General · EdgepediaOctober 10, 20267 min read

Anatoly Karatsuba

Anatoly Alexeevich Karatsuba (Анатолий Алексеевич Карацуба) was a Soviet and Russian mathematician at the Steklov Institute of Mathematics who discovered the first fast multiplication algorithm, disproving a conjecture of Andrey Kolmogorov that schoolbook multiplication was optimal, and who made major contributions to analytic number theory, including the density of zeros of the Riemann zeta function on short intervals of the critical line.¹^•² The 1960 invention of what is now called Karatsuba multiplication founded the theory of fast algorithms, and the algorithm is used for arbitrary-precision arithmetic in the GMP library today.³^•⁴

Key fact	Detail
Discovery	In autumn 1960, within a week of Kolmogorov stating his n² conjecture at a Moscow State University seminar, Karatsuba found a multiplication method with complexity O(n^log₂3), log₂3 ≈ 1.5849¹
Core identity	AB = A₁B₁·2^(2n) + (A₁B₁ + A₀B₀ − (A₀−A₁)(B₀−B₁))·2^n + A₀B₀, needing three half-size multiplications instead of four⁵
Publication	A. Karatsuba and Yu. Ofman, "Multiplication of Multiplace Numbers on Automata", Doklady Akad. Nauk SSSR 145, No. 2, pp. 293–294, submitted 13 February 1962¹
Practical range	Karatsuba is the fastest algorithm in GMP for operands of roughly 300–10,000 bits; the GMP threshold can be as low as 10 limbs⁶^•⁴
Zeta result	For H = T^(27/82+ε), the interval (T, T+H) contains at least cH·ln T zeros of odd order of ζ(1/2+it)⁷
Career	Head of the Department of Number Theory, Steklov Institute, from 1983; supervisor of 15 Ph.D. students²
Awards	P. L. Chebyshev award of the AN USSR (1981), I. M. Vinogradov award of the RAS (2001), Meritorious Science Worker of Russia (1999)²

Life and career

Karatsuba graduated from the Faculty of Mechanics and Mathematics of Moscow State University in 1959. His 1962 Ph.D. thesis, "Rational trigonometric sums of a special form and their applications", was written under N. M. Korobov, and his 1966 Habilitation thesis was "Method of trigonometric sums and the theorems on the mean value".² He worked at MSU from February 1962, and from 1966 his main place of employment was the Steklov Institute of Mathematics (MIAN) of the Russian Academy of Sciences, where he became head of the Department of Number Theory in 1983.²

He supervised 15 Ph.D. students, 7 of whom later obtained D.Sci. Habilitation degrees; the Steklov obituary names Grigory Kolesnik, Sergei Voronin, Gennady Arkhipov, Vladimir Chubarikov, and Maxim Korolev among them.²^•⁸ The French popular-science journal Pour la Science in 2000 called Karatsuba multiplication "one of the most useful results in mathematics".⁸

The Karatsuba multiplication algorithm and the Kolmogorov conjecture

The refutation. Karatsuba, then a 23-year-old student in the audience, set out to prove the conjecture; within a week he found that the algorithm he hoped would yield a lower bound instead gave an upper estimate of order n^log₂3, with log₂3 = 1.5849..., disproving it.¹^•⁹ Kolmogorov was greatly agitated by the refutation, reported Karatsuba's method himself at the next seminar meeting, and at that point the seminar was terminated; the Steklov obituary records the same episode as Kolmogorov closing his seminar in surprise.¹^•⁸

The identity. Split each 2n-digit factor into halves A = A₁·2^n + A₀ and B = B₁·2^n + B₀. The schoolbook method needs four half-size products; Karatsuba's formula

$AB = A_1B_1 \cdot 2^{2n} + \bigl(A_1B_1 + A_0B_0 - (A_0-A_1)(B_0-B_1)\bigr) \cdot 2^{n} + A_0B_0$

needs only three, plus additions and subtractions, giving the recurrence K(2n) ≤ 3K(n) + O(n).⁵ Solving the recurrence gives the exponent log₂3 ≈ 1.585 instead of 2. The method was later called "Divide and Conquer", also known as "Binary Splitting" and the "Dichotomy Principle", and its appearance was the starting point of the theory of fast computations.³

An unusual publication. The result appeared as A. Karatsuba and Yu. Ofman, "Multiplication of Multiplace Numbers on Automata", Doklady Akad. Nauk SSSR vol. 145, No. 2, pp. 293–294, submitted 13 February 1962; Karatsuba wrote that he learned of the article only when he was given its reprints.¹

Later multiplication methods and real-world implementations

The line Karatsuba opened continued quickly. A. L. Toom proposed the first generalization in 1963, dividing each multiplier into r blocks, and Cook refined it; the survey notes that the Schönhage–Strassen-type complexity estimate was previously obtained by Karatsuba but not published.⁵ In 1971 Schönhage and Strassen constructed an algorithm with M(n) = O(n log n log log n), based on the fast Fourier transform, the best known bound at the time of Karatsuba's 1995 memoir.¹ As of 2024 the best known asymptotic upper bound is Harvey and van der Hoeven's O(n log n), which superseded Fürer's 2007 bound; the historical sequence runs from long multiplication O(n²) to Karatsuba 1962 at O(n^(log 3/log 2)) ≈ n^1.58, Knuth 1969, Schönhage–Strassen 1971, and Fürer 2007.¹⁰

Where Karatsuba still wins. Asymptotic complexity alone does not determine real-world speed. Per GMP documentation and 2025 benchmarks, Karatsuba is the fastest practical algorithm for operands of roughly 300–10,000 bits, Toom-3 for 10,000–100,000 bits, and Schönhage–Strassen above about 10^6·log n bits; the Harvey–van der Hoeven O(n log n) algorithm is classified as "Galactic", asymptotically best but not practical on current hardware.⁶ In GMP, Karatsuba multiplication is asymptotically O(N^1.585), the exponent being log(3)/log(2), representing 3 multiplies each 1/2 the size of the inputs; the Karatsuba threshold MUL_TOOM22_THRESHOLD can be as little as 10 limbs, and the squaring threshold is usually about twice the multiplication threshold.⁴ GMP computes the term (x₁−x₀)·(y₁−y₀) as an absolute value and uses the sign to choose to add or subtract.⁴

Work in analytic number theory

Karatsuba's doctoral work was in the theory of trigonometric sums, and his research areas included the Selberg problem, zeros of linear combinations of Dirichlet L-series, distribution of zeros of the Riemann zeta function on short intervals of the critical line, and the multidimensional Dirichlet divisor problem.²

Zeros on short intervals. His best-known zeta theorem states that for any ε > 0 there exists c = c(ε) > 0 such that for T ≥ T₀(ε) and H = T^(27/82+ε), the interval (T, T+H) contains at least cH·ln T zeros of odd order of ζ(1/2+it), where N₀(T) counts odd-order zeros in (0, T).⁷ In 1985 he also published a survey in Russian Mathematical Surveys covering estimates for exponential sums and theorems on zeros of the zeta function, including the effect of "closing in" on zeros as T grows.¹¹ His textbook Basic Analytic Number Theory, with Melvyn B. Nathanson, was published by Springer and includes a chapter on the density of zeros of the zeta function and the distribution of primes in short intervals.¹²

The p-adic method and exponential sums. He built a p-adic method in the theory of trigonometric sums, leading to new bounds for Dirichlet L-series and a new p-adic proof of Vinogradov's mean value theorem.⁸ Maxim Korolev's 2017 overview of Karatsuba's work from the early 1990s to 2008 lists the prime number theorem, the divisor problem, exponential sums, "Karatsuba's phenomenon", modular hyperbolas, Dirichlet characters, fractional moments, and Kloosterman sums among his subjects.¹³

Complexity theory and pseudorandomness

The appearance of Divide and Conquer was, in Karatsuba's own account, the starting point of the theory of fast computations, and he devoted a 1995 Steklov Institute memoir, "The Complexity of Computations", to the field.³^•¹ A later line of research constructs pseudorandom number generators from zeta-function-based one-way functions, relying on the theorem that a one-to-one one-way function implies the existence of a pseudorandom generator in the sense of Blum–Micali and Yao.¹⁴

By the numbers

n² versus n^1.585. Schoolbook multiplication of two n-digit numbers takes about n² digit operations; Karatsuba's recurrence K(2n) ≤ 3K(n) + O(n) reduces the exponent to log₂3 ≈ 1.5849.³^•⁵
Three half-size multiplications. The saving comes from computing two half-size products fewer per split: 3 instead of 4.⁵
GMP threshold. Karatsuba switches on at MUL_TOOM22_THRESHOLD, which can be as little as 10 limbs (a limb is one machine word of the internal representation).⁴
Practical range. Roughly 300–10,000 bits, per 2025 benchmarks; Toom-3 takes over at 10,000 bits and Schönhage–Strassen far above.⁶
27/82. The exponent in the interval length H = T^(27/82+ε) of his zero-density theorem; the interval (T, T+H) then holds at least cH·ln T odd-order zeros.⁷
15 doctoral students, 7 of whom obtained D.Sci. degrees.²

References

Topic: Encyclopedia › Physical world and mathematics › Physical and mathematical scientists › Mathematicians and statisticians › Number theorists › Analytic number theorists

Initially written Oct 10, 2026 · Edited: Oct 11, 2026; Oct 15, 2026

Notice something wrong?