---
title: Dynamically-Weighted SA Constrained CASSCF
url: https://www.emergentmind.com/topics/dynamically-weighted-state-averaged-constrained-casscf
type: topic
---

# Dynamically-Weighted SA Constrained CASSCF

Dynamically-weighted state-averaged constrained CASSCF is a family of multiconfigurational electronic-structure methods that combines state averaging, geometry-dependent state weights, and explicit constraints on the active space in order to obtain smooth, chemically meaningful descriptions of charge-transfer and related excited-state problems. In this framework, the orbital optimization is not performed for a single state alone, nor for a rigid fixed-weight average; instead, the weights evolve with the instantaneous energetic structure, while constraints enforce a targeted subspace such as donor–acceptor balance or impurity character. The resulting methodology appears in several closely related forms, including electron-transfer and hole-transfer eDSC/hDSC, the model-Hamiltonian DW-SA-cCASSCF(2,2), the multi-fragment eDSCn/hDSCn extension, and solver-independent constrained orbital-optimization variants for state-averaged CAS calculations [2409.14631, 2303.11488, 2508.21136, 2606.17761].

## 1. Conceptual foundations

State-averaged CASSCF can be formulated as an ensemble theory in which the objective is the state-averaged energy
\[
\langle H\rangle = \mathrm{Tr}(\Gamma H),
\]
with ensemble density matrix
\[
\Gamma = \sum_I p_I |\Phi_I\rangle\langle \Phi_I|.
\]
Within that perspective, the choice of weights is not a minor numerical detail; it determines the invariance properties of the self-consistent solution and the extent to which a diabatic or chemically localized representation can be obtained without breaking self-consistency [1407.6063].

In the standard microcanonical form of SA-CASSCF, all states in the support are equally weighted, so \(\Gamma = \frac{1}{M}I\) on the support. This makes the solution invariant under any unitary transformation acting locally within that supported subspace. The same paper also describes a canonical-ensemble SA-CASSCF formulation with Boltzmann-like weights,
\[
\Gamma = \frac{e^{-\beta H}}{Z},
\]
and emphasizes that the effective “electronic inverse temperature” \(\beta\) is an energy-scale parameter rather than a physical temperature. The importance of that analysis for dynamically weighted constrained methods is that it clarifies why weighting matters for diabatic analysis and for maintaining a stable, interpretable state manifold [1407.6063].

The dynamically weighted constrained methods address a practical limitation of ordinary SA-CASSCF in charge-transfer problems: the relevant diabatic states can swap order as geometry changes, a fixed average can overemphasize the wrong state, and rigid averaging can produce discontinuous orbital solutions or poor balancing of states. The constrained component is introduced because state averaging alone does not guarantee that the active orbitals remain attached to the desired donor–acceptor or impurity subspace. This suggests that dynamically weighted state averaging and explicit active-space constraints are complementary rather than redundant design choices [2508.21136, 2303.11488].

## 2. Variational structure and weighting schemes

In the two-state electron-transfer and hole-transfer formulations, the total objective is
\[
E_{\rm tot} = w_1 E_1 + w_2 E_2,
\]
with dynamic weights
\[
w_1(\Delta E) = 1 - \frac{1-e^{-\Delta E/T}}{2\Delta E/T}, \qquad
w_2(\Delta E) = \frac{1-e^{-\Delta E/T}}{2\Delta E/T},
\]
where \(\Delta E = E_2 - E_1\) and \(E_2 > E_1\). The derivatives entering the SCF gradient are
\[
w_1' = 1-\frac{1}{2}e^{-\Delta E/T}, \qquad
w_2' = \frac{1}{2}e^{-\Delta E/T}.
\]
The intended effect is explicit in the formulation: when the two states are far apart, the lower state dominates; when they become close in energy, the excited state receives more weight, improving balance near crossings [2409.14631, 2602.07746].

The multi-state generalization eDSCn/hDSCn uses
\[
E_{\rm tot} = \sum_{j=1}^{M^c} w_j E_j,
\]
with
\[
w_j = \frac{a_j}{\sum_j a_j}, \qquad
a_j = \frac{1-e^{-\Delta E_j/T}}{\Delta E_j/T}, \qquad
\Delta E_j = E_j - E_1.
\]
The corresponding differential form is
\[
dE_{\rm tot} = \sum_j w_j' \, dE_j,
\]
with explicit \(w_j'\) expressions used in the orbital optimization. Here again, the “temperature” parameter controls how sharply the lowest state is favored, and the stated purpose is to improve smoothness and balance across a charge-transfer manifold [2508.21136].

A related but distinct dynamical-weighting strategy appears in the constrained-orbital-optimization framework for state-averaged CAS calculations. There, the state-averaged energy
\[
E_{\mathrm{SA}}=\sum_\alpha w_\alpha E_\alpha[U]
\]
is replaced by a dynamically reweighted form
\[
E^{\mathrm{DW}}=\sum_{i}^{N}\omega_i\big(E^{\mathrm{CAS},i}\big)\,E^{\mathrm{CAS},i},
\]
with weights that are equal for the lowest \(L\) states, smoothly tapered for nearby higher states, and set to zero beyond an energy cutoff \(\alpha\). The paper presents this as a way to reduce discontinuities and fluctuations in state-averaged potential-energy curves when state character changes along a coordinate [2606.17761].

## 3. Constraints, active spaces, and target electronic manifolds

The defining constrained condition in the original charge-transfer formulations enforces equal projection of the active space onto donor and acceptor fragments:
\[
{\rm Tr}\!\left(\mathbf{P}_L \mathbf{P}_{\rm active} - \mathbf{P}_R \mathbf{P}_{\rm active}\right)=0.
\]
In equivalent projector notation for the two active-space density matrices,
\[
{\rm Tr}\!\left[\mathbf{Q}(\mathbf{P}_1+\mathbf{P}_2)\right]=0, \qquad \mathbf{Q}=\mathbf{P}_L-\mathbf{P}_R.
\]
This constraint is designed to suppress local excitations and retain a charge-transfer active space [2409.14631, 2602.07746].

The minimal active spaces are deliberately problem-adapted. For eDSC, the method targets one unpaired electron moving between two active orbitals, corresponding to constrained CASSCF(1,2). For hDSC, the analogous hole-transfer case corresponds to constrained CASSCF(3,2). In the multisite extension, the same logic becomes CASSCF\((1,M^c)\) for one electron and CASSCF\((2M^c-1,M^c)\) for one hole, with all configurations chosen to be mutually single excitations of each other. The compactness of this configuration manifold is treated as central to the method’s efficiency [2409.14631, 2508.21136].

For molecule-on-metal and impurity problems, the constraint takes a different but closely related form. In the partially localized DW-SA-cCASSCF(2,2) implementation, the two active orbitals \(t,u\) are constrained so that together they contain one unit of impurity population:
\[
\sum_{\mu\in\text{impurity}}\langle t|d_\mu^\dagger d_\mu|t\rangle+
\langle u|d_\mu^\dagger d_\mu|u\rangle=1.
\]
A more restrictive fully localized alternative is also defined,
\[
\sum_{\mu\in\text{impurity}}\langle t|d_\mu^\dagger d_\mu|t\rangle=1,\qquad
\sum_{\mu\in\text{impurity}}\langle u|d_\mu^\dagger d_\mu|u\rangle=0,
\]
but the partially localized form is described as smoother and more robust, especially in the two-site case [2303.11488].

The multisite formalism generalizes the fragment-balance condition to
\[
{\rm Tr}\left(\hat{P}^1\hat{P}^{\rm active}\right)=
{\rm Tr}\left(\hat{P}^2\hat{P}^{\rm active}\right)=\cdots=
{\rm Tr}\left(\hat{P}^{M^f}\hat{P}^{\rm active}\right),
\]
or equivalently
\[
{\rm Tr}\left((\hat{P}^j-\hat{P}^{j+1})\hat{P}^{\rm active}\right)=0, \qquad
j=1,\dots,M^f-1.
\]
The paper also introduces a general constraint kernel
\[
\bm{Q}^j = \sum_{i=1}^{M^f} q_{ij}\hat{P}^i,
\]
to accommodate different topologies and boundary conditions [2508.21136].

## 4. Optimization algorithms and numerical implementation

A major development in this area is the reformulation of the constrained optimization into an unconstrained SCF problem plus a constrained non-self-consistent-field subproblem. In the eDSC/hDSC algorithm, the SCF-like residual is written in commutator form as
\[
\mathbf{V}_{\rm DIIS} = [\mathbf{M}_0,\mathbf{P}_0] + [\mathbf{M}_1-\lambda \mathbf{Q},\mathbf{P}_1] + [\mathbf{M}_2-\lambda \mathbf{Q},\mathbf{P}_2],
\]
and DIIS is used to converge the outer fixed-point iteration. The inner nSCF step minimizes the auxiliary objective
\[
E_{\rm aux}= {\rm Tr}[\mathbf{M}_0\mathbf{P}_0] + {\rm Tr}[\mathbf{M}_1\mathbf{P}_1] + {\rm Tr}[\mathbf{M}_2\mathbf{P}_2]
\]
subject to the remaining constraint, with the orbitals parameterized as
\[
\mathbf{C}=\mathbf{C}_0 e^{\mathbf A}.
\]
This DIIS-SQP decomposition reduces the overall computational cost by a factor of 8 to 20 relative to directly using SQP; the paper reports reductions of about 10–20× for phenoxyl–phenol and 8–20× for amidinium–carboxylate, with an overall cost approximately twice the cost of a single ROHF calculation for the same geometry [2409.14631].

The same paper identifies a saddle-point pathology near \(D_1\)–\(D_0\) crossings and proposes a sign-flip stabilization in selected off-diagonal Fock-matrix elements before building the effective matrices. The stated purpose is to remove an unwanted branch where the optimizer can satisfy the DIIS condition at a saddle rather than the intended CASSCF stationary point. In the reported calculations, the inner nSCF step typically requires only 10 to fewer than 100 nSCF steps per SCF step and accounts for less than 5% of the total wall time [2409.14631].

The multisite extension preserves the DIIS-SQP strategy. Its formal constrained Lagrangian includes idempotency, orthogonality, and fragment-balance terms,
\[
\mathcal{L} = \sum_{j=1}^{M^c} w_jE_j
- \sum_{j=0}^{M^c} {\rm Tr}\left[\bm{\epsilon}^j((\bm{P}^j)^2-\bm{P}^j)\right]
- \sum_{j>i\geq 0}^{M^c}{\rm Tr}\left[\bm{\sigma}^{ij}(\bm{P}^i\bm{P}^j+\bm{P}^j\bm{P}^i)\right]
- {\rm Tr}\left[\left(\sum_{j=1}^{M^f-1}\lambda_j\bm{Q}^j\right)\left(\sum_{j=1}^{M^c}\bm{P}^j\right)\right].
\]
The authors describe this algorithmic combination as efficient and explicitly frame it as a route toward nonadiabatic dynamics once gradients and derivative couplings are available [2508.21136].

A more abstract constrained-orbital-optimization formulation separates the correlated electronic solver from the orbital optimizer and updates orthonormal molecular orbitals on the Stiefel manifold using implicit steepest descent,
\[
U_{k+1}(\tau)=\pi\!\left((I_n+\tau_k A(U_k))^{-1}U_k\right).
\]
In that framework, a modified DIIS stores orbital matrices and residuals \(\Delta U^k = U^{k+1}-U^k\), then re-orthonormalizes the extrapolated orbitals via polar decomposition. For CO-CAS on LiF, the paper reports that without DIIS only the smallest basis case converged within the maximum 200 macro-iterations, whereas with DIIS all tested basis sets converged within 50 macro-iterations [2606.17761].

## 5. Open-shell systems, spin–orbit coupling, and diabatization

The extension to odd-electron systems with spin–orbit coupling generalizes eDSC/hDSC from real orbitals to complex-valued Kramers-restricted spinor orbitals. In this setting, the relevant low-energy states are Kramers doublets, and the method uses a four-configuration description to represent ground–excited curve crossings between pairs of Kramers-restricted doublet states. The spinor orbitals are written as
\[
\chi_i(\mathbf r,\omega) = \psi_i^\alpha(\mathbf r)\alpha(\omega) + \psi_i^\beta(\mathbf r)\beta(\omega),
\]
with density and Fock matrices containing four spin blocks. SOC is incorporated directly through a Breit–Pauli one-electron spin-orbit Hamiltonian, and the orbital coefficients are updated by
\[
\mathbf C = \mathbf C_0 e^{\mathbf A},
\]
with \(\mathbf A\) constrained to preserve Kramers structure [2602.07746].

The purpose of dynamic weighting in this open-shell SOC setting remains the same as in the spin-free case: to bias the orbital optimization smoothly toward the physically relevant charge arrangement as the geometry changes. The associated active-space constraint still enforces balanced projection onto the donor and acceptor fragments, while the use of complex spinors makes the method compatible with explicit SOC. The numerical examples show smooth potential-energy surfaces and rapid SCF convergence across a wide range of SOC strengths, including scaled SOC factors \(\eta = 1, 50, 100, 200\); the avoided-crossing gap is reported to increase with SOC strength and to scale approximately quadratically with the SOC scaling parameter [2602.07746].

A subsequent development treats the eDSC/hDSC-generated adiabatic doublets as the input manifold for a separate diabatization step that localizes both charge and spin. The adiabatic-to-diabatic transformation is
\[
\ket{\xi_i} = \sum_{j}^{n} \ket{\psi_j} U_{ji},
\]
with \(\mathbf U\) a complex unitary matrix restricted to preserve Kramers structure. Charge localization is achieved first by maximizing dipole moments over the four-state manifold,
\[
f_1(\mathbf U)=\sum_i \sum_{\alpha=x,y,z} \left| \langle \xi_i \mid \hat{\mu}_\alpha \mid \xi_i \rangle \right|^2 + \left| \langle \bar{\xi}_i \mid \hat{\mu}_\alpha \mid \bar{\xi}_i \rangle \right|^2,
\]
followed by spin localization within each Kramers pair through
\[
f_2^i(\mathbf U)=\sum_{\alpha=x,y,z} \left| \langle \xi_i \mid \hat{S}_\alpha \mid \xi_i \rangle \right|^2 + \left| \langle \bar{\xi}_i \mid \hat{S}_\alpha \mid \bar{\xi}_i \rangle \right|^2.
\]
The optimization proceeds by Jacobi sweeps and is described as effectively equivalent to approximate joint diagonalization applied to charge and spin [2602.07750].

In that workflow, dynamically weighted state-averaged constrained CASSCF is explicitly the state-generation layer, while the complex-unitary localization is the post-SCF diabatization layer. The paper emphasizes that the resulting diabatic basis maintains exact Kramers pairing, preserves time-reversal symmetry, and yields a slowly varying pseudospin texture along the reaction coordinate. It also notes a practical limitation: because the localization is performed in a finite projected subspace, the spin expectation values are pseudospin quantities rather than exact conserved-spin observables [2602.07750].

## 6. Applications, performance, and limitations

The method family has been applied to several distinct classes of problems. In the Anderson impurity and Anderson–Holstein model setting, DW-SA-cCASSCF(2,2) is used to obtain a ground-state and an excited-state potential-energy surface for electron transfer at a metal surface. The reported outcome is that both surfaces are smooth, incorporate states with charge-transfer character, and the ground-state impurity population is in reasonable agreement with numerical renormalization group theory for correlated one-site cases. The same paper is explicit that the implementation is currently computationally inefficient, is demonstrated on model Hamiltonians rather than full ab initio surfaces, and does not yet provide gradients, derivative couplings, diabatic state construction, or bath-induced couplings for dynamics [2303.11488].

In the original odd-electron charge-transfer benchmarks, implemented in a development version of Q-Chem, the accelerated eDSC/hDSC algorithm was tested on phenoxyl–phenol with hDSC and amidinium–carboxylate with eDSC. The practical conclusion is primarily algorithmic: the DIIS-SQP scheme makes constrained charge-transfer calculations sufficiently inexpensive to be considered for future trajectory-based nonadiabatic simulations once gradients and derivative couplings are available [2409.14631].

The multisite extension eDSCn/hDSCn was demonstrated on a finite Su–Schrieffer–Heeger chain. After constrained optimization and Boys diabatization, the active orbitals become localized \(\pi\)-bond-like diabats, and the diabatic couplings decay approximately exponentially with distance; for the 28-carbon chain, the fitted decay constant is reported as
\[
\beta = 0.36~\text{\AA}^{-1}.
\]
The paper also stresses that a naive equal-projection constraint does not work well for the finite chain because of edge effects, motivating the use of mirror-symmetric generalized constraints [2508.21136].

In the solver-independent constrained-orbital-optimization framework, LiF bond stretching serves as the primary state-averaged benchmark for dynamical weighting. The paper reports that plain SA-CO-CAS produces noticeable fluctuations, whereas DW-CO-CAS yields smoother potential-energy curves and lower energies than SA-CASSCF, with the ground state lowered by about 10 mEh and the first excited state lowered by about 1 mEh in the CAS(6,6)/6-31G example. The same framework is also applied to H\(_2\)O and pyrazine, where orbital optimization lowers energies relative to fixed-orbital references and can recover lower-energy stationary solutions when conventional CASSCF converges to a higher-energy local solution [2606.17761].

Several limitations recur across the literature. The minimal active spaces that make these methods efficient are problem-specific and not universal; CASSCF(2,2) is stated to become insufficient when more than two electrons are strongly entangled in the weakly bonded two-site impurity case [2303.11488]. The open-shell SOC formulation preserves smoothness and time-reversal symmetry but produces pseudospin observables within a projected subspace rather than exact conserved-spin quantities [2602.07750]. Across multiple papers, gradients and derivative couplings are identified as the immediate next step for nonadiabatic dynamics applications [2303.11488, 2409.14631, 2508.21136].

Taken together, these developments define dynamically weighted state-averaged constrained CASSCF as a targeted methodology for cases in which fixed-weight SA-CASSCF is too rigid and unconstrained CASSCF is too ambiguous. Its central strategy is consistent across implementations: a small active manifold is chosen to represent the relevant charge-transfer sector, explicit constraints keep that manifold physically anchored, and dynamic weights adapt the orbital optimization to changing state energetics without sacrificing smoothness.

Source: https://www.emergentmind.com/topics/dynamically-weighted-state-averaged-constrained-casscf