Planted problems in fine-grained complexity

Worst-case conjectures are not enough for cryptography; planting is how you get the rest

Motivation

Fine-grained complexity is organized around a few worst-case conjectures — that \(k\)-SAT needs \(2^{n}\) time, that \(k\)-Orthogonal Vectors needs \(n^{k}\) — and it derives a large body of conditional lower bounds from them. Cryptography cannot use those conjectures directly. It needs three things they do not supply: average-case hardness rather than worst-case, instances that can be sampled together with a solution, and a solution that is unique so the sampler knows what it planted.

Planting is the standard way to get all three, and the design question is what kind of planting leaves the problem hard. A careless planting introduces statistical signatures that make the instance easy to recognize long before it is easy to solve; a good one is invisible to everything short of an actual solver.

The payoff on offer is fine-grained public-key cryptography. The field has key exchange from fine-grained assumptions, due to LaVigne, Lincoln and Vassilevska Williams, and not much beyond it.

Provenance and history

David Kühnemann, Adam Polak and Alon Rosen, The Planted Orthogonal Vectors Problem (IACR ePrint 2025/780), propose a planting for \(k\)-OV and conjecture it is hard on average.

The construction plants a solution among vectors with i.i.d. \(p\)-biased entries, for an appropriately chosen \(p\), so that the planted solution is the unique one. Its distinguishing property is a marginal one: any subset of strictly fewer than \(k\) vectors has exactly the same distribution as in the unplanted model. The planting is invisible below its own arity — not approximately, exactly.

That property is not just cosmetic. It is what powers their search-to-decision reduction: split a set in two and resample one half, and by \((k-1)\)-wise independence the resampled instance follows the model distribution, so a decision oracle reveals whether the planted vector survived. Recursing gives binary search for the planted solution at a cost of \(O(\log n)\) in time and error.

Parameter lattice

What is conjectured, and at what arity an attack could live.

Regime Status
Worst-case \(k\)-OV \(n^{k-o(1)}\) conjectured; central to fine-grained complexity
Planted decision \(k\)-OV, algorithms seeing \(< k\) vectors provably no advantage — marginals identical
Planted decision \(k\)-OV, general c/0098 — open
Planted search \(k\)-OV reduces to decision, \(O(\log n)\) overhead (Theorem 12)

Note the parameter mismatch across the two results, which the source does not flag: Conjecture 10 is stated for \(\alpha(n) = \omega(1)\), while the search-to-decision reduction requires \(\alpha(n) = \mathrm{polylog}(n)\).

What kind of assumption this is

Worth being explicit, since the distinction governs whether a statement belongs in an archive of open problems at all. This is a newly proposed average-case assumption, not a question the literature had already posed. It is recorded here because the underlying object is canonical and the planting is analysed rather than merely asserted — the marginal property and the search-to-decision reduction are theorems. It is not recorded as settled evidence of hardness: one paper without an attack is weak evidence, and the natural place for an attack is whether the uniqueness of the planted solution, which is what makes the planting useful, is itself a handle.

Statements in this hub