QD-RSPT: Effective Hamiltonians for Near-Degenerate States

• QD-RSPT is a framework that treats multiple nearly degenerate low-energy states simultaneously by constructing an effective Hamiltonian within a designated model space.
• It employs model-space partitioning and block-diagonalization techniques, including Bloch–RS, van Vleck, and Schrieffer–Wolff formulations, to accurately capture strong state mixing.
• Applications span excited-state electronic structure, core-level spectroscopy, and strong-field dynamics, yielding significant computational speedups and high precision.

Quasi-degenerate Rayleigh–Schrödinger perturbation theory (QD-RSPT) is the extension of stationary Rayleigh–Schrödinger perturbation theory to situations in which a set of low-energy states is nearly degenerate and strongly mixed by the perturbation. Instead of correcting one isolated state at a time, QD-RSPT introduces a model space containing the relevant near-degenerate states, treats the complementary space perturbatively, and constructs an effective Hamiltonian whose spectrum reproduces the targeted part of the full problem. In contemporary usage, this framework appears in Bloch–Rayleigh–Schrödinger and van Vleck formulations, generalized van Vleck perturbation theory, Schrieffer–Wolff or Löwdin block diagonalization, non-orthogonal effective Hamiltonians, and state-specific iterative decoupling schemes (Bindech et al., 19 Sep 2025, Brouder et al., 2010, Day et al., 2024, Kang et al., 20 May 2025).

1. Formal setting and model-space partitioning

The standard starting point is a Hamiltonian split

$H = H_0 + V,$

with $H_0$ exactly solvable. QD-RSPT differs from non-degenerate RSPT by replacing a single reference state with a finite-dimensional model space $M$, spanned by low-energy states whose unperturbed energies are close to each other and separated from the rest of the spectrum by a finite gap. With projector

$P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$

the Hilbert space is decomposed into a model sector and an external sector, and the central object becomes an effective Hamiltonian acting in $P$ (Brouder et al., 2010).

In wave-operator language, the exact eigenvectors are written as $|\Phi_i\rangle=\Omega|\phi_i\rangle$, with $P\Omega=P$ and $\Omega P=\Omega$. The effective Hamiltonian is then

$H_{\rm eff}=PH\Omega,$

and its eigenvalues coincide with the exact eigenvalues associated with states having nonzero projection onto the model space (Brouder et al., 2010). The governing equation for the wave operator is the Kvasnička–Lindgren equation

$[\Omega,H_0]=V\Omega-\Omega V\Omega,$

or, with $H_0$0, a Riccati equation for the off-block component $H_0$1 (Brouder et al., 2010).

A closely related formulation appears in generalized van Vleck perturbation theory. There one block-diagonalizes the full Hamiltonian by a similarity transformation,

$H_0$2

so that diagonalizing the small block $H_0$3 recovers the desired energies (Neville et al., 2022). This block-diagonal viewpoint is also the basis of the Schrieffer–Wolff and Löwdin formulations treated algorithmically in modern implementations (Day et al., 2024).

The defining feature of quasi-degeneracy is therefore not exact degeneracy but the need to treat several strongly coupled low-energy states simultaneously. In practical schemes, the model space may be chosen from configuration state functions, reference-space eigenvectors, atom–photon Floquet states, or state-specific non-orthogonal determinants, depending on the physical problem (Bindech et al., 19 Sep 2025, Kang et al., 20 May 2025, Bruhnke et al., 3 Jun 2026).

2. Effective Hamiltonians and second-order structure

The canonical second-order QD-RSPT object is the model-space effective Hamiltonian. In the Bloch–Rayleigh–Schrödinger form used explicitly for low-energy electronic states,

$H_0$4

where $H_0$5 index the model space and $H_0$6 the external space (Bindech et al., 19 Sep 2025). This is the prototypical quasi-degenerate RS effective Hamiltonian: multi-state in $H_0$7, energy-denominator mediated through $H_0$8.

A symmetric second-order effective Hamiltonian also appears in generalized van Vleck perturbation theory: $H_0$9 with state-specific resolvents

$M$0

The average over $M$1 and $M$2 yields a symmetric result appropriate to the quasi-degenerate setting (Neville et al., 2022).

One recurrent issue is Hermiticity. Second-order Bloch effective Hamiltonians in multi-state problems can be non-Hermitian. A practical remedy is symmetrization,

$M$3

which was explicitly related to the canonical van Vleck QDPT expansion in a state-specific iterative decoupling scheme (Bindech et al., 19 Sep 2025). By contrast, some effective-Hamiltonian constructions are intrinsically non-Hermitian. In intense light–matter interaction, the exact effective Hamiltonian

$M$4

is generically non-Hermitian because it arises from a non-unitary similarity transformation $M$5, and the projected eigenvectors in the model space become non-orthogonal (Bruhnke et al., 3 Jun 2026).

Non-orthogonal QDPT makes this point explicit. For spin–orbit coupling in L-edge spectroscopy, a quasi-degenerate manifold of independently optimized core-hole determinants is treated by a generalized eigenvalue problem,

$M$6

with $M$7 the overlap matrix of the non-orthogonal model functions. Here the effective Hamiltonian is represented directly in a non-orthogonal basis rather than an orthonormalized model space (Kang et al., 20 May 2025). A common misconception is therefore that QD-RSPT necessarily yields a Hermitian, orthonormal reduced problem; in the modern literature, Hermitian and non-Hermitian realizations coexist, and the choice is formulation-dependent.

3. Organization of the perturbation series

The formal expansion of QD-RSPT can be written compactly in terms of the wave operator. A notable result is the representation

$M$8

where $M$9 is the set of planar binary trees with $P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$0 inner vertices (Brouder et al., 2010). Each tree determines one operator term, including its product of perturbation matrix elements and energy denominators. This gives a closed description of the general term of the Rayleigh–Schrödinger series for quasi-degenerate systems and supports resummations over specific tree classes such as left combs and right-normalized trees (Brouder et al., 2010).

A more elementary route uses Epstein–Nesbet partitioning. With diagonal perturbation elements absorbed into

$P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$1

the exact coefficient and energy equations generate both RS and BW perturbation series. A variant of Brillouin–Wigner perturbation theory was formulated recursively so that, after identifying a quasi-degenerate manifold and diagonalizing its inner block, the outer-space amplitudes and energy are updated iteratively. In the quasi-degenerate case, the second-order energy becomes

$P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$2

which was identified with the second-order multireference perturbation expression in the spirit of Chen, Davidson, and Iwata (Lee et al., 2013).

Modern algorithmic formulations emphasize asymptotic efficiency. Pymablock recasts quasi-degenerate perturbation theory as a polynomial unitary transformation $P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$3 and separates operators into selected and remaining parts. Its central recursions solve Sylvester-type equations for the off-diagonal generator and construct the effective Hamiltonian with optimal asymptotic scaling, while allowing any number of subspaces. In the two-block case, the resulting transformation was shown to be identical to the conventional Schrieffer–Wolff transformation in the standard gauge (Day et al., 2024).

A distinct but related development is state-specific iterative QD-RSPT. For a zeroth-order candidate state $P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$4, the model space is enlarged according to

$P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$5

so that states with large $P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$6 are incorporated into the model space and the second-order effective Hamiltonian is rebuilt and diagonalized repeatedly. The resulting optimized zeroth-order states are frozen in an “Optimized Zeroth-Order” space, and a Brillouin–Wigner correction is then applied state by state (Bindech et al., 19 Sep 2025). This makes the quasi-degenerate structure adaptive and target-specific rather than fixed a priori.

4. Representative physical realizations

In excited-state electronic structure, QD-RSPT underlies perturbative replacements for very large configuration-interaction diagonalizations. DFT/MRCI(2) applies generalized van Vleck perturbation theory with Epstein–Nesbet partitioning to the DFT/MRCI Hamiltonian, replacing the diagonalization of a large DFT/MRCI matrix by that of a small effective Hamiltonian. For Thiel’s test set of 28 organic molecules and 472 excited states, the reported root mean squared deviation from full DFT/MRCI was $P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$7, with 99% of excitation energies within $P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$8 and speedups of $P = \sum_i |i\rangle\langle i|,\qquad Q=1-P,$9–$P$0 for the largest systems; for large molecules, the estimated savings can reach three orders of magnitude (Neville et al., 2022).

State-specific QD-RSPT combined with Brillouin–Wigner correction has been used for low-lying singlet states of LiH and the H$P$1 ring. The method reports all RMS errors below $P$2 Hartree for three LiH singlet states and between $P$3 and $P$4 Hartree for three H$P$5 singlet states, while the BW equations converge in less than five iterations over all geometries. In H$P$6, the first and second singlet excited states become degenerate at $P$7 and $P$8, and the targeted enlargement of the model space recovers two numerically degenerate levels (Bindech et al., 19 Sep 2025).

In core-level spectroscopy, non-orthogonal QDPT serves as the bridge between scalar-relativistic OO-DFT or ROHF core-hole states and spin–orbit coupled spectra. For L-edge core electron binding energies, a quasi-degenerate manifold of six $P$9 core-hole determinants is built and then mixed by spin–orbit coupling through a NOCI generalized eigenvalue problem. Using the DCB screened 1-electron SOC operator, NO-QDPT/SCAN was reported to be accurate to about $|\Phi_i\rangle=\Omega|\phi_i\rangle$0 for L-edge CEBEs of molecules containing 3rd-row atoms, while becoming less accurate for 4th-row elements starting in the middle of the 3d transition-metal series (Kang et al., 20 May 2025).

In strong-field and quantum-optical settings, QD-RSPT provides effective Hamiltonians for Floquet problems with essential and non-essential atom–photon states. A recent formulation reconciles QD-RSPT with adiabatic elimination, derives a quasi-degenerate extension of adiabatic elimination, and shows excellent agreement with Floquet calculations at high intensities in both the low- and high-frequency regimes (Bruhnke et al., 3 Jun 2026). The same effective-Hamiltonian logic also appears in multiblock $|\Phi_i\rangle=\Omega|\phi_i\rangle$1 models, superconducting-qubit dispersive shifts, and low-energy reductions of large tight-binding Hamiltonians implemented with Pymablock (Day et al., 2024).

5. Convergence, singularities, and intruder states

The convergence of any RS-based expansion is controlled by singularities in the complex perturbation plane. For a finite-dimensional Hermitian family

$|\Phi_i\rangle=\Omega|\phi_i\rangle$2

the eigenvalues are branches of an algebraic function, and their only singularities are branch points $|\Phi_i\rangle=\Omega|\phi_i\rangle$3 where two eigenvalues coalesce. The dominant branch point determines the radius of convergence of the Taylor series for the chosen eigenbranch (Kvaal et al., 2010). A numerical procedure based on a generalized eigenvalue problem in a Kronecker-product space was proposed to approximate the complete set of such singularities, including the dominant one, without using the perturbation coefficients themselves (Kvaal et al., 2010).

This perspective is directly related to the intruder-state problem in QD-RSPT. In DFT/MRCI(2), intruders occur when a Q-space configuration becomes nearly degenerate with a model-space state so that

$|\Phi_i\rangle=\Omega|\phi_i\rangle$4

The adopted intruder-state-avoidance modification replaces

$|\Phi_i\rangle=\Omega|\phi_i\rangle$5

with $|\Phi_i\rangle=\Omega|\phi_i\rangle$6, thereby damping small denominators while keeping large-gap denominators essentially unchanged (Neville et al., 2022).

State-specific decoupling schemes address the same issue by changing the model space rather than modifying denominators. In H$|\Phi_i\rangle=\Omega|\phi_i\rangle$7, local enrichment of the model space near the crossing angles was described as improving the conditioning of the BW correction, whereas a naïve fixed-reference perturbative strategy would likely fail (Bindech et al., 19 Sep 2025). In intense light–matter interaction, a practical validity criterion is

$|\Phi_i\rangle=\Omega|\phi_i\rangle$8

over bound $|\Phi_i\rangle=\Omega|\phi_i\rangle$9-space states, with $P\Omega=P$0; this is the multilevel generalization of the usual coupling-over-detuning condition (Bruhnke et al., 3 Jun 2026).

A second misconception is that QD-RSPT only repairs small denominators after they appear. The recent literature instead treats model-space construction, block structure, and singularity avoidance as intrinsic parts of the theory. This suggests that convergence and conditioning are governed as much by the geometry of the selected subspace as by perturbative order.

6. Scope, limitations, and current directions

The present landscape shows QD-RSPT as both a formal theory and a computational technology, but also one with clear limits. State-specific RS–BW schemes currently use second-order QD-RSPT effective Hamiltonians and second-order BW energy corrections; higher-order BW was not implemented, although the reported enhanced convergence was taken to suggest that higher-order BW energy expansion might be included in the future (Bindech et al., 19 Sep 2025). In NO-QDPT for spin–orbit coupling, the treatment is restricted to a small quasi-degenerate manifold with a screened one-electron SOC operator, and the deterioration for heavier elements indicates the breakdown of that approximation (Kang et al., 20 May 2025).

Algorithmic generalizations also come with structural assumptions. Pymablock requires that the remaining part of an operator have no matrix elements within any eigensubspace of $P\Omega=P$1; otherwise the Sylvester equations and energy denominators are not well defined (Day et al., 2024). Singularity analyses further indicate that the location and character of branch points can be basis-dependent in discretized many-body models, so front-door and back-door intruder classifications need not be invariant under basis change (Kvaal et al., 2010).

There are also indications of directions not yet fully realized. A supersymmetric expansion algorithm gives a sum-free reformulation of non-degenerate RSPT and does not explicitly develop quasi-degenerate perturbation theory, but its structure was described as pointing toward a possible extension with matrix-valued superpotentials (Napsuciale et al., 14 Jan 2026). A plausible implication is that future QD-RSPT work may continue to move away from explicit sum-over-states formulas toward more structured operator equations, adaptive subspace methods, and non-orthogonal effective-Hamiltonian constructions.

In that broader sense, QD-RSPT now spans several interconnected themes: model-space partitioning, Hermitian and non-Hermitian effective Hamiltonians, recursive and graphical series organizations, intruder-state control, and scalable implementations for spectroscopy, excited states, condensed-matter reductions, and strong-field Floquet dynamics (Brouder et al., 2010, Day et al., 2024, Bruhnke et al., 3 Jun 2026).