε_t(h_t) \le ε_t\) and \(δ_t(h_t,a_t) \le δ_t\), so \(α_t(h_t) \le α_t^{\mathrm{unif}}\).

Corollary 9.1 (Uniform sub-optimality of certainty equivalence) Under \(\eqref{eq:ce-Lip-c}\)\(\eqref{eq:ce-Lip-P}\), for every history \(h_t\), \[ V^{π^{\mathrm{CE}}}_t(h_t) - V^\star_t(h_t) \le 2 α_t^{\mathrm{unif}}. \]

NoteProof

The constant envelopes \(α_t(\,\cdot\,) \equiv α_t^{\mathrm{unif}}\) are feasible for Theorem 9.2, since \(ε_τ(h_τ) \le ε_τ\) and \(δ_τ(h_τ,a_τ) \le δ_τ\) for all histories and actions. Alternatively, the same scalars \((ε_t,δ_t)\) are uniform AIS tolerances in the sense of Definition 8.1, and Theorem 8.1 / Corollary 8.1 (with \(ρ_{\mathfrak F} = \operatorname{Lip}\)) give the identical bound. Either route is a corollary of the history-local development: one takes \(\sup_{h_t}\) after the pointwise (AP1)/(AP2) estimates.

TipWhen the corollary is enough

If estimation quality is essentially constant across histories (e.g. uniform bounded observation noise below), \(α_t(h_t)\) and \(α_t^{\mathrm{unif}}\) coincide up to the usual upper bounds, and Corollary 9.1 is the readable statement. When quality varies (erasures, intermittent degradation), Theorem 9.2 is strictly sharper on good histories.

9.7 Example: bounded observation noise

Example 9.1 (Bounded observation noise) Suppose \(\ALPHABET Y = \ALPHABET S\) and the observation always lands within radius \(r\) of the state: \[ d_{\ALPHABET S}(Y_t, S_t) \le r \qquad\text{for all realizations.} \] Take the estimator to be the last observation, \(σ_t(h_t) = y_t\). Then \[ η_t(h_t) = \EXP\bigl[ d_{\ALPHABET S}(S_t, y_t) \bigm| h_t \bigr] \le r \] for every history, so \(η_t \le r\) as well. Under the Lipschitz assumption \(\eqref{eq:ce-Lip-c}\)\(\eqref{eq:ce-Lip-P}\), \[ ε_t(h_t) \le r L^c_t, \qquad δ_t(h_t, a_t) \le r\bigl(1 + L^P_t\bigr), \] and likewise \(ε_t \le r L^c_t\), \(δ_t \le r(1 + L^P_t)\). Theorem 9.2 and Corollary 9.1 therefore both yield \[ V^{π^{\mathrm{CE}}}_t(h_t) - V^\star_t(h_t) \le 2 r\, L^{\ALPHABET M}_t, \] where \[ L^{\ALPHABET M}_t \coloneqq \sum_{τ=t}^{T} \Bigl( L^c_τ + \bigl(1 + L^P_τ\bigr)\, \operatorname{Lip}(\hat V^\star_{τ+1}) \Bigr) \] depends only on the smoothness of \(\ALPHABET M\) (with \(\hat V^\star_{T+1} = 0\)). As \(r \to 0\) (vanishing observation noise), the CE policy with \(σ_t(h_t) = y_t\) approaches optimality at a linear rate in \(r\).

NoteVariant: intermittent degradation

The same argument applies when observations are usually accurate but occasionally degraded. Suppose \(d_{\ALPHABET S}(Y_t, S_t) \le r\) on a “good” event and \(d_{\ALPHABET S}(Y_t, S_t) \le R\) otherwise. Then \(η_t(h_t) \le r\) on good histories and \(η_t(h_t) \le R\) on bad ones, so \(α_t(h_t)\) tracks the realized degradation through the conditional expectations in \(δ_t\) and in \(\eqref{eq:ce-alpha-local}\). This is the same mechanism as the packet-drop signed envelopes in lecture 8.

9.8 Example: learning / adaptive control

Example 9.2 (Learning with unknown parameters) Consider a parameterized fully observed MDP \(\ALPHABET M_X(θ)\) with state space \(\ALPHABET X\), action space \(\ALPHABET A\), time-invariant dynamics \(P_{X,θ}\), and time-invariant per-step cost \(c_θ\), where the unknown parameter \(θ \in Θ\) is drawn from a known prior, independent of the process noise. The controller observes the physical state and past actions (and, typically, realized costs), but not \(θ\).

This is a POMDP with joint state \(S_t = (X_t, θ)\). We assume that \(\ALPHABET X\) and \(Θ\) are metric spaces with metrics \(d_{\ALPHABET X}\) and \(d_Θ\), and define the metric on the product space as \[ d_{\ALPHABET S}\bigl((x,θ), (x',θ')\bigr) = d_{\ALPHABET X}(x,x') + d_Θ(θ,θ'). \] The associated fully observed model \(\ALPHABET M\) has state \((X_t, θ)\); its optimal policy \(\hat π^\star\) depends on both. Equivalently, \(\hat π^\star_t(x,θ)\) is the optimal policy of \(\ALPHABET M_X(θ)\) evaluated at \(x\).

Assume joint Lipschitz continuity in \((x,θ)\): there exist \(L^c, L^P \ge 0\) such that for all \(x,x' \in \ALPHABET X\), \(θ,θ' \in Θ\), and \(a \in \ALPHABET A\), \[\begin{align*} \ABS{c_θ(x,a) - c_{θ'}(x',a)} &\le L^c \bigl( d_{\ALPHABET X}(x,x') + d_Θ(θ,θ') \bigr), \\ W_1\bigl( P_{X,θ}(\cdot \mid x,a),\, P_{X,θ'}(\cdot \mid x',a) \bigr) &\le L^P \bigl( d_{\ALPHABET X}(x,x') + d_Θ(θ,θ') \bigr). \end{align*}\]

Take the estimator \(σ_t(h_t) = (x_t, \hat θ_t)\), where \(\hat θ_t\) is any estimate of \(θ\) from \(h_t\) (for example the MMSE estimate \(\EXP[θ \mid h_t]\)). The certainty equivalent policy is \[ π^{\mathrm{CE}}_t(h_t) = \hat π^\star_t\bigl(x_t, \hat θ_t\bigr). \] Because \(X_t\) is observed perfectly, \[ η_t(h_t) = \EXP\bigl[ d_Θ(θ, \hat θ_t) \bigm| H_t = h_t \bigr]. \] Under the Lipschitz assumption above, Theorem 9.1 specializes to \[ ε_t(h_t) = L^c\, η_t(h_t), \qquad δ_t(h_t, a_t) = L^P\, η_t(h_t) + \EXP\bigl[ η_{t+1}(H_{t+1}) \bigm| h_t, a_t \bigr], \] and Theorem 9.2 yields \[ V^{π^{\mathrm{CE}}}_t(h_t) - V^\star_t(h_t) \le 2 α_t(h_t) \] with \(α_t\) as in \(\eqref{eq:ce-alpha-local}\). The gap is therefore controlled by the local quality of parameter estimation along the realized history. If \(η_t(h_t)\) decays sufficiently fast (for example exponentially in \(t\)), the resulting upper bound remains useful even as \(T \to ∞\).

9.9 Some remarks

9.9.1 When CE is the wrong structure

Certainty equivalence is not appropriate for every POMDP. If the action set includes a costly sensing action that reveals the state, a CE policy never chooses to sense: its action is always \(\hat π^\star_t(\hat s_t)\) for the fully observed MDP \(\ALPHABET M\), which has no reason to pay for information it already treats as free and perfect.

This is exactly the tension in the when-to-observe example of lecture 7. There the optimal policy sometimes pays \(λ\) to reset the belief to a point mass. A CE rule with \(σ_t\) equal to the last observation (or any estimator that ignores the option value of sensing) will not reproduce that tradeoff. The AIS bound remains valid, but \(η_t(h_t)\) need not be small for such an estimator, and a different structured policy (one that occasionally senses) can do better.

9.9.2 CE with state abstraction

The same argument applies with a state abstraction \(φ \colon \ALPHABET S \to \hat{\ALPHABET S}\): estimate \(φ(S_t)\), solve the abstract MDP \(\widehat{\ALPHABET M}\) from lecture 5, and play \(\hat π^\star_t(σ_t(h_t))\). Abstraction error and filter error add inside an effective \(η_t(h_t)\); the sample-path AIS envelopes again supply the gap, and taking \(\sup_{h_t}\) recovers a uniform corollary. We do not derive the details here.

Notes

The generalized CE principle and the examples follow Bozkurt et al. (2025). The policy-gap step uses the sample-path AIS envelopes of lecture 8 (Beyond sup-norm; Theorem 6.4), based on Subramanian et al. (2022) and Bozkurt et al. (2026); the uniform corollary also matches Theorem 8.1.

References

Bozkurt, B., Mahajan, A., Nayyar, A., and Ouyang, Y. 2025. Sub-optimality bounds for certainty equivalent policies in partially observed systems. Preprint.
Bozkurt, B., Mahajan, A., Nayyar, A., and Ouyang, Y. 2026. Sample path dependent bounds for approximate information state.
Subramanian, J., Sinha, A., Seraj, R., and Mahajan, A. 2022. Approximate information state for approximate planning and reinforcement learning in partially observed systems. Journal of Machine Learning Research 23, 12, 1–83. Available at: http://jmlr.org/papers/v23/20-1165.html.