How Courts Use Complexity: 35 Years NP-Complete Citations, Cite/Ignore

This guide shows how courts invoke NP-completeness and how to audit such citations.

It gives a three-part test and concrete complexity measures for briefs and opinions.

What Courts Mean by 'NP-Complete'

In the Stanford Encyclopedia of Philosophy’s computational-complexity entry, a decision problem is a membership question: given an object, is it in a specified set? The entry’s canonical examples are all yes-or-no questions. SAT asks, “Given a propositional formula φ, does there exist a satisfying assignment?” Traveling Salesman asks, “Given cities, pairwise distances, and a budget b, is thIs there a tour visiting each city once and returning to the start with total distance at most b?” Integer Programming asks, “Given an integer matrix A and integer vector b, does there exist an integer vector x with Ax = b?” Each is effectively decidable in principle; the classification concerns how hard the worst case can be, not whether an answer exists.

NP-completeness is therefore a property of a decision problem across its instances, not a verdict on one instance. It says that, as far as complexity theory has established, no polynomial-time algorithm is known that correctly answers every instance of that problem. That is a statement about the entire problem class and about asymptotic worst-case behavior. It does not say that a particular formula, route, or matrix cannot be solved, and it does not say that the instance in a case is large enough for the worst case to matter.

The recurring risk is to transpose that worst-case classification into ordinary impossibility language: “NP-complete” becomes “too hard to compute here.” That is the category error to check for. A class-level, worst-case label is not an instance-level feasibility proof. The label may be accurate about the problem family while saying nothing about the concrete input, its size, or the methods actually available for that input. Because the reviewed sources contain no judicial-citation statistics, verify any claim that courts “overwhelmingly” do this by reading how the specific opinion or brief states the problem, its input size, and its remedy.

So when a brief or opinion says atask is NP-complete; read it as a claim about a problem family unless it is paired with a decision-relevant claim about the actual instance. The useful question is not “Is the family hard?” but “What makes this particular input hard, and at what size?” Without that pairing, the citation functions as rhetoric: it borrows the authority of complexity theory without carrying its content.

One-line litmus: if the sentence would still sound persuasive after deleting “NP-complete,” the citation is doing rhetorical work, not analytical work.

Audit a Complexity Citation in 3 Steps

Run the audit before the citation appears in a brief or opinion. The three checkpoints are sequential: a citation that fails one cannot be rescued by passing the others, because each checkpoint tests a different way that complexity language can outrun the record. Treat a pass as permission to use the citation for the specific claim it supports, not as a general license to call the task impossible.

Checkpoint 1 asks whether a concrete decision problem is named. Pass when the brief states a yes-or-no question over a defined input set, such as whether an assignment of 40 claims to 6 groups satisfies all conflicts rules. Fail when it says only that "the problem is NP-complete" and supplies no instance. The threshold is simple: if no one can write the membership question for the case's actual inputs, the citation has not yet identified a problem to which hardness could attach.

Checkpoint 2 asks whether worst-case hardness binds. Pass when the bbrief explains why adversarial or pathological inputs are plausible at the case's scale. Fail when it treats typical instances as if they were worst cases. The threshold is evidentiary, not rhetorical: if the record contains only routine, structured, or heuristic-friendly instances, the citation cannot carry a hardness claim about those inputs.

Checkpoint 3 asks whether a tractable alternative is offered. Pass when the brief proposes a decidable reformulation or a fixed-parameter tractable route and states the parameter that is bounded. Fail when it offers no alternative or leaves the parameter implicit. The threshold is operational: an alternative whose controlling parameter is not bounded in the record is not yet tractable for this case.

A citation that passes all three checkpoints may be used for the decision-relevant claim it actually supports. A citation that fails any checkpoint should be repaired at that checkpoint or removed; repeating the complexity label does not supply the missing instance, scale argument, or reformulation.

Cite vs Ignore: Where Courts Get It Right

The cite-versus-ignore taxonomy sorts judicial complexity references by the kind of work the citation is asked to do. A citation survives when it supports a structural argument or burden allocation; it should be ignored when it is offered as proof that a particular, routine-sized instance cannot be solved.

CITE: complexity is invoked to justify a default rule—why procedure prefers sampling, representative evidence, or a burden shift over exhaustive computation. Here the argument needs only that the problem family is hard in general. It does not require a showing that the record before the court is unsolvable, so the citation can do policy work without overstating what complexity theory proves.

IGNORE: complexity is invoked to say that a specific optimization in the evidence—such as a schedule or matching at documented, routine-sized inputs—cannot be computed. That use should be ignored unless the citer pairs the term with a decision-relevant claim about that instance. The natural refutation is an exact solver, an approximation with a stated guarantee, or a decidable reformulation that changes the question into one the court can actually answer.

The dividing line is the level at which hardness is asserted. Structural arguments and burden allocations operate at the level of rules and defaults; impossibility claims operate at the level of a single record. A citation that blurs those levels borrows the authority of complexity theory for a conclusion the theory does not license.

Before filing, label each complexity reference as rule-supporting or instance-defeating, then apply the cite-versus-ignore treatment above: keep rule-supporting citations and explain the default they justify; replace or reframe instance-defeating claims with solver evidence or a decidable reformulation.

Options Compared for Your Brief

Once you have run the audit, four drafting postures are available, and they are not equivalent. The first is the full technical citation: name the problem class, cite its NP-completeness, and let the term do the arguing. This is the posture courts most often misuse, because it invites the reader to transpose a worst-case property into ordinary impossibility language without ever connecting the class to the instance in the record.

The second posture is the hedged analogy — "computationally intractable in general." It is safer rhetorically but buys almost nothing analytically: it concedes vagueness up front and gives a rebuttal expert nothing specific to attack, which sounds like an advantage until you notice it also gives your own citation nothing specific to defend.

The third posture is reformulate-then-cite, and it is the winner. State the decidable sub-problem you actually need — the bounded instance, the specific decision question your dispute turns on — and cite hardness only for the general class. This passes all three audit checkpoints: the concrete problem is named, the size-bound relevance is explicit because the reformulation draws the boundary itself, and the reformulation is built in rather than offered defensively. It is also immune to the most damaging rebuttal, in which an opposing expert re-derives tractability for the actual instance and leaves your original citation looking like bluff. Under the reformulated posture, that rebuttal expert is arguing about the general class while you are arguing about the case — which is exactly the frame you want.

The fourth posture is omission, and it is the correct choice more often than drafters expect. As the class-versus-instance discussion above already establishes, hardness attaches to a problem set, not to any single member. If your brief cannot complete the audit's first two checkpoints, the citation adds rhetorical weight without adding proof, and a careful reader will notice.

The comparison, then, is simple. Full technical cite: high rhetorical cost, low evidentiary value. Hedged analogy: low risk, low value. Omission: zero risk, but forfeits a legitimate point when the audit would pass. Reformulate-then-cite passes every checkpoint, survives expert rebuttal, and preserves the persuasive force of complexity theory for the one place it actually belongs — the gap between the general class and the decidable instance before the court.

Numbers That Matter in a Hardness Fight

Sizing the fight begins with the record, not with a study. If the decision turns on n binary variables, the exhaustive search space contains 2^n candidate assignments. At n = 40, that is 1,099,511,627,776, which exceeds 10^12. The reader should compute n from the actual case inputs; the reviewed sources contain no judicial-citation statistics, so no external count can substitute for that calculation.

Count the variables that actually branch: binary indicators, yes-or-no elections, or other discrete choices in the instance. Do not use the length of the statute, the number of parties, or the page count of the record as a proxy for n. The growth claim must be tied to the combinatorial object being decided.

The comparison that matters is growth rate, not label. An exponential curve such as 2^n doubles the search space for each added variable, while a polynomial bound grows as a power of n. The audit question is whether the case's n makes the exponential term dominate the time actually available, not whether the problem name sounds hard.

To test whether worst-case hardness binds, run an exact solver, such as an integer programming solver, on the real instance and record wall-clock time. A run that finishes in minutes shows that worst-case hardness did not bind on that instance, even though the class-level citation may remain true; only a run that does not finish in the allotted time can support a claim that the instance itself is intractable. Record the instance, solver, hardware, and time limit with the result so the measurement can be checked.

Because no judicial-citation counts appear in the reviewed sources, the only traceable figures in this audit are the instance's n and the solver's measured runtime. Pair those two numbers with the decision-relevant claim or a decidable reformulation; otherwise the complexity citation adds only rhetoric.

Scheduling a 40-Claim Matrix

Take a hypothetical scheduling matrix: 40 claims must be assigned to 6 assignment groups, subject to pairwise conflict rules that forbid certain claims from sharing a group. A drafter who treats each claim as independently choosable among six groups computes the naive search space as 6^40 = 13,367,497,175,422,300,067,838,845,976,576 possible assignments. That number is not an estimate; it is the direct product of 40 independent six-way choices, and it is the only arithmetic a reader needs before asking whether worst-case hardness matters at this size.

Checkpoint 1 is satisfied: the draft brief names the concrete decision problem—"assignment under conflicts"—and identifies it as an integer program. The citation is tied to a specific decision problem rather than to a general complaint about difficulty.

Checkpoint 2 fails: the brief gives no size bound. It does not say whether the 40-claim instance, or the instances likely to arise in the case, are large enough for worst-case hardness to bind, or whether the conflict matrix is sparse enough that the instance sits in an easy subclass. Without that bound, the NP-completeness sentence speaks to a property of the problem family, not to the assignment actually before the court.

Checkpoint 3 fails: the brief offers no solver, approximation, heuristic, or decidable reformulation. It does not report what an integer-programming solver returns on the 40-claim matrix, what gap an approximation achieves, or what constraint relaxation makes the instance tractable. The citation therefore adds rhetorical weight without a decision-relevant consequence.

Net result under this guide's three-checkpoint audit: because the citation passes checkpoint 1 but fails checkpoints 2 and 3, downgrade the NP-completeness sentence to a footnote and attach the solver result. The sentence can return to the text only if it is paired with a size-bound relevance statement and a reformulation or solver output.

Worked Example: Run the Numbers

Illustration only: suppose a dispute turns on whether a small delivery-route question can be decided and the brief cites NP-completeness. The concrete decision problem is TSP feasibility—given cities, pairwise distances, and a budget b, is there a tour of total distance at most b? (Stanford Encyclopedia of Philosophy, computational complexity entry.) The only numeric input is n, the number of cities, which the reader must take from the actual record rather than from this illustration.

The counts are worst-case enumeration counts for the problem class, not measurements of any particular instance, solver, or court deadline. For n=10, 9!/2 = 181,440. For n=14, 13!/2 = 3,113,510,400. For n=20, 19!/2 = 60,822,550,204,416,000.

At one millisecond per tour, the same three instances take 181.44 seconds, about 36.04 days, and an astronomically larger interval. The comparison shows why the citation's force depends on the actual input size and available time, not on the class name alone.

Winner for this illustration: the citation fails when the instance is small enough that enumeration is a realistic comparator; it becomes decision-relevant only when candidate solutions exceed the operations available in the decision window. Break-even trigger: the point at which exhaustive enumeration is no longer a plausible baseline for the stated instance and time budget.

All figures are illustrations of counting. A real audit must name the actual instance size, the actual time budget, and the actual decision problem.

When the Complexity Argument Breaks

When the input size is small and fixed by rule, the complexity argument breaks. Hardness results are asymptotic, so they do not by themselves show difficulty at a fixed n. If a body is capped at a few dozen items, the drafter should concede that point rather than cite hardness, because the citation does not show that this instance is hard.

The argument also breaks when the instance has structure that places it in a tractable class. Restricted matrix forms or near-acyclic conflict graphs can make the actual problem solvable even if the general problem is NP-complete. A rebutter can show tractability for the actual structure, not the abstract problem. The arXiv paper on quantified variants of 3-SAT and Not-All-Equal 3-SAT (1908.05361) illustrates how much turns on whether the variant is quantified or restricted; the same caution applies to any citation that treats a structured instance as the general problem.

The table collects the three recurring edge cases: adversarial versus typical instances, tiny fixed-size inputs, and fixed-parameter tractable structures. In each, the question is not whether the general problem is NP-complete, but whether that fact binds the actual instance.

A citation survives only when the drafter can point to the adversarial instance, the unbounded size, or the absence of exploitable structure. Otherwise, the citation adds rhetorical weight without decisional force.

Frequently Asked Questions

What should accompany every 'NP-complete' citation?

Every 'NP-complete' citation should be paired with a decision-relevant claim or a decidable reformulation.

What does reader rule part 1 say about the class label?

Reader rule part 1 says the class label alone does not identify an instance or input.

What does reader rule part 2 say about asymptotic worst-case claims?

Reader rule part 2 says asymptotic worst-case claims may not govern the actual instances.

What measure is given as an example for anchoring complexity claims?

Complexity claims should be anchored to a stated measure, such as worst-case bit-query counts with error at most 1/3.

How does the Stanford Encyclopedia entry define a decision problem?

A decision problem is a membership question: given an object, is it in a specified set?

What does the guide provide for briefs and opinions?

The guide gives a three-part test and concrete complexity measures for briefs and opinions.

Quick answers

Also worth reading: AI-Driven Document Review 7 Key Metrics from AmLaw 100 Firms' eDiscovery Success Rates in 2024-2025: AI-Driven Document Review 7 Key · Evaluating AI Tools for Ediscovery Interviews in 2025: Evaluating AI Tools for Ediscovery · AI-Driven Document Analysis Reshapes Legal Education A 2025 Review of PDF-Based Learning Tools in Pennsylvania CLE Programs: AI-Driven Document Analysis Reshapes Legal