Hypothesis Selection Problem

• The hypothesis selection problem is the process of choosing among competing models, experiments, or sensors based on data, where selection is integrated with downstream statistical inference.
• It encompasses a variety of formulations including active testing, sequential sensing, information-source selection, and privacy-aware distributional methods.
• Recent advances focus on adaptive strategies, robust optimization, and methods to correct for selection bias in multiple testing and ordered inference.

Searching arXiv for recent and foundational papers related to hypothesis selection. The hypothesis selection problem concerns procedures that choose among competing hypotheses, models, experiments, or information sources using data, while optimizing statistical reliability, decision time, resource usage, privacy, or other task-specific criteria. Across the literature, the term covers several related but nonidentical formulations: selecting the true state from a finite hypothesis set under controlled sensing (Kartik et al., 2019); selecting among candidate distributions so as to approximate an unknown data-generating law (Gopi et al., 2020); selecting information sources or sensors that best support Bayesian classification under cost or robustness constraints (Bhargav et al., 2024, Bhargav et al., 20 Feb 2025); selecting a stopping point in an ordered testing sequence (G'Sell et al., 2013); and selecting only some families of hypotheses for follow-up inference while correcting for selection bias (Benjamini et al., 2011). A unifying theme is that the selection step is itself part of the inferential problem: the criterion for choosing a hypothesis, model, family, sensor, or experiment must be analyzed jointly with the downstream decision rule and its error guarantees.

1. Conceptual scope and formal variants

In its broadest form, the hypothesis selection problem asks how to choose an element of a finite or structured candidate set using observations from an unknown stochastic mechanism. One canonical formulation considers a finite hypothesis set

$$\mathcal{X} = \{0,1,\ldots,M-1\}, \quad X \in \mathcal{X},$$

with prior $\rho_1 \in \Delta\mathcal{X}$, a finite action or experiment set $\mathcal{U}$, and an observation model

$p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$

where the decision maker chooses experiments and then selects a hypothesis or declares an inconclusive outcome (Kartik et al., 2019). In this active-testing view, hypothesis selection is inseparable from experiment design.

A second formulation treats hypothesis selection as choosing one distribution from a finite class

$\mathcal{Q} = \{q_1,\dots,q_k\}$

to approximate an unknown distribution $p$ or $h$. The objective is to output $\hat q \in \mathcal{Q}$ satisfying a total-variation guarantee of the form

$$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$

with high probability (Gopi et al., 2020, Pour et al., 2023, Kamath et al., 19 Sep 2025, Aliakbarpour et al., 1 Jun 2025). In this setting, the emphasis is on sample complexity, privacy constraints, approximation factor, and computational complexity.

A third formulation arises in sequential sensing. There, one must identify the true hypothesis while also selecting which sensor or experiment to query at each step. In the stationary randomized setting of sequential detection,

$H_k,\quad k\in\{0,\dots,M-1\},$

sensor $\rho_1 \in \Delta\mathcal{X}$0 produces observations with density $\rho_1 \in \Delta\mathcal{X}$1, and the control variable is a distribution $\rho_1 \in \Delta\mathcal{X}$2 over sensors, with expected decision time

$\rho_1 \in \Delta\mathcal{X}$3

The resulting optimization is a resource-allocation problem over information sources embedded within sequential hypothesis testing (0909.1801). In the online setting, sensor choice can depend on the full observation history and is derived from a Bayesian optimal stopping problem (Li et al., 2016).

A fourth formulation is selective multiple testing. Here the data are first used to choose which families of hypotheses are “interesting,” and only those selected families are tested further. The challenge is that the selection step induces bias, so ordinary within-family error control does not imply control over the selected families. The target quantity becomes the expected average error over selected families,

$\rho_1 \in \Delta\mathcal{X}$4

rather than an unconditional familywise average (Benjamini et al., 2011).

A fifth formulation appears in ordered testing and sequential model selection. With ordered nulls

$\rho_1 \in \Delta\mathcal{X}$5

the admissible rejection set is restricted to a prefix $\rho_1 \in \Delta\mathcal{X}$6. Selecting $\rho_1 \in \Delta\mathcal{X}$7 is then equivalent to selecting a model or stopping point along a path, subject to FDR control (G'Sell et al., 2013). This suggests that the hypothesis selection problem includes data-dependent choice of model complexity, not only choice among mutually exclusive states.

These variants are not interchangeable, but they share a common structure: a candidate set, a data-acquisition or evidence-aggregation mechanism, a selection rule, and a formal performance criterion.

2. Active and sequential hypothesis selection

In active hypothesis testing with fixed horizon $\rho_1 \in \Delta\mathcal{X}$8, the agent chooses experiments $\rho_1 \in \Delta\mathcal{X}$9 over $\mathcal{U}$0 stages and then outputs $\mathcal{U}$1, where $\mathcal{U}$2 denotes “inconclusive” (Kartik et al., 2019). Two formulations are central.

In the asymmetric formulation, one fixes a particular hypothesis $\mathcal{U}$3 and minimizes the probability of incorrectly declaring $\mathcal{U}$4,

$\mathcal{U}$5

subject to the correct-inference constraint

$\mathcal{U}$6

with $\mathcal{U}$7 and $\mathcal{U}$8 (Kartik et al., 2019). The optimal exponent is characterized by

$\mathcal{U}$9

and

$p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$0

This is an active Chernoff–Stein analogue (Kartik et al., 2019).

In the symmetric formulation, the objective is to minimize the overall misclassification probability

$p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$1

subject to $p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$2 for all $p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$3. The exponent becomes

$p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$4

so the hardest hypothesis dominates the asymptotic rate (Kartik et al., 2019).

A notable feature of this framework is its treatment of confidence. With posterior

$p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$5

confidence in hypothesis $p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$6 is

$p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$7

The increment in confidence is exactly the log-likelihood ratio between the trajectory law under $p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$8 and the law under $p_i^u(y) \doteq \mathbb{P}(Y_n=y \mid X=i, U_n=u),$9 (Kartik et al., 2019). Decision rules then become threshold tests on posterior log-odds, and experiment-selection strategies are designed to maximize expected confidence growth.

The paper also departs from classical randomized open-loop strategies by constructing fully deterministic and adaptive experiment-selection policies that satisfy a criterion based on

$\mathcal{Q} = \{q_1,\dots,q_k\}$0

where

$\mathcal{Q} = \{q_1,\dots,q_k\}$1

These deterministic adaptive strategies are asymptotically optimal and can perform substantially better in finite horizon than open-loop randomized strategies (Kartik et al., 2019). This suggests that, in active settings, the hypothesis selection problem is fundamentally a joint inference-and-control problem rather than a static classification problem.

Sequential sensing adds another layer. In the sensor-selection formulation with randomized stationary policies, the fusion center observes one sensor at a time and runs SPRT or MSPRT. For binary hypotheses, the expected sample size scales as

$\mathcal{Q} = \{q_1,\dots,q_k\}$2

so the denominator is an effective information rate under the sensor mixture $\mathcal{Q} = \{q_1,\dots,q_k\}$3 (0909.1801). For multi-hypothesis MSPRT, the analogous rate is

$\mathcal{Q} = \{q_1,\dots,q_k\}$4

The resulting optimization problems are linear-fractional in $\mathcal{Q} = \{q_1,\dots,q_k\}$5, and for generic binary problems the optimal randomized policy uses at most two sensors, while in the $\mathcal{Q} = \{q_1,\dots,q_k\}$6-hypothesis case an optimal policy uses at most $\mathcal{Q} = \{q_1,\dots,q_k\}$7 sensors (0909.1801).

The online usage-constrained variant replaces static mixing by belief-dependent sensor choice. There are $\mathcal{Q} = \{q_1,\dots,q_k\}$8 sensors, binary hypotheses, and usage constraints

$\mathcal{Q} = \{q_1,\dots,q_k\}$9

Lagrange multipliers convert these constraints into per-sample costs

$p$0

and the problem becomes a Bayesian optimal stopping problem on the posterior $p$1 (Li et al., 2016). In finite horizon, the optimal stopping boundaries are time-varying; in infinite horizon, the optimal procedure is an SPRT with constant thresholds and a stationary sensor-selection policy $p$2, where $p$3 is the accumulated log-likelihood ratio (Li et al., 2016). This suggests that constrained hypothesis selection admits an exact MDP formulation whose state variable is posterior belief.

3. Information-source and sensor selection as hypothesis selection

The sensor-selection literature makes explicit a broader interpretation of hypothesis selection: selecting experiments or information sources that best discriminate among hidden states. In the stationary randomized sequential setting, the decision time under hypothesis $p$4 takes the form

$p$5

where $p$6 is processing time and $p$7 is an information constant derived from KL divergences and decision thresholds (0909.1801). Three criteria are emphasized: conditioned decision time, worst-case decision time, and average decision time. Their structure yields sharp support-size results: one sensor suffices for conditioned optimization, at most two sensors suffice for generic binary worst-case or average optimization, and at most $p$8 sensors suffice in the $p$9-hypothesis average case under a generic rank condition (0909.1801). The selection problem is therefore low-dimensional even when the sensor network is large.

More recent work generalizes the objective from expected decision time to misclassification penalties. In centralized Bayesian classification with finite hypothesis set

$h$0

and information sources

$h$1

each source $h$2 has likelihood $h$3, and conditional independence implies

$h$4

for any selected set $h$5 (Bhargav et al., 2024, Bhargav et al., 20 Feb 2025). The crucial object is the observationally equivalent set

$h$6

which is equivalently the set of hypotheses whose induced observation distributions coincide with that of $h$7 under $h$8 (Bhargav et al., 2024, Bhargav et al., 20 Feb 2025). Under Bayesian learning, posterior mass eventually concentrates uniformly on $h$9 when the true state is $\hat q \in \mathcal{Q}$0, and all hypotheses outside this set have posterior going to zero (Bhargav et al., 20 Feb 2025).

A misclassification penalty matrix

$\hat q \in \mathcal{Q}$1

encodes non-uniform costs, with $\hat q \in \mathcal{Q}$2, $\hat q \in \mathcal{Q}$3, and each row summing to one (Bhargav et al., 2024, Bhargav et al., 20 Feb 2025). Two metrics then arise. The maximum-penalty metric for true state $\hat q \in \mathcal{Q}$4 is

$\hat q \in \mathcal{Q}$5

and the aggregate objective is

$\hat q \in \mathcal{Q}$6

Under distinctness assumptions on each row of $\hat q \in \mathcal{Q}$7, these set functions are approximately submodular, with submodularity ratio

$\hat q \in \mathcal{Q}$8

where $\hat q \in \mathcal{Q}$9 and $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$0 are minimal and maximal rowwise penalty gaps (Bhargav et al., 2024). This yields greedy guarantees for minimum-cost and budget-constrained information selection, but the guarantees can be weak when penalty gaps are small.

To address this, the total-penalty metric is introduced: $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$1 Unlike the maximum-penalty objective, $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$2 is submodular, and sums of these functions remain submodular (Bhargav et al., 2024, Bhargav et al., 20 Feb 2025). This supports near-optimal greedy algorithms under standard submodular set cover and submodular maximization analyses (Bhargav et al., 2024). A plausible implication is that the total-penalty formulation provides a structurally more stable surrogate for cost-sensitive hypothesis selection than the exact worst-case penalty.

The robust extension introduces adversarial deletions or failures. If the designer selects $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$3 with $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$4, and an adversary can remove up to $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$5 selected sources, the robust objective becomes

$$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$6

or its submodular surrogate version with $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$7 replacing $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$8 (Bhargav et al., 20 Feb 2025). The robust greedy algorithms build an “oblivious” or “attack” set together with a greedy remainder and obtain curvature-dependent approximation guarantees (Bhargav et al., 20 Feb 2025). This places hypothesis selection within the broader theory of resilient combinatorial optimization.

4. Distributional hypothesis selection and privacy

A major line of work studies hypothesis selection for a finite class of candidate distributions under total variation distance. In the non-private setting, the sample complexity is logarithmic in the number of candidates, and recent work shows that the optimal approximation factor is $$d_{\mathrm{TV}}(p,\hat q) \le C \cdot \min_{q\in\mathcal{Q}} d_{\mathrm{TV}}(p,q) + \alpha,$$9 (Aliakbarpour et al., 1 Jun 2025). The quality criterion is

$H_k,\quad k\in\{0,\dots,M-1\},$0

with $H_k,\quad k\in\{0,\dots,M-1\},$1 being the best possible approximation factor in the standard hypothesis-selection model (Aliakbarpour et al., 1 Jun 2025).

Under local differential privacy, the picture changes sharply. In the LDP setting, users each hold one sample from the unknown distribution and must locally randomize before communication. The problem is to output $H_k,\quad k\in\{0,\dots,M-1\},$2 with

$H_k,\quad k\in\{0,\dots,M-1\},$3

using only $H_k,\quad k\in\{0,\dots,M-1\},$4-LDP access (Gopi et al., 2020, Pour et al., 2023, Kamath et al., 19 Sep 2025). The foundational lower bound is that any $H_k,\quad k\in\{0,\dots,M-1\},$5-LDP algorithm for $H_k,\quad k\in\{0,\dots,M-1\},$6-wise simple hypothesis testing requires

$H_k,\quad k\in\{0,\dots,M-1\},$7

samples, even with full interaction (Gopi et al., 2020). This is an exponential deterioration in the dependence on $H_k,\quad k\in\{0,\dots,M-1\},$8 relative to the $H_k,\quad k\in\{0,\dots,M-1\},$9 non-private or centrally private rate (Gopi et al., 2020). A key conceptual point is that local privacy prevents aggressive sample reuse across many pairwise comparisons.

The first general LDP algorithms for agnostic selection achieved $\rho_1 \in \Delta\mathcal{X}$00 samples but used $\rho_1 \in \Delta\mathcal{X}$01 rounds of interaction (Gopi et al., 2020). They reduce the problem to maximum selection with adversarial comparators and use private Scheffé tests to simulate pairwise comparisons. For the general case, the guarantee is

$\rho_1 \in \Delta\mathcal{X}$02

with sample complexity

$\rho_1 \in \Delta\mathcal{X}$03

and $\rho_1 \in \Delta\mathcal{X}$04 rounds (Gopi et al., 2020). This demonstrated that interaction can reduce round complexity and achieve near-linear-in-$\rho_1 \in \Delta\mathcal{X}$05 sample complexity.

The subsequent interactive LDP result established the optimal sample complexity

$\rho_1 \in \Delta\mathcal{X}$06

for constant failure probability, while still allowing only

$\rho_1 \in \Delta\mathcal{X}$07

rounds (Pour et al., 2023). The algorithm BOKSERR is built from three stages—Boosted Knockout, Boosted Sequential Round-Robin, and a final MDE-variant—and the analysis introduces the notion of critical queries. An SQ algorithm is said to rely on only a small number of critical queries if correctness depends on only a small subset of the many queries it asks. This permits SQ-to-LDP simulation with a $\rho_1 \in \Delta\mathcal{X}$08 term governed by the number of critical queries rather than by the total number of queries (Pour et al., 2023). The resulting theorem gives an $\rho_1 \in \Delta\mathcal{X}$09-LDP algorithm using

$\rho_1 \in \Delta\mathcal{X}$10

samples, with probability at least $\rho_1 \in \Delta\mathcal{X}$11, and guarantee

$\rho_1 \in \Delta\mathcal{X}$12

(Pour et al., 2023). The approximation factor $\rho_1 \in \Delta\mathcal{X}$13 is larger than the non-private optimum $\rho_1 \in \Delta\mathcal{X}$14, but the sample complexity is optimal for interactive LDP.

A later non-adaptive improvement introduces the Scheffé graph (Kamath et al., 19 Sep 2025). Vertices correspond to unordered pairs $\rho_1 \in \Delta\mathcal{X}$15, hence to signed Scheffé sets $\rho_1 \in \Delta\mathcal{X}$16. There is a directed edge $\rho_1 \in \Delta\mathcal{X}$17 when

$\rho_1 \in \Delta\mathcal{X}$18

so the test associated with one pair approximately distinguishes another pair (Kamath et al., 19 Sep 2025). A dominating set in this graph yields a smaller set $\rho_1 \in \Delta\mathcal{X}$19 of functionals sufficient for a relaxed minimum distance estimator. The resulting non-adaptive $\rho_1 \in \Delta\mathcal{X}$20-LDP algorithm uses

$\rho_1 \in \Delta\mathcal{X}$21

samples and returns $\rho_1 \in \Delta\mathcal{X}$22 satisfying

$\rho_1 \in \Delta\mathcal{X}$23

(Kamath et al., 19 Sep 2025). This improves over the naive $\rho_1 \in \Delta\mathcal{X}$24 non-interactive strategy and suggests that structural reuse among Scheffé tests can partially substitute for interaction.

Under central differential privacy, the situation is more favorable. The recent nearly-linear-time private algorithm achieves the optimal approximation factor $\rho_1 \in \Delta\mathcal{X}$25 while remaining polylogarithmic in sample complexity and nearly linear in the number of hypotheses (Aliakbarpour et al., 1 Jun 2025). It combines empirical semi-distances, updateable proxy distances $\rho_1 \in \Delta\mathcal{X}$26, the exponential mechanism over hypotheses with small proxy distance, and a stable quantile-based score used inside a sparse-vector routine to find prompting hypotheses (Aliakbarpour et al., 1 Jun 2025). The final theorem gives sample complexity

$\rho_1 \in \Delta\mathcal{X}$27

and nearly-linear running time in $\rho_1 \in \Delta\mathcal{X}$28, resolving an open question on whether one can simultaneously obtain optimal factor $\rho_1 \in \Delta\mathcal{X}$29, polylogarithmic sample complexity, and nearly-linear time under central DP (Aliakbarpour et al., 1 Jun 2025). This suggests a strong separation between local and central privacy for hypothesis selection.

5. Selection bias, multiple testing, and ordered selection

The hypothesis selection problem also appears when the data are used to select which hypotheses will be tested or reported. In the multiple-family setting, hypotheses are partitioned into families, each family $\rho_1 \in \Delta\mathcal{X}$30 having p-values

$\rho_1 \in \Delta\mathcal{X}$31

and the entire collection is $\rho_1 \in \Delta\mathcal{X}$32 (Benjamini et al., 2011). If families are selected based on the same data that will later be used for testing inside them, neither ordinary within-family error control nor pooled control over all hypotheses guarantees control over the selected families (Benjamini et al., 2011).

The central target becomes

$\rho_1 \in \Delta\mathcal{X}$33

where $\rho_1 \in \Delta\mathcal{X}$34 is the random set of selected families and $\rho_1 \in \Delta\mathcal{X}$35 is a within-family error measure such as $\rho_1 \in \Delta\mathcal{X}$36, $\rho_1 \in \Delta\mathcal{X}$37, or $\rho_1 \in \Delta\mathcal{X}$38 (Benjamini et al., 2011). The paper proves that if the selection rule is simple and p-values across families are independent, then testing each selected family at level

$\rho_1 \in \Delta\mathcal{X}$39

controls $\rho_1 \in \Delta\mathcal{X}$40 (Benjamini et al., 2011). For general selection rules, the adjustment uses the minimal compatible number of selected families

$\rho_1 \in \Delta\mathcal{X}$41

and the corresponding level

$\rho_1 \in \Delta\mathcal{X}$42

(Benjamini et al., 2011). This is a direct solution to a selective-inference version of hypothesis selection: one must correct the inferential target to account for the data-driven selection of which hypotheses are examined at all.

Ordered testing gives a different selection-bias problem. One observes ordered p-values $\rho_1 \in \Delta\mathcal{X}$43 and is restricted to rejecting a contiguous prefix $\rho_1 \in \Delta\mathcal{X}$44. The goal is to choose $\rho_1 \in \Delta\mathcal{X}$45 with FDR control

$\rho_1 \in \Delta\mathcal{X}$46

(G'Sell et al., 2013). The ForwardStop rule uses

$\rho_1 \in \Delta\mathcal{X}$47

and sets

$\rho_1 \in \Delta\mathcal{X}$48

with FDR control under independent null p-values even when signals and nulls are arbitrarily interspersed (G'Sell et al., 2013). StrongStop uses a reverse-time Rènyi representation: $\rho_1 \in \Delta\mathcal{X}$49 and then applies a BH-type condition

$\rho_1 \in \Delta\mathcal{X}$50

Under perfect separation—nonnulls first, nulls last—it controls FWER and hence FDR (G'Sell et al., 2013). TailStop further specializes to harmonic null statistics

$\rho_1 \in \Delta\mathcal{X}$51

and attains exact FDR in that ideal model (G'Sell et al., 2013). These procedures formalize model-path selection as an ordered hypothesis selection problem.

A related but conceptually distinct setting is “Testing One Hypothesis Multiple times,” where a nuisance parameter appears only under the alternative. For each $\rho_1 \in \Delta\mathcal{X}$52 one computes a local statistic $\rho_1 \in \Delta\mathcal{X}$53, then selects

$\rho_1 \in \Delta\mathcal{X}$54

and asks whether the supremum

$\rho_1 \in \Delta\mathcal{X}$55

is unexpectedly large under the null (Algeri et al., 2018). In the multidimensional extension, the global p-value is approximated via expected Euler characteristics of excursion sets of the limiting random field,

$\rho_1 \in \Delta\mathcal{X}$56

This is another form of post-selection correction: the alternative is chosen after scanning over a continuum of nuisance parameters, so the null distribution must be that of the maximum over the search domain (Algeri et al., 2018).

6. Hypothesis sets as beliefs in partial observability

In robotic manipulation under partial observability, the hypothesis selection problem appears as active disambiguation of an object’s pose. The underlying POMDP state is

$\rho_1 \in \Delta\mathcal{X}$57

where $\rho_1 \in \Delta\mathcal{X}$58 is the target-object pose and $\rho_1 \in \Delta\mathcal{X}$59 is the environment structure (Sankaran et al., 2015). Because explicit Bayesian filtering over this state is intractable, the belief is replaced by a discrete hypothesis set

$\rho_1 \in \Delta\mathcal{X}$60

where $\rho_1 \in \Delta\mathcal{X}$61 is a perception routine that outputs candidate object poses consistent with the current observation $\rho_1 \in \Delta\mathcal{X}$62 (Sankaran et al., 2015). The update is not Bayesian recursion but recomputation: $\rho_1 \in \Delta\mathcal{X}$63

Action selection is based on features of $\rho_1 \in \Delta\mathcal{X}$64. The policy is

$\rho_1 \in \Delta\mathcal{X}$65

learned from expert demonstrations using SVM classifiers (Sankaran et al., 2015). In minesweeper, the hypothesis set consists of all H-structure placements consistent with observed cells, and an inverse distance transform on the hypothesis set provides the feature map. In the robotic tabletop task, hypotheses are pose candidates generated from clustered RGB-D data, projected to the support surface, again yielding a feature map that drives local action selection (Sankaran et al., 2015).

Termination occurs when

$\rho_1 \in \Delta\mathcal{X}$66

at which point the target pose is deemed fully resolved and the terminal action $\rho_1 \in \Delta\mathcal{X}$67 can be executed (Sankaran et al., 2015). This is an explicitly set-valued hypothesis selection process: the agent acts to reduce the cardinality of a feasible hypothesis set until only one remains. A plausible implication is that many active-perception problems admit tractable set-valued approximations of belief-space planning, provided the perception module can regenerate sufficiently informative candidate sets.

A superficially different but structurally related instance appears in speech recognition with light user feedback. Here the “hypotheses” are alternative transcriptions $\rho_1 \in \Delta\mathcal{X}$68, and the recognition system is treated as a policy

$\rho_1 \in \Delta\mathcal{X}$69

mapping acoustic features $\rho_1 \in \Delta\mathcal{X}$70 to distributions over word sequences (Kato et al., 2017). The user is shown two hypotheses $\rho_1 \in \Delta\mathcal{X}$71 and $\rho_1 \in \Delta\mathcal{X}$72 and selects the better one, giving a binary reward $\rho_1 \in \Delta\mathcal{X}$73. The resulting hypothesis-selection gradient is

$\rho_1 \in \Delta\mathcal{X}$74

Thus the selected hypothesis is upweighted and the rejected one is downweighted (Kato et al., 2017). In experiments, this pairwise-selection feedback improved WER over unsupervised adaptation and remained effective under simulated user errors (Kato et al., 2017). This suggests that pairwise preference elicitation is a viable hypothesis-selection signal even when full labels are unavailable.

7. Structural themes, limitations, and open directions

Several structural themes recur across these formulations.

First, many objective functions are built from divergences or likelihood ratios. In sequential sensing, KL divergences define effective information rates (0909.1801, Li et al., 2016). In active fixed-horizon testing, the exponents are max–min KL values $\rho_1 \in \Delta\mathcal{X}$75 (Kartik et al., 2019). In distributional selection, total variation distance and Scheffé sets determine achievable approximation guarantees (Gopi et al., 2020, Pour et al., 2023, Kamath et al., 19 Sep 2025, Aliakbarpour et al., 1 Jun 2025). In selective multiple testing, the object of control is not distance to truth but an error functional averaged over selected units (Benjamini et al., 2011).

Second, the role of adaptivity varies sharply by setting. Fully deterministic adaptive experiment selection improves finite-horizon active testing (Kartik et al., 2019). Online belief-dependent sensing is exactly optimal under sensor usage constraints (Li et al., 2016). By contrast, in locally private distributional selection, interaction mainly affects sample complexity and round complexity rather than the final decision rule, and the most recent non-adaptive result still lags the interactive optimum in $\rho_1 \in \Delta\mathcal{X}$76-dependence (Pour et al., 2023, Kamath et al., 19 Sep 2025).

Third, structural sparsity is common. Optimal randomized sensor policies use small supports—one, two, or at most $\rho_1 \in \Delta\mathcal{X}$77 sensors (0909.1801). Robust information selection is amenable to greedy algorithms because the relevant set functions are weakly submodular or submodular (Bhargav et al., 2024, Bhargav et al., 20 Feb 2025). The Scheffé-graph approach compresses the full $\rho_1 \in \Delta\mathcal{X}$78 pairwise comparison structure into a dominating set of size $\rho_1 \in \Delta\mathcal{X}$79 (Kamath et al., 19 Sep 2025). This suggests that large candidate spaces often possess latent combinatorial redundancy.

Fourth, selection itself changes the inferential target. In family-wise selection, one must control error over selected families, not all families (Benjamini et al., 2011). In ordered testing, the stopping index $\rho_1 \in \Delta\mathcal{X}$80 is the inferential object, and p-values must be calibrated for prefix selection (G'Sell et al., 2013). In TOHM, the null distribution must be that of the supremum over the search space, not that of any single local test (Algeri et al., 2018).

The limitations are equally recurrent. Independence assumptions are strong in ordered FDR control (G'Sell et al., 2013) and in several LDP constructions (Gopi et al., 2020, Pour et al., 2023). Exact model knowledge is assumed in sensor-selection and Bayesian information-selection papers (0909.1801, Li et al., 2016, Bhargav et al., 2024, Bhargav et al., 20 Feb 2025). Weak submodularity guarantees can be fragile when penalty gaps are small (Bhargav et al., 2024). Locally private methods still face a substantial gap between interactive and non-interactive complexity (Pour et al., 2023, Kamath et al., 19 Sep 2025). Central-DP nearly-linear-time selection attains the optimal factor $\rho_1 \in \Delta\mathcal{X}$81 but retains polynomial dependence on $\rho_1 \in \Delta\mathcal{X}$82 and additional logarithmic overheads (Aliakbarpour et al., 1 Jun 2025). In robotic hypothesis-set planning, the belief approximation discards posterior weights and depends heavily on the quality of the perception map $\rho_1 \in \Delta\mathcal{X}$83 (Sankaran et al., 2015).

Open directions follow naturally from these tensions. A plausible open question is whether non-interactive locally private hypothesis selection can achieve $\rho_1 \in \Delta\mathcal{X}$84 sample complexity with a constant-factor approximation comparable to the interactive optimum; current barriers show that new structural ideas would be needed beyond the triangular properties of the Scheffé graph (Kamath et al., 19 Sep 2025). Another is whether the approximation factor or round complexity in interactive LDP can be improved while keeping optimal sample complexity (Pour et al., 2023). In active sensing, extending exact belief-state solutions from binary to general multi-hypothesis constrained settings remains challenging (Li et al., 2016). In robust information selection, connecting surrogate submodular objectives more tightly to the original worst-case penalty objective is still largely empirical (Bhargav et al., 20 Feb 2025). In selective inference, broader dependence models and richer selection rules remain open (Benjamini et al., 2011, G'Sell et al., 2013).

Taken together, these results show that the hypothesis selection problem is not a single theorem or algorithmic template but a family of closely related inferential design problems. What unifies them is the necessity of treating selection as part of inference rather than as a preprocessing step. Whether the selected object is a hypothesis, a model, a distribution, a family, a sensor set, or a stopping point, the governing question is the same: how to use data to choose among alternatives while preserving quantitative guarantees on error, approximation, robustness, privacy, or efficiency.