Nearly optimal lattice simulation by product formulas

I. NONRIGOROUS ERROR ANALYSIS BY BAKER-CAMPBELL-HAUSDORFF FORMULA

• Even-odd grouping of nearest-neighbor terms gives a first-order product formula e^-itHeven e^-itHodd for approximating evolution under H.The ideal evolution is e^-itH.Source: §I · p. 8
• Jordan, Lee, and Preskill analyzed product-formula error scaling by applying the BCH formula, then expanding the Taylor series and ignoring all higher-order terms, obtaining the desired first-order product-formula error scaling [1, 2, Eq.(3)].Source: §I · p. 8
• For a pth-order Taylor term, the nested-commutator norm scales as O(nt^p) with constant nesting depth, but as O(n^pt^p) beyond n layers.The latter dependence is superlinear in n and therefore does not yield the desired error scaling [2, Eq.(3)]; see also [3, Appendix B].Source: §I · p. 8
• Local error analysis instead rigorously bounds the lowest-order error, with the triangle inequality forming part of the argument.The section notes that analogous, more complicated error expressions extend to higher-order product formulas, discussed in Sections II–IV.Source: §I · p. 8

II. CANONICAL PRODUCT FORMULAS AND ORDER CONDITIONS

• A canonical formula for H = A + B is built from s stages e^(-itb_jB)e^(-ita_jA), with real coefficients bounded in magnitude by u.The parameters s, p, and u specify the number of stages, order, and coefficient bound, respectively.Source: §II · p. 9
• Lie-Trotter and Suzuki constructions can be expressed canonically, and the (2k)th-order Suzuki formula has s ≤ 2·5^(k−1), p = 2k, and u = 1 [2, Eq.(2)].Identity padding accommodates stage orderings that do not initially match the canonical form.Source: §II · p. 9
• Order p is equivalent to an error scaling of O(t^(p+1)) and to the vanishing of derivatives through order p at t = 0.The section applies this equivalence to the product formula's difference from the exact evolution.Source: §II · pp. 9–10
• For H=A+B with Hermitian A and B, Theorem 1 says four conditions are equivalent for canonical product formula S(t), including smooth operator-valued R(t) and T(t) whose derivatives at zero vanish through order p−1.Source: §II · p. 11

III. SIMPLIFIED LOCAL ERROR REPRESENTATION

• Direct Taylor expansion of R(t), or of T(t) = S(t)†R(t), gives the correct scaling in t but not the correct dependence on n.The product-formula error first has an integral representation involving R(t).Source: §III · p. 11
• Unitary conjugation rewrites the integrand as a linear combination of nested commutators with unitary conjugations, whose summand count and nesting depth do not depend on n or t.The construction computes R(t) = d/dt S(t) − (−iH)S(t), then seeks T(t) such that R(t) = S(t)T(t).Source: §III · pp. 11–12
• The representation builds on the canonical form S(t) = S_s(t) ··· S_1(t), with stages S_j(t) = e^(−itb_jB)e^(−ita_jA) and real coefficients a_j and b_j.Here the ideal evolution is E(t) = e^(−it(A+B)).Source: §III · pp. 11, 13
• For an s-stage canonical formula for H = A + B, Theorem 2 gives an integral representation of the product-formula error and a further result when the formula is pth order.The supplied theorem passages omit the displayed formulas, so their specific bounds are not stated here.Source: §III · p. 13
• The authors contrast their unitary-conjugation derivation with Descombes and Thalhammer [5], whose calculation overlooked a time-dependent term and used recursively defined integrals with hard-to-unravel combinatorial structure.They use Theorem 2 in their subsequent product-formula analysis.Source: §III · p. 13

• The analysis introduces adjoint mappings in Section IV A and uses them in Section IV B to bound product-formula error.Source: §IV · p. 13

A. Adjoint mappings

• Ad denotes conjugation by an invertible matrix, while ad denotes the commutator transformation for an arbitrary operator.Source: A. Adjoint mappings · p. 13
• A differentiation rule for Ad and ad is introduced to support Taylor expansions of multivariate functions.Source: A. Adjoint mappings · p. 14
• A higher-order differentiation rule is stated for an operator-valued function with infinitely many derivatives.Source: A. Adjoint mappings · p. 14
• The abbreviated adjoint representation omits operator identities and retains only time variables, with its own differentiation rule.Source: A. Adjoint mappings · p. 14

• For H = A + B, the canonical s-stage product formula has real coefficients and an error with an integral representation.Source: B. Error analysis of the pth-order… · p. 15
• Because terms of order p − 1 or less vanish, the error scaling follows by bounding the order-p integral Taylor remainder.Source: B. Error analysis of the pth-order… · p. 15
• O(nt^(p+1)) bounds the product-formula error for lattice Hamiltonians when s, p, and u are constant.Source: B. Error analysis of the pth-order… · p. 15
• Locality bounds the nested commutators: each ad contributes a factor 2u, while support grows by two qubits per Ad and remains at most 4k − 2.Source: B. Error analysis of the pth-order… · p. 16

Section V A

• Section V A introduces canonical formulas for time-dependent Hamiltonian simulation and studies their order conditions.Source: §V · p. 16

Section V B

• Section V B analyzes the local-error structure of time-dependent product formulas.Source: §V · p. 16

A. Time-dependent canonical formulas and order conditions

• Assuming infinitely differentiable H(t), the time-ordered evolution is differentiable and its derivatives are described using recursively defined operators T_j(t).The paper notes that infinite differentiability may be relaxed, but adopts it for a cleaner presentation.Source: A. Time-dependent canonical formulas… · p. 17
• Each stage applies A(t) and B(t) at sampled times tα_k and tβ_k, defining a time-dependent canonical product formula with s stages.The stages use real coefficients a_k, b_k, α_k, and β_k.Source: A. Time-dependent canonical formulas… · p. 18
• The residual R_T(t) yields an integral representation of product-formula error via variation of parameters [8, Theorem 4.9].A further operator T_T is introduced to obtain the correct error scaling in n and t.Source: A. Time-dependent canonical formulas… · pp. 18–19
• The order conditions are equivalent to matching evolution derivatives at t=0 or to vanishing derivatives of residual functions through order p−1.Theorem 3 states four equivalent conditions; the supplied passages specify derivative matching and residual-function conditions.Source: A. Time-dependent canonical formulas… · pp. 19–20

B. Time-dependent local error representation

• The representation assumes infinitely differentiable A(t) and B(t), but the smoothness condition can be relaxed using techniques from.Source: B. Time-dependent local error… · p. 20
• To expose the error’s dependence on the operators, the derivation conjugates the residual R_T(t) by S_T(t) and expands T_T(t) = S_T(t)†R_T(t).A direct Taylor expansion of R_T(t) gives the time scaling but does not readily show the correct dependence on n.Source: B. Time-dependent local error… · p. 20
• Theorem 4 represents the error of a time-dependent canonical s-stage product formula as an integral for H(t) = A(t) + B(t).Its stages use real coefficients and evaluate A and B at time-scaled arguments; the theorem also states a further result when the formula is pth order.Source: B. Time-dependent local error… · pp. 21–22
• The pth-order result follows from the order conditions and Taylor’s theorem with integral remainder, as in Theorem 2.Source: B. Time-dependent local error… · p. 22

VI. EMPIRICAL PERFORMANCE

• The comparison uses a one-dimensional nearest-neighbor Heisenberg model with uniformly random magnetic-field coefficients, accuracy ϵ = 10^-3, and simulation time t = n.The study uses open boundary conditions and normalizes the Hamiltonian by 4, simulating for time 4t.Source: §VI · p. 22
• HHKL uses blocks of 7 and 14 qubits, each simulated by a fourth-order product formula; the pure fourth-order formula serves as the comparison.The block size is fixed at ℓ = 7, and the pure fourth-order formula is selected for a fair comparison.Source: §VI · pp. 22–23
• Both algorithms are ancilla-free, but pure product formulas have a better constant prefactor while HHKL has better asymptotic scaling.HHKL adds negative Hamiltonian terms to compensate for Lieb-Robinson decomposition error, whereas pure product formulas simulate the original lattice Hamiltonian without that overhead.Source: §VI · pp. 22–23
• For first-order product formulas, even-odd and X-Y-Z orderings are consistent with r = O(nt^2) = O(n^3), but even-odd performs better empirically.The comparison uses the even-odd pattern and the X-Y-Z pattern of.Source: §VI · p. 23, Figure 2 · p. 24