// deduxit © Fabian Lukas Grubmüller 2026
// Licensed under the EUPL
//! Problem families with known verdicts, generated at any size: the hard
//! families of the literature (3-Partition as a Horn program and as
//! Lincoln and Winkler's two-literal sequent, Matsuoka's Partition, random
//! QBF under a lock-and-key encoding into MALL, Petri-net counters), the
//! families where one engine is known to be slow (wide contexts under
//! Mix, wide sequents of distinct or shared atoms), and the slow cases of
//! the exponentials, and first-order families: the Lambek calculus in
//! first-order MILL, a chain of instances of a quantified clause under
//! `!`, and a witness chosen among many. Some families have a growing set
//! ([`Family::growth`]) and a claim ([`Family::claim`]): what the growth of
//! the time to decide them shows about one engine or another prover,
//! among them Franzén's contraction formulas and Cook's pigeonhole
//! principle from Dyckhoff's benchmarks, Euler tours from Chaudhuri's
//! benchmarks of the inverse method, a clause used many times on one
//! branch, and reachability through a cycle at first order. Each instance says whether it is provable, by
//! construction or by solving the combinatorial problem it encodes, never
//! by running an engine. Random families are seeded by their size and
//! index, so an instance is the same on every run.
//!
//! # Examples
//!
//! ```
//! use deduxit::linear::families;
//! use deduxit::linear::search::{Options, Verdict, prove};
//!
//! let family = families::find("partition-no").unwrap();
//! let instance = family.instance(3, 0)?;
//! assert!(!instance.provable);
//! let outcome = prove(&instance.sequent, instance.mode, &Options::default())?;
//! assert!(matches!(outcome.verdict, Verdict::Unprovable(_)));
//! # Ok::<(), deduxit::Error>(())
//! ```
//!
//! Needs the cargo feature `parse` (on by default).
use crate::{
Error,
linear::{Mode, Sequent},
};
/// One problem of a family.
#[derive(Clone, Debug)]
#[non_exhaustive]
pub struct Instance {
/// The family's name, the size and, for random families, the index:
/// `qbf/8#2`.
pub name: String,
/// The sequent.
pub sequent: Sequent,
/// The mode the sequent is meant for.
pub mode: Mode,
/// Whether the sequent is provable in its mode.
pub provable: bool,
/// The per-branch copy bound under which a proof exists, for the
/// families with exponentials; `None` for the others.
pub copies: Option<u32>,
/// The witness depth under which a proof exists, for the families
/// whose witnesses deepen with the size; `None` for the others, which
/// the default depth serves.
pub witness_depth: Option<u32>,
}
/// A family of problems, indexed by a size.
#[derive(Debug)]
#[non_exhaustive]
pub struct Family {
/// The name, as [`find`] takes it.
pub name: &'static str,
/// What the family is and what its size counts.
pub summary: &'static str,
/// The sizes a benchmark runs by default: from easy to past what the
/// engines decide in minutes.
pub sizes: &'static [u32],
/// How many instances each size has: one for the constructed families,
/// several for the random ones and for the constructed families of
/// several shapes, whose summary says what each index is.
pub instances: u32,
/// The least size the family has an instance of.
pub least: u32,
/// Whether its sizes are the powers of two.
pub powers_of_two: bool,
/// What the family shows: the strength or weakness of an engine or
/// another prover it is chosen for, and what a run is expected to
/// show; empty for a family that has no growing set.
pub claim: &'static str,
/// The sizes of its growing set, from easy to the ceiling where every
/// engine that admits it has met the time limit of a minute or a
/// bound; geometric where the time grows polynomially, by one where it
/// grows exponentially. Empty for a family that has none.
pub growth: &'static [u32],
/// Builds the instance of a size and an index.
generate: fn(u32, u32) -> Instance,
}
impl Family {
/// Returns the instance of the given size and index (below
/// [`instances`](Self::instances)), the same on every call.
///
/// # Errors
///
/// [`Error::FamilySize`] for a size the family has no instance of:
/// below [`least`](Self::least) (a 3-Partition bin smaller than four,
/// a QBF of fewer than three variables, a solvable Partition of fewer
/// than two items or an unsolvable one of fewer than three), or no
/// power of two where [`powers_of_two`](Self::powers_of_two) asks one
/// (a counter's tokens).
pub fn instance(&self, size: u32, index: u32) -> Result<Instance, Error> {
if size < self.least || (self.powers_of_two && !size.is_power_of_two()) {
return Err(Error::FamilySize {
family: self.name,
size,
least: self.least,
powers_of_two: self.powers_of_two,
});
}
let mut instance = (self.generate)(size, index);
instance.name = if self.instances > 1 {
format!("{}/{size}#{index}", self.name)
} else {
format!("{}/{size}", self.name)
};
Ok(instance)
}
}
/// Every family, in the order a benchmark runs them.
pub static FAMILIES: &[Family] = &[
Family {
name: "3-partition-yes",
summary: "3-Partition as a Horn program with a `&` over the bins per item (MALL), two bins of the size, solvable",
sizes: &[4, 6, 8, 12],
instances: 1,
least: 4,
powers_of_two: false,
claim: "",
growth: &[],
generate: |b, _| {
let items = solvable_triples(b);
instance(three_partition(&items, 2, b), Mode::CLASSICAL, true)
},
},
Family {
name: "3-partition-no",
summary: "3-Partition as a Horn program with a `&` over the bins per item (MALL), two bins of the size, unsolvable",
sizes: &[4, 5],
instances: 1,
least: 4,
powers_of_two: false,
claim: "",
growth: &[],
generate: |b, _| {
instance(
three_partition(&unsolvable_triples(b), 2, b),
Mode::CLASSICAL,
false,
)
},
},
Family {
name: "3-partition-mll-yes",
summary: "3-Partition as Lincoln and Winkler's two-literal MLL sequent, two bins of the size, solvable",
sizes: &[4, 6, 8, 12],
instances: 1,
least: 4,
powers_of_two: false,
claim: "",
growth: &[],
generate: |b, _| {
let items = solvable_triples(b);
instance(three_partition_mll(&items, 2, b), Mode::CLASSICAL, true)
},
},
Family {
name: "3-partition-mll-no",
summary: "3-Partition as Lincoln and Winkler's two-literal MLL sequent, two bins of the size, unsolvable",
sizes: &[4, 5, 6, 8],
instances: 1,
least: 4,
powers_of_two: false,
claim: "",
growth: &[],
generate: |b, _| {
instance(
three_partition_mll(&unsolvable_triples(b), 2, b),
Mode::CLASSICAL,
false,
)
},
},
Family {
name: "partition-yes",
summary: "Matsuoka's Horn encoding of Partition (MLL) with this many random items, solvable",
sizes: &[4, 5, 6, 7, 12, 16, 20, 24, 28],
instances: 1,
least: 2,
powers_of_two: false,
claim: "",
growth: &[],
generate: |n, _| instance(partition(&partition_items(n, true)), Mode::CLASSICAL, true),
},
Family {
name: "partition-no",
summary: "Matsuoka's Horn encoding of Partition (MLL) with this many random items, unsolvable",
sizes: &[3, 4, 5, 9, 12, 14, 15],
instances: 1,
least: 3,
powers_of_two: false,
claim: "",
growth: &[],
generate: |n, _| {
instance(
partition(&partition_items(n, false)),
Mode::CLASSICAL,
false,
)
},
},
Family {
name: "qbf",
summary: "random 3-QBF with this many alternating variables (∃ first), 1.5 clauses per variable with two existentials each, under a lock-and-key encoding into MALL",
sizes: &[8, 12, 16, 20, 24, 32, 40, 44, 48],
instances: 4,
least: 3,
powers_of_two: false,
claim: "",
growth: &[],
generate: |n, index| {
let (formula, valid) = qbf(n, index);
instance(formula, Mode::CLASSICAL, valid)
},
},
// Size 0 is the empty sequent, unprovable without Mix: the wide
// families start at one pair.
Family {
name: "wide-m1",
summary: "(a1 ⅋ b1) ⊗ … ⊗ (ak ⅋ bk), ~a1 ⊗ ~b1, …, ~ak ⊗ ~bk over distinct atoms (MLL), provable; the size is k",
sizes: &[8, 16, 24, 32, 256, 2048],
instances: 1,
least: 1,
powers_of_two: false,
claim: "the inverse method past the focused engine's limits: with every atom once with each sign the linking is forced, so where the focused engine stops at its memory bound (4 096 pairs) the default hands the sequent over to the inverse method, which decides it in about quadratic time, three to four times as fast as the net engine, which takes it over under Mix",
growth: &[256, 512, 1024, 2048, 4096, 8192, 16384, 32768],
generate: |k, _| instance(wide(k, 1), Mode::CLASSICAL, true),
},
Family {
name: "wide-m2",
summary: "the wide sequent with every atom shared by two pairs: every literal twice",
sizes: &[8, 16, 24, 32, 256, 2048],
instances: 1,
least: 1,
powers_of_two: false,
claim: "",
growth: &[],
generate: |k, _| instance(wide(k, 2), Mode::CLASSICAL, true),
},
Family {
name: "wide-m3",
summary: "the wide sequent with every atom shared by three pairs: every literal three times",
sizes: &[12, 24, 30, 36, 256, 1024, 2048],
instances: 1,
least: 1,
powers_of_two: false,
claim: "",
growth: &[],
generate: |k, _| instance(wide(k, 3), Mode::CLASSICAL, true),
},
Family {
name: "wide-m4",
summary: "the wide sequent with every atom shared by four pairs: every literal four times",
sizes: &[12, 24, 28, 32, 36, 256, 1024, 2048],
instances: 1,
least: 1,
powers_of_two: false,
claim: "",
growth: &[],
generate: |k, _| instance(wide(k, 4), Mode::CLASSICAL, true),
},
Family {
name: "mix",
summary: "(ai ⊗ bi) ⊕ 0, (~ai ⊗ ~bi) ⊕ 0 for this many i under Mix (MALL, where no count equation refutes it at once), unprovable",
sizes: &[4, 6, 8, 9, 10, 11],
instances: 1,
least: 0,
powers_of_two: false,
claim: "the inverse method on Mix: it saturates in a few sequents at every size, where the focused engine alone grows about eightfold per pair as it meets the partitions of the context; the default gives the inverse method its turn after 2²⁴ units of the focused engine's work",
growth: &[4, 6, 8, 10, 12, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096],
generate: |k, _| instance(mix(k), Mode::CLASSICAL.with_mix(), false),
},
Family {
name: "counter",
summary: "the Petri-net counter !(c0 ⊗ c0 ⊸ c1), …, c0^n ⊢ cL with n tokens, n a power of two (MELL), provable within log2 n copies",
sizes: &[2, 4, 8, 16, 32, 64],
instances: 1,
least: 0,
powers_of_two: true,
claim: "the Horn engine's strength: a counter decided as Petri-net reachability in about linear time, where the focused engine forced grows steeply, about as the cube of the tokens, and the inverse method forced does not finish from 16 tokens on",
growth: &[
2, 4, 8, 16, 32, 64, 128, 256, 1024, 4096, 16384, 65_536, 262_144,
],
generate: |n, _| {
let (sequent, levels) = counter(n, false);
Instance {
copies: Some(levels),
..instance(sequent, Mode::CLASSICAL, true)
}
},
},
Family {
name: "counter-over",
summary: "the counter with the unreachable goal cL ⊗ c0 (MELL), unprovable, searched to log2 n copies",
sizes: &[2, 4, 8, 16, 32, 64],
instances: 1,
least: 0,
powers_of_two: true,
claim: "",
growth: &[],
generate: |n, _| {
let (sequent, levels) = counter(n, true);
Instance {
copies: Some(levels),
..instance(sequent, Mode::CLASSICAL, false)
}
},
},
Family {
name: "growing",
summary: "!(a ⊸ a ⊗ a), a ⊢ ?b (MELL), unprovable, a context that grows with every copy; the size is the copy bound",
sizes: &[16, 64, 256, 1024],
instances: 1,
least: 0,
powers_of_two: false,
claim: "",
growth: &[],
generate: |copies, _| Instance {
copies: Some(copies),
..instance(parse("!(a -o a * a), a |- ?b"), Mode::CLASSICAL, false)
},
},
Family {
name: "chain",
summary: "!(p0 ⊸ p1 & p2), …, !(p(k−1) ⊸ pk & p(k+1)), p0 ⊢ p(k+1) (MELL, ILL-shaped), provable within k + 2 copies; the size is k",
sizes: &[16, 64, 128, 256],
instances: 1,
// `p0 ⊢ p1` at size 0 is unprovable, against the claim.
least: 1,
powers_of_two: false,
claim: "the focused engine's memo on repeated subgoals: the `&` of each clause offers two ways on, so Fibonacci-many paths lead to the goal and meet the same subgoals, which the memo answers once (about half the stable sequents are hits), so the time grows polynomially where without the memo it passes a minute at 64; the default deepening (`--copies none`, about k/2 + 1 levels) costs nothing more, the memo answering each level's repeated subgoals; the inverse method forced is faster still, though no row of the dispatch gives it a turn on a classical sequent without Mix",
growth: &[16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192],
generate: |k, _| {
let clauses: Vec<String> = (0..k)
.map(|i| format!("!(p{i} -o p{} & p{})", i + 1, i + 2))
.collect();
let text = format!("{}, p0 |- p{}", clauses.join(", "), k + 1);
Instance {
copies: Some(k + 2),
..instance(parse(&text), Mode::CLASSICAL, true)
}
},
},
Family {
name: "additive",
summary: "A ⊢ A for A a complete tree of this depth alternating & and ⊕ over distinct atoms (the additive path), provable",
sizes: &[8, 12, 14, 16],
instances: 1,
least: 0,
powers_of_two: false,
claim: "",
growth: &[],
generate: |depth, _| {
let formula = additive_tree(depth, 0, &mut 0);
instance(
parse(&format!("{formula} |- {formula}")),
Mode::CLASSICAL,
true,
)
},
},
Family {
name: "lambek-fo",
summary: "Lambek sequents in Moot and Piazza's translation into first-order MILL, a formula at a pair of string positions; the size is k: #0 the chain a0, a0 \\ a1, …, a(k−1) \\ ak ⊢ ak, provable, #1 its first two formulas exchanged, unprovable, #2 the chain ak / a(k−1), …, a1 / a0, a0 ⊢ ak, provable, #3 a0, …, a(k−1) ⊢ a0 • … • a(k−1), provable, #4 its goal reversed, unprovable",
sizes: &[16, 64, 256, 512],
instances: 5,
// The reversed product of one formula is the product itself.
least: 2,
powers_of_two: false,
claim: "LinearOne's home ground: the Lambek calculus in first-order MILL, the fragment LinearOne's proof-net search is written for; the two-sided first-order search is faster at small sizes and meets its recursion limit of 2 048 at 512 formulas on #0 and #2 and at 1 024 on #3 and #4",
growth: &[8, 16, 32, 64, 128, 256, 512, 1024, 2048],
generate: |k, index| {
let (hypotheses, goal, provable) = lambek_shape(k, index);
instance(
parse(&lambek_first_order(&hypotheses, &goal)),
Mode::INTUITIONISTIC,
provable,
)
},
},
// #0 is proved by n copies of the clause, instantiated at z, s(z), …;
// fewer do not reach sⁿ(z). #1 is unprovable: forgetting linearity
// turns a proof into one of classical first-order logic, and the model
// whose domain is the closed terms, with p true of the terms sⁱ(z)
// alone, satisfies the hypotheses and not p(b). No refuter decides it
// (MELL1 is undecidable), so a search ends at the copy bound.
Family {
name: "chain-fo",
summary: "p(z), !∀x.(p(x) ⊸ p(s(x))) ⊢ p(sⁿ(z)) (MELL1), #0 provable within n copies, and #1 the goal p(b), unprovable, searched to n copies; the size is n",
sizes: &[16, 64, 256, 512],
instances: 2,
least: 1,
powers_of_two: false,
claim: "the first-order search's instances of one clause: n instances on one branch, decided in polynomial time until the recursion limit of 2 048, met near 400 copies, while #1 ends at its copy bound; LinearOne refuses it, its fragment being MILL1",
growth: &[16, 32, 64, 128, 256, 384, 512, 1024],
generate: |n, index| {
let goal = if index == 0 {
format!("p({}z{})", "s(".repeat(n as usize), ")".repeat(n as usize))
} else {
"p(b)".to_owned()
};
let text = format!("p(z), !forall x. (p(x) -o p(s(x))) |- {goal}");
// The last step binds x to sⁿ⁻¹(z).
Instance {
copies: Some(n),
witness_depth: Some(n - 1),
..instance(parse(&text), Mode::CLASSICAL, index == 0)
}
},
},
// The provable instances take an as the witness, and in #2 the other
// p(ai) for the factors ∃x.p(x). In #1 and #3 a cut-free proof links
// the goal's q(w) to the only q given, so w = b, and then p(b) to some
// p(ai), which differs from it: unprovable.
Family {
name: "witness-fo",
summary: "a witness among n: #0 p(a1) & … & p(an), q(an) ⊢ ∃x.(p(x) ⊗ q(x)) (MALL1), provable, #1 the same with q(b), unprovable, #2 p(a1), …, p(an), q(an) ⊢ ∃x.(p(x) ⊗ q(x)) ⊗ ∃x.p(x) ⊗ … with n − 1 factors ∃x.p(x) (MILL1), provable, #3 the same with q(b), unprovable",
sizes: &[4, 8, 10, 12],
instances: 4,
least: 1,
powers_of_two: false,
claim: "a witness among n: the first-order search tries the instances in turn, and its split steps on the MILL1 refutation #3 grow exponentially (seconds at ten) where LinearOne's proof-net search stays near flat, a pool of threads dividing that work; LinearOne refuses the MALL1 shapes #0 and #1",
growth: &[2, 4, 6, 8, 10, 12, 14, 16],
generate: |n, index| {
let alternatives: Vec<String> = (1..=n).map(|i| format!("p(a{i})")).collect();
let q = if index % 2 == 0 {
format!("q(a{n})")
} else {
"q(b)".to_owned()
};
let pair = "(exists x. (p(x) * q(x)))";
let text = if index < 2 {
format!("{}, {q} |- {pair}", alternatives.join(" & "))
} else {
let mut goal = vec![pair];
goal.extend((1..n).map(|_| "(exists x. p(x))"));
format!("{}, {q} |- {}", alternatives.join(", "), goal.join(" * "))
};
instance(parse(&text), Mode::INTUITIONISTIC, index % 2 == 0)
},
},
// #0 is proved by using the clause n + 1 times on one branch: choosing
// pi ⊸ f puts pi into the context for the next use, and once every pi
// is there the tensor takes them. #1 is unprovable: forgetting
// linearity and `!` gives ILTP's SYJ209, false in the Kripke model of
// two worlds r ≤ u with p2, …, p(n) forced at both and p1 and f at u
// alone (¬¬p1 is forced at r, the clause at both, f not at r).
Family {
name: "contraction",
summary: "Franzén's formulas in ILL: !((p1 ⊗ … ⊗ p(n)) ⊕ (p1 ⊸ f) ⊕ … ⊕ (p(n) ⊸ f) ⊸ f) ⊢ f, #0 provable within n + 1 copies (ILTP's SYJ203), #1 with ((p1 ⊸ 0) ⊸ 0) ⊸ f for p1 ⊸ f, unprovable (SYJ209), searched to n + 1 copies; the size is n",
sizes: &[2, 4, 8],
instances: 2,
least: 1,
powers_of_two: false,
claim: "the inverse method's strength on a clause used n + 1 times on one branch: the backward search finds the proof in exponentially many stable sequents and the inverse method in about n²/2 sequents, the default giving it its turn after 2²⁴ units of work; on the non-theorem #1 the backward search ends at its copy bound and the inverse method saturates, up to 16 atoms within a minute",
growth: &[2, 4, 8, 16, 32, 64, 128, 256, 512],
generate: |n, index| {
let atoms: Vec<String> = (1..=n).map(|i| format!("p{i}")).collect();
let mut alternatives = vec![format!("({})", atoms.join(" * "))];
alternatives.extend(atoms.iter().enumerate().map(|(i, p)| {
if i == 0 && index == 1 {
format!("((({p} -o 0) -o 0) -o f)")
} else {
format!("({p} -o f)")
}
}));
let text = format!("!(({}) -o f) |- f", alternatives.join(" + "));
Instance {
copies: Some(n + 1),
..instance(parse(&text), Mode::INTUITIONISTIC, index == 0)
}
},
},
// #0: each pigeon's ⊕ splits the proof, and on every branch two of the
// n + 1 pigeons share one of the n holes, whose disjunct the goal
// takes, the other atoms weakened. #1: with as many pigeons as holes
// the classical reading is false where pigeon i sits in hole i alone,
// and affine ILL proves nothing classically false.
Family {
name: "pigeonhole",
summary: "Cook's pigeonhole principle in intuitionistic affine MALL: o(p,1) ⊕ … ⊕ o(p,n) for each pigeon p ⊢ the ⊕ over the holes h and the pairs p < q of o(p,h) ⊗ o(q,h), #0 with n + 1 pigeons, provable (ILTP's SYJ202), #1 with n pigeons, unprovable; the size is the holes n",
sizes: &[2, 3, 4],
instances: 2,
least: 2,
powers_of_two: false,
claim: "the two-sided engine's weakness: the pigeons' ⊕ on the left split the search into nⁿ⁺¹ branches, which the inverse method forced avoids until its memory bound, while the default gives the inverse method no turn on a sequent without exponentials; #1, false classically, is refuted by its classical reading once the backward search has ended",
growth: &[2, 3, 4, 5, 6, 7, 8],
generate: |n, index| {
let pigeons = if index == 0 { n + 1 } else { n };
let hypotheses: Vec<String> = (0..pigeons)
.map(|p| {
let holes: Vec<String> = (0..n).map(|h| format!("o{p}_{h}")).collect();
format!("({})", holes.join(" + "))
})
.collect();
let mut pairs = Vec::new();
for h in 0..n {
for p in 0..pigeons {
for q in p + 1..pigeons {
pairs.push(format!("(o{p}_{h} * o{q}_{h})"));
}
}
}
let text = format!("{} |- {}", hypotheses.join(", "), pairs.join(" + "));
instance(parse(&text), Mode::INTUITIONISTIC.with_affine(), index == 0)
},
},
// A cut-free proof of x, E ⊢ y, E edges between atoms, is a trail from
// x to y using every edge of E once: an axiom is the empty trail, `&`
// on the left picks a direction, and `⊸` on the left sends x to the
// premise proving its argument (a premise without an atom proves
// nothing) and its result to the other. So #0 is provable, the
// square of a cycle being connected with every degree four (Euler),
// and #1 is not: without the edge v1v2 those two vertices have degree
// three.
Family {
name: "euler",
summary: "Euler tours in IMALL: v0, (va ⊸ vb) & (vb ⊸ va) for every edge ab ⊢ v0, a closed trail from v0 through every edge once, on the square of a cycle of n vertices (vi adjacent to vi±1 and vi±2); #0 provable, #1 without the edge v1v2, unprovable; the size is n",
sizes: &[5, 6, 8],
instances: 2,
least: 5,
powers_of_two: false,
claim: "the focused engines' context splits: each ⊸ divides the remaining edges between its premises and each & picks a direction, so the exhaustive search on a graph without an Euler circuit (#1) grows exponentially, about twofold per vertex, where a circuit (#0) is found at once; Chaudhuri's inverse method was measured on complete graphs, whose edges grow too fast for a curve",
growth: &[5, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24],
generate: |n, index| {
let mut edges: Vec<(u32, u32)> = (0..n)
.flat_map(|i| [1, 2].map(|d| (i.min((i + d) % n), i.max((i + d) % n))))
.filter(|&edge| index == 0 || edge != (1, 2))
.collect();
edges.sort_unstable();
let edges: Vec<String> = edges
.iter()
.map(|(a, b)| format!("((v{a} -o v{b}) & (v{b} -o v{a}))"))
.collect();
let text = format!("v0, {} |- v0", edges.join(", "));
instance(parse(&text), Mode::INTUITIONISTIC, index == 0)
},
},
// m − 1 uses of the clause turn the one `a` into m, all on the branch
// before the goal's tensor splits them; a proof of ILL is one of
// classical linear logic, so #1 is provable too.
Family {
name: "duplicate",
summary: "!(a ⊸ a ⊗ a), a ⊢ a ⊗ … ⊗ a with m factors, provable within m − 1 copies on one branch, #0 intuitionistic (IMELL), #1 classical (MELL), so that the provers of either run it; the size is m",
sizes: &[2, 4, 8],
instances: 2,
least: 1,
powers_of_two: false,
claim: "the other provers' caps on contraction: one clause used m − 1 times on one branch, past lolli's two uses of a clause, APLL's bound of 3, llprover's threshold of 5 and the Maude prover's copy bound of 5, while deduxit's Horn engine decides it in about linear time",
growth: &[
2, 4, 8, 16, 32, 64, 128, 256, 1024, 4096, 16384, 65_536, 262_144,
],
generate: |m, index| {
let text = format!("!(a -o a * a), a |- {}", power("a", m, " * "));
let mode = if index == 0 {
Mode::INTUITIONISTIC
} else {
Mode::CLASSICAL
};
Instance {
copies: Some(m - 1),
..instance(parse(&text), mode, true)
}
},
},
// #0 is proved by n uses of the clause along the path n0, …, n(n−1),
// z, nested on one branch, at whose end the first edge is a copy of
// its own: n + 1 copies. #1 is unprovable: forgetting linearity, the model whose domain is
// the constants, with e true of the edges given and r of the cycle's
// nodes alone, satisfies the hypotheses and not r(z). No refuter
// decides it in general, so a search ends at a bound or the limit.
Family {
name: "reach-fo",
summary: "reachability through a cycle (IMELL1): !e(n0,n1), …, !e(n(n−1),n0), r(n0), !∀x∀y.(e(x,y) ⊗ r(x) ⊸ r(y)) ⊢ r(z), #0 with the exit !e(n(n−1),z), provable within n + 1 copies, #1 without it, unprovable, searched to n + 1 copies; the size is the cycle's nodes n, one a loop",
sizes: &[2, 4],
instances: 2,
least: 1,
powers_of_two: false,
claim: "the first-order memo's loop check: a search through a cycle comes back to an instance it has met, which the memo recognises, so the exit is found (without the memo the search passes a minute at 8 nodes, with it at 12); the time still grows steeply with the cycle; #1, the cycle without an exit, ends at its copy bound",
growth: &[2, 3, 4, 5, 6, 8, 10, 12, 16],
generate: |n, index| {
let mut hypotheses: Vec<String> = (0..n)
.map(|i| format!("!e(n{i}, n{})", (i + 1) % n))
.collect();
if index == 0 {
hypotheses.push(format!("!e(n{}, z)", n - 1));
}
let text = format!(
"{}, r(n0), !forall x. forall y. (e(x, y) * r(x) -o r(y)) |- r(z)",
hypotheses.join(", ")
);
Instance {
copies: Some(n + 1),
..instance(parse(&text), Mode::INTUITIONISTIC, index == 0)
}
},
},
];
/// Returns the family of that name.
pub fn find(name: &str) -> Option<&'static Family> {
FAMILIES.iter().find(|family| family.name == name)
}
/// Encodes a 3-Partition instance as a linear Horn program in the style of
/// Kanovich's encodings: bin `j` offers `size` units `bj` and three slots
/// `tj`; item `i` is a `&` over the bins of a clause taking its units and
/// a slot from that bin and producing `di`; the goal is the tensor of every
/// `di`. Every resource must be used exactly once, so the sequent is
/// provable if and only if the items split into triples of sum `size`,
/// one per bin.
pub fn three_partition(items: &[u32], bins: u32, size: u32) -> Sequent {
let mut hypotheses = Vec::new();
for j in 1..=bins {
hypotheses.extend((0..size).map(|_| format!("b{j}")));
hypotheses.extend((0..3).map(|_| format!("t{j}")));
}
for (i, &a) in items.iter().enumerate() {
let clauses: Vec<String> = (1..=bins)
.map(|j| format!("({} * t{j} -o d{i})", power(&format!("b{j}"), a, " * ")))
.collect();
hypotheses.push(clauses.join(" & "));
}
let goal: Vec<String> = (0..items.len()).map(|i| format!("d{i}")).collect();
parse(&format!(
"{} |- {}",
hypotheses.join(", "),
goal.join(" * ")
))
}
/// Encodes a 3-Partition instance as Lincoln and Winkler's two-literal MLL
/// sequent `⊢ k ⊗ (~c ⅋ … ⅋ ~c), …, ((~k ⅋ ~k ⅋ ~k) ⅋ (c ⊗ … ⊗ c)) ⊗ …`:
/// an item of size `s` is `k ⊗` a par of `s` literals `~c`, a bin is three
/// `~k` beside a tensor of `size` literals `c`, and the bins are joined by
/// `⊗`, so that each bin's premise takes three items of total `size`.
pub fn three_partition_mll(items: &[u32], bins: u32, size: u32) -> Sequent {
let mut roots: Vec<String> = items
.iter()
.map(|&s| format!("k * ({})", power("~c", s, " | ")))
.collect();
let bin = format!("((~k | ~k | ~k) | ({}))", power("c", size, " * "));
roots.push(power(&bin, bins, " * "));
parse(&format!("|- {}", roots.join(", ")))
}
/// Encodes a Partition instance as a Horn sequent, as Matsuoka does: every
/// item `a_i` of size `s_i` may be turned into `s_i` units `b` or `s_i`
/// units `c`; half the total of each buys the items back once, and half
/// the total of each buys the goal `e`. Every clause is used exactly once,
/// so the sequent is provable if and only if the sizes split into two
/// halves of equal sum.
///
/// # Panics
///
/// If the sizes add up to an odd total.
pub fn partition(sizes: &[u32]) -> Sequent {
let total: u32 = sizes.iter().sum();
assert_eq!(total % 2, 0, "the total is even");
let items: Vec<String> = (1..=sizes.len()).map(|i| format!("a{i}")).collect();
let mut hypotheses = vec![items.join(" * ")];
for unit in ["b", "c"] {
for (i, &size) in sizes.iter().enumerate() {
hypotheses.push(format!("(a{} -o {})", i + 1, power(unit, size, " * ")));
}
}
let half = format!(
"({} * {})",
power("b", total / 2, " * "),
power("c", total / 2, " * ")
);
hypotheses.push(format!("({half} -o {})", items.join(" * ")));
hypotheses.push(format!("({half} -o e)"));
parse(&format!("{} |- e", hypotheses.join(", ")))
}
/// The counter program as Petri-net reachability: `tokens` tokens `c0`,
/// and reusable clauses `!(ci ⊗ ci ⊸ c(i+1))` doubling up to the goal
/// `cL`, `L` the logarithm of `tokens`; with `over`, the goal is `cL ⊗
/// c0`, which no firing reaches. Returns the sequent and `L`, the copies
/// a branch to one token of the goal takes.
///
/// # Panics
///
/// If `tokens` is not a power of two.
pub fn counter(tokens: u32, over: bool) -> (Sequent, u32) {
assert!(tokens.is_power_of_two(), "{tokens} tokens");
let levels = tokens.trailing_zeros();
let mut hypotheses: Vec<String> = (0..levels)
.map(|i| format!("!(c{i} * c{i} -o c{})", i + 1))
.collect();
hypotheses.extend((0..tokens).map(|_| "c0".to_owned()));
let extra = if over { " * c0" } else { "" };
let text = format!("{} |- c{levels}{extra}", hypotheses.join(", "));
(parse(&text), levels)
}
/// The wide sequent `⊢ (a1 ⅋ b1) ⊗ … ⊗ (ak ⅋ bk), ~a1 ⊗ ~b1, …, ~ak ⊗ ~bk`
/// with pair `i` using the atoms of index `i / shared`, so that every
/// literal occurs `shared` times (the last atoms fewer when `shared` does
/// not divide `k`). Provable: identifying atoms keeps a proof a proof.
pub fn wide(k: u32, shared: u32) -> Sequent {
let pairs: Vec<String> = (0..k)
.map(|i| format!("(a{0} | b{0})", i / shared))
.collect();
let mut roots = vec![pairs.join(" * ")];
roots.extend((0..k).map(|i| format!("~a{0} * ~b{0}", i / shared)));
parse(&format!("|- {}", roots.join(", ")))
}
/// The unprovable sequent `⊢ (a1 ⊗ b1) ⊕ 0, (~a1 ⊗ ~b1) ⊕ 0, …` of `k`
/// pairs: every axiom linking of a pair closes a cycle through its two
/// `⊗`, with Mix too, and the `⊕ 0` makes it MALL, where the count
/// equation that refutes the bare tensors at once does not hold, so that
/// a search under Mix meets the partitions of the context.
pub fn mix(k: u32) -> Sequent {
let roots: Vec<String> = (0..k)
.map(|i| format!("(a{i} * b{i}) + 0, (~a{i} * ~b{i}) + 0"))
.collect();
parse(&format!("|- {}", roots.join(", ")))
}
/// A random closed 3-QBF and its truth value: `n` variables quantified
/// alternately, `∃` first, and `⌈1.5 n⌉` clauses of three distinct
/// variables of which at least two are existential (Gent and Walsh's
/// model A, which keeps random instances from being false outright), with
/// random signs; seeded by `n` and `index`. The formula is encoded into
/// MALL by locks and keys: a `∀x` is `(~tx & ~fx) ⅋ S`, putting one of the
/// two values into the context on each premise; an `∃x` is `((~tx ⅋ ~kx)
/// ⊕ (~fx ⅋ ~kx)) ⅋ (kx ⊗ S)`, whose rest `S` needs the key `~kx` that
/// only the choice of a value releases, so the choice precedes every
/// later quantifier; the matrix is the `&` of the clauses, a clause the
/// `⊕` of `(tx ⊗ ⊤)` or `(fx ⊗ ⊤)` over its literals, provable when the
/// context holds a value that satisfies one of them.
///
/// # Panics
///
/// For fewer than three variables, which leave no clause to draw.
pub fn qbf(n: u32, index: u32) -> (Sequent, bool) {
assert!(n >= 3, "a QBF of {n} variables has no clause of three");
let mut rng = Rng((u64::from(n) << 32) | u64::from(index));
let existential = |v: u32| v.is_multiple_of(2);
let mut clauses: Vec<[(u32, bool); 3]> = Vec::new();
while clauses.len() < (3 * n).div_ceil(2) as usize {
let x = rng.below(n);
let y = rng.below(n);
let z = rng.below(n);
let vars = [x, y, z];
if x == y || y == z || x == z || vars.iter().filter(|&&v| existential(v)).count() < 2 {
continue;
}
clauses.push(vars.map(|v| (v, rng.below(2) == 0)));
}
let valid = qbf_value(n, &clauses, &mut Vec::new());
let literal =
|(v, positive): (u32, bool)| format!("({}{v} * top)", if positive { "t" } else { "f" });
let matrix: Vec<String> = clauses
.iter()
.map(|clause| format!("({})", clause.map(literal).join(" + ")))
.collect();
let mut formula = format!("({})", matrix.join(" & "));
for v in (0..n).rev() {
formula = if existential(v) {
format!("(((~t{v} | ~k{v}) + (~f{v} | ~k{v})) | (k{v} * {formula}))")
} else {
format!("((~t{v} & ~f{v}) | {formula})")
};
}
(parse(&format!("|- {formula}")), valid)
}
/// Evaluates the QBF of [`qbf`] under the values of the first variables.
fn qbf_value(n: u32, clauses: &[[(u32, bool); 3]], values: &mut Vec<bool>) -> bool {
let v = values.len() as u32;
if v == n {
return clauses
.iter()
.all(|clause| clause.iter().any(|&(x, sign)| values[x as usize] == sign));
}
let mut outcomes = [true, false].into_iter().map(|value| {
values.push(value);
let result = qbf_value(n, clauses, values);
values.pop();
result
});
if v.is_multiple_of(2) {
outcomes.any(|b| b)
} else {
outcomes.all(|b| b)
}
}
/// Six item sizes for two bins of size `b` (at least four) that split into
/// the triples `1, 2, b − 3`.
fn solvable_triples(b: u32) -> [u32; 6] {
assert!(b >= 4, "bins of {b}");
[1, 2, b - 3, 1, 2, b - 3]
}
/// Six item sizes for two bins of size `b` (at least four) that admit no
/// 3-Partition: the item `b − 1` needs two more items of total one.
fn unsolvable_triples(b: u32) -> [u32; 6] {
assert!(b >= 4, "bins of {b}");
[1, 1, 1, b - 1, 1, b - 3]
}
/// Returns `n` random item sizes from 1 to `n` with an even total whose
/// Partition instance is solvable or not as asked: the first seed, counted
/// from one derived from `n` and `solvable`, whose sizes answer so.
fn partition_items(n: u32, solvable: bool) -> Vec<u32> {
// Two items of sizes one or two always split, one never does.
assert!(n >= 2 + u32::from(!solvable), "{n} items");
let mut rng = Rng((u64::from(n) << 1) | u64::from(solvable));
loop {
let sizes: Vec<u32> = (0..n).map(|_| 1 + rng.below(n)).collect();
let total: u32 = sizes.iter().sum();
if total % 2 == 1 {
continue;
}
// Subset sums, as a bit per reachable sum.
let mut reachable = vec![false; total as usize + 1];
reachable[0] = true;
for &s in &sizes {
for sum in (s..=total).rev() {
reachable[sum as usize] |= reachable[(sum - s) as usize];
}
}
if reachable[total as usize / 2] == solvable {
return sizes;
}
}
}
/// A complete tree of the given depth over the atoms `x<next>`, `&` at even
/// levels and `⊕` at odd ones.
fn additive_tree(depth: u32, level: u32, next: &mut u32) -> String {
if depth == level {
*next += 1;
return format!("x{}", *next - 1);
}
let left = additive_tree(depth, level + 1, next);
let right = additive_tree(depth, level + 1, next);
let connective = if level.is_multiple_of(2) { "&" } else { "+" };
format!("({left} {connective} {right})")
}
/// A formula of the Lambek calculus, which [`lambek_first_order`]
/// translates.
#[derive(Clone, Debug)]
enum Lambek {
/// An atom by its name.
Atom(String),
/// `B \ A`: the argument `B`, wanted on the left, and the result `A`.
Under(Box<Lambek>, Box<Lambek>),
/// `A / B`: the result `A` and the argument `B`, wanted on the right.
Over(Box<Lambek>, Box<Lambek>),
/// `A • B`.
Product(Box<Lambek>, Box<Lambek>),
}
impl Lambek {
/// The atom `a<i>`.
fn atom(i: u32) -> Self {
Self::Atom(format!("a{i}"))
}
/// The formula in the syntax of the Lambek calculus this crate's
/// ordered parse reads, every binary formula in parentheses.
#[cfg(test)]
fn text(&self) -> String {
match self {
Self::Atom(name) => name.clone(),
Self::Under(b, a) => format!("({} \\ {})", b.text(), a.text()),
Self::Over(a, b) => format!("({} / {})", a.text(), b.text()),
Self::Product(a, b) => format!("({} * {})", a.text(), b.text()),
}
}
/// Moot and Piazza's translation of the formula at the string
/// positions `x` to `y` into first-order MILL: an atom `p` is the
/// predicate `p(x, y)`, `A / B` is `∀z.(B(y, z) ⊸ A(x, z))`, `B \ A` is
/// `∀z.(B(z, x) ⊸ A(z, y))` and `A • B` is `∃z.(A(x, z) ⊗ B(z, y))`,
/// each `z` a variable of its own (`v<n>`, numbered by `fresh`).
fn first_order(&self, x: &str, y: &str, fresh: &mut u32) -> String {
if let Self::Atom(name) = self {
return format!("{name}({x}, {y})");
}
*fresh += 1;
let z = format!("v{fresh}");
match self {
Self::Atom(_) => unreachable!("handled above"),
Self::Over(a, b) => format!(
"(forall {z}. ({} -o {}))",
b.first_order(y, &z, fresh),
a.first_order(x, &z, fresh)
),
Self::Under(b, a) => format!(
"(forall {z}. ({} -o {}))",
b.first_order(&z, x, fresh),
a.first_order(&z, y, fresh)
),
Self::Product(a, b) => format!(
"(exists {z}. ({} * {}))",
a.first_order(x, &z, fresh),
b.first_order(&z, y, fresh)
),
}
}
}
/// The Lambek sequent of shape `index` at size `k` (at least two) of the
/// family `lambek-fo`, and whether it is provable. The verdicts hold of
/// the translations directly, by the positions alone, since the positions
/// `i0`, …, `in` are distinct constants and a cut-free proof of
/// multiplicative linear logic without units links every atom of its
/// conclusion's instance to one of the same predicate, arguments and
/// opposite side:
///
/// - the chain is proved with every `∀` instantiated by `i0`, the chain
/// of `/` with every `∀` instantiated by the last position;
/// - the exchanged chain is not: `a0 \ a1` at `i0, i1` asks for `a0(z, i0)`,
/// and the only `a0` given stands at `i1, i2`;
/// - the product is proved with the positions as the witnesses;
/// - the reversed product is not: its leftmost atom `a(k−1)` stands at
/// `i0` and the only one given at `i(k−1)`.
fn lambek_shape(k: u32, index: u32) -> (Vec<Lambek>, Lambek, bool) {
let under = |i: u32| Lambek::Under(Box::new(Lambek::atom(i)), Box::new(Lambek::atom(i + 1)));
let product = |order: &mut dyn Iterator<Item = u32>| {
let first = Lambek::atom(order.next().expect("two atoms"));
order.fold(first, |left, i| {
Lambek::Product(Box::new(left), Box::new(Lambek::atom(i)))
})
};
match index {
0 | 1 => {
let mut hypotheses = vec![Lambek::atom(0)];
hypotheses.extend((0..k).map(under));
if index == 1 {
hypotheses.swap(0, 1);
}
(hypotheses, Lambek::atom(k), index == 0)
}
2 => {
let mut hypotheses: Vec<Lambek> = (1..=k)
.rev()
.map(|i| Lambek::Over(Box::new(Lambek::atom(i)), Box::new(Lambek::atom(i - 1))))
.collect();
hypotheses.push(Lambek::atom(0));
(hypotheses, Lambek::atom(k), true)
}
3 => (
(0..k).map(Lambek::atom).collect(),
product(&mut (0..k)),
true,
),
_ => (
(0..k).map(Lambek::atom).collect(),
product(&mut (0..k).rev()),
false,
),
}
}
/// The two-sided text of the Lambek sequent `hypotheses ⊢ goal` in Moot
/// and Piazza's translation: the `i`th hypothesis at the positions `i`
/// to `i + 1`, the goal from the first position to the last, each
/// position a constant `i<n>`.
fn lambek_first_order(hypotheses: &[Lambek], goal: &Lambek) -> String {
let mut fresh = 0;
let translated: Vec<String> = hypotheses
.iter()
.enumerate()
.map(|(i, h)| h.first_order(&format!("i{i}"), &format!("i{}", i + 1), &mut fresh))
.collect();
let goal = goal.first_order("i0", &format!("i{}", hypotheses.len()), &mut fresh);
format!("{} |- {goal}", translated.join(", "))
}
/// `n` copies of `formula` joined by `separator`.
fn power(formula: &str, n: u32, separator: &str) -> String {
vec![formula; n as usize].join(separator)
}
/// Parses a sequent the generators built.
fn parse(text: &str) -> Sequent {
// A large size is the caller's to ask for: no bound on occurrences or
// terms.
let unbounded = crate::Limits::default()
.with_occurrences(None)
.with_terms(None);
Sequent::parse_within(text, &unbounded)
.unwrap_or_else(|e| panic!("a generated sequent parses: {e}"))
}
/// An instance with a name to be filled in by [`Family::instance`].
fn instance(sequent: Sequent, mode: Mode, provable: bool) -> Instance {
Instance {
name: String::new(),
sequent,
mode,
provable,
copies: None,
witness_depth: None,
}
}
/// SplitMix64: a small seeded generator, so that a random instance is a
/// function of its seed.
struct Rng(u64);
impl Rng {
/// Returns a number below `n`.
fn below(&mut self, n: u32) -> u32 {
self.0 = self.0.wrapping_add(0x9E37_79B9_7F4A_7C15);
let mut z = self.0;
z = (z ^ (z >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
z = (z ^ (z >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
((z ^ (z >> 31)) % u64::from(n)) as u32
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::linear::search::{Options, Verdict, prove};
/// The smallest instances of every family are decided as the family
/// says, and every proof checks; the random families are the same on
/// every call. The unsolvable Horn 3-Partition is left out: refuting
/// its smallest instance takes a minute, which is what the benchmarks
/// measure.
#[test]
fn verdicts_as_constructed() {
for family in FAMILIES.iter().filter(|f| f.name != "3-partition-no") {
for index in 0..family.instances {
let instance = family.instance(family.sizes[0], index).unwrap();
let options = Options::default()
.with_copies(Some(instance.copies.unwrap_or(Options::DEFAULT_COPIES)))
.with_witness_depth(
instance
.witness_depth
.unwrap_or(Options::DEFAULT_WITNESS_DEPTH),
);
let outcome = prove(&instance.sequent, instance.mode, &options).unwrap();
match &outcome.verdict {
Verdict::Proved(proof) => {
assert!(instance.provable, "{} proved", instance.name);
proof.check(instance.mode).unwrap();
}
Verdict::Unprovable(_) => {
assert!(!instance.provable, "{} refuted", instance.name)
}
// Exponential families without a proof are bounded.
Verdict::Unknown(_) => assert!(
!instance.provable && instance.copies.is_some(),
"{}: {:?}",
instance.name,
outcome.verdict
),
}
assert_eq!(
instance.sequent,
family.instance(family.sizes[0], index).unwrap().sequent
);
}
}
}
/// The QBF encoding agrees with the formula's truth value on many
/// small random instances, true and false ones alike.
#[test]
fn qbf_encoding() {
let mut valid = 0;
for n in 3..=7 {
for index in 0..12 {
let (sequent, expected) = qbf(n, index);
let outcome = prove(&sequent, Mode::CLASSICAL, &Options::default()).unwrap();
assert_eq!(
outcome.verdict.proof().is_some(),
expected,
"qbf({n}, {index})"
);
assert!(expected || matches!(outcome.verdict, Verdict::Unprovable(_)));
valid += usize::from(expected);
}
}
assert!((10..50).contains(&valid), "{valid} of 60 valid");
}
/// The Lambek sequents behind `lambek-fo` have the verdicts the
/// family claims for their translations, as the planar search of the
/// Lambek calculus decides them, and the translations are decided so
/// by the first-order search: the translation keeps provability.
#[test]
fn lambek_translations_keep_the_planar_verdicts() {
let family = find("lambek-fo").unwrap();
for k in 2..=5 {
for index in 0..family.instances {
let (hypotheses, goal, provable) = lambek_shape(k, index);
let written: Vec<String> = hypotheses.iter().map(Lambek::text).collect();
let text = format!("{} |- {}", written.join(", "), goal.text());
let original = Sequent::parse_ordered(
&text,
crate::linear::Order::Cyclic,
&crate::Limits::default(),
)
.unwrap();
let planar = prove(&original, Mode::LAMBEK, &Options::default()).unwrap();
assert_eq!(planar.verdict.proof().is_some(), provable, "{text}");
let translated = family.instance(k, index).unwrap();
let first_order =
prove(&translated.sequent, translated.mode, &Options::default()).unwrap();
assert_eq!(
first_order.verdict.proof().is_some(),
provable,
"{}",
translated.sequent
);
}
}
}
/// A size a family has no instance of is refused, never a panic.
#[test]
fn refuses_sizes_without_instances() {
for (name, size, least, powers_of_two) in [
("qbf", 2, 3, false),
("wide-m1", 0, 1, false),
("3-partition-no", 3, 4, false),
("partition-no", 2, 3, false),
("counter", 6, 0, true),
] {
let family = find(name).unwrap();
assert_eq!(
family.instance(size, 0).map(|i| i.name),
Err(Error::FamilySize {
family: family.name,
size,
least,
powers_of_two
}),
"{name}"
);
}
// Every family answers every size from 0 up past its least with an
// instance or this refusal, never a panic: a wrong `least` was one.
for family in FAMILIES {
for size in 0..=family.least + 2 {
let instance = family.instance(size, 0);
assert!(
matches!(instance, Ok(_) | Err(Error::FamilySize { .. })),
"{} {size}",
family.name
);
}
}
}
}