betaFeedback

WorksPhysics and Astronomy

Efficient simulation of second-order phase transitions in quantum anharmonic materials

Abstract

When a crystal undergoes a second-order structural phase transition, such as in ferroelectrics, Peierls, and charge-density waves, the diverging fluctuations of the order parameter lead to the break- down of the standard phonon quasiparticle picture. Simulating these highly anharmonic regimes is notoriously challenging, as methods such as molecular dynamics suffer from a critical slowdown near the transition point, while the harmonic approximation fails dramatically at saddle points of the energy landscape stabilized by quantum or thermal ionic fluctuations. This work introduces a new approach, based on the variational free-energy principle, to predict critical long-range behavior and dynamical spectra in strongly anharmonic systems, even when quantum ionic fluctuations dominate. The proposed framework builds upon the stochastic self- consistent harmonic approximation but reduces its computational scaling with the number of atoms, N, from O(N^6) to O(N^2) and the memory requirement from O(N^4) to O(N). We benchmark the method on the prototypical lead-free metal-halide perovskite CsSnI3, a promising candidate for photovoltaic engineering, simulating its phase stability and Raman spectrum near the phase transition, where the breakdown of the quasiparticle picture becomes evident. We demonstrate the effectiveness of the method by computing the full free-energy Hessian and the critical temperature in a supercell with 1080 atoms. Such simulations would have required tens of thousands of years with the legacy approach; it is now feasible in a few hours on consumer hardware.

Introduction

Second-order structural phase transitions play a fundamental role in statistical physics and materials science, thanks to the peculiar properties of the critical transition point: a thermodynamic state where the excitation energy of the order parameter vanishes. At this critical point, fluctuations in the order parameter grow with system size, and the system’s response to perturbations involving the order parameter diverges. This behavior is desirable for many technological applications as it results in a macroscopic response to small fields[1, 2]. Moreover, the large fluctuations in the order parameter interact with the rest of the system, introducing strong phonon-phonon scattering. This is exploited in thermal transport, where critical fluctuations slow down heat propagation, resulting in a drop in thermal conductivity while retaining almost unchanged electronic properties. Prototypical examples include group IV–VI alloys such as SnTe, GeTe[3, 4], SnSe[5], and SnS[6], some of which are among the most efficient bulk thermoelectric materials known today. In several of these materials, the criticality is hidden by quantum fluctuations of nuclei, as in the case of SrTiO3_3[7, 8], KTaO3_3[9], and PbTe[10, 11], which serve as interesting playgrounds for nonlinear phononics[12, 13, 14, 15, 16], where a ferroelectric state can be persistently induced on the picosecond timescale by applying a sufficiently strong mid-IR/THz pulse.

Simulating critical points is extremely challenging for two reasons: (i) the correlation length of the order parameters decays polynomially, meaning that extremely large simulation cells are required to properly account for this divergence, and (ii) the zero excitation energy of the order parameter corresponds to a divergent time period of its oscillation, which results in the well-known critical slowdown of molecular dynamics simulations[17]. To overcome these issues, several mean-field approaches have been introduced[18, 19, 20, 21, 22]. Among these, the self-consistent phonons[23] and the stochastic self-consistent harmonic approximation (SSCHA)[24, 25] have encountered great success, as they also account for ionic quantum effects and do not rely on molecular dynamics (or path-integral molecular dynamics) trajectories affected by the aforementioned issues. The key idea behind the SSCHA is to approximate the quantum ionic density matrix as the most general Gaussian (which includes interatomic covariances), and optimize its average position and covariance matrix to minimize the free energy.

In particular, the SSCHA can exploit crystal symmetries to compute the thermodynamic conditions at the critical point using the elegant Landau theory of phase transitions: we define a positional free energy F\mathcal{F} as a function of the order parameter Δ\Delta, and assess the sign of the F\mathcal{F} second derivative around Δ=0\Delta= 0[26]. This approach has been applied successfully to systems with strong quantum and thermal anharmonic fluctuations, including high-pressure hydrogen[27, 28, 29], hydrides and hydrates[30, 31, 32, 33], thermoelectric materials[5, 6, 34], transition metal oxides[7, 35] bending rigidity in graphene[36], and it has unveiled the ionic entropic origin of charge-density wave melting in transition-metal dichalcogenides[37, 38, 39].

However, the calculation of the positional curvature of the free energy, i.e., the free energy Hessian, requires the knowledge of 3-phonon and 4-phonon scattering tensors, which are expensive to compute and memory-heavy to store without approximated compression techniques[40, 41, 42, 43, 44, 45]. As a result, the nonperturbative computational of the full free energy Hessian within the SSCHA scales as O(N6)O(N^{6}), with NN the number of atoms in the simulation cell, limiting its application to supercells with less than 50 atoms, too small to obtain converged results. Phase transitions have therefore been mostly investigated while ignoring 4-phonon scattering, using the so-called “bubble” approximation[26, 46], which is considered state-of-the-art for nonperturbative phonon calculations.

In this work, we introduce two algorithmic innovations that drastically reduce the computational cost of inverting the high-rank anharmonic scattering tensors. The first is a reformulation of the free-energy Hessian problem as the inversion of a static Liouvillian operator via iterative solvers, completely bypassing the explicit construction of the fourth-order force-constant tensor. The second is a reformulation of the entire SSCHA linear response in momentum space. For the full Hessian, the resulting computational scaling is reduced from O(N6)→O(N2)O(N^{6}) \to O(N^{2}), while the memory requirement of the algorithm drops from O(N4)→O(N)O(N^{4}) \to O(N), where NN indicates the number of atoms in the supercell. Notably, for a 6×6×66 \times6 \times6 supercell, we achieve a speedup on the order of 10910^{9}, enabling simulations that were simply impossible before. This dramatic speedup unlocks full mode mixing in free-energy Hessians and phonon spectral functions, bypassing the traditional bubble approximation at a fraction of the legacy computational cost.

In Sec. II we review the theoretical context of the free energy Hessian, its link to the Landau theory of phase transitions, and discuss the origin of the computational bottleneck in state-of-the-art implementations. Then, Sec. III and III B present the details of the new algorithm introduced in this work. The method’s scalability with system size in both time and memory is benchmarked on a simple gold crystal (FCC) and compared with the previous state of the art (§ 5 A). Then, we apply the method to a real-case material, CsSnI3_3, computing the phase diagram and showing that it correctly identifies the phase transition.

The same formalism also applies to the evaluation of phonon spectral functions within the dynamical extension of the static SSCHA, the Time-Dependent Self-Consistent Harmonic Approximation (TD-SCHA)[47, 48]. Therefore, we characterise the breakdown of the phonon quasiparticle picture in the simulated Raman spectrum of CsSnI3_3, with a supercell exceeding 1000 atoms (§ 5 B).

The Free Energy Hessian

The positional free-energy Hessian is the crucial quantity in the Landau theory of phase transitions for identifying a critical point[26]. The positional free energy F(Δ)\mathcal{F}(\Delta) is a function of the order parameter whose curvature goes to zero at the critical temperature. A displacive phase transition is such that the order parameter can be expressed as a linear combination of the average atomic position: Δ=∑aeΔa⟨ra−ra(0)⟩\Delta= \sum_a e_{\Delta}^{a}\langle r_a-r_a^{(0)}\rangle, where rr is the vector of the Cartesian coordinates of each atom in the cell (the index aa run over both xx, yy, zz and atomic indices), r(0)r^{(0)} indicates the average positions of atoms in the high-symmetry (high temperature) phase, and eΔe_{\Delta} is the atomic displacement pattern in the broken symmetry phase (bold symbols indicate vectors and matrices). Thus, Δ=0\Delta=0 for T>TcT>T_c, while the lattice centroids deform following eΔe_{\Delta} below the phase transition.

The Landau free energy F(Δ)\mathcal{F}(\Delta) can be rigorously defined in quantum mechanics[46] by constraining Δ\Delta on the equilibrium density matrix, while minimizing the global free energy functional under the Gibbs variational principle:

F(Δ)=min⁡ρ^{F[ρ^]+λ(⟨∑aeΔa(ra−ra(0))⟩ρ−Δ)},(1)\mathcal{F}(\Delta)=\min_{\hat{\rho}}\left\{F[\hat{\rho}]+\lambda\left(\left\langle\sum_a e_{\Delta}^{a}(r_a-r_a^{(0)})\right\rangle_{\rho}-\Delta\right)\right\}, \tag*{(1)}

and λ\lambda is chosen to impose the Δ\Delta constrain by making F(Δ)\mathcal{F}(\Delta) stationary. Here, ρ^\hat{\rho} is the trial density matrix and F[ρ^]F[\hat{\rho}] is the free energy functional defined

F[ρ^]=⟨H^⟩ρ−TS[ρ^],(2)F[\hat{\rho}]=\langle\hat{H}\rangle_{\rho}-TS[\hat{\rho}], \tag*{(2)}
⟨H^⟩ρ=Tr⁡[ρ^H^],S[ρ^]=−kb⟨ln⁡ρ^⟩ρ,\langle\hat{H}\rangle_{\rho}=\operatorname{Tr}\left[\hat{\rho}\hat{H}\right],\qquad S[\hat{\rho}]=-k_b\langle\ln\hat{\rho}\rangle_{\rho},

and H^\hat{H} is the Born-Oppenheimer nuclear Hamiltonian. In practice, if the specific symmetry-breaking pattern eΔe_{\Delta} is unknown, we can define the positional free energy on all centroids r~i\tilde{r}_i as

F(r~1,⋯ ,r~n)=min⁡ρ^[F[ρ^]+∑iλi(⟨ri⟩ρ−r~i)].(3)\mathcal{F}(\tilde{r}_1,\cdots,\tilde{r}_n)=\min_{\hat{\rho}}\left[F[\hat{\rho}]+\sum_i\lambda_i\left(\langle r_i\rangle_{\rho}-\tilde{r}_i\right)\right]. \tag*{(3)}

Here λ1,…,λn\lambda_1,\ldots,\lambda_n respect the stationarity condition. Then, a critical point occurs when F(r~1,⋯ ,r~n)\mathcal{F}(\tilde{r}_1,\cdots,\tilde{r}_n) is at a saddle point, i.e., the curvature of the free energy is flat along at least one direction, and the resulting symmetry-breaking displacement is related to the associated polarization vector. At the critical point, this condition is reached at the high-symmetry configuration, r~=r(0)\tilde{r}=r^{(0)}.

For the sake of simplicity, from now on we use mass-rescaled positions, defined as

Ra=mara(4)R_a=\sqrt{m_a}r_a \tag*{(4)}

(see Appendix A for details on how any quantity is mass-rescaled).

Eq. (3) is exact in principle but impractical to solve. The stochastic self-consistent harmonic approximation (SSCHA) restricts the variational landscape to Gaussian density matrices, allowing the explicit minimization of Eq. (3)[24, 25]. The SSCHA name stands for the equivalence between the Gaussian equilibrium density matrix ρ^\hat{\rho} and the auxiliary harmonic Hamiltonian that has ρ^\hat{\rho} as the equilibrium solution. Thus, the SSCHA density matrix is parametrized by the average positions of all atoms R\mathcal{R} and the dynamical matrix D\mathbf{D} of the auxiliary Hamiltonian, which are optimized to minimize the free energy. At the minimum of the free energy, they satisfy the following self-consistent relations[24, 25]:

Dab=⟨d2VdRadRb⟩ρ,Ra=⟨Ra⟩ρ.(5)D_{ab}=\left\langle\frac{d^{2}V}{dR_{a}dR_{b}}\right\rangle_{\rho},\qquad\mathcal{R}_{a}=\left\langle R_{a}\right\rangle_{\rho}. \tag*{(5)}

The auxiliary SCHA dynamical matrix allows us to define the free phonon propagation, with poles at the dressed harmonic frequencies ωμ\omega_{\mu} that diagonalize D\mathbf{D}:

∑beμbDab=ωμ2eμa.(6)\sum_{b}e_{\mu}^{b}D_{ab}=\omega_{\mu}^{2}e_{\mu}^{a}. \tag*{(6)}

Bianco et al.[26] derived an elegant equation using SSCHA perturbation theory to compute the positional free energy Hessian:

d2FdRdR=D+D(3)Λ[1−D(4)Λ]−1D(3),(7)\frac{d^{2}F}{d\mathcal{R}d\mathcal{R}}=\mathbf{D}+\overset{(3)}{\mathbf{D}}\boldsymbol{\Lambda}\left[\mathbf{1}-\overset{(4)}{\mathbf{D}}\boldsymbol{\Lambda}\right]^{-1}\overset{(3)}{\mathbf{D}}, \tag*{(7)}

where D(3)\overset{(3)}{\mathbf{D}} and D(4)\overset{(4)}{\mathbf{D}} are the 3, and 4-rank mass-rescaled force constants averaged over the SSCHA Gaussian density

D(3)abc=⟨d3(V−V)dRadRbdRc⟩ρ,(8)\overset{(3)}{D}_{abc}=\left\langle\frac{d^{3}(V-\mathcal{V})}{dR_{a}dR_{b}dR_{c}}\right\rangle_{\rho}, \tag*{(8)}
D(4)abcd=⟨d4(V−V)dRadRbdRcdRd⟩ρ,(9)\overset{(4)}{D}_{abcd}=\left\langle\frac{d^{4}(V-\mathcal{V})}{dR_{a}dR_{b}dR_{c}dR_{d}}\right\rangle_{\rho}, \tag*{(9)}

where V(R)\mathcal{V}(\mathbf{R}) is the Harmonic potential of the auxiliary hamiltonian:

V(R)=12∑ab(Ra−Ra)Dab(Rb−Rb).\mathcal{V}(\mathbf{R})=\frac{1}{2}\sum_{ab}(R_{a}-\mathcal{R}_{a})D_{ab}(R_{b}-\mathcal{R}_{b}).

Λ\boldsymbol{\Lambda} is the static limit of the two-phonon free propagator:

Λabcd=−∑μνeμaeμbeνceνd4ωμων{1+nμ+nνωμ+ων−nμ−nνωμ−ωνfor ωμ≠ων,1+2nμ2ωμ−dndω∣ω=ωμfor ωμ=ων.(10)\Lambda_{abcd}=-\sum_{\mu\nu}\frac{e_{\mu}^{a}e_{\mu}^{b}e_{\nu}^{c}e_{\nu}^{d}}{4\omega_{\mu}\omega_{\nu}}\left\{\begin{array}{ll} \dfrac{1+n_{\mu}+n_{\nu}}{\omega_{\mu}+\omega_{\nu}}-\dfrac{n_{\mu}-n_{\nu}}{\omega_{\mu}-\omega_{\nu}} & \text{for }\omega_{\mu}\ne\omega_{\nu},\\ \dfrac{1+2n_{\mu}}{2\omega_{\mu}}-\left.\dfrac{dn}{d\omega}\right|_{\omega=\omega_{\mu}} & \text{for }\omega_{\mu}=\omega_{\nu}. \end{array}\right. \tag*{(10)}

Here, nμ=1/(eωμ/kBT−1)n_{\mu}=1/(e^{\omega_{\mu}/k_{\mathrm{B}}T}-1) is the Bose-Einstein distribution function, and we assume Hartree atomic units where ℏ=1\hbar=1.

The inversion in Eq. (7) is equivalent to an RPA-like resummation of the ladder diagrams reported in Figure 1, where D(3)\overset{(3)}{\mathbf{D}} and D(4)\overset{(4)}{\mathbf{D}} appear as three-phonon and four-phonon scattering vertices, while D\mathbf{D} is the inverse non-interacting one-phonon propagator.

Diagrammatic representation of the free energy Hessian as an RPA-like resummation

Figure 1. Diagrammatic representation of the free energy Hessian as an RPA-like resummation. Wavy lines are free SCHA phonon propagators: the static one-phonon free propagator is the inverse of the dynamical matrix D\mathbf{D}, while the two-phonon propagator is the Λ\boldsymbol{\Lambda} tensor (Eq. (10)). Triangles (▹\triangleright) denote three-phonon vertices D(3)\overset{(3)}{\mathbf{D}}: each triangle has three corners where phonon lines attach (corresponding to the indices to which it can be contracted). Squares (□\square) denote four-phonon vertices D(4)\overset{(4)}{\mathbf{D}}: each square has four corners connecting to the two legs of the two-phonons propagator Λ\boldsymbol{\Lambda} on either side. These are the lowest-order diagrams of the geometric expansion of the inversion in Eq. (7).

Truncating the series at the first perturbative term (i.e., neglecting D(4)\overset{(4)}{\mathbf{D}}) corresponds to the widely employed bubble approximation, where the name takes inspiration from the bubble form of the only surviving self-energy diagram

d2FdRdR≈bubbleD+D(3)ΛD(3).(11)\frac{d^{2}F}{d\mathcal{R}d\mathcal{R}}\mathrel{\mathop{\approx}^{\mathrm{bubble}}}\mathbf{D}+\overset{(3)}{\mathbf{D}}\boldsymbol{\Lambda}\overset{(3)}{\mathbf{D}}. \tag*{(11)}

While elegant, implementing Eq. (7) as it stands presents two critical problems: (i) the inversion of a dense N2×N2N^{2}\times N^{2} matrix, and (ii) the storage of the dense high-rank tensor D(4)\overset{(4)}{\mathbf{D}} (Eq. (9)). For this reason, the software that implements the free energy Hessian often relies on approximations, like series truncation to lowest order (Eq. (11)) or employing compressed sensing techniques for the storage of high-rank force constant tensors[43, 49, 50].

Here, we devise a different approach that completely avoids materializing either D(3)\overset{(3)}{\mathbf{D}} or D(4)\overset{(4)}{\mathbf{D}} while maintaining the exact equivalence to the full Eq. (7). The key insight is to reformulate Eq. (7) as a linear system, and solve it iteratively by applying only sparse matrix-vector products.

Positional free energy Hessian as a sparse linear system

Let us define a vector space of possible perturbations of the equilibrium Gaussian density matrix to static external perturbations as

∣ψ⟩=(R(1)γ(1)).(12)|\psi\rangle=\begin{pmatrix}\mathcal{R}^{(1)}\\ \gamma^{(1)}\end{pmatrix}. \tag*{(12)}

where the space is composed in two sectors: R(1)\mathcal{R}^{(1)} and Y(1)\mathcal{Y}^{(1)}. R(1)\mathcal{R}^{(1)} has dimension 3N3N and represents the perturbation on the average positions of atoms (NN atoms and 3 Cartesian directions). Y(1)\mathcal{Y}^{(1)} has dimension 3N(3N+1)/23N(3N+1)/2, and represents the perturbation of the Gaussian covariance matrix (thus a symmetric (3N)×(3N)(3N)\times(3N) matrix). The positional free energy Hessian can be computed by solving a linear system in this vector space as

L∣ϕk⟩=∣δk⟩(13)\mathcal{L}|\phi_k\rangle=|\delta_k\rangle \tag*{(13)}

where

L=(DD(3)D(3)D(4)−Λ−1),∣δk⟩=(δk0).(14)\mathcal{L}= \begin{pmatrix} \mathcal{D} & \mathcal{D}^{(3)}\\ \mathcal{D}^{(3)} & \mathcal{D}^{(4)}-\Lambda^{-1} \end{pmatrix}, \qquad |\delta_k\rangle= \begin{pmatrix} \delta_k\\ 0 \end{pmatrix}. \tag*{(14)}

The L\mathcal{L} is an operator in this vector space determining how the density matrix propagates. Lastly, the δk\delta_k is a 3N3N vector whose all elements are 0 except the kk-th index, which is 1. In the SSCHA language, the R(1)\mathcal{R}^{(1)} and Y(1)\mathcal{Y}^{(1)} blocks correspond, respectively, to the propagation of one and two phonons.

The free energy Hessian can be obtained from the inverse of L\mathcal{L} by solving the linear system in Eq. (13): let Xk\mathbf{X}^k be the one-phonon block of the solution of Eq. (13), then the kk-th column of the inverse of the positional free energy Hessian is Xk\mathbf{X}^k.

∣ϕk⟩=(XkYk)(15)|\phi_k\rangle= \begin{pmatrix} \mathbf{X}^k\\ \mathbf{Y}^k \end{pmatrix} \tag*{(15)}

Iterating over kk, we can reconstruct the full inverse of the free energy Hessian (i.e., the static susceptibility) and, therefore, the Hessian itself:

G=(X11X12⋯X13NX21⋱⋯X23N⋮⋮⋱⋮X3N1X3N2⋯X3N3N)(16)\mathcal{G}= \begin{pmatrix} X_1^1 & X_1^2 & \cdots& X_1^{3N}\\ X_2^1 & \ddots& \cdots& X_2^{3N}\\ \vdots& \vdots& \ddots& \vdots\\ X_{3N}^1 & X_{3N}^2 & \cdots& X_{3N}^{3N} \end{pmatrix} \tag*{(16)}
d2FdRadRb=(G−1)ab.(17)\frac{d^2F}{dR_a dR_b}=(\mathcal{G}^{-1})_{ab}. \tag*{(17)}

The proof of the equivalence between Eq. (17) and Eq. (7) is reported in detail in Appendix B.

The linear system in Eq. (13) can be solved efficiently with the GMRES algorithm, once we have a practical recipe to apply the L\mathcal{L} operator on a generic ∣ψ⟩|\psi\rangle vector. The GMRES algorithm should be chosen over conjugate gradient approaches because the L\mathcal{L} matrix is not positive definite when the free-energy Hessian is a saddle point. The details of the numerical solver are discussed in Appendix § 13.

Efficient implementation

Recasting the free-energy Hessian calculation as a linear system allows us to exploit efficient iterative algorithms developed for this purpose (see Appendix § 13).

However, inverting Eq. (13) is convenient only if the operator L\mathcal{L} is sparse, which is not the case for a general system in Eq. (14), where all phonons interact. However, here we prove that applying the operator L\mathcal{L} to a generic vector ∣ψ⟩|\psi\rangle can be achieved in O(N)O(N) operations, i.e., without materializing D(4)\mathcal{D}^{(4)} and D\mathcal{D}. A similar technique was already introduced for the dynamical Lanczos algorithm of TD-SCHA[47].

We decompose L=L(har)+L(anh)\mathcal{L}=\mathcal{L}^{(\mathrm{har})}+\mathcal{L}^{(\mathrm{anh})}, where

L(har)=(D00−Λ−1),L(anh)=(0D(3)D(3)D(4)).(18)\mathcal{L}^{(\mathrm{har})}= \begin{pmatrix} \mathcal{D} & 0\\ 0 & -\Lambda^{-1} \end{pmatrix}, \qquad \mathcal{L}^{(\mathrm{anh})}= \begin{pmatrix} 0 & \mathcal{D}^{(3)}\\ \mathcal{D}^{(3)} & \mathcal{D}^{(4)} \end{pmatrix}. \tag*{(18)}

The harmonic part L(har)\mathcal{L}^{(\mathrm{har})} is diagonal in the phonon polarization basis. The anharmonic part L(anh)\mathcal{L}^{(\mathrm{anh})} is the expensive contribution.

The application of L(anh)\mathcal{L}^{(\mathrm{anh})} on the two blocks of a generic ∣ψ⟩|\psi\rangle vector can be written as

L(anh)∣ψ⟩=([L(anh)∣ψ⟩]R(1)[L(anh)∣ψ⟩]Y(1)).\mathcal{L}^{(\mathrm{anh})}|\psi\rangle= \begin{pmatrix} [\mathcal{L}^{(\mathrm{anh})}|\psi\rangle]_{\mathcal{R}^{(1)}}\\ [\mathcal{L}^{(\mathrm{anh})}|\psi\rangle]_{\mathcal{Y}^{(1)}} \end{pmatrix}.

Each component of this vector is explicitly expressed as

[L(anh)∣ψ⟩]Ra(1)=∑bcDabc(3)Ybc(1).(19)[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle]_{\mathcal{R}^{(1)}_a} = \sum_{bc}\mathcal{D}^{(3)}_{abc}\mathcal{Y}^{(1)}_{bc}. \tag*{(19)}
[L(anh)∣ψ⟩]Yab(1)=∑cDabc(3)Rc(1)+∑cdDabcd(4)Ycd(1).(20)[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle]_{\mathcal{Y}^{(1)}_{ab}} = \sum_c\mathcal{D}^{(3)}_{abc}\mathcal{R}^{(1)}_c + \sum_{cd}\mathcal{D}^{(4)}_{abcd}\mathcal{Y}^{(1)}_{cd}. \tag*{(20)}

This expression is still computationally extremely expensive.

However, the SSCHA defines the high-rank tensors D(3)\mathcal{D}^{(3)} and D(4)\mathcal{D}^{(4)} as stochastic averages over displacements and forces (see ref. [26]).

Dabc(3)=−∑hkYahYbk⟨uhukuc⟩ρ(21)\mathcal{D}^{(3)}_{abc}=-\sum_{hk}\mathcal{Y}_{ah}\mathcal{Y}_{bk}\langle u_hu_ku_c\rangle_{\rho} \tag*{(21)}
Dabcd(4)=−∑hklYahYbkYcl⟨uhukulfd⟩ρ(22)\mathcal{D}^{(4)}_{abcd}=-\sum_{hkl}\mathcal{Y}_{ah}\mathcal{Y}_{bk}\mathcal{Y}_{cl}\langle u_hu_ku_lf_d\rangle_{\rho} \tag*{(22)}

where uu is the centroid displacement for each configuration in the ensemble (R−R)(\mathbf{R}-\mathcal{R}) and ff is the BO force subtracted from the auxiliary harmonic force:

ua=(Ra−Ra),fa=−dVdRa+∑cDacuc.u_a=(R_a-\mathcal{R}_a), \qquad f_a=-\frac{dV}{dR_a}+\sum_c\mathcal{D}_{ac}u_c.

The average ⟨⋅⟩ρ\langle\cdot\rangle_{\rho} is performed over the equilibrium SSCHA Gaussian density matrix. Note that both ff and uu are mass-rescaled quantities (see Appendix A). Analogously, Y\mathcal{Y} is the (mass-rescaled) inverse covariance matrix:

Yab=∑μεμaεμb2ωμ2nμ+1.(23)\mathcal{Y}_{ab}=\sum_{\mu}\varepsilon^a_{\mu}\varepsilon^b_{\mu}\frac{2\omega_{\mu}}{2n_{\mu}+1}. \tag*{(23)}

To compact the notation of the high-rank tensor in Eq. (21) and Eq. (22), we can define a new vector v\mathbf{v} as the product between Υ\boldsymbol{\Upsilon} and the displacement u\mathbf{u}

va=∑hΥahuh.v_a = \sum_h \Upsilon_{ah}u_h.

The advantage of Eq. (21) and Eq. (22) is that the high-rank tensors can be written as averages of the outer product between vectors. Replacing Eq. (21) and Eq. (22) inside Eq. (19) and Eq. (20), we get

[L(anh)∣ψ⟩]Ra(1)=−⟨va∑bcvbfcΥbc(1)⟩ρ,(24)\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{\mathcal{R}^{(1)}_a} = -\left\langle v_a \sum_{bc}v_b f_c \Upsilon_{bc}^{(1)}\right\rangle_\rho, \tag*{(24)}
[L(anh)∣ψ⟩]Υab(1)=−⟨vavb∑cfcRc(1)⟩ρ−⟨vavb∑cdvcfdΥcd(1)⟩ρ,(25)\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{\Upsilon_{ab}^{(1)}} = -\left\langle v_av_b\sum_c f_c\mathcal{R}_c^{(1)}\right\rangle_\rho-\left\langle v_av_b\sum_{cd}v_cf_d\Upsilon_{cd}^{(1)}\right\rangle_\rho, \tag*{(25)}

By introducing a configuration dependent weight w[ψ](R)w[\psi](\mathbf{R}), the application of L(anh)\mathcal{L}^{(\mathrm{anh})} becomes

[L(anh)∣ψ⟩]Ra(1)=⟨w[ψ]fa⟩ρ\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{\mathcal{R}^{(1)}_a} = \left\langle w[\psi]f_a\right\rangle_\rho
[L(anh)∣ψ⟩]Υab(1)=−⟨w[ψ]d2(V−V)dR~adR~b⟩ρ,(26)\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{\Upsilon_{ab}^{(1)}} = -\left\langle w[\psi]\frac{d^2(V-\mathcal{V})}{d\widetilde{R}_a d\widetilde{R}_b}\right\rangle_\rho, \tag*{(26)}
w[ψ](R)=∑aRa(1)∂ln⁡ρ∂Ra+2∑abhkΥab(1)ΥahΥbk∂ln⁡ρ∂Υhk.(27)w[\psi](\mathbf{R}) = \sum_a \mathcal{R}_a^{(1)}\frac{\partial\ln\rho}{\partial\mathcal{R}_a}+2\sum_{abhk}\Upsilon_{ab}^{(1)}\Upsilon_{ah}\Upsilon_{bk}\frac{\partial\ln\rho}{\partial\Upsilon_{hk}}. \tag*{(27)}

We recall that ρ\rho is the Gaussian equilibrium density matrix from the SSCHA Hamiltonian, i.e.

ρ(R)=det⁡Υ(2π)3Nexp⁡[−12∑ab(Ra−Ra)Υab(Rb−Rb)].(28)\rho(\mathbf{R})=\sqrt{\frac{\det\Upsilon}{(2\pi)^{3N}}}\exp\left[-\frac{1}{2}\sum_{ab}(R_a-\mathcal{R}_a)\Upsilon_{ab}(R_b-\mathcal{R}_b)\right]. \tag*{(28)}

Eq. (26) has the great advantage of never materializing the high-rank force-constants D(3)\mathbf{D}^{(3)} and D(4)\mathbf{D}^{(4)}, providing a much faster implicit way to define them; each matrix-vector product therefore requires only O(Nmodes2)O(N_{\mathrm{modes}}^2) operations per configuration, compared to O(Nmodes4)O(N_{\mathrm{modes}}^4) for explicit construction of D\mathbf{D}. This trick was first introduced in the TD-SCHA Lanczos algorithm from ref. [47]. The weight w[ψ]w[\psi] can be decomposed into the contributions given by the centroid block R(1)\mathcal{R}^{(1)} and the two-phonons block Υ(1)\Upsilon^{(1)}:

w[ψ](R)=w[R(1)](R)+w[Υ(1)](R)w[\psi](\mathbf{R})=w[\mathcal{R}^{(1)}](\mathbf{R})+w[\Upsilon^{(1)}](\mathbf{R})
w[R(1)](R)=−∑aRa(1)va(29)w[\mathcal{R}^{(1)}](\mathbf{R})=-\sum_a\mathcal{R}_a^{(1)}v_a \tag*{(29)}
w[Υ(1)](R)=−∑abΥab(1)vavb(30)w[\Upsilon^{(1)}](\mathbf{R})=-\sum_{ab}\Upsilon_{ab}^{(1)}v_av_b \tag*{(30)}

where the derivative of the normalization factor in Eq. (28) can be omitted, as it leads to a weight independent of the atomic position that averages to zero in the equilibrium SSCHA result. In particular, the ⟨w[R(1)]fa⟩ρ\left\langle w[\mathcal{R}^{(1)}]f_a\right\rangle_\rho is zero in the SSCHA equilibrium and can be omitted, resulting in

[L(anh)∣ψ⟩]Ra(1)=⟨w[Υ(1)]fa⟩ρ\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{\mathcal{R}^{(1)}_a}=\left\langle w[\Upsilon^{(1)}]f_a\right\rangle_\rho
[L(anh)∣ψ⟩]Υab(1)=−⟨(w[R(1)]+w[Υ(1)])d2(V−V)dRadRb⟩ρ(31)\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{\Upsilon_{ab}^{(1)}}=-\left\langle\left(w[\mathcal{R}^{(1)}]+w[\Upsilon^{(1)}]\right)\frac{d^2(V-\mathcal{V})}{dR_a dR_b}\right\rangle_\rho \tag*{(31)}

Indeed, the permutation symmetry of the D(3)\mathbf{D}^{(3)} and D(4)\mathbf{D}^{(4)} indices can be exploited by averaging the results obtained by swapping the force ff with the vv variables. Applying Eq. (31) is still very expensive, as the ∣δk⟩|\delta_k\rangle perturbation breaks translational symmetry. Therefore, translational symmetries inside the supercell must be imposed a posteriori by averaging the result among symmetry-equivalent perturbations, leading to a still unfavorable O(N3)O(N^3) scaling: NN per translation times N2N^2 for the w[Υ(1)]w[\Upsilon^{(1)}] and the d(V−V)/dRadRbd(V-\mathcal{V})/dR_a dR_b. The promising linear scaling is achieved in Sec. III B, where we reformulate Eq. (31) in reciprocal space. There, momentum conservation reduces the computational cost by a factor of Nq2N_q^2, where NqN_q is the qq-space mesh, thus leading to linear scaling (N=NqNucN=N_qN_{\mathrm{uc}}, where NucN_{\mathrm{uc}} is the number of atoms in the primitive cell and fixed).

TD-SCHA in momentum basis

Linear scaling with supercell size can be achieved thanks to momentum conservation. This property can be easily enforced on the vector ∣ψ⟩|\psi\rangle when the external perturbation is monochromatic, therefore avoiding averaging over all translational replicas of Eq. (31). Moreover, the momentum basis also sparsifies L(anh)\mathcal{L}^{(\mathrm{anh})} thanks to the Bloch theorem. Let us start with the definition of a vector in momentum space. Let vκ(R)v_\kappa(\mathbf{R}) be a real-space vector, where κ\kappa is the atom-cartesian index in the primitive cell, while R\mathbf{R} is a lattice vector indicating on which supercell the κ\kappa atom is. vκ(R)v_\kappa(\mathbf{R}) transforms in vκ(q)v_\kappa(\mathbf{q}) as

vκ(q)=1Nq∑Re−2πiR⋅qvκ(R).(32)v_\kappa(\mathbf{q})=\frac{1}{\sqrt{N_q}}\sum_{\mathbf{R}}e^{-2\pi i\mathbf{R}\cdot\mathbf{q}}v_\kappa(\mathbf{R}). \tag*{(32)}

This phase convention also defines the form of the dynamical matrix in q\mathbf{q} space:

Dκκ′(q)=∑Re−2πiR⋅qDκ,κ′(R),(33)D_{\kappa\kappa'}(\mathbf{q})=\sum_{\mathbf{R}}e^{-2\pi i\mathbf{R}\cdot\mathbf{q}}D_{\kappa,\kappa'}(\mathbf{R}), \tag*{(33)}

where Dκκ′(R)D_{\kappa\kappa'}(\boldsymbol{R}) is the force-constant matrix between atom κ\kappa in the reference primitive cell and atom κ′\kappa' in another primitive cell. The vector R\boldsymbol{R} is the lattice vector connecting the origins of the two primitive cells. In q\boldsymbol{q}-space, the dynamical matrix has eigenvalues ωμ2(q)\omega_{\mu}^{2}(\boldsymbol{q}) and eigenvectors eμκ(q)\boldsymbol{e}_{\mu}^{\kappa}(\boldsymbol{q}).

The key advantage of the q\boldsymbol{q}-space formulation arises when we use a monochromatic external perturbation at a given q\boldsymbol{q} vector; then the response must also be at the same q\boldsymbol{q} vector (modulo a reciprocal lattice vector G\boldsymbol{G}). This means that the perturbed centroids oscillate only at momentum q\boldsymbol{q}, reducing the degrees of freedom in the R(1)\mathcal{R}^{(1)} sector from the atoms in the supercell (3Nsc3N_{\mathrm{sc}}) to the atoms on the primitive cell (3Nuc3N_{\mathrm{uc}}). Even better for the covariance matrix perturbation Υ(1)\Upsilon^{(1)}, which, when transformed in Fourier space, becomes:

Υab(1)(q1,q2)=1Nq∑R1,R2e−2πi(q2⋅R2+q1⋅R1)Υab(1)(R1,R2).(34)\Upsilon_{ab}^{(1)}(\boldsymbol{q}_{1},\boldsymbol{q}_{2})=\frac{1}{N_{q}}\sum_{\boldsymbol{R}_{1},\boldsymbol{R}_{2}}e^{-2\pi i(\boldsymbol{q}_{2}\cdot\boldsymbol{R}_{2}+\boldsymbol{q}_{1}\cdot\boldsymbol{R}_{1})}\Upsilon_{ab}^{(1)}(\boldsymbol{R}_{1},\boldsymbol{R}_{2}). \tag*{(34)}

The crucial observation is that Υ(1)(q1,q2)\Upsilon^{(1)}(\boldsymbol{q}_{1},\boldsymbol{q}_{2}) is excited from the centroid sector R(1)(q)\mathcal{R}^{(1)}(\boldsymbol{q}) only through the three-phonon vertex D(3)D^{(3)}, see Eq. (20). In q\boldsymbol{q} space, D(3)D^{(3)} must conserve the momentum so that the sum of the incoming q\boldsymbol{q} (coming from R(1)\mathcal{R}^{(1)}) must be equal to the sum of the outgoing momenta q1+q2\boldsymbol{q}_{1}+\boldsymbol{q}_{2} (modulo a reciprocal lattice vector G\boldsymbol{G})

q1+q2=q+G.(35)\boldsymbol{q}_{1}+\boldsymbol{q}_{2}=\boldsymbol{q}+\boldsymbol{G}. \tag*{(35)}

The application of L(anh)\mathcal{L}^{(\mathrm{anh})} on the covariance matrix sector reads

[L(anh)∣ψ⟩]abΥ(1)(q1,q2)=∑c,q3Dabc(3)(q1,q2,−q3)Rc(1)(q3)+∑cd,q3,q4Dabcd(4)(q1,q2,−q3,−q4)Υcd(1)(q3,q4).(36)\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{ab}^{\Upsilon^{(1)}(\boldsymbol{q}_{1},\boldsymbol{q}_{2})} =\sum_{c,\boldsymbol{q}_{3}}D_{abc}^{(3)}(\boldsymbol{q}_{1},\boldsymbol{q}_{2},-\boldsymbol{q}_{3})\mathcal{R}_{c}^{(1)}(\boldsymbol{q}_{3}) +\sum_{cd,\boldsymbol{q}_{3},\boldsymbol{q}_{4}}D_{abcd}^{(4)}(\boldsymbol{q}_{1},\boldsymbol{q}_{2},-\boldsymbol{q}_{3},-\boldsymbol{q}_{4})\Upsilon_{cd}^{(1)}(\boldsymbol{q}_{3},\boldsymbol{q}_{4}). \tag*{(36)}

The two processes can be depicted in diagrams of Figure 2 and 3, where R(1)(q)\mathcal{R}^{(1)}(\boldsymbol{q}) represents a one-phonon propagation while Υ(1)(q1,q2)\Upsilon^{(1)}(\boldsymbol{q}_{1},\boldsymbol{q}_{2}) represents the two-phonons propagation. If the original perturbation is monochromatic in q\boldsymbol{q}, and the centroid is constrained only on R(1)(q)\mathcal{R}^{(1)}(\boldsymbol{q}), and imposing the momentum conservation rule inside the D(3)D^{(3)} and D(4)D^{(4)}, we immediately see that Υ(1)\Upsilon^{(1)} is constrained only to be defined at q1,q−q1\boldsymbol{q}_{1},\boldsymbol{q}-\boldsymbol{q}_{1} (for any possible q1\boldsymbol{q}_{1})

Three-phonon scattering vertex diagram

Figure 2. Three-phonon scattering vertex D(3)D^{(3)}, represented as a triangle whose three corners connect to phonon lines. A perturbation at wave vector q\boldsymbol{q} (left vertex) decays into two phonons with momenta q1\boldsymbol{q}_{1} and q2\boldsymbol{q}_{2} (right vertices) satisfying momentum conservation q1+q2=q+G\boldsymbol{q}_{1}+\boldsymbol{q}_{2}=\boldsymbol{q}+\boldsymbol{G}.

[L(anh)∣ψ⟩]abΥ(1)(q1,q−q1)=∑cDabc(3)(q1,q−q1,−q)Rc(1)(q)+∑cd,q2Dabcd(4)(q1,q−q1,−q2,q2−q)Υcd(1)(q2,q−q2).(37)\left[\mathcal{L}^{(\mathrm{anh})}|\psi\rangle\right]_{ab}^{\Upsilon^{(1)}(\boldsymbol{q}_{1},\boldsymbol{q}-\boldsymbol{q}_{1})} =\sum_{c}D_{abc}^{(3)}(\boldsymbol{q}_{1},\boldsymbol{q}-\boldsymbol{q}_{1},-\boldsymbol{q})\mathcal{R}_{c}^{(1)}(\boldsymbol{q}) +\sum_{cd,\boldsymbol{q}_{2}}D_{abcd}^{(4)}(\boldsymbol{q}_{1},\boldsymbol{q}-\boldsymbol{q}_{1},-\boldsymbol{q}_{2},\boldsymbol{q}_{2}-\boldsymbol{q})\Upsilon_{cd}^{(1)}(\boldsymbol{q}_{2},\boldsymbol{q}-\boldsymbol{q}_{2}). \tag*{(37)}

The reduction in both the degrees of freedom of R(1)\mathcal{R}^{(1)} and Υ(1)\Upsilon^{(1)}, as well as the drop of one q3\boldsymbol{q}_{3} summation between

Eq. (36) to (37) is the key advantage of using q\boldsymbol{q}-space basis. Indeed, this reduction is only possible if we consider monochromatic perturbations, which is, however, quite general and true for most physically relevant observables (e.g., Raman and IR spectroscopy probe the q→0\boldsymbol{q}\to0 response function, inelastic neutron and X-ray scattering measure the dynamical structure factor resolved at specific q\boldsymbol{q} momenta). Indeed, since we require perturbing ∣δκ⟩|\delta_{\kappa}\rangle on all supercell atoms for computing the free energy Hessian in Eq. (13), we can change basis without loss of generality and pick ∣δ~κ(q)⟩|\tilde{\delta}_{\kappa}(\boldsymbol{q})\rangle as monochromatic perturbations.

The momentum conservation, illustrated in Figure 2 and 3, implies that the static Liouvillian L\mathcal{L} is block-diagonal in q\boldsymbol{q}: for each perturbation wave vector q\boldsymbol{q}, the R(1)\mathcal{R}^{(1)}-sector is always only evaluated at the perturbation q\boldsymbol{q}, while the Υ(1)\Upsilon^{(1)}-sector contains two-phonon blocks labeled by the NqN_q momentum-conserving pairs (q1,q2)(\boldsymbol{q}_{1},\boldsymbol{q}_{2}) with q1+q2=q+G\boldsymbol{q}_{1}+\boldsymbol{q}_{2}=\boldsymbol{q}+\boldsymbol{G},

∣ψq⟩=(R(1)(q)Υ(1)(q1,q−q1)).(38)|\psi_{\boldsymbol{q}}\rangle= \begin{pmatrix} \mathcal{R}^{(1)}(\boldsymbol{q})\\ \Upsilon^{(1)}(\boldsymbol{q}_{1},\boldsymbol{q}-\boldsymbol{q}_{1}) \end{pmatrix}. \tag*{(38)}

As a consequence, fixing q\boldsymbol{q}, the perturbation Υ(1)\Upsilon^{(1)} has the same degrees of freedom as a dynamical matrix, i.e., (3Nuc)2Nq(3N_{\mathrm{uc}})^{2}N_q.

Fourier transforming the stochastic implementation presented in Eq. (24) and 25, we get

[L(anh)∣ψq⟩]Ra(1)(q)=⟨w[Rq(1)]fa(i)(q)⟩ρ,(39)\left[\mathcal{L}^{(\mathrm{anh})} \lvert\psi_q\rangle\right]_{\mathcal{R}^{(1)}_a(q)}=\left\langle w\left[\mathcal{R}^{(1)}_q\right]f_a^{(i)}(q)\right\rangle_\rho, \tag*{(39)}
[L(anh)∣ψq⟩]Rab(1)(q1,q−q1)=⟨(w[Rq(1)]+w[Rq1(1)])∑hΓah(q1)uh(i)(q1)fb(i)(q−q1)⟩ρ,(40)\left[\mathcal{L}^{(\mathrm{anh})}\lvert\psi_q\rangle\right]_{\mathcal{R}^{(1)}_{ab}(q_1,q-q_1)} =\left\langle\left(w\left[\mathcal{R}^{(1)}_q\right]+w\left[\mathcal{R}^{(1)}_{q_1}\right]\right) \sum_h\Gamma_{ah}(q_1)u_h^{(i)}(q_1)f_b^{(i)}(q-q_1)\right\rangle_\rho, \tag*{(40)}

where the weight can also be computed in Fourier space by transforming Eq. (27)

wi[Rq(1)]=−∑ahRa(1)(−q)Γah(q)uh(i)(q),(41)w_i\left[\mathcal{R}^{(1)}_q\right]=-\sum_{ah}\mathcal{R}^{(1)}_a(-q)\Gamma_{ah}(q)u_h^{(i)}(q), \tag*{(41)}
wi[Γ(1)]=−∑abhkq2Γab(1)(−q2,q2−q)Γah(q2)uh(i)(q2)Γbk(q−q2)uk(i)(q−q2).(42)w_i\left[\Gamma^{(1)}\right]=-\sum_{abhkq_2}\Gamma^{(1)}_{ab}(-q_2,q_2-q)\Gamma_{ah}(q_2)u_h^{(i)}(q_2)\Gamma_{bk}(q-q_2)u_k^{(i)}(q-q_2). \tag*{(42)}

The formulation in q\mathbf{q} space has enormous advantages: (i) all the sums of atomic indices run on the primitive cell, making the application of L\mathcal{L} scale as O(Nuc2Nq)O(N_{\mathrm{uc}}^2N_q), and (ii) the momentum conservation is automatically enforced, so there is no need to average over all lattice translations. The cost of applying L\mathcal{L}, taking also symmetries into account, scales as O(Nuc2NqNG)O(N_{\mathrm{uc}}^2N_qN_G), against an O(Nuc2Nq3NG)O(N_{\mathrm{uc}}^2N_q^3N_G) for the same implementation in real space (NGN_G is the number of symmetries in the point group). Notably, this is a better scaling even than evaluating the perturbative diagrams individually, since any D(4)\mathcal{D}^{(4)} vertex introduces an additional summation over qq, as depicted in the lowest-order D(4)\mathcal{D}^{(4)} RPA diagram (Figure 5) already carries an intrinsic Nq2N_q^2 momentum sum. Indeed, evaluating the RPA series perturbatively introduces an additional power to the NqN_q scaling per order (Figure 6) which quickly becomes intractable. The inversion of the L\mathcal{L} operator in qq-space is the best approach compared to any perturbative expansion of diagrams.

Fourth-order contribution to the free energy Hessian in momentum space

Figure 5. Fourth-order contribution D(3)ΛD(4)ΛD(3)\mathcal{D}^{(3)}\Lambda\mathcal{D}^{(4)}\Lambda\mathcal{D}^{(3)} to the free energy Hessian in momentum space. An incoming phonon of quasi-momentum q\mathbf{q} splits at a cubic vertex into the pair (q1,q2)(\mathbf{q}_1,\mathbf{q}_2), which propagates through Λ\Lambda, is scattered by the quartic vertex D(4)\mathcal{D}^{(4)} into the pair (q3,q4)(\mathbf{q}_3,\mathbf{q}_4), and recombines at the second cubic vertex into an outgoing phonon of the same quasi-momentum. Quasi-momentum is conserved at each vertex up to a reciprocal lattice vector G\mathbf{G}.

Generic diagram in the Taylor expansion in powers of $D^{(4)}$

Figure 6. Generic diagram in the Taylor expansion in powers of D(4)D^{(4)} of the inverse in Eq. (7). The term represents a contribution of type D(3)ΛD(4)Λ⋯ΛD(4)ΛD(3)D^{(3)}\Lambda D^{(4)}\Lambda\cdots\Lambda D^{(4)}\Lambda D^{(3)} to the free energy Hessian in eq. (7) in momentum space, containing NN quartic vertices and N+1N+1 two-phonon propagators Λ\Lambda. The incoming phonon of quasi-momentum qq splits at the first cubic vertex into a phonon pair which is repeatedly scattered by the quartic vertices before recombining at the second cubic vertex. Each Λ\Lambda carries a pair (q2k−1,q2k)(q_{2k-1},q_{2k}) whose total quasi-momentum equals qq up to a reciprocal lattice vector, so that the resummation of all such terms is exactly the geometric series solved by the iterative algorithm.

Benchmarking and scaling

Scaling analysis

In Figure 7, we report the measured computational time and memory scaling with system size for the free-energy-Hessian calculation in an FCC gold structure. We compare the direct evaluation with the new linear-system method in momentum space presented in this work (section III). For each method, we show two curves: one corresponding to the bubble approximation (Eq. (11)), and one to the full fourth-order expression (Eq. (7)).

Figure 7. Computational time and memory peak consumption required to evaluate the free-energy Hessian as functions of the number of atoms in a FCC gold crystal (one atom per primitive cell), for both the bubble approximation (Eq. (11)) and the full RPA inversion (Eq. (7)). Black markers and lines correspond to the direct (legacy) formulation, which is extremely prohibitive to use for large systems since it requires the explicit construction of third- and fourth-order tensors. Red markers and lines correspond to the qq-space formulation introduced in Sec. III B, demonstrating the qualitative change in scaling laws achieved by the novel approach without introducing any additional approximation (the results are comparable to machine precision for the system size reachable for the legacy algorithm). The observed scaling laws are consistent with those reported in Table I. All data points were obtained for FCC gold using ensembles of nc=400n_\mathrm{c} = 400 configurations at T=100T = 100K. Peak memory usage is estimated by monitoring the process during the evaluation of the Free Energy Hessian (details in Appendix § 15). The comparison to algorithms developed in similar legacy codes is discussed in Appendix § 16.

Since FCC Gold has one atom per primitive cell, the number of atoms coincides with the number of qq-points: N=NqN=N_q. The measured time and memory scaling is consistent with the theoretical estimates in Table I, and demonstrates the fundamental importance of the Q-space formulation for simulating large systems that are entirely inaccessible through the direct approach.

The direct evaluation based on Eqs. (21) and (22), implemented in the SSCHA code until now (legacy[25, 26]), requires the full materialization of the D(3)\mathcal{D}^{(3)} and D(4)\mathcal{D}^{(4)} tensors, which needs respectively O(Nq3)O(N_q^3) and O(Nq4)O(N_q^4) memory. To compute each element of the tensor, we perform a stochastic average over the ensemble of ncn_c configurations. So the time needed to materialize the tensors scales as O(ncNq3)O(n_cN_q^3) and O(ncNq4)O(n_cN_q^4). This is also consistent with the calculation of the same quantity by different approaches, like fitting from an ensemble of configurations a polynomial potential, as implemented in TDEP[21], ALAMODE[50], hiPhive[43] (more details about comparison with other legacy approaches are in Appendix G).

Remember that computing the full free energy Hessian directly involves the inversion of the rank-4 tensor

CodeTimeMemory
Legacy SSCHA (D(4)D^{(4)})O(Nuc6Nq6+ncNuc4Nq4)O(N_{\mathrm{uc}}^{6}N_{q}^{6}+n_{c}N_{\mathrm{uc}}^{4}N_{q}^{4})O(Nuc4Nq4)O(N_{\mathrm{uc}}^{4}N_{q}^{4})
Legacy SSCHA (Bubble approx.)O(Nuc3Nq4+ncNuc3Nq3)O(N_{\mathrm{uc}}^{3}N_{q}^{4}+n_{c}N_{\mathrm{uc}}^{3}N_{q}^{3})O(Nuc3Nq3)O(N_{\mathrm{uc}}^{3}N_{q}^{3})
Best theoretical Bubble approx.O(Nuc3Nq2)O(N_{\mathrm{uc}}^{3}N_{q}^{2})O(Nuc3Nq2)O(N_{\mathrm{uc}}^{3}N_{q}^{2})
TD-SCHA Lanczos[47]O(ncNuc3Nq4)O(n_{c}N_{\mathrm{uc}}^{3}N_{q}^{4})O(ncNucNq)O(n_{c}N_{\mathrm{uc}}N_{q})
This workO(ncNuc3Nq2)O(n_{c}N_{\mathrm{uc}}^{3}N_{q}^{2})O(ncNucNq)O(n_{c}N_{\mathrm{uc}}N_{q})

Table 1. Computational scaling for the full free energy Hessian. NqN_q: number of qq-points, NucN_{\mathrm{uc}}: number of atoms in the primitive cell, ncn_c: stochastic configurations. All iterative methods assume a system-size-independent number of solver steps. Notably, both the TD-SCHA Lanczos and this work are the only references where the cost of computing the bubble or the full RPA has exactly the same scaling, as they are differentiated by considering or neglecting w[Υq(1)]w[\Upsilon_q^{(1)}] inside Eq. (40), which shares the same scaling with the bubble only calculation. For comparison with other algorithms by different codes, see Appendix § 16.

[1−D(4)Λ]−1,(43)\left[\mathbb{1}-D^{(4)}\Lambda\right]^{-1}, \tag*{(43)}

where we interpret the inversion by grouping indices two-by-two to transform a rank-4 tensor in a rank-2 matrix, i.e. D(ab)(cd)(4)D^{(4)}_{(ab)(cd)} becomes a matrix of dimensions (3Nq)2×(3Nq)2(3N_q)^2 \times(3N_q)^2. Inverting Eq. (43) requires O[(Nq2)3]=O(Nq6)O[(N_q^2)^3]=O(N_q^6) operations. For the system sizes considered in the benchmark, the stochastic construction of D(4)D^{(4)}, which also carries the factor ncn_c, may remain a significant or even dominant contribution to the total computational time. The crossover between the O(ncNq4)O(n_cN_q^4) from D(4)D^{(4)} materialization and the O(Nq6)O(N_q^6) arising from the dense matrix inversion is investigated in Appendix § 15.

Conversely, without any further approximation, the q-space approach introduced in Sec. III B has a time complexity of O(Nq2)O(N_q^2) and linear memory O(Nq)O(N_q). Notably, the same scaling applies to both the bubble approximation and the full Hessian, as they differ only by a small prefactor (Figure 7).

The impact of this reduction in terms of memory and time costs is dramatic. Extrapolating the legacy algorithm to the same FCC gold crystal with 1728 atoms (a 12×12×1212\times12\times12 supercell) would require approximately 5 PB of RAM and more than 300,000 years of computational time. A calculation started by the first Neanderthals would still be running today. In contrast, using the new qq-space linear-system formulation, we evaluated the same quantity on a laptop using only 1 GB of RAM and approximately 3 hours of computational time.

Displacive phase transition in metal-halide perovskite CsSnI3_3

We now demonstrate the capabilities of the approach on a realistic, strongly anharmonic system where converged free-energy Hessians were previously out of reach. CsSnI3_3 is a lead-free metal-halide perovskite of great interest for photovoltaic and optoelectronic applications. Above ∼426 K\sim426\,\mathrm{K} it adopts the cubic Pm3ˉmPm\bar{3}m black-perovskite structure (B-α\alpha phase), which on cooling transforms through a sequence of displacive transitions driven by octahedral tilting from cubic (B-α\alpha), tetragonal (B-β\beta) to finally orthorhombic (B-γ\gamma)[34, 51]. These transitions are governed by zone-boundary soft modes at the M=(12120)M=\left(\frac{1}{2}\frac{1}{2}0\right) and R=(121212)R=\left(\frac{1}{2}\frac{1}{2}\frac{1}{2}\right) points of the cubic Brillouin zone, whose energy is renormalized by anomalously large thermal ionic fluctuations[34, 46]. Resolving the onset of these instabilities requires evaluating the free-energy Hessian on a qq-mesh dense enough to sample the zone boundary and its surroundings. Crucially, the direct inversion of Eq. (7) is in practice limited to the 2×2×22 \times2 \times2 supercell (40 atoms), which commensurates only the bare MM and RR points and cannot resolve the dispersion of the unstable branches; a severe finite-size limitation for a critical phenomenon whose correlation length grows near the transition.

We model the Born–Oppenheimer (BO) surface with a NequIP machine-learning interatomic potential (MLIP) trained on PBEsol DFT data. The details of the training set, model parameters, and quality of the fit are discussed in Appendix § 14.

Exploiting the new qq-space algorithm, we compute the full free-energy Hessian of cubic CsSnI3_3 retaining the four-phonon scattering (i.e., overcoming the bubble approximation) as a function of temperature on supercells up to 6×6×66 \times6 \times6 (1080 atoms, 3240 modes, Nq=216N_q = 216), using a stochastic ensemble of ∼32 000\sim32\,000 configurations. Figure 8 reports the lowest Hessian frequency (square root of the eigenvalue) at MM and RR versus temperature for the 2×2×22 \times2 \times2 (N=40N = 40), 4×4×44 \times4 \times4 (N=320N = 320), and 6×6×66 \times6 \times6 (N=1080N = 1080) supercells.

Figure 8. Convergence of the CsSnI3 free-energy Hessian with supercell size. Lowest square-root eigenvalue of the full (four-phonon) free-energy Hessian at the MM (left) and RR (right) zone-boundary points of the cubic Pm3ˉmPm\bar{3}m phase, as a function of temperature, for the 2×2×22 \times2 \times2 (N=40N = 40), 4×4×44 \times4 \times4 (N=320N = 320), and 6×6×66 \times6 \times6 (N=1080N = 1080) supercells. Negative values (shaded region) denote imaginary frequencies, i.e. a structural instability of the cubic phase (negative Hessian eigenvalue). The small 2×2×22 \times2 \times2 cell—the practical limit of the direct inversion of Eq. (7)—strongly overestimates the phase stability, whereas the 4×4×44 \times4 \times4 and 6×6×66 \times6 \times6 results converge onto each other. The dash-dotted line marks the experimental cubic-to-tetragonal transition temperature[34].

The supercell dependence is dramatic: the 2×2×22 \times2 \times2 cell (i.e., the only size accessible to direct inversion) strongly

overestimates the stability of the cubic phase, predicting both soft modes to harden through zero already near ∼450\sim450 K. Enlarging the cell to 4×4×44 \times4 \times4 softens the modes by up to 55 cm−1^{-1}, bringing the transition temperature to about 600 K, while the 6×6×66 \times6 \times6 data confirm that the 4×4×44 \times4 \times4 result is essentially converged, the residual difference being comparable to the stochastic noise. This level of convergence of the four-phonon free-energy Hessian on a kilo-atom supercell, impossible with any previously available method, is the central result of this work.

Unfortunately, achieving proper supercell convergence reveals that the agreement with experiments reported in previous work[34] likely emerges from a fortuitous error cancellation between the GGA exchange-correlation potential and finite-size effects. In fact, semi-local functionals are well-documented to overestimate the energy barrier for octahedral tilting in metal halide perovskites[52], thereby overestimating the transition temperatures compared to experiments. Consequently, these results underscore the need to revise the CsSnI3_3 phase diagram by combining accurate hybrid functionals with the large-scale supercells enabled by this work.

While our results agree up to double-precision rounding to the legacy algorithm on the small 40-atom cells, we cannot directly compare the two approaches on the 320- and 1080-atom systems, as they are out of reach for the legacy algorithm. Thus, we tracked the phase transition by relaxing the low-symmetry phase (B-β\beta) within SSCHA, monitoring the order parameter as a function of temperature (lattice vectors and octahedral tilting). Figure 9(a) compares the thermal expansion of the B-α\alpha and B-β\beta phases as a function of the supercell size: above the 4×4×44 \times4 \times4 q-mesh, the volume per formula unit is essentially independent of supercell size. The B-β\beta volume expansion converges toward the B-α\alpha at about 600 K, in good agreement with the temperature at which the B-α\alpha phase becomes stable in the MM point. Figure 9(b) reports the two structural parameters related to this phase transition: the tetragonal cell strain (c/ac/a) and the tilting of the II octahedra around the cc axis. Both parameters converge clearly to the cubic value at about 600 K, providing an additional independent quality check of the phase transition estimated from the free energy Hessian.

Structural changes of the cubic and tetragonal phases of perovskite CsSnI3

Figure 9. Structural changes of the cubic (B-α\alpha) and tetragonal (B-β\beta) phase of perovskite CsSnI3\mathrm{CsSnI}_{3}. (a) Thermal expansion. The cubic lattice thermal expansion shows that it is well converged with a 4x4x4 supercell (corresponding to 320 atoms) and completely overlaps with the 6x6x6 supercell (1080 atoms). The tetragonal B-β\beta volume merges with the cubic one at 600 K. The data are compared with experimental data by Chung et al.[51] (figure S4). (b) Structural parameters of B-β\beta phase. Both the strain deformation from cubic and the tilting angle of the octahedra converge toward the expected cubic value (0) at 600 K, confirming that it is the converged phase-transition temperature. The B-β\beta 4x4x4 supercell contains 640 atoms, as its primitive cell is twice the cubic one.

The method introduced here can also be applied to dynamical response functions. Indeed, the efficient evaluation of the L(anh)\mathcal{L}^{(\mathrm{anh})} operator represents the main bottleneck of the Lanczos algorithm within the time-dependent stochastic self-consistent harmonic approximation (TD-SSCHA)[48]. To demonstrate the effectiveness of this approach, we compute the Raman spectrum of CsSnI3_3 just below the β→α\beta\to\alpha transition (at 550 K), where the soft mode responsible for the symmetry breaking at MM in the high-temperature cell becomes Raman active in the broken symmetry phase B-β\beta (Figure 10). This soft mode is strongly anharmonic and accounts for the broad peak observed in the unpolarized Raman spectrum. In the Stokes scattering regime, the Raman spectrum is calculated from the polarizability correlation function C(ω)C(\omega), which is related to the imaginary part of the polarizability response function via the fluctuation-dissipation theorem:

Simulated power-average Raman spectrum of the B-$\beta$ phase of CsSnI$_3$ at 550 K

Figure 10. Simulated power-average Raman spectrum of the B-β\beta phase of CsSnI3_3 at 550 K (Stokes). The β→α\beta\to\alpha transition occurs at ∼\sim600 K within the DFT exchange-correction employed. The soft phonon mode is responsible for most of the signal visible in Raman. In the inset, we report the imaginary part of the polarizability correlation function, which does not account for the Bose scattering peak of the Stokes, enhancing the low-frequency region of the spectrum. The soft mode appears as a large, featureless boson-like peak, evidencing the overdamped nature of the oscillation and its strongly non-Lorentzian shape due to the anharmonicity near the critical point.

C(ω)=−1+n(ω)πIm⁡χunpol(ω).(44)C(\omega) = -\frac{1+n(\omega)}{\pi}\operatorname{Im}\chi_{\mathrm{unpol}}(\omega). \tag*{(44)}

The unpolarized Raman response function χunpol(ω)\chi_{\mathrm{unpol}}(\omega) is obtained by orientational averaging over the isotropic and anisotropic invariants, as discussed in Appendix H.

Discussion

There are multiple approaches to compute the static and dynamical response functions of highly anharmonic materials. SSCHA is an approximate framework that constrains the ansatz for the (dynamical) density matrix to a Gaussian. The ground truth can be computed, in principle, with path-integral molecular dynamics in the static limit, while for exact quantum dynamical response functions things are much more challenging, as exact quantum methods are hampered by the phase problem, which causes the estimation to scale exponentially. These approaches are extremely computationally heavy and challenging when large-scale simulations are required to converge finite-size effects, such as in the presence of displacive phase transitions. To solve this problem, multiple alternatives have been developed. The temperature-dependent effective potential (TDEP) method can fit the static response functions to classical or quantum MD trajectories[18, 21, 53, 54, 55], while self-consistent phonons (SCP)[23, 43, 45, 50] is a mean-field approach similar to the SSCHA, where a phonon self-energy is solved self-consistently. While a detailed comparison across methods is beyond the scope of this work, as already deeply discussed in ref. [46], we just briefly summarize what differentiates the SSCHA: i) SSCHA is a quantum theory, thus correctly capturing quantum nuclear fluctuations, crucial in the presence of light atoms, strong covalent bonds, or at cryogenic temperatures (in contrast to methods that rely on classical MD). ii) The SSCHA does not approximate the BO energy landscape, which retains its full Taylor expansion. Therefore, the SSCHA automatically includes high-order terms, as the phonon-scattering vertices are dressed with higher-order anharmonicity, and always accounts for mixing of polarization vectors (mode-mixing) and anharmonic-driven structural relaxation (the centroids R\mathcal{R} and the lattice via Gibbs free energy minimization[56]). As a consequence, high-order diagrams arising from dressing vertices, often neglected in SCP (like the quintic-third-order bubble[57]) or dynamical TDEP [53] (like the full dynamical RPA self-energy from Figure 6) are automatically included.

The resulting frequency-dependent self-energy allows the SSCHA to describe system spectral functions that deviate from a Lorentzian shape, even when the phonon quasiparticle picture breaks down. In contrast, SSCHA (and its dynamical extension[47, 58]) is limited to Gaussian fluctuations, i.e., the centroids and cross-phonon covariance matrix, which correspond to one-phonon and two-phonon propagators with mode-mixing. Therefore, dynamical diagrams where we have more than two phonon lines propagating at the same time (vertical cut) are forbidden. Among them, the Saturn diagram [48] is not included in the SSCHA self-energy.

We have removed the principal computational barrier to evaluating the full SSCHA free-energy Hessian in periodic crystals. The novel linear-system momentum-space formulation scales linearly with the qq-mesh for each perturbation, enabling the simulation of large-scale systems and systematically overcoming the bubble approximation. This is not only extremely beneficial for identifying second-order phase transitions, which is the main focus of this work, but also enables the calculation of dynamical response functions within the TD-SCHA, as the same reformulation also applies as it is to the time-dependent Lanczos algorithm. Therefore, this work paves the way for simulating phase diagrams and spectral properties in strongly anharmonic materials, where quantum fluctuations play a major role, and supercell-based methods like Path-Integral Molecular Dynamics are computationally prohibitive for dense qq-point grids. A prime example is high-pressure hydrogen, where the anharmonicity is extremely strong, and the bubble approximation has been demonstrated to fail[28]; hydrogen-bond symmetrization in ice[32] and clathrates[33]. It opens a direct route to phase boundaries and momentum-resolved vibrational spectra in systems where four-phonon physics and long correlation lengths must be treated together; the computational wall that once defined the accessible physics has become a controllable convergence parameter.

Acknowledgments

The authors acknowledge EUROHPC and CINECA under the ISCRA initiative for access to HPC resources.

Code Availability Statement

The algorithm described in this work is implemented in the open-source tdscha package[25, 47], and will be published under the GPLv3 license in version 1.7 of the SSCHA software suite (planned for autumn 2026). Starting from version 1.7, this will be the default approach in the python-sscha package for free-energy Hessian calculations, with an opt-out strategy allowing the user to revert to the legacy algorithm. Documentation, tutorials, and news about future schools are regularly updated on the SSCHA website (www.sscha.eu).

Data Availability Statement

All the data represented in this work are available from the corresponding author upon reasonable request.

Mass-rescaled quantities

To simplify equations, it is convenient to work in mass-rescaled quantities, as introduced in ref. [47]. In particular, all stems from the definition of mass-rescaled position, as done in Eq. (4). This converts to forces in the following way:

fa=−dVdRa=−1madVdra=fama(45)f_a=-\frac{dV}{dR_a}=-\frac{1}{\sqrt{m_a}}\frac{dV}{dr_a}=\frac{\mathfrak{f}_a}{\sqrt{m_a}} \tag*{(45)}

where faf_a is the standard Born-Oppenheimer force, i.e., the derivative of the ionic energy landscape with respect to the atomic position. Analogously, the force constants matrices are defined as

Φab=d2Vdradrb,(3)Φabc=d3Vdradrbdrc(46)\Phi_{ab}=\frac{d^2V}{dr_a dr_b},\qquad{}^{(3)}\Phi_{abc}=\frac{d^3V}{dr_a dr_b dr_c} \tag*{(46)}
(4)Φabcd=d4Vdradrbdrcdrd.(47){}^{(4)}\Phi_{abcd}=\frac{d^4V}{dr_a dr_b dr_c dr_d}. \tag*{(47)}

The relationship to the D\mathbf{D}, (3)D{}^{(3)}\mathbf{D}, and (4)D{}^{(4)}\mathbf{D} defined in the paper is

Dab=Φabmamb,(3)Dabc=(3)Φabcmambmc,(48)D_{ab}=\frac{\Phi_{ab}}{\sqrt{m_a m_b}},\qquad{}^{(3)}D_{abc}=\frac{{}^{(3)}\Phi_{abc}}{\sqrt{m_a m_b m_c}}, \tag*{(48)}
(4)Dabcd=(4)Φabcdmambmcmd.(49){}^{(4)}D_{abcd}=\frac{{}^{(4)}\Phi_{abcd}}{\sqrt{m_a m_b m_c m_d}}. \tag*{(49)}

In the same way, the mass-rescaled inverse covariance matrix Υ\Upsilon is defined in the following way:

Ψab=⟨(ra−ra(0))(rb−rb(0))⟩ρΨab=⟨uaub⟩ρ=Ψabmamb(50)\Psi_{ab}=\left\langle(r_a-r_a^{(0)})(r_b-r_b^{(0)})\right\rangle_\rho\qquad\Psi_{ab}=\langle u_a u_b\rangle_\rho=\frac{\Psi_{ab}}{\sqrt{m_a m_b}} \tag*{(50)}

then, the mass-rescaled inverse covariance matrix is

Υ=(Ψ)−1.(51)\Upsilon=(\Psi)^{-1}. \tag*{(51)}

Equivalence proof

Here we prove that the solution of the linear system Eq. (13) yields the free energy Hessian as in Eq. (16). The explicit linear system in Eq. (13) that we need to solve for XkX^k and YkY^k is

{DXk+(3)DYk=δk(3)DXk+(4)DYk−Λ−1Yk=0(52)\begin{cases} D X^k + {}^{(3)}D Y^k = \delta_k \\ {}^{(3)}D X^k + {}^{(4)}D Y^k - \Lambda^{-1}Y^k = 0 \tag*{(52)} \end{cases}

We first find YkY^k from the second equation of (B1)

((4)D−Λ−1)Yk=−(3)DXk\left({}^{(4)}D-\Lambda^{-1}\right)Y^k=-{}^{(3)}D X^k
Yk=−((4)D−Λ−1)−1(3)DXk.(53)Y^k=-\left({}^{(4)}D-\Lambda^{-1}\right)^{-1}{}^{(3)}D X^k. \tag*{(53)}

Substituting in the first equation, we get

[D−(3)D((4)D−Λ−1)−1(3)D]Xk=δk,(54)\left[D-{}^{(3)}D\left({}^{(4)}D-\Lambda^{-1}\right)^{-1}{}^{(3)}D\right]X^k=\delta_k, \tag*{(54)}

which can be solved by inverting the term in the parentheses:

Xk=[D−(3)D((4)D−Λ−1)−1(3)D]−1δk.(55)X^k=\left[D-{}^{(3)}D\left({}^{(4)}D-\Lambda^{-1}\right)^{-1}{}^{(3)}D\right]^{-1}\delta_k. \tag*{(55)}

Since the susceptibility matrix G\mathcal{G} has XkX^k as kk-th column, the full matrix is obtained replacing the δk\delta_k vector with an identity matrix, therefore

G=[D−(3)D((4)D−Λ−1)−1(3)D]−1.(56)\mathcal{G}=\left[D-{}^{(3)}D\left({}^{(4)}D-\Lambda^{-1}\right)^{-1}{}^{(3)}D\right]^{-1}. \tag*{(56)}

Since from Eq. (17) the positional free energy Hessian is the inverse of G\mathcal{G}, then we have

G−1=D−(3)D((4)D−Λ−1)−1(3)D(57)\mathcal{G}^{-1}=D-{}^{(3)}D\left({}^{(4)}D-\Lambda^{-1}\right)^{-1}{}^{(3)}D \tag*{(57)}

which is very similar to the expression of Eq. (7) derived in [26]. In fact, they are exactly equivalent, as

((4)D−Λ)−1=−(Λ−1−(4)D)−1=−[(1−(4)DΛ)Λ−1]−1=−Λ(1−(4)DΛ)−1.(58)\begin{aligned} \left({}^{(4)}D-\Lambda\right)^{-1}=-\left(\Lambda^{-1}-{}^{(4)}D\right)^{-1} \\ &=-\left[(\mathbb{1}-{}^{(4)}D\Lambda)\Lambda^{-1}\right]^{-1} \\ &=-\Lambda(\mathbb{1}-{}^{(4)}D\Lambda)^{-1}. \tag*{(58)} \end{aligned}

Substituting Eq. (58) into Eq. (57), we recover Eq. (7)

G−1=D+(3)DΛ(1−(4)DΛ)−1(3)D=d2FdR dR.(59)\mathcal{G}^{-1}=D+{}^{(3)}D\Lambda(\mathbb{1}-{}^{(4)}D\Lambda)^{-1}{}^{(3)}D=\frac{d^2F}{dR\,dR}. \tag*{(59)}

Notably, the variable YkY^k, obtained from the solution of the linear system and discarded, has a physical meaning: it represents the response of the two-phonon covariance matrix to a perturbation of the average atomic positions kk, capturing how quantum fluctuations relax when the structure is deformed.

Equivalence proof using diagramatic formalism

Here we present the equivalence proof by using the diagrammatic formulation of SCHA [48] introduced in Figure 1, Figure 2, Figure 3. The main advantage of the diagrammatic representation is that it makes the contraction pattern of the indices in third- and fourth-order tensors more transparent and intuitive. In this section, we represent the fourth-order tensor (4)D{}^{(4)}D by a red square, ■\blacksquare. Its two vertices on the left indicate that the first pair of tensor indices is combined into a single composite index of dimension (3N)2(3N)^2, while the two vertices on the right represent the second pair of indices, which is combined in the same way. So, it is an (3N)2×(3N)2(3N)^2 \times(3N)^2 matrix.

Four-phonon scattering vertex diagram

Figure 3. Four-phonon scattering vertex D(4)D^{(4)}, represented as a square whose four corners connect to phonon lines. Two incoming phonons with momenta q1\boldsymbol{q}_{1} and q2\boldsymbol{q}_{2} (left corners) scatter into two outgoing phonons with momenta q3\boldsymbol{q}_{3} and q4\boldsymbol{q}_{4} (right corners), satisfying momentum conservation q1+q2=q3+q4+G\boldsymbol{q}_{1}+\boldsymbol{q}_{2}=\boldsymbol{q}_{3}+\boldsymbol{q}_{4}+\boldsymbol{G}.

The two-phonon propagator Λ\Lambda, similarly, is represented by a double line, \doublebar, whereas the dynamical matrix DD is represented by an inverse of a single line, −1\rule{0.7em}{0.4pt}^{-1}. The inversion comes from the fact that the dynamical matrix is the inverse of the Phonon static Green function.

We represent the third-order tensor (3)D{}^{(3)}D by using yellow triangles. A right-pointing triangle, ▶\blacktriangleright, indicates that the two indices on the left are combined into a single composite index of dimension (3N)2(3N)^2. So that the dimension of ▶\blacktriangleright are (3N)2×3N(3N)^2 \times3N. Conversely, a left-pointing triangle, ◀\blacktriangleleft, indicates that the two indices on the right are combined into the same composite index, so that the dimension is 3N×(3N)23N \times(3N)^2. Notice that one is the transpose of the other ▶=(◀)T\blacktriangleright=(\blacktriangleleft)^T. More generally, the transposition corresponds to a left-right reflection of the associated diagram (exchanging the matrix indices corresponds to flip the two matrix indices, and so flipping the diagram).

For example, take the product (3)DΛ(3)D{}^{(3)}D\Lambda{}^{(3)}D. The result is a 3N×3N3N \times3N matrix, and any element can be written explictly as:

((3)DΛ(3)D)ab=∑(hk)(lm)(3)Da(hk)Λ(hk)(lm)(3)D(lm)b.(60)\left({}^{(3)}D\Lambda{}^{(3)}D\right)_{ab} =\sum_{(hk)(lm)}{}^{(3)}D_{a(hk)}\Lambda_{(hk)(lm)}{}^{(3)}D_{(lm)b}. \tag*{(60)}

We putted inside the brackets (⋅)(\cdot) the couples of indices which are combined and flattened (in this case, (hk)(hk) and (lm)(lm)). This product can be easily written diagramatically as:

(3)DΛ(3)D=▶–◀(61){}^{(3)}D\Lambda{}^{(3)}D=\blacktriangleright\text{--}\blacktriangleleft \tag*{(61)}

Now it is clear which indices are summed just by looking at the diagrams. In this formalism, the free energy Hessian in eq. (7) is an RPA summation (Figure 1):

d2FdR dR=−1+(diagram)(1−(diagram))−1(diagram)=−1+(diagram)+(diagram)(diagram)+⋯(62)\frac{d^2F}{dR\,dR} = -\mathbf{1} + \text{(diagram)}( \mathbf{1} - \text{(diagram)} )^{-1}\text{(diagram)} = -\mathbf{1} + \text{(diagram)} + \text{(diagram)}\text{(diagram)} + \cdots \tag*{(62)}

The operator L\mathcal{L} reads:

L=(−1(diagram)(diagram)(diagram)−(diagram))(63)\mathcal{L} = \begin{pmatrix} -\mathbf{1} & \text{(diagram)}\\ \text{(diagram)} & \text{(diagram)}-\text{(diagram)} \end{pmatrix} \tag*{(63)}

The dimensions of L\mathcal{L} are 3N(1+3N)×3N(1+3N)3N(1+3N) \times3N(1+3N). We saw that the solutions ∣ϕk⟩|\phi_k\rangle of the 3N3N linear systems

L∣ϕk⟩=(δk0),with k=1,…,3N\mathcal{L}|\phi_k\rangle= \begin{pmatrix} \delta_k\\ 0 \end{pmatrix}, \qquad\text{with } k=1,\ldots,3N

can be separated in two parts XkX^k which is a vector containing 3N3N elements, and YkY^k, which contains (3N)2(3N)^2 elements. This block structure has a transparent diagrammatic counterpart: XkX^k is drawn as an object with a single leg, and YkY^k as an object with two legs, the number of legs matching the number of phonon indices carried by each block.

Xk≡(one-leg diagram)k∈R3N(64)X^k \equiv\text{(one-leg diagram)}^k \in\mathbb{R}^{3N} \tag*{(64)}
Yk≡(two-leg diagram)k∈R(3N)2(65)Y^k \equiv\text{(two-leg diagram)}^k \in\mathbb{R}^{(3N)^2} \tag*{(65)}
(−1(diagram)(diagram)(diagram)−(diagram))((one-leg diagram)k(two-leg diagram)k)=(δk0)(66)\begin{pmatrix} -\mathbf{1} & \text{(diagram)}\\ \text{(diagram)} & \text{(diagram)}-\text{(diagram)} \end{pmatrix} \begin{pmatrix} \text{(one-leg diagram)}^k\\ \text{(two-leg diagram)}^k \end{pmatrix} = \begin{pmatrix} \delta_k\\ 0 \tag*{(66)} \end{pmatrix}

So, the linear system becomes:

−1 (one-leg diagram)k+(diagram) (two-leg diagram)k=δk(diagram) (one-leg diagram)k+((diagram)−(diagram)−1)(two-leg diagram)k=0(67)\begin{aligned} -\mathbf{1}\,\text{(one-leg diagram)}^k+\text{(diagram)}\,\text{(two-leg diagram)}^k&=\delta_k\\ \text{(diagram)}\,\text{(one-leg diagram)}^k+ \left(\text{(diagram)}-\text{(diagram)}^{-1}\right)\text{(two-leg diagram)}^k&=0 \tag*{(67)} \end{aligned}

From the second equation, putting 1=(diagram)(diagram)−1\mathbf{1}=\text{(diagram)}\text{(diagram)}^{-1}:

(two-leg diagram)k=−((diagram)−(diagram)−1)−1(diagram) (one-leg diagram)k=−[(diagram)(1−(diagram))(diagram)−1]−1(diagram) (one-leg diagram)k=−(diagram)(1−(diagram))−1(one-leg diagram)k(68)\begin{aligned} \text{(two-leg diagram)}^k &=-\left(\text{(diagram)}-\text{(diagram)}^{-1}\right)^{-1}\text{(diagram)}\,\text{(one-leg diagram)}^k\\ &=-\left[\text{(diagram)}\left(\mathbf{1}-\text{(diagram)}\right)\text{(diagram)}^{-1}\right]^{-1}\text{(diagram)}\,\text{(one-leg diagram)}^k\\ &=-\text{(diagram)}\left(\mathbf{1}-\text{(diagram)}\right)^{-1}\text{(one-leg diagram)}^k \tag*{(68)} \end{aligned}

Putting it inside the first equation:

(−1+(diagram)(1−(diagram))−1(diagram))(one-leg diagram)k=δk(69)\left(-\mathbf{1}+\text{(diagram)}\left(\mathbf{1}-\text{(diagram)}\right)^{-1}\text{(diagram)}\right)\text{(one-leg diagram)}^k=\delta_k \tag*{(69)}

So, the different solutions (one-leg diagram)k\text{(one-leg diagram)}^k of the kk−th linear system are the columns of the inverse of the free energy Hessian:

d2FdR dR=((one-leg diagram)1 (one-leg diagram)2 ⋯ (one-leg diagram)3N)−1(70)\frac{d^2F}{dR\,dR}= \left( \text{(one-leg diagram)}^1\ \text{(one-leg diagram)}^2\ \cdots\ \text{(one-leg diagram)}^{3N} \right)^{-1} \tag*{(70)}

Iterative solver with harmonic preconditioner

We solve Eq. (13) using the GMRES algorithm[47]. The operator L\mathcal{L} is not guaranteed to be positive definite (negative eigenvalues signal a structural instability), so standard conjugate-gradient methods are not applicable in general.

To accelerate convergence, we employ a preconditioner M=(L(har))−1M=(\mathcal{L}^{(\mathrm{har})})^{-1}. Since L(har)\mathcal{L}^{(\mathrm{har})} is diagonal in the phonon basis, its inverse is trivial:

M=(D−100−Λ)(71)M= \begin{pmatrix} D^{-1} & 0\\ 0 & -\Lambda \end{pmatrix} \tag*{(71)}

This preconditioner is exact in the harmonic limit and captures the dominant energy scales, effectively preconditioning away the large frequency spread present near the phase transition, where the frequency of the order parameter shows a softening. Notably, without preconditioning, the condition number CC of this system is dominated by Λ\Lambda exactly like the SSCHA minimization[56]:

C∼(ωmax⁡ωmin⁡)γγ=3kBT≪ω,γ=4kBT≫ωC\sim\left(\frac{\omega_{\max}}{\omega_{\min}}\right)^\gamma \qquad \gamma=3\quad k_{\mathrm{B}}T\ll\omega, \qquad \gamma=4\quad k_{\mathrm{B}}T\gg\omega

Therefore, preconditioning is essential to achieve convergence of the GMRES algorithm in few steps. Empirically, we observed in all tests that 10–30 steps are required to converge the GMRES algorithm to machine precision, mostly independent of the mesh size, with small variation depending on the nearness of the critical point and whether D(4)D^{(4)} is accounted for or neglected (if D(4)D^{(4)} is neglected, we observed systematically less than 10 steps are needed to converge). We do not rule out the possibility that highly anharmonic low-symmetry phases may require more steps to converge, due to the need to relax the polarization vectors compared to the auxiliary SSCHA dynamical matrix. Notably, this algorithm fully preserves mode-mixing, i.e., polarization vectors are free to change both in the SSCHA minimization and in the calculation of the free energy Hessian.

Fit of the CsSnI₃ Born-Oppenheimer energy landscape

The BO energy landscape of CsSnI₃ used for the test has been fitted using as training data the first-principles calculations performed in ref. [34], using the PBEsol exchange-correlation functional with the Quantum ESPRESSO software suite. The fit is performed with nequip version 0.6.2, including in the cost function energies, forces, and stress on the structures sampled from the 40-atom supercell in the B-α\alpha, B-β\beta, B-γ\gamma and Y-δ\delta phases at various temperatures between 250 K and 450 K. The training set is composed of 1200 independent structures sampled from the SSCHA converged density matrices (computed fully ab initio in ref. [34]), while the test set consists of 320. The model is an E(3)-equivariant message-passing neural network with four interaction layers, rotational features up to angular order lmax⁡=1l_{\max}=1, 32 independent equivariant features per atom, and a cutoff radius of 5.0 Å in which interatomic distances are expanded in a basis of 8 radial functions. The model achieves a force root-mean-square error below 50 meV Å−1^{-1}. Gated activation functions (SiLU for even, tanh for odd parities) are used as nonlinearities, and the loss function combines forces (unit weight), total energy per atom (unit weight), and stress (weight 0.1). Training uses the Adam optimizer with AMSGrad, a learning rate of 0.001 and batch size 8, with early stopping upon validation-loss saturation. The quality of the fit is independently assessed by investigating the Harmonic phonon, as displayed in Figure 11, where we compare the DFT reference harmonic phonons to the fitted machine-learned interatomic potential (MLIP).

Harmonic phonon dispersion of cubic CsSnI3

Figure 11. Harmonic phonon dispersion of cubic CsSnI3_3 computed with the NequIP MLIP (red dashed) compared to the underlying DFT reference (blue). The potential faithfully reproduces the harmonic spectrum, including the dynamically unstable branches (imaginary frequencies plotted as negative) at the MM and RR zone-boundary points that drive the displacive transitions.

We explicitly verified that, after relaxing with the SSCHA at finite temperatures on the 40-atom cell, the MLIP yields results perfectly consistent with the full DFT reference; both the auxiliary phonon frequencies and free energy Hessian root eigenvalues agree by less than 2 cm−1^{-1} in the 8 qq points commensurate with the 2x2x2 supercell, further validating the fidelity of the machine-learned model for the SSCHA calculation.

Details about the scaling analysis

Scaling data which are used to produce Figure 7 come from a measurement of time and memory peak of the process in the call of the only functions responsible of the evaluation of the Hessian in the SSCHA and TDSCHA code.

All memory data have offsets due to ensemble loaded and preparation. For the legacy algorithm, most of the memory consumption is associated with the Hessian evaluation. In contrast, for the new linear-system-based method, memory usage is mainly determined by Julia’s initialization and dependencies, which require approximately 800 MB of RAM, with fluctuations of about 20 MB caused by the Julia Garbage Collector, and by the ensembles, whose memory scales also linearly with NN. Consequently, the peak-memory estimate for the proposed algorithm is affected by greater measurement noise. Notably, the total measured peak memory usage for the 12 × 12 × 12 system (N=1728N=1728) was only approximately 1.2 GB, which is about one order of magnitude lower than the memory required by the legacy algorithm for the much smaller 4 × 4 × 4 system (N=64N=64).

We toggled these offset to see the scaling of the only Hessian evaluation algorithms in the log-log scale in Figure 7. We present also the raw data of the total memory peak of the process in Figure 13.

Raw data of total memory peaks

Figure 13. Raw data of the total memory peaks during the evaluation of the Free Energy Hessian with the method presented in this paper. The inset plot is a zoom in linear-linear scale.

The theoretical lines are the result of laws of the type

ya(N)=aNα,y_a(N)=aN^\alpha,

where α\alpha is fixed to be the expected scaling. This curve becomes linear in the log-log scale:

ηa(ξ)=log⁡a+αξ\eta_a(\xi)=\log a+\alpha\xi

With ηa≡log⁡ya\eta_a\equiv\log y_a and ξ≡log⁡N\xi\equiv\log N. The only exception is in the Legacy full Hessian time evaluated with the legacy algorithm, which is fitted on a non-linear function

Ta,bLegacy,Full(N)=aN4+bN6T_{a,b}^{Legacy,Full}(N)=aN^4+bN^6

Notice that the values of these coefficients depends on the specific system, the number of configurations of the ensembles, and the specific implementation. Since only three data points were computed for this fit, in Figure 12 we provide an additional validation using gold with broken cubic symmetry. This setup allowed us to include two additional data points between the 3×3×3 and the 4×4×4 supercell, providing a more robust validation of the fitted curve and, consequently, of the computational time extrapolation to larger systems.

Figure 12. Validation of the aN4+bN6aN^{4}+bN^{6} time dependence on 5 data points of the legacy algorithm in the evaluation of the full Free energy Hessian in a non-cubic Gold. Breaking the cubic symmetries allowed us to add 2 data points between relative to 3x3x4 and 3x4x4 non-cubic supercells. The behavior confirms the correctness of the theoretical prediction in time for the legacy algorithm.

Scaling of TDEP, ALAMODE and Phono3py codes

Here, report the scaling with the number of q points NqN_q of some of the most commonly employed software codes for computing the bubble approximation, and briefly discuss the algorithmic differences with the SSCHA method presented here.

We assume that no approximation is employed on the shape of Dˉ(3)\bar{D}^{(3)}. In the language of the codes, this means that

CodeTime scalingMemory scaling
ALAMODE[50]O(ncNuc2Nq4)\mathcal{O}(n_{\mathrm{c}}N_{\mathrm{uc}}^{2}N_{\mathrm{q}}^{4})O(ncNuc4Nq3)\mathcal{O}(n_{\mathrm{c}}N_{\mathrm{uc}}^{4}N_{\mathrm{q}}^{3})
phono3py[41]O(Nuc3Nq4)\mathcal{O}(N_{\mathrm{uc}}^{3}N_{\mathrm{q}}^{4})O(Nuc3Nq3)\mathcal{O}(N_{\mathrm{uc}}^{3}N_{\mathrm{q}}^{3})
TDEP*[21]O(Nuc6Nq6)\mathcal{O}(N_{\mathrm{uc}}^{6}N_{\mathrm{q}}^{6})O(Nuc6Nq4)\mathcal{O}(N_{\mathrm{uc}}^{6}N_{\mathrm{q}}^{4})

caption: TABLE II. Scaling with the number of qq points of different codes that implement the fitting of the high-order force-constant tensors and the calculation within the bubble approximation (Eq. (11)).

:::

high-rank force constant matrices are computed without any real-space cutoff, and the parameter reduction only comes from symmetries and acoustic sum rules. Also, we assume that the number of configurations employed, ncn_{\mathrm{c}}, does not grow with NqN_{\mathrm{q}}. This latter assumption is not true in general, as the absence of a cutoff makes the number of unknowns in Dˉ(3)\bar{D}^{(3)} grow with the cell size, and therefore, a larger number of force evaluations (ncn_{\mathrm{c}}) is required to determine the Dˉ(3)\bar{D}^{(3)} elements with the same accuracy.

All state-of-the-art codes compute and materialize the high-rank force constant tensors. In particular, for the Bubble approximation, the Dˉ(3)\bar{D}^{(3)}. The codes TDEP[21], ALAMODE[50], and Hiphive[43] obtain Dˉ(3)\bar{D}^{(3)} from fitting the force residuals through a linear system of the form

Ax=fA x = f

where xx are the unknown symmetry-independent entries of Dˉ(3)\bar{D}^{(3)}, thus it has a length of the order O(Nuc3Nq2)\mathcal{O}(N_{\mathrm{uc}}^{3}N_{\mathrm{q}}^{2}), while ff are the first-principles forces in each configuration inside the ensemble (a vector of size 3NucNqnc3N_{\mathrm{uc}}N_{\mathrm{q}}n_{\mathrm{c}}). The aforementioned codes differ in how the system is solved. TDEP solves a linear system of the form

A†Ax=A†f.A^{\dagger}A x = A^{\dagger}f.

This requires materializing A†AA^{\dagger}A, which occupies O(Nuc6Nq4)\mathcal{O}(N_{\mathrm{uc}}^{6}N_{\mathrm{q}}^{4}) in memory, and scales as the number of unknowns to the third power in time, i.e., O(Nuc9Nq6)\mathcal{O}(N_{\mathrm{uc}}^{9}N_{\mathrm{q}}^{6}). Its dimensions are from the dimension of all forces in the ensemble times the number of elements in the Dˉ(3)\bar{D}^{(3)}: O(ncNucNq)×O(Nuc3Nq2)\mathcal{O}(n_{\mathrm{c}}N_{\mathrm{uc}}N_{\mathrm{q}}) \times\mathcal{O}(N_{\mathrm{uc}}^{3}N_{\mathrm{q}}^{2}).

The ALAMODE code[50] also builds a fitting table AA. However, differently from TDEP, it does not form the A†AA^{\dagger}A matrix, but performs the singular value decomposition (SVD) of the AA table directly via LAPACK to invert the system. The computational cost for the SVD decomposition is the square of the row elements times the number of columns:

SVG:O[(ncNucNq)2(Nuc3Nq2)]=O(nc2Nuc5Nq4)\mathrm{SVG}: \qquad\mathcal{O}\left[(n_{\mathrm{c}}N_{\mathrm{uc}}N_{\mathrm{q}})^{2}(N_{\mathrm{uc}}^{3}N_{\mathrm{q}}^{2})\right] = \mathcal{O}(n_{\mathrm{c}}^{2}N_{\mathrm{uc}}^{5}N_{\mathrm{q}}^{4})

Finally, Hiphive[43] project the fitting table AA on the basis of symmetry allowed constrains (including the ASR). While this approach is very powerful when approximated constraints enforcing sparsity are introduced, it determines an unfavorable NqN_{\mathrm{q}} scaling when no cutoff is introduced. In fact, the matrix product between AA and the constraint map (a rank-2 array of size (Nuc3Nq2)×(Nuc3Nq2)(N_{\mathrm{uc}}^{3}N_{q}^{2}) \times(N_{\mathrm{uc}}^{3}N_{q}^{2})) determines the scaling as O(ncNuc7Nq5)\mathcal{O}(n_{c}N_{\mathrm{uc}}^{7}N_{q}^{5}).

All these approaches largely benefit from introducing a sparsity assumption, like a fixed cutoff or a compressed sensing technique, which strongly suppresses the NqN_{q} scaling at the cost of a controlled approximation. Notably, such an assumption could also be employed in the SSCHA, further optimizing the current algorithm, but it is beyond the scope of this work.

Then, these codes perform the bubble approximation to compute the free energy Hessian (like ALAMODE[50]), or to compute the phonon spectral function (like phono3py[41] and TDEP[21]). TDEP, ALAMODE and also phono3py compute the bubble diagram in Figure 4 by iterating over all external q\boldsymbol{q} O(Nq)\mathcal{O}(N_{q}), all internal q1\boldsymbol{q}_{1} O(Nq)\mathcal{O}(N_{q}), then summing over every real-space atom triplet D\mathcal{D}, which grows, without cutoff, as O(Nuc3Nq2)\mathcal{O}(N_{\mathrm{uc}}^{3}N_{q}^{2}) (the specificity on how the D\mathcal{D} is represented changes in the codes, but the scaling is the same). Therefore, the resulting cost of the Bubble calculation is O(Nuc3Nq4)\mathcal{O}(N_{\mathrm{uc}}^{3}N_{q}^{4}). Notably, this is the same cost as the SSCHA legacy algorithm in the bubble calculation. Also in this case, introducing a cutoff could significantly reduce the inner loop, as triplets of atoms far away can be approximated to zero.

Diagrammatic representation of the SSCHA phonon’s bubble term

Figure 4. Diagrammatic representation of the SSCHA phonon’s bubble term D(3)ΛD(3)\mathcal{D}^{(3)}\Lambda\mathcal{D}^{(3)}. An incoming phonon of quasi-momentum q\mathbf{q} decays into two intermediate phonons q1\mathbf{q}_1 and q2\mathbf{q}_2, which recombine at the second vertex into an outgoing phonon of the same quasi-momentum, with q1+q2=q+G\mathbf{q}_1+\mathbf{q}_2=\mathbf{q}+\mathbf{G}.

Notably, the TDEP free energy Hessian is exact when extracted from converged path-integral molecular dynamics trajectories, including interactions that go beyond what all other methods here described[46]. However, if self-consistent TDEP is used, the D\mathcal{D} must be corrected with the full Hessian equation in Eq. (7), like for the SSCHA. Moreover, the bubble approximation is introduced to compute the dynamical spectrum anyway.

It is worth noting that the novel algorithm here introduced not only beats all SOTA codes by at least a factor Nq3N_{q}^{3} in scaling, but at this cost also goes far beyond the bubble approximation, introducing new physics with a much lower computational cost.

Unpolarized Raman within TD-SSCHA

The TD-SSCHA code can efficiently compute diagonal response functions, i.e., response functions whose probe and pump observables coincide. For Raman, the response function is the polarizability αxy\alpha_{xy} defined as the derivative of the total energy to an external static electric field:

αxy=∂2E∂Ex∂Ey(72)\alpha_{xy} = \frac{\partial^{2}\mathcal{E}}{\partial E_{x}\partial E_{y}} \tag*{(72)}

where E\mathcal{E} is an external electric field and E\mathcal{E} is the ground-state electronic total energy. The perturbation vector on the TD-SCHA one-phonon channel is

pk=∑abϵa(in)ϵb(out)∂αab∂Rk(73)\boldsymbol{p}_{k} = \sum_{ab}\epsilon_{a}^{(\mathrm{in})}\epsilon_{b}^{(\mathrm{out})}\frac{\partial\alpha_{ab}}{\partial R_{k}} \tag*{(73)}

where ϵ^\hat{\epsilon} is the versor of incoming and outgoing radiation (Raman is a scattering experiment). The perturbation vector p\boldsymbol{p} is used as the initial state ∣ψ⟩|\psi\rangle of the Lanczos algorithm, whose details are reported elsewhere[47]. The unpolarized Raman response function χunpol(ω)\chi_{\mathrm{unpol}}(\omega) can be computed by averaging over the Placzek invariants. We use the following form of Placzek invariants to compute only diagonal perturbations in the Lanczos algorithm: each run employs a symmetry-adapted Raman perturbation vector dk\boldsymbol{d}_{k} (k=1,…,7k=1,\ldots,7), defined in terms of the Cartesian polarizability components

eijk=∂αij∂Rke_{ijk} = \frac{\partial\alpha_{ij}}{\partial R_{k}}

Finally, the 7 perturbations along the Placzek invariants are

dk={13(exx+eyy+ezz),k=1,12(exx−eyy),k=2,12(exx−ezz),k=3,12(eyy−ezz),k=4,3exy,k=5,3exz,k=6,3eyz,k=7.(74)\boldsymbol{d}_{k} = \begin{cases} \frac{1}{3}(e_{xx}+e_{yy}+e_{zz}), & k=1,\\ \frac{1}{\sqrt{2}}(e_{xx}-e_{yy}), & k=2,\\ \frac{1}{\sqrt{2}}(e_{xx}-e_{zz}), & k=3,\\ \frac{1}{\sqrt{2}}(e_{yy}-e_{zz}), & k=4,\\ \sqrt{3}e_{xy}, & k=5,\\ \sqrt{3}e_{xz}, & k=6,\\ \sqrt{3}e_{yz}, & k=7. \end{cases} \tag*{(74)}

The total unpolarized response function χunpol(ω)\chi_{\mathrm{unpol}}(\omega) is then expressed as a linear combination of the 7 diagonal Green’s functions χk(ω)=⟨dk∣G(ω)∣dk⟩\chi_{k}(\omega)=\langle d_{k}|\mathcal{G}(\omega)|d_{k}\rangle:

χunpol(ω)=45χ1(ω)+7∑k=27χk(ω)(75)\chi_{\mathrm{unpol}}(\omega) = 45\chi_{1}(\omega) + 7\sum_{k=2}^{7}\chi_{k}(\omega) \tag*{(75)}

where χ1(ω)\chi_{1}(\omega) corresponds to the isotropic contribution, while χ2(ω),…,χ7(ω)\chi_{2}(\omega),\ldots,\chi_{7}(\omega) collectively account for the anisotropic invariant.

References

  1. [3]Đ. Dangić, O. Hellman, S. Fahy, and I. Savić, npj Computational Materials 7, 1 (2021), number: 1.
  2. [4]Đ. Dangić, S. Fahy, and I. Savić, Physical Review B 106, 134113 (2022).
  3. [5]U. Aseginolaza, R. Bianco, L. Monacelli, L. Paulatto, M. Calandra, F. Mauri, A. Bergara, and I. Errea, Physical Review Letters 122, 075901 (2019).
  4. [6]U. Aseginolaza, R. Bianco, L. Monacelli, L. Paulatto, M. Calandra, F. Mauri, A. Bergara, and I. Errea, Physical Review B 100, 214307 (2019).
  5. [7]C. Verdi, L. Ranalli, C. Franchini, and G. Kresse, Physical Review Materials 7, L030801 (2023).
  6. [8]D. Shin, S. Latini, C. Schäfer, S. A. Sato, U. De Giovannini, H. Hübener, and A. Rubio, Physical Review B 104, L060103 (2021).
  7. [9]L. Ranalli, C. Verdi, M. Zacharias, J. Evens, F. Giustino, and C. Franchini, Electron Mobilities in SrTiO$_3$ and KTaO$_3$: Role of Phonon Anharmonicity, Mass Renormalization and Disorder (2024), arXiv:2407.18771 [cond-mat].arxiv.org/abs/2407.18771
  8. [10]A. H. Romero, E. K. U. Gross, M. J. Verstraete, and O. Hellman, Physical Review B 91, 214310 (2015).
  9. [11]G. A. S. Ribeiro, L. Paulatto, R. Bianco, I. Errea, F. Mauri, and M. Calandra, Physical Review B 97, 014306 (2018).
  10. [12]X. Li, T. Qiu, J. Zhang, E. Baldini, J. Lu, A. M. Rappe, and K. A. Nelson, Science 364, 1079 (2019).
  11. [13]T. F. Nova, A. S. Disa, M. Fechner, and A. Cavalleri, Science 364, 1075 (2019).
  12. [14]A. A. Melnikov, Y. G. Selivanov, and S. V. Chekalin, Anharmonic coherent dynamics of the soft phonon mode of a PbTe crystal (2023), arXiv:2309.06549 [cond-mat, physics:physics].DOI
  13. [15]B. Cheng, P. L. Kramer, Z.-X. Shen, and M. C. Hoffmann, Physical Review Letters 130, 126902 (2023).
  14. [16]F. Libbi, A. Johansson, B. Kozinsky, and L. Monacelli, Science Advances 11, eadw1634 (2025).
  15. [17]P. Souvatzis, O. Eriksson, M. I. Katsnelson, and S. P. Rudin, Physical Review Letters 100, 095901 (2008).
  16. [18]O. Hellman, I. A. Abrikosov, and S. I. Simak, Physical Review B 84, 180301 (2011).
  17. [19]P. Souvatzis, O. Eriksson, M. Katsnelson, and S. P. Rudin, Computational Materials Science 44, 888 (2009).
  18. [20]A. van Roekeeghem, J. Carrete, and N. Mingo, Computer Physics Communications 263, 107945 (2021).
  19. [21]F. Knoop, N. Shulumba, A. Castellano, J. P. A. Batista, R. Farris, M. J. Verstraete, M. Heine, D. Broido, D. S. Kim, J. Klarbring, I. A. Abrikosov, S. I. Simak, and O. Hellman, Journal of Open Source Software 9, 6150 (2024).
  20. [22]F. Bottin, J. Bieder, and J. Bouchet, Computer Physics Communications 254, 107301 (2020).
  21. [23]T. Tadano and S. Tsuneyuki, Physical Review B 92, 054301 (2015), number of pages: 10.
  22. [24]I. Errea, M. Calandra, and F. Mauri, Physical Review B 89, 064302 (2014).
  23. [25]L. Monacelli, R. Bianco, M. Cherubini, M. Calandra, I. Errea, and F. Mauri, Journal of Physics: Condensed Matter 33, 363001 (2021).
  24. [26]R. Bianco, I. Errea, L. Paulatto, M. Calandra, and F. Mauri, Physical Review B 96, 014111 (2017).
  25. [27]M. Borinaga, I. Errea, M. Calandra, F. Mauri, and A. Bergara, Physical Review B 93, 174308 (2016), number of pages: 8.
  26. [28]L. Monacelli, I. Errea, M. Calandra, and F. Mauri, Nature Physics 17, 63 (2020).
  27. [29]L. Monacelli, M. Casula, K. Nakano, S. Sorella, and F. Mauri, Nature Physics 19, 845 (2023).
  28. [30]I. Errea, M. Calandra, C. J. Pickard, J. R. Nelson, R. J. Needs, Y. Li, H. Liu, Y. Zhang, Y. Ma, and F. Mauri, Nature 532, 81 (2016).
  29. [31]I. Errea, F. Belli, L. Monacelli, A. Sanna, T. Koretsune, T. Tadano, R. Bianco, M. Calandra, R. Arita, F. Mauri, and J. A. Flores-Livas, Nature 578, 66 (2020).
  30. [32]M. Cherubini, L. Monacelli, B. Yang, R. Car, M. Casula, and F. Mauri, Physical Review B 110, 014112 (2024).
  31. [33]L. Monacelli, Maria Rescigno, Alasdair Nicholls, Umbertoluca Ranieri, Simone Di Cataldo, and Livia Eleonora Bove, Physical Review B 112, 10.1103/1cgl-mklx (2025).
  32. [34]L. Monacelli and N. Marzari, Chemistry of Materials 35, 1702 (2023).
  33. [35]L. Ranalli, C. Verdi, L. Monacelli, G. Kresse, M. Calandra, and C. Franchini, Advanced Quantum Technologies 6, 2200131 (2023).
  34. [36]U. Aseginolaza, J. Diego, T. Cea, R. Bianco, L. Monacelli, F. Libbi, M. Calandra, A. Bergara, F. Mauri, and I. Errea, Nature Physics 20, 1288 (2024).
  35. [37]J. S. Zhou, R. Bianco, L. Monacelli, I. Errea, F. Mauri, and M. Calandra, 2D Materials 7, 045032 (2020).
  36. [38]R. Bianco, I. Errea, L. Monacelli, M. Calandra, and F. Mauri, Nano Letters 19, 3098 (2019).
  37. [39]J. Diego, A. H. Said, S. K. Mahatha, R. Bianco, L. Monacelli, M. Calandra, F. Mauri, K. Rossnagel, I. Errea, and S. Blanco-Canosa, Nature Communications 12, 598 (2021).
  38. [40]F. Zhou, W. Nielson, Y. Xia, and V. Ozoliņš, Physical Review Letters 113, 185501 (2014).
  39. [41]A. Togo, L. Chaput, and I. Tanaka, Physical Review B 91, 094306 (2015).
  40. [42]F. Zhou, B. Sadigh, D. Åberg, Y. Xia, and V. Ozoliņš, Physical Review B 100, 184309 (2019).
  41. [43]F. Eriksson, E. Fransson, and P. Erhart, Advanced Theory and Simulations 2, 1800184 (2019).
  42. [44]M. Wu, H. Yang, F. Xie, and L. Huang, Materials 16, 7349 (2023).
  43. [45]C. Lin, J. Han, B. Xu, and N. Marzari, npj Computational Materials 10.1038/s41524-026-02163-1 (2026).
  44. [46]L. Monacelli, Physical Review B 112, 014109 (2025).
  45. [47]L. Monacelli and F. Mauri, Physical Review B 103, 104305 (2021).
  46. [48]A. Siciliano, L. Monacelli, G. Caldarelli, and F. Mauri, Physical Review B 107, 174307 (2023).
  47. [49]A. Togo, L. Chaput, and I. Tanaka, Physical Review B 91, 094306 (2015), number of pages: 31.
  48. [50]T. Tadano, Y. Gohda, and S. Tsuneyuki, Journal of Physics: Condensed Matter 26, 225402 (2014).
  49. [51]I. Chung, J.-H. Song, J. M. Androulakis, C. D. Malliakas, H. Li, A. J. Freeman, J. T. Kenney, and M. G. Kanatzidis, Journal of the American Chemical Society 134, 8579 (2012).
  50. [52]W. Kaiser, M. Carignano, A. A. Alothman, E. Mosconi, A. Kachmar, W. A. Goddard, and F. D. Angelis, The Journal of Physical Chemistry Letters 12, 11886 (2021).
  51. [53]A. Castellano, J. P. A. Batista, and M. J. Verstraete, The Journal of Chemical Physics 159, 234501 (2023).
  52. [54]J. A. Thomas, J. E. Turney, R. M. Iutzi, C. H. Amon, and A. J. H. McGaughey, Physical Review B 81, 081411 (2010).
  53. [55]E. Fransson, M. Slabanja, P. Erhart, and G. Wahnström, Advanced Theory and Simulations 4, 000240 (2021).
  54. [56]L. Monacelli, I. Errea, M. Calandra, and F. Mauri, Physical Review B 98, 024106 (2018).
  55. [57]Y. Xia, Quantum-Anharmonicity-Assisted Three-Phonon Scattering: A Previously Overlooked Same-Order Channel to Four-Phonon Scattering (2026), arXiv:2606 24409 [cond-mat.mtrl-sci].
  56. [58]J.-M. Lihm and C.-H. Park, Physical Review Research 3, L032017 (2021).

Paper details

vlnc:W_rrp78nes
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents