THC-sRI-CC2: Hybrid Reduced-Scaling CC2 Method

• The paper demonstrates that THC-sRI-CC2 mitigates stochastic errors by deterministically factorizing the Coulomb term and applying stochastic resolution for the exchange term.
• The method couples tensor hypercontraction and stochastic estimation to maintain an overall O(N³) scaling, improving energy and gradient accuracy in CC2 computations.
• Benchmark results indicate a 5–10× noise reduction in energy derivatives compared to sRI-CC2, enabling efficient excited-state and nonadiabatic dynamics simulations.

THC-sRI-CC2 is a noise-reduced formulation of stochastic resolution of identity applied to CC2 in which tensor hypercontraction (THC) is used to decouple the expensive Coulomb term and stochastic resolution of the identity (sRI) is used to treat the time-determining exchange term, while preserving an overall $O(N^3)$ scaling with system size $N$ (Zhao et al., 26 Sep 2025). The method is designed to mitigate the substantial stochastic errors that conventional sRI-CC2 can introduce in energy derivatives, particularly in contexts such as molecular dynamics simulations, and benchmarks report greater accuracy and markedly reduced stochastic noise than conventional sRI-CC2 at identical computational samplings (Zhao et al., 26 Sep 2025).

1. Definition and computational rationale

The sRI approximation reduces the computational scaling of CC2 from $O(N^5)$ to $O(N^3)$, where $N$ is a measure of system size, but its inherent stochastic noise can introduce substantial errors in energy derivatives (Zhao et al., 26 Sep 2025). THC-sRI-CC2 addresses this limitation by combining two distinct low-scaling decompositions within the same CC2 workflow.

In this formulation, the expensive Coulomb term, which scales as $O(N^4)$, is decoupled via THC, while the time-determining exchange term with an $O(N^5)$ cost is addressed through the sRI scheme, collectively yielding an overall $O(N^3)$ scaling (Zhao et al., 26 Sep 2025). The central design choice is therefore not a uniform stochastic treatment of all contractions, but a hybrid partitioning in which the deterministic THC factorization removes the dominant source of noise from the Coulomb contribution and leaves stochastic sampling only in the exchange contribution.

This suggests that the defining feature of THC-sRI-CC2 is not merely reduced asymptotic cost, but a redistribution of approximation error: deterministic low-rank compression for the Coulomb structure and tunable stochastic estimation for the exchange structure.

2. CC2 equations and the key contraction

The method is built from the CC2 working equations. With $|HF\rangle$ denoting the Hartree–Fock determinant, the CC cluster operator is defined as

$T = T_1 + T_2,$

with

$N$0

and

$N$1

The CC2 ground-state energy and amplitude equations follow from the similarity-transformed Hamiltonian

$N$2

Projecting onto singles and doubles yields

$N$3

$N$4

and

$N$5

Once the doubles amplitudes $N$6 are found by solving the doubles equation, one back-substitutes to obtain the singles amplitudes and then the CC2 energy (Zhao et al., 26 Sep 2025). In practical form, the key tensor contraction appearing in $N$7 or in excited-state Jacobian contractions is

$N$8

This contraction separates naturally into a Coulomb-like part and an exchange-like part. The hybrid structure of THC-sRI-CC2 follows directly from this decomposition: the Coulomb contribution is factorized deterministically, while the exchange contribution is estimated stochastically.

3. Tensor hypercontraction for the Coulomb contribution

Tensor HyperContraction factorizes a four-index electron-repulsion quantity $N$9 by introducing four low-rank interpolation matrices and a small core matrix:

$O(N^5)$0

A common choice, identified as Interpolative Separable Density Fitting, is

$O(N^5)$1

so that

$O(N^5)$2

All five tensor factors $O(N^5)$3 scale as $O(N^5)$4 elements, and $O(N^5)$5 (Zhao et al., 26 Sep 2025). Inserting this factorization into the Coulomb part of the key CC2 contraction and regrouping yields an $O(N^5)$6 form:

$O(N^5)$7

where

$O(N^5)$8

and $O(N^5)$9 run over occupied orbitals while $O(N^3)$0 and $O(N^3)$1 are active indices.

The stated implementation precomputes the $O(N^3)$2 intermediates

$O(N^3)$3

then

$O(N^3)$4

so that all sums reduce to $O(N^3)$5 steps (Zhao et al., 26 Sep 2025). The significance of this deterministic treatment is that the Coulomb part no longer contributes stochastic variance.

4. Stochastic resolution of the identity for the exchange contribution

The exchange contraction in the key CC2 tensor expression involves two distinct four-index integrals and is formally $O(N^3)$6 (Zhao et al., 26 Sep 2025). THC-sRI-CC2 treats this part with stochastic resolution of the identity.

The sRI procedure draws $O(N^3)$7 random auxiliary-space vectors $O(N^3)$8, for $O(N^3)$9, satisfying

$N$0

For each $N$1, one builds the stochastic RI tensor

$N$2

so that

$N$3

The exchange contribution is then written as

$N$4

which, after rearranging inner sums over $N$5, is realized in $N$6 (Zhao et al., 26 Sep 2025).

Variance analysis is given explicitly. For each fixed index tuple $N$7,

$N$8

and by standard sampling theory

$N$9

Thus the per-element stochastic error is $O(N^4)$0 and is tunable by $O(N^4)$1 (Zhao et al., 26 Sep 2025). In this respect, THC-sRI-CC2 retains the controllable-error character of sRI while confining it to the exchange sector.

5. Scaling and noise characteristics

The asymptotic cost analysis is stated in terms of a system-size measure $O(N^4)$2, such as the number of basis functions or electrons (Zhao et al., 26 Sep 2025). For the THC steps associated with the Coulomb term, building $O(N^4)$3 and $O(N^4)$4 scales as

$O(N^4)$5

building $O(N^4)$6 and $O(N^4)$7 scales as

$O(N^4)$8

and the final contraction scales as

$O(N^4)$9

For the sRI steps associated with the exchange term, forming $O(N^5)$0 costs

$O(N^5)$1

since $O(N^5)$2, and the exchange contraction scales as

$O(N^5)$3

Because $O(N^5)$4, $O(N^5)$5, and $O(N^5)$6 are all proportional to $O(N^5)$7 and treated as linear in $O(N^5)$8, every stage grows as $O(N^5)$9, giving overall $O(N^3)$0 scaling (Zhao et al., 26 Sep 2025).

The noise-reduction argument is correspondingly explicit. In pure sRI-CC2, both Coulomb and exchange carry stochastic error $O(N^3)$1. In THC-sRI-CC2, the Coulomb part is entirely deterministic, so the only fluctuation stems from the much smaller exchange contribution. Denoting

$O(N^3)$2

one has in sRI-CC2

$O(N^3)$3

whereas in THC-sRI-CC2

$O(N^3)$4

Numerically $O(N^3)$5, so eliminating the Coulomb noise often reduces the total $O(N^3)$6 by an order of magnitude. Quantitatively,

$O(N^3)$7

A plausible implication is that the principal practical gain of the method is not only asymptotic efficiency, but also improved stability of derivative-level quantities at fixed sampling.

6. Benchmarks and stated scope

The reported benchmarks compare conventional sRI-CC2, THC-sRI-CC2, and RI-CC2 (Zhao et al., 26 Sep 2025).

Quantity	sRI-CC2	THC-sRI-CC2
H$O(N^3)$8O excitation energy, $O(N^3)$9	$|HF\rangle$0 eV	$|HF\rangle$1 eV
Benzene excitation energy, $|HF\rangle$2	$|HF\rangle$3 eV	$|HF\rangle$4 eV
H$|HF\rangle$5O $|HF\rangle$6 ground-state gradient, $|HF\rangle$7	$|HF\rangle$8	$|HF\rangle$9
Benzene $T = T_1 + T_2,$0 ground-state gradient, $T = T_1 + T_2,$1	$T = T_1 + T_2,$2	$T = T_1 + T_2,$3

For the same excitation-energy benchmark, the RI-CC2 reference values are $T = T_1 + T_2,$4 eV for H$T = T_1 + T_2,$5O and $T = T_1 + T_2,$6 for benzene (Zhao et al., 26 Sep 2025). The benchmark summary further states “$T = T_1 + T_2,$7 reduced by $T = T_1 + T_2,$8” and “abs. error $T = T_1 + T_2,$9 eV” for excitation energies. For ground-state gradients, the summary states “$N$00–$N$01 noise reduction,” and excited-state gradients and derivative couplings are reported to exhibit similar $N$02–$N$03 drops in stochastic error at identical $N$04 (Zhao et al., 26 Sep 2025).

Scaling timings on all-E olefins are reported as

$N$05

with crossover around $N$06–$N$07 electrons, beyond which the cubic methods become faster by factors of $N$08–$N$09 at $N$10 (Zhao et al., 26 Sep 2025).

The stated implications for application are direct. By combining the mathematical rigor of THC for Coulomb decoupling with the flexibility of sRI for exchange, THC-sRI-CC2 delivers full $N$11 scaling for energies, gradients, and derivative couplings; noise levels comparable to pure RI-CC2 with $N$12–$N$13 fewer stochastic samples; and wall-clock times $N$14–$N$15 faster than RI-CC2 for $N$16 (Zhao et al., 26 Sep 2025). The same source states that this opens the door to routine CC2-quality excited-state and nonadiabatic dynamics in systems of several hundred atoms, with systematically controllable stochastic error and no loss in formal accuracy, and further establishes a general THC-sRI hybrid strategy for the development of reduced-scaling electronic structure methods.