Capped-Interval Heuristic Overview

• Capped-interval heuristic is a structural pattern that restricts the search space via a predefined cap, converting global problems into bounded local decision rules.
• It finds applications in methods such as simplex projection, robust optimization with LP duality, temporal indexing, and iterative solvers, each adapting the cap to enforce feasibility.
• These heuristics leverage mathematical properties like monotonicity, KKT conditions, and interval geometry to improve computational efficiency and decision-making.

“Capped-interval heuristic” (Editor's term) designates a family of methods in which a cap or interval restriction is imposed first, and the resulting admissible set is then searched, projected onto, sampled from, or certified by a secondary rule. In the cited literature, the expression is not used uniformly: some works study literal capped intervals or box constraints, some cap the effect of interval uncertainty on regret, and some use interval geometry to classify nodes or candidate updates. This suggests a unifying interpretation in which a capped-interval heuristic is not a single algorithmic primitive but a recurrent structural pattern for enforcing feasibility, limiting uncertainty, or restricting search while retaining a tractable optimization or decision step (Wang et al., 2015, Assunção et al., 2016, Wang et al., 22 Jun 2026, Zhang et al., 2022).

1. Taxonomy and common structure

Across the sources, the same schematic move recurs: define a bounded interval-like region, identify admissible states or candidates within it, and then optimize only over that restricted region or certify whole regions at once. In some papers the cap is literal, as in $0 \le x_i \le 1$ or a payoff cap $\epsilon$; in others it is operational, as when LP dual information “caps” regret, or node rectangles in corner space cap the possible coordinates of intervals. A plausible abstraction is that capped-interval heuristics convert a global problem into a bounded local decision rule whose correctness or usefulness derives from monotonicity, KKT structure, LP duality, or interval geometry (Wang et al., 2015, Assunção et al., 2016, Wang et al., 22 Jun 2026, Zhang et al., 2022, Bacci et al., 2017, Kyprianou et al., 2012, Li et al., 2016).

Context	Capped object	Operational rule
Capped simplex projection	$0 \le x_i \le 1$, $\sum_i x_i = s$	sort, search $(a,b)$, compute $\gamma$
Interval min–max regret	interval costs $[l_i,u_i]$	replace exact inner optimum by LP-dual lower bound
Disk interval indexes	node rectangles in corner space	classify nodes as must, may, or prune
Nonlinear Kaczmarz	capped index sets $\mathcal U_k,\mathcal I_k$	sample or block-update only inside the capped set
BPPC on interval graphs	bin capacity $B$	interval coloring, then TAIL-EXCHANGE and INSERTION
Capped optimal stopping	payoff cap $\epsilon$	stop when excursion exceeds $\epsilon$0
Interval separation	movement threshold $\epsilon$1	compare $\epsilon$2 with the cap

The significance of this taxonomy is methodological rather than terminological. The papers do not define a single named framework, but they repeatedly exploit the same idea: once the feasible or relevant region has been “capped,” one can either derive an exact finite procedure, as in capped simplex projection, or obtain a compact heuristic surrogate, as in LP-based min–max regret and greedy capped Kaczmarz.

2. Projection-based feasibility restoration

The most literal capped-interval construction in the sources is Euclidean projection onto the capped simplex, defined by

$\epsilon$3

The feasible set is the intersection of the unit cube $\epsilon$4 and the hyperplane $\epsilon$5, and the problem is a strictly convex quadratic program, hence the projection is unique. The paper derives an exact algorithm from the KKT conditions. After sorting $\epsilon$6 in ascending order, the solution has three regimes,

$\epsilon$7

with some coordinates fixed at $\epsilon$8, some fixed at $\epsilon$9, and the remainder strictly interior. For the interior block, $0 \le x_i \le 1$0, where

$0 \le x_i \le 1$1

and $0 \le x_i \le 1$2 is validated by KKT-derived inequalities on $0 \le x_i \le 1$3, $0 \le x_i \le 1$4, $0 \le x_i \le 1$5, and $0 \le x_i \le 1$6. The implementation sorts once, precomputes partial sums $0 \le x_i \le 1$7, scans candidate $0 \le x_i \le 1$8, and reconstructs the projected vector in sorted order before permuting back. The complexity is $0 \le x_i \le 1$9 overall, with sorting $\sum_i x_i = s$0, prefix sums $\sum_i x_i = s$1, and the nested search $\sum_i x_i = s$2 (Wang et al., 2015).

The same paper explicitly interprets this projection as a projected-gradient or proximal step. Given an unconstrained update $\sum_i x_i = s$3, the next iterate is

$\sum_i x_i = s$4

In that sense, the capped-interval heuristic is exact rather than approximate: feasibility restoration is performed by a one-shot Euclidean projector onto box constraints plus a global sum constraint. The paper also notes that $\sum_i x_i = s$5 generalizes to $\sum_i x_i = s$6 for any $\sum_i x_i = s$7 by scaling, so the same mechanism covers non-unit caps.

3. Interval uncertainty and regret capping

A different use of the motif appears in interval data min–max regret combinatorial optimization. The deterministic problem is

$\sum_i x_i = s$8

while the robust version allows interval costs $\sum_i x_i = s$9. For a feasible $(a,b)$0, the maximum regret occurs in the solution-induced scenario $(a,b)$1, where $(a,b)$2 if $(a,b)$3 and $(a,b)$4 otherwise. The robust objective can therefore be written as

$(a,b)$5

For interval robust-hard problems, the inner optimization is already NP-hard, and the standard exact formulation requires an exponential number of cuts, each separated by solving the classical NP-hard problem. The LP-based heuristic replaces the exact inner minimum by the optimal value $(a,b)$6 of an LP relaxation and then embeds the dual constraints directly: $(a,b)$7 subject to $(a,b)$8, $(a,b)$9, $\gamma$0, and $\gamma$1. Because $\gamma$2, this produces an upper bound on the true regret. The paper explicitly states that the heuristic “caps (bounds) the effect of interval uncertainty on regret” by substituting an LP lower bound for the exact inner optimum; if the classical problem has a totally unimodular constraint matrix and integer right-hand side, the bound is tight and the heuristic solution is optimal (Assunção et al., 2016).

The framework is instantiated for the Restricted Robust Shortest Path and Robust Set Covering problems. In both cases, the workflow is: build a strong LP relaxation of the classical problem, derive its dual, construct a robust heuristic MILP with binary solution variables and continuous dual variables, solve that MILP once, and then compute the exact robustness cost of the final solution by solving one classical NP-hard instance. Computationally, this avoids the cut-generation regime of LB-Benders while improving markedly over the AMU 2-approximation. For R-RSP, the reported average gaps are approximately $\gamma$3 versus $\gamma$4 in one benchmark and approximately $\gamma$5 versus $\gamma$6 in another; on hard instances, the heuristic sometimes gives better upper bounds than LB-Benders within a $\gamma$7s limit. For RSC, the heuristic always has lower average gaps than AMU and can even beat LB-Benders on the most challenging KZ instances when the exact method times out or stalls.

4. Geometric caps in temporal interval indexing

In temporal databases, the cap is geometric rather than algebraic. Intervals $\gamma$8 are mapped to corner-space coordinates such as $\gamma$9, with $[l_i,u_i]$0, or $[l_i,u_i]$1, with $[l_i,u_i]$2. Under the Increasing Ending Time assumption, the paper develops two disk-based indexes: CEB, organized by center and endpoint, and TIDE, organized by duration and endpoint. Both use a two-layer architecture in which a top B$[l_i,u_i]$3-tree partitions the $[l_i,u_i]$4- or $[l_i,u_i]$5-axis and each leaf points to a bottom B$[l_i,u_i]$6-tree on $[l_i,u_i]$7. Because insertions arrive in non-decreasing order of end time, both top and bottom trees are append-only. The central query-processing optimization is the distinction between nodes that “must contain results” and nodes that “may contain results.” Each data node has a capped region given by its key ranges, for example

$[l_i,u_i]$8

and range queries become regions in corner space bounded by a horizontal line and a slash or diagonal line. If the node rectangle lies entirely inside the query region, all records are reported directly; if it lies entirely outside, the node is pruned; otherwise its records are inspected individually. The paper explicitly identifies this must/may distinction as the essence of the capped-interval heuristic in this setting (Wang et al., 22 Jun 2026).

The IET assumption is structural. It makes bottom trees append-only, stabilizes $[l_i,u_i]$9, and creates immutable nodes whose caps no longer change. That stability is what permits safe must-report and prune decisions. In CEB, the query region for $\mathcal U_k,\mathcal I_k$0 is

$\mathcal U_k,\mathcal I_k$1

while in TIDE the overlap condition becomes $\mathcal U_k,\mathcal I_k$2 inside a bottom tree. The paper gives concrete node tests for both structures, including break conditions and must-report conditions based on $\mathcal U_k,\mathcal I_k$3, $\mathcal U_k,\mathcal I_k$4, $\mathcal U_k,\mathcal I_k$5, and $\mathcal U_k,\mathcal I_k$6. Empirically, CEB and TIDE outperform state-of-the-art competitors in index size and insertion speed, and TIDE is always faster in query processing, sometimes by orders of magnitude. The reported speedups are attributed largely to the ability to certify whole nodes geometrically rather than reading them from disk.

5. Adaptive capped candidate sets in iterative solvers

In greedy capped nonlinear Kaczmarz methods, the cap is imposed on the row-selection mechanism for solving $\mathcal U_k,\mathcal I_k$7, where $\mathcal U_k,\mathcal I_k$8 is continuously differentiable. The base update is the nonlinear Kaczmarz projection

$\mathcal U_k,\mathcal I_k$9

The capped methods alter only the choice of $B$0. In DR-CNK, the threshold is

$B$1

and the eligible set is

$B$2

Sampling is then residual-weighted inside $B$3. In RD-CNK, the threshold is

$B$4

with capped set

$B$5

and probabilities proportional to $B$6 inside $B$7. The paper presents these as “two-level greedy” mechanisms: one rule defines the cap, and a second rule defines the probability distribution within the cap. Block variants DB-CNK and RB-CNK simply use the entire capped set as a block and update with a pseudoinverse step (Zhang et al., 2022).

Under the local tangential cone condition and existence of a solution $B$8 with $B$9, the paper proves expected linear convergence for DR-CNK and RD-CNK, and shows that their convergence factors satisfy

$\epsilon$0

Thus the cap is not merely a filtering device; it improves the expected contraction factor by excluding low-impact rows. The numerical section reports large iteration reductions on the Brown almost linear function, including approximately $\epsilon$1 for a $\epsilon$2 problem and approximately $\epsilon$3 for a $\epsilon$4 problem, while DB-CNK and RB-CNK often converge in one block step on those tests. On regularized GLM instances, both single-sample and multi-sample capped methods outperform Kaczmarz-TCS and Block TCS in iteration counts and CPU time.

6. Interval-structured local search and motion capping

For the Bin Packing Problem with Conflicts on interval graphs, the cap is the bin capacity $\epsilon$5, while conflicts are induced by interval overlap. The heuristic BN exploits interval-graph structure via a leftmost maximum clique $\epsilon$6 and its associated coordinate $\epsilon$7. Phase I builds a coloring with $\epsilon$8 subsets, assigning each interval of $\epsilon$9 to a different subset, then processing intervals left of $\epsilon$00 by decreasing right endpoint and intervals right of $\epsilon$01 by increasing left endpoint. On the left side, the chosen subset minimizes

$\epsilon$02

where $\epsilon$03 estimates the weight of empty space to the right of color $\epsilon$04. Phase II restores capacity feasibility by TAIL-EXCHANGE, which swaps interval tails at a coordinate $\epsilon$05 between a heavy and a light subset, and INSERTION, which moves an interval $\epsilon$06 from a heavy subset into a light subset, a heavy subset, or a new subset, with the explicit goal of making the recipient as full as possible without exceeding $\epsilon$07. The total complexity is $\epsilon$08. Over $\epsilon$09 instances, BN typically outperforms the adapted M heuristics for large bins, while M shows better or similar performance for small bins (Bacci et al., 2017).

A movement-capped variant arises in interval separation on a line. The exact problem is to move intervals so that no two overlap while minimizing the maximum moving distance. The paper reduces the two-direction problem to a one-direction problem: if $\epsilon$10 is the optimal rightward max-displacement, then an optimal two-direction solution is obtained by shifting every interval in the one-direction optimum left by $\epsilon$11. The key combinatorial object is an optimal list $\epsilon$12, and once such an order is known, the left-possible placement strategy yields the optimal configuration in $\epsilon$13 time. The hard part is order selection. The paper shows that it is sufficient to maintain at most $\epsilon$14 candidate lists, prune dominated lists using exchange arguments and dominance lemmas, and compute an optimal solution in $\epsilon$15 time and $\epsilon$16 space, which matches an $\epsilon$17 lower bound. This suggests a capped-interval interpretation in which a displacement threshold $\epsilon$18 can be tested by computing $\epsilon$19 and checking whether $\epsilon$20, while the candidate-list logic also provides a template for simpler capped movement heuristics based on local order updates and pruning (Li et al., 2016).

7. Capped stopping regions in stochastic control

In capped optimal stopping for the running maximum of a spectrally negative Lévy process $\epsilon$21, the cap acts directly on the payoff. With running maximum

$\epsilon$22

the value function is

$\epsilon$23

The cap $\epsilon$24 has two structural consequences stated explicitly in the paper: once the running maximum hits $\epsilon$25, the spatial component of the payoff is frozen at $\epsilon$26, and because of discounting there is then no upside from waiting. The optimal rule is excursion-based: stop when the excursion below the current maximum is too deep,

$\epsilon$27

so the continuation interval at level $\epsilon$28 is $\epsilon$29. The boundary $\epsilon$30 is characterized by the ODE

$\epsilon$31

with $\epsilon$32. The paper further shows that the slope near the cap depends on whether $\epsilon$33 has bounded or unbounded variation, and reformulates the solution through a version of Peskir’s maximality principle: among solutions of the ODE, the relevant boundary is the minimal one in the excursion-depth parameterization (Kyprianou et al., 2012).

As an overview, this is the most literal state-space version of the capped-interval heuristic. For each running-maximum level $\epsilon$34, the method defines a continuation interval in the state variable $\epsilon$35; the interval shrinks to zero at the cap; and stopping occurs when the process exits that interval from below. The same high-level logic reappears, with different technical machinery, in the other domains surveyed here: KKT projection onto capped boxes, LP-dual capping of interval regret, node-rectangle certification in interval indexes, capped candidate sets in nonlinear Kaczmarz, and interval-structured local search under a bin-cap or movement cap. The common theme is that a cap converts an unconstrained or globally coupled problem into a bounded decision geometry, and the heuristic or exact method derives its power from exploiting the resulting structure rather than from generic optimization alone.