Regrouped Pauli Expansions in Quantum Preconditioning

• Regrouped Pauli expansions are algebraic representations that consolidate phase-adjusted products of Pauli operators into unique terms, reducing computational overhead.
• They improve normalization in block-encoding and randomized quantum solvers by aggregating products with cancelling phases, thereby lowering the overall coefficient weight.
• This technique enhances quantum linear system solvers by reducing circuit-depth proxies and mitigating the normalization overhead of traditional preconditioning methods.

Regrouped Pauli expansions are Pauli-basis representations obtained after multiplying Pauli-structured operators and then collecting all products that yield the same Pauli word, up to phases in $\{\pm1,\pm i\}$. In the formulation developed for Pauli-structured preconditioning of quantum linear system solvers, the regrouped expansion of a preconditioned operator $PA$ is the natural input for both direct block-encodings and randomized Pauli algorithms, because regrouping can reduce the apparent size of the Pauli list and can reveal cancellation among coefficients, thereby lowering the Pauli coefficient weight (Nie et al., 1 Jun 2026).

1. Algebraic definition of regrouping

Let $n=\log_2 N$. Any $n$-qubit Hermitian operator $A\in\mathbb{C}^{N\times N}$ admits a Pauli-basis expansion

$$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$

where $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$. Its Pauli coefficient weight is

$w(A)=\sum_{\ell=1}^{L}|a_\ell|.$

A Pauli-structured preconditioner is written as

$$P=\sum_{m=1}^{M}\beta_m P_m,\qquad P_m\in\mathcal{V}_n,\; \beta_m\in\mathbb{R},$$

with

$w(P)=\sum_{m=1}^{M}|\beta_m|.$

The brute-force product is

$PA$0

Because each product $PA$1 is a phase $PA$2 times some Pauli word $PA$3, the product can be rewritten as a regrouped Pauli expansion

$PA$4

where each distinct $PA$5 collects the sum of all phase-weighted coefficients that produce it. In compact form,

$PA$6

The essential operation is therefore not truncation but algebraic consolidation: identical Pauli words are collected after multiplication. This distinguishes regrouped Pauli expansions from a merely enumerated product list of size $PA$7. The 2026 analysis treats this regrouping as a structural transformation of the operator representation, not only as a bookkeeping convenience (Nie et al., 1 Jun 2026).

2. Size reduction and coefficient-weight bounds

The basic quantitative properties of regrouped Pauli expansions are given by explicit bounds on the number of distinct Pauli terms and on the total coefficient weight. For left preconditioning, the regrouped operator $PA$8 satisfies

$PA$9

and

$n=\log_2 N$0

For symmetric preconditioning, the corresponding operator $n=\log_2 N$1 satisfies analogous bounds.

Operator	Distinct-term bound	Weight bound
$n=\log_2 N$2	$n=\log_2 N$3	$n=\log_2 N$4
$n=\log_2 N$5	$n=\log_2 N$6	$n=\log_2 N$7

These bounds are obtained by first considering the brute-force sum over all products and then applying triangle-inequality regrouping. The term-count inequalities show that regrouping never increases the formal support beyond the naive product size. The weight inequalities are more consequential algorithmically, because $n=\log_2 N$8 enters normalization parameters in block-encoding and complexity proxies in randomized Pauli solvers. A plausible implication is that the practical value of regrouping depends less on the nominal count $n=\log_2 N$9 than on how strongly the phase-weighted coefficient sums cancel when identical Pauli words are collected (Nie et al., 1 Jun 2026).

3. Consequences for direct block-encoding

For a Hermitian Pauli sum $n$0, one may choose a normalization factor $n$1 and build a $n$2-encoding. Applied to a regrouped preconditioned operator, a natural choice is

$n$3

The effective normalization becomes

$n$4

and the corresponding effective condition-number-like parameter is

$n$5

This representation matters because naive composition of separate block-encodings of $n$6 and $n$7 forces a normalization $n$8. Direct regrouped encoding can therefore be strictly better whenever

$n$9

The abstract frames this as a response to a specific limitation: in quantum linear system algorithms, the potential advantage of preconditioning may be offset by the normalization overhead incurred by composing separate block-encodings of the system matrix and the preconditioner. Regrouped Pauli expansions identify a regime in which Pauli-structured preconditioning can reduce the effective complexity parameters of quantum algorithms, rather than merely improving the classical condition number (Nie et al., 1 Jun 2026).

4. Randomized Pauli linear-system solvers

In the randomized solver of Wang–McArdle–Berta, the per-sample circuit-depth proxy is

$A\in\mathbb{C}^{N\times N}$0

For a regrouped preconditioned operator, this gives

$A\in\mathbb{C}^{N\times N}$1

Preconditioning strictly reduces this proxy if and only if

$A\in\mathbb{C}^{N\times N}$2

or equivalently

$A\in\mathbb{C}^{N\times N}$3

The paper further notes that one can combine the regrouping weight bound

$A\in\mathbb{C}^{N\times N}$4

with the lower bound

$A\in\mathbb{C}^{N\times N}$5

to certify improvements. Numerical experiments on a finite-dimensional synthetic benchmark show reductions in norm-aware direct block-encoding diagnostics and in the randomized QLS per-sample depth proxy. In this setting, regrouped Pauli expansions function as the algebraic mechanism by which coefficient-weight reduction is translated into algorithmic improvement criteria (Nie et al., 1 Jun 2026).

5. Distinction from simultaneous-measurement partitioning

A common source of confusion is to conflate regrouped Pauli expansions with Pauli-string partitioning for simultaneous measurement. The latter addresses a different task. In a typical variational quantum eigensolver, one estimates

$A\in\mathbb{C}^{N\times N}$6

and direct measurement of each $A\in\mathbb{C}^{N\times N}$7 separately leads to $A\in\mathbb{C}^{N\times N}$8 circuit runs. If a collection of Pauli strings commute pairwise, they can be measured simultaneously in one basis-change circuit, so one seeks a partition of $A\in\mathbb{C}^{N\times N}$9 into the smallest number of commuting cliques (Kurita et al., 2022).

Kurita et al. formulate one clique-finding step as a quadratic binary optimization with variables $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$0,

$$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$1

where $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$2 if $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$3 and $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$4 commute and $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$5 if they anticommute. Minimizing this cost picks a largest-possible subset of mutually commuting strings. The method is mapped to an Ising Hamiltonian and implemented on a Digital Annealer. For $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$6, the worst-case time complexity is $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$7; for $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$8, Kurita et al. show $$A=\sum_{\ell=1}^{L} a_\ell A_\ell,\qquad A_\ell\in\mathcal{V}_n,\; a_\ell\in\mathbb{R},$$9 in the worst case. On Fujitsu’s second-generation Digital Annealer with $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$0, the study investigates up to $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$1 Pauli strings and reports that the reduction factor $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$2 can be $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$3 at maximum (Kurita et al., 2022).

The distinction is structural. In the 2026 preconditioning setting, regrouping means collecting identical Pauli words generated by operator multiplication. In the 2022 measurement setting, regrouping means partitioning a set of Pauli strings into commuting families. Both reduce downstream cost, but they optimize different objects: coefficient-weighted operator representations in one case, measurement-circuit counts in the other.

6. Geometric organization of commuting families

A different but related line of work studies maximal commuting Pauli families through finite geometry. For the generalized Pauli group of a qubit–qudit system with $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$4, each operator modulo overall phase is represented by a vector

$\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$5

equipped with the standard symplectic form. Two operators commute exactly when the symplectic form vanishes: $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$6 (Saniga et al., 2011).

In this geometry, a point is a maximal commuting set, equivalently an additive Lagrangian subspace of dimension $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$7. Including the identity, such subgroups have size $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$8, so as sets of non-identity operators they contain $\mathcal{V}_n=\{I,X,Y,Z\}^{\otimes n}$9 elements. Two distinct points are collinear when the corresponding Lagrangian subspaces intersect in a hyperplane, giving exactly $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$0 common non-trivial operators. Saniga and Planat report that the full incidence structure comprises $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$1 copies of the generalized quadrangle of order two, the “doily,” forming $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$2 pencils; points split into ordinary and exceptional types; and the exceptional subgeometry of the $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$3-case is isomorphic to the full geometry of the $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$4-case. They stress, however, that these generic properties were inferred from computer-handled cases $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$5, and that a rigorous, computer-free proof for $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$6 remained a mathematical challenge (Saniga et al., 2011).

This geometric framework supplies a structural language for commuting-side regrouping. The constructive recipe given for arbitrary $w(A)=\sum_{\ell=1}^{L}|a_\ell|.$7 proceeds by symplectic encoding, finding a maximal isotropic subspace through a chosen operator, collecting all terms whose labels lie in that subspace, and iterating until no terms remain. This suggests a complementary perspective on regrouped Pauli structures: coefficient-side regrouping is naturally expressed through Pauli-product closure and coefficient aggregation, whereas commuting-side regrouping is naturally expressed through Lagrangian subspaces, doilies, and pencils.