Decision-Estimation Decomposition Overview

• Decision-Estimation Decomposition is a principle that separates task-specific decision rules from the methods used to estimate underlying latent quantities.
• It underpins frameworks such as minimax tradeoffs in reinforcement learning, invariance in stochastic choice, and architectural decoupling in multi-stakeholder evaluation.
• Empirical and theoretical studies show that decoupling estimation from decision-making improves scalability, performance bounds, and fairness in complex systems.

Decision-estimation decomposition denotes a family of principles that separate a task-specific decision rule from the estimation, information-acquisition, aggregation, or structural reduction needed to implement that rule. In recent work, this separation appears in several technically distinct forms: as a minimax tradeoff between suboptimality and divergence in sequential decision making, as an invariance axiom for stochastic choice under irrelevant appended decisions, and as explicit architectural decoupling of utility estimation from aggregation in multi-stakeholder evaluation (Chen et al., 2022, Sandomirskiy et al., 2023, Zheng et al., 26 May 2026). Taken together, these works suggest a common pattern: the “decision” component specifies how actions or outputs are ranked once relevant latent quantities are fixed, while the “estimation” component specifies how those quantities are learned, aggregated, or isolated from irrelevant structure.

1. Conceptual scope and terminological usage

A recurrent formulation is the classic idea, stated explicitly in recent multi-stakeholder alignment work, that decision making can be separated from estimating uncertain quantities: first estimate the relevant latent quantities as accurately as possible, then apply a fixed decision rule or aggregation rule (Zheng et al., 26 May 2026). In sequential decision making, this appears as a tradeoff between a task loss such as regret or suboptimality and an information term such as KL, Hellinger, or squared error divergence. In stochastic choice, it appears as invariance of marginal choice predictions to the inclusion of unrelated product-separated decisions. In optimization and representation learning, it appears as a decomposition of a large decision layer into smaller subproblems, or as a split between task-relevant directions and orthogonal residual structure.

A common misconception is that the phrase always denotes a single formal theorem. Several papers state the opposite. Decision-Directed Data Decomposition is described as conceptually aligned with decision-estimation decomposition but “not a formal Neyman–Pearson or statistical decision-theory decomposition”; the networked infrastructure paper states that it does not use the phrase in the classical sense, but formalizes a closely related idea; and the Lagrangian decomposition paper says that it does not explicitly use the term, while remaining structurally close to it (Davis et al., 2019, Diamond et al., 6 May 2026, Way et al., 7 Jun 2026). The literature therefore uses the label at multiple levels of generality: exact theorem, minimax complexity principle, architectural design rule, or analogy.

2. Decision-estimation tradeoffs in sequential decision making

In reinforcement learning and structured online decision making, the decomposition is formalized by a minimax objective that subtracts an estimation or information term from a decision loss. The generalized DEC framework writes a learning goal as

$$-Gdec_{\gamma}(\mathcal M,\mu) := \inf_{(p,q)\in\mathcal A} \sup_{M\in\mathcal M} \Big\{ SubOpt_M(p,q)-\gamma\,\mathbb E_{\pi\sim p,\tilde M\sim\mu} D^2(M(\pi),\tilde M(\pi)) \Big\},$$

thereby subsuming no-regret RL, PAC RL, reward-free learning, all-policy model estimation, and preference-based RL within a single decision-versus-divergence template (Chen et al., 2022). In this formulation, the first term is the decision side—task-specific suboptimality—while the second term is the estimation side—how much the current action distribution reveals about competing models.

The corresponding generic algorithmic pattern is Estimation-to-Decisions (E2D). It maintains a randomized model estimator $\mu^t$, solves a minimax problem for the current learning goal, samples an exploration policy, observes data, and updates $\mu^{t+1}$ by Tempered Aggregation (Chen et al., 2022). The resulting bounds preserve the same decomposition at the level of performance guarantees: a complexity term proportional to the relevant DEC variant and an estimation term controlled by the online estimator. The same paper also proves that generalized DEC is not merely sufficient but also a lower bound, so the decomposition is both algorithmic and complexity-theoretic.

A saddle-point reformulation sharpens this viewpoint. The average-constrained DEC is defined as

$$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$

where the min-player selects a sampling distribution and the max-player selects a confusing model mixture (Kirschner et al., 2024). This replaces an offset parameter by a confidence radius, restores strong duality on the adversary side, and yields Anytime-E2D, which optimizes the exploration-exploitation tradeoff online rather than only in the analysis. The same work emphasizes that informative but suboptimal actions can be optimal under this objective, because the decision layer is evaluated jointly with model discrimination.

3. Optimistic, partition-aware, and information-gain refinements

A major refinement replaces symmetric estimation requirements by optimism-aware ones. In model-free RL with value-function approximation, the optimistic DEC framework proves

$\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$

where the estimation error includes both a discrepancy term and an optimism bonus

$$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$

(Foster et al., 2022). The paper’s main structural result is that optimism helps for asymmetric divergences, especially Bellman-type discrepancies tailored to value-function learning, but not materially for symmetric divergences such as Hellinger distance. This sharply distinguishes when the estimation side can be weakened without damaging the decision side.

A second refinement makes the latent object itself coarser. In hybrid environments with fixed dynamics and adversarial rewards, the model aggregation approach partitions the joint space $M\times\Pi$ into cells $\Phi$, assumes the environment stays inside one unknown cell, and derives the regret bound

$$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$

equivalently

$$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$

This makes the tradeoff explicit: fine partitions increase $\mu^t$0 but can reduce decision difficulty, while coarse partitions do the opposite (Liu et al., 9 Feb 2025).

A third refinement removes optimism altogether. Dig-DEC is defined as a model-free coefficient with two information terms,

$\mu^t$1

The removal of optimism allows exploration to be driven purely by information gain and extends the framework to adversarial environments without explicit reward estimators (Liu et al., 10 Oct 2025). The same paper proves $\mu^t$2, gives a toy example where optimistic DEC suffers $\mu^t$3 but Dig-DEC achieves $\mu^t$4, obtains the first model-free regret bounds for hybrid MDPs with bandit feedback, and improves model-free rates from $\mu^t$5 to $\mu^t$6, from $\mu^t$7 to $\mu^t$8, and from $\mu^t$9 to $\mu^{t+1}$0 in the respective estimation settings (Liu et al., 10 Oct 2025).

4. Axiomatic decomposition in stochastic choice

In stochastic choice theory, the decomposition appears as an invariance principle over product menus. A decision problem is modeled as a menu $\mu^{t+1}$1, with finite action set $\mu^{t+1}$2 and outcome map $\mu^{t+1}$3, and a stochastic choice rule is a mapping

$\mu^{t+1}$4

(Sandomirskiy et al., 2023). For two unrelated decisions $\mu^{t+1}$5 and $\mu^{t+1}$6, the product menu is

$\mu^{t+1}$7

The key axiom is independence of irrelevant decisions (IID): $\mu^{t+1}$8 equivalently,

$\mu^{t+1}$9

This formalizes the claim that an appended unrelated decision can be dropped without changing the marginal prediction on the decision of interest.

The paper shows that monotonicity, continuity, and IID characterize exactly the mixed-logit family. Multinomial logit is

$$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$0

and mixed logit is

$$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$1

The main theorem states that if $$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$2 satisfies monotonicity, continuity, and IID on $$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$3, then $$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$4 for some probability measure $$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$5 supported on $$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$6 (Sandomirskiy et al., 2023). If IID is strengthened to decomposability,

$$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$7

then the only possibilities are pure multinomial logit rules.

This formulation is decomposition in a strict axiomatic sense. Conditional on a latent logit intensity $$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$8, decision behavior is logit; the estimation or aggregation layer is the mixing distribution $$\operatorname{acdec}_\epsilon(f) = \min_{\mu\in\mathcal P(\Pi)} \max_{\nu\in\mathcal P(\mathcal M)} \left\{ \mu \Delta \nu :\ \mu I_f \nu \le \epsilon^2 \right\},$$9. The proof strategy—extension to infinite products, exchangeability, De Finetti’s theorem for partially exchangeable processes, and representation as multinomial logit with random $\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$0—shows that invariance to irrelevant decisions imposes a global stochastic structure, not merely a local restriction (Sandomirskiy et al., 2023). A further consequence is that even though IID is stated only on product menus, the theorem implies a mixed-logit rule on all menus, including non-product menus.

5. Decomposing utility estimation from aggregation in multi-stakeholder alignment

In multi-stakeholder LLM alignment, the decomposition is explicit and architectural. The ideal target reward is a cardinal welfare function

$\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$1

but a holistic judge instead produces

$\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$2

where $\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$3 is utility-estimation noise, $\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$4 is implicit weight drift, and $\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$5 is residual score-level noise (Zheng et al., 26 May 2026). The paper identifies a distinct aggregation-specific instability—weighting noise—arising when one model simultaneously estimates per-stakeholder utilities and chooses the trade-off rule among them.

The variance decomposition makes the source of the instability explicit. Under zero-mean noise and independence assumptions,

$\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$6

The crucial point is that Term II is not reducible by merely improving the judge’s scoring skill unless the aggregation rule is fixed. Under exchangeable zero-sum drift,

$\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$7

so instability grows with weight-drift variance, cross-stakeholder utility dispersion, and the stakeholder-count factor $\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$8.

DecompR addresses this by separating the two stages. First, aggregation weights are fixed from query structure before candidate scoring. A difficulty proxy

$\text{Regret} \le \operatorname{ODec}_\gamma(\cM)\cdot T + \gamma\cdot \EstOptD^{\mathrm{Full}},$9

induces query-fixed weights

$$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$0

Second, per-role utilities are estimated independently,

$$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$1

This removes candidate-dependent weight drift by construction (Zheng et al., 26 May 2026).

The empirical evidence is aligned with the theory. On 60 seed multi-stakeholder travel-planning responses with stakeholder counts $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$2, presentation-induced score variance increases with stakeholder count for medium- and high-quality responses; variant types affecting stakeholder salience and trade-off framing are most unstable; and decomposed adaptive scoring has growth $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$3 versus $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$4 for decomposed uniform scoring, isolating the effect of online weight drift (Zheng et al., 26 May 2026). The paper also reports weight-induced shifts at $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$5 of about $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$6 of within-group score standard deviation on average, about $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$7 at the $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$8th percentile, and up to $$\gamma^{-1}\bigl(f_{M^\star}(\pi^\star)-f_{M_t}(\pi^\star)\bigr)$$9 in the worst case. In end-to-end GRPO on MR-TravelBench, DecompR performs best on Group Utility, Group Fairness, and Preference Completeness, and is the only reward construction reported to improve both utility and fairness.

6. Representation- and optimization-level decompositions

Decision-Directed Data Decomposition (D4) is a supervised linear decomposition of a representation into a task-relevant component and an orthogonal residual. Given $M\times\Pi$0, labels $M\times\Pi$1, and a learner of the form $M\times\Pi$2, D4 defines

$M\times\Pi$3

with $M\times\Pi$4 (Davis et al., 2019). Sequentially removing orthonormal decision directions $M\times\Pi$5 yields

$M\times\Pi$6

The paper is explicit that this is not classical statistical decision-versus-estimation separation; rather, it is a representation-level split between a decision-relevant subspace and a complementary residual. Its applications include post-hoc removal of gender information from image representations while leaving age prediction almost unchanged initially, improved generalization under a designed spurious correlation shift, and state-of-the-art debiasing behavior on word embeddings (Davis et al., 2019).

A different optimization-level construction inserts Lagrangian decomposition into decision-focused learning. The original predict-then-optimize problem is

$M\times\Pi$7

with $M\times\Pi$8 predicted from features (Way et al., 7 Jun 2026). The method duplicates variables across constraints, relaxes agreement constraints with multipliers $M\times\Pi$9, and obtains the dual bound

$\Phi$0

thereby replacing one large constrained problem with one main subproblem plus $\Phi$1 independent smaller subproblems. Training uses decomposed losses $\Phi$2 or $\Phi$3, fixes multipliers to $\Phi$4 to avoid bilevel differentiation, and supports single- and multiple-decomposition variants. The framework is designed as a wrapper around SPO+ and IMLE, and for quadratic portfolio optimization also admits an Exact variant with a closed-form main subproblem (Way et al., 7 Jun 2026).

The empirical claims are strictly about scalability and regret. On multi-dimensional knapsack and quadratic portfolio benchmarks, LD-based methods achieve the best or near-best test-set relative regret, handle instances up to 300 items with 10 constraints and 400 assets, and are reported to scale to roughly $\Phi$5–$\Phi$6 larger instances than typical prior work (Way et al., 7 Jun 2026). In the multiple-decomposition knapsack setting, using 10 workers reduces offline multiplier computation from 4218 seconds to 462 seconds, about a $\Phi$7 speedup. The paper also notes a limitation: because the losses are based on a relaxed problem, primal solution quality on the original constrained optimization problem is not guaranteed.

7. Exact local decomposition and domain-specific formulations

Several domain-specific papers instantiate the same logic as an exact structural reduction. In structural dynamic estimation of networked infrastructure, the Bellman operator becomes exactly separable under fixed group membership, localized transition dependence,

$\Phi$8

and additive separability of utility,

$\Phi$9

The theorem implies that the full value function is additive,

$$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$0

so the joint structural likelihood is preserved exactly while the Bellman step decomposes into independent group-level problems (Diamond et al., 6 May 2026). In the Titan GPU application, this makes fully structural estimation feasible over 14,344 GPU node locations grouped into 479 cages, yields significant localized interaction estimates $$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$1 and $$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$2, and shows that removing lagged replacement activity reduces replacements by 191 units while removing the failure mechanism reduces them by 276 units.

In RIS-assisted MIMO estimation, the decomposition is spectral rather than dynamic. Each RIS element is modeled as a rank-one keyhole-like channel,

$$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$3

and the full cascaded channel is

$$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$4

Two EVDs isolate the receive-side and transmit-side factors: $$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$5 from which $$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$6 and $$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$7 are reconstructed separately (Zegrar et al., 2020). The paper emphasizes reduced estimation time overhead: by a factor of $$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$8 in the first scheme and $$\mathbb E[Reg(\pi^\star_{1:T})] \le \frac{\log |\Phi|}{\eta} + T\max_{\rho \in \Delta(\Phi)} \max_{\nu\in\Delta(\Psi)} \min_{p\in\Delta(\Pi)} AIR^\Phi_{\rho,\eta}(p,\nu),$$9 in the enhanced scheme; for $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$0, $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$1, and $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$2, the reductions are about $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$3 and more than $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$4, respectively, with NMSE evaluated over 10,000 random channel realizations.

A behavioral-control analogue appears in optimal investment with herd behaviour. The follower’s optimal risky investment is decomposed as

$$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$5

where $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$6 and $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$7 are the agents’ standalone Merton-optimal decisions and

$$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$8

is interpreted as the follower’s investment opinion (Wang et al., 2024). This is not estimation in the statistical sense; it is a rational decision decomposition that isolates the herd effect into a single weight. The paper derives comparative statics such as $$\mathbb E[Reg] \;\lesssim\; \underbrace{\frac{\log|\Phi|}{\eta}}_{\text{estimation complexity}} + \underbrace{T\cdot DEC^{KL}_\eta(M,\Phi)}_{\text{decision complexity}}.$$9, $\mu^t$00, and, under a stated condition on $\mu^t$01, $\mu^t$02 and $\mu^t$03. Numerical experiments on Dow Jones Industrial Average data from Jan. 1972 to Dec. 2022 confirm that the optimal decision remains within the convex hull of the two rational benchmarks and moves toward the expert as the herd coefficient increases (Wang et al., 2024).

Across these formulations, the unifying feature is not a single canonical definition but a recurring separation principle. Sometimes the separation is exact and theorem-level, as with mixed-logit characterization or block-separable Bellman operators; sometimes it is a regret decomposition into decision complexity and estimation complexity; sometimes it is a modeling architecture that fixes aggregation and isolates per-component estimation; and sometimes it is only a closely related structural analogy. The literature therefore treats decision-estimation decomposition less as one object than as a general methodology for isolating what must be estimated from what must be decided.