betaFeedback

WorksPhysics and Astronomy

Fer expansion converges almost twice as far as Magnus for qubits

Abstract

The Fer expansion writes the propagator of a time-dependent Hamiltonian as a product of unitary exponentials whose error decreases doubly exponentially with the number of factors, while the Magnus expansion improves by one order at a time. Its convergence, however, was guaranteed only below a radius of about 2.60582.6058, smaller than the Magnus radius ππ. We establish a guaranteed convergence threshold of about 6.21066.2106, or 1.98 π1.98\,π, for a qubit. This is almost twice the optimal universal Magnus threshold ππ; the bound on a single Fer step is sharp, while the exact convergence radius remains open.

Introduction

Time-dependent Hamiltonians are of fundamental relevance in quantum physics. A driven qubit is the basic model of magnetic resonance and of quantum computing [1, 2, 3], and simulating time-dependent Hamiltonian dynamics is a central task for quantum algorithms [4, 5, 6, 7]. More generally, evolution with respect to parameters other than time also arises in a variety of quantum settings [8, 9, 10]. For a time-dependent Hamiltonian H(t)H(t) this dynamics is given formally by the time-ordered exponential U(t)=Texp⁡[−i∫0tds H(s)]U(t)=\mathcal{T}\exp[-i\int_0^t ds\,H(s)], and expanding it in powers of HH gives the Dyson series [11]. The solution is only formal, since every truncation of the series is in general not unitary. How to reorganize it so that unitarity is preserved at every order is an old question, with two standard answers.

The first answer collects the whole Dyson series into one exponential, U(t)=eΩ(t)U(t)=e^{\Omega(t)}, where the anti-Hermitian operator Ω(t)\Omega(t) is a series of nested commutators of HH at different times. This is the Magnus expansion [12, 13, 14]. Each truncation is unitary by construction, but convergence of the series is not guaranteed for arbitrary times. Convergence is guaranteed when ∫0tds ∥H(s)∥<π\int_0^t ds\,\lVert H(s)\rVert<\pi [15, 16], where ∥⋅∥\lVert\cdot\rVert denotes the operator norm, and π\pi is the best possible constant even for a qubit [17].

The second answer leaves the propagator as an ordered product of unitary exponentials. Examples include the Trotter–Suzuki formulas, built from compositions of exponentials and widely used in digital quantum simulation [18, 19, 20], and the Zassenhaus expansion [21] and its continuous version due to Wilcox [22, 23]. In the latter expansions, the jj-th exponent is of order jj in the Hamiltonian, and successive factors add one order of accuracy at a time.

The Fer expansion [24, 25] lies between these two answers. As in the Magnus expansion, only commutators of HH appear; as in the product formulas, the propagator is an ordered product. It is not, however, an expansion in powers of HH. It extracts one exponential at a time: the exponent is the time integral of an effective Hamiltonian, and what remains obeys a genuine Schrödinger equation, with a new effective Hamiltonian built from the previous one. The new Hamiltonian is of second order in the previous one, so that each new factor roughly doubles the order in time to which the product reproduces the propagator, instead of adding one [26, 27]. The expansion was rediscovered in numerical analysis by Iserles [26] and later recognized as a Lie-group method [27, 28]. It is used to build effective Hamiltonians in solid-state nuclear magnetic resonance [29, 30], in the same spirit as average-Hamiltonian theory [31, 32] and dynamical decoupling [33], and to construct geometric integrators [28, 34, 35].

Figure 1. One step of the Fer recursion as a map Kj(t)↦Kj+1(t)K_j(t)\mapsto K_{j+1}(t). The classical map Ψ1(x)\Psi_1(x) (red, dashed) and the finite-angle map Ψ2(x)\Psi_2(x) (blue, dashed) of Eq. (10) hold for Hermitian Hamiltonians of any dimension; the map Ψqubit(x)\Psi_{\rm qubit}(x) of Eq. (56) (violet) is the exact worst case of one step for a qubit. The guaranteed convergence thresholds are the positive fixed points, where the maps cross the diagonal (open circles).

This fast convergence, however, is guaranteed only within a finite region. For a Hermitian Hamiltonian, the product was proved to converge when ∫0tds ∥H(s)∥<2\int_0^t ds\,\lVert H(s)\rVert<2 [25], and this value was recently raised to about 2.60582.6058 [36]. Both values are still smaller than the Magnus radius π\pi. They come from the same argument, which follows a single number, the time integral Kj(t)K_j(t) of the norm of the jj-th effective Hamiltonian Hj(t)H_j(t), and bounds Kj+1(t)K_{j+1}(t) in terms of Kj(t)K_j(t) by combining two inequalities. The first bounds the norm of Hj+1(t)H_{j+1}(t) by the norm of Hj(t)H_j(t) times a factor that increases with the norm of the time integral of Hj(t)H_j(t). The second bounds the norm of this time integral by Kj(t)K_j(t). The first inequality rests on an estimate for a single unitary conjugation that cannot be improved [37]. For a qubit, however, the two inequalities are never tight at the same time. The norm of the time integral of Hj(t)H_j(t) grows as fast as Kj(t)K_j(t) only when Hj(t)H_j(t) is proportional to it, and then Hj(t)H_j(t) commutes with its time integral and gives no contribution to Hj+1(t)H_{j+1}(t). A better radius thus requires keeping track of the correlation between an effective Hamiltonian and its own time integral.

In this work we take this correlation into account exactly for a qubit, and show that the guaranteed threshold becomes more than twice as large as the general value of about 2.60582.6058. For a qubit, the effective Hamiltonians and their time integrals can be written as vectors of R3\mathbb{R}^3, and each step of the recursion is an explicit map of these vectors [38]. The size of Hj+1(t)H_{j+1}(t) turns out to be the size of the component of Hj(t)H_j(t) orthogonal to its time integral, multiplied by a weight that depends only on the norm of the time integral. The time integral draws a curve of length Kj(t)K_j(t) that starts at the origin, and the worst case of one step becomes a maximization over the shape of this curve, which we solve exactly. It gives the sufficient condition for convergence

∫0tds λmax⁡(s)−λmin⁡(s)2<ρqubit≈6.2106(1)\int_0^t \mathrm{d}s\,\frac{\lambda_{\max}(s)-\lambda_{\min}(s)}{2}<\rho_{\mathrm{qubit}}\approx6.2106 \tag*{(1)}

for every qubit Hamiltonian, with λmax⁡(s)\lambda_{\max}(s) and λmin⁡(s)\lambda_{\min}(s) its two eigenvalues. For a traceless Hamiltonian the integrand is ∥H(s)∥\lVert H(s)\rVert. Notice that the guaranteed threshold ρqubit\rho_{\mathrm{qubit}} is about 1.98 π1.98\,\pi, almost twice the optimal universal Magnus threshold π\pi. The one-step bound is sharp, but this does not establish that ρqubit\rho_{\mathrm{qubit}} is the exact convergence radius of the Fer expansion.

The paper is organized as follows. Section II recalls the Fer recursion and the argument that turns a one-step bound into a sufficient condition for convergence. Section III writes one step for a qubit and gives its exact size. Section IV finds the worst case of one step: it writes the size of one step in terms of the radial speed of the curve (Sec. IV A), reduces the worst case to a problem on curves (Sec. IV B), and solves it (Sec. IV C): the Euler–Lagrange equation gives a candidate profile, and an inequality proves that it gives the worst case. Section IV D shows a qubit Hamiltonian that reaches it. Section V derives the corresponding sufficient condition for convergence and an error bound, and Sec. VI concludes.

The Fer expansion

We briefly recall the construction of the Fer expansion. Given a time-dependent Hamiltonian H(t)=H†(t)H(t)=H^\dagger(t), the time-evolution operator U(t)U(t) satisfies the Schrödinger equation (ℏ=1\hbar=1)

ddtU(t)=−iH(t)U(t),U(0)=1.(2)\frac{\mathrm{d}}{\mathrm{d}t}U(t)=-\mathrm{i}H(t)U(t),\qquad U(0)=\mathbb{1}. \tag*{(2)}

The Fer expansion [24, 25] writes it as an ordered product of exponentials,

U(t)=eF1(t)eF2(t)⋯eFj(t)Uj+1(t),(3)U(t)=\mathrm{e}^{F_1(t)}\mathrm{e}^{F_2(t)}\cdots\mathrm{e}^{F_j(t)}U_{j+1}(t), \tag*{(3)}

where the remainder Uj+1(t)U_{j+1}(t) obeys a Schrödinger equation of the same form, with the effective Hamiltonian

Hj+1(t)=e−Fj(t)Hj(t)eFj(t)−∫01dλ e−λFj(t)Hj(t)eλFj(t),(4)H_{j+1}(t)=\mathrm{e}^{-F_j(t)}H_j(t)\mathrm{e}^{F_j(t)} -\int_0^1\mathrm{d}\lambda\,\mathrm{e}^{-\lambda F_j(t)}H_j(t)\mathrm{e}^{\lambda F_j(t)}, \tag*{(4)}

where

Fj(t)=−i∫0tds Hj(s)(5)F_j(t)=-\mathrm{i}\int_0^t\mathrm{d}s\,H_j(s) \tag*{(5)}

and H1(t)≡H(t)H_1(t)\equiv H(t).

The size of any Hamiltonian is measured by

Kj(t)≡∫0tds ∥Hj(s)∥,(6)K_j(t)\equiv\int_0^t\mathrm{d}s\,\lVert H_j(s)\rVert, \tag*{(6)}

and the truncation error is bounded by the size of the first Hamiltonian that has been left out [25],

∥U(t)−eF1(t)⋯eFN(t)∥≤KN+1(t).(7)\left\lVert U(t)-\mathrm{e}^{F_1(t)}\cdots\mathrm{e}^{F_N(t)}\right\rVert\leq K_{N+1}(t). \tag*{(7)}

The expansion converges if Kj(t)K_j(t) goes to zero as the number of factors grows.

For Hermitian Hamiltonians of any dimension, Refs. [25] and [36] prove two one-step bounds of the form

∥Hj+1(t)∥≤ψi(∥Fj(t)∥)∥Hj(t)∥,(8)\lVert H_{j+1}(t)\rVert\leq\psi_i(\lVert F_j(t)\rVert)\lVert H_j(t)\rVert, \tag*{(8)}

with the classical coefficient ψ1(x)=x\psi_1(x)=x of Ref. [25] and the finite-angle coefficient of Ref. [36],

ψ2(x)={2(1−cos⁡(x))x,0≤x≤π/2,2−π−2x,x≥π/2.(9)\psi_2(x)= \begin{cases} \dfrac{2(1-\cos(x))}{x}, & 0\leq x\leq\pi/2,\\ 2-\dfrac{\pi-2}{x}, & x\geq\pi/2. \end{cases} \tag*{(9)}

Since ∥Fj(t)∥≤Kj(t)\lVert F_j(t)\rVert\leq K_j(t) and ψi(x)\psi_i(x) is increasing, integrating Eq. (8) over time gives

Kj+1(t)≤Ψi(Kj(t)),Ψi(x)≡∫0xdu ψi(u).(10)K_{j+1}(t)\leq\Psi_i(K_j(t)),\qquad\Psi_i(x)\equiv\int_0^x\mathrm{d}u\,\psi_i(u). \tag*{(10)}

Each map satisfies Ψi(x)<x\Psi_i(x)<x below its first positive fixed point ρ\rho, which is the guaranteed convergence threshold of Refs. [25] and [36]: ρ=2\rho=2 for Ψ1(x)\Psi_1(x) and ρ\rho about 2.60582.6058 for Ψ2(x)\Psi_2(x). When K1(t)<ρK_1(t)<\rho the sequence Kj(t)K_j(t) decreases, and its limit can only be zero: the product (3) converges as the number of factors goes to infinity, with an error after NN factors not larger than the NN-fold iterate ΨiN(K1(t))\Psi_i^N(K_1(t)).

One step for a qubit

From now on we consider a qubit. In this case every effective Hamiltonian Hj(t)H_j(t) and every exponent Fj(t)F_j(t) is fixed by a vector of R3\mathbb{R}^3, and one step of the recursion (4)–(5) becomes an explicit map between vectors [38]. We first write the Hamiltonians as vectors (§ 1 A), then derive the map (§ 1 B), and finally read from it the exact size of one step (§ 1 C).

Hamiltonians as vectors

We consider a qubit Hamiltonian H(t)=h(t)⋅TH(t)=\mathbf{h}(t)\cdot\mathbf{T}, where h(t)∈R3\mathbf{h}(t)\in\mathbb{R}^3 and T=(T1,T2,T3)\mathbf{T}=(T_1,T_2,T_3) is the vector of Pauli matrices, with [Ta,Tb]=2i εabcTc[T_a,T_b]=2i\,\varepsilon_{abc}T_c [39]. As shown in Ref. [38], the Fer recursion never leaves this algebra, and every effective Hamiltonian and every exponent can be written as

Hj(t)=hj(t)⋅T,Fj(t)=−ifj(t)⋅T,(11)H_j(t)=\mathbf{h}_j(t)\cdot\mathbf{T},\qquad F_j(t)=-i\mathbf{f}_j(t)\cdot\mathbf{T}, \tag*{(11)}

with h1(t)≡h(t)\mathbf{h}_1(t)\equiv\mathbf{h}(t) and fj(t)=∫0tds hj(s)\mathbf{f}_j(t)=\int_0^t\mathrm{d}s\,\mathbf{h}_j(s). Since (x⋅T)2=∥x∥21(\mathbf{x}\cdot\mathbf{T})^2=\lVert\mathbf{x}\rVert^2\mathbb{1} for every x∈R3\mathbf{x}\in\mathbb{R}^3, the operator norm of x⋅T\mathbf{x}\cdot\mathbf{T} is the Euclidean length ∥x∥\lVert\mathbf{x}\rVert of x\mathbf{x}. We write hj(t)≡∥hj(t)∥h_j(t)\equiv\lVert\mathbf{h}_j(t)\rVert and fj(t)≡∥fj(t)∥f_j(t)\equiv\lVert\mathbf{f}_j(t)\rVert for the lengths of the two vectors, so that Eq. (11) gives

∥Hj(t)∥=hj(t),∥Fj(t)∥=fj(t).(12)\lVert H_j(t)\rVert=h_j(t),\qquad\lVert F_j(t)\rVert=f_j(t). \tag*{(12)}

The size (6) of the effective Hamiltonian is then the time integral of the length of hj(t)\mathbf{h}_j(t),

Kj(t)=∫0tds hj(s).(13)K_j(t)=\int_0^t\mathrm{d}s\,h_j(s). \tag*{(13)}

This gives a geometric picture that we use throughout. As time goes on, the vector fj(t)\mathbf{f}_j(t) draws in R3\mathbb{R}^3 a curve that starts at the origin, since fj(0)=0\mathbf{f}_j(0)=0, and whose velocity is f˙j(t)=hj(t)\dot{\mathbf{f}}_j(t)=\mathbf{h}_j(t). Its distance from the origin is fj(t)=∥Fj(t)∥f_j(t)=\lVert\mathbf{F}_j(t)\rVert, and by Eq. (13) its length up to the time tt is Kj(t)K_j(t). Since a curve is never shorter than the segment that joins its end points, fj(t)≤Kj(t)f_j(t)\leq K_j(t), which is the inequality ∥Fj(t)∥≤Kj(t)\lVert F_j(t)\rVert\leq K_j(t) used in Sec. II.

The Fer map

One step of the Fer recursion maps the vector hj(t)\mathbf{h}_j(t) into the vector hj+1(t)\mathbf{h}_{j+1}(t), and in this subsection we write this map explicitly. For a qubit the conjugations in Eq. (4) are rotations of R3\mathbb{R}^3, given explicitly by the Rodrigues formula [39]. Inserting it into Eq. (4), one shows [38] that the new vector hj+1(t)\mathbf{h}_{j+1}(t) is a combination of two auxiliary vectors. The first is the cross product of fj(t)\mathbf{f}_j(t) with hj(t)\mathbf{h}_j(t), and the second is the cross product of the first with fj(t)\mathbf{f}_j(t),

pj+1(t)≡fj(t)×hj(t),(14)\mathbf{p}_{j+1}(t)\equiv\mathbf{f}_j(t)\times\mathbf{h}_j(t), \tag*{(14)}
qj+1(t)≡pj+1(t)×fj(t).(15)\mathbf{q}_{j+1}(t)\equiv\mathbf{p}_{j+1}(t)\times\mathbf{f}_j(t). \tag*{(15)}

In terms of these vectors the map reads

hj+1(t)=α(fj(t)) pj+1(t)+β(fj(t)) qj+1(t),(16)\mathbf{h}_{j+1}(t)=\alpha(f_j(t))\,\mathbf{p}_{j+1}(t)+\beta(f_j(t))\,\mathbf{q}_{j+1}(t), \tag*{(16)}

with the weights

α(x)=1−cos⁡(2x)2x2−sin⁡(2x)x,(17)\alpha(x)=\frac{1-\cos(2x)}{2x^2}-\frac{\sin(2x)}{x}, \tag*{(17)}
β(x)=cos⁡(2x)x2−sin⁡(2x)2x3.(18)\beta(x)=\frac{\cos(2x)}{x^2}-\frac{\sin(2x)}{2x^3}. \tag*{(18)}

Figure 2 shows the map (16) in the geometric picture of § 1 A. Up to the time tt the curve fj(s)\mathbf{f}_j(s), 0≤s≤t0\leq s\leq t, has length Kj(t)K_j(t) and ends at the distance fj(t)f_j(t) from the origin, where its velocity hj(t)\mathbf{h}_j(t) is tangent to it. We split the velocity into a radial part along fj(t)\mathbf{f}_j(t), hj∥(t)≡[fj(t)⋅hj(t)]fj(t)/fj2(t)\mathbf{h}_j^{\parallel}(t)\equiv[\mathbf{f}_j(t)\cdot\mathbf{h}_j(t)]\mathbf{f}_j(t)/f_j^2(t), and an orthogonal part hj⊥(t)≡hj(t)−hj∥(t)\mathbf{h}_j^{\perp}(t)\equiv\mathbf{h}_j(t)-\mathbf{h}_j^{\parallel}(t), which lies in the plane Πj(t)\Pi_j(t) orthogonal to fj(t)\mathbf{f}_j(t). The radial part drops out of the cross product (14), so that pj+1(t)=fj(t)×hj⊥(t)\mathbf{p}_{j+1}(t)=\mathbf{f}_j(t)\times\mathbf{h}_j^{\perp}(t) is the orthogonal part turned by a right angle around fj(t)\mathbf{f}_j(t) and multiplied by fj(t)f_j(t). Expanding the double cross product (15) gives instead qj+1(t)=fj2(t)hj(t)−[fj(t)⋅hj(t)]fj(t)=fj2(t)hj⊥(t)\mathbf{q}_{j+1}(t)=f_j^2(t)\mathbf{h}_j(t)-[\mathbf{f}_j(t)\cdot\mathbf{h}_j(t)]\mathbf{f}_j(t)=f_j^2(t)\mathbf{h}_j^{\perp}(t). Both auxiliary vectors lie in Πj(t)\Pi_j(t), and so does their combination hj+1(t)\mathbf{h}_{j+1}(t) [38]. One step of the recursion thus erases the radial part of hj(t)\mathbf{h}_j(t), while it rotates the orthogonal part around fj(t)\mathbf{f}_j(t) and changes its length.

The size of one step

The two terms of Eq. (16) are orthogonal, since qj+1(t)\mathbf{q}_{j+1}(t) is parallel to the orthogonal part hj⊥(t)\mathbf{h}_{j}^{\perp}(t), while pj+1(t)\mathbf{p}_{j+1}(t) is hj⊥(t)\mathbf{h}_{j}^{\perp}(t) turned by a right angle. Their lengths are ∥pj+1(t)∥=fj(t)∥hj⊥(t)∥\lVert\mathbf{p}_{j+1}(t)\rVert=f_j(t)\lVert\mathbf{h}_{j}^{\perp}(t)\rVert and ∥qj+1(t)∥=fj2(t)∥hj⊥(t)∥\lVert\mathbf{q}_{j+1}(t)\rVert=f_j^2(t)\lVert\mathbf{h}_{j}^{\perp}(t)\rVert, so that by Pythagoras the length of hj+1(t)\mathbf{h}_{j+1}(t) is

hj+1(t)=η(fj(t))∥hj⊥(t)∥,(19)h_{j+1}(t)=\eta(f_j(t))\lVert\mathbf{h}_{j}^{\perp}(t)\rVert, \tag*{(19)}

with the weight

η2(x)≡x2[α2(x)+x2β2(x)]=1−sin⁡(2x)x+sin⁡2(x)x2.(20)\begin{aligned} \eta^2(x)\equiv x^2\left[\alpha^2(x)+x^2\beta^2(x)\right] \\ &=1-\frac{\sin(2x)}{x}+\frac{\sin^2(x)}{x^2}. \tag*{(20)} \end{aligned}

This is exactly what one step of the recursion does to the size of a qubit Hamiltonian. In Figure 2, the new vector hj+1(t)\mathbf{h}_{j+1}(t) is the orthogonal part hj⊥(t)\mathbf{h}_{j}^{\perp}(t) rotated around fj(t)\mathbf{f}_j(t) and multiplied by the weight η(fj(t))\eta(f_j(t)), which depends only on the distance fj(t)f_j(t) of the curve from the origin. The weight η(x)\eta(x), shown in Figure 3, increases from η(0)=0\eta(0)=0 up to its maximum, reached at

Figure 2. The curve fj(s)\mathbf{f}_j(s), 0≤s≤t0\leq s\leq t (green, dash-dotted), has length Kj(t)K_j(t) and ends at the distance fj(t)f_j(t) from the origin OO. At its end point the velocity hj(t)\mathbf{h}_j(t) (brown) is tangent to the curve, along the brown dashed line, and it splits into the radial part hj∥(t)\mathbf{h}_j^{\parallel}(t) (orange), along fj(t)\mathbf{f}_j(t), and the orthogonal part hj⊥(t)\mathbf{h}_j^{\perp}(t) (violet), in the plane Πj(t)\Pi_j(t) orthogonal to fj(t)\mathbf{f}_j(t) (red). One step of the Fer recursion erases the radial part: the new vector hj+1(t)\mathbf{h}_{j+1}(t) (blue) lies in Πj(t)\Pi_j(t), and it is the orthogonal part rotated around fj(t)\mathbf{f}_j(t) and multiplied by γ(fj(t))\gamma(f_j(t)).

Figure 3. The two functions that determine the optimal radial profile in Sec. IV C 1. The weight η(x)\eta(x) of Eq. (20) (violet) reaches 1 for the first time at x∘x^\circ (open circle), increases up to its maximum η⋆\eta^\star at x⋆x^\star (red, dashed), then oscillates around 1 and tends to 1 for large xx (dotted line). The length L(x)L(x) of Eq. (41) (black) is defined for 0≤x<x⋆0\leq x<x^\star: it grows from 0, as πx/2\pi x/2 for small xx, and diverges at x=x⋆x=x^\star.

x⋆≈2.04,η⋆≡η(x⋆)≈1.26.(21)x^\star\approx2.04,\qquad\eta^\star\equiv\eta(x^\star)\approx1.26. \tag*{(21)}

For larger xx the weight oscillates around 1 and approaches 1 as x→∞x\to\infty.

The worst case of one step

As in Refs. [25, 36], our goal is to express the size Kj+1(t)K_{j+1}(t) of the next effective Hamiltonian in terms of the size Kj(t)K_j(t) of the current one, and then to look for the fixed points of the resulting map, which yield a sufficient condition for convergence.

An exact formula for one step

For a qubit, Eq. (19) already gives the size of one step exactly, but in terms of the component of hj(t)\mathbf{h}_j(t) orthogonal to fj(t)\mathbf{f}_j(t). In this subsection we rewrite it in terms of the speed hj(t)h_j(t) of the curve fj(t)\mathbf{f}_j(t), whose time integral is Kj(t)K_j(t), and of its distance fj(t)f_j(t) from the origin.

We differentiate fj2(t)=fj(t)⋅fj(t)f_j^2(t)=\mathbf{f}_j(t)\cdot\mathbf{f}_j(t) with f˙j(t)=hj(t)\dot{\mathbf{f}}_j(t)=\mathbf{h}_j(t), and obtain the speed at which the curve fj(t)\mathbf{f}_j(t) moves away from the origin, the radial speed

f˙j(t)=fj(t)⋅hj(t)fj(t),(22)\dot{f}_j(t)=\frac{\mathbf{f}_j(t)\cdot\mathbf{h}_j(t)}{f_j(t)}, \tag*{(22)}

which is the component of the velocity hj(t)\mathbf{h}_j(t) along the unit vector fj(t)/fj(t)\mathbf{f}_j(t)/f_j(t). A component of a vector is never longer than the vector itself, ∣fj(t)⋅hj(t)∣≤fj(t)hj(t)|\mathbf{f}_j(t)\cdot\mathbf{h}_j(t)|\leq f_j(t)h_j(t) by the Cauchy–Schwarz inequality, and therefore ∣f˙j(t)∣≤hj(t)|\dot{f}_j(t)|\leq h_j(t): the distance from the origin cannot change faster

The worst case as a problem on curves

than the curve moves, and it changes at the speed hj(t)h_j(t) only when hj(t)\mathbf{h}_j(t) is parallel to fj(t)\mathbf{f}_j(t). The radial part hj∥(t)\mathbf{h}_{j}^{\parallel}(t) of Sec. IIIB has length ∣f˙j(t)∣|\dot{f}_j(t)|, so that by Pythagoras the orthogonal part has length ∥hj⊥(t)∥=[hj2(t)−(f˙j(t))2]1/2\lVert\mathbf{h}_{j}^{\perp}(t)\rVert=[h_j^2(t)-(\dot{f}_j(t))^2]^{1/2}. Together with Eqs. (12) and (19), this yields the exact size of one step,

hj+1(t)=η(fj(t))hj2(t)−(f˙j(t))2,(23)h_{j+1}(t)=\eta(f_j(t))\sqrt{h_j^2(t)-(\dot{f}_j(t))^2}, \tag*{(23)}

and, integrating over time and using hj(s)=K˙j(s)h_j(s)=\dot{K}_j(s) from Eq. (13),

Kj+1(t)=∫0tds η(fj(s))K˙j2(s)−(f˙j(s))2.(24)K_{j+1}(t)=\int_0^t ds\,\eta(f_j(s))\sqrt{\dot{K}_j^2(s)-(\dot{f}_j(s))^2}. \tag*{(24)}

At the times when fj(t)=0f_j(t)=0 the radial speed is not defined, but both sides of Eq. (23) vanish, since η(0)=0\eta(0)=0.

In Eq. (24) we now take as integration variable the length σ\sigma traveled by the curve up to the time ss,

σ≡Kj(s)=∫0sds′ ∥f˙j(s′)∥,(25)\sigma\equiv K_j(s)=\int_0^s ds'\,\lVert\dot{\mathbf{f}}_j(s')\rVert, \tag*{(25)}

and describe the curve by its radial profile r(σ)r(\sigma), the distance from the origin after the length σ\sigma has been traveled,

r(σ)≡fj(s).(26)r(\sigma)\equiv f_j(s). \tag*{(26)}

Changing the integration variable of Eq. (24) from ss to σ\sigma, with dσ=K˙j(s) ds\mathrm{d}\sigma=\dot{K}_j(s)\,\mathrm{d}s and f˙j(s)=dr(σ)dσK˙j(s)\dot{f}_j(s)=\frac{\mathrm{d}r(\sigma)}{\mathrm{d}\sigma}\dot{K}_j(s), gives

Kj+1(t)=∫0Kj(t)dσ η(r(σ))1−(dr(σ)dσ)2(27)K_{j+1}(t)=\int_{0}^{K_j(t)}\mathrm{d}\sigma\,\eta(r(\sigma))\sqrt{1-\left(\frac{\mathrm{d}r(\sigma)}{\mathrm{d}\sigma}\right)^2} \tag*{(27)}

where the square root is real: by Eq. (22) the radial speed is at most the speed, ∣dfj(s)/ds∣≤hj(s)=K˙j(s)\lvert\mathrm{d}f_j(s)/\mathrm{d}s\rvert\le h_j(s)=\dot{K}_j(s), and comparing with dfj(s)/ds=(dr(σ)/dσ)K˙j(s)\mathrm{d}f_j(s)/\mathrm{d}s=(\mathrm{d}r(\sigma)/\mathrm{d}\sigma)\dot{K}_j(s) one has ∣dr(σ)/dσ∣≤1\lvert\mathrm{d}r(\sigma)/\mathrm{d}\sigma\rvert\le1, that is, along the curve the distance from the origin cannot change by more than the length traveled. Since it is a distance that starts from fj(0)=0f_j(0)=0, every radial profile therefore satisfies the constraints

r(0)=0,r(σ)≥0,∣dr(σ)dσ∣≤1.(28)r(0)=0,\qquad r(\sigma)\ge0,\qquad\left\lvert\frac{\mathrm{d}r(\sigma)}{\mathrm{d}\sigma}\right\rvert\le1. \tag*{(28)}

For a curve of length Σ\Sigma we call admissible the radial profiles that satisfy these constraints, and denote their set by

XΣ≡{r:[0,Σ]→R:r(0)=0, r(σ)≥0, ∣r(σ)−r(τ)∣≤∣σ−τ∣ ∀ σ,τ}.(29)X_\Sigma\equiv\left\{r:[0,\Sigma]\to\mathbb{R}:r(0)=0,\ r(\sigma)\ge0,\ \lvert r(\sigma)-r(\tau)\rvert\le\lvert\sigma-\tau\rvert\ \forall\,\sigma,\tau\right\}. \tag*{(29)}

The last condition is the last constraint of Eq. (28) written without derivatives: it also admits profiles with corners, which have a derivative at almost every σ\sigma. Equation (27) shows that Kj+1(t)K_{j+1}(t) depends on Kj(t)K_j(t) only through the upper limit of integration and on the Hamiltonian only through the radial profile r(σ)r(\sigma) of the curve.

Solving the problem on curves

Equation (27) is an equality: for a given Hamiltonian, Kj+1(t)K_{j+1}(t) is exactly the integral on its right-hand side. Hamiltonians with the same Kj(t)K_j(t) have in general different radial profiles, and knowing Kj(t)K_j(t) alone we do not know which one applies. As in Sec. II, however, we look for a bound on Kj+1(t)K_{j+1}(t) that involves Kj(t)K_j(t) only. Such a bound must hold for every radial profile allowed by the constraints (28), that is, in the worst case, and we want it to be as small as possible, since a smaller bound gives a larger convergence radius. We therefore fix the length Kj(t)K_j(t) to an arbitrary value Σ\Sigma and look for the largest value of the functional

F[r]≡∫0Σdσ L(σ,r(σ),r˙(σ))(30)\mathcal{F}[r]\equiv\int_{0}^{\Sigma}\mathrm{d}\sigma\,\mathcal{L}(\sigma,r(\sigma),\dot{r}(\sigma)) \tag*{(30)}

over the admissible profiles r∈XΣr\in X_\Sigma of Eq. (29), with the final distance r(Σ)r(\Sigma) free. Its Lagrangian is

L(σ,r(σ),r˙(σ))≡η(r(σ))1−r˙(σ)2,(31)\mathcal{L}(\sigma,r(\sigma),\dot{r}(\sigma))\equiv\eta(r(\sigma))\sqrt{1-\dot{r}(\sigma)^2}, \tag*{(31)}

where r˙(σ)≡dr(σ)/dσ\dot{r}(\sigma)\equiv\mathrm{d}r(\sigma)/\mathrm{d}\sigma is the radial speed, the derivative of the profile with respect to the length traveled. In this subsection the dot is a derivative with respect to σ\sigma and no longer with respect to time, while the prime denotes the derivative of a function with respect to its argument, as in η′(x)≡dη(x)/dx\eta'(x)\equiv\mathrm{d}\eta(x)/\mathrm{d}x. The integrand contains two competing factors. The weight η(r(σ))\eta(r(\sigma)) is large only when the curve is far from the origin, at distances close to x⋆x^\star, while the square root is large only when the distance r(σ)r(\sigma) changes slowly.

We find the largest value in two parts, as is usual in the calculus of variations [40, 41, 42]. In the first part (Sec. IV C1) we assume that the largest value exists and is reached by a regular profile, and the Euler–Lagrange equation gives a candidate profile rˉ(σ)\bar{r}(\sigma). These assumptions only serve to find the candidate, and nothing is proved there. In the second part (Sec. IV C2) we prove, with an inequality that holds for every admissible profile, that the candidate gives the largest value. This also shows that the largest value exists.

A candidate from the Euler–Lagrange equation

Suppose that the largest value of F[r]\mathcal{F}[r] is reached by a profile r∈XΣr\in X_\Sigma. We compare it with the nearby profiles r(σ)+εw(σ)r(\sigma)+\varepsilon w(\sigma), where ε\varepsilon is a small number of either sign and w(σ)w(\sigma) is a differentiable function with w(0)=0w(0)=0, so that the starting point does not change, and we suppose that these profiles are admissible. The final value w(Σ)w(\Sigma) is free, as r(Σ)r(\Sigma) is. Along this family, F\mathcal{F} becomes a function of ε\varepsilon alone, φ(ε)≡F[r+εw]\varphi(\varepsilon)\equiv\mathcal{F}[r+\varepsilon w], which is largest at ε=0\varepsilon=0, so that its derivative vanishes there.

We compute this derivative by differentiating under the integral at fixed σ\sigma and setting ε=0\varepsilon=0, which gives

dφ(ε)dε∣ε=0=−∫0Σdσ η(r(σ))r˙(σ)1−r˙(σ)2w˙(σ)+∫0Σdσ η′(r(σ))1−r˙(σ)2w(σ).(32)\left.\frac{\mathrm{d}\varphi(\varepsilon)}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} =-\int_{0}^{\Sigma}\mathrm{d}\sigma\,\frac{\eta(r(\sigma))\dot{r}(\sigma)}{\sqrt{1-\dot{r}(\sigma)^2}}\dot{w}(\sigma) +\int_{0}^{\Sigma}\mathrm{d}\sigma\,\eta'(r(\sigma))\sqrt{1-\dot{r}(\sigma)^2}w(\sigma). \tag*{(32)}

In the first integral w˙(σ)\dot{w}(\sigma) is multiplied by −P(σ)-P(\sigma), with

P(σ)≡η(r(σ))r˙(σ)1−r˙(σ)2.(33)P(\sigma)\equiv\frac{\eta(r(\sigma))\dot{r}(\sigma)}{\sqrt{1-\dot{r}(\sigma)^2}}. \tag*{(33)}

An integration by parts moves the derivative from w˙(σ)\dot{w}(\sigma) to P(σ)P(\sigma). This requires r(σ)r(\sigma) to have two continuous derivatives, since P(σ)P(\sigma) contains r˙(σ)\dot{r}(\sigma). The boundary term at σ=0\sigma=0 vanishes, because w(0)=0w(0)=0, while the one at σ=Σ\sigma=\Sigma remains:

dφ(ε)dε∣ε=0=−P(Σ)w(Σ)+∫0Σdσ P˙(σ)w(σ)+∫0Σdσ η′(r(σ))1−r˙(σ)2w(σ).(34)\left.\frac{\mathrm{d}\varphi(\varepsilon)}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} =-P(\Sigma)w(\Sigma)+\int_{0}^{\Sigma}\mathrm{d}\sigma\,\dot{P}(\sigma)w(\sigma) +\int_{0}^{\Sigma}\mathrm{d}\sigma\,\eta'(r(\sigma))\sqrt{1-\dot{r}(\sigma)^2}w(\sigma). \tag*{(34)}

This derivative vanishes for every w(σ)w(\sigma), and we use this fact in two steps. We first take w(Σ)=0w(\Sigma)=0. The boundary term disappears, and the two integrals together vanish for every w(σ)w(\sigma) that vanishes at both ends. By the fundamental lemma of the calculus of variations the sum of their integrands then vanishes at every σ\sigma,

P˙(σ)=−η′(r(σ))1−r˙(σ)2.(35)\dot{P}(\sigma)=-\eta^\prime(r(\sigma))\sqrt{1-\dot{r}(\sigma)^2}. \tag*{(35)}

Computing P˙(σ)\dot{P}(\sigma) from the definition (33) with the product rule, Eq. (35) becomes the Euler–Lagrange equation

η(r(σ))r¨(σ)=−η′(r(σ))[1−r˙(σ)2].(36)\eta(r(\sigma))\ddot{r}(\sigma)=-\eta^\prime(r(\sigma))\left[1-\dot{r}(\sigma)^2\right]. \tag*{(36)}

We then take w(Σ)≠0w(\Sigma)\ne0. The two integrals in Eq. (34) cancel, only the boundary term is left, and P(Σ)=0P(\Sigma)=0. Excluding the case r(Σ)=0r(\Sigma)=0, in which the curve ends at the origin and the weight vanishes, this gives r˙(Σ)=0\dot{r}(\Sigma)=0: since the final distance is free, the profile reaches it with zero radial speed.

The Lagrangian (31) does not depend on σ\sigma explicitly, so that the Euler–Lagrange equation has a constant of motion, a fact known as the Beltrami identity [40, 42]. With Eq. (33), it is E(σ)≡L(σ,r(σ),r˙(σ))+r˙(σ)P(σ)=η(r(σ))[1−r˙(σ)2]−1/2E(\sigma)\equiv\mathcal{L}(\sigma,r(\sigma),\dot{r}(\sigma))+\dot{r}(\sigma)P(\sigma)=\eta(r(\sigma))[1-\dot{r}(\sigma)^2]^{-1/2}. Computing it at σ=Σ\sigma=\Sigma, where r˙(Σ)=0\dot{r}(\Sigma)=0, we get

η(r(σ))1−r˙(σ)2=η(r(Σ)).(37)\frac{\eta(r(\sigma))}{\sqrt{1-\dot{r}(\sigma)^2}}=\eta(r(\Sigma)). \tag*{(37)}

Solving for r˙(σ)\dot{r}(\sigma), with the positive root since the profile moves away from the origin, we find the radial profile

r˙(σ)=1−η2(r(σ))η2(r(Σ)).(38)\dot{r}(\sigma)=\sqrt{1-\frac{\eta^2(r(\sigma))}{\eta^2(r(\Sigma))}}. \tag*{(38)}

At σ=Σ\sigma=\Sigma, Eq. (38) gives r˙(Σ)=0\dot{r}(\Sigma)=0, the condition found above for a free final distance. Separating the variables in Eq. (38) and integrating from 00 to σ\sigma, we obtain the length that the profile needs to reach the distance r(σ)r(\sigma),

σ=∫0r(σ)duη(r(Σ))η2(r(Σ))−η2(u).(39)\sigma=\int_{0}^{r(\sigma)}\mathrm{d}u\frac{\eta(r(\Sigma))}{\sqrt{\eta^2(r(\Sigma))-\eta^2(u)}}. \tag*{(39)}

At σ=Σ\sigma=\Sigma, Eq. (39) becomes a condition on the final distance,

L(r(Σ))=Σ,(40)L(r(\Sigma))=\Sigma, \tag*{(40)}
L(x)≡∫0xduη(x)η2(x)−η2(u).(41)L(x)\equiv\int_{0}^{x}\mathrm{d}u\frac{\eta(x)}{\sqrt{\eta^2(x)-\eta^2(u)}}. \tag*{(41)}

As shown in Figure 3, L(x)L(x) increases from zero and diverges as xx approaches x⋆x^\star. The condition L(r(Σ))=ΣL(r(\Sigma))=\Sigma therefore fixes a single final distance r(Σ)r(\Sigma) in (0,x⋆)(0,x^\star).

In summary, the maximizing candidate profile rˉ(σ)\bar{r}(\sigma) must belong to XΣ\mathcal{X}_{\Sigma} and, moreover, satisfy

rˉ˙(σ)=1−η2(rˉ(σ))η2(rˉ(Σ)),L(rˉ(Σ))=Σ.(42)\dot{\bar{r}}(\sigma)=\sqrt{1-\frac{\eta^2(\bar{r}(\sigma))}{\eta^2(\bar{r}(\Sigma))}},\qquad L(\bar{r}(\Sigma))=\Sigma. \tag*{(42)}

The first equation contains the Euler–Lagrange equation (36). At the end of the profile, the first and the second equation give, respectively,

rˉ˙(Σ)=0,rˉ(Σ)∈(0,x⋆).(43)\dot{\bar{r}}(\Sigma)=0,\qquad\bar{r}(\Sigma)\in(0,x^\star). \tag*{(43)}

We have not justified any of the assumptions made in this part: the candidate still has to be shown to give the largest value. The second part does this without any of these assumptions, and uses the previous conditions only as a starting point.

The candidate gives the largest value

We now prove that F[r]≤F[rˉ]\mathcal{F}[r]\le\mathcal{F}[\bar{r}] for every admissible profile r(σ)r(\sigma). Nothing of the first part is used, except the candidate itself.

First step: monotone profiles below x⋆x^\star. The first of Eqs. (42) and the second of Eqs. (43) suggest which profiles to consider. The final distance of the candidate lies below x⋆x^\star, and on [0,x⋆][0,x^\star] the weight η(x)\eta(x) is strictly increasing, so that the argument of the square root in Eq. (42) is nonnegative only if the final distance is the largest one. We now obtain both properties for a generic admissible profile r(σ)r(\sigma): we first cut it at x⋆x^\star and then replace it with its running maximum (Figure 4),

Radial profiles

Figure 4. The first step of the proof in Sec. IV C 2, for an example radial profile. The profile r(σ)r(\sigma) differs from r1(σ)r_1(\sigma) of Eq. (44) only where it exceeds x⋆x^\star (red): there it is the black dashed line, and r1(σ)=x⋆r_1(\sigma)=x^\star. The profile r1(σ)r_1(\sigma) differs from r2(σ)r_2(\sigma) of Eq. (45) (violet) only where it lies below the largest value reached before: there it is the blue dashed line, and r2(σ)r_2(\sigma) stays at that value. The arrows show the two modifications. On both the derivative vanishes and the integrand of Eq. (30) does not decrease, which gives Eq. (46).

r1(σ)≡min⁡{r(σ),x⋆},(44)r_1(\sigma)\equiv\min\{r(\sigma),x^\star\}, \tag*{(44)}
r2(σ)≡max⁡0≤τ≤σr1(τ).(45)r_2(\sigma)\equiv\max_{0\le\tau\le\sigma}r_1(\tau). \tag*{(45)}

Where r1(σ)r_1(\sigma) differs from r(σ)r(\sigma), the profile r1(σ)r_1(\sigma) stays at the distance x⋆x^\star: its derivative vanishes, and its weight η(x⋆)\eta(x^\star) is the largest value of η(x)\eta(x). Where r2(σ)r_2(\sigma) differs from r1(σ)r_1(\sigma), the profile r2(σ)r_2(\sigma) stays at the largest distance reached before: its derivative vanishes, and its weight is not smaller, because η(x)\eta(x) is increasing on [0,x⋆][0,x^\star]. In both cases the square root takes its largest value 1, so that at almost every σ\sigma the integrand of Eq. (30) does not decrease, and

F[r]≤F[r1]≤F[r2].(46)\mathcal{F}[r] \le\mathcal{F}[r_1] \le\mathcal{F}[r_2]. \tag*{(46)}

The profile r2(σ)r_2(\sigma) is still admissible, it never decreases, and it never exceeds x⋆x^\star. We call xf≡r2(Σ)x_{\mathrm f}\equiv r_2(\Sigma), in [0,x⋆][0,x^\star], its final distance. It is the largest distance from the origin reached by the curve, cut at x⋆x^\star, and in general it differs from the final distance r(Σ)r(\Sigma) of the original profile. Since r2(σ)r_2(\sigma) never decreases, 0≤r2(σ)≤xf0\le r_2(\sigma)\le x_{\mathrm f} for every σ\sigma. If xf=0x_{\mathrm f}=0, then r2(σ)r_2(\sigma) vanishes and F[r]≤F[r2]=0≤F[rˉ]\mathcal{F}[r]\le\mathcal{F}[r_2]=0\le\mathcal{F}[\bar r], so that from now on we take xf>0x_{\mathrm f}>0, and η(xf)>0\eta(x_{\mathrm f})>0.

Second step: an inequality between numbers. We consider the Lagrangian (31) at a position qq in (0,xf](0,x_{\mathrm f}] and a velocity vv in [0,1][0,1], two numbers that will play the role of r(σ)r(\sigma) and r˙(σ)\dot r(\sigma), so that L(σ,q,v)=η(q)1−v2\mathcal{L}(\sigma,q,v)=\eta(q)\sqrt{1-v^2}. At fixed qq this is a concave function of vv, so that its graph lies below each of its tangent lines (Figure 5). We take the tangent at the velocity that the candidate would have at the position qq, that is

Figure 5. The second step of the proof in Sec. IV C 2. At fixed qq, the Lagrangian L(σ,q,v)=η(q)1−v2\mathcal{L}(\sigma,q,v)=\eta(q)\sqrt{1-v^2} is concave in vv (violet), and its graph lies below its tangent line ℓ(v)\ell(v) of Eq. (48) at the point v0v_0 of Eq. (47) (red, dashed; open circle). The tangent line takes the value η(xf)\eta(x_{\mathrm f}) at v=0v=0 (open circle). At every vv the vertical distance between the two is the right-hand side of Eq. (49) minus its left-hand side, which is therefore nonnegative.

v0≡1−η2(q)η2(xf).(47)v_0\equiv\sqrt{1-\frac{\eta^2(q)}{\eta^2(x_{\mathrm f})}}. \tag*{(47)}

The tangent line at v0v_0 is

ℓ(v)≡∂L(σ,q,v)∂v∣v=v0(v−v0)+L(σ,q,v0)=η(xf)−vη2(xf)−η2(q),(48)\begin{aligned} \ell(v)&\equiv\left.\frac{\partial\mathcal{L}(\sigma,q,v)}{\partial v}\right|_{v=v_0}(v-v_0)+\mathcal{L}(\sigma,q,v_0)\\ &=\eta(x_{\mathrm f})-v\sqrt{\eta^2(x_{\mathrm f})-\eta^2(q)}, \tag*{(48)} \end{aligned}

and by concavity L(σ,q,v)≤ℓ(v)\mathcal{L}(\sigma,q,v)\le\ell(v) for every vv in [0,1][0,1], with equality only for v=v0v=v_0. This is the inequality between numbers

η(q)1−v2+vη2(xf)−η2(q)≤η(xf).(49)\eta(q)\sqrt{1-v^2}+v\sqrt{\eta^2(x_{\mathrm f})-\eta^2(q)}\le\eta(x_{\mathrm f}). \tag*{(49)}

For q=0q=0 the weight vanishes and Eq. (49) reads vη(xf)≤η(xf)v\eta(x_{\mathrm f})\le\eta(x_{\mathrm f}), which holds for every vv in [0,1][0,1].

Third step: integration. We set q=r2(σ)q=r_2(\sigma) and v=r˙2(σ)v=\dot r_2(\sigma) in Eq. (49), which is allowed since 0≤r2(σ)≤xf0\le r_2(\sigma)\le x_{\mathrm f} and 0≤r˙2(σ)≤10\le\dot r_2(\sigma)\le1, and integrate from σ=0\sigma=0 to σ=Σ\sigma=\Sigma. The first term gives F[r2]\mathcal{F}[r_2], and the right-hand side gives η(xf)Σ\eta(x_{\mathrm f})\Sigma. Changing the integration variable to u=r2(σ)u=r_2(\sigma), the second term becomes an integral from 0 to xfx_{\mathrm f}, which depends only on xfx_{\mathrm f}. We write it as I(xf)I(x_{\mathrm f}), with

I(x)≡∫0xdu η2(x)−η2(u).(50)I(x)\equiv\int_0^x\mathrm{d}u\,\sqrt{\eta^2(x)-\eta^2(u)}. \tag*{(50)}

Together with the first step,

F[r]≤F[r2]≤η(xf)Σ−I(xf).(51)\mathcal{F}[r]\le\mathcal{F}[r_2]\le\eta(x_{\mathrm f})\Sigma-I(x_{\mathrm f}). \tag*{(51)}

The two terms of this bound have a simple meaning. The first, η(xf)Σ\eta(x_{\mathrm f})\Sigma, is what one would obtain if the curve could spend its whole length at the distance xfx_{\mathrm f}. The second,

I(xf)I(x_{\mathrm f}), is the price paid for the part of the length spent moving away from the origin to reach that distance.

Fourth step: the worst final distance. The distance xfx_{\mathrm f} depends on the profile, and we remove it by taking the largest value of g(x)≡η(x)Σ−I(x)g(x)\equiv\eta(x)\Sigma-I(x) over [0,x⋆][0,x^\star]. To find it, we differentiate g(x)g(x). In I(x)I(x) the variable xx appears both in the upper limit and in the integrand. The upper limit gives the integrand at u=xu=x, which vanishes, and the integrand gives

dI(x)dx=∫0xdu η(x)η′(x)η2(x)−η2(u)=η′(x)L(x),(52)\frac{\mathrm{d}I(x)}{\mathrm{d}x}=\int_0^x\mathrm{d}u\,\frac{\eta(x)\eta'(x)}{\sqrt{\eta^2(x)-\eta^2(u)}}=\eta'(x)L(x), \tag*{(52)}

so that

dg(x)dx=η′(x)[Σ−L(x)].(53)\frac{\mathrm{d}g(x)}{\mathrm{d}x}=\eta'(x)\left[\Sigma-L(x)\right]. \tag*{(53)}

In the interval (0,x⋆)(0,x^\star) we have η′(x)>0\eta'(x)>0, and L(x)L(x) is increasing. By the second of Eqs. (42), rˉ(Σ)\bar r(\Sigma) is the solution of L(x)=ΣL(x)=\Sigma, so that g(x)g(x) increases for x<rˉ(Σ)x<\bar r(\Sigma) and decreases for x>rˉ(Σ)x>\bar r(\Sigma). Its largest value on [0,x⋆][0,x^\star] is g(rˉ(Σ))g(\bar r(\Sigma)), and η(xf)Σ−I(xf)≤η(rˉ(Σ))Σ−I(rˉ(Σ))\eta(x_{\mathrm f})\Sigma-I(x_{\mathrm f})\le\eta(\bar r(\Sigma))\Sigma-I(\bar r(\Sigma)) for every admissible profile.

The four steps together give, for every admissible profile,

F[r]≤F[r2]≤η(xf)Σ−I(xf)≤η(rˉ(Σ))Σ−I(rˉ(Σ)),(54)\begin{aligned} \mathcal{F}[r]&\le\mathcal{F}[r_2]\le\eta(x_{\mathrm f})\Sigma-I(x_{\mathrm f})\\ &\le\eta(\bar r(\Sigma))\Sigma-I(\bar r(\Sigma)), \tag*{(54)} \end{aligned}

where the first inequality comes from the first step, the second from the second and third steps, and the third from the fourth step.

Fifth step: the candidate reaches the bound. We now take r=rˉr=\bar r in Eq. (54). The candidate is admissible, since it starts at the origin and its slope (38) lies between 0 and 1. It never decreases and stays below rˉ(Σ)<x⋆\bar{r}(\Sigma) < x^\star, so that the first step leaves it unchanged and xf=rˉ(Σ)x_{\mathrm{f}} = \bar{r}(\Sigma). By Eq. (38), at every σ\sigma its radial speed is the point of tangency v0v_0 of the second step, with q=rˉ(σ)q = \bar{r}(\sigma) and xf=rˉ(Σ)x_{\mathrm{f}} = \bar{r}(\Sigma), so that Eq. (49) is an equality and the third step gives F[rˉ]=η(rˉ(Σ))Σ−I(rˉ(Σ))\mathcal{F}[\bar{r}] = \eta(\bar{r}(\Sigma))\Sigma- I(\bar{r}(\Sigma)). Finally, xf=rˉ(Σ)x_{\mathrm{f}} = \bar{r}(\Sigma) is the point where g(x)g(x) is largest. Every inequality in Eq. (54) is therefore an equality for the candidate. Each of the two conditions found in the first part makes one inequality an equality: Eq. (38) the inequality between numbers of the second step, and L(rˉ(Σ))=ΣL(\bar{r}(\Sigma)) = \Sigma the choice of the final distance in the fourth step.

The chain (54) and the fifth step show that the largest value of F[r]\mathcal{F}[r] exists, that it is reached by the candidate, and that it equals η(rˉ(Σ))Σ−I(rˉ(Σ))\eta(\bar{r}(\Sigma))\Sigma- I(\bar{r}(\Sigma)). With Σ=Kj(t)\Sigma= K_j(t), Eq. (27) then gives

Kj+1(t)≤Ψqubit(Kj(t)),(55)K_{j+1}(t) \le\Psi_{\mathrm{qubit}}(K_j(t)), \tag*{(55)}
Ψqubit(Kj(t))≡max⁡0≤x≤x⋆[η(x)Kj(t)−I(x)],(56)\Psi_{\mathrm{qubit}}(K_j(t)) \equiv\max_{0 \le x \le x^\star} \left[\eta(x)K_j(t) - I(x)\right], \tag*{(56)}

where the maximum is reached at x=rˉ(Kj(t))x = \bar{r}(K_j(t)), fixed by L(rˉ(Kj(t)))=Kj(t)L(\bar{r}(K_j(t))) = K_j(t).

A qubit Hamiltonian that reaches the bound

We now show that the bound (55) is reached by a qubit Hamiltonian, so that Ψqubit(Kj(t))\Psi_{\mathrm{qubit}}(K_j(t)) is the exact worst case of one step. By Eq. (27), K2(t)K_2(t) depends on the Hamiltonian only through the radial profile and the length K1(t)K_1(t), so that every Hamiltonian whose curve has length K1(t)=ΣK_1(t) = \Sigma and a radial profile that satisfies Eqs. (42) reaches the bound. An example can be read directly from Eqs. (42), because the first of them already contains the two components of a velocity of length one. Squaring it, we obtain

(rˉ˙(σ))2+[η(rˉ(σ))η(rˉ(Σ))]2=1.(57)\left(\dot{\bar{r}}(\sigma)\right)^2 + \left[\frac{\eta(\bar{r}(\sigma))}{\eta(\bar{r}(\Sigma))}\right]^2 = 1. \tag*{(57)}

We therefore take a curve traveled at speed one, h1(t)=1h_1(t) = 1, so that K1(t)=tK_1(t) = t and the length traveled σ\sigma is the time, with tt in [0,Σ][0,\Sigma]. By Eq. (22), the radial part of the velocity is the derivative of the distance from the origin, which is now rˉ˙(t)\dot{\bar{r}}(t). Since the speed is one, the orthogonal part has length 1−(rˉ˙(t))2\sqrt{1-(\dot{\bar{r}}(t))^2}, which is the square root in Eq. (27), and by Eq. (57) with σ=t\sigma= t this length is η(rˉ(t))/η(rˉ(Σ))\eta(\bar{r}(t))/\eta(\bar{r}(\Sigma)).

It remains to choose the direction of the orthogonal part. We take the simplest case, a curve in a plane, with an angle θ(t)\theta(t) such that θ(0)=0\theta(0) = 0,

f1(t)=rˉ(t)(cos⁡θ(t),sin⁡θ(t),0).(58)\mathbf{f}_1(t) = \bar{r}(t)(\cos\theta(t),\sin\theta(t),0). \tag*{(58)}

Figure 6. The curve f1(t)\mathbf{f}_1(t) of the Hamiltonian (59) for Σ=ρqubit\Sigma= \rho_{\mathrm{qubit}} (green). The curve is numerically plotted as described in Sec. IV D. At two points the velocity h1(t)\mathbf{h}_1(t) (brown) is split into its radial part h1∥(t)\mathbf{h}^{\parallel}_1(t) (orange) and its orthogonal part h1⊥(t)\mathbf{h}^{\perp}_1(t) (violet). The curve ends on the circle of radius rˉ(Σ)\bar{r}(\Sigma) (red, dashed), where h1(Σ)\mathbf{h}_1(\Sigma) has no radial part.

The Hamiltonian is then H(t)=h1(t)⋅TH(t) = \mathbf{h}_1(t)\cdot\mathbf{T}, with

h1(t)=rˉ˙(t)(cos⁡θ(t),sin⁡θ(t),0)+η(rˉ(t))η(rˉ(Σ))(−sin⁡θ(t),cos⁡θ(t),0),(59)\mathbf{h}_1(t) = \dot{\bar{r}}(t)(\cos\theta(t),\sin\theta(t),0) + \frac{\eta(\bar{r}(t))}{\eta(\bar{r}(\Sigma))}(-\sin\theta(t),\cos\theta(t),0), \tag*{(59)}

where the first term is the radial part of the velocity and the second is its orthogonal part (Fig. 6). The angle θ(t)\theta(t) is fixed by h1(t)=f˙1(t)\mathbf{h}_1(t) = \dot{\mathbf{f}}_1(t), which follows from f1(t)=∫0tds h1(s)\mathbf{f}_1(t) = \int_0^t \mathrm{d}s\,\mathbf{h}_1(s). The derivative of Eq. (58) has the same first term as Eq. (59), and its second term is rˉ(t)θ˙(t)(−sin⁡θ(t),cos⁡θ(t),0)\bar{r}(t)\dot{\theta}(t)(-\sin\theta(t),\cos\theta(t),0). The two second terms coincide if rˉ(t)θ˙(t)=η(rˉ(t))/η(rˉ(Σ))\bar{r}(t)\dot{\theta}(t) = \eta(\bar{r}(t))/\eta(\bar{r}(\Sigma)), that is, if

θ(t)=∫0tds η(rˉ(s))η(rˉ(Σ))rˉ(s).(60)\theta(t) = \int_0^t \mathrm{d}s\,\frac{\eta(\bar{r}(s))}{\eta(\bar{r}(\Sigma))\bar{r}(s)}. \tag*{(60)}

The integrand is finite also at s=0s = 0, where it equals 1/η(rˉ(Σ))1/\eta(\bar{r}(\Sigma)), because η(x)/x\eta(x)/x tends to 1 for small xx. For Fig. 6 we set Σ=ρqubit\Sigma= \rho_{\mathrm{qubit}} and determine the final radius from L(rˉ(Σ))=ΣL(\bar{r}(\Sigma)) = \Sigma. Since the speed is one, σ=t\sigma= t: we numerically evaluate Eq. (39) to obtain rˉ(t)\bar{r}(t), then evaluate the integral in Eq. (60) to obtain θ(t)\theta(t). These two functions give the curve f1(t)\mathbf{f}_1(t) through Eq. (58).

Equation (58) shows that the curve is a kind of spiral, since its distance from the origin and its angle both grow along it. By the first of Eqs. (42), rˉ˙(t)>0\dot{\bar{r}}(t) > 0 for t<Σt < \Sigma, because rˉ(t)<rˉ(Σ)\bar{r}(t) < \bar{r}(\Sigma) and η(x)\eta(x) increases up to x⋆x^\star, and by Eq. (60), θ˙(t)>0\dot{\theta}(t)>0. Dividing the two derivatives, the growth of the distance for a unit of angle is

rˉ˙(t)θ˙(t)=rˉ(t)η2(rˉ(Σ))−η2(rˉ(t))η(rˉ(t))(61)\frac{\dot{\bar{r}}(t)}{\dot{\theta}(t)}=\frac{\bar{r}(t)\sqrt{\eta^2(\bar{r}(\Sigma))-\eta^2(\bar{r}(t))}}{\eta(\bar{r}(t))} \tag*{(61)}

which is positive for t<Σt<\Sigma and vanishes at t=Σt=\Sigma. The curve leaves the origin in the radial direction, since η(0)=0\eta(0)=0 makes the orthogonal part vanish, and by Eq. (43) it reaches the circle of radius rˉ(Σ)\bar{r}(\Sigma) tangentially, staying inside it.

The curve (58) has length K1(Σ)=ΣK_1(\Sigma)=\Sigma and radial profile rˉ(σ)\bar{r}(\sigma), so that Eq. (27) gives K2(Σ)=F[rˉ]K_2(\Sigma)=\mathcal{F}[\bar{r}], and by the fifth step of Sec. IV C 2 this is η(rˉ(Σ))Σ−I(rˉ(Σ))=Ψqubit(Σ)\eta(\bar{r}(\Sigma))\Sigma-I(\bar{r}(\Sigma))=\Psi_{\mathrm{qubit}}(\Sigma). The unit speed, the plane and the sign of θ˙(t)\dot{\theta}(t) are choices that only serve to make the example explicit.

A sufficient condition for convergence

The fixed point

As in Sec. II, the bound (55) makes the expansion converge when Ψqubit(Kj(t))<Kj(t)\Psi_{\mathrm{qubit}}(K_j(t))<K_j(t), that is, when η(x)Kj(t)−I(x)<Kj(t)\eta(x)K_j(t)-I(x)<K_j(t) for every xx in [0,x⋆][0,x^\star]. For the values of xx with η(x)≤1\eta(x)\leq1 this always holds, because I(x)>0I(x)>0. For the values with η(x)>1\eta(x)>1 it holds when Kj(t)<I(x)/[η(x)−1]K_j(t)<I(x)/[\eta(x)-1]. A sufficient condition for convergence is therefore Kj(t)<ρqubitK_j(t)<\rho_{\mathrm{qubit}}, where ρqubit\rho_{\mathrm{qubit}} is the smallest of these thresholds:

ρqubit=min⁡x∘<x≤x⋆I(x)η(x)−1≈6.2106(62)\rho_{\mathrm{qubit}}=\min_{x^\circ<x\leq x^\star}\frac{I(x)}{\eta(x)-1}\approx6.2106 \tag*{(62)}

where x∘x^\circ, about 1.1656 (Figure 3), is the point where η(x)\eta(x) first reaches 1. Dividing by π\pi, we get

ρqubitπ≈1.9769(63)\frac{\rho_{\mathrm{qubit}}}{\pi}\approx1.9769 \tag*{(63)}

Convergence and error bound

Every effective Hamiltonian of a qubit is again a qubit Hamiltonian, so the bound Kj+1(t)≤Ψqubit(Kj(t))K_{j+1}(t)\leq\Psi_{\mathrm{qubit}}(K_j(t)) holds at every step. The map Ψqubit(Σ)\Psi_{\mathrm{qubit}}(\Sigma) is nondecreasing, as the maps Ψi(x)\Psi_i(x): by Eq. (56) it is the largest of the functions η(x)Σ−I(x)\eta(x)\Sigma-I(x), none of which decreases with Σ\Sigma, since η(x)≥0\eta(x)\geq0. The argument of Sec. II therefore applies with Ψqubit(Σ)\Psi_{\mathrm{qubit}}(\Sigma) in place of Ψi(x)\Psi_i(x). So far we have considered Hamiltonians H(t)=h(t)⋅TH(t)=\boldsymbol{h}(t)\cdot\mathbf{T}. A general qubit Hamiltonian also contains a multiple of the identity, H(t)=h0(t)1+h(t)⋅TH(t)=h_0(t)\mathbb{1}+\boldsymbol{h}(t)\cdot\mathbf{T} with h0(t)h_0(t) real. This term commutes with everything: it changes only the first factor of the product (3), by the global phase exp⁡[−i∫0tds h0(s)]\exp[-i\int_0^t ds\,h_0(s)], and leaves all the effective Hamiltonians Hj(t)H_j(t) with j≥2j\geq2 unchanged [36]. We conclude that

for every qubit Hamiltonian the Fer expansion converges in operator norm whenever

∫0tds ∥h(s)∥<ρqubit(64)\int_0^t ds\,\lVert\boldsymbol{h}(s)\rVert<\rho_{\mathrm{qubit}} \tag*{(64)}

with the truncation error bounded by

∥U(t)−eF1(t)⋯eFN(t)∥≤ΨqubitN(∫0tds ∥h(s)∥).(65)\left\lVert U(t)-e^{F_1(t)}\cdots e^{F_N(t)}\right\rVert\leq\Psi_{\mathrm{qubit}}^{N}\left(\int_0^t ds\,\lVert\boldsymbol{h}(s)\rVert\right). \tag*{(65)}

Since η′(x)\eta'(x) decreases from 1 to 0 on [0,x⋆][0,x^\star], changing variable from uu to η(u)\eta(u) in I(x)I(x) gives I(x)≥πη2(x)/4I(x)\geq\pi\eta^2(x)/4, the area of a quarter disk of radius η(x)\eta(x). Therefore η(x)Σ−I(x)≤η(x)Σ−πη2(x)/4≤Σ2/π\eta(x)\Sigma-I(x)\leq\eta(x)\Sigma-\pi\eta^2(x)/4\leq\Sigma^2/\pi for every xx in [0,x⋆][0,x^\star], that is, Ψqubit(Σ)≤Σ2/π\Psi_{\mathrm{qubit}}(\Sigma)\leq\Sigma^2/\pi. Once Kj(t)K_j(t) drops below π\pi, this gives Kj+1(t)≤Kj2(t)/πK_{j+1}(t)\leq K_j^2(t)/\pi, and the remainder decays doubly exponentially.

Half-width and comparison

The eigenvalues of H(s)H(s) are h0(s)±∥h(s)∥h_0(s)\pm\lVert\boldsymbol{h}(s)\rVert, so ∥h(s)∥\lVert\boldsymbol{h}(s)\rVert is the half-width of the spectrum, [λmax⁡(s)−λmin⁡(s)]/2[\lambda_{\max}(s)-\lambda_{\min}(s)]/2, and Eq. (64) is the condition (1). In terms of the operator norm, ∥H(s)∥=∣h0(s)∣+∥h(s)∥\lVert H(s)\rVert=|h_0(s)|+\lVert\boldsymbol{h}(s)\rVert. For a traceless Hamiltonian Eq. (64) therefore reads ∫0tds ∥H(s)∥<ρqubit\int_0^t ds\,\lVert H(s)\rVert<\rho_{\mathrm{qubit}}, and for a general one this condition is sufficient.

The bound ρqubit\rho_{\mathrm{qubit}} should be compared with the values 2 of Ref. [25] and about 2.6058 of Ref. [36], which hold for Hermitian generators of any dimension, and with the radius π\pi of the Magnus expansion [15, 16]. An example is the rotating field of Ref. [17], h(s)=γ(cos⁡(2s),0,−sin⁡(2s))\boldsymbol{h}(s)=\gamma(\cos(2s),0,-\sin(2s)) with 0≤s≤π0\leq s\leq\pi and γ>0\gamma>0, whose cumulative norm at the final time is γπ\gamma\pi. The Magnus series diverges for every γ>1\gamma>1, that is, as soon as the cumulative norm exceeds π\pi, while by Eq. (64) the Fer expansion converges for every γ<ρqubit/π\gamma<\rho_{\mathrm{qubit}}/\pi, about 1.98. For a qubit, the guaranteed Fer threshold in cumulative norm is thus about 1.98 times the optimal universal Magnus threshold.

Conclusions

Our analysis shows how the geometry of qubit dynamics can sharpen the convergence criterion for the Fer expansion. The key is the relation between each effective Hamiltonian and its own time integral: the component that increases the norm of the integral is precisely the component erased by the next step. The norm of the integral and the size of the next step therefore cannot be maximized independently. For a qubit this competition is described exactly by Eq. (23), which turns the search for the worst case of one step into a variational problem on the radial profile of a curve.

The resulting one-step bound is reached by a qubit Hamiltonian, so that the convergence threshold obtained here is optimal among those derived by iterating a universal bound based only on the time integral of the norm of each effective Hamiltonian. We have not proved that it coincides with the exact convergence radius of the Fer expansion for a qubit: a larger guaranteed threshold would require information beyond this universal one-step bound.

Extending this approach to systems with more than two levels requires a richer geometric description and whether this additional information can yield a stronger general convergence guarantee in the operator norm remains open.

Acknowledgments

We thank Kazuya Yuasa (Department of Physics, Waseda University, Tokyo, Japan) and Daniel Burgarth (Department Physik, Friedrich-Alexander-Universität Erlangen–Nürnberg, Erlangen, Germany), with whom our work on the Fer expansion started, for many discussions on the Fer expansion in general and for the ideas that came out of them.

Use of AI

Claude Code (Anthropic), with the model Claude Opus 5.5, was used under the direction of the authors as a research tool in this work. It played an important role in suggesting mathematical tools for the argument. The interpretation of teh results, their development and all the remaining work are due to the authors. They build on our previous work on the Fer expansion [36, 38], within a research program on the Fer expansion and related product expansions that both authors have pursued for several years. Claude was also used to edit the manuscript, in the last revision with the model Claude Fable 5.2. The authors checked all derivations, numerical values, and figures, and take full responsibility for the content.

References

References

  1. [1]F. Bloch, Nuclear induction, Phys. Rev. 70, 460 (1946).
  2. [2]R. P. Feynman, F. L. Vernon, and R. W. Hellwarth, Geometrical representation of the Schrödinger equation for solving maser problems, J. Appl. Phys. 28, 49 (1957).DOI
  3. [3]L. M. K. Vandersypen and I. L. Chuang, NMR techniques for quantum control and computation, Rev. Mod. Phys. 76, 1037 (2005).DOI
  4. [4]M. Kieferová, A. Scherer, and D. W. Berry, Simulating the dynamics of time-dependent Hamiltonians with a truncated Dyson series, Phys. Rev. A 99, 023414 (2019).DOI
  5. [5]D. W. Berry, A. M. Childs, Y. Su, X. Wang, and N. Wiebe, Time-dependent Hamiltonian simulation with L¹-norm scaling, Quantum 4, 254 (2020).DOI
  6. [6]P. A. Moreno Casares, M. S. Zini, and J. M. Arrazola, Quantum simulation of time-dependent Hamiltonians via commutator-free quasi-Magnus operators, Quantum 8, 1567 (2024).DOI
  7. [7]D. Fang, D. Liu, and R. Sarkar, Time-dependent Hamiltonian simulation via Magnus expansion: algorithm and superconvergence, Commun. Math. Phys. 406, 128 (2025).DOI
  8. [8]L. Bagnasacco, F. Taddei, and V. Giovannetti, Formal integration of electron scattering processes via separation of dynamical and geometric contributions, J. Phys. A: Math. Theor. 58, 505302 (2025).DOI
  9. [9]O. Oreshkov, T. A. Brun, and D. A. Lidar, Scheme for fault-tolerant holonomic computation on stabilizer codes, Phys. Rev. A 80, 022325 (2009).DOI
  10. [10]E. Sjöqvist, D. M. Tong, L. Mauritz Andersson, B. Hessmo, M. Johansson, and K. Singh, Non-adiabatic holonomic quantum computation, New J. Phys. 14, 103035 (2012).DOI
  11. [11]F. J. Dyson, The S matrix in quantum electrodynamics, Phys. Rev. 75, 1736 (1949).DOI
  12. [12]W. Magnus, On the exponential solution of differential equations for a linear operator, Commun. Pure Appl. Math. 7, 649 (1954).DOI
  13. [13]S. Klarsfeld and J. A. Oteo, Recursive generation of higher-order terms in the Magnus expansion, Phys. Rev. A 39, 3270 (1989).pubmed.ncbi.nlm.nih.gov/9901627
  14. [14]S. Blanes, F. Casas, J. A. Oteo, and J. Ros, The Magnus expansion and some of its applications, Phys. Rep. 470, 151 (2009).
  15. [15]P. C. Moan and J. Niesen, Convergence of the Magnus series, Found. Comput. Math. 8, 291 (2008).
  16. [16]F. Casas, Sufficient conditions for the convergence of the Magnus expansion, J. Phys. A: Math. Theor. 40, 15001 (2007).DOI
  17. [17]G. Lakos, Convergence estimates for the Magnus expansion II. C*-algebras, arXiv:1910.03328, 2019.arxiv.org/abs/1910.03328
  18. [18]M. Suzuki, Generalized Trotter’s formula and systematic approximants of exponential operators and inner derivations with applications to many-body problems, Commun. Math. Phys. 51, 183 (1976).DOI
  19. [19]S. Lloyd, Universal quantum simulators, Science 273, 1073 (1996).
  20. [20]A. M. Childs, Y. Su, M. C. Tran, N. Wiebe, and S. Zhu, Theory of Trotter error with commutator scaling, Phys. Rev. X 11, 011020 (2021).DOI
  21. [21]F. Casas, A. Murua, and M. Nadininic, Efficient computation of the Zassenhaus formula, Comput. Phys. Commun. 183, 2386 (2012).DOI
  22. [22]R. M. Wilcox, Exponential operators and parameter differentiation in quantum physics, J. Math. Phys. 8, 962 (1967).DOI
  23. [23]A. Arnal, F. Casas, C. Chiralт, and J. A. Oteo, A unifying framework for perturbative exponential factorizations, Mathematics 9, 637 (2021).
  24. [24]F. Fer, Résolution de l’équation matricielle dU/dt = pU par produit infini d’exponentielles matricielles, Bull. Cl. Sci., Acad. R. Belg. 44, 818 (1958).DOI
  25. [25]S. Blanes, F. Casas, J. A. Oteo, and J. Ros, Magnus and Fer expansions for matrix differential equations: the convergence problem, J. Phys. A: Math. Gen. 31, 259 (1998).DOI
  26. [26]A. Iserles, Solving linear ordinary differential equations by exponentials of iterated commutators, Numer. Math. 45, 183 (1984).DOI
  27. [27]A. Iserles, H. Z. Munthe-Kaas, S. P. Nørsett, and A. Zanna, Lie-group methods, Acta Numer. 9, 215 (2000).DOI
  28. [28]A. Zanna, Collocation and relaxed collocation for the Fer and the Magnus expansions, SIAM J. Numer. Anal. 36, 1145 (1999).DOI
  29. [29]P. K. Madhu and N. D. Kurur, Fer expansion for effective propagators and Hamiltonians in NMR, Chem. Phys. Lett. 418, 235 (2006).DOI
  30. [30]E. S. Mananga, On the Fer expansion: Applications in solid-state nuclear magnetic resonance and physics, Phys. Rep. 608, 1 (2016).DOI
  31. [31]U. Haeberlen and J. S. Waugh, Coherent averaging effects in magnetic resonance, Phys. Rev. 175, 453 (1968).DOI
  32. [32]E. S. Mananga and T. Charpentier, Introduction of the Floquet-Magnus expansion in solid-state nuclear magnetic resonance spectroscopy, J. Chem. Phys. 135, 044109 (2011).DOI
  33. [33]L. Viola, E. Knill, and S. Lloyd, Dynamical decoupling of open quantum systems, Phys. Rev. Lett. 82, 2417 (1999).DOI
  34. [34]F. Casas, Fer’s factorization as a symplectic integrator, Numer. Math. 74, 283 (1996).DOI
  35. [35]A. Zanna, The Fer expansion and time-symmetry: a Strang-type approach, Appl. Numer. Math. 39, 435 (2001).
  36. [36]L. Bagnasacco and V. Giovannetti, Improved convergence radius of the Fer expansion for Hermitian generators, arXiv:2609.28302, 2026.arxiv.org/abs/2609.28302
  37. [37]J. G. Stampfli, The norm of a derivation, Pac. J. Math. 33, 737 (1970).DOI
  38. [38]L. Bagnasacco, D. Burgarth, V. Giovannetti, and K. Yuasa, The Fer expansion of a driven spin under σu(2) operations cannot terminate beyond first order, arXiv:2610.27979, 2026.
  39. [39]J. J. Sakurai and J. Napolitano, Modern Quantum Mechanics, 3rd ed. (Cambridge University Press, Cambridge, 2020).
  40. [40]I. M. Gelfand and S. V. Fomin, Calculus of Variations (Prentice-Hall, Englewood Cliffs, NJ, 1963), reprinted by Dover, Mineola, NY, 2000.
  41. [41]G. Buttazzo, M. Giaquinta, and S. Hildebrandt, One-dimensional Variational Problems: An Introduction (Clarendon Press, Oxford, 1998).
  42. [42]M. Gobbino, Corso di Elementi di Calcolo delle Variazioni: Stampato integrale delle lezioni, A.A. 2015/2016 (Università di Pisa, Pisa, 2016), lecture notes in Italian.

Paper details

vlnc:W_ap4usbhs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Claude Code, Claude Fable 5.2

View statement
Looking for related papers…
Contents