Breaking Barriers: The First Truly Subquadratic 3SUM Algorithm
This presentation reveals a breakthrough in fine-grained complexity theory: a new algorithm for sparse matrix products that refutes long-standing algorithmic conjectures. By exploiting the hidden structure of lopsided sparse triangle problems, the authors achieve the first polynomial speedup for integer 3SUM and All-Pairs Shortest Paths, demonstrating that reductions used to prove hardness can also propagate algorithmic improvements across an entire problem class.Script
For decades, researchers believed that finding three numbers that sum to zero in a list of integers required quadratic time. This paper shatters that assumption with the first truly subquadratic algorithm for 3SUM on polynomially bounded integers.
The key insight is that many hard problems reduce to a lopsided triangle structure: imagine a graph with two large vertex sets and a very thin middle layer. When you only care about triangles touching certain pairs from the outer sets, you can exploit the middle layer's thinness in ways that balanced instances never allow.
The algorithm adapts a recursive multiplication identity due to Schönhage. Each output entry depends on many leaves in the recursion tree, but here is where the magic happens: high-order leaves are shared across thousands of different outputs. By computing only the union of all needed leaves rather than recomputing for each output independently, the algorithm evaluates fewer than one leaf per entry.
Through known reductions, the lopsided triangle speedup propagates to the headline problems. Exact Triangle improves to n to the 2.9983, which cascades to 3SUM at n to the 1.9992 and All-Pairs Shortest Paths at n to the 2.9995. Each reduction preserves only a fraction of the original saving, making the reduction network both powerful and lossy.
The method works only when the middle dimension is very thin, specifically when D is at most N to the 0.1204. This threshold comes from balancing the cost of encoding inputs against the savings from leaf sharing, and it represents a limitation of the construction rather than a fundamental barrier. Balanced triangle instances and most dynamic matrix problems remain untouched.
This result refutes the integer 3SUM and APSP hypotheses in their standard form, but many related conjectures remain standing. The Strong Exponential Time Hypothesis, Orthogonal Vectors, and higher-order SUM problems lack the specific structure exploited here. Whether similar algebraic techniques can crack these barriers is wide open. To dive deeper into this breakthrough and create your own video explanations of cutting-edge research, visit EmergentMind.com.
