Graph-Based Block-Diagonalization (GBBD)

• Graph-Based Block-Diagonalization (GBBD) is a matrix reorganization method that uses a graph representation to partition a sparse Hamiltonian into independent blocks based on connectivity.
• The method enables block-wise diagonalization by converting non-zero Hamiltonian elements into graph edges, facilitating efficient extraction of low-energy spectra in molecular systems.
• GBBD extends beyond quantum chemistry to applications like sparse PCA and graph-spectrum analysis, though its effectiveness hinges on the inherent sparsity and structure of the matrix.

Graph-Based Block-Diagonalization (GBBD) denotes a family of matrix-reorganization methods in which a graph induced by a matrix, operator, or partition is used to expose independent blocks that can be treated separately. In the formulation developed for full configuration interaction (FCI) Hamiltonians of molecular systems, each reduced-basis configuration is represented as a node, each non-zero Hamiltonian matrix element defines an edge, and the connected components of the resulting graph determine the independent blocks of a permuted Hamiltonian (Park et al., 31 Jul 2025). Applied to linear hydrogen chains from $\mathrm{H}_2$ to $\mathrm{H}_{12}$, this construction yielded exact low-energy spectra up to the tested limits, with excellent agreement with direct reduced-Hamiltonian diagonalization for chains up to $\mathrm{H}_{10}$ (Park et al., 31 Jul 2025). In a broader technical sense, the same label also appears in graph-spectrum and covariance-thresholding settings, where graph partitions or connected components again drive a block decomposition (Xu et al., 28 Nov 2025).

1. Definition and matrix-to-graph correspondence

In the molecular GBBD construction, the starting point is the many-electron Hamiltonian in second quantization, expressed in a finite one-electron basis of $n$ molecular orbitals. The full Fock-space dimension is $L=2^n$, but the Hamiltonian is first restricted to the fixed-electron-number sector with $m$ electrons, giving a reduced Hilbert-space dimension

$L_{\mathrm{red}}=\binom{n}{m}.$

Each basis state $|X\rangle$ in this reduced FCI space is treated as a graph vertex, and an undirected edge is introduced whenever the Hamiltonian matrix element between two basis states is non-zero: $(i,j)\in E \iff H_{ij}\neq 0,\quad i\neq j.$ The resulting graph $G=(V,E)$ is typically disconnected because the reduced Hamiltonian is extremely sparse (Park et al., 31 Jul 2025).

The disconnected components of this graph are the central structural objects. If $\mathrm{H}_{12}$0 decomposes as a disjoint union of connected components $\mathrm{H}_{12}$1, then there exists a permutation of the determinant basis for which the reduced Hamiltonian becomes block diagonal,

$\mathrm{H}_{12}$2

with one block for each connected component. In this formulation, GBBD is not defined by a group-theoretic symmetry reduction beyond fixed particle number. The only symmetry explicitly imposed is the restriction to the $\mathrm{H}_{12}$3 sector; the subsequent partition is determined by the actual connectivity pattern of the Hamiltonian in the chosen determinant basis (Park et al., 31 Jul 2025).

The sparsity underlying this graph construction arises from the locality of one- and two-electron integrals together with the Slater–Condon rules, according to which $\mathrm{H}_{12}$4 is non-zero only when determinants differ by at most two orbital occupations. This makes the graph representation natural rather than merely heuristic: vanishing off-block matrix elements are exactly the statement that no path of non-zero couplings links the corresponding basis subsets (Park et al., 31 Jul 2025).

2. Construction of blocks and spectral workflow

The Hamiltonian construction in the hydrogen-chain study uses OpenFermion to generate the ab initio molecular Hamiltonian, construct its second-quantized form, and map it to a matrix in the Fock-space basis. After restricting to the fixed-$\mathrm{H}_{12}$5 sector, only the indices and values of the non-zero matrix elements are stored, yielding a sparse representation of the reduced Hamiltonian (Park et al., 31 Jul 2025).

The graph is then assembled from these non-zero entries. NetworkX is used to find connected components through networkx.connected_components, and the basis is reordered so that all configurations belonging to the same component are contiguous. The reduced Hamiltonian is thereby reconstructed in a block-diagonalized form in which each block, or cluster, can be treated independently and can have a different size (Park et al., 31 Jul 2025).

The low-energy eigenspectrum is obtained blockwise. Each block matrix is solved independently, and the union of the resulting eigenvalues is sorted to recover the global low-energy sector. For the systems studied, the ground state and first excited state were found in the largest and second-largest block matrices, respectively. When multiple states must be extracted from the same block, the method introduces a penalty-modified Hamiltonian

$\mathrm{H}_{12}$6

where $\mathrm{H}_{12}$7 is a previously computed eigenstate and $\mathrm{H}_{12}$8 is a penalty weight. This raises the energy of the known state and enforces orthogonality while searching for the next state within the same block (Park et al., 31 Jul 2025).

The eigensolver workflow in the hydrogen-chain application is hybrid. A Quantum Eigensolver on D-Wave hardware is used to obtain block ground states, and the converged quantum result is then supplied as the initial vector of a classical Lanczos step for error reduction. In the reported $\mathrm{H}_{12}$9 example, around 100 measurements were needed for convergence, with sampling error of order $\mathrm{H}_{10}$0. The paper also states that the QE algorithm exhibits linear scaling $\mathrm{H}_{10}$1 with respect to the matrix size, where $\mathrm{H}_{10}$2 denotes the block dimension (Park et al., 31 Jul 2025).

An important structural simplification follows from the disconnected support of different blocks. Eigenvectors belonging to different blocks are automatically orthogonal in the full reduced space, so explicit orthonormality constraints are only needed for successive states extracted within the same block, not across blocks (Park et al., 31 Jul 2025).

3. Hydrogen-chain application and physical results

The explicit benchmark system for molecular GBBD is the family of linear hydrogen chains

$\mathrm{H}_{10}$3

The study varies the H–H bond length and analyzes ground and excited energies as functions of bond length, with a recurring optimal bond length of approximately

$\mathrm{H}_{10}$4

at which the ground-state energy per $\mathrm{H}_{10}$5 molecule is minimized for all chains considered (Park et al., 31 Jul 2025).

The block structure grows with system size but remains substantially smaller than the full Fock-space Hamiltonian. The full Hamiltonian dimension scales as $\mathrm{H}_{10}$6, the reduced fixed-$\mathrm{H}_{10}$7 dimension as $\mathrm{H}_{10}$8, and the blocks associated with the ground-state and first-excited-state configurations also increase exponentially with the number of $\mathrm{H}_{10}$9 units, though they remain significantly smaller than the full Hamiltonian itself. These dominant blocks are also described as highly sparse (Park et al., 31 Jul 2025).

Accuracy was assessed against exact diagonalization of the reduced Hamiltonian using SciPy. For chains up to $n$0, GBBD ground and excited energies showed excellent agreement with exact results. For $n$1, direct comparison to exact reduced-Hamiltonian diagonalization was not possible on the reported hardware because of memory limitations caused by the massive Hamiltonian matrix, but GBBD still produced low-energy spectra on the individual blocks (Park et al., 31 Jul 2025).

Several physical trends were extracted from the resulting spectra. At large bond lengths, the energy per $n$2 molecule converges to the same value across chain lengths, reflecting complete separation of the H atoms. At the optimal bond length, the ground-state energy per $n$3 increases, becoming less negative, as the number of $n$4 molecules increases, which the paper interprets as indicating that longer $n$5 chains are less stable than the isolated molecule. The energy gap

$n$6

decreases with increasing chain length; the paper notes that in the limit of an infinite chain, $n$7 may converge to zero, implying metallic behavior (Park et al., 31 Jul 2025).

The low-energy spectrum also becomes denser with increasing chain length. At $n$8, the energy gaps between several eigenvalues in $n$9 are more compressed than those in $L=2^n$0. The authors further state that the dominant low-energy eigenstates are localized within the largest connected components of the reduced configuration space, linking the graph decomposition to the effective localization of low-energy physics in determinant space (Park et al., 31 Jul 2025).

4. Structural interpretation and common misconceptions

A common misconception is to treat GBBD as simply another name for symmetry-based block diagonalization. In the molecular formulation this is not how the method is defined. Fixed particle number is imposed first, but the subsequent partition is produced by graph connectivity of the actual sparsity pattern. The paper explicitly does not invoke spin symmetry, point-group symmetry, or translational symmetry in the construction. The partition can therefore be finer than conventional symmetry sectors whenever the determinant-space graph disconnects for reasons not captured by standard representation-theoretic labels (Park et al., 31 Jul 2025).

A second misconception is that GBBD removes the exponential complexity of exact many-body methods. It does not. The largest blocks still grow exponentially with system size, and the paper explicitly treats this as a limitation: GBBD delays but does not remove the exponential wall. Its gain is structural and practical rather than asymptotic in the strongest sense. Memory usage is reduced by fixed-$L=2^n$1 restriction, sparse storage, and blockwise processing, but very large blocks can still become intractable (Park et al., 31 Jul 2025).

A third misconception is that disconnected graph components are guaranteed. They are not. The method is effective when the reduced Hamiltonian is sparse and decomposes into disconnected components. If the graph is connected, there is only one block, and GBBD yields no reduction beyond the fixed-particle-number restriction. The paper does not offer a full classification of when a determinant graph is disconnected or nearly connected; it states only the constructive criterion implied by the graph itself (Park et al., 31 Jul 2025).

These points clarify the proper scope of the method. GBBD is best understood as a sparsity-pattern decomposition of a matrix in a chosen basis. This suggests a distinction between two kinds of structure: symmetry-imposed invariance, which is basis-independent in a representation-theoretic sense, and connectivity-imposed decomposability, which depends on the realized zero pattern of the operator in the working basis. In the hydrogen-chain study, GBBD exploits the latter (Park et al., 31 Jul 2025).

5. Broader technical meanings of GBBD

The label GBBD is not confined to molecular Hamiltonians. It also appears in graph-spectrum theory and sparse covariance decomposition, where the graph object differs but the core operation remains the extraction of block structure from graph connectivity or graph-induced partitions.

Domain	Graph object	Resulting block structure
FCI Hamiltonians	Determinant graph with edges for $L=2^n$2	Connected components become Hamiltonian blocks (Park et al., 31 Jul 2025)
Sparse PCA	Feature graph with edges where $L=2^n$3 after thresholding	Connected components become principal submatrix blocks (Pia et al., 2024)
Generalized block diagonal Laplacian spectrum	Degree-partition blocks $L=2^n$4	Orthogonal block-preserving similarity on partition subspaces (Xu et al., 28 Nov 2025)

In sparse PCA, the block-diagonalization framework is explicitly graph-based. A covariance matrix $L=2^n$5 is thresholded at level $L=2^n$6, producing $L=2^n$7, and a graph is built on feature indices with an edge whenever $L=2^n$8. Connected components of this graph define the blocks of a permuted block-diagonal approximation $L=2^n$9. The paper proves that thresholding to $m$0 already achieves the optimal largest block size $m$1 in time $m$2, and that the resulting framework reduces the runtime of an underlying sparse-PCA solver to

$m$3

with reported average speedup factor $m$4 and average approximation error $m$5 for exact sparse-PCA experiments (Pia et al., 2024).

In graph-spectrum theory, the phrase is used in a more algebraic sense. The generalized block diagonal Laplacian spectrum retains only the diagonal block all-ones matrices $m$6 associated with a degree partition $m$7, replacing the $m$8 matrices $m$9 of the generalized block Laplacian spectrum by only $L_{\mathrm{red}}=\binom{n}{m}.$0 matrices. The corresponding multivariate matrix

$L_{\mathrm{red}}=\binom{n}{m}.$1

is real symmetric for real parameters, hence has real spectrum. The central characterization states that equality of the associated multivariate characteristic polynomials for two graphs is equivalent to the existence of an orthogonal matrix $L_{\mathrm{red}}=\binom{n}{m}.$2 satisfying

$L_{\mathrm{red}}=\binom{n}{m}.$3

so the spectral data encode a partition-preserving orthogonal similarity (Xu et al., 28 Nov 2025).

This broader usage suggests that GBBD should be treated as a family of graph-induced decomposition principles rather than as one fixed algorithm. In one setting, connectivity of a sparse determinant graph defines blocks directly; in another, thresholded covariance connectivity identifies decoupled feature groups; in another, a prescribed graph partition induces a block-diagonal spectral family whose invariants are studied through orthogonal similarity (Park et al., 31 Jul 2025).

6. Limitations, extensions, and relation to adjacent block-diagonalization literatures

The clearest limitation in the molecular setting is the dependence on sparsity and decomposability. If the reduced Hamiltonian graph is fully connected, GBBD does not reduce the problem. Even when it does, the largest blocks can still grow exponentially, so the method is a structural compression rather than a full complexity-theoretic resolution (Park et al., 31 Jul 2025).

A second limitation emerges in the graph-spectrum literature. For undirected graphs, the generalized block diagonal Laplacian spectrum shows that diagonal block all-ones matrices $L_{\mathrm{red}}=\binom{n}{m}.$4 are sufficient to recover the same orthogonal-similarity characterization previously obtained from all $L_{\mathrm{red}}=\binom{n}{m}.$5. For digraphs with Hermitian adjacency matrices, the authors explicitly state that they fail to discard the off-diagonal block all-one matrices. In that sense, block-diagonal information alone is sufficient in the undirected real-symmetric setting but not, in the present proof framework, in the directed Hermitian setting (Xu et al., 28 Nov 2025).

A third limitation concerns empirical scope. The molecular paper develops GBBD for the FCI Hamiltonian of molecular systems in general, but its direct evidence is confined to linear hydrogen chains. The claim of broader applicability is therefore conceptual rather than benchmarked across diverse chemical systems (Park et al., 31 Jul 2025).

The relation to non-graph block-diagonalization literatures is complementary rather than redundant. Simultaneous block diagonalization via congruence, for example, can detect hidden invariant subspaces not visible to permutation-based sparsity partitions, because its blocks arise from central idempotents in a Jordan-algebraic center rather than from connected components of an explicit sparsity graph (Fang et al., 3 Mar 2025). This suggests that GBBD, in the strict graph sense, is strongest when zero patterns themselves are physically or statistically meaningful, while congruence- or similarity-based approaches can reveal finer algebraic decompositions even when the sparsity graph is connected.

A plausible implication is that graph-derived blocks may serve as preprocessing units for more elaborate solvers. In the molecular case, the paper itself mentions selected CI, DMRG, and tensor-network methods only as possible broader context rather than implemented hybrids, so any such use remains prospective rather than demonstrated (Park et al., 31 Jul 2025). What is already established is narrower and concrete: GBBD is an exact block extraction procedure whenever graph connectivity truly disconnects the operator in the chosen basis, and it can turn a single large spectral problem into a family of smaller ones whose physical content remains transparent.