Stochastic Boolean Function Evaluation

• Stochastic Boolean Function Evaluation is a sequential testing process that determines a Boolean function's value using cost-sensitive queries under known probability distributions.
• The methodology employs adaptive decision trees with certificate-based stopping rules to reduce testing costs without fully revealing the input.
• Applications span operations research and learning theory, with techniques addressing adaptivity gaps and submodular-goal approximations for efficient evaluations.

Searching arXiv for recent and foundational papers on SBFE. Stochastic Boolean Function Evaluation (SBFE) is the problem of determining the value of a known Boolean function $f$ on an unknown input $x\in\{0,1\}^n$ by sequentially testing variables, where each test incurs a cost and $x$ is drawn from a known probability distribution. In the standard formulation, the bits are independent and the objective is to minimize expected total testing cost; the process stops as soon as the observed partial assignment is sufficient to determine $f(x)$ (Hellerstein et al., 2022). The problem has also been studied in Operations Research as “sequential testing” of Boolean functions and in learning theory in the context of learning with attribute costs (Deshpande et al., 2013). More general formulations replace the product distribution by a conditional probability model over truth assignments and represent adaptive strategies as decision diagrams rather than trees (Zong et al., 23 Jun 2026).

1. Formal model and certificate semantics

In the standard SBFE model, one is given a Boolean function

$f:\{0,1\}^n\to\{0,1\},$

a positive cost vector

$c=[c_1,\ldots,c_n], \qquad c_i>0,$

and a probability vector

$p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$

with an unknown random input $x=(x_1,\ldots,x_n)\in\{0,1\}^n$ whose bits are independent and satisfy $\Pr(x_i=1)=p_i$ (Hellerstein et al., 2022). Writing $x\sim p$, the corresponding product distribution is

$x\in\{0,1\}^n$0

The only way to learn $x\in\{0,1\}^n$1 is to test variable $x\in\{0,1\}^n$2, paying cost $x\in\{0,1\}^n$3 (Hellerstein et al., 2022).

The stopping condition is certificate-based rather than full revelation. A set of tested indices determines $x\in\{0,1\}^n$4 if

$x\in\{0,1\}^n$5

Equivalently, in the partial-assignment language used for symmetric-function SBFE, a partial assignment $x\in\{0,1\}^n$6 is a certificate if all full assignments extending $x\in\{0,1\}^n$7 give the same function value (Hellerstein et al., 2022, Gkenosis et al., 2021). Thus SBFE seeks not necessarily to reveal all bits, but only enough bits to certify the function value.

The distinction between evaluation and related one-sided tasks is significant. In evaluation, the algorithm is not told $x\in\{0,1\}^n$8 and must continue until the observed partial assignment is a certificate. In verification, the claimed value $x\in\{0,1\}^n$9 is given in advance, and the algorithm only has to verify that claim. For symmetric Boolean functions, $x$0 always, equality is known for $x$1-of-$x$2 functions and for all symmetric functions in the unit-cost case, but arbitrary costs can yield $x$3 (Gkenosis et al., 2021).

2. Strategy models, objective functions, and representations

An adaptive SBFE strategy is naturally represented by a Boolean decision tree. Each internal node tests some variable, the outgoing edges correspond to observed outcomes, and each leaf is labeled by the determined function value (Gkenosis et al., 2021). For a fixed input $x$4, if $x$5 denotes the total testing cost incurred by strategy $x$6 until $x$7 is determined, then the expected cost under the product distribution is

$x$8

The optimal adaptive cost and optimal non-adaptive cost are

$x$9

where $f(x)$0 and $f(x)$1 are the sets of adaptive and non-adaptive strategies, respectively (Hellerstein et al., 2022).

A non-adaptive strategy is much more restrictive: it is simply a permutation of the variables. Tests are executed in that fixed order, but evaluation still stops early once $f(x)$2 becomes determined (Hellerstein et al., 2022). In the decision-tree view, every node at the same level tests the same variable. This distinction is algorithmically consequential because adaptive strategies can have lower expected cost, whereas non-adaptive strategies can be stored in linear space and may better exploit parallel resources; by contrast, adaptive strategies may require up to exponential space to store (Hellerstein et al., 2022).

More general exact formulations represent strategies as finite rooted DAGs. For a problem instance $f(x)$3, where $f(x)$4 is a Boolean formula, $f(x)$5 is a conditional probability distribution on variables, and $f(x)$6 is a positive cost function, a decision diagram is a directed acyclic graph whose internal nodes query variables and whose leaves are $f(x)$7 or $f(x)$8 once the current partial valuation makes the formula valid or unsatisfiable (Zong et al., 23 Jun 2026). In that setting, the expected cost of a diagram is formally

$f(x)$9

which generalizes the usual recursive expectation over a product distribution (Zong et al., 23 Jun 2026).

3. Submodular-goal-value framework

A central algorithmic framework reduces SBFE to Stochastic Submodular Set Cover (SSSC). The key object is a utility function

$f:\{0,1\}^n\to\{0,1\},$0

that is monotone and submodular, starts at $f:\{0,1\}^n\to\{0,1\},$1, and reaches a goal value $f:\{0,1\}^n\to\{0,1\},$2 exactly on partial assignments that contain a certificate (Deshpande et al., 2013, Bach et al., 2017). In this language, a goal function for $f:\{0,1\}^n\to\{0,1\},$3 satisfies

$f:\{0,1\}^n\to\{0,1\},$4

and the minimum such $f:\{0,1\}^n\to\{0,1\},$5 is the goal value $f:\{0,1\}^n\to\{0,1\},$6 (Bach et al., 2017). Once such a function is available, Adaptive Greedy yields an $f:\{0,1\}^n\to\{0,1\},$7-approximation for the corresponding SBFE instance (Bach et al., 2017).

The framework is structurally informative because goal value is tied to several Boolean complexity measures. Deshpande et al.’s basic bounds, recalled in the goal-value paper, include

$f:\{0,1\}^n\to\{0,1\},$8

hence

$f:\{0,1\}^n\to\{0,1\},$9

The same paper proves that if $c=[c_1,\ldots,c_n], \qquad c_i>0,$0 depends on $c=[c_1,\ldots,c_n], \qquad c_i>0,$1 variables, then $c=[c_1,\ldots,c_n], \qquad c_i>0,$2, and gives exact formulas for several important classes: $c=[c_1,\ldots,c_n], \qquad c_i>0,$3 for AND, OR, and XOR; $c=[c_1,\ldots,c_n], \qquad c_i>0,$4 for $c=[c_1,\ldots,c_n], \qquad c_i>0,$5-of-$c=[c_1,\ldots,c_n], \qquad c_i>0,$6; and $c=[c_1,\ldots,c_n], \qquad c_i>0,$7 for read-once functions (Bach et al., 2017). It also proves the universal upper bound $c=[c_1,\ldots,c_n], \qquad c_i>0,$8, while showing that this can still be too large to yield compelling approximation factors (Bach et al., 2017).

The same structural lens leads to concrete SBFE algorithms. For CDNF formulas, one constructs separate utility functions for falsified DNF terms and satisfied CNF clauses and combines them to obtain an $c=[c_1,\ldots,c_n], \qquad c_i>0,$9-approximation, where $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$0 is the number of CNF clauses and $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$1 the number of DNF terms; the same argument gives an $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$2-approximation for decision trees with $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$3 leaves (Deshpande et al., 2013). For linear threshold formulas with integer coefficients, the $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$4-value construction gives an $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$5-approximation, where $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$6, and Adaptive Dual Greedy strengthens this to a 3-approximation (Deshpande et al., 2013).

For symmetric Boolean functions, the goal-value method is particularly effective. A graph-based goal-function construction implies that the goal value of any symmetric Boolean function is strictly less than $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$7, which yields a polynomial-time $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$8-approximation via Adaptive Greedy (Gkenosis et al., 2021). The same paper also gives a simple polynomial-time $p=[p_1,\ldots,p_n], \qquad 0<p_i<1,$9-approximation, where $x=(x_1,\ldots,x_n)\in\{0,1\}^n$0 is the number of blocks of 0’s and 1’s in the value-vector representation of the symmetric function (Gkenosis et al., 2021).

The main limitation of the framework is that goal value can be exponentially large even for natural functions. The goal-value paper proves exponential exact values for certain read-once DNFs, and the SBFE/SSSC paper shows that the $x=(x_1,\ldots,x_n)\in\{0,1\}^n$1-value approach cannot yield sublinear approximations for some simple classes because any assignment-feasible utility function must have exponentially large $x=(x_1,\ldots,x_n)\in\{0,1\}^n$2 (Bach et al., 2017, Deshpande et al., 2013). This limitation motivated more specialized algorithms such as Adaptive Dual Greedy for linear threshold evaluation (Deshpande et al., 2013).

4. Function classes and algorithmic landscape

The algorithmic landscape of SBFE is highly class-dependent. Some classes admit exact polynomial-time algorithms; others admit only approximation guarantees, and in some cases those guarantees are measured against expected certificate cost rather than optimal evaluation cost (Allen et al., 2013).

Function class	Setting	Representative guarantee
Linear threshold formulas	Integer coefficients	3-approximation; also $x=(x_1,\ldots,x_n)\in\{0,1\}^n$3 via goal value (Deshpande et al., 2013)
CDNF formulas / decision trees	CNF with $x=(x_1,\ldots,x_n)\in\{0,1\}^n$4 clauses, DNF with $x=(x_1,\ldots,x_n)\in\{0,1\}^n$5 terms / $x=(x_1,\ldots,x_n)\in\{0,1\}^n$6 leaves	$x=(x_1,\ldots,x_n)\in\{0,1\}^n$7 / $x=(x_1,\ldots,x_n)\in\{0,1\}^n$8 (Deshpande et al., 2013)
Symmetric Boolean functions	Arbitrary positive costs, independent $x=(x_1,\ldots,x_n)\in\{0,1\}^n$9	Polynomial-time $\Pr(x_i=1)=p_i$0-approximation and $\Pr(x_i=1)=p_i$1-approximation (Gkenosis et al., 2021)
Monotone $\Pr(x_i=1)=p_i$2-DNF	Arbitrary costs, arbitrary product distribution	$\Pr(x_i=1)=p_i$3-approximation w.r.t. expected certificate cost (Allen et al., 2013)
Monotone $\Pr(x_i=1)=p_i$4-term DNF	Arbitrary costs, arbitrary product distribution	$\Pr(x_i=1)=p_i$5-approximation; exact $\Pr(x_i=1)=p_i$6 for constant $\Pr(x_i=1)=p_i$7 (Allen et al., 2013)
$\Pr(x_i=1)=p_i$8-of-$\Pr(x_i=1)=p_i$9 non-adaptive	Unit-cost case	PTAS (Nielsen et al., 8 Jul 2025)

For monotone $x\sim p$0-DNF and monotone $x\sim p$1-term DNF, the DNF-evaluation paper develops a modified round-robin protocol that combines a 0-certificate subroutine and a 1-certificate subroutine under arbitrary costs. For monotone $x\sim p$2-DNF this yields a polynomial-time approximation within $x\sim p$3 of expected certificate cost; for monotone $x\sim p$4-term DNF it yields a polynomial-time approximation within

$x\sim p$5

and an exact polynomial-time dynamic program for constant $x\sim p$6, with running time $x\sim p$7 (Allen et al., 2013). In the unit-cost uniform-distribution case, the same paper gives an exact $x\sim p$8-time algorithm for monotone $x\sim p$9-term DNF, which is polynomial when $x\in\{0,1\}^n$00 (Allen et al., 2013).

For $x\in\{0,1\}^n$01-of-$x\in\{0,1\}^n$02 functions, the adaptive side is classical: an exact polynomial-time algorithm is known, and both the symmetric-function paper and the 2025 non-adaptive paper treat this as one of the best-understood SBFE special cases (Gkenosis et al., 2021, Nielsen et al., 8 Jul 2025). The 2025 paper complements this by showing that the unit-cost non-adaptive problem admits a PTAS, stated as the first PTAS for an SBFE problem (Nielsen et al., 8 Jul 2025).

An important methodological caution is that expected certificate cost can be a weak proxy for optimal expected evaluation cost. The DNF-evaluation paper proves that for read-once DNF formulas with sufficiently long terms under the unit-cost uniform-distribution setting,

$x\in\{0,1\}^n$03

for any fixed $x\in\{0,1\}^n$04 satisfying the stated term-length condition (Allen et al., 2013). This shows that approximation factors relative to $x\in\{0,1\}^n$05 need not directly describe approximation relative to $x\in\{0,1\}^n$06.

5. Adaptivity, non-adaptivity, and adaptivity gaps

The adaptivity gap formalizes the benefit of allowing the next test to depend on previous outcomes. For a function class $x\in\{0,1\}^n$07, it is defined as

$x\in\{0,1\}^n$08

and every SBFE problem has adaptivity gap at most $x\in\{0,1\}^n$09, since testing variables in increasing order of $x\in\{0,1\}^n$10 gives a non-adaptive strategy within a factor $x\in\{0,1\}^n$11 of the optimal adaptive strategy (Hellerstein et al., 2022).

The known gap landscape is sharply nonuniform across function classes. For symmetric Boolean functions and linear threshold functions, prior work cited in the adaptivity-gap paper shows $x\in\{0,1\}^n$12 gaps via the stochastic score classification framework (Hellerstein et al., 2022). By contrast, that paper proves large lower bounds for formula classes. Its summary table gives the following results: for read-once DNF, the gap is $x\in\{0,1\}^n$13 under unit costs and the uniform distribution, $x\in\{0,1\}^n$14 under unit costs, and $x\in\{0,1\}^n$15 under the uniform distribution; for read-once formulas, it proves

$x\in\{0,1\}^n$16

for unit costs and equal probabilities; and for general DNF it proves $x\in\{0,1\}^n$17 under unit costs and the uniform distribution and $x\in\{0,1\}^n$18 under the uniform distribution with arbitrary costs (Hellerstein et al., 2022).

The $x\in\{0,1\}^n$19-of-$x\in\{0,1\}^n$20 case is again exceptional. The 2025 non-adaptive paper proves that the adaptivity gap of SBFE on $x\in\{0,1\}^n$21-of-$x\in\{0,1\}^n$22 functions is exactly 2, improving the previously known tight lower bound of $x\in\{0,1\}^n$23 for the unit-cost variant (Nielsen et al., 8 Jul 2025). The same paper shows that despite this worst-case separation, the optimal non-adaptive unit-cost policy can be approximated arbitrarily well in polynomial time via a PTAS (Nielsen et al., 8 Jul 2025).

The practical motivation for studying non-adaptivity is not only theoretical comparison. Non-adaptive strategies use linear space, whereas adaptive strategies may require exponential space to store, and non-adaptive orders may better exploit parallel resources (Hellerstein et al., 2022). The adaptivity-gap results therefore quantify a concrete trade-off between expected-cost optimality and implementational simplicity.

6. Generalizations, structured variants, and adjacent topics

Recent work has broadened SBFE beyond the standard independent-product setting. One direction studies exact optimization for arbitrary Boolean formulas under general conditional probability models. In that formulation, an instance is $x\in\{0,1\}^n$24, where $x\in\{0,1\}^n$25 is a propositional formula, $x\in\{0,1\}^n$26 is a positive observation-cost function, and $x\in\{0,1\}^n$27 is a conditional probability distribution induced by a joint distribution over truth assignments. The 2026 decision-diagram paper presents what it claims is the first practical exact branch-and-bound algorithm for this level of generality, with pruning, caching, and variable-selection heuristics, and proves that calculating the expected cost of the optimal decision diagram is $x\in\{0,1\}^n$28-hard, while the function problem is in FPSPACE and the bounded-cost decision problem is in $x\in\{0,1\}^n$29 (Zong et al., 23 Jun 2026).

A second direction exploits explicit combinatorial structure. The Limited Query Graph Connectivity Test is a graph-structured sequential-query model in which hidden edge states are queried until one identifies either an $x\in\{0,1\}^n$30-$x\in\{0,1\}^n$31 path of On edges or an $x\in\{0,1\}^n$32-$x\in\{0,1\}^n$33 cut of Off edges, with an additional budget $x\in\{0,1\}^n$34 after which the process stops regardless of whether connectivity is established. When the query limit is removed, the paper states that the problem is a special case of monotone SBFE. It proves $x\in\{0,1\}^n$35-hardness of computing the minimum expected number of queries and gives an empirically scalable exact algorithm based on iterative certificate generation and tree growth, together with strong heuristics and anytime lower bounds (Guo et al., 2023).

A third direction addresses correlations, though initially in the one-sided certification setting rather than full evaluation. The 2026 SBFC paper studies monotone Boolean function certification under correlated Bernoulli variables and shows that, for $x\in\{0,1\}^n$36-of-$x\in\{0,1\}^n$37, SBFE reduces essentially to two certification problems, one on active elements and one on inactive elements. It gives a polynomial-time non-adaptive $x\in\{0,1\}^n$38-approximation for arbitrary matroid certification under arbitrary distributions, a polynomial-time non-adaptive 4.642-approximation for both $x\in\{0,1\}^n$39-of-$x\in\{0,1\}^n$40 SBFC and SBFE under arbitrary distributions, a 2-approximation for non-adaptive $x\in\{0,1\}^n$41-of-$x\in\{0,1\}^n$42 SBFE under conditional negative association, and adaptive $x\in\{0,1\}^n$43 and $x\in\{0,1\}^n$44 approximations for graph and hypergraph probing models generated by latent independent variables (Ghuge et al., 3 Apr 2026). A plausible implication is that correlation does not eliminate approximation uniformly; rather, approximation becomes contingent on exploiting either function structure or distribution structure.

Finally, SBFE should not be conflated with exact Boolean-function representation and evaluation in the absence of uncertainty. Tarau’s “Pairing Functions, Boolean Evaluation and Binary Decision Diagrams in Prolog” is not about SBFE in the usual sense of adaptive test sequencing under uncertainty. Instead it studies natural-number encodings of truth tables and binary decision diagrams, proves that Boolean evaluation of the resulting BDD reconstructs exactly the same natural number as the structural inverse encoding, and derives ranking and unranking procedures for BDDs and reduced BDDs (0808.0555). This makes it adjacent foundational material for exact Boolean-function representation, compact symbolic evaluation, and benchmark generation, but not a contribution to the core stochastic problem of cost-sensitive adaptive variable querying.