Source-linked AI summary

Efficient Clifford+T approximation of single-qubit operators

Peter Selinger

arXiv:1212.6253v2quant-ph

TL;DR

Approximating arbitrary single-qubit operators with a discrete Clifford+T set requires efficient algorithms with near-optimal gate scaling. The paper gives a randomized, ancilla-free algorithm whose expected runtime is polynomial in log(1/ε), achieving linear-logarithmic T-counts and near-optimal z-rotation counts, while relying on a prime-distribution hypothesis and remaining specialized to Clifford+T.

  • Problem

    The paper addresses efficient approximation of arbitrary single-qubit operators using Clifford+T without ancillas, extending beyond prior approaches that used ancillas or had worse asymptotic scaling.

  • Method

    The paper presents an efficient randomized algorithm that approximates arbitrary SU(2) elements with products of Clifford+T operators.

  • Results

    For z-rotations the T-count is K + 4 log2(1/ε), for arbitrary SU(2) it is K + 12 log2(1/ε), and the exponent improves from Solovay–Kitaev’s c approximately 3.97 to c = 1.

  • Takeaways & Limitations

    The algorithm provides asymptotically optimal logarithmic gate-complexity scaling and is within an additive constant of optimal for certain z-rotations.

  • Takeaways & Limitations

    Expected polynomial runtime depends on an unproven prime-distribution hypothesis, and the algorithm is specialized to Clifford+T; the z-rotation lower bound assumes single-qubit Clifford+T synthesis without auxiliary techniques.

Abstract

from arXiv · show

We give an efficient randomized algorithm for approximating an arbitrary element of $SU(2)$ by a product of Clifford+$T$ operators, up to any given error threshold $ε>0$. Under a mild hypothesis on the distribution of primes, the algorithm's expected runtime is polynomial in $\log(1/ε)$. If the operator to be approximated is a $z$-rotation, the resulting gate sequence has $T$-count $K+4\log_2(1/ε)$, where $K$ is approximately equal to $10$. We also prove a worst-case lower bound of $K+4\log_2(1/ε)$, where $K=-9$, so that our algorithm is within an additive constant of optimal for certain $z$-rotations. For an arbitrary member of $SU(2)$, we achieve approximations with $T$-count $K+12\log_2(1/ε)$. By contrast, the Solovay-Kitaev algorithm achieves $T$-count $O(\log^c(1/ε))$, where $c$ is approximately $3.97$.

1 Introduction

The paper develops an efficient randomized, ancilla-free Clifford+T approximation algorithm for single-qubit operators, improving on Solovay–Kitaev scaling and approaching optimal T-counts for z-rotations. Its expected polynomial runtime depends on an unproven hypothesis about prime distributions, and its scope is specialized to Clifford+T.

  • Problem and setting: The paper targets approximate decomposition of arbitrary single-qubit operators using the fault-tolerant Clifford+T universal gate set.The approximation is required because discrete universal gate sets cannot generally represent arbitrary operators exactly.
  • Main results: K + 4 log2(1/ε) T-count is achieved for z-rotations, with K approximately 10, matching a worst-case lower bound up to an additive constant.The lower bound has K = −9 for certain z-rotations.
  • Main results: K + 12 log2(1/ε) T-count is achieved for arbitrary members of SU(2).The algorithm applies to any requested error threshold ε > 0.
  • Comparison with prior work: The method reduces the gate-complexity exponent from Solovay–Kitaev’s c approximately 3.97 to c = 1.Solovay–Kitaev produces T-count O(log^c(1/ε)).
  • Prior work and scope: The paper addresses the open question of whether asymptotically optimal efficient approximation can be achieved without ancillas.Earlier Kliuchnikov–Maslov–Mosca results used a fixed, small number of ancillas initialized and returned to |0⟩.
  • Limitations: The number-theoretic algorithm is specialized to Clifford+T rather than arbitrary universal gate sets.The paper notes that the approach is unlikely to work for arbitrary universal gate sets.
  • Limitations: The expected polynomial runtime is contingent on an unproven hypothesis about the distribution of primes.The hypothesis is described as heuristically true but not proven to the author’s knowledge.
  • Limitations: The z-rotation lower bound applies only to products of single-qubit Clifford+T operators, since ancillas, resource states, or online measurement can yield smaller T-counts.This boundary concerns the specific synthesis model studied in the paper.

2 Overview of the algorithm

The algorithm samples algebraic candidates in a ring that characterizes exactly synthesizable Clifford+T operators, then constructs an ε-close unitary and exactly synthesizes it. The resulting T-count is controlled by the denominator exponent selected for the approximation.

  • Algebraic representation: The algorithm works with elements of the ring 2^k(aω^3 + bω^2 + cω + d), where k is natural and a, b, c, d are integers.This ring provides the algebraic representation used to construct candidate approximations.
  • Exact synthesis: A unitary is exactly synthesizable over Clifford+T without ancillas if and only if all matrix entries belong to D[ω].The required T-count is at most 2k, where k is the minimal exponent representing all entries in the relevant ring form.
  • Approximation construction: For a target z-rotation, the method chooses k and randomized u ∈ Z[ω] so that a suitable algebraic approximation can be found with relatively high probability.The construction also involves solving a Diophantine equation for the required coefficients.
  • Approximation construction: The constructed unitary satisfies ||U − R_z(θ)|| ≤ ε and has coefficients in D[ω], enabling exact Clifford+T synthesis.The exact synthesis step gives a T-count of at most 2k.

3 Some number theory

This section introduces the algebraic-number rings and operations used by the approximation algorithm. It establishes Euclidean-domain structure, divisibility and norm properties, and square criteria for units.

  • Algebraic rings: The paper works with the rings Z, Z[√2], Z[i], and Z[ω], with Z[ω] containing the others.Z[ω] is the ring of cyclotomic integers of degree 8.
  • Conjugation: Conjugation provides automorphisms of Z[ω], including complex conjugation and a second map acting nontrivially on Z[ω] and Z[√2].These maps are used to define the relevant norms and conjugates.
  • Norms: The norms N_i and N_ω are nonnegative, multiplicative, and characterize units through values ±1.The cyclotomic norm is expressed using the coefficients of an element of Z[ω].
  • Euclidean domains: These rings are Euclidean domains, so greatest common divisors can be computed by Euclid’s algorithm in O(log |N(t)|) divisions.Divisibility is independent of the chosen ring among the rings considered.
  • Units: A unit in Z[√2] is a square exactly when it and its conjugate are both nonnegative.The criterion follows from the sign characterization of the unit exponents.

4 A Diophantine equation

The paper reduces a norm equation in Z[ω] to algebraic constructions based on prime splitting and greatest common divisors. Under explicit positivity and parity conditions, it proves existence and efficient randomized computability of a solution.

  • Equation: The central task is solving a relative norm equation in Z[ω], and the paper develops the solution method in detail.The equation is connected to splitting fully split primes in Z[ω].
  • Existence theorem: Under the stated conditions on ξ and p=ξ•ξ, there exists t∈Z[ω] satisfying the equation, and an efficient randomized algorithm computes it.The conditions include x odd, y even, and nonnegative ξ and ξ•.
  • Proof: The proof eliminates the cases s†s∼1 and s†s∼ξ^2, leaving s†s∼ξ as the required case.The excluded cases contradict divisibility or norm properties.
  • Final construction: After obtaining s†s∼ξ, the argument shows the remaining unit is a square and constructs t by multiplying s by its square root.The resulting element satisfies t†t=ξ.
  • Algorithmic properties: Only the computation of h is randomized; the remaining arithmetic, including gcd, division, square-root, and multiplication steps, is deterministic.The square-root computation reduces to solving a quadratic equation over the integers.

5 Approximations in Z[

This section proves that algebraic integers in Z[√2] can be found inside prescribed regions of the two-dimensional conjugate embedding. The coverage results provide existence and efficient computation, including parity-controlled solutions.

  • Uniqueness: If the product of interval widths is below 1, a region contains at most one element of Z[√2].The result bounds the number of algebraic-integer solutions in a conjugate-coordinate rectangle.
  • Existence: If the product of interval widths is at least (1+√2)^2, every such rectangle contains an element of Z[√2].The proof defines a coverage property and constructs a solution algorithmically.
  • Geometry: The geometric interpretation places conjugate pairs on a nonuniform grid and searches for grid points inside rotated rectangles.The two coordinates correspond to an element and its conjugate.
  • Coverage proof: The proof uses symmetry, rescaling by powers of λ=1+√2, and a finite candidate check to establish coverage.The construction transforms interval widths while preserving the existence property.
  • Efficiency: The coverage construction is efficient because it uses no iterative search beyond computing a power of λ in O(|log(δ)|) steps.The stated computational effort is at most O(log^2(δ)).
  • Sharpness: The bounds are sharp, with examples attaining the uniqueness and existence thresholds.The examples show regions containing exactly two or four solutions at the relevant boundaries.
  • Parity control: A rescaled corollary guarantees solutions with either even or odd first coefficient, while retaining efficient computability.This parity control is obtained by transformations of the coverage result.

6 Approximation up to ε

The approximation problem is reduced to finding algebraic candidates inside an ε-region and then solving a norm equation for the remaining matrix entry. Geometric coverage supplies many candidates once k is large enough.

  • Approximation setup: For fixed θ and ε, the goal is to choose k, u, and t so that a Clifford+T-form matrix is unitary and ε-close to R_z(θ).Unitarity requires u†u+t†t=2^k.
  • ε-region: The ε-region has width proportional to ε^2, so approximation becomes a constrained search in the unit disk.The region is represented geometrically as a subset of the disk.
  • ε-region: The operator-norm error condition is equivalent to requiring the normalized u to lie in the ε-region R_ε.The reduction uses the dot product between the normalized u and z=e^{-iθ/2}.
  • Two-step reduction: After selecting u, the second step is to find t∈Z[ω] satisfying t†t+u†u=2^k.The remaining norm equation is handled by the Diophantine solver from the preceding section.
  • Candidates: A candidate normalized u lies in R_ε and satisfies parity and nonnegativity conditions on the associated ξ and ξ•.A prime candidate additionally requires p=ξ•ξ to be prime.
  • Candidate generation: If k≥C+2 log_2(1/ε), with C=5, the construction produces at least floor(4^k ε^2) candidates efficiently.The candidates are obtained by partitioning a parallelogram inside R_ε and applying the Z[√2] coverage results.
  • Geometric construction: The parallelogram used for candidate selection is contained within R_ε, ensuring that every selected point meets the approximation constraint.Its boundaries are defined by two dot-product lines and two horizontal lines.

7 The main algorithm

The algorithm constructs Clifford+T approximations to z-rotations and extends them to arbitrary SU(2) operators by decomposing them into three rotations. It includes randomized candidate generation, exact synthesis, and a determinant-based scope limitation for U(2).

  • Approximating a z-rotation: The algorithm combines candidate generation with Theorem 12 to search for a ring element satisfying the required norm equation.Candidates are tried until a valid solution is found; random generation is also permitted.
  • Approximating a z-rotation: A successful solution yields an ε-approximation of Rz(θ), which is converted into Clifford+T gates using exact synthesis.The sequence may optionally be reduced to Matsumoto-Amano normal form.
  • Approximating arbitrary gates: An arbitrary U ∈ SU(2) is decomposed as Rz(β) H Rz(γ) H Rz(δ), with each rotation approximated separately to error ε/3.This reduces arbitrary single-qubit approximation to three z-rotation approximations connected by Clifford gates.
  • Scope: The method directly targets SU(2); for U(2), arbitrary determinants are approximable only when the determinant is a power of ω unless arbitrary global phases are included.Otherwise, approximation is possible only up to a global phase.

8 Complexity analysis

The analysis relates denominator exponent to T-count, obtains logarithmic gate complexity, and bounds runtime under a prime-distribution hypothesis. Randomization also permits parallel execution and theoretical seeding safeguards termination.

  • Gate complexity: 2k − 3 ≤ n ≤ 2k bounds the minimal T-count n for a Clifford+T operator with denominator exponent k.This relation converts the algorithm’s denominator choice into a circuit-length bound.
  • Gate complexity: K + 4 log2(1/ε) bounds the T-count for z-rotations, while arbitrary SU(2) operators require K + 12 log2(1/ε).The constants are approximately 10.09 for z-rotations and 30.26 for arbitrary SU(2).
  • Time complexity: Under Hypothesis 29, the expected runtime is O(k^4) = O(log^4(1/ε)).The assumption states that candidate primality is at least as likely as primality among comparable odd numbers.
  • Time complexity: The randomized algorithm can be parallelized because candidates are chosen independently.Independent candidate trials provide a direct source of parallelism.
  • Seeding: Seeding can theoretically avoid unlucky candidate sets, but for z-rotations it approximately doubles the T-count to K + 8 log2(1/ε).For arbitrary gates, a different seed leaves the asymptotic T-count unchanged up to a small additive constant; in practice, seeding appears unnecessary.

9 Lower bounds

Counting arguments give a general lower bound, while a sharper construction establishes a worst-case bound for z-rotations. The z-rotation result places the algorithm within an additive constant of optimal for selected angles.

  • General lower bound: K + 3 log2(1/ε) is a general lower bound derived from counting Clifford+T circuits against the ε-ball covering number of SU(2).The argument uses the three-dimensional structure of SU(2) and the number of circuits available at T-count n.
  • General lower bound: Most operators require T-counts close to or above the general lower bound, although the bound does not automatically apply to z-rotations.Z-rotations form a one-dimensional submanifold rather than the full three-dimensional SU(2) space.
  • Z-rotation lower bound: K′ + 4 log2(1/ε), with K′ = −9, is a worst-case lower bound for ε-approximating some z-rotation.The proof constructs an angle whose ε-region excludes the relevant boundary points.
  • Scope of the lower bound: The sharper z-rotation lower bound applies only to carefully chosen worst-case operators, while achieving the typical lower-bound order remains plausible rather than proved.A proposed algorithmic variation is sketched, but its details are left for future work.

10 Overclocking

Overclocking lowers the algorithm’s parameter k to seek shorter circuits. A fixed additive reduction increases runtime only by a constant factor, whereas multiplicative overclocking remains an unresolved optimization.

  • Overclocking: Overclocking decreases the prescribed value of k, and practical estimates suggest that smaller k can still produce valid decompositions.The paper names this practice by analogy with running processors above their designed clock speed.
  • Additive overclocking: A fixed additive reduction of C increases runtime by a fixed factor without affecting validity or big-O time complexity.The success probability decreases by a factor independent of ε, while the asymptotic complexity remains unchanged.
  • Multiplicative overclocking: Multiplicative overclocking may target K + 3 log2(1/ε) for most angles, but the current algorithm cannot efficiently exploit the resulting thin ε-regions.The paper leaves refinement for sparsely populated regions to future work.

11 Experimental results

Experiments demonstrate practical high-precision decompositions, including ε = 10^-100 with 1338 T-gates, while runtime and candidate processing grow polynomially as ε decreases.

  • The experiments include approximating Rz(π/128) for several ε values on one 3.40GHz Intel i5-3570 CPU core.
  • Runtime per candidate increases polynomially as ε decreases because the algorithm processes larger integers and real numbers.
  • The number of candidates tried per run is approximately proportional to k, consistent with the prediction in equation (34).This provides empirical evidence supporting Hypothesis 29.
  • ε = 10^-100 was achieved using only 1338 T-gates, outperforming Solovay-Kitaev at comparable T-count.Solovay-Kitaev with comparable T-count achieves less than ε = 10^-10.

12 Conclusions

The paper presents an efficient Clifford+T decomposition algorithm that performs well theoretically and practically, producing shorter circuits than the comparable ancilla-based approach by a constant factor.

  • The algorithm is reported as the most efficient to date for decomposing SU(2) elements into Clifford+T gates at arbitrarily small ε.
  • ε = 10^-100 decompositions use fewer than 1400 T-counts in practice.
  • The resulting decompositions are shorter by a constant, non-negligible factor than those from the ancilla-based Kliuchnikov–Maslov–Mosca algorithm.Both algorithms have the same optimal asymptotic big-O complexity and are based on solving a Diophantine equation.
Loading 1212.6253v2…