Position-Threshold Rules: A Cross-Domain Schema

• Position-threshold rules are decision frameworks that trigger actions when an ordered measure exceeds a predefined threshold, with applications in stopping problems, classification, and percolation.
• They simplify complex decision tasks by decomposing problems into order comparisons and surplus evaluations, enabling precise analysis in areas such as prophet inequalities and online sample selection.
• These rules adapt to domain-specific orderings—temporal, spatial, or structural—and provide optimal or competitive performance in diverse settings including ergodic trading, voting systems, and graph recognition.

Position-threshold rules are families of decision rules in which an action is triggered by comparing an ordered coordinate to one or more thresholds. The coordinate can be a payoff value observed online, the number of previously selected samples, a scalar signal, the number of infected neighbors in a specified neighborhood, an alternative’s position in an ordered domain, a spatial displacement from a vortex center, or a vertex weight. The term is therefore not uniform across literatures: in some settings it denotes explicitly position-indexed thresholds, while in others it denotes threshold policies defined on an ordered state space or geometry (Zhang, 19 May 2026, Bach et al., 2010, Lederer, 5 Sep 2025, Duminil-Copin et al., 2023, Penn, 8 Apr 2025, Lovas et al., 2021, Ravanmehr et al., 2016).

1. Formal scope and recurring templates

Across the cited literatures, the common structural motif is an order plus a threshold comparison. What changes is the underlying order: temporal order in stopping problems, selection count in online sampling, signal order in classification, geometric order on $\mathbb{Z}^2$, the exogenous order of alternatives in voting, or latent one-dimensional weights in graph formation.

Domain	Ordered object	Threshold form
Prophet inequality	value level	stop at first $X_i > \tau$
Online sample selection	acceptance position $i$	accept iff sample lies in $S_i$
Endogenous classification	score $x$	$\delta(x)=\mathbf{1}_{x\ge\tau}$, $\mathbf{1}_{x\le\tau}$, or two-cut
Bootstrap percolation	neighborhood positions	infect iff at least $\theta$ sites in $x+\mathcal N$ are infected
Interval-domain voting	alternative $x_i$ in a total order	choose left-most $X_i > \tau$0 with $X_i > \tau$1
PT graphs	vertex weights $X_i > \tau$2	edge iff $X_i > \tau$3 and $X_i > \tau$4

The definitions are correspondingly heterogeneous. In the classical prophet inequality, the relevant note studies a single-threshold rule with a threshold $X_i > \tau$5 that is the same at every arrival index. In online sample selection, a threshold rule is a nested sequence $X_i > \tau$6, so the threshold depends on how many samples have already been accepted. In interval-domain voting, a position-threshold rule is parametrized by a weight vector $X_i > \tau$7 and a nonincreasing threshold vector $X_i > \tau$8, and it aggregates individual interval reports through a collective position function $X_i > \tau$9. In equilibrium classification, the optimal rules are single-threshold, negative-threshold, or two-cut rules on a scalar score under monotone likelihood ratio conditions. In bootstrap percolation, a threshold rule with neighborhood $i$0 and threshold $i$1 activates a site when at least $i$2 specified relative positions are already active (Zhang, 19 May 2026, Bach et al., 2010, Lederer, 5 Sep 2025, Penn, 8 Apr 2025, Penn et al., 11 Nov 2025, Duminil-Copin et al., 2023).

This suggests that “position-threshold rule” is best understood as a cross-domain schema rather than a single formal object. The schema is minimal—order, threshold, trigger—but the induced mathematics ranges from stopping-time analysis and ergodic control to metastability, social choice, and graph recognition.

2. Online stopping and selection

In the classical prophet inequality, one observes independent nonnegative random variables $i$3 online and compares the payoff of a nonanticipating stopping rule with the prophet benchmark $i$4, where $i$5. The note "Threshold Rules for the Classical Prophet Inequality" analyzes the rule “stop at the first $i$6” through the decomposition

$i$7

This threshold/surplus decomposition certifies several deterministic thresholds—median, half-mean, and balanced-surplus—and also a randomized threshold distributed as the maximum. In each case the guarantee is $i$8. The note is explicit that all rules it analyzes are position-independent: the same $i$9 is used at every index, and there is no sequence $S_i$0 in the paper (Zhang, 19 May 2026).

The online sample-selection literature uses the phrase differently. In "Threshold rules for online sample selection" a threshold rule is specified by a nested sequence of acceptance sets $S_i$1; when the process has already selected $S_i$2 samples, the next sample is accepted iff it lies in $S_i$3. This is an explicitly position-based rule because the threshold depends only on the selection count $S_i$4, not on calendar time or on previously observed values. The paper gives model-specific instantiations: for scalar quality on $S_i$5, $S_i$6 with $S_i$7; for a binary-tree coverage model, $S_i$8; and for skyline selection under a partial order, $S_i$9. The resulting tradeoff results are strong: for power-law distributions on $x$0, the rule $x$1 achieves an $x$2 approximation with constant $x$3 independent of the power-law parameter $x$4; in the binary-tree model, any threshold rule with $x$5 achieves $x$6 competitive ratio in expectation; and in skyline models, broad classes of such rules achieve $x$7 competitive ratio in expectation with respect to gap (Bach et al., 2010).

Taken together, these two online literatures separate two technically distinct notions. One is the value-threshold rule with a fixed $x$8; the other is the acceptance-position rule with an indexed family $x$9. Both are threshold-based, but only the latter is intrinsically position-dependent.

3. Equilibrium classification and endogenous behavior

In outcome-performative classification, the rule changes the behavior it attempts to classify. In "Optimal classification with endogenous behavior", an individual chooses $\delta(x)=\mathbf{1}_{x\ge\tau}$0, signals are drawn from $\delta(x)=\mathbf{1}_{x\ge\tau}$1, and the classifier commits to an integrable rule $\delta(x)=\mathbf{1}_{x\ge\tau}$2. The central incentive statistic is

$\delta(x)=\mathbf{1}_{x\ge\tau}$3

which determines equilibrium compliance through $\delta(x)=\mathbf{1}_{x\ge\tau}$4. Under continuity, full support, and strict MLRP of $\delta(x)=\mathbf{1}_{x\ge\tau}$5 relative to $\delta(x)=\mathbf{1}_{x\ge\tau}$6, the main theorem states that for any integrable classifier $\delta(x)=\mathbf{1}_{x\ge\tau}$7, there exists either a threshold rule $\delta(x)=\mathbf{1}_{x\ge\tau}$8 or a negative threshold rule $\delta(x)=\mathbf{1}_{x\ge\tau}$9 that is at least as accurate as $\mathbf{1}_{x\le\tau}$0. The paper also gives a numerical example in which the globally accuracy-maximizing classifier is negative-threshold: with $\mathbf{1}_{x\le\tau}$1, $\mathbf{1}_{x\le\tau}$2, and $\mathbf{1}_{x\le\tau}$3, the best positive threshold has accuracy about $\mathbf{1}_{x\le\tau}$4 with compliance about $\mathbf{1}_{x\le\tau}$5, whereas the best negative threshold has accuracy about $\mathbf{1}_{x\le\tau}$6 with compliance about $\mathbf{1}_{x\le\tau}$7 (Penn, 8 Apr 2025).

"Classification in Equilibrium: Structure of Optimal Decision Rules" extends this equilibrium perspective. The baseline Stackelberg model again uses a scalar signal and MLRP, but allows quota constraints and, in an extension, cheating behavior $\mathbf{1}_{x\le\tau}$8. In the quota-constrained baseline, the optimal classifier is a threshold or negative threshold rule. In the cheating model, when $\mathbf{1}_{x\le\tau}$9, $\theta$0, and $\theta$1 belong to an exponential family, the optimal classifier is either a threshold rule or a two-cut rule. The two-cut rules are of two types: an inner rule that rewards a middle band $\theta$2, and an outer rule that rewards the complement of a middle band. The examples are deliberately counterintuitive. One outer two-cut rule improves accuracy from $\theta$3 to $\theta$4 while reducing cheating from $\theta$5 to $\theta$6; one inner two-cut rule improves accuracy from $\theta$7 to $\theta$8 while maintaining compliance at $\theta$9 and inducing cheating of about $x+\mathcal N$0 (Penn et al., 11 Nov 2025).

A common misconception is that threshold optimality in classification necessarily means “reward higher scores.” These papers show that this is false once behavior is endogenous. Threshold simplicity survives, but monotonicity need not.

4. Spatial, lattice, and chemical threshold dynamics

In two-dimensional bootstrap percolation, a position-threshold rule is defined by a finite neighborhood $x+\mathcal N$1 and an integer threshold $x+\mathcal N$2: a site becomes infected when at least $x+\mathcal N$3 sites in $x+\mathcal N$4 are infected. For symmetric threshold rules $x+\mathcal N$5, the geometry of $x+\mathcal N$6 determines the difficulty exponent

$x+\mathcal N$7

The paper proves that every isotropic symmetric threshold rule with a two-dimensional convex symmetric neighborhood has a sharp metastability transition: $x+\mathcal N$8 Here the threshold is literally attached to specified lattice positions, and the resulting asymptotics are governed by a Wulff-type variational principle for $x+\mathcal N$9 (Duminil-Copin et al., 2023).

A closely related but physically distinct usage appears in the study of twisted light. For a trapped $x_i$0 ion interacting with a vortex beam, the transition amplitude has a Bessel factor $x_i$1, so at the vortex center $x_i$2 only channels satisfying $x_i$3 survive, and near the axis the amplitude scales as $x_i$4. The paper identifies a “prenumbra” crossover radius of about $x_i$5 for $x_i$6 and about $x_i$7 for $x_i$8, and reports agreement between experiment and theory at a level of better than $x_i$9. In this usage, the rule is a position-dependent angular-momentum selection rule with threshold-like radial suppression rather than a discrete decision policy (Afanasev et al., 2017).

Step chemical reaction networks provide a computational realization of threshold logic. In "Computing Threshold Circuits with Bimolecular Void Reactions in Step Chemical Reaction Networks", step CRNs with only $X_i > \tau$00 void rules simulate threshold formulas with $X_i > \tau$01 species, $X_i > \tau$02 steps, and $X_i > \tau$03 volume, and simulate threshold circuits with $X_i > \tau$04 species, $X_i > \tau$05 steps, and $X_i > \tau$06 volume. Under restricted gate-wise simulation, the paper proves a matching exponential lower bound on required volume for simulating threshold circuits, so the construction is optimal in that regime (Anderson et al., 2024).

5. Voting, network models, and trading strategies

In the interval domain of social choice, alternatives are linearly ordered as $X_i > \tau$07, and each voter reports an interval $X_i > \tau$08. A position-threshold rule is defined by an individual position function $X_i > \tau$09, a collective position function

$X_i > \tau$10

and a nonincreasing threshold vector $X_i > \tau$11. The chosen alternative is

$X_i > \tau$12

that is, the left-most alternative whose collective position exceeds its threshold. The main characterization states that a voting rule on the interval domain is robust, anonymous, unanimous, reinforcing, and right-biased continuous iff it is a position-threshold rule. The same paper singles out the endpoint-median rule—corresponding to $X_i > \tau$13 and $X_i > \tau$14—as the unique position-threshold rule satisfying strong unanimity and the majority criterion (Lederer, 5 Sep 2025).

Paired Threshold graphs replace sequential or social positions by latent vertex weights. A graph is PT if there exist thresholds $X_i > \tau$15 and weights $X_i > \tau$16 such that

$X_i > \tau$17

This combines a sum-threshold with a difference-threshold. Connected PT graphs are characterized either as unit interval graphs or as graphs admitting a distance decomposition $X_i > \tau$18 in which $X_i > \tau$19 has a total vicinal preorder and each $X_i > \tau$20 for $X_i > \tau$21 is a clique with a total preorder $X_i > \tau$22. The paper also gives a polynomial-time recognition algorithm with worst-case $X_i > \tau$23 complexity (Ravanmehr et al., 2016).

In ergodic trading with threshold strategies, the thresholds are two-sided and govern portfolio position changes. The fluctuation process is $X_i > \tau$24, and the trading rule uses $X_i > \tau$25 with alternating crossing times

$X_i > \tau$26

This generates repeated buy-low/sell-high cycles. Under a minorization condition, zero mean, and one-sided boundedness of Markovian increments, the associated cycle chain $X_i > \tau$27 is uniformly ergodic, its law converges in total variation at a geometric rate, and bounded functionals such as per-trade utilities satisfy a strong law of large numbers. Here “position-threshold” refers to thresholds on portfolio position changes induced by the fluctuation process rather than to score cutoffs or choice rules (Lovas et al., 2021).

6. Cross-domain themes and recurrent misunderstandings

A first recurrent misunderstanding is terminological. Position-threshold rules are not uniformly “position-dependent.” The prophet-inequality note is explicit that its analysis concerns a single, shared threshold $X_i > \tau$28 and that no position-dependent sequence $X_i > \tau$29 appears. By contrast, online sample selection and interval-domain voting make position indexing intrinsic: the threshold depends on the number selected so far or on the alternative’s ordinal location (Zhang, 19 May 2026, Bach et al., 2010, Lederer, 5 Sep 2025).

A second misunderstanding is that threshold rules are necessarily monotone in the direction of better evidence. Outcome-performative classification disproves this sharply: a negative threshold rule can maximize accuracy, and more general equilibrium models admit inner and outer two-cut rules. The threshold family remains low-dimensional and interpretable, but “high signal $X_i > \tau$30 favorable decision” is not preserved once the rule reshapes behavior (Penn, 8 Apr 2025, Penn et al., 11 Nov 2025).

A third theme is that threshold rules often owe their tractability to decomposability. In prophet inequalities, expected payoff splits into a threshold part and a surplus part. In bootstrap percolation, the metastable scale is encoded by directional difficulties and a droplet-growth variational problem. In PT graphs, paired threshold inequalities induce a layered decomposition. In step CRNs, gate-wise threshold logic becomes analyzable because the chemistry respects a depth decomposition and dedicated gate species (Zhang, 19 May 2026, Duminil-Copin et al., 2023, Ravanmehr et al., 2016, Anderson et al., 2024).

A final commonality is that threshold simplicity does not imply weak expressivity. Single-threshold rules attain the sharp $X_i > \tau$31 constant in the classical prophet inequality; position-indexed threshold rules achieve constant-factor or $X_i > \tau$32 tradeoffs in online selection; convex symmetric position-threshold rules in bootstrap percolation exhibit sharp metastability; interval-domain position-threshold rules admit a full axiomatization; and deletion-only step CRNs still simulate threshold formulas and circuits, albeit with exponential volume for general circuits under gate-wise simulation (Zhang, 19 May 2026, Bach et al., 2010, Duminil-Copin et al., 2023, Lederer, 5 Sep 2025, Anderson et al., 2024).

In that sense, position-threshold rules form a broad methodological class: they compress complex decision problems into order comparisons against one or more thresholds, while retaining enough structure to support exact characterization, asymptotic analysis, or optimal design.