Tensor Product Selected CI (TPSCI)

• TPSCI is a selected configuration interaction method that builds the many-electron basis from tensor products of correlated local cluster states.
• It partitions the active space into chemically meaningful clusters and incorporates local correlations via methods like cluster mean-field, yielding a compact and interpretable representation.
• Its adaptive CIPSI-like selection workflow, excited-state extensions, and HPC implementations demonstrate both practical efficiency and robust chemical insights.

Searching arXiv for Tensor Product Selected Configuration Interaction and closely related work. Tensor Product Selected Configuration Interaction (TPSCI) is a selected configuration interaction method in which the global many-electron basis is built from tensor products of correlated local cluster states rather than ordinary Slater determinants. In the chemistry formulation developed from 2020 onward, the active space is partitioned into chemically meaningful clusters, local correlation is incorporated into the cluster basis through cluster mean-field or related constructions, and a CIPSI-like adaptive search is then performed over tensor-product configurations. The resulting representation is designed to be compact and chemically interpretable, particularly when strong correlation is predominantly local and the remaining problem is inter-cluster entanglement, spin coupling, and charge delocalization (Abraham et al., 2020).

1. Historical development and scope

TPSCI was introduced in 2020 as a deterministic selected-CI method in a basis of cluster state tensor products, with the stated goal of exploiting local molecular structure to reduce the number of selected-CI variables required for strongly correlated systems. The original presentation emphasized modified Hubbard models, bond breaking in N$_2$ and F$_2$, and planar $\pi$-conjugated systems with active spaces up to $42$ electrons in $42$ orbitals (Abraham et al., 2020).

The method was generalized in 2023 to molecular excited states. That extension retained the tensor-product cluster basis and CIPSI-like selection machinery, but targeted multiple roots, introduced excitonic initial spaces such as TPS-single exciton (TPS-SE), and combined the variational selection with state-specific PT2 and optional state-averaged HOSVD updates. The same work reported a more efficient Julia implementation based on ClusterMeanField.jl and FermiCG.jl (Braunscheidel et al., 2023).

A 2024 study placed particular emphasis on wavefunction analysis and chemical interpretability. It argued that TPSCI compresses correlated wavefunctions into a smaller number of physically meaningful local many-body configurations and illustrated this through direct inspection of TPS coefficients, Bloch effective Hamiltonians, and cluster correlation functions in systems ranging from a tetracene tetramer to an open-shell Cr$_2$ complex and hexabenzocoronene (Braunscheidel et al., 2024).

In 2025, TPSCI was benchmarked against DMRG for exchange coupling constants in six transition-metal systems, including dinuclear Cr, Fe, and Mn complexes and a tetranuclear Ni-cubane. That study focused on the method’s multistate character, the behavior of magnetic exchange constants $J$, and the limitations imposed by truncation of local cluster states (Bachhar et al., 18 Aug 2025).

A distinct 2025 implementation paper treated a tensor-product bitstring SCI/TPSCI framework oriented toward distributed-memory classical HPC. That work is explicitly described as an implementation and scaling paper for the tensor-product selected configuration interaction idea used in IBM’s 2024 “QC+HPC” scheme, and its central contribution is a distributed SCI engine rather than a new selection criterion (Xu et al., 13 Mar 2025).

2. Formal structure of the tensor-product basis

TPSCI starts from the usual active-space Hamiltonian,

$$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$

and partitions the active orbitals into clusters $I,J,\dots$ chosen so that interactions are strongest within a cluster and weaker between clusters. A global basis function is then a tensor product of local cluster states,

$$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$

and the exact wavefunction is expanded in that basis. Because each cluster state is itself a linear combination of local determinants, intra-cluster correlation is built into the basis before the global selected-CI step (Bachhar et al., 18 Aug 2025).

The Hamiltonian is partitioned by cluster content as

$_2$0

with at most four-cluster terms because the ab initio Hamiltonian contains only two-body interactions. Matrix elements are expressed through precomputed local operator tensors $_2$1, and non-active clusters contribute only overlaps. This produces a TPS analogue of Slater–Condon sparsity, but the matrix elements are more expensive than in determinant CI because they require tensor contractions over local operator matrices (Abraham et al., 2020).

The local basis is commonly generated from cluster mean-field (cMF) theory. For cluster $_2$2,

$_2$3

where $_2$4 is the 1RDM of cluster $_2$5. The resulting cMF product state is the cluster analogue of a Hartree–Fock reference, and the generalized Brillouin condition expresses its stationarity with respect to local excitations (Braunscheidel et al., 2024).

Practical TPSCI calculations truncate the local cluster spaces. Each cluster has many Fock sectors, and the method retains only a limited number $_2$6 of the lowest cMF eigenstates per allowed sector, often within a particle-number window controlled by $_2$7. To preserve spin symmetry during truncation, the missing $_2$8 components are generated by applying $_2$9 to a seed state, starting from $\pi$0 for even-electron clusters and $\pi$1 for odd-electron clusters. In the 2025 transition-metal study, the cMF optimization yielded a spin-pure restricted open-shell reference (RO-cMF), which was important for open-shell complexes (Bachhar et al., 18 Aug 2025).

A recurrent point in the literature is that TPSCI is not simply a one-particle basis rotation. The many-body basis itself is changed: each basis object is a correlated local many-body state, so a single tensor-product configuration can subsume correlation that would require many determinants in ordinary SCI (Abraham et al., 2020).

3. Selected-CI workflow, perturbation theory, and multistate extensions

The core TPSCI algorithm is CIPSI-like. In the chemistry formulation, the workflow is: diagonalize $\pi$2 in the current variational $\pi$3-space of tensor-product configurations; apply $\pi$4 to the current variational eigenvectors to identify couplings to the external $\pi$5-space; estimate first-order amplitudes, commonly via Löwdin partitioning; add configurations whose amplitudes exceed a threshold; and iterate until the variational space stops growing. The first-order interacting space may be screened through thresholds such as $\pi$6, and configuration selection may be based on absolute first-order coefficients, in line with traditional SCI practice (Braunscheidel et al., 2023).

After convergence of the variational space, TPSCI typically applies a second-order correction. In the 2020 formulation, the PT2 correction is evaluated from the full $\pi$7 action after setting $\pi$8; in the 2023 excited-state framework, the PT2 is state-specific and implemented with a Löwdin-type barycentric partitioning based on a cMF zeroth-order Hamiltonian. Because the external space can be very large, the 2023 work uses a batched PT2 algorithm organized over FockConfig blocks (Abraham et al., 2020).

The excited-state generalization targets multiple low-lying roots simultaneously. For $\pi$9 target states, the workflow is: define a reference $42$0-space, diagonalize in $42$1, perturbatively search $42$2, optionally update the cluster basis via HOSVD, and compute state-specific PT2 corrections. For weakly coupled clusters, the initial reference can be constructed from excitonic configurations in which one cluster is locally excited and all others remain in their ground states; this is the TPS-single exciton starting space. The paper explicitly notes that this qualitative starting point can miss charge-transfer configurations, double excitations, and more entangled higher-rank excitonic states, so iterative expansion of $42$3 remains essential (Braunscheidel et al., 2023).

TPSCI also supports local many-body basis updates through higher-order singular value decomposition. In this step, cluster reduced density matrices are built from an approximate TPSCI wavefunction, diagonalized, and used to rotate the cluster basis. For multiple states, the 2023 work uses a state-averaged cluster RDM. HOSVD is presented as a local many-body basis rotation rather than a truncation of the full TPS space, and “HOSVD bootstrapping” denotes the practical workflow of running a cheaper TPSCI, rotating the cluster basis, tightening thresholds, and rerunning (Braunscheidel et al., 2023).

4. Interpretability and physical analysis

A defining feature of TPSCI is that its basis states are local many-body objects with explicit chemical content. The 2024 interpretability study highlights local singlets and triplets, local excitons, biexcitons, oxidation-state configurations, charge-transfer states, and local excitation states as examples of physically meaningful labels that can be attached directly to dominant TPS coefficients. This is contrasted with determinant expansions, in which physical meaning is often obscured when many determinants carry similar weights (Braunscheidel et al., 2024).

The same study uses Bloch effective Hamiltonians to downfold near-exact TPSCI states into physically motivated model spaces. If $42$4 spans a model space with large overlap with the targeted TPSCI states, then the Bloch effective Hamiltonian provides dressed matrix elements that quantify physically meaningful couplings within that space. In the tetracene and Cr$42$5 examples, this analysis was used to discuss effective singlet-exciton/biexciton coupling and effective spin couplings, respectively (Braunscheidel et al., 2024).

Cluster correlation functions provide a complementary diagnostic:

$42$6

The 2024 paper evaluates covariances of local particle number $42$7, spin projection $42$8, local spin $42$9, and the excitation projector $42$0. These observables reveal how charge, spin, and excitation entanglement are distributed across clusters (Braunscheidel et al., 2024).

The interpretability claims in the TPSCI literature are therefore not limited to compactness alone. They are tied to specific analysis tools: direct coefficient inspection, downfolded effective Hamiltonians, and cluster-resolved covariances. This suggests that TPSCI is intended not only as a numerical compression scheme but also as a representation in which local chemical language can be read directly from the wavefunction (Braunscheidel et al., 2024).

5. Applications and benchmark behavior

The 2020 paper demonstrated TPSCI on three benchmark classes. In a modified Hubbard model with $42$1, TPSCI became increasingly compact as inter-cluster coupling weakened, and for the 64-site case TPSCI with 2735 variables was reported as comparable in quality to DMRG with $42$2. In N$42$3 and F$42$4 bond breaking, TPSCI reached chemical accuracy with much smaller variational spaces than SHCI or ASCI; for example, in cc-pVDZ N$42$5 at equilibrium the listed dimensions were SHCI $42$6, TPSCI $42$7, and ASCI $42$8, while in F$42$9 at equilibrium SHCI used $_2$0 and TPSCI $_2$1. In large planar $_2$2-systems, including hexabenzocoronene with $_2$3, TPSCI remained competitive with SHCI while using much smaller variational spaces (Abraham et al., 2020).

The 2023 excited-state generalization treated polycyclic aromatic hydrocarbons labeled P1–P5 and a tetracene tetramer with a $_2$4 active space partitioned into four $_2$5 clusters. For P1–P4 the paper computed 16 excited states; for P5, 24 excited states. In P5, the TPSCI variational dimension was $_2$6, compared with $_2$7 for SHCI, and TPSCI could compute 25 roots while SHCI reached 13 roots due to memory constraints. For the tetracene tetramer, TPSCI reported 31 low-lying eigenstates, including the ground state, 4 triplet states, 4 singlet bright states, and 18 biexcitonic states (Braunscheidel et al., 2023).

The 2024 interpretability paper examined three chemically distinct applications. In a tetracene tetramer, TPSCI recovered large-active-space singlet-fission states and used Bloch effective Hamiltonians and local excitation analysis to characterize bright states and $_2$8 biexcitons. In a tris-hydroxy-bridged Cr(III) dimer with a $_2$9 active space and five clusters, TPSCI computed low-lying $J$0 states and extracted $J$1 values around $J$2 to $J$3 cm$J$4. In hexabenzocoronene with the full $J$5-space $J$6, seven clusters of six orbitals, and local bases generated with EST, the TPSCI+PT2 energy was only about $J$7 mH away from the extrapolated limit, the variational space was about 114k TPS states, and the extrapolated total energy was reported as $J$8 au (Braunscheidel et al., 2024).

The 2025 transition-metal benchmark compared TPSCI with DMRG for six systems: $J$9 with $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$0, $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$1 with $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$2, $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$3 with $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$4 and $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$5, $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$6 with $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$7 and $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$8, $$\hat{H} = h_{pq}\hat{p}^\dagger\hat{q} + \tfrac{1}{2}\langle pq|rs\rangle \hat{p}^\dagger\hat{q}^\dagger\hat{s}\hat{r},$$9 with $I,J,\dots$0, and a tetranuclear Ni-cubane with $I,J,\dots$1. Across these systems, DMRG consistently gave lower variational energies than TPSCI, but TPSCI $I,J,\dots$2 values were generally within about $I,J,\dots$3–$I,J,\dots$4 cm$I,J,\dots$5 of DMRG in the more converged cases. The same study emphasized a practical advantage of TPSCI: because several spin states arise naturally within the same cluster basis, one can perform direct $I,J,\dots$6-extrapolation against PT2 contributions, often with smaller regression errors than subtracting independently extrapolated absolute energies (Bachhar et al., 18 Aug 2025).

6. Limitations, failure modes, and large-scale implementations

TPSCI’s limitations are stated explicitly across the literature. The 2020 paper notes that cluster choice matters substantially, that the memory cost of local tensors is significant, that exact cluster spaces become difficult beyond roughly six orbitals per cluster, that PT2 is expensive, and that the then-current code was in Python (Abraham et al., 2020). The 2023 excited-state paper adds that matrix elements are more expensive than in determinant SCI, that the method requires a meaningful cluster partitioning, that larger clusters will need approximate local solvers such as RASCI, and that non-degenerate PT2 can struggle near degeneracies, motivating a future quasidegenerate perturbation theory (Braunscheidel et al., 2023).

The 2025 transition-metal benchmark identifies the central limitation of the current chemistry implementation as cluster-state truncation. Retaining the $I,J,\dots$7 lowest cMF eigenstates per local sector is not entanglement-aware, because cMF has no knowledge of inter-cluster entanglement. As a result, important low-energy physics may reside in local states that are not among the lowest cMF eigenstates. Several failure modes are listed explicitly: in the larger $I,J,\dots$8 active space, increasing the active space made TPSCI energies worse because local truncation became more severe; in $I,J,\dots$9, agreement with DMRG in the smaller active space deteriorated in the larger active space; for complex B, increasing $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$0 from 200 to 250 barely changed $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$1, suggesting that the chosen local basis was not systematically converging the relevant physics; and for the Ni-cubane, direct convergence of the full spectrum was not possible on a single node, so a reduced effective-Hamiltonian strategy from a single $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$2 manifold was used (Bachhar et al., 18 Aug 2025).

That study also discusses possible remedies. Embedded Schmidt truncation (EST) is identified as a promising entanglement-based improvement, but it was not usable there because it breaks local $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$3 symmetry. The stated requirements for future progress are spin-preserving, entanglement-aware truncation schemes, distributed-memory implementation, and possibly a spin-constrained HOSVD rotation of cluster states (Bachhar et al., 18 Aug 2025).

The distributed-memory direction is developed in the separate 2025 SCI/TPSCI implementation paper. There, the CI vector is distributed by $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$4-string blocks rather than replicated on every node; the Hamiltonian action for Davidson iterations is organized around the tensor-product determinant basis

$$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$5

with one-sided MPI communication through MPI_GET, no global synchronization inside each fetch/computation step, and on-the-fly construction of off-diagonal Hamiltonian elements while storing only diagonals. The paper reports $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$6 determinants for N$$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$7 in aug-cc-pVDZ under $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$8, $$|\Phi\rangle = |\alpha\rangle_I |\beta\rangle_J \cdots |\delta\rangle_N,$$9 determinants for CN in cc-pVTZ under $_2$00, and discusses the $_2$01-determinant scale in the context of the largest runs enabled by the framework. On Fugaku, where nodes have 32 GiB each, the distributed design was essential; for CN, storing even one CI vector in single precision would require about $_2$02 TB on a single node (Xu et al., 13 Mar 2025).

A common misunderstanding is that lower variational energies alone determine the quality of all observables. The exchange-coupling benchmark shows a more specific picture: DMRG variational energies are lower because TPSCI is restricted by local-state truncation, yet relative spin-state energetics and extracted $_2$03 values can still remain close. Conversely, the same benchmark makes clear that this agreement is not guaranteed and must be checked carefully through convergence with respect to $_2$04, PT2 behavior, and sensitivity of adjacent spin gaps to non-Heisenberg effects (Bachhar et al., 18 Aug 2025).