A quadratic-scaling algorithm with guaranteed convergence for quantum coupled-channel calculations Authors: Hubert J. Jóźwiak, Md Muktadir Rahman, Timur V. Tscherbul Published: 2026 ABSTRACT Rigorous quantum dynamics calculations provide essential insights into complex scattering phenomena across atomic and molecular physics, chemical reaction dynamics, and astrochemistry. However, the application of the gold-standard quantum coupled-channel (CC) method has been fundamentally constrained by a steep cubic scaling of computational cost [O(N3)]. Here, we develop a general, rigorous, and robust method for solving the time-independent Schrödinger equation for a single column of the scattering S-matrix with quadratic scaling [O(N2)] in the number of channels. The Weinberg-regularized Iterative Series Expansion (WISE) algorithm resolves the divergence issues affecting iterative techniques by applying a regularization procedure to the kernel of the multichannel Lippmann-Schwinger integral equation. The method also explicitly incorporates closed-channel effects, including those responsible for multichannel Feshbach resonances. We demonstrate the power of this approach by performing rigorous calculations on He + CO and CO + N2 collisions, achieving exact quantum results with quadratic scaling guaranteed by a contour-integral construction. Our results establish a highly scalable computational paradigm, enabling state-to-state quantum scattering computations for complex molecular systems. INTRODUCTION Quantum collision dynamics of atoms, molecules, and nuclei is central to a vast array of fields ranging from atomic, molecular, and optical (AMO) physics (1–3) to chemical reaction dynamics (4–7), astrochemistry (8, 9), and nuclear physics (10). Precision calculations of scattering observables—such as state-to-state cross sections and reaction rates—provide essential quantitative insights into complex quantum phenomena abundant in these fields. Examples include intricate mechanisms of chemical reactions (4), the broadening of spectral lines in the atmospheres of Earth (11–14), Solar System planets (15) and exoplanets (16–19), the fate of water and small polyatomic molecules in the interstellar medium (8, 9), and the mechanisms of ultracold molecular collisions (6, 7, 20–29) and atomic few-body recombination (2, 30–32). A key challenge across these fields is the need to rigorously model ever-larger, more complex systems that remain out of reach of current computational methods. Time-independent coupled-channel (CC) calculations are the established gold standard for elucidating quantum collision dynamics in AMO physics, chemical reaction dynamics, and astrochemistry (4, 5). CC methods are uniquely suited for precision calculations of state-to-state observables (33), particularly at the low collision energies relevant for ultracold chemistry and astrochemical modeling. However, the rigorous application of CC methodology is fundamentally limited by a steep cubic scaling [O(N3)] of computational cost with the number of collision channels, N. While powerful time-dependent wave packet methods do achieve quadratic scaling [O(N2)], they are inherently ill-suited for calculating state-to-state collision dynamics at low collision energies due to the extensive spatial grids and propagation times required. As a result, complex quantum phenomena of major fundamental and applied importance, such as H2O + H2O or H2O + CH3OH collisions (34), remain completely intractable using rigorous quantum scattering methodology. Early efforts to bypass the cubic scaling barrier, most notably by Thomas (35, 36), demonstrated that the CC equations could be formally solved for a single column of the scattering S-matrix with O(N2) complexity. This reduction from cubic to quadratic scaling is possible in principle because the dynamical information contained within a single S-matrix column is sufficient to yield complete state-to-state scattering observables for molecules prepared in a specific quantum state—a scenario ubiquitous in ultracold chemistry, astrochemistry, and crossed-molecular beam collision experiments. However, despite its promise, Thomas’s iterative approach (35, 36) suffers from two critical limitations that have likely precluded its widespread adoption. First, it lacks a mechanism to incorporate closed channels, which are essential for numerical convergence and for properly capturing threshold and Feshbach resonance phenomena. Second, and more fundamentally, the underlying iterative scheme frequently diverges when applied to the deep and highly anisotropic interaction potentials characteristic of complex molecular systems (35, 36). Here, we overcome these long-standing methodological barriers to deliver a robust quadratic-scaling algorithm for solving the time-independent Schrödinger equation for a single column of the scattering S-matrix. Our approach is based on the iterative solution of the multichannel Lippmann-Schwinger (LS) equation and introduces two fundamental innovations. First, unlike previous iterative techniques, it rigorously incorporates closed channels, which are essential for numerical convergence of scattering observables. Second, to eliminate long-standing divergence issues affecting previously developed iterative expansions (35–37), we formulate a general regularization procedure in which the Weinberg eigenvalues of the LS kernel that cause the iterative series to diverge are identified and removed. The resulting Weinberg-regularized Iterative Series Expansion (WISE) algorithm guarantees numerical convergence of scattering observables regardless of the strength or anisotropy of the interaction potential. We demonstrate the robustness and quadratic scaling of the WISE algorithm by applying it to multichannel He + CO and CO + N2 collisions on ab initio potential energy surfaces (PESs). These results enable rigorous quantum scattering calculations on a wide array of molecular systems previously considered intractable, contributing to our fundamental understanding of the intricate quantum dynamics of complex molecular collisions and paving the way for large-scale numerical simulations in ultracold molecular physics, chemical reaction dynamics, and astrochemistry. RESULTS Theory Our starting point is the time-independent Schrödinger equation for the scattering of two molecules, HˆΨ=EΨ. The Hamiltonian of the collision complex is given byHˆ=−ℏ22μ∇r→2+Hˆint(ξ)+Vˆ(r→,ξ)(1)where μ is the reduced mass of the complex, r→ is the vector connecting the centers of mass of the colliding molecules, ξ denotes all internal coordinates, and Vˆ is the interaction potential. The term Hˆint describes the internal structure of the isolated monomers. In the framework of the CC approach, the total wave function Ψ is expanded in a complete basis of scattering channels (38)ΨJMJ(r→,ξ)=1r∑γ,LuγLJMJ(r)YγLJMJ(rˆ,ξ)(2)where r is the magnitude of r→, and rˆ=r→/r specifies its orientation in space. The channel basis functions YγLJMJ(ξ,rˆ) are simultaneous eigenfunctions of the internal Hamiltonian, Hˆint (indexed by γ), the end-over-end rotational angular momentum Lˆ2 (indexed by L), and the total angular momentum squared Jˆ2. The total angular momentum of the collision pair J and its projection on the quantization axis MJ are good quantum numbers in the absence of external fields. For the specific case of atom-molecule collisions, the internal index γ corresponds to the rotational angular momentum of the molecule, denoted as j. For collisions between two molecules, γ≡{jA,jB,jAB} comprises individual rotational angular momenta (jA,jB) and their vector sum (jAB). Substituting the Eq. 2 into the Schrödinger equation yields the standard set of CC equations for the radial expansion coefficients, uγLJMJ(r)d2dr2+kλ2−Lλ(Lλ+1)r2−Uλλ(r)uλ(r)=∑λ′≠λUλλ′(r)uλ′(r)(3) Here, we have introduced a collective channel index λ≡{γ,L} and suppressed the J and MJ labels for brevity. Further, kλ2=2μ(E−Eλ)/ℏ2 is the squared wave vector for channel λ and U(r)=(2μ/ℏ2)V(r) is the scaled interaction potential matrix. In Eq. 3, the terms on the left-hand side define the reference Hamiltonian, Hˆλ, for each channel, which includes the centrifugal term and the diagonal part of the interaction potential. The sum on the right-hand side contains couplings (Uλλ′) between different channels. In practice, the infinite expansion in Eq. 2 and the summation in Eq. 3 are truncated to include only a finite set of N strongly coupled channels to ensure convergence. Standard methods solve the CC Eq. 3 by propagating a matrix of N linearly independent solution vectors outward from the classically forbidden region (39–41). Matching this matrix to the proper boundary conditions yields the full scattering matrix S. Because this propagation involves matrix-matrix operations at every grid step, the computational cost scales as O(N3), effectively limiting solvable problems to a few tens of thousands of channels—to our knowledge, the largest published CC calculation to date involved N≃18,850 channels (42). Instead of solving the system of coupled second-order differential Eq. 3 directly, as done in previous work (39–41), we recast it in the form of the LS integral equation (35, 36, 43) for a single-column solution vector u→(r) corresponding to a specific incoming channel λ0u→(r)=u→0(r)+Ku→(r)(4)which can be solved iteratively using the standard Born series, u→(r)=∑n=0∞Knu→0(r) (43). Because the iterative computation involves only matrix-vector operations (i.e., the repeated application of K to u→0), it scales quadratically as O(MN2), where M is the number of iterations. For M≪N, this offers a substantial advantage over standard direct propagation techniques (39–41). However, as noted above, the naive iterative approach—originally developed by Thomas (35, 36)—suffers from two major limitations that have likely hindered its widespread use: (i) the inability to rigorously account for closed-channel effects and (ii) the divergence of the Born series for strong or attractive potentials. Here, we describe how the WISE framework overcomes these challenges through the robust, divergence-free inclusion of closed channels and the spectral regularization of the kernel matrix K. In Eq. 4, the unperturbed source term u→0(r) imposes the incoming-wave boundary condition; its components are u0,λ(r)=δλλ0xλ(r), where xλ(r) is the regular solution of the single-channel Schrödinger equation for the reference Hamiltonian Hˆλd2dr2+kλ2−Lλ(Lλ+1)r2−Uλλ(r)xλ(r)=0(5)with the boundary condition limr→0xλ(r)→0. Its asymptotic form depends on the channel energy. For open channels (kλ2>0), it takes the formxλ(r)=kλrcosϕλjL(kλr)−sinϕλnL(kλr)(6)which is appropriate when the interaction potential is negligible compared to the centrifugal term and involves the spherical Bessel (jL) and von Neumann (nL) functions; ϕλ is the elastic phase shift for the diagonal potential Uλλ(r). Since the Schrödinger equation is a second-order differential equation, there exists a second, linearly independent solution to Eq. 5, yλ(r), diverging at the origin and behaving asymptotically asyλ(r)=−ieiδλrjL(kλr)+inL(kλr)(7) The action of the integral operator K in Eq. 4 is defined by the kernelKu→(r)=∫0∞G(r,r′)Uoff(r′)u→(r′)dr′(8) Here, the matrix Uoff contains only the off-diagonal elements of the potential matrix, and G(r,r′) is the Green’s function matrix constructed explicitly from the reference solutions of Eq. 5Gλλ′(r,r′)=δλλ′xλ(r<)yλ(r>)(9)where r<=min(r,r′) and r>=max(r,r′). In practice, the radial coordinate is discretized onto a grid of Ng points, {r1,r2,…,rNg}. This discretization transforms the integral operator K into a nonsymmetric matrix of dimension (N⋅Ng)×(N⋅Ng)Kαβ=K(λ,i)(λ′,i′)=Gλλ(ri,ri′)Uλλ′off(ri′)wi′(10)where α and β are composite indices flattening the channel and grid dimensions, e.g., α=(λ,i)∈{1,…,N⋅Ng}, and wi′ are the quadrature weights associated with the grid points. The sought-after vector u→(r) determines the single column of the S-matrix corresponding to the incoming channel λ0 (35, 36)Sλλ0=ei(ϕλ+ϕλ0)δλλ0−2ikλkλ0∫0∞xλ(r)∑λ′≠λUλλ′off(r)uλ′(r)dr(11) In the following, we describe how the WISE framework overcomes the two primary challenges of the naive iterative approach (35, 36): the rigorous inclusion of closed channels and the guaranteed convergence of the iterative series through spectral regularization. Robust integration of closed channels The first major limitation of the naive iterative approach is the omission of energetically inaccessible (closed) channels with kλ2<0. In molecular scattering, these channels must be rigorously included in the basis set to achieve converged results and capture the full dimensionality of the interaction potential (44–46). Closed channels are also responsible for fundamental scattering phenomena, such as Feshbach resonances (1, 23–27). In practice, including closed channels in CC calculations can be challenging because their irregular wave functions grow exponentially in the classically forbidden region, making standard propagation schemes notoriously unstable, and necessitating advanced techniques—such as those based on the log derivative of the multichannel wave function (39–41)—to maintain numerical stability. To overcome this limitation, we developed a stable ratio-propagation method for constructing the Green’s function matrix G. First, we solve the single-channel reference equation (Eq. 5) by propagating the ratio of the regular solutions Qλ(x)(ri)=xλ(ri−1)/xλ(ri) outward from the origin using the renormalized Numerov algorithm (40), starting with Qλ(x)(r1)=0. For asymptotically open channels, we match this ratio to the asymptotic form in Eq. 6 to determine the elastic phase shift, ϕλ. Simultaneously, we determine the ratio of irregular solutions, Qλ(y)(ri)=yλ(ri+1)/yλ(ri), by starting at the final grid point, rNg, using the asymptotic form of yλ. For open channels, this is straightforward using Eq. 7. For closed channels, however, the boundary conditions are completely different—the regular and irregular solutions must behave as modified Bessel functions of the first (IL+12) and second (KL+12) kindxλ(r)=iL+1π2κλrαλIL+12(κλr)+βλKL+12(κλr),yλ(r)=−i−(L+1)κλ2πκλrαλ−1KL+12(κλr)(12)where κλ=−kλ2. The regular solution must be finite at the origin, which fixes the ratio of the coefficients αλ and βλ. Direct evaluation of these coefficients in terms of Qλ(x) is unstable due to the exponential growth of IL+1/2. To circumvent this problem, we derive a stable expression for the Green’s function at the final grid point, Gλλ(rNg,rNg), that depends only on the ratio of the regular solution, Qλ(x), and ratios and products of modified Bessel functionsGλλ′(rNg,rNg)=−δλλ′rNg(1−Qλ(x)(rNg)−rNg−1rNgIL+12(κrNg−1)IL+12(κrNg)Qλ(x)(rNg)−rNg−1rNgKL+12(κrNg−1)KL+12(κrNg))IL+12(κrNg)KL+12(κrNg)(13) Note that Gλλ′(r,r)→−1/κλ as r→∞. With the asymptotic value of Gλλ(rN,rN) established, the full Green’s function matrix elements on the spatial diagonal (r=r′) are generated by inward propagationGλλ′(ri−1,ri−1)=δλλ′Qλ(x)(ri)Gλλ′(ri,ri)Qλ(y)(ri−1)−1(14) Once the diagonal elements of G are known, the off-diagonal elements (r≠r′) are evaluated using the separability of the Green’s functionGλλ(ri,ri′)=Gλλ(ri,ri)∏k=i′+1iQλ(x)(rk)for i′i(15) This procedure allows us to construct the full Green’s function matrix without explicitly evaluating the exponentially growing wave functions, providing a robust approach for incorporating asymptotically closed channels into iterative solvers of the LS equation. Armed with this approach, we can now elucidate the effect of closed channels on the convergence of the iterative Born series. To this end, consider a reduced-dimensional model of rotational energy transfer in He + CO collisions involving two open channels (j=0,L=0 and j=1,L=1), see Methods for details. The model describes rotational relaxation of CO molecules in collisions with 4He atoms at a collision energy of 5 cm−1 and J = 0. The left of Fig. 1B shows the convergence of the S-matrix elements for the j=1→0 collisional transition computed using the iterative approach. We observe that in the absence of closed channels, convergence to the exact CC result is reached after ≃10 iterations. Notably, adding a single closed channel (j=2,L=2) causes the Born series to diverge catastrophically. This instability is the second major limitation of iterative methods in quantum scattering theory (43). We note that this divergence commonly occurs even in the absence of closed channels (35, 36) and can be mitigated by using, e.g., optimized initial guess vectors (35), Padé approximants (37), and block-diagonal Green’s functions (36). However, to our knowledge, none of these approaches guarantees convergence to the exact solution, particularly in the strong interaction regime. Schematic of the WISE framework: Overcoming closed-channel divergence with spectral regularization. (A) A pictorial representation of molecular collisions (top) and the key equations of the WISE method: the LS equation for a single solution vector, u→, the standard Born series expansion, and the regularization procedure, where the kernel K is decomposed into a regular part KR and a divergent part KD. (B) Convergence of the squared S-matrix element ∣Sj=1,L=1;j=0,L=0∣2 for the reduced-dimensional model of He + CO scattering. The left (“Born Series”) compares the convergence of the Born series with only open channels (black circles, left axis, rapid convergence) against one with an added closed channel (red squares, right axis, logarithmic scale), where the series diverges catastrophically. The right (“Regularized Series”) demonstrates that spectral regularization restores stable convergence in the presence of closed channels. (C) Conceptual schematic of the regularization procedure and key equations. The integral operator K is analyzed spectrally; “red” components represent the subspace of divergent eigenvalues (Weinberg eigenvalues with ∣η∣≥1). These are projected out to form the regularized remainder KR (blue blocks), which has a spectral radius within the unit circle, ensuring the convergence of the iterative series. Spectral regularization and guaranteed convergence To address the second major limitation and ensure convergence of the iterative Born series, we consider the spectral radius of the LS kernel, ρ(K)=maxk∣ηk∣, where {ηk} are the eigenvalues of K (43). These eigenvalues, formally introduced in the context of nuclear scattering by Weinberg (47) and recently revisited as a diagnostic tool for nucleon-nucleon interactions (48), serve as a valuable indicator of the “perturbativeness” of the multichannel interaction potential. Specifically, the Born series converges only if ρ(K)<1; the presence of Weinberg eigenvalues lying outside the unit circle in the complex plane (∣ηk∣>1) signals the divergence of the series. We note that the number of these divergent Weinberg eigenvalues is always finite (47, 49), provided the interaction potential operator Vˆ is short ranged (true for most molecular and nuclear systems, excluding those described by Coulomb interactions). For these systems, the symmetrized scattering kernel KˆS=Vˆ1/2GˆVˆ1/2, which has the same eigenvalues as our LS kernel Kˆ, is a compact operator (49). According to the Riesz-Schauder theorem (50), the spectrum of a compact operator consists of a countable set of eigenvalues with no limit point in the complex plane except at zero. Thus, the number of divergent Weinberg eigenvalues approaches a constant in the large N limit for a fixed scattering Hamiltonian. Fixing the Hamiltonian and converging with respect to the basis size N is the standard approach in molecular quantum dynamics calculations (41), which we adopt throughout this work. Figure 2 visualizes the structure of the discretized K matrix and its spectrum for the two- and three-channel models of He + CO scattering. The matrix exhibits a distinctive asymmetric structure defined by the interplay between the Green’s function and the interaction potential in Eq. 10. Three distinct regions are visible in Fig. 2 (A and C). At very short range (column indices below 100), the matrix elements are small because the regular solution component x(r) of the Green’s function vanishes as r→0. At intermediate range (column indices ≃100 to 200), a bright vertical strip appears where the interaction potential is strongest, and the Green’s function is nonzero. Last, at long range (column indices above 200), the matrix elements decay to zero as the potential decays, despite the oscillatory nature of the open-channel Green’s function. The visible asymmetry—where columns, and not rows, dominate in the interaction region—arises because the interaction potential Uoff(r′) weighs the integration variable (r′, columns) but not the observation variable (r, rows). Weinberg eigenvalue analysis of the kernel matrix K. (A and B) Structure of the discretized kernel matrix K for the reduced-dimensional He + CO model, plotted as absolute values ∣Kαβ∣. The axes represent composite indices flattening both channel and grid dimensions. The column axis is truncated at index 500 to highlight the relevant interaction region. The vertical stripe corresponds to the short-range region where the coupling potential is substantial. (C and D) Corresponding spectra of Weinberg eigenvalues, ηk, in the complex plane. The solid black line indicates the unit circle. (A and C) Two-open-channel model. All eigenvalues (blue dots) lie strictly within the unit circle (∣ηk∣<1), ensuring convergence of the Born series. (B and D) Three-channel model (two open channels plus one closed channel). Although the spatial structure of K appears qualitatively similar to (A), the inclusion of the closed channel pushes four eigenvalues (red dots) outside the unit circle (∣ηk∣≥1), causing the standard iterative methods to diverge. In the two-channel case shown in Fig. 2C, this structure yields a spectrum entirely contained within the unit circle (∣ηk∣<1), ensuring convergence of the Born series. In the three-channel case (see Fig. 2D), the inclusion of the closed channel fundamentally alters the spectrum by pushing four Weinberg eigenvalues outside the unit circle (∣ηk∣≥1) and leading to the divergence observed in the right of Fig. 1B. We attribute this to a substantial enhancement of the magnitude of K-matrix elements at short range (Fig. 2B) caused by the inclusion of the closed channel. The lack of convergence of the Born series is thus caused by the change in spectral properties of the kernel matrix K introduced by the additional closed channel. The above analysis motivates a regularization procedure to restore convergence of the iterative Born series. This is accomplished by spectrally decomposing the kernel operator K=KR+KD into a divergent part, KD, which spans the subspace of eigenvectors with ∣ηk∣≥1, and a well-behaved regularized part, KR (43). The divergent part is defined by a separable kernel constructed from the right (v→k) and left (w→k†) eigenvectors of KKDu→(r)=∑k∈{∣ηk∣≥1}ηkv→k(r)∫ dr′w→k†(r′)u→(r′)(16) By construction, the spectral radius of the regularized kernel is less than unity, ρ(KR)<1. We can therefore reformulate the LS equation to treat the divergent part exactly while solving for the remainder with a convergent Born series. To this end, we first define the regularized source term, u→0(R), and the regularized right eigenvectors of K, v→k(R)u→0(R)(r)=(1−KR)−1u→0(r)(17)v→k(R)(r)=(1−KR)−1v→k(r)(18)where 1 is the unit matrix. These quantities are computed using the iterative series based on the regularized kernel KR. The regularized series is therefore guaranteed to converge. The full solution, u→(r), is then expressed asu→(r)=u→0(R)(r)+∑k∈{∣ηk∣≥1}ηkckv→k(R)(r)(19)where the coefficients ck represent the projection of the solution onto the divergent subspaceck=∫w→k†(r′)u→(r′)dr′(20) To determine these coefficients, we project Eq. 19 onto the set of left eigenvectors, w→i†. Crucially, as shown in Fig. 2, the number of divergent Weinberg eigenvalues outside the unit circle is much smaller than the total number of eigenvalues of K. The projection thus yields a small system of linear equations, ∑k∈{∣ηk∣≥1}Mikck=bi, wherebi=∫w→i†(r′)u→0(R)(r′)dr′,Mik=δik−ηk∫w→i†(r′)v→k(R)(r′)dr′(21) This framework guarantees convergence regardless of the interaction strength or the presence of closed channels while retaining the favorable matrix-vector operational scaling. As shown in the right in Fig. 1B, the regularization procedure tames the exponential divergence of the Born series for the three-channel model of He + CO collisions with one closed channel, bringing the results in perfect agreement with exact CC calculations. Summary of the algorithm The practical implementation of the WISE algorithm proceeds in three distinct stages, as illustrated in Fig. 3. First, we construct the reference Green’s function by solving a set of independent single-channel Schrödinger equations for all channels. Regular and irregular solutions are determined in the form of stable ratios, Qλ(x)(r) and Qλ(y)(r), by forward and backward propagations, respectively, on the radial grid ri. These ratios determine the Green’s function matrix via Eqs. 14 and 15, enabling the computation of the discretized kernel operator K. Crucially, the algorithm (see Methods) does not require the explicit construction or storage of the full matrices G, Uoff, or K. Instead, these are computed on-the-fly, defining the action of K on an arbitrary trial vector as a matrix-vector operation. Flowchart of the WISE algorithm. Stage 1 begins by solving a set of independent, single-channel Schrödinger equations to construct the reference Green’s function matrix (Eqs. 14 and 15). This setup enables the efficient computation of matrix-vector products involving the discretized kernel operator Kˆ. Stage 2 constitutes the algorithmic core: The subspace of divergent Weinberg eigenvalues is identified, the regularization procedure is invoked to determine the expansion coefficients ck, and the resulting convergent Born series is summed iteratively. The total scattering wave function is then reconstructed via Eq. 19. Last, in stage 3, the scattering matrix elements are extracted from the wave function (Eq. 11). Second, we identify the subspace of divergent Weinberg eigenvalues (∣ηk∣≥1) using the Arnoldi iterative algorithm. This requires two passes: First, we compute the nD divergent eigenvalues and their corresponding right eigenvectors, v→k(r), by iterating on K. To fully project out the divergent subspace, we also require the left eigenvectors, w→k†. Because K is complex and nonsymmetric, these are obtained by finding the right eigenvectors of the Hermitian conjugate operator, K†, associated with the conjugate eigenvalues ηk∗. We next invoke the regularization procedure by constructing and solving the reduced system of linear equations in Eq. 21. This yields the expansion coefficients ck necessary to rigorously account for the contribution of the divergent Weinberg eigenvalues, thereby ensuring a convergent iterative series. Because the number of divergent eigenvalues is typically small, this step adds negligible overhead to the overall O(MN2) scaling. The terms entering this equation—specifically the action of the regularized source term, u→0(R), and the regularized right eigenvectors of K, v→k(R)—are evaluated using the standard Born series, which is now guaranteed to converge. Notably, the evaluation of v→k(R) via Eq. 18 is computationally inexpensive: Because KR acts as a null operator on the divergent subspace (KRv→k≈0), the series truncates immediately. This property implies that the matrix M in Eq. 21 is predominantly diagonal and highly sparse, facilitating the use of efficient sparse linear solvers in cases where the dimension of the divergent subspace is large. With the coefficients ck determined, the full wave function is reconstructed via Eq. 19, and the scattering matrix elements are extracted from Eq. 11 in stage 3 (see Fig. 3), which requires the explicit form of the regular solution xλ(r) only for the open channels, consistent with the definition of the S matrix. We note that the divergent eigenvalues and eigenvectors need only to be determined once per symmetry block (i.e., for a fixed total angular momentum and spatial parity). Because the kernel K depends only on the system Hamiltonian and the reference Green’s function, its divergent subspace is independent of the incoming channel. Consequently, the iterative procedure can be efficiently applied to all asymptotically open channels within the block. For each new initial state, one simply updates the source term u→0, evaluates its corresponding expansion coefficients ck, and computes the action of the regularized kernel KR via the Born series. The Arnoldi-based implementation enjoys O(N2) scaling in the limit of large N, where the number of divergent Weinberg eigenvalues nD saturates in accordance with the Riesz-Schauder theorem (see Methods for a numerical illustration of the “Riesz-Schauder limit” for He + CO). In the preasymptotic regime, however, nD can grow linearly with N, and the cost of the Arnoldi step is then (nDN2) = (N3). To obtain (N2) scaling that does not depend on the behavior of nD (N), we recast the construction of the spectral projector onto the subspace of divergent Weinberg eigenvalues, PˆD (see Methods), as a matrix-free contour integral (51, 52). Because the contour ΓD is a fixed boundary enclosing the divergent subspace, evaluated at a fixed number of quadrature nodes with an O(N2) matrix-vector product, the computational cost is O(N2) by construction and is independent of nD regardless of whether nD saturates, grows linearly, or is a significant fraction of N. The only quantity that must remain bounded in N is the iteration count of the inner linear solves at each quadrature node; we demonstrate this property directly (see Methods). Given this bounded-iteration property, the contour-integration construction removes the nD dependence of the algorithmic complexity and yields (N2) scaling for a fixed scattering Hamiltonian. Application The WISE framework is broadly applicable to challenging problems in quantum molecular collision dynamics, cold and ultracold chemistry, and astrochemistry. To demonstrate its capabilities, we first apply the algorithm to a benchmark system known for its rich resonant structure (53–56): cold collisions of CO molecules with He atoms beyond the reduced-dimensional model described above. Accurate knowledge of low-temperature He + CO collision rates is essential for modeling the chemical properties of the interstellar medium, where CO is a primary tracer molecule (56). Figure 4A shows the total integral cross section for rotational de-excitation (j=1→0) in He + CO collisions. Our method perfectly recovers the reference CC result, including Feshbach resonances near the channel thresholds (53–56). To highlight that we are truly operating in a regime not accessible to previous iterative approaches (35, 36), we additionally present the results obtained with only open channels included. The energy dependence of the “open-channel-only” cross section is smooth and fails to reproduce the resonances, confirming that the inclusion of closed channels is crucial for an accurate description of collision dynamics at the state-to-state, energy-resolved level. Validation and application of the WISE method. (A) Integral cross sections for rotational relaxation (j=1→0) in He + CO collisions plotted as a function of collision energy. The standard CC reference solution (solid blue line) is perfectly reproduced by the WISE method (blue circles). Note that the WISE calculations were performed on a dense energy grid; only a subset of points is shown for clarity. To highlight the physical importance of closed channels, the gray curve shows the results of CC calculations including only open channels, which do not reproduce the rich Feshbach resonance structure that arises directly from closed-channel couplings. (B) Integral cross sections for rotational de-excitation of CO (jCO=7→6) in collisions with N2 (jN2=6). This system represents a computationally demanding regime relevant to atmospheric modeling, demonstrating the method’s stability for anisotropic molecule-molecule interactions with high channel densities over a broad energy range. Having validated the WISE approach for rotationally inelastic He + CO scattering in the resonant regime, we now demonstrate its broader applicability by addressing a much more complex system: collisions of CO with molecular nitrogen (N2). The CO-N2 interaction is much more anisotropic than that of He-CO (57), serving as a stringent test of our algorithm’s stability in systems with high densities of states. Beyond its theoretical complexity, the CO-N2 system is of paramount importance to Earth science. Carbon monoxide is a primary atmospheric pollutant that serves as a crucial tracer of anthropogenic activity, specifically fossil fuel combustion (58–60). Furthermore, while CO is not itself a direct greenhouse gas, it plays a critical indirect role in the climate system: By reacting with hydroxyl radicals, CO depletes the primary sink of methane, thereby increasing the atmospheric lifetime of this potent greenhouse gas (61, 62). Consequently, global monitoring of CO via satellite remote sensing is essential. To accurately retrieve CO column densities from spectral observations, one must account for perturbations caused by collisions with dominant atmospheric constituents: N2 and O2. Specifically, these collisions lead to pressure broadening and shift of the CO spectral lines (63). Modeling these effects from first principles allows for minimizing systematic errors in retrieval algorithms (64). The gold standard for such modeling is CC calculations based on high-quality ab initio PESs (11). However, the combination of small rotational constants of CO and N2, strong potential anisotropy, and the large number of partial waves required at thermal energies makes fully converged CC calculations prohibitively expensive. Consequently, first-principles studies of N2- and O2-perturbed spectra of CO have been limited to only the lowest rotational transitions (13, 65, 66). To explore the robustness of the WISE algorithm when scaled to these computationally demanding regimes, we compute inelastic cross sections for the rotational de-excitation of CO (j=7→6) in collisions with N2 in its most populated rotational state (jN2=6). We specifically select this transition because the j = 7 level corresponds to the peak population of CO at terrestrial temperatures (∼300 K), making it physically representative of the dominant collision dynamics in the atmosphere. This setup creates a scattering problem of substantially greater complexity than He + CO collisions: The number of channels increases by an order of magnitude. Given the standard cubic scaling, this corresponds to a 1000-fold increase in computational cost. Figure 4B shows that our WISE algorithm successfully replicates exact CC results for the de-excitation cross section across a broad range of collision energies (1 to 1000 cm−1). The agreement is excellent over the entire energy range. Scaling A critical advantage of the WISE method is its favorable scaling with the number of scattering channels, N. Upon discretizing the radial coordinate onto a grid, the operator K becomes a nonsymmetric matrix of dimension (N⋅Ng)×(N⋅Ng). While finding its full spectrum would be computationally prohibitive, our regularization scheme requires only the small subset of divergent eigenvalues with ∣ηk∣≥1. These can be found efficiently using sparse iterative eigensolvers, such as the Arnoldi algorithm (67). To demonstrate the scaling of our algorithm in a realistic setting, we performed a series of benchmarks using the CO-N2 system described above. We fixed the collision energy to 10 cm−1 and systematically increased the basis set size, generating a set of coupled equations ranging from N = 12 to N = 497. The resulting performance is presented in Fig. 5. The total central processing unit (CPU) time shown in Fig. 5A scales quadratically with the number of channels, breaking the cubic bottleneck of the standard approaches based on propagating the scattering wave function. This cost is dominated by the iterative search for the divergent eigenvalues and the corresponding right and left eigenvectors (Fig. 5B). Computational scaling of the WISE algorithm. The wall-clock time (in arbitrary units) is plotted as a function of the total number of scattering channels, N. Red dashed lines represent least-squares fits to the data: quadratic (∝N2) for (A) and (B) and linear (∝N) for (C) and (D). (A) Total time to solution, showing the overall scaling performance. (B) Time required to identify the subspace of divergent Weinberg eigenvalues with ∣ηk∣≥1 using Arnoldi iterations. (C) Time required to converge the Born series for the regularized source vector u→0(R). (D) Time required to converge the Born series for the regularized eigenvectors v→k(R). We now consider the computational scaling of the regularization steps. Figure 5C shows the time required to converge the regularized source term, u→0(R)(r), which typically requires a sequence of multiple iterations. In contrast, Fig. 5D shows the cumulative time to compute the regularized right eigenvectors of K, v→k(R)(r), for the entire divergent subspace. Despite involving multiple vectors, this step is extremely fast. Because KR acts effectively as a null operator on these eigenvectors (KRv→k≈0), the Born series truncates almost immediately. Consequently, the scaling observed in this panel is driven not by Born iterations but rather by the increase in the number of divergent eigenvalues nD with the number of channels N. Last, the solution of the sparse linear system for the coefficients ck makes a negligible contribution to the total computational time, even with a standard direct solver used in the present implementation (see Methods). The cost of this step depends only on the number of divergent Weinberg eigenvalues, which can become large (nD≃2×105) for systems involving an extremely large number of channels (N≃105), as estimated below. For these systems, the sparse structure of M (Eq. 21) would allow for a further reduction in computational cost using sparse solvers. DISCUSSION Quantum CC calculations are an essential tool widely used in atomic, molecular, chemical, and nuclear physics to elucidate complex scattering phenomena. They are also instrumental in atmospheric chemistry and astrochemistry (8, 68), underlying quantitative modeling of spectral line shapes (11, 12, 15) and astrochemical reaction networks (8, 9). All previous algorithms for solving CC equations scale cubically with the number of scattering channels, making these calculations extremely computationally intensive and motivating the development of numerous approximate techniques, ranging from the coupled-states approximation (69) to mixed quantum-classical methods (34, 70). The accuracy of these approximate techniques is often difficult to estimate. While the need for a practical low-scaling iterative algorithm for solving CC equations has been recognized for some time (35–37, 71), progress has been hindered by the lack of reliable treatment of closed channels and notorious divergence issues with iterative expansions (35, 36). Here, we overcome these long-standing barriers by developing a practical low-scaling algorithm for solving CC equations for a single column of the scattering matrix. A key part of the WISE algorithm is the regularization procedure in which the scattering kernel K is split into the regular and divergent parts using the computed spectrum of Weinberg eigenvalues. The regular part is treated perturbatively via Born iterations, whereas the divergent part is explicitly solved using sparse matrix inversion. These features ensure numerical convergence of scattering observables and enable the WISE algorithm to readily handle closed channels, making it a promising tool for now intractable scattering problems in molecular physics and chemical reactions dynamics. We have applied the WISE algorithm to atom-molecule (He + CO) and molecule-molecule (CO + N2) collisions of relevance to astrochemistry and atmospheric chemistry, demonstrating quadratic scaling (Fig. 5) and good agreement with benchmark CC calculations (Fig. 4). While these problems involve hundreds of coupled channels, the current limit for conventional algorithms stands at a few tens of thousands of channels (42). To illustrate the capability of the WISE approach to go beyond this limit, consider, e.g., cold H2O + H2O collisions at interstellar temperatures (≤100 K), which are now intractable at the exact CC level even in the rigid-rotor approximation. Because of the high anisotropy of the water dimer PES (72, 73), these calculations could easily involve hundreds of thousands of scattering channels. Using N=105 and 100 optimized radial quadrature points per channel leads to a 10 M × 10 M K-matrix, which can be realistically solved using modern iterative (e.g., Arnoldi) solvers, given the sparse structure of K. The number of Weinberg eigenvalues outside of the unit circle can be estimated as nD≤105 based on our preliminary CO-N2 calculations. A single WISE calculation would therefore require the inversion of a complex nD×nD M-matrix. This matrix needs to be inverted only once, as opposed to hundreds of times in conventional CC calculations (41). Note that the M-matrix is sparse, facilitating the use of highly efficient iterative solvers, such as the generalized minimal residual (GMRES) algorithm (46). Using these techniques to extend the WISE algorithm to N≃105 is now in progress. Last, we note that Weinberg eigenvalues can be used as a valuable diagnostic tool for complex multichannel scattering dynamics, as already demonstrated for nucleon-nucleon interactions in nuclear physics (48). In particular, Weinberg eigenvalue spectra provide insight into the perturbativeness of an interaction PES, i.e., the extent to which scattering observables for the PES are well described by a convergent perturbative expansion. The perturbativeness of a given interaction potential is strongly related to the amount of computational resources required for convergence of scattering observables (48). Here, we analyze Weinberg eigenvalues in the context of multichannel molecular scattering. While the physical interpretation of diverging Weinberg eigenvalues merits further study, our calculations show that they can serve as valuable indicators of strong-coupling phenomena, as their number increases notably in the presence of closed channels and highly anisotropic intermolecular interactions. While the WISE framework fundamentally improves the algorithmic complexity of the time-independent quantum scattering problem, its current practical implementation presents several specific challenges. When compared to standard CC propagators, which benefited from decades of low-level optimization, the integral equation approach introduces a different demand on computational resources. Because the WISE approach evaluates the Lippman-Schwinger equation over the entire spatial domain rather than propagating the differential solution sector-by-sector, it intrinsically requires large memory storage. Constructing the Green’s function on the fly necessitates storing the ratios of the regular and irregular solutions at every point on the spatial grid. Furthermore, the Arnoldi spectral regularization procedure requires keeping nD left and right eigenvectors of K (each of size N×Ng) in active memory. We found that using a sufficiently high density of radial grid points in the short-range region provides multiple advantages by (i) ensuring an accurate representation of the interaction PES; (ii) accurately resolving the oscillatory behavior of the short-range Green’s function; and (iii) reducing the number of diverging Weinberg eigenvalues, nD. To fully exploit these advantages without inflating the overall memory cost, future implementations will benefit from the use of optimized spatial grids, such as Gauss-Lobatto quadrature (46) and finite-element discrete-variable representations (74) to minimize the total number of grid points Ng. Last, isolating the divergent subspace introduces its own challenges. While the fundamental guarantee of the O(N2) asymptotic scaling relies on the Riesz-Schauder theorem for compact operators (50), which dictates that nD must eventually saturate, there exists an intermediate range of channels where nD can grow linearly with N. In this specific intermediate regime, standard Arnoldi iterative solvers exhibit unfavorable cubic scaling, which can be overcome using matrix-free spectral projection operators constructed via contour interaction (see Methods). METHODS PESs and Hamiltonians For the CO-He system, we use the ab initio PES developed by Peterson and McBane (75) at the coupled cluster singles and doubles with perturbative triples [CCSD(T)] level of theory. We use the effective two-dimensional (2D) potential calculated for the ground vibrational state (v = 0) of CO, denoted as Vv=0(r,θ), where r is the distance between the CO center of mass and the He atom, and θ is the Jacobi angle. The angular dependence of the CO-He PES is expanded in a basis of 20 Legendre polynomials. The internal Hamiltonian Hˆint describes the CO molecule as a rigid rotorHˆCO=Bejˆ2−Dejˆ4(22)with the rotational constant Be=1.92251 cm−1 and the centrifugal distortion constant De=6.1193×10−6 cm−1 (76). For the CO-N2 system, we use the 4D PES computed by Cybulski et al. (57) using the CCSD(T) method with an aug-cc-pVQZ basis set augmented with midbond functions. The intramolecular bond lengths were frozen at their vibrationally averaged ground-state values (rCO=2.13201,a0 and rN2=2.07397 a0). The full interaction potential is expanded in bispherical harmonics as detailed in (13). The internal Hamiltonian is the sum of two rigid rotor Hamiltonians, Hˆint=HˆCO+HˆN2, with the rotational constants of CO and N2 taken from (76). Basis sets and channel definitions For the CO-He system, calculations are performed in the space-fixed basis set defined by the coupling of the rotational angular momentum jˆ with the end-over-end orbital angular momentum Lˆ to form the total angular momentum Jˆ. The internal index used in Eq. 2 is γ≡{j}. For the reduced-dimensional model (Fig. 4), calculations were restricted to the J = 0 block at a collision energy of Ekin=5 cm−1. The two-channel model includes only the open channels (j=0,L=0 and j=1,L=1). The three-channel model adds the lowest closed channel (j=2,L=2). The fully converged CO-He calculations presented in Fig. 4 used an extended basis set including all rotational states up to jmax=6 and total angular momenta up to Jmax=9, in addition to all the allowed L values. To explicitly verify that the WISE algorithm converges with respect to the basis size exactly as standard CC calculations do, we track values of the scattering matrix as closed channels are incrementally added. The top of Fig. 6 shows the results for the j=1→0 de-excitation transition in He + CO collisions at a collision energy of 5 cm−1. As a representative example, we examine the J = 0 block of positive parity p=(−1)j+L=+1. Under these conditions, only two channels are asymptotically open: j=0,L=0 (denoted as channel 0) and j=1,L=1 (denoted as channel 1). We plot the squared moduli of the elastic (∣S11∣2) and inelastic (∣S01∣2) matrix elements as a function of the total number of channels N kept in the basis. The sequence begins at N = 2 (open channels only) and adds closed channels sequentially up to j=19,L=19. We observe that the WISE results perfectly trace the benchmark CC values throughout the entire convergence pattern. The iterative scheme exhibits no instabilities or loss of accuracy even in the presence of a large number of closed channels: At N = 20, where the ratio of closed to open channels is 9:1, the WISE and CC results are exactly the same. Convergence of the WISE algorithm with respect to the number of scattering channels. (Top) Convergence of the transition probabilities (∣Sij∣2) as a function of the number of channels, N, for He + CO scattering at a collision energy of 5 cm−1 (J=0,p=1). The plot displays both the elastic (∣S11∣2) and inelastic (∣S01∣2) transitions between the two asymptotically open channels: j=0,L=0 (channel 0) and j=1,L=1 (channel 1). Solid lines indicate the benchmark values obtained using standard CC calculations, while the symbols represent the results computed using the WISE algorithm. (Bottom) Number of diverging Weinberg eigenvalues, nD, as a function of the total number of channels for the same system. The dashed line highlights the regime in which nD scales linearly with N. In the case of the CO-N2 system, the basis is formed by coupling the rotational angular momenta of CO (jCO) and N2 (jN2) to a resultant jAB, which is then coupled with L to form J. The internal index γ≡{jCO,jN2,jAB}. The cross sections for the jCO=7→6 transition in collisions with N2(jN2=6) were computed for J = 0 with the basis set that included jCO = 0 to 14 and jN2=4,6, and 8 (note that coupling between even and odd rotational levels of a homonuclear molecule via the CO-N2 PES is forbidden due to nuclear spin symmetry). This results in a system of 248 (even parity) and 203 (odd parity) CC equations, ensuring convergence of the de-excitation cross section to within 20%. To analyze the computational scaling, we performed a series of CO-N2 calculations at a fixed collision energy of 10 cm−1. We defined a minimal basis set containing only the initial and final states (jCO∈6,7,jN2=6) and systematically expanded the number of channels by adding the rotational states of both monomers (expanding to jCO<6,jCO>7 and jN2≠6). This procedure generates a set of scattering problems with total channel counts ranging from 12 to 497. Numerical implementation of the iterative algorithm The radial grid and basis sets used in the WISE calculations are identical to those used in the benchmark CC production runs. For He + CO collisions, the kernel matrix (Eq. 10) is discretized on a radial grid spanning r∈[3.4,20.0] a0 with Ng=1661 points (Δr=0.01 a0). For CO + N2 collisions, we use a grid of Ng=1581 points ranging from 4.2 to 20.0 a0 (Δr=0.01 a0), which ensures strict convergence (<1%) of benchmark CC calculations (40). For cold and ultracold collisions, the outer range of radial integration may have to be extended to ≃100 a0 or longer, which can be readily handled by specialized Gauss-Lobatto or finite-element methods. The quadrature weights wi′ entering the definition of the discretized K matrix in Eq. 10 follow the composite trapezoidal rule: wi=Δr for internal points i∈{2,Ng−1} and wi=Δr/2 for the endpoints i∈{1,Ng}. A critical feature of the WISE algorithm is that the full kernel matrix K is never explicitly constructed or stored in memory, which would otherwise impose a prohibitive O(N2Ng2) memory cost. Instead, we treat K as a matrix-free linear operator, defined solely by its action on a trial vector. In our implementation, this is achieved using the LinearOperator abstraction in SciPy, which interfaces directly with sparse eigensolvers. The subset of eigenvalues required for regularization is computed using the Implicitly Restarted Arnoldi Method, as implemented in the Arnoldi Package (ARPACK) library (67). This allows us to selectively converge only the divergent eigenvalues with magnitudes ∣η∣≥1, avoiding the cost of a full spectral decomposition. Once the divergent subspace is identified, the regularization coefficients ck are determined by solving a system of linear Eq. 21 of dimension nD×nD, where nD is the number of divergent Weinberg eigenvalues. Since the system is sparse, and the set of equations has to be solved only once, the total computational cost introduced by this stage is negligible. The set of equations is solved using the standard LU (Lower-Upper) decomposition driver for general complex matrices (ZGESV) from the LAPACK (Linear Algebra Package) library, accessed via NumPy’s linear algebra wrapper. The Born series for the regularized remainder is evaluated via fixed-point iteration. At each step n, the solution is updated as u→(n)=u→0+KRu→(n−1), and the single column of the S-matrix is computed according to Eq. 11. Convergence is monitored by explicitly tracking the maximum absolute change in the scattering matrix elements between consecutive iterations, ϵ=max∣Sij(n)−Sij(n−1)∣. The iterations are terminated when ϵ falls below a specified tolerance threshold, set to 10−6 in the present work. Because the spectral radius of the regularized operator ρ(KR) is strictly less than 1, this iterative sequence is formally guaranteed to converge. However, the practical speed of convergence is governed by the largest remaining Weinberg eigenvalues. Eigenvalues with magnitudes strictly less than, but close to, unity can slow down the convergence of the Born series. This bottleneck can be easily circumvented by extending the regularized subspace to project out not only the strictly divergent eigenvalues (∣η∣>1) but also those that merely impede convergence. In the present calculations leading to Figs. 4 and 5, we optimize this by setting the exclusion threshold to ∣η∣≥0.9. The efficiency of this optimized regularization scheme is illustrated in Fig. 7, which presents the number of iterations M required to converge the single column of the S-matrix as a function of the number of channels, N for the N2 + CO calculations used to demonstrate scaling of the WISE algorithm on Fig. 5. While there is no strictly monotonic pattern (M is scattered between 125 and 225, depending on the specific spectrum of K for a given set of channels), it is clear that M does not grow continuously with N. For a small number of channels, the number of iterations can exceed the basis size (M>N), making the iterative approach less practical for simple systems. Crucially, however, as the basis set size increases, the required number of iterations saturates. For the largest number of channels considered here (N = 497), convergence is achieved with M≈218 steps, clearly demonstrating that in the asymptotic limit of large N, the iterative part of the algorithm operates within the M