Orthogonal Vectors Conjecture (OVC)

• Orthogonal Vectors Conjecture is the assertion that no truly subquadratic algorithm exists for finding orthogonal pairs in high-dimensional binary vectors under specified conditions.
• It underpins a wide range of conditional lower bounds and fine-grained reductions, impacting problems like Max-IP, Subset Sum, and various circuit complexity models.
• Recent advances in deterministic and randomized algorithms, as well as studies on average-case hardness, highlight both the current barriers and open research directions within this complexity paradigm.

The Orthogonal Vectors Conjecture (OVC) occupies a central role in fine-grained complexity theory, positing that certain seemingly simple problems in moderate to high dimensions fundamentally resist truly subquadratic algorithms. These conjectures bridge lower bound methodology, algorithmic barriers, and reductions across a diverse spectrum of computational problems within P. Recent advances have clarified both the technical boundaries imposed by the OVC and its connections to conjectured hardness for an equivalence class of related problems, circuit lower bounds, and average-case variants.

1. Formal Definition and Statement of the OVC

The Orthogonal Vectors problem (OV) is defined as follows: given two sets $A,B \subseteq \{0,1\}^d$, each of size $n$, the task is to determine whether there exists a pair $(a,b) \in A \times B$ such that the dot product $\langle a, b \rangle = 0$. An equivalent formulation considers $n$ vectors $v_1,\ldots,v_n \in \{0,1\}^d$ and asks if $\exists\,i \neq j$ with $\langle v_i, v_j \rangle = 0$.

The Orthogonal Vectors Conjecture asserts that for every constant $\epsilon > 0$, there is a constant $c = c(\epsilon)$ so that OV with $n$0 requires $n$1 time, i.e., there does not exist an $n$2-time algorithm for such dimensions. In the regime $n$3, the trivial quadratic algorithm is believed to be essentially optimal, up to $n$4-factor improvements (Kane et al., 2017, Chen et al., 2018, Nederlof et al., 2020).

For the $n$5-Orthogonal Vectors (k-OV) problem, which generalizes the search to $n$6 sets of vectors, the corresponding conjecture posits the absence of $n$7-time algorithms for $n$8 (Kühnemann et al., 30 Apr 2025).

2. Central Role in Fine-Grained Complexity and Equivalence Classes

The OVC underpins conditional lower bounds for numerous P-time problems. It serves as the canonical hardness assumption for fine-grained reductions, similar in spirit to NP-completeness in classical complexity. Dozens of tight $n$9 lower bounds hinge on OVC, with prominent reductions to batch Partial Match, batch Subset Queries, Hamming Nearest Neighbors, and exact or approximate versions of vector inner product problems (Chen et al., 2018).

Recent work establishes that OV in $(a,b) \in A \times B$0 dimensions is truly-subquadratic equivalent (under fine-grained reductions) to the following problems (Chen et al., 2018):

Problem	Definition (brief)	Complexity barrier (under OVC)
OV	Exists $(a,b) \in A \times B$1 with $(a,b) \in A \times B$2?	$(a,b) \in A \times B$3
Max-IP / Min-IP	Max/min inner product over pairs	$(a,b) \in A \times B$4
Exact-IP	Exists $(a,b) \in A \times B$5 with $(a,b) \in A \times B$6?	$(a,b) \in A \times B$7
Approximate Max-IP / Min-IP	Constant-factor approximation (e.g. $(a,b) \in A \times B$8)	$(a,b) \in A \times B$9
$\langle a, b \rangle = 0$0-Closest/Furthest Pair	$\langle a, b \rangle = 0$1-approximation	$\langle a, b \rangle = 0$2
Data structure (Partial Match/NNS)	Polynomial space, $\langle a, b \rangle = 0$3 query time	Polylogarithmic blowup in dimension

Any breakthrough (even mild acceleration) on one member cascades to the entire equivalence class.

3. Circuit Complexity, Algorithmic Techniques, and Lower Bounds

The OVC has been validated in several computational models. For branching programs, AC$\langle a, b \rangle = 0$4 formulas, constant-fan-in formulas, and formulas with symmetric gates (unbounded fan-in), the best known lower bounds almost match trivial quadratic upper bounds up to polylogarithmic factors (Kane et al., 2017). Specifically, for $\langle a, b \rangle = 0$5 vectors in $\langle a, b \rangle = 0$6:

• Branching program complexity: $\langle a, b \rangle = 0$7
• Boolean formula complexity (fan-in $\langle a, b \rangle = 0$8): $\langle a, b \rangle = 0$9
• Symmetric function wire complexity: $n$0

These results are established via input restriction methods distinct from classical random restrictions, leveraging hard functions on the "middle layer" and a shrink-and-count strategy.

Crucially, these barriers delineate what is achievable by current circuit and communication-cover paradigms. For example, no OR-covering or rank-1 linear-circuit techniques can yield OV algorithms faster than $n$1 (decision) or $n$2 (counting), and pushing below these rates would require fundamentally new circuit cancellation techniques (Alman et al., 17 Sep 2025, Nederlof et al., 2020).

4. State-of-the-Art Algorithmic Barriers and Their Significance

Recent advances provide nearly optimal deterministic and randomized algorithms for OV in moderate dimension but fail to break the OVC barrier. For $n$3 and $n$4 vectors, Alman and Li construct depth-2 linear circuits for the OV-related disjointness matrix with degree about $n$5 for the decision problem and $n$6 for the counting version, yielding:

• Decision: $n$7
• Counting: $n$8

These results derandomize previous approaches and, through Strassen's asymptotic spectrum duality, show that current circuit-based techniques are essentially optimal in this regime (Alman et al., 17 Sep 2025). For denser regimes where $n$9, new approaches are required to achieve truly subquadratic time.

For the specific structured case where all vectors have Hamming weight $v_1,\ldots,v_n \in \{0,1\}^d$0, specialized algorithms achieve $v_1,\ldots,v_n \in \{0,1\}^d$1 time for $v_1,\ldots,v_n \in \{0,1\}^d$2 input vectors, still far from subquadratic for superlogarithmic $v_1,\ldots,v_n \in \{0,1\}^d$3 (Nederlof et al., 2020).

5. Reductions, Conditional Hardness, and Consequences of Refuting OVC

The credibility of the OVC is reinforced by its position in the fine-grained reductions web. Falsifying the OVC would yield breakthrough polynomial-factor speedups for other problems previously believed intractable to such improvements. For example, if moderate-dimension OVC fails (i.e., there exists an $v_1,\ldots,v_n \in \{0,1\}^d$4 algorithm for OV with $v_1,\ldots,v_n \in \{0,1\}^d$5 for some $v_1,\ldots,v_n \in \{0,1\}^d$6), then:

• There exist $v_1,\ldots,v_n \in \{0,1\}^d$7 and fixed $v_1,\ldots,v_n \in \{0,1\}^d$8 such that Zero-Weight-$v_1,\ldots,v_n \in \{0,1\}^d$9-Clique and Min-Weight-$\exists\,i \neq j$0-Clique in $\exists\,i \neq j$1-hypergraphs with $\exists\,i \neq j$2 vertices can be solved in $\exists\,i \neq j$3 randomized time.
• The satisfiability of sparse $\exists\,i \neq j$4 circuits (threshold circuits of at most $\exists\,i \neq j$5 wires and depth $\exists\,i \neq j$6) with $\exists\,i \neq j$7 input variables can be computed in $\exists\,i \neq j$8 time.

Both consequences represent polynomial improvements on longstanding brute force bounds, and so their absence is interpreted as strong evidence for the OVC (Abboud et al., 2018). The OVC is also implied by the Weighted Clique conjecture.

The reduction from Subset Sum to OV, and the tight connection between improving OV algorithms and breaking the Schroeppel-Shamir $\exists\,i \neq j$9 barrier for Subset Sum, further entangles OVC with exponential-time complexities in NP-hard problem domains (Nederlof et al., 2020).

6. Average-Case Variants, Planted Hardness, and Cryptographic Implications

A recent direction is the study of average-case hardness via planted solutions. In the $\langle v_i, v_j \rangle = 0$0-OV variant, a distribution over planted instances is defined where the planted solution is unique and all $\langle v_i, v_j \rangle = 0$1-marginals coincide with the model distribution (i.i.d. $\langle v_i, v_j \rangle = 0$2-biased vectors). The conjecture is that any $\langle v_i, v_j \rangle = 0$3-time algorithm cannot distinguish model from planted instances with any nontrivial advantage, nor efficiently recover the planted solution, in superlogarithmic dimension (Kühnemann et al., 30 Apr 2025). Importantly:

• The construction ensures $\langle v_i, v_j \rangle = 0$4-wise independence across marginals, preventing known attack strategies.
• Efficient search-to-decision reductions show that average-case distinguishing is as hard as recovering the solution.

This property suggests planted $\langle v_i, v_j \rangle = 0$5-OV as a potential foundation for fine-grained cryptographic primitives due to its resistance to circuit lower bound-based attacks. However, no unconditional lower bounds are currently known for these average-case instances.

7. Open Problems, Barriers, and Research Directions

Several open avenues remain:

• New algorithms for OV in the dense regime: All current “rectangle/circuit” or rebalancing methods are known to be optimal up to polylogarithmic factors for moderate $\langle v_i, v_j \rangle = 0$6; fundamentally new algorithmic or circuit-theoretic paradigms are required for meaningful improvements (Alman et al., 17 Sep 2025).
• Derandomization and generalization: Further progress may involve derandomizing communication-cover constructions or extending techniques to broader input support patterns (Nederlof et al., 2020).
• Connecting planted and worst-case complexity: The validity of planted OVC instances as average-case hard problems connects fine-grained complexity to cryptographic hardness (Kühnemann et al., 30 Apr 2025).
• Fine-grained barriers in P: The OVC, alongside SETH, is central to understanding and conjecturing the quadratic and exponential time barriers for broad classes of polynomial-time problems.

Continued research explores derandomization, transferability to geometric and data-structure settings, and the possibility of leveraging OVC to establish cryptographic hardness in regimes where strong worst-case complexity assumptions are plausible.