Orthogonal Vectors Conjecture

• Orthogonal Vectors Conjecture is a fine-grained complexity hypothesis stating that no algorithm can solve the OV problem in n^(2-o(1)) time for dimensions d = ω(log n).
• The conjecture underpins many conditional lower bounds by linking the OV problem's complexity to SETH and related problems in theoretical computer science.
• Recent algorithmic advances reduce the exponential dependence on dimension d, while rigorous lower bounds in restricted computational models reinforce the conjecture's significance.

The Orthogonal Vectors Conjecture (OVC) concerns the complexity of the Orthogonal Vectors (OV) problem, a central object in fine-grained complexity theory with far-reaching implications for polynomial-time algorithms and conditional lower bounds within P. The OV problem asks, given n Boolean vectors of dimension d, whether there exists a pair that are orthogonal, i.e., have no coordinate where both are 1. The OVC posits that, for dimensions exceeding ω(log n), no algorithm can solve OV in $n^{2-o(1)}$ time, and that this barrier reflects a fundamental computational hardness even for nonuniform models such as branching programs and Boolean formulas. The conjecture is tightly linked to the Strong Exponential Time Hypothesis (SETH) and underpins many hardness results for problems across theoretical computer science.

1. Formal Definition and Statement of the Conjecture

Given $n$ vectors $v_1, \dots, v_n \in \{0,1\}^d$, the OV problem is to decide whether there exist distinct $i, j \in [n]$ such that $\langle v_i, v_j \rangle = 0$, where $\langle \cdot,\cdot \rangle$ denotes the standard inner product. This can be equivalently phrased as the existence of two disjoint sets among the sets corresponding to the vectors' supports.

The OV Conjecture states: For every $\varepsilon > 0$, there exists $c(\varepsilon) = \omega(1)$ such that for $d = c(\varepsilon)\cdot \log n$, no algorithm can solve OV in $n^{2-\varepsilon}$ time. Equivalently, for all $d = \omega(\log n)$, every algorithm for OV requires time $n^{2-o(1)}$ (Kane et al., 2017, Dürr et al., 15 Jul 2025).

For the $k$-Orthogonal Vectors (k-OV) generalization, given $k$ sets $U_1,\dots,U_k \subseteq \{0,1\}^d$, each of size $n$, the k-OV problem asks whether there exist $u_1 \in U_1$, ..., $u_k \in U_k$ such that for all coordinates $j \in [d]$, $\prod_{\ell=1}^k u_{\ell}[j]=0$. The k-OV Conjecture posits that for $d = \omega(\log n)$, no $n^{k-o(1)}$-time algorithm exists (Kühnemann et al., 30 Apr 2025).

2. Algorithmic Landscape and Lower Bounds

Baseline deterministic algorithms include $O(n^2 d)$ time pairwise checking and $O(2^d n)$ time algorithms using Boolean zeta transforms for small $d$. The core fine-grained complexity regime sets $d = \Theta(\log n)$ or larger.

Assuming SETH, it is established that no $2^{o(d)} n^{2-\varepsilon}$-time algorithm can solve 2-OV for any fixed $\varepsilon>0$ and $d \gg \log n$ (Dürr et al., 15 Jul 2025, Abboud et al., 2018). In moderate or low dimensions, prior to recent advancements, the best known upper bound was $O(2^d n)$.

A foundational result of Kane and Williams proves that, for large enough $n, d$, every branching program evaluating OV must have size $\tilde{\Omega}(n \cdot \min(n,2^d))$, which matches known constructions up to polylogarithmic factors. The same bound holds for DeMorgan formulas of fixed fan-in and for formulas of arbitrary symmetric gates (counted by wire complexity) (Kane et al., 2017).

These models, therefore, evidence that under OVC, $n^2$ (up to polylog factors) is a true barrier for OV in $\mathbf{P}$, not just for specific computational architectures but for broad classes of nonuniform models as well.

3. Techniques for Proving Circuit Lower Bounds

The lower bounds for OV in restricted models are established via an input-restriction method, differing significantly from classical random restriction arguments. Central to the proof is the construction of hard functions on the "middle" Hamming layer (i.e., weight-$d/2$ strings). For any formula or branching program solver $F$ for OV, a pigeonhole argument identifies at least one vector whose bits occur infrequently, and by fixing all other vectors, one can enforce computation of a hard $d$-bit function on that vector. By exploiting the combinatorial structure of functions on the middle layer, it is shown that such $F$ must have size at least $\tilde{\Omega}(n \cdot \min(n, 2^d))$ (Kane et al., 2017).

In all three models—branching programs, constant fan-in formulas, and unbounded fan-in symmetric-gate formulas—this approach yields lower bounds that essentially match those for any explicit function known so far (Nechiporuk's bound).

4. Recent Algorithmic Progress

Significant recent advances have lowered the exponential dependence on $d$ in OV algorithms for small dimensions. Williams showed that the Boolean disjointness matrix admits a decomposition of equality-rank $O(1.35^d)$, yielding a $\tilde{O}(1.35^d n)$ time algorithm (Dürr et al., 15 Jul 2025). Further improvements, based on combinatorial sampling certificate techniques and computer-aided optimization, reduce this to $\tilde{O}(1.25^d n)$ and, with numerical tuning, to $O(1.16^d n)$ (Dürr et al., 15 Jul 2025). These approaches exploit entropy-based block partitioning and leverage information-theoretic arguments about the sparsity of orthogonal pairs.

For $k$-OV, analogous algorithmic tools yield $O(2^{(1-\varepsilon_k)d}n)$ runtime upper bounds for each fixed $k\ge2$, and under the Set Cover Conjecture, these are optimal in the sense that reducing the base of the exponent would contradict established hardness for set cover (Dürr et al., 15 Jul 2025).

5. Average-Case Hardness and Planted Distributions

While OVC holds robustly in the worst-case setting, the behavior differs dramatically on random inputs. Kane and Williams prove that for product distributions (i.i.d. Bernoulli input vectors), OV can be decided for almost all instances by $AC^0$ formulas of size $O(n^{2-\epsilon_p})$ for some $\epsilon_p>0$, demonstrating that average-case OV is easy under mild random distributions. This rules out lower-bound proofs for OVC relying on independent random restrictions (Kane et al., 2017).

In contrast, the planted k-OV model introduces a specific hard distribution in which a unique solution is planted among vectors with appropriately chosen $p$-biased noise, preserving $(k-1)$-wise independence. It is conjectured that distinguishing between such planted and random instances requires time $n^{k-o(1)}$ on average, tightly paralleling the worst-case conjecture and opening avenues for fine-grained cryptographic applications. Search-to-decision reductions based on binary search and counter-based amplifications further reinforce the plausibility of average-case hardness (Kühnemann et al., 30 Apr 2025).

6. Complexity-Theoretic Implications and Reductions

The OVC is a cornerstone assumption in fine-grained complexity, underlying hardness results for numerous problems in $\mathbf{P}$ via succinct reductions. Its failure would yield truly faster algorithms for long-standing open problems: for example, $O(n^{(1-\varepsilon)k})$-time algorithms for Min-Weight $k$-Clique in hypergraphs and SAT for sparse threshold circuits (TC$^1$), which are currently far from reach (Abboud et al., 2018). It is shown that SETH implies OVC, and the Weighted Clique conjecture in turn implies OVC.

Reductions are established using encoding constructions: k-OV can be reduced to 2-OV in $O(n^{\lceil k/2\rceil} D)$ time and Min-Weight $k$-Clique can be encoded via Boolean vector representations, indicating that improvement in OV would cascade into breakthroughs for clique detection and improved circuit SAT for TC$^1$ (Abboud et al., 2018).

The OV conjecture's role in fine-grained reductions makes it central for establishing equivalences or separations in the complexity of natural polynomial-time problems, and its linkages to the Set Cover Conjecture and k-SUM further solidify its status in computational complexity theory (Dürr et al., 15 Jul 2025).

7. Open Questions and Research Directions

Central open questions include:

• Determining the precise relationship between OVC, APSP, and 3SUM, and whether one can base OV-hardness directly on other fundamental conjectures such as APSP or 3SUM.
• Understanding the threshold regimes of parameters ($p$, $\alpha(n)$) that make the planted average-case k-OV instances hardest for algorithms, particularly in the context of fine-grained cryptography (Kühnemann et al., 30 Apr 2025).
• Improving the depth-dependence in the reduction from threshold circuits to CNF-SAT that follows from OVC refutations (Abboud et al., 2018).
• Constructing cryptographic primitives such as public-key encryption from average-case hard planted k-OV, and determining whether sub-$n^k$ average-case algorithms exist for these distributions.

Progress in these areas would have deep ramifications for both complexity theory and cryptographic applications, potentially reshaping our understanding of hardness within $\mathbf{P}$ and beyond.