---
title: Projective Cluster-Additive Transformation (PCAT)
url: https://www.emergentmind.com/topics/projective-cluster-additive-transformation-pcat
type: topic
---

# Projective Cluster-Additive Transformation (PCAT)

Projective Cluster-Additive Transformation (PCAT) is a projection-based, cluster-additive transformation for quantum lattice models that block-diagonalizes sectors of $\mathcal{H}=\mathcal{H}_0+V$ or $\mathcal{H}=\mathcal{H}_0+\lambda V$ into subspaces adiabatically connected to eigenspaces of $\mathcal{H}_0$. Its defining purpose is to produce effective Hamiltonians that remain additive on disconnected clusters, which is the prerequisite for perturbative and non-perturbative linked-cluster expansions. In recent numerical linked-cluster expansion (NLCE) workflows, including hybrid quantum-classical pipelines, PCAT serves as the post-processing step that converts finite-cluster low-energy information into intensive, cluster-additive one–quasi-particle effective Hamiltonians whose Fourier transforms yield thermodynamic-limit dispersions [2303.04774; 2511.06623; 2605.28599].

## 1. Formal setting and conceptual scope

PCAT is formulated for local lattice Hamiltonians whose unperturbed part admits a block structure in quasi-particle number or, more generally, in degenerate eigenspaces. In the foundational formulation, the Hilbert space decomposes as
$$
\mathscr{H}=\bigoplus_{n=0}^N \mathscr{H}_0^n,
\qquad
\mathcal{H}_0=\bigoplus_{n=0}^N \mathcal{H}_0^n,
$$
with an ordering of unperturbed energies $e_0^m\leq e_0^n$ for $m\leq n$. The objective is to block-diagonalize the full Hamiltonian into sectors adiabatically connected to these subspaces, so that
$$
\mathcal{H}_{\mathrm{eff}}=\bigoplus_{n=0}^N \mathcal{H}_{\mathrm{eff}}^n.
$$
The relevant projectors satisfy
$$
\mathds{1}=\sum_n P_n=\sum_n \bar P_n,
$$
where $P_n$ projects onto an eigenspace of $\mathcal{H}_0$ and $\bar P_n$ onto the adiabatically connected eigenspace of $\mathcal{H}$ [2303.04774].

In this setting, “projective” refers to constructions that use projectors onto eigenspaces of $\mathcal{H}_0$ and the adiabatically connected eigenspaces of $\mathcal{H}$, while “cluster-additive” refers to the requirement that the effective Hamiltonian respects additivity on disconnected clusters. For disconnected clusters $A$ and $B$,
$$
\mathcal{H}_{\mathrm{eff}(A\cup B)}
=
\mathcal{H}_{\mathrm{eff}(A)}\otimes \mathds{1}_B
+
\mathds{1}_A\otimes \mathcal{H}_{\mathrm{eff}(B)}.
$$
This property is not an auxiliary convenience. It is the structural condition that allows linked-cluster expansions to be formulated exclusively on connected subgraphs.

The later one–quasi-particle formulations adopt the same logic in a more specialized setting. There one considers a finite connected cluster $C$ with cluster Hamiltonian $H_C=\sum_{\alpha\in C} h_\alpha$, solves for a low-energy effective Hamiltonian that is block-diagonal in quasi-particle number, enforces cluster additivity on the transformation, and then embeds and Fourier-transforms the one–quasi-particle block to obtain the dispersion $\epsilon(k)$ or $\omega(\vec k)$ in the thermodynamic limit. This suggests that PCAT is best understood as a compatibility condition between block-diagonalization and linked-cluster summation rather than as a stand-alone diagonalization scheme.

## 2. Construction of the transformation

The defining modification introduced by PCAT is the subtraction of all projections onto lower-energy subspaces before orthonormalization. For the single-particle sector in the foundational construction,
$$
\ket{\tilde{\Psi}_0}=\ket{\Psi}_0,
$$
and
$$
\ket{\tilde{\Psi}_1}
=
\ket{\Psi}_1
-
\left(1/\langle 0|\Psi_0\rangle\right)\langle 0|\Psi_1\rangle \ket{\Psi}_0,
$$
which implies
$$
P_0\ket{\tilde{\Psi}_1}=0.
$$
For general target sector $n$, one introduces
$$
R=\sum_{m<n} P_m,
$$
and, assuming a non-singular subtraction operator $Y_{n-1}$, defines modified states by subtracting lower-energy components so that
$$
R\ket{\tilde{\Psi}_n}=0.
$$
In the quantum-algorithm formulation this same idea is written directly in projector form as
$$
|\widetilde{\Psi}^n\rangle
=
\left(\mathbb{I}-\bar R^n (R^n \bar R^n R^n)^{-1}R^n\right)|\Psi^n\rangle,
$$
with $R^n=\sum_{m<n}P^m$ [2303.04774; 2511.06623].

Once the lower-energy projections have been removed, PCAT constructs the blockwise correction from modified overlaps. In the one–quasi-particle sector,
$$
\tilde O^{[1]}_{ij}
=
\langle \Phi_i^{[1]}|\tilde\Psi_j^{[1]}\rangle,
$$
and the correction is the symmetric Löwdin orthonormalization
$$
V^{[1]}
=
\tilde O^{[1]\dagger}
\left(\tilde O^{[1]}\tilde O^{[1]\dagger}\right)^{-1/2}.
$$
Equivalently, in block $n$,
$$
U_{\mathrm{PCAT}}^n
=
U^n \tilde X^{n\dagger}
\big(\tilde X^n \tilde X^{n\dagger}\big)^{-1/2}.
$$
The resulting effective Hamiltonian is obtained by conjugation with the PCAT unitary or partial isometry. In the one–quasi-particle case,
$$
H_{\mathrm{eff}}^{[1]}
=
V^{[1]\dagger} H^{[1]} V^{[1]} - E^{[0]},
$$
or, cluster by cluster,
$$
H^{1}_{\mathrm{eff},C}
=
\big(\tilde H_C\big)^{[11]}-E_C^0.
$$

A central structural result is that PCAT depends only on the target sector and all lower-energy eigenspaces. The 2023 construction states this explicitly as a minimal basis requirement relative to other linked-cluster block-diagonalization schemes. The 2025 quantum-algorithm paper reformulates the same point operationally: PCAT uses only low-energy eigenspace information, namely energies, overlaps with unperturbed states, and low-energy Hamiltonian matrix elements, without tomography of higher sectors. A plausible implication is that PCAT separates the difficulty of preparing a low-energy subspace from the distinct difficulty of enforcing cluster additivity on that subspace.

## 3. Cluster additivity and linked-cluster expansions

The reason PCAT is necessary appears most transparently on disconnected clusters. If $A$ and $B$ do not share sites or bonds, then the microscopic Hamiltonian factorizes, and exact eigenstates satisfy the product structure
$$
|\Psi_{A\cup B}^{[n+m]}\rangle
=
|\Psi_A^{[n]}\rangle\otimes |\Psi_B^{[m]}\rangle.
$$
Approximate block-diagonalization can spoil this structure by leaking lower-sector components into the target manifold. PCAT removes those components, and on disconnected clusters the modified product states factorize:
$$
|\widetilde{\Psi}^{n+m}_{A\cup B}\rangle
=
|\widetilde{\Psi}^n_A\rangle\otimes |\widetilde{\Psi}^m_B\rangle.
$$
This is the core mechanism by which PCAT restores cluster additivity [2511.06623].

For one–quasi-particle quantities, cluster additivity also requires subtraction of the ground-state energy. The effective one–quasi-particle Hamiltonian is defined as
$$
H_{\mathrm{eff}}^{[1]}
=
H_{\mathrm{eff}}^{[11]}-E^{[0]}.
$$
The subtraction is necessary to make the one–quasi-particle block intensive and additive: on $A\cup B$, matrix elements describing an excitation on one disconnected component would otherwise carry the ground-state energy of the other component. In the earlier linked-cluster formulation this same point appears as the additive decomposition of the ground-state-energy-reduced one-particle effective Hamiltonian, following the observation that $e^0(A\cup B)=e^0(A)+e^0(B)$ for non-degenerate ground states [2303.04774].

Once cluster additivity holds, NLCE proceeds by inclusion–exclusion. For a property $P$ evaluated on clusters,
$$
\bar P_C
=
P_C-\sum_{C'\subsetneq C} w_{C'/C}\,\bar P_{C'},
\qquad
P=\sum_C w_C\,\bar P_C.
$$
In the excited-state pipeline, $P$ is the real-space matrix of the one–quasi-particle effective Hamiltonian. Effective hoppings are extracted from its matrix elements and embedded across the lattice. Their Fourier transform yields the thermodynamic-limit dispersion. On rectangular clusters of the square lattice,
$$
\omega(\vec{k})
=
\sum_{L_m,L_n}\sum_{\vec\mu,\vec\nu}
e^{i\vec k\cdot(\vec\nu-\vec\mu)}
\,
\bar H^{[1]}_{\mathrm{eff},L_m\times L_n,\vec\mu,\vec\nu},
$$
while in one dimension the rectangular expansion reduces to subtracting the contributions of consecutive cluster sizes [2605.28599].

The failure mode is equally explicit: if the transformation is not cluster-additive, spurious inter-cluster matrix elements appear, unphysical hopping between disconnected clusters is generated, and NLCE diverges. This point is emphasized both in the abstract mathematical construction and in later one–quasi-particle benchmarks.

## 4. Relation to established effective-Hamiltonian methods

PCAT was introduced as a generalization of the minimal transformation known under several names: Takahashi’s transformation, Schrieffer–Wolff transformation, des Cloizeaux effective Hamiltonian, canonical van Vleck effective Hamiltonian, and the two-block orthogonalization method. The minimal projective transformation is written as
$$
T=\sum_n T_n,
\qquad
T_n=\bar P_n P_n
\left(\sum_m P_m \bar P_m P_m\right)^{-1/2}.
$$
In the two-block setting this is equivalent, order by order, to the Schrieffer–Wolff direct rotation. It has minimal norm, but the 2023 analysis shows that it fails to be cluster-additive for excitations in general [2303.04774].

PCAT resolves that failure by modifying the underlying eigenvectors before the Löwdin-type orthonormalization. The added ingredients relative to the minimal projective transformation are projectors onto lower-energy eigenspaces. This is the sense in which PCAT retains a minimal-basis character while restoring linked-cluster additivity. The 2023 paper contrasts this with the multi-block orthogonalization method and perturbative continuous unitary transformations, which are genuine linked-cluster methods but typically require a larger operator basis [2303.04774].

The later one–quasi-particle literature rephrases the same issue in physically transparent terms. In the pure transverse-field Ising model, the global $\mathbb{Z}_2$ parity separates ground-state and one–quasi-particle sectors, so
$$
\langle \Phi^{[0]}|\Psi_i^{[1]}\rangle=0.
$$
In that special case the PCAT correction in the one–quasi-particle sector vanishes, and a good single-unitary block-diagonalization already has a clean block structure. By contrast, adding a longitudinal field breaks parity, mixes the ground-state and one–quasi-particle sectors, and causes canonical two-block decoupling to violate cluster additivity; the resulting effective Hamiltonian can exhibit unphysical hopping between disconnected clusters [2511.06623]. A common misconception is therefore that any accurate low-energy block-diagonalization is sufficient for NLCE. The published constructions show that accuracy of the low-energy subspace and cluster additivity of the transformation are distinct requirements.

## 5. Variational, adiabatic, and hardware realizations

In the NLCE+VQE formulation, each cluster $C$ is assigned a single unitary $U_C$ that approximately block-diagonalizes the low-energy subspace containing the ground state and the one–quasi-particle sector. For the transverse-field Ising model with or without longitudinal field, the Hamiltonian variational ansatz (HVA) is chosen to mirror the Hamiltonian structure:
$$
U_{\mathrm{HVA},C}(\boldsymbol{\theta})
=
\prod_{\ell=1}^{L}
\exp\!\Big(i \sum_{\langle \nu,\mu\rangle}\theta^{XX}_{\nu,\mu,\ell} X_\nu X_\mu \Big)
\exp\!\Big(i \sum_{\nu} \theta^{X}_{\nu,\ell} X_\nu \Big)
\exp\!\Big(i \sum_{\nu} \theta^{Z}_{\nu,\ell} Z_\nu \Big),
$$
with XX–X–Z ordering. Benchmarks used $L=\lceil N/2\rceil$ or $L=N$ layers. Two cost-function families were employed. The trace-minimization cost promotes the correct low-energy block without forcing internal diagonalization, while the variance-based cost vanishes iff the subspace is an invariant eigenspace and thus directly penalizes couplings from the transformed one–quasi-particle sector to the rest of the Hilbert space [2511.06623].

After optimization, the cluster workflow measures three $(N+1)\times(N+1)$ matrices: overlaps with unperturbed states, overlaps among prepared low-energy states, and the Hamiltonian matrix in the prepared basis. These measurements scale as $O(N^2)$ per cluster. Classical post-processing then orthogonalizes the subspace when necessary through a generalized eigenvalue problem, constructs the modified overlap, and applies the PCAT correction to the one–quasi-particle block. The same paper emphasizes that the PCAT post-processing is independent of the cluster solver and applies to VQE, phase estimation, adiabatic preparation, and other methods that can prepare or characterize the low-energy subspace [2511.06623].

The hardware realization of NLCE+QA implements the same logic on a 20-qubit trapped-ion quantum processing unit. It considers both adiabatic state preparation (ASP) and VQE trained on a classical device, while the final expectation values are obtained from the QPU. A distinctive feature is the CX-test, introduced as an alternative to the Hadamard test for overlap estimation. In PCAT, the global prefactor $\langle 0|U^\dagger|0\rangle$ cancels in $V^{[1]\dagger}H^{[1]}V^{[1]}$, so it need not be measured. Ground-state energies are obtained via sample-based quantum diagonalization (SQD). The one–quasi-particle matrices remain $N\times N$, the measurement cost remains $O(N^2)$, and the paper gives explicit circuit counts such as
$$
n_{\rm circ}=2N^2+2(2N^2-N)+1
$$
for a one-dimensional TFIM chain, with additional circuits in the longitudinal-field case [2605.28599].

## 6. Benchmarks, regimes of validity, and limitations

The original PCAT paper benchmarks the method in the low-field ordered phase of the square-lattice transverse-field Ising model for single spin-flips and two spin-flip bound states. It reports perturbative linked-cluster series for the single spin-flip gap $\Delta$ up to order $24$ and the two spin-flip bound-state gap $\Delta_{\mathrm{bs}}$ up to order $22$. In its non-perturbative implementation, exact diagonalization is performed on connected clusters and Wynn extrapolations are used across cluster sizes. The NLCE for $\Delta$ converges up to $h\approx 0.5$, Wynn extrapolations extend slightly further, but both break down before $h_c\approx 0.7610$. Finite-size scaling at $h_c$ indicates algebraic behavior with exponent $\kappa\approx -0.51$ [2303.04774].

In the high-field polarized phase, the 2025 NLCE+VQE study benchmarks one–quasi-particle dispersions for the transverse-field Ising model in one and two dimensions, and with longitudinal field. In one dimension at $h=1$, the exact dispersion is
$$
\omega_{\mathrm{exact}}(k)=2\sqrt{1+J^2-2J\cos k}.
$$
At $J_c=1$, NLCE+VQE with $L=\lceil N/2\rceil$ reproduces NLCE+ED and the exact dispersion, including the closing gap at $k=0$, with rapid convergence up to $N_{\max}=16$. In two dimensions on the square lattice, at $J_c=0.328$, NLCE+VQE with $L=\lceil N/2\rceil$ matches NLCE+ED and high-order series benchmarks from pCUT; at $N_{\max}=14$ the difference between NLCE+VQE and NLCE+ED is $\Delta\omega\lesssim 10^{-2}$, and all four tested cost functions converge identically within $\lesssim 10^{-3}$ variations, below NLCE truncation errors of $\sim 5\times 10^{-2}$. For the one-dimensional TFIM with longitudinal field at $J=0.5$ and $h_l=0.5$, Schrieffer–Wolff-only effective Hamiltonians show unphysical oscillations and non-additive long-range hopping, whereas the PCAT correction restores smooth dispersions and NLCE convergence. The same benchmarks report optimizer trapping for variance-based costs under ground-state initialization, with residual $C_{\mathrm{var}}\approx 0.5$ and $\gtrsim 25$–$35\%$ weight in $2$QP/$3$QP sectors; near-zero initialization yields $C_{\mathrm{var}}\approx 0.01$ and subspace infidelity $\sim 10^{-4}$, while trace-based costs remain robust [2511.06623].

The trapped-ion NLCE+QA study addresses the separate question of whether current hardware can supply expectation values accurate enough to survive PCAT’s non-linear classical post-processing. PCAT involves matrix inversion and matrix square roots, and these operations amplify measurement noise. The paper quantifies this by Monte Carlo propagation with $M_{\rm mc}=10^4$ samples. Experiments used $2000$ shots per matrix element, totaling $\sim 5\times 10^6$ shots across all runs on a $20$-qubit trapped-ion QPU with single-/two-qubit gate fidelities $99.96\%/98.47\%$. For one-dimensional TFIM at $J/h=0.3$, statevector NLCE+VQE overlaps perfectly with the analytic dispersion and QPU NLCE+VQE approaches it closely; at $J/h=0.8$ and especially $J/h=1$, noise amplification becomes more visible. In the one-dimensional TFIM with longitudinal field at $J/h=0.5$, $h_l=0.1$, the modified-state correction is mathematically required, but numerically $\Delta_{ij}\approx 4\times 10^{-4}$, below the sampling-noise floor at $2000$ shots per element and therefore negligible in practice. The ladder benchmark at $J/h=0.2$ and $N_{\max}=8$ resolves two one–quasi-particle bands with moderate upward bias. The paper attributes the characteristic upward shift and vertical compression of the dispersion to effective depolarization rates satisfying $p_H=p_E+\delta$ with $\delta>0$, and concludes that ASP requires one to two orders of magnitude lower gate depolarization rates than current reference values to be reliably usable without strong mitigation [2605.28599].

Across these formulations, the domain of validity is consistently stated. PCAT is most natural in gapped phases with well-defined quasi-particles and adiabatic connectivity of the target sector. It requires a non-singular subtraction operator in the general multiparticle construction. Near criticality, under avoided level crossings, or when particle decay processes render the subtraction ill-conditioned, the relevant singular values can drop sharply and NLCE convergence can deteriorate. In quantum-algorithm settings, inaccurate identification of the low-energy subspace, optimizer trapping, and hardware noise can degrade decoupling quality and violate additivity. Extending the method to higher-particle sectors is possible in principle, but increases the subspace dimension and measurement overhead.

Source: https://www.emergentmind.com/topics/projective-cluster-additive-transformation-pcat