Higher-Order Methods for Hamiltonian Engineering Pulse Sequence Design

Matthew Tyler1 ,∗ Hengyun Zhou1 ,∗ Leigh S. Martin1 , Nathaniel Leitao1 , and Mikhail D. Lukin1†
1
Department of Physics, Harvard University, Cambridge, Massachusetts 02138, USA
We introduce a framework for designing Hamiltonian engineering pulse sequences that systematically ac-
counts for the effects of higher-order contributions to the Floquet-Magnus expansion. Our techniques result
in simple, intuitive decoupling rules, despite the higher-order contributions naively involving complicated, non-
local-in-time commutators. We illustrate how these rules can be used to efficiently design improved Hamiltonian
engineering pulse sequences for a wide variety of tasks, such as dynamical decoupling, quantum sensing, and
quantum simulation.

I. INTRODUCTION AND MOTIVATION average Hamiltonian theory. First, we find that the frame rep-
arXiv:2303.07374v1 [quant-ph] 13 Mar 2023

resentation employed in Ref. [7, 9, 30] still provides a con-

The effective control of many-body quantum dynamics is venient way to describe the pulse sequence and contributions,
an important challenge in the emerging field of quantum sci- resulting in analytical decoupling rules for higher-order terms.
ence and technology, with wide-ranging applications in quan- As an example, in Fig. 1(c) we illustrate how first-order Mag-
tum computation [1, 2], quantum sensing [3], and quan- nus terms involving disorder and Heisenberg interactions have
tum simulation [4, 5]. One of the key tools for controlling a simple geometric interpretation in analogy with dipoles, and
such many-body quantum dynamics is Hamiltonian engineer- in Fig. 1(d) we illustrate how first-order Magnus terms in-
ing [6–19], in which a train of pulses transform the original volving Ising and Heisenberg interactions have a similar in-
system Hamiltonian into a desired target Hamiltonian for var- terpretation as balancing the center of mass along a given
ious applications. Indeed, from the inception of such tech- axis. Second, we find that although there exist additional
niques in early NMR work to the present day, Hamiltonian en- cross-terms, the majority of finite pulse duration effects can
gineering has enabled high resolution spectroscopy [6, 8, 20– still be described as a simple extension of the effective free
22], high sensitivity metrology [23], as well as the realization evolution time [9], making it easy to build in robustness to
of exotic Floquet phases of matter [24–26]. sequence design. Finally, we extend the principle of pulse
One of the key tools for performing Hamiltonian engi- cycle decoupling [7] to more general pulse sequences and
neering is average Hamiltonian theory [27]. Here, the engi- Hamiltonians, beyond those where the zeroth-order average
neered Hamiltonian is approximated by the time-average of Hamiltonian vanishes. We use this to show how decoupling
interaction-picture Hamiltonians with respect to the control rules can be significantly simplified for pulse sequences that
pulses. This allows the effective engineering of many-body are composed of common motifs, such as spin echoes [31]
Hamiltonians, even in the case where only global manipula- or WAHUHA blocks [6] (Fig. 1(a,b)), resulting in time-local
tion of spins is accessible, as is the case in many large-scale decoupling conditions even for higher-order Magnus contri-
quantum systems [5, 28]. Moreover, design rules that system- butions. Together, these techniques allow us to find higher-
atically take into account robustness against various imper- order robust pulse sequences with substantially improved per-
fections can be derived [7, 9], enabling robust pulse sequence formance for a variety of dynamical decoupling, Hamiltonian
design as well. engineering and quantum sensing applications, as discussed
Despite the success of techniques based on average Hamil- here and in the accompanying paper, Ref. [32].
tonian theory, large variations in performance still exist among This paper is organized as follows: in Sec. II, we review
the different sequences obtained, suggesting that higher-order our representation of the pulse sequence and associated in-
contributions in the full Magnus expansion may play an im- teraction picture Hamiltonian, as well as existing decoupling
portant role. Existing works treating higher-order contribu- rules for the zeroth order effective Hamiltonian. In Sec. III,
tions often rely purely on symmetrization, or treat the higher- we utilize this representation to provide general expressions
order terms on a case-by-case basis [7, 9, 29]. However, find- for the higher-order Magnus contributions and present sys-
ing general conditions for the cancellation of higher-order tematic decoupling conditions for higher-order terms. We
Magnus terms can be non-trivial, as the expressions involve then analyze the structures present in these decoupling rules
commutators that are non-local in time. in Sec. IV, finding significant simplifications for commonly-
In this paper, we systematically analyze higher-order Mag- found pulse sequence structures in both the disorder-dominant
nus contributions to effective Hamiltonians, providing a gen- and interaction-dominant regimes. We also tabulate the re-
eral toolset for pulse sequence design in interacting spin sys- sulting decoupling rules, and provide a pictorial depiction
tems in the form of concise decoupling rules. Despite the non- of them. In Sec. V, we provide further details on the ef-
local nature of the commutators involved in higher-order con- ficient numerical screening of pulse sequences, resulting in
tributions, we are still able to generalize many results from high-performance pulse sequences for dynamical decoupling,
quantum sensing, and quantum simulation. Finally, in Sec. VI
we conclude with a discussion of further extensions and fu-
∗ These authors contributed equally to this work ture directions of the formalism. A summary of the notation
† lukin@physics.harvard.edu adopted in this manuscript can be found in Appendix. A.
 2

px (p/2)y (p/2)x py (p/2)y (p/2)x Assuming ideal, infinitely fast rotation pulses, and that the
(a) (b)
combined rotation unitary is identity Pn ⋯P1 = I, we can
(-p/2)x py px (-p/2)y write the total unitary evolution as

FX
+1 -1 T
U (T ) = T exp(− ∫ iH̃(t)dt) ≈ exp(−iH (0) T ), (2)

FY
+1 -1
0

X
FZ
+1 +1 -1 -1

F
where T indicates time-ordering and T is the Floquet pe-

Y
+1 -1

F
Rapid Echo Rapid Symmetrization-1
riod. For pulse separations much shorter than the dynamical

Z
+1

F
(c) timescale of the system, we can conveniently write the effec-
tive Hamiltonian to leading order as
Net Dipole Cancellation T
1
(d) H (0) = ∫ H̃(t1 )dt1 . (3)
T 0
For general system Hamiltonians satisfying the secular ap-
proximation (rotating wave approximation under a strong
Row Frame Balance quantizing field) [9], the interaction picture Hamiltonian
H̃s (t) can be uniquely determined by transformations of the
FIG. 1. Visualization of representative first-order cancellation S z operator (we will refer to these as toggling “frames”)
rules. (a) A rapid spin echo block, which cancels disorder-disorder
and disorder-Ising terms. (b) Rapid symmetrization block, which S̃ z (t) = Uc† (t)S z Uc (t) = ∑ Fµ (t)S µ , (4)
cancels Ising-Ising terms. (c) Cancellation of net dipoles along each µ
axis, which cancels disorder-Heisenberg terms. (d) Balancing the
weight in each row, which cancels Ising-Heisenberg terms. where S µ is a basis for the spin system, e.g. the Pauli spin
operators for qubits, and we have defined the coefficients

Fµ (t) = 2Tr[S µ S̃ z (t)]. (5)

II. GENERAL FRAMEWORK AND REVIEW OF
EXISTING RESULTS Assuming ideal, instantaneous pulses (the case of finite
pulse effects and other associated imperfections are discussed
II.1. General Framework and Frame Representation in Sec. III.4), we can express the preceding information in the
form of a single 4 × N matrix, where each element Fµ,k corre-
sponds to the coefficient Fµ (t) during the kth free evolution
We begin by introducing our method to represent the pulse time, and the last row contains the free evolution time dura-
sequence and associated average Hamiltonian, which will tion.
greatly simplify the analysis of effective Hamiltonians and
As a concrete example, let us consider a spin-1/2 system,
finite pulse effects compared to the conventional representa-
where each pulse Pk is assumed to be a π/2 pulse around
tion of individual pulses. We will adopt the toggling-frame
±x̂, ŷ. Note that π pulses can be viewed as two consecutive
sequence representation (also known as the Mansfield repre-
π/2 pulses, with zero time separation in between. With S µ
sentation) used in Choi et al. [9, 30], which focuses on how
chosen to be the Pauli basis, a spin echo can be represented as
operators are transformed under the pulses, rather than the ap-
plied pulses themselves. In addition to being a complete and
concise representation of the pulse sequence, this represen- ⎛0 0⎞
F ⎜0 0⎟
tation also has the additional advantage that it leads to sim- ( ) =⎜ ⎟, (6)
τ echo ⎜+1 −1⎟
ple decoupling conditions that are amenable to fast numerical ⎝τ τ⎠
screening.
Consider a pulse sequence composed of n global spin- while the WAHUHA decoupling sequence for dipolar interac-
rotation pulses {P1 , ⋯, Pn } acting on a system with native tions [6] can be expressed as
Hamiltonian H, with a free evolution time τk preceding the
kth pulse Pk . The interaction picture Hamiltonian with re-
⎛0 0 +1 +1 0 0⎞
spect to the ideal control pulses can then be written as F ⎜0 +1 0 0 +1 0⎟
( ) =⎜ ⎟. (7)
τ WAHUHA ⎜+1 0 0 0 0 +1⎟
H̃(t) = Uc† (t)⊗m HUc (t)⊗m , (1) ⎝τ τ τ τ τ τ⎠

where H̃(t) is the interaction picture Hamiltonian at time t, Pictorially, we can represent the first three rows of the matrix
Uc (t) is the single spin rotation due to the control field (e.g. by the blocks in Fig. 1(a,b), in which a yellow(green) block
Uc (t) = Pk Pk−1 ⋯P1 right after the kth pulse), and m is the indicates a +1(-1) value along the given axis (row) at a given
number of spins in the system. time (column). We illustrate more advanced versions of spin
 3

echoes and WAHUHA blocks in Fig. 1(a,b), in both pulse no- periodically-driven system, and extend the analysis beyond
tation and the frame matrix notation utilized here. In the pre- the average Hamiltonian described in Eq. (3).
ceding examples, we have neglected finite pulse duration ef- The total unitary over a single Floquet cycle can be ex-
fects, but they can be easily treated by specifying an additional pressed in terms of a time-independent effective Hamiltonian
intermediate toggling frame with zero time duration. U(T ) = exp(−iHeff T ), where in the fast-driving limit, the ef-
This representation allows us to easily express the interac- fective Hamiltonian can be written via the Magnus expansion
tion picture Hamiltonian H̃(t). For example, for a spin-1/2 up to order l as Heff ≈ ∑lk=0 H (k) , with
dipolar-interacting many-body spin system with on-site disor-
der, the system Hamiltonian can be written as 1 T
H (0) = ∫ H(t1 )dt1 , (14)
T 0
Hdip = ∑ hi Siz + ∑ Jij (Six Sjx + Siy Sjy − 2Siz Sjz ) T t1
H (1) =
−i
i ij
∫ dt1 ∫ dt2 [H(t1 ), H(t2 )], (15)
z 2T 0 0
= ∑ hi Si + ∑ Jij (S⃗i ⋅ S⃗j − 3Siz Sjz ), (8) T t1 t2
1
i ij H (2) = ∫ dt1 ∫ dt2 ∫ dt3
6T 0 0 0
where hi is the on-site disorder strength for spin i, and Jij is ([H(t1 ), [H(t2 ), H(t3 )]] + [H(t3 ), [H(t2 ), H(t1 )]]) .
the dipolar interaction between spins i and j. With our repre- (16)
sentation, for sequences composed of π/2 or π pulses around
±x̂, ŷ, the interaction picture Hamiltonian during the kth free Higher-order terms are more complex, involving progres-
evolution time can be easily expressed as sively deeper nested commutators, but in the fast-driving limit
they will be relatively suppressed, and we can focus on the
H̃dip,k = ∑ Fµ,k hi Siµ + ∑ Jij S⃗i ⋅ S⃗j − 3 ∑ Fµ,k
2
Jij Siµ Sjµ , leading order terms above.
iµ ij ijµ
Plugging in Eq. (10) and separating the time-independent
(9)
operator commutation relation information from the time in-
where we have organized the terms according to how they tegrals, we have
transform with Fµ,k . Using these expressions, it is easy to ver-
1 α T
ify that the spin echo cancels disorder, since ∑k Fµ,k τk = 0, H (0) = ∑ O ∫ dt1 cα (t1 ), (17)
while the WAHUHA pulse sequence fully symmetrizes (de- α T 0
2
couples) dipolar interactions, since ∑k Fµ,k τk is the same for T t1
H (1) = ∑
−i α β
all µ. [O , O ] ∫ dt1 ∫ dt2 cα (t1 )cβ (t2 ),
α,β 2T 0 0
Motivated by these considerations, for any secular Hamil-
(18)
tonian, we will organize the interaction picture Hamiltonian
1
in terms of how the operators transform as the toggling frame H (2) = ∑ [Oα , [Oβ , Oγ ]] ∭ dt1 dt2 dt3
changes. Let us write α,β,γ 6T 0≤t3 ≤t2 ≤t1 ≤T
α (cα (t1 )cβ (t2 )cγ (t3 ) + cα (t3 )cβ (t2 )cγ (t1 )). (19)
H̃(t) = ∑ cα (t)O , (10)
α
This allows us to reduce the computation of the Magnus ex-
where cα (t) are time-dependent coefficients encoding the pansion to the evaluation of a few integrals on the c(t) co-
frame transformations of a general operator basis set Oα (we efficients, which can in turn be readily phrased as algebraic
will use greek letters to denote labels of operator sets in the conditions on the set of frame transformations.
remainder of the paper). For the example above in Eq. (8), we
can write out the individual terms in the summation as
II.3. Review of Zeroth Order Rules
O0 = ∑ Jij S⃗i ⋅ S⃗j , c0 (t) = 1, (11)
ij

O 1,µ
= ∑ hi Siµ , c1,µ (t) = Fµ,k , (12) Using the preceding framework, we can readily write down
i conditions for the cancellation or symmetrization of various
O2,µ = 3 ∑ Jij Siµ Sjµ , 2
c2,µ (t) = Fµ,k , (13) zeroth order average Hamiltonian terms, see also Ref. [9] for
ij details. For example, plugging Eqs. (11-13) into Eq. (17), and
assuming ideal, instantaneous pulses, we can easily see that
which clearly illustrates how the various terms in the Hamil- on-site disorder is cancelled when ∑k Fµ,k τk = 0 for each axis
tonian transform differently with the toggling frames. µ = x̂, ŷ, ẑ, while interactions are symmetrized into a Heisen-
2
berg Hamiltonian or cancelled when ∑k Fµ,k τk is equal for
all different µ.
II.2. Magnus Expansion More importantly, as shown in Ref. [9], these conditions
can be readily generalized to the case with pulse imperfec-
With this general representation framework in hand, we tions. The primary effect of finite pulse durations is to ex-
will now briefly review the Magnus expansion, which pro- tend the effective free evolution times in each frame, as most
vides a useful tool to calculate the effective dynamics of the of the terms generated by π/2 rotations can be written as an
 4

average of the Hamiltonian before and after the pulse. How- respectively, and cross terms between the two sequences, re-
ever, there will be additional terms arising from rotation angle sulting in
errors or interaction cross-terms during rotations, which give
rise to additional chirality or parity conditions between neigh- 1 (1) (0) (0) 1 (1)
H (1) = HA + ( )[T HB , T HA ] + HB ,
−i
(20)
boring frames [9]. 2 4T 2
The simple, time-local nature (all rules only involve neigh-
(0,1)
boring frames) of these decoupling conditions enabled ef- where HA,B are the zeroth (first) order effective Hamiltoni-
ficient design and screening of pulse sequences. Indeed, ans during pulse sequences A and B.
using these simple decoupling conditions, novel pulse se- The key observation of pulse cycle decoupling is that if the
quences with improved decoupling performance have been (0) (0)
commutator [HA , HB ] vanishes, then the first-order con-
found, leading to the demonstration of the first solid-state AC
tribution fully decouples into the sum of that in each individ-
magnetometer that operates beyond the limits of spin-spin in-
ual block, regardless of the details. In prior work [7], this
teractions [23]. However, the further extension of such tech-
was achieved by making one of the average Hamiltonians van-
niques to incorporate higher-order Magnus contributions and
ish, thus causing the commutator to automatically vanish as
improve performance is at first sight challenging, given the
well. For more general Hamiltonians, however, this no longer
time-non-local nature of higher-order Magnus terms, which
directly applies, since the Heisenberg interaction is invariant
involve commutators between all times of a Floquet cycle. In
under global rotations and cannot be cancelled with a global
the following, we demonstrate how this challenge can be sys-
drive [18].
tematically overcome by utilizing the structure of commonly-
used pulse sequences. Despite this challenge, we find that we can still make use
of the pulse cycle decoupling principle in many scenarios be-
yond the case where the Hamiltonian vanishes. First, even
if the total Hamiltonian does not vanish, pulse cycle decou-
III. SYSTEMATIC ANALYSIS OF HIGHER-ORDER pling can still apply to individual terms. For example, if a
MAGNUS TERMS given block fully decouples disorder (the rapid echo blocks in
Fig. 1(a)), then any first-order terms involving disorder will
With the basic formalism in hand, we now turn to the sys- not have cross terms between this block and other parts of
tematic extension of these decoupling rules from zeroth-order the sequence, simplifying the design. Second, if the interac-
to higher-order. First, we will describe the pulse cycle de- tion is transformed into the same form in two separate blocks,
(0)
coupling principle [7] and extend it to the case of more gen- then although HA,B are both nonzero, they still commute, and
eral interactions, which serves as a useful tool to decompose so the pulse cycle decoupling principle still applies (see e.g.
non-local higher-order terms into local blocks. We will then Fig. 2). Thus, even if the interaction has a Heisenberg com-
systematically derive expressions for first- and second-order ponent that cannot be cancelled, the cross-term is still zero
Magnus contributions in the general case, assuming ideal because [∑ij Jij Si ⋅ Sj , ∑ij Jij Si ⋅ Sj ] = 0. This insight gen-
pulses. Finally, we briefly describe how the treatment can be eralizes the pulse cycle decoupling principle to cases in which
readily generalized to the case with finite pulse durations, pri- one desires to engineer a non-zero target Hamiltonian, signif-
marily by extending the effective duration of free evolution icantly expanding its applicability.
times, with more details given in Appendix. B. While we have illustrated the pulse cycle decoupling princi-
ple at first order, the same methods also apply at higher-order
by generalizing the arguments in Ref. [7]. For example, the
III.1. Pulse Cycle Decoupling second-order Magnus contribution can be expressed as a sum
of commutators between zeroth and first-order terms, and thus
if lower orders are fully symmetrized, then the second-order
In order to simplify sequence analysis, it is helpful to be
Magnus contribution will also separate into independent, local
able to break down larger pulse sequences into smaller blocks
blocks.
and analyze them independently. In this section, we show how
common motifs used in sequence design—in the form of spin
echoes or interaction symmetrization—allow us to decompose
higher-order contributions into a sum of independent, local III.2. First-Order Decoupling
pieces, no longer requiring non-local correlators between ar-
bitrary locations and thus significantly simplifying the design. We now return to analyze the structure of higher-order
Our results are applicable even to some cases where the sym- Magnus terms directly and derive decoupling rules for vari-
metrization results in a residual Heisenberg interaction Hamil- ous contributions. The expressions here will be derived in full
tonian, thus extending the existing methods [7] of pulse cycle generality, without making use of the pulse cycle decoupling
decoupling to new and experimentally important regimes. principle, although we will use this to further simplify the ex-
Let us proceed by examining the first-order contribution pressions in following sections.
when sequentially applying two sequences A and B of equal In order to better understand the structure of the first-
length T . We can split the first-order Magnus contribution in order Magnus contributions and derive decoupling rules, let
Eq. (15) into integrals within the first and second sequence us rewrite the preceding expressions into a form that relates
 5

(a) Disorder Decoupling (b) Interaction Decoupling The first term in the parenthesis can be interpreted as a product
between the cα coefficient during a given frame and the inte-
gral of zeroth order average Hamiltonians up to the center of
the frame. The last term is simply the product of two zeroth-
order contributions over the entire Floquet period, which will
vanish when zeroth-order decoupling rules are satisfied. Geo-
metrically, this expression can be understood as rewriting the
triangular integration area in Eq. (18) into a sum over thin col-
A B A B umn slices.
(c) Full Decoupling We emphasize that these results apply to all first-order con-
tributions, illustrating the common structure found in the de-
coupling of many different types of terms. By keeping track
of the running sum, the evaluation of this expression now re-
quires only linear time, as opposed to the naive quadratic com-
plexity. In addition, although Eq. (18) still contains products
of coefficients that are non-local in time, in Sec. IV we shall
A B see that in many cases of interest, it can be reduced into sim-
ple, local decoupling rules.
𝐻"! 𝐻#!
! !
𝐻 = + No cross term 𝐻"#
We also generalize these results to the case with finite pulse
durations in Sec. III.4. The primary effect, similar to the
FIG. 2. Pulse Cycle Decoupling. Example pulse blocks where the zeroth-order case [9], is to lengthen the effective duration of
pulse cycle decoupling principle can be applied, and first-order con- each free evolution time by an amount proportional to the
tributions separate into independent, local terms. (a) A spin echo pulse duration. There will be additional cross terms that we
cancels disorder, and thus HA = HB = 0, satisfying the pulse
(0) (0)
tabulate in the appendix, but they are generally smaller.
cycle decoupling condition. Note that in this case, a single block
being fully decoupled is sufficient for the standard pulse cycle de-
coupling condition to hold. (b) Frame symmetrization between the III.3. Second-Order Decoupling
different axes symmetrizes the interaction into a Heisenberg form,
so even if the two blocks have a residual zeroth order interaction
Hamiltonian, the zeroth order contributions commute, and thus the We can now apply the same formalism to the second-order
pulse cycle decoupling condition still applies. (c) Combining spin Magnus contributions. As we show in Appendix. D, by re-
echoes and frame symmetrization yields pulse cycle decoupling for ordering the integrals, we can re-express the second-order
the full Hamiltonian. contribution in terms of the zeroth- and first-order contribu-
tions at different times. Let us define the first-order contribu-
tion from time t1 to t2 of the operator [Oβ , Oγ ] as
them to zeroth-order Magnus contributions and makes clear (1)
how terms can be cancelled. cβ,γ (t1 , t2 ) = t1 <tb∬<ta <t2 (cβ (ta )cγ (tb ) − cβ (tb )cγ (ta )),
Denoting the zeroth order contribution up to a given time as (23)
t
cα (t) ≡ ∫0 cα (t1 )dt1 , we can rewrite Eq. (18) as
¡ where we have dropped the integrand dta dtb for notational
simplicity here and below. We can rewrite the expression as
H (1) =
−i α β
∑ [O , O ] follows
T α>β
T 1
×∫
1
cα (t1 )cβ (t1 )dt1 − cα (T )cβ (T ). (21) H (2) = α β γ
∑ ([O , [O , O ]])
0 ¡ 2¡ ¡ 6T α<β<γ
T
Focusing first on the case of instantaneous, ideal pulses (the (1) (1)
×∫ dt1 cα (t1 ) (cβ,γ (0, t1 ) − cβ,γ (t1 , T )) . (24)
more general case will be treated in Sec. III.4), our toggling- 0
frame Hamiltonian becomes piecewise-constant in time, al-
lowing us to replace integrals with summations. We can then Thus, we see a very similar structure as at first order,
define the discrete frame equivalents of the terms in the previ- wherein the second-order term can also be expressed as a sim-
ous section, letting cα,k = cα (t) for t ∈ [tk − τk /2, tk + τk /2] ple integral of lower-order products, enabling formulation of
and cα,k ≡ ∑l<k cα,l τl , and find simple decoupling rules.
¡
H (1) = ∑ [Oα , Oβ ]
β<α
III.4. Robustness Conditions
n
1 1
× ( ∑ cα,k τk (cβ,k + τk cβ,k ) − cα,n+1 cβ,n+1 ), We now extend these results to the case with finite pulse
k=1 ¡ 2 2¡ ¡ durations, and describe how to incorporate robustness to these
(22) effects into the sequence design. We will focus our attention
 6

(a) type terms and other terms. Note that there will be no such
𝑐́!,# 𝑐!,# 𝑐̀!,# 𝑐́%,& 𝑐%,& 𝑐̀%,&
cross-terms for the disorder part of our Hamiltonian. Thus,
although the magnitude of this term can in principle scale as
O(τp τ ), in practice the coefficients are small for disorder-
𝑞',#,&
dominated systems, and we will analyze this in detail instead
𝜏# 𝜏( 𝜏& in Appendix B.
Frame 1 Hamiltonian

(b)
≈ Frame 2 Hamiltonian

IV. HIGHER-ORDER DECOUPLING RULES

IV.1. Summary of General Rules

FIG. 3. Evolution of frames with finite pulse durations. (a) Il- We now utilize the results from the preceding section to
lustration of the coefficients of two different frames in a typical se- derive concrete decoupling rules for various important higher-
quence. With finite pulse durations, the coefficients have additional
order contributions.
tails on the two sides as well as cross-terms in overlapping regions.
(b) The leading order effect of this is an extension of the free evolu- Plugging in different Hamiltonian terms into the expres-
tion durations, while still treating each frame independently. sions derived in Sec. III.2,III.3, we arrive at the decoupling
rules in Tab. I. As higher-order terms originate from commu-
tators between different terms, we label the cancellation rules
with all operators involved. The table includes two types of
on the dominant contribution, which we find to be a simple ex- contributions: first, there are the main terms that will appear
tension of the effective free evolution time by an amount pro- even with ideal, infinitesimally short pulses, together with cor-
portional to the pulse duration. A full treatment of the finite rections to their effective duration due to finite pulse dura-
pulse effects, including additional sub-leading cross terms, tions; second, we include terms that come purely from the fi-
can be found in Appendix. B. nite pulse duration, in the form of the overlapping pulse term
As shown in Fig. 3, with a finite pulse duration, the coeffi- derived in Eq. (25). There is one additional type of term, as
cient cα (t) of each term of the Hamiltonian will consist of a mentioned in the preceding section, that involves interaction
ramp up (the preceding rotation), free evolution in the frame, cross terms qρ during continuous rotations. We omit them
a ramp down (the following rotation), as well as some addi- from this table, since they do not appear for disorder terms
tional cross-terms between the frames. that are dominant in our experiments, but we discuss them in
The primary effect of finite pulse effects is illustrated in more detail in Appendix B. Note also that since the Magnus
Fig. 3(b), in which the effective duration of each frame is expansion is not invariant with respect to cyclic permutations
lengthened by the integral of ćα,k and c̀α,k over time. This of the pulse sequence, there are modifications to terms relating
is a simple extension of the calculation for the zeroth-order to the first and last frames in the complete expression. How-
case, and scales as O(τ τp ), where τ is the free evolution time ever, we neglect them from this table, both to simplify nota-
and τp is the pulse duration. We incorporate this into the main tion, as well as due to the fact that after many Floquet cycles
term in the decoupling rule table, Tab. I, as described in more we expect the contributions from these boundary terms to be
detail in the next section. diminished.
In addition to this dominant term, there are contributions While the decoupling rules are somewhat more complicated
that scale as O(τp2 ) or higher powers. First, we have to treat than the zeroth-order rules derived in Ref. [9], many of them
the overlap of the frames, which gives rise to the additional nonetheless have simple geometric intuitions (Fig. 1), and can
term: be further simplified for many common scenarios. Moreover,
the decoupling rules can often be satisfied with simple local
∬ c̀α,k (θ1 )ćβ,k+1 (θ2 ) − ćα,k+1 (θ2 )c̀β,k (θ1 ). (25) motifs, further simplifying the pulse sequence design task.
0<θ2 <θ1 <π/2 For example, in many cases, by using the pulse cycle de-
coupling principle described in Sec. III.1, one can apply the
These terms correspond to the overlap of the ramp up of one same rules as zeroth-order sequence design, except requiring
frame with the ramp down of another. We note that we get a the cancellation on a much faster timescale. These consider-
positive effect from the ramp down of one into the ramp up ations are summarized in the last column of Tab. I. We will
of the other, and a negative effect from the ramp up of one now go through a few representative examples in more detail,
with the ramp down of the other. These are the main finite and explain how to interpret and simplify the rules.
pulse effects to each of the correction terms, and are shown in
column 4 of Tab. I.
The final contribution originates from first-order contribu-
tions involving interaction cross-terms qρ . More specifically, IV.2. Fast Echo Cancellation for Disorder-Related Terms
during the continuous rotation from an XX Hamiltonian to a
Y Y Hamiltonian, XY -type terms are generated; the qρ cross- Let us now focus our attention on first-order rules related to
terms come from first-order cross-terms between these XY - disorder-disorder and disorder-Ising terms.
 7

# Decoupling Effect Algebraic Condition (Ideal Pulse) Algebraic Condition (Finite Pulse Correction) Local Cancellation Condition
2τ 2
1 Disorder-Disorder 2 ∑n
k=1 Fµ,k (τk +
4 v
τ ) F<k
π p
− F̄ µ F̄ v ( πp ) (1 − π4 ) ∑ηk=1 Fµ,k Fv,k+1 − Fv,k Fµ,k+1 Fast Echo
2πp 2 π
2 Disorder-Ising 2 ∑n
k=1 Fµ,k (τk +
4
π p
v
τ ) I<k − F̄ µ I¯v ( π
) (4 − 32 ) ∑n
k=1 Fµ,k ∣Fv,k+1 ∣ − ∣Fv,k ∣ Fµ,k+1 Fast Echo
2τp 2 π 2
3 Ising-Ising ∑n v v
k=1 ∣Fµ,k ∣ (τk + τp ) (I<k − I>k ) ( π
) ( 32 − 14 ) ∑n
k=1 ∣Fµ,k ∣ ∣Fv,k+1 ∣ − ∣Fv,k ∣ ∣Fµ,k+1 ∣ Block Symmetrization

4 T
4 Disorder-Heisenberg ∑n
k=1 Fµ,k (τk + τ ) (tk
π p
− 2
) 0 Dipole Cancellation

4 T
5 Ising-Heisenberg ∑n
k=1 ∣Fµ,k ∣ (τk + τ ) (tk
π p
− 2
) 0 Row Balancing

6 Heisenberg-Heisenberg 0 0 Automatically Satisfied

1 ν,ρ
7 2nd-Order Disorder 2 ∑n
k=1 Fµ,k (τk + τ ) F<k
π p
− F µ F ν,ρ Fast Echo

TABLE I. Higher-order cancellation rules. Each row describes a different higher-order contribution, described in the second column. The
third column gives the expressions that need to vanish to cancel the largest contributions to a given term, including contributions from free
evolution periods and frame lengthening corrections due to finite pulse durations. The fourth column gives further finite pulse corrections. The
q-cross terms are ignored, only giving an additional correction term in terms involving the Ising interaction. The last column gives the local
ν
cancellation condition that eliminates that type of error. F<k = ∑l<k Fν,l (τl + π4 τp ), I<k
ν
= ∑l<k ∣Fν,l ∣(τl + τp ), F̄ ν = ∑k Fν,k (τk + π4 τp ),
I = ∑k ∣Fν,k ∣(τk + τp ).
¯ν

To derive the first and second conditions in Tab. I, which We also include an example of a second-order rule involv-
hold with full generality and no restrictions on the frame ma- ing disorder only in Tab. I. As one can see, the structure bears
trix, let us examine the structure of the first-order Magnus con- many similarities with the first-order contributions. In the case
tribution shown in Eq. (22). There, we found that a generic where the pulse sequence is composed of fast echoes, we can
first-order Magnus contribution can be rewritten as a prod- further simplify the expressions.
uct between the current frame contribution of one term and
the cumulative contribution of another term, together with a
factor corresponding to the total zeroth order contribution of
IV.3. Block Symmetrization for Ising-Ising Terms
both terms. Plugging the disorder and Ising expressions in
Eqs. (11-13) into Eq. (22) results in the rules in the third col-
umn of Tab. I. Moving on to the first-order Ising-Ising terms (row 3 of
Due to this common structure, we see in Tab. I that the Tab. I), we see that the structure of the expression again has
decoupling of first-order disorder-disorder and disorder-Ising many similarities as above. However, here we have grouped
contributions are almost identical, except replacing one term the terms slightly differently, since the zeroth-order sum I¯ν =
from scaling with Fµ,k to be scaling instead as Fµ,k 2
= ∣Fµ,k ∣ ∑k ∣Fkν ∣(τk + τp ) will always be nonzero. Expressed in this
for Fµ,k = 0, ±1. We can thus decouple the primary contri- way, first-order Ising-Ising interactions are decoupled by en-
bution of both of these first-order effects with the same pulse suring that for every pair of axes µ and ν, the cumulative oc-
sequence block, simply by arranging the frames to form fast currences of ν frames before and after each µ frame are equal,
spin echoes. These fast spin echoes are illustrated in Fig. 1(a). i.e. the appearance of the two frames is balanced.
To see this, first note that the commutator pre-factor implies Based on this result, we find that a simple motif is to per-
that there will be a nonzero contribution only when µ ≠ ν. form a mirror symmetrization [30] of the frames within each
With a spin echo block, the contribution from Fµ,k flips and block, as illustrated in Fig. 1(b). Note that since the coefficient
ν ν of the Ising contribution is identical regardless of the sign of
cancels, while the sum F<k or I<k is along a different axis and
thus remains unchanged during this time. Moreover, the av- the frame, only the relative frame ordering matters and not
erage zeroth order disorder Hamiltonian over the entire pulse the sign. If within each block the frames are balanced along
sequence will also vanish due to the spin echo blocks. Thus, the x̂, ŷ and ẑ directions, and mirror symmetrization in terms
the main term rules in conditions 1 and 2 are satisfied in Tab. I. of ordering is performed, then using the pulse cycle decou-
In column 4 of Tab. I, we derived additional corrections to pling principle, the first-order Ising-Ising contribution will be
the expressions, originating from the pulse-induced overlaps cancelled. Here, contrary to global mirror symmetrization, we
in Eq. (25). Interestingly, we find again that the different terms find that symmetrization within local blocks can also be a use-
share some common structures, where they can be related to ful tool to effectively cancel certain first-order contributions.
each other simply by replacing Fν,k by ∣Fν,k ∣. Moreover, we The finite pulse duration correction terms for the first-order
find that the finite pulse duration corrections for first-order Ising-Ising contribution are shown in the fourth column of the
disorder-disorder terms are proportional to that of rotation an- Tab. I. They resemble the other correction terms, but with ad-
gle errors at zeroth-order [9], making it automatically satisfied ditional absolute value signs, and can be easily incorporated
if the latter has been incorporated into sequence design. as decoupling rules in a similar fashion.
 8

IV.4. Dipole Cancellation and Row Balancing for V. DETAILED SEQUENCE DESIGN PROCEDURE
Heisenberg-Related Terms
We now utilize the preceding insights to design higher-
Let us now examine terms related to the Heisenberg inter- order pulse sequences for various applications, focusing on
action S⃗ ⋅ S.
⃗ As noted in Eq. (11), the Heisenberg interaction the case of interacting spin ensembles dominated by on-site
is invariant under frame transformations. Thus, the first-order disorder [23, 28]. The result is a pulse sequence that decou-
expression resulting from the Heisenberg term and a different ples all zeroth-order and first-order contributions in the Mag-
term will be the commutator between a constant term (Heisen- nus expansion, and is robust against disorder to second or-
berg), and a term that depends on the frame transformations der, which we name DROID-R2D2 (Disorder RObust Interac-
(other). The inclusion of the time integrals then result in the tion Decoupling - Robust To Disorder 2nd order), and a pulse
expressions in row 4-6 of Tab. I, where there are no first-order sequence that achieves similar results but also has interest-
contributions between two Heisenberg Hamiltonians because ing AC field sensing capabilities [32]. These pulse sequences
they are identical and thus commute. In this case, because were crucial for a variety of our recent experiments in dynami-
the Heisenberg Hamiltonian is invariant, the frame length ex- cal decoupling, quantum metrology [32], and Hamiltonian en-
tension becomes exact, and we do not need to include any gineering [33]. We will illustrate the complete design proce-
additional finite pulse corrections in the table. dure in detail, and mention a few practical tricks to improve
To explore in more detail what the resulting rules mean, let the efficiency of sequence screening and to examine larger de-
us first recall the expressions for cancelling disorder at zeroth sign spaces of pulse sequences.
order. Here, we found that in order for disorder to be cancelled
at zeroth order, we require the average disorder along each 1. Choose target decoupling rules
axis to vanish, i.e.
The first step is to determine the set of decoupling rules that
n
4 should be satisfied by the desired sequence. The choice of this
∑ Fµ,k (τk + τp ) = 0. (26) set is usually informed by several factors: First, the target ap-
k=1 π
plication may influence which terms need to be decoupled.
For example, if we wish to study many-body dynamics in a
A useful physical analogy to interpret this expression is to as- disordered system, we may wish to preserve the disorder term
sociate a positive(negative) charge with Fµ,a = +1(−1). The while engineering interactions. Alternatively, if we are inter-
zeroth-order decoupling condition then dictates that the aver- ested in quantum sensing, then there may be additional design
age charge is 0. rules that are imposed to maximize sensitivity.
Generalizing the analogy to first-order terms, the first-order Second, the experimental system characteristics may in-
term of a given Hamiltonian contribution with the invari- form which contributions are most important to decouple.
ant Heisenberg Hamiltonian will be proportional to the given As an example, dense electronic spin ensembles, such as
Hamiltonian, weighted by its location in time in the sequence. nitrogen-vacancy (NV) centers and nitrogen (P1) defects in
This is because of the integration limits in Eq. (18), where the diamond, or rare earth ions, typically have much larger disor-
relative ordering of the time variable values of the two Hamil- der than interactions [23, 34, 35]. Thus, it is much more im-
tonians determines the sign of the expression. Furthering the portant to address disorder-related effects to higher-order than
electromagnetic analogy, this results in a distance weighting interaction-related effects. The relative importance of differ-
factor from the center of the pulse sequence timing. Thus, ent contributions can be made more quantitative by using our
the first-order expression resembles the expression of a dipole, expressions for various terms to estimate the typical total mag-
with charge given by Fµ,k at each time point and distance be- nitude of each of the Hamiltonian terms.
ing the distance in time to the center of the sequence. Can- Finally, for a given pulse sequence, we can also diagnose
celling this contribution requires the net dipole along each axis the dominant residual term by examining a cluster of a few
to vanish, as illustrated in Fig. 1(c). We note that this intuition spins, typically two or three, and computing the exact unitary
was key to improving decoupling pulse sequence performance for a set of disorder and interaction values. Taking the ma-
in Sec. V.1, and led to insights regarding the dichotomy be- trix log of the unitary yields the exact effective Hamiltonian,
tween AC field sensing and decoupling for existing pulse se- and performing polynomial fits of the dominant terms with
quences described in Ref. [32]. respect to the disorder and interaction strengths informs us
Similarly, we can also analyze the expression for cross- which type of contribution is the largest, as well as the order
terms between Ising interactions and Heisenberg interactions, at which it contributes in the Magnus expansion. For exam-
simply by replacing the general charge by a non-negative ple, in Fig. 4(a), we find that the dominant error terms for
charge value (∣Fµ,k ∣). With only positive charges, the con- the existing DROID sequence from Ref. [23] are the XZ and
dition can also be alternatively viewed geometrically as bal- ZX components of the Hamiltonians, when decomposed in the
ancing frame weights in each row; as illustrated in Fig. 1(d), Pauli basis. In Fig. 5(a,b), we find that the dominant scaling
one can imagine a fulcrum placed at the middle of a sequence, of this term is linear in both the disorder strength and inter-
and for a given axis, placing a weight whenever the frame is action strength, suggesting that it originates from a first-order
along this axis (regardless of it being a positive or negative cross-term between them. This motivated us to systematically
frame); the rule then becomes that the row would balance. include decoupling rules that target this effect.
 9

10
-5
(a) 10
-6 (b) 20 10 -6
(a) DROID 20
5
Error Magnitude

XZ Magnitude
XZ Magnitude
10 10
DROID
0
R2D2
0 0

-5 0 0.02 0.04
0 2 4
X Disorder (2 MHz) Interaction (2 MHz)
Y I
Z Z (c) (d)
Y 10 -5 10 -6
10 -5 I X 2 5
5

XZ Magnitude

YI Magnitude
Error Magnitude

0
(b) DROID-R2D2 0
-2
0 -5
-4

-6 -10
-5 -0.5 0 0.5 1 1.5 -0.5 0 0.5 1 1.5
XXZ Parameter XXZ Parameter
X
Y I
Z Z FIG. 5. Error scaling for various pulse sequences. (a,b) Magnitude
Y of XZ error term in the effective Hamiltonian as a function of disor-
I X
der (a) and interactions (b) for the DROID sequence (blue circles)
and the DROID-R2D2 sequence (red squares, much smaller errors).
The dominant scaling is linear for DROID. (c,d) Magnitude of XZ
FIG. 4. Comparison between decoupling sequence performance. (c) and YI (d) error terms in the effective Hamiltonian for differ-
We compare the effective error Hamiltonian extracted from the ma- ent XXZ Hamiltonians, where DROID again shows a systematically
trix log of a two-spin unitary for disorder strengths 2π × (4, 0) MHz larger error and stronger Hamiltonian dependence.
and interactions 2π ×35 kHz, for the pulse sequences DROID (a) and
DROID-R2D2 (b). The ideal Heisenberg portion of the Hamiltonian
has been subtracted out. Bars represent coefficients of the effective
Hamiltonian decomposed in the Pauli basis. We find that the error is a set of imposed decoupling rules. The number of possible
much smaller for our new sequence DROID-R2D2. frame sets without any additional constraints is combinatori-
ally large. For example, even a sequence consisting of 12 free
evolution times connected by π/2 pulses, including interme-
In practice, we search for sequences by randomly enumer- diate frames for the finite pulse durations (e.g. the frame half
ating those of a fixed length that satisfy a chosen set of rules way through π pulses or composite π/2 pulses), admits ap-
(see below for a description of how to efficiently enforce proximately 423 ≈ 1014 distinct pulse sequences (each frame
rules). We iterate the preceding error diagnosis step several is connected to 4 other frames by π/2 pulses, and the first
times by identifying the dominant contributions for typical frame is fixed to be +ẑ). However the vast majority of these
pulse sequences, and adding in new rules to fully decouple sequences will not satisfy our rules. Therefore it is essential
them. Each addition of a new dominant rule eliminates the to enumerate only sequences that satisfy them.
most poorly performing sequences, and increases the proba- For the disorder-dominated interacting NV ensembles we
bility of enumerating a sequence with high coherence time; work with, we choose to impose the following structures to
see Fig. 2 of the accompanying paper [32]. efficiently pre-screen pulse sequences: we require that all
Following this procedure leads us to include the follow- frames, including both free evolution and pulse frames, come
ing decoupling rules for our disorder-dominated NV center in spin echo pairs, in order to echo out disorder on the fastest
ensemble: decoupling of all zeroth-order conditions, as de- possible timescale. In addition, we require an equal num-
scribed in Ref. [9]; decoupling of all first-order conditions ber of elements along each row, so as to symmetrize inter-
involving at least one factor of disorder, including disorder- actions. Finally, we impose the “dipole” rules for first-order
disorder cross-terms, disorder-Ising and disorder-Heisenberg disorder-Heisenberg cross-terms, by requiring there to be an
cross-terms, for both free evolution times and pulses; second- equal number of +− and −+ spin echoes.
order disorder-disorder-disorder cross-terms, for both free In order to directly restrict the search space to candidate
evolution times and pulses. frame sequences that satisfy the above rules, we separately
enumerate the locations of X, Y and Z spin echo pairs, and
2. Efficiently construct candidate frame sets enumerate the echo ordering signs (i.e. whether the echo
frames has the ordering +− or −+) of both free evolution
With the set of target decoupling rules in hand, we now dis- frames and finite pulse frames. We then combine these pieces
cuss how to efficiently enumerate pulse sequences satisfying of information to generate candidate frame sequences, impos-
 10

ing the additional constraint that each frame must be distinct primarily cross terms between disorder and first-order Heisen-
from the two neighboring frames, to ensure that a π/2 pulse is berg interactions, the residual errors when examining the ef-
applied and the pulse error calculation is accurate. fective Hamiltonian are much reduced (Fig. 4(b)).
Moreover, we can adapt this pulse sequence to perform
3. Screen frame sets using decoupling rules Hamiltonian engineering by adjusting the frame durations
along the x̂, ŷ and ẑ axes [33], resulting in a tunable inter-
Having generated candidate frame sets that already have a action Hamiltonian
number of rules enforced by construction, we now proceed
1+λ x x 1 − 2λ z z
to screen through them by applying the remaining decoupling HXXZ = ∑ Jij [ (Si Sj + Siy Sjy ) + Si Sj ],
rules. In order to speed up the screening process, the key in- ij 3 3
sight is to transform the original rules into a vectorized form, (27)
such that fast matrix computation can be performed, signifi-
cantly reducing the run time. This is achieved by labeling the where λ is a coefficient that tunes the XXZ Hamiltonian. We
frames as 1 to 6 for +x,+y,+z,−x,−y,−z, and noting that the find that our techniques also significantly reduce the error in
rules become simple cumulative sums of index matching re- engineering a wide range of generic XXZ Hamiltonians. In
sults when expressed in this fashion. We then further simplify Fig. 5(c,d), we see that both two-body and single-body im-
them based on known decoupling structures (e.g. the rapid perfection terms are much smaller across a wide range of
spin echoes built into the sequence). Moreover, when evalu- different Hamiltonians, which can help improve the fidelity
ating some of the higher-order expressions in full generality, of Hamiltonian engineering and reduce systematic artifacts.
we can keep track of the cumulative integral of lower-order These techniques can be readily generalized to engineer XYZ
terms [36], which reduces the time complexity of computing Hamiltonians, with different coefficients in front of each term,
many such terms to linear in the sequence length, rather than or even more complex many-body Hamiltonians.
a higher polynomial scaling. The same techniques can also be used to design pulse se-
quences for improved quantum sensing, as we explain in more
4. Verify performance and further optimization detail in Ref. [32]. The key insight is that current pulse se-
quences for quantum sensing [23], which periodically flip the
To optimize the performance of the pulse sequences, we spin along each axis with the same frequency as the target
further symmetrize the pulse sequence to reduce higher-order signal, will always result in a violation of the “net dipole can-
error contributions. Here, for dynamical decoupling, we em- cellation” rule in Fig. 1(c) for first-order disorder-Heisenberg
ploy the symmetrization used in Ref. [9], where the frames are terms. This imposes a fundamental trade-off between sensitiv-
repeated twice, but the frame ordering is reversed and sign of ity and decoupling quality for current pulse sequences. With
all frames flipped in the second repetition. For the quantum this insight from higher-order decoupling rules, we are able to
metrology pulse sequences designed here and in Ref. [32], design the new pulse sequence DIRAC2 (DIsorder Robust AC
this symmetrization will affect the magnetic field sensitiv- sensing with period 2), as shown in Fig. 6(c), which circum-
ity, and consequently we employ a mirror-symmetrization in- vents this issue by targeting a sensing signal half the frequency
stead, where the frame ordering is reversed but the sign is not of frame flipping, thereby fully cancelling all first-order Mag-
flipped in the second repetition. nus contributions while also increasing the rate of spin echo
Finally, we numerically simulate the performance of these decoupling, leading to better performance. See Ref. [32] for a
pulse sequences to identify the ones with the longest decou- more detailed description.
pling timescales. Effective Hamiltonian extraction using the
matrix log of the unitary can identify dominant error terms
for the pulse sequences employed, and the whole design pro- VI. DISCUSSION AND CONCLUSION
cedure can be repeated with an improved rule set. For the
above final set of decoupling rules, we no longer find a single We have developed a general framework for dynamical
contribution that dominates over the others, instead seeing a Hamiltonian engineering that includes higher-order consid-
competition between several different contributions. erations. Contrary to the naive expectation, we found that
many higher-order decoupling conditions can still have sim-
ple, intuitive interpretations, particularly when the pulse se-
V.1. Resulting Pulse Sequences quence is designed to have certain structures in it. We an-
alytically derived a number of decoupling rules for higher-
Using the decoupling rules described above, we designed order contributions, and used them to design robust pulse se-
pulse sequences for dynamical interaction decoupling, many- quences in disorder-dominated systems for dynamical decou-
body physics, and quantum metrology. pling, Hamiltonian engineering, and quantum sensing, signif-
For dynamical decoupling and Hamiltonian engineering, icantly improving upon state-of-the-art pulse sequences.
one of the best pulse sequences we identified, DROID-R2D2, While we have focused on the application of our tech-
is shown in Fig. 6(b). We find that compared to the previ- niques to the case of electronic spin ensembles, where dis-
ous best pulse sequence DROID [23], as shown in Fig. 6(a), order is much larger than spin-spin interactions, we believe
that had significant residual first-order cross terms (Fig. 4(a)), that our techniques can be applied to disparate systems such
 11

(a) DROID

Fx
Fy
Fz
(b) DROID-R2D2

Fx
Fy
Fz
px p/2x py p/2y
(c) DIRAC2

Fx
Fy
Fz

FIG. 6. Details of higher-order pulse sequences. Pulse representation (top) and frame representation (bottom) of various pulse sequences
found using the higher-order design rules developed in this paper. (a) DROID [23], the previous best pulse sequence for dynamical decoupling
and quantum metrology in disorder-dominated interacting spin ensembles. (b) DROID-R2D2, a new pulse sequence designed for dynamical
decoupling and Hamiltonian engineering, in which all first-order contributions are fully cancelled and disorder is cancelled to second order.
See Ref. [33] for an application of this pulse sequence to many-body XXZ dynamics. (c) DIRAC2, a new pulse sequence that has similar
characteristics, but is designed for improved quantum sensing, see Ref. [32] for details and experimental demonstration of the improved
sensitivity compared to the best known sequences.

as NMR [8, 37], simply by changing which rules are empha- ing pulse sequence design, with broad applications in dynam-
sized and included at higher-order. It may also be interesting ical decoupling, quantum many-body physics, and quantum
to further extend the techniques to even higher-order than the metrology.
ones that we have considered here [38], or to examine alter-
native expansions beyond the Magnus expansion [39, 40]. In
our formalism, the contributions from higher-order terms are
decomposed into an operator commutation portion, and a por-
ACKNOWLEDGEMENTS
tion that relates to the frame matrix and pre-factors. This also
makes the extension to higher-spin systems relatively straight-
forward, and can be combined with recent methods for robust We thank J. Choi, A. Douglas, H. Gao, N. Maskara, P. Peng,
Hamiltonian engineering with higher-spin systems [41, 42]. M. Yu for helpful discussions. This work was supported in
With these further improvements, we believe that our frame- part by CUA, HQI, NSSEFF, ARO MURI, DARPA DRINQS,
work presents a key tool for advanced Hamiltonian engineer- Moore Foundation GBMF-4306, NSF PHY-1506284.

[1] T. D. Ladd, F. Jelezko, R. Laflamme, Y. Nakamura, C. Monroe, [9] J. Choi, H. Zhou, H. S. Knowles, R. Landig, S. Choi, and M. D.
and J. L. O’Brien, Nature 464, 45 (2010). Lukin, Physical Review X 10, 031002 (2020).
[2] L. M. K. Vandersypen and I. L. Chuang, Reviews of Modern [10] L. Viola, E. Knill, and S. Lloyd, Physical Review Letters 82,
Physics 76, 1037 (2005). 2417 (1999).
[3] C. L. Degen, F. Reinhard, and P. Cappellaro, Reviews of Mod- [11] K. Khodjasteh and D. A. Lidar, Physical Review Letters 95,
ern Physics 89, 035002 (2017). 180501 (2005).
[4] I. M. Georgescu, S. Ashhab, and F. Nori, Reviews of Modern [12] G. S. Uhrig, Physical Review Letters 98, 100504 (2007).
Physics 86, 153 (2014). [13] G. A. Álvarez, D. Suter, and R. Kaiser, Science 349, 846
[5] I. Bloch, J. Dalibard, and S. Nascimbène, Nature Physics 8, (2015).
267 (2012). [14] K. X. Wei, C. Ramanathan, and P. Cappellaro, Physical Review
[6] J. S. Waugh, L. M. Huber, and U. Haeberlen, Physical Review Letters 120, 070501 (2018).
Letters 20, 180 (1968). [15] K. X. Wei, P. Peng, O. Shtanko, I. Marvian, S. Lloyd, C. Ra-
[7] D. P. Burum and W. K. Rhim, The Journal of Chemical Physics manathan, and P. Cappellaro, Physical Review Letters 123,
71, 944 (1979). 090605 (2019).
[8] D. G. Cory, J. B. Miller, and A. N. Garroway, Journal of Mag- [16] D. Hayes, S. T. Flammia, and M. J. Biercuk, New Journal of
netic Resonance (1969) 90, 205 (1990). Physics 16, 083027 (2014).
 12

[17] A. Ajoy and P. Cappellaro, Physical Review Letters 110, and M. D. Lukin, Physical Review Letters 121, 023601 (2018).
220503 (2013). [29] D. Cory, J. Miller, R. Turner, and A. Garroway, Molecular
[18] S. Choi, N. Y. Yao, and M. D. Lukin, Physical Review Letters Physics 70, 331 (1990).
119, 183603 (2017). [30] P. Mansfield, Journal of Physics C: Solid State Physics 4, 1444
[19] H. Haas, D. Puzzuoli, F. Zhang, and D. G. Cory, New Journal (1971).
of Physics 21 (2019), 10.1088/1367-2630/ab4525. [31] E. L. Hahn, Physical Review 80, 580 (1950).
[20] W. Rose, H. Haas, A. Q. Chen, N. Jeon, L. J. Lauhon, D. G. [32] See Accompanying Paper.
Cory, and R. Budakian, Physical Review X 8, 011030 (2018). [33] L. S. Martin, H. Zhou, N. T. Leitao, N. Maskara, O. Makarova,
[21] C. P. Slichter, Principles of magnetic resonance, Vol. 1 H. Gao, Q.-Z. Zhu, M. Park, M. Tyler, H. Park, S. Choi,
(Springer Science & Business Media, 2013). and M. D. Lukin, arXiv preprint arXiv:2209.09297 (2022),
[22] M. Mehring, Principles of high resolution NMR in solids 10.48550/arxiv.2209.09297.
(Springer Science & Business Media, 2012). [34] C. Zu, F. Machado, B. Ye, S. Choi, B. Kobrin, T. Mittiga,
[23] H. Zhou, J. Choi, S. Choi, R. Landig, A. M. Douglas, J. Isoya, S. Hsieh, P. Bhattacharyya, M. Markham, D. Twitchen, A. Jar-
F. Jelezko, S. Onoda, H. Sumiya, P. Cappellaro, H. S. Knowles, mola, D. Budker, C. R. Laumann, J. E. Moore, and N. Y. Yao,
H. Park, and M. D. Lukin, Physical Review X 10, 031003 Nature 597, 45 (2021).
(2020). [35] B. Merkel, P. Cova Fariña, and A. Reiserer, Physical Review
[24] S. Choi, J. Choi, R. Landig, G. Kucsko, H. Zhou, J. Isoya, Letters 127, 030501 (2021).
F. Jelezko, S. Onoda, H. Sumiya, V. Khemani, C. von Keyser- [36] N. Khaneja, T. Reiss, C. Kehlet, T. Schulte-Herbrüggen, and
lingk, N. Y. Yao, E. Demler, and M. D. Lukin, Nature 543, 221 S. J. Glaser, Journal of Magnetic Resonance 172, 296 (2005).
(2017). [37] P. Peng, X. Huang, C. Yin, L. Joseph, C. Ramanathan,
[25] J. Zhang, G. Pagano, P. W. Hess, A. Kyprianidis, P. Becker, and P. Cappellaro, arXiv preprint arXiv:2102.13161 (2021),
H. Kaplan, A. V. Gorshkov, Z. X. Gong, and C. Monroe, Nature 10.48550/arxiv.2102.13161.
551, 601 (2017). [38] M. Hohwy and N. C. Nielsen, The Journal of Chemical Physics
[26] N. H. Lindner, G. Refael, and V. Galitski, Nature Physics 7, 106, 7571 (1997).
490 (2011). [39] K. R. Mote, V. Agarwal, and P. Madhu, Progress in Nuclear
[27] U. Haeberlen and J. S. Waugh, Physical Review 175, 453 Magnetic Resonance Spectroscopy 97, 1 (2016).
(1968). [40] E. S. Mananga, Physics Reports 608, 1 (2016).
[28] G. Kucsko, S. Choi, J. Choi, P. C. Maurer, H. Zhou, R. Landig, [41] H. Zhou et al., in preparation.
H. Sumiya, S. Onoda, J. Isoya, F. Jelezko, E. Demler, N. Y. Yao, [42] N. Leitao et al., in preparation.

Appendix A: Conventions

See Tab. II for a summary of the conventions employed in this manuscript.

Appendix B: Derivation of First-Order Magnus Formalism

To develop the full expression at first order, we extend the formalism developed in Ref. [9]. To keep the expressions fully
general, we do not restrict to a specific qubit Hamiltonian here, and specialize to the dipolar Hamiltonian only in the following
sections.
Following Ref. [9], we can separate the evolution into free evolution periods and evolution during pulses. We will write the
coefficient of a given operator Oα during the k-th free evolution period as cα (t) = cα,k , and during the π/2 pulse after the k-th
free evolution period as

t t t
cα (t) = c̀α,k ( ) + qα,k,k+1 ( ) + ćα,k+1 ( ). (B1)
r r r

Here, r is the rate of angular precession under the applied pulses, and the rotation angles during the π/2 pulse are given by
θ = t/r. The first term c̀α,k (θ) describes the finite pulse duration contribution from the k-th frame that precedes the pulse, while
ćα,k+1 (θ) describes the contribution from the (k + 1)-th frame that follows the pulse. qα,k,k+1 (θ) is an additional cross-term
between the two frames that arises for certain types of interaction terms. Note that similar to Ref. [9], in our pulse sequence
composed of π/2 pulses and π pulses, we treat each π pulse as a combination of two π/2 pulses with zero free evolution time in
between.
As a concrete example to illustrate these terms, let us consider a rotation that transformed the S z operator into S x , i.e.
S̃ (θ) = cos θS z + sin θS x . For an Ising interaction HI = JSiz Sjz , the time-dependent operator would be
z

H̃I (t) = J[cos2 θSiz Sjz + sin θ cos θ(Six Sjz + Siz Sjx ) + sin2 θSix Sjx ]. (B2)

The three terms in the parenthesis correspond to the c̀α,k (θ), qα,k,k+1 (θ) and ćα,k+1 (θ) terms, respectively.
 13

T Floquet period
tk Midpoint time of a the k-th free evolution period
τk Duration of the k-th free evolution period
τp Duration of π/2 pulses
Hdip Dipole Hamiltonian
H̃(t) Interaction picture Hamiltonian
H̃k Interaction picture Hamiltonian during the k-th free evolution period
Heff Time-independent effective Hamiltonian
(k)
H k-th order Magnus term for the effective Hamiltonian
hi On-site disorder strength for spin i
Ji,j Interaction strength between spins i and j
Pk k-th global spin rotation pulse
Uc (t) Single spin rotation unitary due to the control field
U (T ) Unitary operator for the evolution over one full Floquet cycle
S̃ (t)
z
Interaction picture S z vector at time t under an ideal sequence
Sµ Spin basis (e.g. Pauli spin operators)
Fµ (t) The S µ coefficient of S̃ z at time t
Fµ,k The S µ coefficient of S̃ z during free evolution frame k
µ
F<k Accumulated disorder through pulse k, ∑j<k Fµ,j (τj + π4 τp )
µ
F>k Accumulated disorder after pulse k, ∑j>k Fµ,j (τj + π4 τp )
µ
I<k Accumulated Ising interaction through pulse k, ∑j<k ∣Fµ,j ∣(τj + τp )
µ
I>k Accumulated Ising interaction after pulse k, ∑j>k ∣Fµ,j ∣(τj + τp )
ν,ρ
F<k Accumulated first-order disorder-disorder effect through pulse k, ∑kl=1 Fν,l (τl + π4 τp )(F<l
ρ
− F>l
ρ
)
µ
k=1 Fµ,k (τk + π τp )
Total zeroth-order disorder effect, ∑n 4
F
µ
I k=1 ∣Fµ,k ∣(τk + τp )
Total zeroth-order Ising interaction, ∑n
ν,ρ
k=1 Fν,k (τk + π τp )F<k
ρ
F Total first-order disorder-disorder effect ∑n 4

Oα Operator basis for interaction-picture Hamiltonian

cα (t) Time-dependent coefficients of Oα in the Hamiltonian
cα (t) Accumulation of cα (t1 ) over t1 ∈ [0, t]
¡
cµ,k Discrete version of cµ (t) in the kth free evolution frame
ćµ,k (θ) cµ (t) during the pulse preceding the kth free evolution time, as a function of the rotation angle θ ∈ [0, π/2]
c̀µ,k (θ) cµ (t) during the pulse following the kth free evolution time, as a function of the rotation angle θ ∈ [0, π/2]
r Angular rotation rate during a pulse
Cµ,k Total accumulation of cµ (t) over t ∈ [tk − τk /2 − τp , tk + τk /2 + τp ]
cµ,k Accumulated coefficient, ∑l<k Cµ,l
¡
qα,k,k+1 (θ) Cross terms during pulses that involve both preceding and following frames k and k + 1
Af ree First-order contribution from free evolution frames
(1)
Cα,β,k Frame k’s first order effect with itself
Pk,k+1,α,β Finite pulse correction for kth pulse

TABLE II. Summary of conventions employed in this paper.

 14

With this representation in hand, we proceed by rewriting the integral in Eq. (21) as a summation over the distinct blocks. Let
us examine the first term, which integrates all cβ terms occurring temporally before cα :
T t1
A=∫ dt1 cα (t1 ) ∫ dt2 cβ (t2 ) (B3)
0 0

To compute this, we first define the integral of the coefficient of a given frame, including its finite pulse duration effects:
π/2 τk π/2
Cα,k = r ∫ ćα,k (θ)dθ + ∫ cα,k dt + r ∫ c̀α,k (θ)dθ. (B4)
0 0 0

This can be viewed as a simple extension of the effective free evolution time.
We can then decompose the inner integral in Eq. (B3) into three parts (ignoring additional contributions from qα,k,k+1 -terms
for now): a contribution from previous, non-overlapping free evolution times, together with their surrounding pulses (Af ree ); a
contribution from integrating both time variables within the same free evolution period, corresponding to first-order contributions
(1)
within the same frame (Cα,β,k ); and a further correction arising from the pulse overlaps of neighboring free evolution times
(Pk,k+1,α,β ). This is illustrated in Fig. 7.
More concretely, the first term describes contributions where t1 lies within the k-th frame, and t2 originates from an earlier
frame. As most of these contributions will be temporally non-overlapping, we can factorize these contributions as
n k−1
Af ree = ∑ Cα,k ∑ Cβ,j . (B5)
k=1 j=1

The next term describes contributions where both t1 and t2 come from the k-th frame, with t2 < t1 ∈ [tk−1 + τk−1
2
, tk+1 − τk+1
2
],
i.e. the first-order Magnus contribution of a frame with itself. We can explicitly write this as
τk+1
tk+1 − t1
(1) 2
Cα,β,k = ∫ τ
cα (t1 )dt1 ∫ τk−1
cβ (t2 )dt2 . (B6)
tk−1 + k−1
2 tk−1 + 2

Note that this term is usually zero in our case, as the commutator [Oα , Oβ ] vanishes when α = β, and otherwise cα , cβ are both
non-zero within the same free evolution period only when they originate from different types of non-commuting Hamiltonians,
e.g. one coming from local S z disorder, and the other coming from Heisenberg interactions.
Finally, we have additional corrections Pk,k+1,α,β that arise from the overlap in terms due to the pulses: for the k-th frame,
we overcounted the overlap contribution with the previous (k − 1)-th frame by assuming that the k-th frame came completely
after the (k − 1)-th frame, but undercounted the overlap contribution with the next (k + 1)-th frame. Explicit calculation shows
that this results in a first order pulse correction that is related to both the preceding and subsequent frame:
π/2 θ1 π/2 π/2
Pk,k+1,α,β = ∫ c̀α,k (θ1 )rdθ1 ∫ ćβ,k+1 (θ2 )rdθ2 − ∫ ćα,k+1 (θ1 )rdθ1 ∫ c̀β,k (θ2 )rdθ2 . (B7)
0 0 0 θ1

Thus, neglecting all terms that directly dependent on multiple pulses (see q-terms below), we can express the integral in a
clean manner as
n−1 n
(1)
A = Af ree + ∑ Pk,k+1,α,β + ∑ Cα,β,k . (B8)
k=1 k=1

For certain interaction terms such as the Ising Hamiltonian, there is an additional contribution we need to keep track of, the
qα,k,k+1 (θ) terms described in Eq. (B2). These terms come from the fact that the Ising interaction transforms as the square of
the frame coefficients, introducing additional cross-terms when expanding the square. These terms are ignored in the main text,
as they are negligible for our disorder-dominated system, but we will analyze them in more detail here.
We can perform a similar decomposition of the terms as above, now adding in the contributions from the q-terms. We can
treat the q-terms as a special type of free evolution frame, and decompose the sum into the three types again, this time keeping
track also of whether the other term is a q-term or a regular free evolution period.
Similar to Eq. (B5), we can evaluate the first-order contributions involving a single q-term and a single free evolution frame
as
n k−1 n k
QAf ree = ∑ Cα,k ∑ Qβ,j,j+1 + ∑ Qα,k,k+1 ∑ Cβ,j , (B9)
k=1 j=1 k=1 j=1
 15

where
π/2
Qα,k,k+1 = ∫ qα,k,k+1 (θ)dθ (B10)
0

is the integral of the q-term during a given pulse.

In analogy to the corrections Pk,k+1,α,β found above, we also have similar corrections here
π/2 θ1 π/2 π/2
QPk,k+1,α,β = ∫ qα,k,k+1 (θ1 )rdθ1 ∫ ćβ,k+1 (θ2 )rdθ2 − ∫ qα,k,k+1 (θ1 )rdθ1 ∫ c̀β,k (θ2 )rdθ2
0 0 0 θ1
π/2 θ1 π/2 π/2
+∫ c̀α,k (θ1 )rdθ1 ∫ qβ,k,k+1 (θ2 )rdθ2 − ∫ ćα,k (θ1 )rdθ1 ∫ qβ,k−1,k (θ2 )rdθ2 . (B11)
0 0 0 θ1

Finally, we also have corrections coming from the first-order contributions between q-terms at different times and in the same
pulse
n k−1 n π/2 π/2
Qself = ∑ Qα,k,k+1 ∑ Qβ,l,l+1 + ∑ ∫ qα,k,k+1 (θ1 )rdθ1 ∫ qβ,k,k+1 (θ2 )rdθ2 . (B12)
k=1 l=1 k=1 0 θ1

Putting all of this together, the final, complete expression for first-order terms is
n−1 n n
(1)
A = Af ree + ∑ Pk,k+1,α,β + ∑ Cα,β,k + QAf ree + Qself + ∑ QPk,k+1,α,β (B13)
k=1 k=1 k=1

Appendix C: Derivation of First-Order Cancellation Rules

We will now apply the preceding general calculations to specific first-order terms, in order to derive first-order decoupling
rules. As we shall see, in many cases of interest, a lot of the terms in Eq. (B13) will drop out, resulting in simple expressions.

1. Disorder-Disorder Rules

Let us start with first-order disorder-disorder contributions, involving commutators between disorder at different times. Since
(1)
this Hamiltonian involves only single-qubit terms, there will be no q-terms. Furthermore, there are no Cα,β,k terms, as the
operator in each frame commutes with itself. We thus have
n−1
Adis−dis = Af ree + ∑ Pk,k+1,α,β . (C1)
k=1

Examining the transformation of the operators for different frames, we have

cα,k → Fµ,k , (C2)
ćα,k (θ) → Fµ,k sin(θ), (C3)
c̀α,k (θ) → Fµ,k cos(θ), (C4)
4τp
Cα,k = Fµ,k (τk + ). (C5)
π
Plugging this into the preceding definitions of the individual terms, we find
n 4τp k−1 4τp
Af ree = ∑ Fµ,k (τk + ) ∑ Fν,l (τl + ), (C6)
k=1 π l=1 π
π 2τp 2
Pk,k+1,dis,dis = (Fµ,k Fν,k+1 − Fµ,k+1 Fν,k )(1 − )( ) , (C7)
4 π
Further plugging this into the full expression Eq. (21) for the first-order disorder-disorder term, we arrive at the full expression
for the main term
n
4 ν 1
∑ Fµ,k (τk + τp )F<k − F̄ µ F̄ ν , (C8)
k=1 π 2
 16

𝑐̀(,& 𝑐́(,'
𝜏$
%
𝐶!,#,$

𝑐̀(,% 𝑐́(,&
𝜏&

%
𝐶!,#,&
𝑐̀(,$ 𝑐́(,%
𝜏'

%
𝐶!,#,'
𝑐̀(,# 𝑐́(,$

𝐴!"##
𝜏%

%
𝐶!,#,%
𝛼
𝜏% 𝜏' 𝜏& 𝜏$
𝑐̀!,# 𝑐́!,$ 𝑐̀!,$ 𝑐́!,% 𝑐̀!,% 𝑐́!,& 𝑐̀!,& 𝑐́!,'

FIG. 7. Visualization of first-order cancellation rule derivation, in the absence of q-terms. The illustration shows the different components
of the first-order expression. The lower triangle of the grid is the first-order integral. In reality, the rising and falling edges of neighboring
frames will overlap, i.e. c̀α,1 is not completely before ćα,2 , which leads to the additional term Pk,k+1,α,β .

and the finite pulse correction

π 2τp n
(1 − )( ) ∑ Fµ,k Fν,k+1 − Fν,k Fµ,k+1 , (C9)
4 π k=1
as described in Tab. I.
Based on these expressions, we can formulate relatively simple rules for their cancellation in sequence design. The expression
Pk,k+1,α,β involves a term that can be rewritten as F⃗k × F⃗k+1 , and thus has the same conditions for cancellation as zeroth-order
rotation angle errors [9].
Due to the rapid spin echo structure found in many decoupling sequences for disorder-dominated systems, e.g. DROID-60 in
Ref. [23], the majority of terms in Af ree are also cancelled in the inner sum, and the only contribution remaining is from the
commutator between a spin echo pair and the intermediate pulse frame that the π pulse uses. To give a concrete example of this
remaining contribution, consider a sequence of two π pulses around X, which implements the following frame transformations
+Z → +Y → −Z → −Y , with +Z and −Z being longer free evolution frames, and +Y and −Y being shorter frames with zero
free evolution time and only pulse effects. The first-order contribution from this will then be proportional to the commutator
between Z and Y , and changes sign both when we flip the sign of one of the operators (e.g. +Z → −Y → −Z → +Y ), as well as
when we switch the order of the operators (e.g. +Y → +Z → −Y → −Z). Thus, this term has the same transformation properties
as a rotation angle error that acts only within such spin echo blocks.

2. Disorder-Heisenberg Rules

The next term we consider is the first-order disorder-Heisenberg contribution, which was the dominant imperfection in the
previous DROID-60 sequence [23] and key to the design of improved sensing sequences such as DIRAC2 [32].
As shown in Eq. (21), we can choose an index ordering where disorder is after Heisenberg interactions, such that α is a
disorder index and β is a Heisenberg interaction index. The case where both are Heisenberg indices gives zero contribution, as
 17

the operator terms are equal to the fixed Heisenberg Hamiltonian and hence commute. As the Heisenberg interaction is invariant
under frame transformations, the coefficients can be chosen to take a particularly simple form:
cβ,k → 1, (C10)
ćβ,k (θ) → 1, (C11)
c̀β,k (θ) → 0, (C12)
Cβ,k = τk + τp . (C13)
Plugging these into the preceding expressions, we find
n 4τp k−1
Af ree = ∑ Fµ,k (τk + ) ∑ (τj + τp ), (C14)
k=1 π j=1
π 2τp 2
Pk,k+1,α,β = Fµ,k ( − 1)( ) , (C15)
2 π
π/2 θ1 τk t1 π/2
(1)
Cα,β,k = Fµ,k [∫ sin(θ1 )rdθ1 ∫ rdθ2 + ∫ dt1 (τp + ∫ dt2 ) + ∫ cos(θ)rdθ(τp + τk )]
0 0 0 0 0

2τp 2 τ2 2τp
= Fµ,k [( ) + τk τp + k + (τp + τk )( )] . (C16)
π 2 π
We can simplify the sum of the last two contributions
(1) 1 2τp
Pk,k+1,α,β + Cα,β,k = Fµ,k (τk τp + τk2 + (2τp + τk )( ))
2 π
1 4τp
= Fµ,k (τp + τk )(τk + ). (C17)
2 π
Adding the corrections together, we get
n 4τp k−1 n 4τp
1
A = ∑ Fµ,k (τk + ) ∑ (τj + τp + τp + τk ) = ∑ Fµ,k tk (τk + ), (C18)
k=1 π j=1 2 k=1 π
where tk is the midpoint of the kth free evolution frame. The remaining term in Eq. (21) can be evaluated to be
1 T
c (T )cβ (T ) = F̄ µ , (C19)
2 ¡α ¡ 2
which combined give us the full algebraic condition for first-order disorder-Heisenberg decoupling
n
4 T
∑ Fµ,k (τk + τp ) (tk − ) . (C20)
k=1 π 2
As described in the main text and in Ref. [32], there is a relatively simple intuition for these contributions, which we visualize
using dipole balancing. If we associate a charge to each frame, with +1(−1) values of Fµ,k being a positive(negative) charge,
then the above expression corresponds to the product of charges (Fµ,k ) with their center-of-mass location (tk − T /2), which
is precisely the definition of a dipole. Thus, geometrically, we can visualize the cancellation of first-order disorder-Heisenberg
contributions as requiring that the net dipole corresponding to a frame configuration to be 0.

3. Disorder-Ising Rules

Next we move on to the Ising contributions, starting with first-order disorder-Ising terms. For this, we use cα,k from the
disorder term, and we use the following for the cβ,k terms:
cβ,k → ∣Fν,k ∣, (C21)
2
ćβ,k (θ) → ∣Fν,k ∣ sin (θ), (C22)
2
c̀β,k (θ) → ∣Fν,k ∣ cos (θ), (C23)
qβ,k,k+1 (θ) → Fν,k Fρ,k+1 sin(θ) cos(θ), (C24)
Cβ,k = ∣Fν,k ∣(τk + τp ), (C25)
τp
Qβ,k,k+1 = Fν,k Fρ,k+1 . (C26)
π
 18

Plugging these into the definitions for the individual terms, we find

n 4τp k−1
Af ree = ∑ Fµ,k (τk + ) ∑ ∣Fν,l ∣ (τl + τp ) , (C27)
k=1 π l=1
π 2 2τp 2
Pk,k+1,dis,isi = (Fµ,k ∣Fν,k+1 ∣ − Fµ,k+1 ∣Fν,k ∣) ( − )( ) . (C28)
4 3 π
(1)
The term Cα,β,k will not contribute, as the disorder and Ising Hamiltonian within the same free evolution time commute with
each other, [S µ ⊗ I, S µ ⊗ S µ ] = 0.
The algebraic conditions in the Tab. I are based on the preceding expressions, and ignore the q-terms. Combining Af ree with
the rest of the terms gives the main term:
n
4 1
∑ Fµ,k (τk +
ν
τp ) I<k − F̄ µ I¯ν . (C29)
k=1 π 2

Summing over pulses in the pulse term gives the finite pulse correction
n 2τp 2 π 2
∑( ) ( − ) (Fµ,k ∣Fν,k+1 ∣ − ∣Fν,k ∣Fµ,k+1 ) . (C30)
k=1 π 4 3

The main term will vanish with the same fast spin echo blocks as that found in the first-order disorder-disorder term, and the
finite pulse correction can be viewed as a simple generalization of, e.g. the chirality condition in Ref. [9].
We now further evaluate the q-terms. As any two adjacent frames will have different operators due to the frame change, we
will have no contribution when ν = ρ. Explicitly plugging into the above expressions gives
n k−1 τp
4
QAf ree = ∑ Fα,k (τk + τp ) ∑ Fν,j Fρ,j+1 , (C31)
k=1 π j=1 π
1 2τp 2
QPk,k+1,dis,isi = (Fµ,k Fν,k Fρ,k+1 − Fµ,k Fν,k−1 Fρ,k ) ( ) , (C32)
6 π
Qself = 0. (C33)

4. Ising-Ising Rules

We will now compute the first-order Ising-Ising term. Using the definitions of the individual terms as in the previous calcula-
tion (taking both α and β to be Ising indices), we find
n k−1
Af ree = ∑ ∣Fµ,k ∣ (τk + τp ) ∑ ∣Fν,l ∣(τl + τp )
k=1 l=1
n
ν
= ∑ ∣Fµ,k ∣ (τk + τp ) I<a , (C34)
k=1
π 2 1 2τp 2
Pk,k+1,isi,isi = (∣Fµ,k ∣∣Fν,k+1 ∣ − ∣Fµ,k+1 ∣∣Fν,k ∣) ( − )( ) (C35)
32 4 π
(1)
Similar to the disorder-disorder case and the disorder-Ising case, we have no contribution from the Cα,β,k terms. Next, we
calculate the contribution from the q-terms. Directly plugging things in
n k−1 τp n τp k
QAf ree = ∑ ∣Fµ,k ∣(τk + τp ) ∑ Fν,k Fρ,k+1 + ∑ Fµ,k Fρ,k+1 ∑ ∣Fν,l ∣(τl + τp ), (C36)
k=1 j=1 π k=1 π l=1
τp2
QPk,k+1,isi,isi = (Fµ,k Fλ,k+1 ∣Fν,k+1 ∣ − Fµ,k Fλ,k+1 ∣Fν,k ∣ + ∣Fµ,k ∣Fν,k Fρ,k+1 − ∣Fµ,k ∣Fν,k−1 Fρ,k ) , (C37)
8π
n τp k−1 τp n τ2
p
Qself = ∑ Fµ,k Fλ,k+1 ∑ Fν,l Fρ,l+1 + ∑ 2 Fµ,k Fλ,k+1 Fν,k Fρ,k+1 . (C38)
k=1 π l=1 π k=1 2π
 19

5. Ising-Heisenberg Rules

Finally, we calculate the first-order Ising-Heisenberg terms. Taking α to be the Ising terms, we can repeat the derivation from
disorder-Heisenberg rules to find that the main terms can be written as
n
T
∑ ∣Fµ,k ∣(τk + τp )(tk − ), (C39)
k=1 2

which simply replaces the Fµ,k in the original disorder-Heisenberg rule by ∣Fµ,k ∣.
Although this term is no longer cancelled by imposing a zero net dipole, we can still formulate a simple condition for it to be
cancelled: if we have “balanced” rows, in which the center of mass of each row is in the middle, then this term will be cancelled.
For example, a simple mirror symmetrization will cancel this first-order term.
For the q-terms, using the expressions found above, we can easily calculate
n τp k
QAf ree = ∑ Fµ,k Fλ,k+1 ∑(τl + τp ), (C40)
k=1 π l=1
π 2τp 2
QPk,k+1,isi,heis = ( ) Fµ,k Fλ,k+1 , (C41)
8 π
Qf ree = 0. (C42)
We note that after summing the pulse term over k, we can add the two expressions together to obtain
n τp k τp
Qtot = ∑ Fµ,k Fλ,k+1 (∑(τl + τp ) + ) . (C43)
k=1 π l=1 2
We note that the second sum is exactly the time at the middle of the pulse, similarly to the other Heisenberg rules. This tells us
that we can cancel this the same way, by balancing the center of mass for terms of the form Fµ,k Fλ,k+1 .

6. Summary of Two Qubit Commutators

We lastly summarize the general commutators of the form [Oα , Oβ ] over a basis of two-qubit operators, {Oα }α , to be defined
shortly. To this end, it is convienent to write a generic two-qubit Hamiltonian in the form of a 4 × 4 matrix A,
3
H(A) = ∑ Aµν σµ ⊗ σν , (C44)
µν=0

where we have defined σµ = (1, σ

⃗ ) as a 4-vector of Pauli operators, including the 2 × 2 identity matrix 1. It follows that a native
symmetric secular Hamiltonian can be specified by the matrix

⎛0 0 0 h2 ⎞
⎜ ⎟
⎜ 0 g0 0 0 ⎟
A=⎜
⎜
⎟,
⎟ (C45)
⎜0 0 g0 0 ⎟
⎜ ⎟
⎝h 0 0 g0 + g1 ⎠
1

parameterized by disorder fields h1 , h2 and Heisenberg/Ising interactions g0 , g1 . Note that we have introduced horizontal and
vertical bars to visually distinguish between interactions and disorder. As explained in the main-text, under global driving
mapping Sz → Fµ Sµ the secular Hamiltonian will depend only on this column vector F . This representation of the 2-qubit
interaction will thus transform as

⎛ h2 F T ⎞
⎜ ⎟
′
⎜ ⎟
A ↦ A (F ) = ⎜
⎜
⎟,
⎟ (C46)
⎜h1 F
⎜ g0 1 + g1 F F T ⎟
⎟
⎝ ⎠
= ∑ cα Aα (F ), (C47)
α
 20

where {Aα (F )}α are judicious choice of operator basis. A particular choice of operators that is convenient to summarize the
commutators is the following
⎧ ⎫
⎪
⎪
⎪
⎪ ⎛ ±F T ⎞ ⎛ ⎞ ⎛ ⎞⎪
⎪
⎪
⎪
⎪
⎪
⎪ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎪
⎪
⎪
⎪ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎪
⎪
{Aα (F )}α = ⎨A± (F ) = ⎜ ⎟ , AH = ⎜ ⎟ , AI (F ) = ⎜ ⎟⎪ . (C48)
⎪ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎬
⎪ ⎜F 0 ⎟ ⎜ 1 ⎟ ⎜ FFT ⎟⎪
⎪
⎪
⎪ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟⎪
⎪
⎪
⎪
⎪
⎪
⎪ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠⎪
⎪
⎪
⎪
⎩ ⎪
⎭
The matrix commutator between the interaction picture Hamiltonians in different frames lifts to a bracket on the basis C matrices,
yielding surprisingly simple “selection rules” for understanding the structure behind the first order Magnus calculation.
Before presenting the result, we define two more interactions

⎛ ⎞
⎜ ⎟
⎜ ⎟
⎜ ⎟
A˜I (F , G) = ⎜ ⎟, (C49)
⎜ T T ⎟
⎜ (F G + GF ) /2 ⎟
⎜ ⎟
⎝ ⎠

⎛ ⎞
⎜ ⎟
⎜ ⎟
AA (F ) = ⎜
⎜
⎟,
⎟ (C50)
⎜ ijk Fk ⎟
⎜ ⎟
⎝ ⎠

where the first one contains the Ising interaction AI (F ) = A˜I (F , F ) as a special case, and the second one is an anti-symmetric
exchange AA (F ).
[Disorder, Disorder] → Disorder

• [Aσ (F), Aσ′ (G)] = 2i Aσσ′ (F × G) σ, σ ′ ∈ ±

[Disorder, Interaction] → Interaction

• [A+ (F), AH ] = 0

• [A+ (F), AI (G)] = 4i A˜I (G, F × G)

• [A− (F), AH ] = −4i AA (F)
• [A− (F), AI (G)] = 2i((F ⋅ G)AA (G) − AA (F))

[Interaction, Interaction] → Disorder

• [AH , AH ] = 0
• [AH , AI (G)] = 0
• [AI (F), AI (G)] = 2i (F ⋅ G) A+ (F × G) = 0 for pulse sequences built from π/2, π pulses

Appendix D: Derivation of Second-Order Decoupling Rules

From Eq. (19), we have that the cancellation condition for the second-order term is given by two integrals

∭ cα (t1 )cβ (t2 )cγ (t3 ) + ∭ cα (t1 )cβ (t2 )cγ (t3 ).
0<t3 <t2 <t1 <T 0<t1 <t2 <t3 <T

In order to see the similarities to the previous order, we write the above integrals as follows
T
∫ dt1 cα (t1 ) (∬ cβ (t2 )cγ (t3 ) + ∬ cβ (t2 )cγ (t3 )) .
0 0<t3 <t2 <t1 t1 <t2 <t3 <T
 21

Noting that the coefficient here is [Oα , [Oβ , Oγ ]], we can again sum the terms which have β and γ switched, as we did with
first order, and derive the following expression

1 T
H (2) = ([Oα , [Oβ , Oγ ]]) ∫ dt1 cα (t1 )⋅ (D1)
6T 0

⋅ (∬ cβ (t2 )cγ (t3 ) − ∬ cβ (t2 )cγ (t3 ) + (D2)

0<t3 <t2 <t1 0<t2 <t3 <t1

+∬ cβ (t2 )cγ (t3 ) − ∬ cβ (t2 )cγ (t3 )) . (D3)

t1 <t2 <t3 <T t1 <t3 <t2 <T

The inner integrals in the above expression are the first order contribution of the β, γ first order term for all times before t1 ,
minus the first order contribution of the β, γ first order term for all times after t1 . Letting
(1)
cβ,γ (t1 , t2 ) = t1 <ta∬<tb <t2 cβ (ta )cγ (tb ) − t1 <tb∬<ta <t2 cβ (ta )cγ (tb ), (D4)

which is exactly the first order contribution between times t1 and t2 of the operator [Oβ , Oγ ], we can rewrite the expression as
follows
1 T
(1) (1)
H (2) = ([Oα , [Oβ , Oγ ]]) ∫ dt1 cα (t1 ) (cβ,γ (0, t1 ) − cβ,γ (t1 , T )) . (D5)
6T 0

We then write
T
(1) (1)
∫ dt1 cα (t1 ) (cβ,γ (0, t1 ) − cβ,γ (t1 , T ))
0
T T
(1) (1) (1)
=∫ dt1 cα (t1 )cβ,γ (0, t1 ) − ∫ dt1 cα (t1 ) (cβ,γ (0, T ) − cβ,γ (0, t1 ))
0 0
T T
(1) (1)
=2 ∫ dt1 cα (t1 )cβ,γ (0, t1 ) − cβ,γ (0, T ) ∫ dt1 cα (t1 ). (D6)
0 0

(1)
We note that this form is identical to the first order case, with cβ (t) replaced by cβ,γ (0, t). We will now perform the same
substitution that we did for first order to convert the integral expression into a summation expression. For simplicity, we restrict
(1)
our discussion to Hamiltonians involving disorder only. In this case, cβ,γ does not change over a free evolution period, i.e.
(1) (1)
looking at cβ,γ (0, ta ) and cβ,γ (0, tb ), for ta , tb in the same free evolution period, the contribution to the overall term is 0, as the
commutator [Oβ , Oγ ] will be 0 during this time. Thus we can write the same approximation for this term as in the 1st order
term
T
(1)
A=∫ dt1 cα (t1 )cβ,γ (0, t1 ) (D7)
0
n−1 n
(2)
= Af ree + ∑ Pk,k+1,α,β,γ + ∑ Cα,β,γ,k , (D8)
k=1 k=1

where
τk+1
tk+1 − t1
(2) 2 (1)
Cα,β,γ,k = ∫ τ
cα (t1 )dt1 ∫ τk−1
cβ,γ (0, t2 )dt2 , (D9)
tk−1 + k−1
2 tk−1 + 2
π/2 θ1 π/2 π/2
(1) (1)
Pk,k+1,α,β,γ = ∫ c̀α,k (θ1 )rdθ1 ∫ ćβ,γ,k+1 (θ2 )rdθ2 − ∫ ćα,k+1 (θ1 )rdθ1 ∫ c̀β,γ,k (θ2 )rdθ2 , (D10)
0 0 0 θ1
n k−1
Af ree = ∑ Cα,k ∑ Cβ,γ,j . (D11)
k=1 j=1

Here Cβ,γ,j = Cβ,j ∑j−1

`=1 Cγ,` , and the angle terms are defined as follows:

(1)
ćβ,γ,k (θ) = ∬ ćβ (θ1 )ćγ (θ2 ) − ∬ ćβ (θ1 )ćγ (θ2 ), (D12)
0≤θ1 ≤θ2 ≤θ 0≤θ2 ≤θ1 ≤θ
(1)
c̀β,γ,k (θ) = ∬ c̀β (θ1 )c̀γ (θ2 ) − ∬ c̀β (θ1 )c̀γ (θ2 ). (D13)
0≤θ1 ≤θ2 ≤θ 0≤θ2 ≤θ1 ≤θ
 22

We note that this means if the two ramp up functions are proportional, i.e. ćγ (θ) = Fγ,k sin(θ), ćβ (θ) = Fβ,k sin(θ), these terms
will always be 0. There are also no q-terms as we are restricting to a disorder Hamiltonian. Thus, the only term left is the free
evolution term with the frame-lengthening correction.
We can now plug in expressions for the explicit terms in the qubit Hamiltonian to calculate the leading second-order effects.
The free evolution period is much like the lower orders
n
4 ν,ρ
Af ree = ∑ Fµ,k (τk + τp )F<k , (D14)
k=1 π
ν,ρ ρ ρ
where F<k = ∑kl=1 Fν,l (τl + π4 τp )(F<l − F>l ) is the first-order contribution given by ν, ρ through time k. By combining Af ree
with the rest of the terms, we obtain the expression in Tab. I
n
4 ν,ρ µ ν,ρ
2 ∑ Fµ,k (τk + τp ) F<k −F F , (D15)
k=1 π
ν,ρ ρ
where F = ∑nk=1 Fν,k (τk + π4 τp )F<k is the total first-order disorder-disorder term between axes ν and ρ.