---
title: Structured Stabilizer Decomposition
url: https://www.emergentmind.com/topics/structured-stabilizer-decomposition
type: topic
---

# Structured Stabilizer Decomposition

Structured stabilizer decomposition refers to a family of constructions that reorganize stabilizer-theoretic objects into components adapted to locality, tensor-product structure, translation invariance, logical-versus-syndrome separation, or low-rank simulation ansätze. Across the cited works, the decomposed object may be a multipartite stabilizer state, an infinite stabilizer process, a bosonic GKP code space, a Clifford or non-Clifford circuit component, or an unknown quantum state whose stabilizer structure is to be recovered from data [2203.04202], [2607.04015], [2210.14919], [1808.00128], [2510.05890]. The resulting representations are correspondingly diverse: tuples of alternating bilinear forms up to simultaneous congruence, generating tableaux over \(F_p(\delta)\), logical–stabilizer subsystem factorizations, and exact or approximate expansions into stabilizer states or stabilizer projectors.

## 1. Formal notions and problem classes

The cited literature does not use a single standardized definition of structured stabilizer decomposition. Instead, several formal notions recur. One class of problems asks for tensor-product decompositions of multipartite stabilizer states under a specified local equivalence relation. In that setting, decomposition means deciding when a state is PLC- or LU-equivalent to \(\ket{\phi}\otimes\ket{\phi'}\), identifying indecomposable building blocks, and organizing those blocks into an entanglement generating set (EGS) [2203.04202], [2507.09416].

A second class of problems treats decomposition as a low-rank linear expansion. For pure states, the exact stabilizer rank is
\[
\chi(\psi)=\min\left\{k:\ |\psi\rangle=\sum_{\alpha=1}^{k} c_\alpha |\phi_\alpha\rangle,\ |\phi_\alpha\rangle\in \mathrm{STAB}_n\right\},
\]
while approximate decomposition is governed by the approximate stabilizer rank \(\chi_\delta(\psi)\) and by the stabilizer extent \(\xi(\psi)\), defined as the minimum squared \(\ell_1\)-norm of decomposition coefficients [1808.00128]. Related operator-level constructions write a non-Clifford unitary as
\[
U=\sum_i c_i \ket{\phi_i}\bra{\phi_i},
\]
with \(\ket{\phi_i}\) stabilizer states, so that branching occurs at the operator level rather than only in a state-vector expansion [2009.05110].

A third class of problems encodes infinite or bosonic stabilizer structure through finite algebraic data. In translation-invariant settings, delayed ZX-diagrams are mapped to generating tableaux \(H=[X\mid Z]\) over \(F_p(\delta)\) satisfying \(XZ^\dag=ZX^\dag\), and then to canonical normal forms [2607.04015]. For GKP codes, the oscillator Hilbert space is factorized as a logical subsystem tensored with a stabilizer subsystem,
\[
\mathcal H=\mathcal L\otimes_{\mathcal G}\mathcal S,
\]
with the defining property that tracing out \(\mathcal S\) reproduces ideal decoding [2210.14919].

A fourth class is algorithmic. Here the target is not merely existence of a decomposition, but efficient recovery of one. For arbitrary \(n\)-qubit states, one formulation seeks
\[
\ket{\psi}=\sum_{i\in[k]}\beta_i\ket{\phi_i}+\alpha\ket{\phi^\perp},
\]
where the extracted part has stabilizer rank \(\operatorname{poly}(1/\varepsilon)\) and the residual has stabilizer fidelity \(<\varepsilon\) [2510.05890]. By contrast, some structurally adjacent results identify nearby stabilizer structure without outputting a linear decomposition; tolerant testing based on Bell-difference or \(U^3\)-type statistics is of that kind [2408.06289].

## 2. Multipartite state decomposition and entanglement structure

For multipartite qubit stabilizer states, the most systematic decomposition framework in the cited literature is the commutation matrix formalism. Fix a partition of the qubits into parties. Choosing a basis of the maximally isotropic stabilizer subspace \(V^\mathcal S\) yields, for each party \(\alpha\), an alternating commutation matrix \(C_\alpha\). PLC equivalence is then reduced to simultaneous congruence of the tuple \((C_\alpha)_\alpha\), and tensor-product decomposition is reduced to simultaneous block diagonalization [2203.04202]. The central criterion is
\[
\ket\psi \text{ is decomposable}
\iff
\exists Q\in \mathrm{GL}(n,\mathbb Z_2)\text{ such that }
Q C_\alpha Q^T = B_\alpha^1 \oplus B_\alpha^2
\quad \forall \alpha.
\]
The same paper also identifies the tuples that actually arise from stabilizer states by the rank-saturation condition
\[
2\,\mathrm{rk}([C_\alpha]_\alpha)=\sum_\alpha \mathrm{rk}(C_\alpha).
\]

This reorganization yields a precise notion of indecomposability: indecomposable stabilizer states correspond exactly to indecomposable tuples of alternating forms under simultaneous congruence. The associated endomorphism ring is defined by
\[
C_\alpha E = E^T C_\alpha\qquad \forall \alpha,
\]
and Fitting’s lemma implies that decomposability is equivalent to the existence of a nontrivial idempotent self-adjoint endomorphism. The same framework gives a sharp structural threshold: the EGS is finite for \(M\le 3\), but infinite for every \(M\ge 4\); for four parties, the paper exhibits an explicit infinite family of indecomposable spiral graph states [2203.04202]. It also establishes a qudit contrast: for prime dimensions \(d\ge 3\), every state decomposes uniquely into indecomposable EGS elements, whereas for qubits the uniqueness of decomposition remains open in general [2203.04202].

Graph states admit a parallel graph-theoretic description. PLC equivalence among graph states is generated by local complementations together with within-party edge additions and removals, denoted LCE. This gives a direct move set on graph representatives, while the commutation matrices supply the finite-field algebraic normal form [2203.04202].

For tripartite pure qudit stabilizer states, a different decomposition theory appears. Any tripartite pure stabilizer state of arbitrary local dimension \(D\) can be reduced by local unitaries to a tensor product of \(p\)-level GHZ states, pairwise \(p\)-level EPR pairs, and unentangled \(p\)-level qudits [2507.09416]. The key invariant is a family of subsystem phase matrices \(M_\alpha\), one for each party, with
\[
M_{\alpha,ij}=r^{(i)T}\Omega_\alpha r^{(j)},\qquad
M_{\alpha,ii}=0,\qquad
M_\alpha+M_\alpha^T=0,\qquad
\sum_\alpha M_\alpha=0.
\]
These matrices encode subsystem commutation phases of stabilizer generators. Dimension reduction occurs when all projected phase matrices vanish modulo \(p\), while GHZ or EPR extraction is triggered by span and intersection conditions on the \(M_\alpha\) [2507.09416].

An important subtlety is that local Clifford unitaries are not sufficient in arbitrary prime-power dimensions. The prime-power case \(D=p^n\) requires local non-Clifford unitaries to access embedded \(p\)-level entanglement structure that the \(D\)-level Clifford group cannot rearrange. The cited examples at \(D=4\) and \(D=9\) are explicitly used to demonstrate that LC equivalence is strictly weaker than LU equivalence for this decomposition problem [2507.09416]. A common misconception is therefore that tripartite stabilizer entanglement normal forms in arbitrary qudit dimension should follow from local Clifford theory alone; the cited results show that this is false in non-squarefree prime-power dimensions.

## 3. Infinite, translation-invariant, and bosonic decompositions

For infinite translation-invariant stabilizer systems, the delayed stabilizer ZX-calculus provides a finite graphical language built by adding a single delay generator to the odd-prime stabilizer ZX-calculus [2607.04015]. The delay has two semantics. In the behavioural semantics, a finite delayed diagram is interpreted through all finite truncations, and two such unrollings are identified when they have the same information content. In the algebraic semantics, the delay is interpreted as multiplication by a formal variable \(\delta\), so that translated Pauli families become rational generating functions in \(F_p(\delta)\) [2607.04015].

The main algebraic object is the generating tableau. A shifted affine Lagrangian subspace is represented by
\[
H=\begin{bmatrix}X & Z\end{bmatrix}\in F_p(\delta)^{n\times 2n},
\qquad
\mathbf a\in F_p(\delta)^{2n},
\]
with
\[
\mathbf a+\ker(H)=S,\qquad XZ^\dag = ZX^\dag,\qquad (-)^\dag=\overline{(-)}^T.
\]
Commutation against all shifts is captured by the shifted symplectic form
\[
\varpi_n(f(\delta),g(\delta)):=\omega_n\bigl(f(\delta),g(\delta^{-1})\bigr).
\]
The paper then gives a canonical tableau form
\[
S= \ker\!\left(
\begin{bmatrix}
I_m & L & \Sigma & 0\\
0 & 0 & -L^\dag & I_{n-m}
\end{bmatrix}
\begin{bmatrix}
\varsigma & 0\\
0 & \varsigma
\end{bmatrix}
\right)+\mathbf a,
\]
with unique data \((m,L,\Sigma,\varsigma,\mathbf a)\) [2607.04015]. This is a canonical structured decomposition of a translation-invariant stabilizer relation into finite algebraic pieces.

The rewrite theory is equally central. Every delayed diagram reduces to graph-like form, then to AP-form, and finally to a unique reduced AP-form by generalized local complementation, pivoting, Schur complementation, and Gaussian elimination over \(F_p(\delta)\). The paper proves soundness, universality, and completeness for the generating tableau semantics. A necessary caveat is that completeness is established for the generating tableau or shifted-affine-Lagrangian semantics, not for the full channel-sequence semantics; the relation to the full infinite stabilizer group is only
\[
\Gamma(\Ext(D))\subseteq StabGrp(D)
\]
[2607.04015]. It would therefore be inaccurate to interpret the completeness theorem as a full behavioural completeness statement.

For bosonic GKP codes, the stabilizer subsystem decomposition gives a different kind of structure. For a code specified by \(\mathcal G=(\Sigma,d,\mathcal P)\), one writes
\[
\mathcal H=\mathcal L\otimes_{\mathcal G}\mathcal S,\qquad
\ket{\psi}\otimes_{\mathcal G}\ket{k}=\ket{\psi,k}_{(\Sigma,d)},
\]
where \(k\in\mathcal P\) labels stabilizer eigenvalues in a primitive cell of the dual lattice [2210.14919]. The basis states \(\ket{\mu,k}_{(\Sigma,d)}\) are displaced ideal codewords and simultaneous stabilizer eigenstates. In this representation, the decisive identity is
\[
\int_{\mathcal P} d^{2n}v\; {}_{\mathcal S}\!\bra{0}\bra{v}\,\hat\rho\,\ket{v}_{\mathcal S}
=
\mathrm{tr}_{\mathcal S}(\hat\rho),
\]
so partial trace over the nonlogical subsystem is equivalent to ideal primitive-cell decoding [2210.14919].

This subsystem factorization is decoder-dependent, since the primitive cell \(\mathcal P\) fixes how displacements are reduced modulo the dual lattice. The paper also gives explicit cell, Gaussian, and dimension transformations between decompositions, and shows that full unfolding recovers Zak-state decompositions. In single-mode square GKP codes, this framework supports efficient logical-noise simulation for loss, Gaussian displacement, and dephasing, avoiding the large Fock cutoffs required by direct Fock-basis numerics [2210.14919].

## 4. State, gate, and circuit decompositions for simulation and synthesis

In classical simulation, structured stabilizer decomposition most commonly means replacing the hard non-Clifford part of a problem by a low-rank stabilizer ansatz. For pure states, the foundational quantitative statement is the sparsification bound
\[
\chi_\delta(\psi)\le 1+\xi(\psi)/\delta^2,
\]
where \(\xi(\psi)\) is the stabilizer extent [1808.00128]. This converts a dense but low-\(\ell_1\)-norm stabilizer expansion into a low-rank approximate decomposition. The same paper proves multiplicativity of \(\xi\) for tensor products of factors acting on at most three qubits and develops exact or optimal decompositions for \(R(\theta)\), \(T\)-type states, and CCZ states and gates. A particularly important structural step is the lifting lemma: if a diagonal resource state \(V|+^t\rangle\) has a stabilizer decomposition into equatorial stabilizer states, then the diagonal unitary \(V\) itself lifts to a sum-over-Cliffords decomposition [1808.00128]. This reorients decomposition from state space to operator space.

A more explicit operator-level version is the stabilizer projector decomposition of unitaries,
\[
U=\sum_i c_i \ket{\phi_i}\bra{\phi_i},
\]
used to build the SPIR and SPC simulation algorithms [2009.05110]. Here only non-Clifford layers are decomposed, while Clifford subcircuits are handled exactly by Gottesman–Knill propagation. The stabilizer projector rank
\[
\kappa(U)=\min\left\{k:\ U=\sum_{i=1}^k c_i \ket{\phi_i}\bra{\phi_i}\right\}
\]
controls cost. SPIR is a polynomial-space recursive algorithm with runtime \(O((2d_{nc})^{k+1}n^3)\) when the largest non-Clifford layer has projector rank \(2^k\), while SPC is an exponential-space contraction algorithm with runtime \(O(d_{nc}\kappa_n^2 n^3)\) [2009.05110]. The cited work emphasizes that these decompositions are exact, gate-specific, and advantageous only in certain dense-non-Clifford regimes.

Structured decomposition also appears at the circuit-normal-form level. Using Bruhat decomposition of the binary symplectic group, arbitrary stabilizer circuits over \(\{H,P,\mathrm{CNOT}\}\) admit the exact 7-stage form
\[
\text{-}C\text{-}CZ\text{-}P\text{-}H\text{-}P\text{-}CZ\text{-}C\text{-},
\]
as well as the 9-stage form
\[
\text{-}C\text{-}P\text{-}C\text{-}P\text{-}H\text{-}C\text{-}P\text{-}C\text{-}P\text{-}
\]
[1705.09176]. The same work proves a depth-\((14n-4)\) implementation on the linear nearest-neighbor architecture and interprets the resulting circuit as a normal form that is optimal in Hadamard count and asymptotically optimal in parameter count [1705.09176].

A subgroup-theoretic alternative avoids explicit symplectic-group factorization and instead derives two normal forms directly inside the Clifford group:
\[
\mathrm{CX}\text{-}\mathrm{CZ}\text{-}P\text{-}Z\text{-}X\text{-}H\text{-}\mathrm{CZ}\text{-}P\text{-}H
\]
and
\[
P\text{-}\mathrm{CX}\text{-}\mathrm{CZ}\text{-}\mathrm{CX}\text{-}Z\text{-}X\text{-}H\text{-}\mathrm{CZ}\text{-}\mathrm{CX}\text{-}P\text{-}H.
\]
In the second form, both \(CZ\) layers are depth \(1\) and together contain at most \(n\) \(CZ\) gates [2012.09224]. The corresponding decomposition of the non-Hadamard subgroup is unique:
\[
Z_vP_bZ_BX_A,
\]
with \(A\in GL(n,\mathbb F_2)\) and \(B\) a symmetric zero-diagonal binary matrix [2012.09224]. This is a circuit-level analogue of structured stabilizer decomposition: the entangling, linear-reversible, phase, and Pauli content are isolated into algebraically distinct layers.

## 5. Learning, testing, and recovering stabilizer structure

Algorithmic recovery of structured stabilizer decompositions for unknown states is developed most explicitly in the learning framework of [2510.05890]. The target is, for arbitrary \(\ket\psi\) and \(\varepsilon>0\), to output
\[
\ket{\psi}=\sum_{i\in[k]}\beta_i\ket{\phi_i}+\alpha\ket{\phi^\perp},
\]
where each \(\ket{\phi_i}\) is a stabilizer state, \(k=\operatorname{poly}(1/\varepsilon)\), and the residual has stabilizer fidelity \(<\varepsilon\). The central technical primitive is self-correction with respect to the class of stabilizer states: from copies of a state having fidelity at least \(\tau\) with some stabilizer state, the algorithm outputs a stabilizer state with fidelity at least \(\tau^C\), for a universal constant \(C>1\) [2510.05890].

The decomposition algorithm iterates this self-corrector on successively prepared residual states, using access to a preparation unitary \(U_\psi\) and its controlled version \(\mathrm{con}U_\psi\). Under the algorithmic polynomial Freiman–Ruzsa conjecture (APFR), the paper gives a polynomial-time protocol for learning a structured decomposition; without APFR, it gives a quasipolynomial-time protocol [2510.05890]. For states of stabilizer extent \(\xi\), the resulting learner runs in time \(\operatorname{poly}(n,\xi^{\log \xi})\) unconditionally and polynomial time under APFR; for stabilizer-rank-\(k\) states, the abstract states an unconditional runtime \(\operatorname{poly}(n,k^{k^2})\) [2510.05890]. In this setting, structured stabilizer decomposition is explicitly an algorithmic analogue of structure-versus-pseudorandomness.

A related but distinct line of work studies certification rather than output decomposition. Polynomial-time tolerant testing of stabilizer states is based on a quantum Gowers-\(3\) norm and on Bell-difference statistics [2408.06289]. The key structural implication is that large \(U^3\)-type signal yields a structured subset \(S'\) of Pauli labels with small doubling,
\[
|2S'|\le (1/\gamma)|S'|,
\]
from which one derives subgroup structure and a stabilizer covering of the relevant Pauli set. This eventually implies existence of a nearby stabilizer state [2408.06289]. The paper is explicit, however, that it is not a paper on stabilizer rank or on explicit linear-combination decompositions of arbitrary states into stabilizer states. Its output is a tester and a Pauli-side covering theory, not a low-rank stabilizer expansion [2408.06289].

An adjacent structural result concerns the class of stabilizer states themselves rather than decompositions of general states. For \(n\)-qubit stabilizer states, the optimal sample complexity of cloning is \(\Theta(n)\), matching the known \(\Theta(n)\) sample complexity of learning [2604.15269]. The proof uses hidden linear subspaces, Bell-basis support, and a structured random purification channel, but it does not study stabilizer rank or sparse expansions of non-stabilizer states [2604.15269]. This clarifies a boundary of the subject: not every efficient structural representation of stabilizer states is a stabilizer decomposition in the rank-theoretic sense.

## 6. Orbit-sensitive finite-copy decompositions and open questions

The most explicit recent finite-copy decomposition formulas are orbit-sensitive stabilizer-rank identities for magic-state orbits [2605.28586]. For qutrits, the cited work treats the Strange, Norrell, Hadamard-eigenstate, and \(T_3\) orbits separately and proves that distinct Clifford orbits can exhibit different stabilizer ranks at small tensor powers. The resulting upper bounds on asymptotic stabilizer-rank exponents are
\[
\gamma_{\mathbb S}\le \log_3(2)/2 \approx 0.316,\qquad
\gamma_{H_3},\gamma_{\mathbb N}\le \log_3(4)/3 \approx 0.421,
\]
all strictly below the prior \( \gamma_{T_3}\le 1/2 \) baseline [2605.28586].

These bounds arise from exact, short decomposition identities. The Strange orbit satisfies \(\chi(\mathbb S^{\otimes 2})=2\), implemented by a difference of two quadratic-phase stabilizer states whose support-selective interference reproduces the target sparse support [2605.28586]. The Hadamard-eigenstate orbit satisfies \(\chi(H_3^{\otimes 3})=4\), using the factorization
\[
\ket{H_3}=\frac{N}{2}(\ket0+\ket+)
\]
to reduce amplitude matching to a handful of zero-versus-nonzero support classes [2605.28586]. The Norrell orbit satisfies \(\chi(\mathbb N^{\otimes 3})=4\), and its 4-copy decomposition uses seven stabilizer states organized by indicator variables \([y_i=2]\) and the count \(n_2(y)\) of coordinates equal to \(2\) [2605.28586]. In the qubit case, the paper gives a direct 4-copy algebraic identity with \(\chi(T^{\otimes 4})=3\), matching the known \(\gamma_T\le \log_2(3)/4\) exponent without using the earlier contracted cat-state construction [2605.28586].

The same paper also proves the first nontrivial \(\Omega(m/\log m)\) asymptotic lower bounds for the Hadamard-eigenstate and Norrell orbits, showing that the finite-copy upper bounds do not yet characterize the true exponent [2605.28586]. Operationally, two copies of \(H_3\) or \(\mathbb N\) can be converted by explicit two-qutrit Cliffords into an injectable phase state with constant success probability, whereas no analogous two-copy conversion to a non-Clifford diagonal gate exists for the Strange orbit [2605.28586]. This demonstrates that good decomposability and good injectability need not coincide.

Several broad open issues remain visible across the cited literature. For qubit multipartite stabilizer states, uniqueness of decomposition into EGS elements is unresolved in general, even though the prime-qudit case \(d\ge 3\) is unique [2203.04202]. For delayed ZX, completeness holds for the generating tableau semantics rather than the full channel-sequence semantics [2607.04015]. For learning arbitrary states, the strongest polynomial-time guarantees still depend on APFR [2510.05890]. For arbitrary qudit dimensions, local Clifford theory is insufficient to capture tripartite LU decomposition at prime powers [2507.09416]. These are not incidental technicalities; they delimit where structured stabilizer decomposition is currently a canonical algebraic theory, where it is a constructive but nonunique synthesis method, and where it remains only partially understood.

Source: https://www.emergentmind.com/topics/structured-stabilizer-decomposition