INTRODUCTION
Microscopic theories of magnetic exchange have developed along several closely related directions. Some questions have been discussed for decades but continue to influence the interpretation of electronic-structure calculations. Others have received less attention in recent applications or have become particularly relevant with the development of new electronic functionals and methods for treating itinerant magnetism. Five aspects are especially important for the present work.
The first is the distinction between rotations of the effective exchange field B \mathbf{B} B and rotations of the magnetization M \mathbf{M} M . Their relation is particularly important in the long-wavelength approximation (LWA), where the leading static energy curvature can agree under specified conditions.
The second concerns the relation between derivatives of the electronic energy and the interactions appearing in an effective spin Hamiltonian. These quantities are often identified, although the mapping is neither automatic nor generally unique. Higher-order magnetic interactions and the choice of reference state can lead to different effective exchange parameters even at the static level. In addition, exchange parameters inferred from static energy curvatures need not coincide with those used to reproduce collective spin dynamics.
The third concerns the extension of microscopic exchange calculations beyond local magnetic potentials. Established methods for calculating interatomic exchange parameters J i j J_{ij} J ij have been developed primarily for local exchange-correlation potentials, although the resulting magnetic interactions can extend over many lattice sites. Recent developments in electronic-structure theory introduce explicitly nonlocal magnetic interactions and dependence on gradients of magnetization and other spin-dependent electronic quantities. These developments raise new questions for calculating J i j J_{ij} J ij , since rotations of the magnetization change not only the equilibrium magnetic potential but also the induced electronic response. A consistent treatment of these contributions is necessary to extend microscopic exchange calculations beyond the local-potential approximation.
The fourth concerns the connection between localized and itinerant magnetism. Rotational descriptions are natural when finite local magnetic moments persist, but magnetic correlations can remain important even when the static local moment vanishes. The distinction between rotations of existing moments and fluctuations of the spin density becomes essential in this limit. These regimes require different magnetic variables: rotations of finite moments in one case and fluctuations of the spin density in the other. Their relation is important when interpreting exchange parameters in itinerant magnets.
The fifth concerns the relation between static magnetic energy curvatures and collective spin dynamics. Static exchange calculations are well established, but the frequency dependence of the electronic response is sometimes neglected when their results are used to describe spin waves. Even in the long-wavelength limit, the frequency-linear term is needed to determine the dynamical normalization. The static and dynamical limits must therefore be distinguished.
These questions have a long history in microscopic theories of magnetic exchange. Early approaches were based primarily on the static electronic response to magnetic perturbations. Ruderman and Kittel [1 ] , Kasuya [2 ] , and Yosida [3 ] established the theory of indirect exchange mediated by conduction electrons. Freeman and collaborators subsequently employed realistic band theory to investigate conduction-electron polarization and magnetic coupling through the static susceptibility of a nonmagnetic reference state [4 , 5 ] . Liu [6 ] extended the Ruderman–Kittel–Kasuya–Yosida (RKKY) approach to spin-polarized electronic systems using a Hubbard-like model, while Cooke [7 ] developed related band-structure formulations. These developments established the importance of electronic response in determining magnetic interactions and provided a basis for subsequent treatments of rotations about magnetized states.
Korenman, Murray, and Prange developed the local-band theory of itinerant ferromagnetism [8 , 9 , 10 , 11 , 12 ] , in which the electronic structure remains locally spin polarized while the magnetization direction varies slowly in space and time. They treated these rotations perturbatively about a magnetized state, rather than expanding in a weak exchange splitting, and studied spin-wave dynamics, nonlinear magnetic fluctuations, and thermodynamics. They described collective rotations in a local spin frame while retaining the electronic response. These studies also established the importance of both spatial and frequency-dependent electronic response in the long-wavelength theory of magnetic excitations. Prange and Korenman subsequently employed a Hubbard–Stratonovich formulation to obtain an effective classical Heisenberg model with exchange interactions expressed through spin-polarized electronic Green functions [10 , 11 ] . Wang, Prange, and Korenman [13 ] later applied this approach to Fe and Ni using realistic electronic bands to calculate interatomic exchange interactions.
Calculating the energy of prescribed moment configurations also brought constraining fields and magnetic torques into microscopic exchange calculations. Edwards [14 ] recognized in 1984 that calculating interatomic exchange interactions requires the electronic energy for specified local-moment orientations, which need not be stationary configurations of the unconstrained electronic system. Transverse constraining fields maintain these orientations while allowing the electronic degrees of freedom to adjust. In the same year, Small and Heine [15 ] developed a complementary torque-based approach within electronic theory, deriving the torque acting on an atomic moment from both the electronic equation of motion and the total energy. They showed how the resulting torques could determine interatomic magnetic couplings for arbitrary moment configurations and applied the method to bcc Fe. These studies made constraining fields and magnetic torques practical tools for exchange calculations. Both approaches are now widely used to calculate noncollinear magnetic structures, magnetic torques, and spin dynamics.
A further development was the multiple-scattering formulation of magnetic exchange introduced by Liechtenstein and Katsnelson [16 , 17 ] . They derived expressions for interatomic exchange coefficients from infinitesimal rotations of the local exchange potentials, conventionally interpreting these coefficients as interactions between rotated magnetic moments. These expressions closely matched those Prange and Korenman obtained earlier, but the multiple-scattering formulation provided a practical framework for realistic electronic-structure calculations. The authors also extended the local force theorem (LFT) to magnetic rotations [18 ] , allowing them to evaluate exchange parameters from changes in the electronic spectrum without performing fully self-consistent calculations for each rotated configuration. This formulation became the basis of what is commonly called the LFT approach to exchange interactions. Subsequent extensions of the coherent potential approximation enabled treatment of chemical and magnetic disorder, while relativistic generalizations provided access to Dzyaloshinskii–Moriya (DMI) interactions and anisotropic exchange [19 , 20 ] . Related formulations incorporating electronic self-energies and many-body effects were later developed for correlated materials [21 ] . These methods are useful, but they leave an important distinction between exchange coefficients obtained by rotating the effective magnetic field and those defined by rotating the magnetic moments. The relation between these quantities is central to the present work.
Antropov [22 ] revisited the microscopic theory of exchange interactions using the inverse magnetic susceptibility, χ − 1 \chi^{-1} χ − 1 , and examined the limitations of the long-wavelength approximation. He showed that conventional exchange expressions, including those based on I χ I I\chi I I χ I , describe rotations of the effective exchange field B \mathbf{B} B rather than the magnetization M \mathbf{M} M , although they had commonly been interpreted as interactions between rotated magnetic moments. The distinction becomes important beyond the LWA, where the two approaches generally give different exchange parameters. Antropov also emphasized that the term linear in frequency and the term proportional to q 2 q^2 q 2 are of the same order for long-wavelength ferromagnetic spin waves, since ω ∝ q 2 \omega\propto q^2 ω ∝ q 2 . Consequently, retaining the spatial term while neglecting the frequency dependence is not a consistent approximation for spin-wave dynamics.
In this work, we develop rotational perturbation theory (PT) for rotations of both the exchange field and the magnetic moments, without invoking the local force theorem for the prescribed-field energy derivatives. Constrained-moment rotations additionally require the stationary electronic response. The theory treats the full exchange splitting and distinguishes the two rotational responses beyond the long-wavelength approximation. The complete PT Hessian also exhibits local transverse-frame covariance and the global spin-rotation Ward identity as consequences of the derivation, rather than as imposed conditions. We also describe the extraction of effective Heisenberg exchange parameters and its limitations, and consider noncollinear magnets, ferromagnetic stiffness, antiferromagnetic dynamics under an adiabatic precession assumption, nonlocal interactions, and frequency-dependent self-energies.
GENERAL ELECTRONIC AND ROTATIONAL PERTURBATION THEORY
We start from a magnetic state with its full exchange splitting and spin texture and rotate it by small angles. The electrons respond to rotations at different sites, producing the coupling between them.
Magnetic representations and general electronic perturbation theory
We distinguish rotations of the exchange field B ( r ) \mathbf{B}(\mathbf{r}) B ( r ) , which change the Kohn–Sham (KS) or another effective electronic Hamiltonian, from constrained rotations of the magnetization M ( r ) \mathbf{M}(\mathbf{r}) M ( r ) , which fix the directions of the local moments. The corresponding electronic Hamiltonian may be written as
H ^ K S = H ^ 0 − σ ⋅ B ( r ) = H ^ 0 − 1 2 σ ⋅ Δ ( r ) . (1) \hat{H}_{\mathrm{KS}}=\hat{H}_0-\boldsymbol{\sigma}\cdot\mathbf{B}(\mathbf{r})=\hat{H}_0-\frac{1}{2}\boldsymbol{\sigma}\cdot\boldsymbol{\Delta}(\mathbf{r}). \tag*{(1)} H ^ KS = H ^ 0 − σ ⋅ B ( r ) = H ^ 0 − 2 1 σ ⋅ Δ ( r ) . (1)
Moment directions are imposed through conjugate constraining fields. We first calculate the energy change under field rotations and then derive the moment-rotation curvature from the inverse susceptibility.
The rotated Hamiltonian is written as H = H 0 + V ( 1 ) + V ( 2 ) + … H=H_0+V^{(1)}+V^{(2)}+\ldots H = H 0 + V ( 1 ) + V ( 2 ) + … , with superscripts denoting orders in the rotation angle. The first-order energy change contains one vertex. At second order, both the direct change in the Hamiltonian and the electronic response to two first-order vertices contribute. We use this static expansion for spin spirals and long-wavelength modes and treat time dependence separately in Sec. II.E.
Rotations of the effective magnetic field
For a noncollinear magnetic state, the magnetic part of the one-electron Hamiltonian is
H m a g = − ∑ i b i e i ⋅ σ (2) H_{\mathrm{mag}}=-\sum_i b_i\mathbf{e}_i\cdot\boldsymbol{\sigma} \tag*{(2)} H mag = − i ∑ b i e i ⋅ σ (2)
where e i \mathbf{e}_i e i is a unit vector specifying the direction of the local exchange field, and 2 b i = Δ i 2b_i=\Delta_i 2 b i = Δ i is its exchange splitting in the local orbital space. In the simplest case, b i b_i b i is a scalar. The rotations are transverse and keep the magnitude fixed,
e i → e i + δ e i (3) \mathbf{e}_i\to\mathbf{e}_i+\delta\mathbf{e}_i \tag*{(3)} e i → e i + δ e i (3)
At a reference state stationary with respect to the unconstrained electronic total-energy functional, the first angular derivative of that total energy vanishes:
∂ E ∂ θ i a ∣ 0 = 0 (4) \left.\frac{\partial E}{\partial\theta_i^a}\right|_0=0 \tag*{(4)} ∂ θ i a ∂ E 0 = 0 (4)
A constrained state can still have a torque. We must then include the constraining fields and the geometric terms when calculating the curvature.
At each site, the transverse plane is spanned by two orthonormal vectors u i 1 , u i 2 \mathbf{u}_i^1,\mathbf{u}_i^2 u i 1 , u i 2 perpendicular to e i \mathbf{e}_i e i . An infinitesimal rotation is described by
δ θ i = ∑ a = 1 , 2 θ i a u i a (5) \delta\theta_i=\sum_{a=1,2}\theta_i^a u_i^a \tag*{(5)} δ θ i = a = 1 , 2 ∑ θ i a u i a (5)
The corresponding tangent vectors are
t i a = u i a × e i (6) \mathbf{t}_i^a=\mathbf{u}_i^a\times\mathbf{e}_i \tag*{(6)} t i a = u i a × e i (6)
so that, to first order,
δ e i = ∑ a θ i a t i a (7) \delta\mathbf{e}_i=\sum_a\theta_i^a\mathbf{t}_i^a \tag*{(7)} δ e i = a ∑ θ i a t i a (7)
A small tilt changes the electronic Hamiltonian to first order. At site i i i , it is
V i ( 1 ) = − b i δ e i ⋅ σ (8) V_i^{(1)}=-b_i\delta\mathbf{e}_i\cdot\boldsymbol{\sigma} \tag*{(8)} V i ( 1 ) = − b i δ e i ⋅ σ (8)
or
V i ( 1 ) = ∑ a θ i a V i a , V i a = − b i t i a ⋅ σ (9) V_i^{(1)}=\sum_a\theta_i^a V_i^a,\qquad V_i^a=-b_i\mathbf{t}_i^a\cdot\boldsymbol{\sigma} \tag*{(9)} V i ( 1 ) = a ∑ θ i a V i a , V i a = − b i t i a ⋅ σ (9)
The rotational vertex is transverse to the local exchange field and retains the full exchange splitting, because the expansion is in the rotation angle rather than in the magnetic potential.
More generally, the vertex is the derivative of the electronic Hamiltonian with respect to the chosen tilt. For a rotation through an angle θ i \theta_i θ i about a fixed transverse axis u i \mathbf{u}_i u i ,
e i ( θ i ) = e i + θ i ( u i × e i ) + θ i 2 2 u i × ( u i × e i ) + O ( θ i 3 ) (10) \mathbf{e}_i(\theta_i)=\mathbf{e}_i+\theta_i(\mathbf{u}_i\times\mathbf{e}_i)+\frac{\theta_i^2}{2}\mathbf{u}_i\times(\mathbf{u}_i\times\mathbf{e}_i)+O(\theta_i^3) \tag*{(10)} e i ( θ i ) = e i + θ i ( u i × e i ) + 2 θ i 2 u i × ( u i × e i ) + O ( θ i 3 ) (10)
Since u i \mathbf{u}_i u i is perpendicular to e i \mathbf{e}_i e i , u i × ( u i × e i ) = − e i \mathbf{u}_i\times(\mathbf{u}_i\times\mathbf{e}_i)=-\mathbf{e}_i u i × ( u i × e i ) = − e i , and the corresponding local second-order perturbation is
V i ( 2 ) = b i 2 θ i 2 e i ⋅ σ (11) V_i^{(2)}=\frac{b_i}{2}\theta_i^2\mathbf{e}_i\cdot\boldsymbol{\sigma} \tag*{(11)} V i ( 2 ) = 2 b i θ i 2 e i ⋅ σ (11)
The direct second-order term is local: it changes the on-site curvature, not the coupling between tilts at different sites.
Electron propagation couples rotations at different sites, and the local term completes the rotational Hessian.
The second derivative of the one-electron energy under imposed exchange-field rotations defines the field-rotation Hessian H B H^B H B . In subsequent field-rotation formulas, H B H^B H B is abbreviated as H H H when no confusion with the constrained-moment Hessian is possible:
H i j B , a b = ∂ 2 E ∂ θ i a ∂ θ j b ∣ 0 (12) H_{ij}^{B,ab}=\left.\frac{\partial^2 E}{\partial\theta_i^a\partial\theta_j^b}\right|_0 \tag*{(12)} H ij B , ab = ∂ θ i a ∂ θ j b ∂ 2 E 0 (12)
so that
δ 2 E = 1 2 ∑ i j ∑ a b θ i a H i j B , a b θ j b (13) \delta^2 E = \frac{1}{2}\sum_{ij}\sum_{ab}\theta_i^a H_{ij}^{B,ab}\theta_j^b \tag*{(13)} δ 2 E = 2 1 ij ∑ ab ∑ θ i a H ij B , ab θ j b (13)
For two independently tilted directions, the second derivative of the electronic grand potential at fixed chemical potential is
H i j B , a b = ∑ n f n ⟨ n ∣ ∂ 2 H ^ ∂ θ i a ∂ θ j b ∣ n ⟩ + 1 2 ∑ n m f n − f m ε n − ε m [ ⟨ n ∣ V i a ∣ m ⟩ ⟨ m ∣ V j b ∣ n ⟩ + ⟨ n ∣ V j b ∣ m ⟩ ⟨ m ∣ V i a ∣ n ⟩ ] . (14) H_{ij}^{B,ab} = \sum_n f_n \left\langle n \left| \frac{\partial^2 \hat{H}}{\partial\theta_i^a \partial\theta_j^b} \right| n \right\rangle+ \frac{1}{2}\sum_{nm}\frac{f_n-f_m}{\varepsilon_n-\varepsilon_m}\left[\langle n|V_i^a|m\rangle\langle m|V_j^b|n\rangle+ \langle n|V_j^b|m\rangle\langle m|V_i^a|n\rangle\right]. \tag*{(14)} H ij B , ab = n ∑ f n ⟨ n ∂ θ i a ∂ θ j b ∂ 2 H ^ n ⟩ + 2 1 nm ∑ ε n − ε m f n − f m [ ⟨ n ∣ V i a ∣ m ⟩ ⟨ m ∣ V j b ∣ n ⟩ + ⟨ n ∣ V j b ∣ m ⟩ ⟨ m ∣ V i a ∣ n ⟩ ] . (14)
(14) contains both contributions to the rotational Hessian: the direct second-order change in the Hamiltonian and the electronic response to two first-order vertices. Virtual electronic transitions generate the off-site interaction.
This form parallels earlier perturbative treatments [1 , 2 , 3 ] and more recent exchange formulas [13 , 16 , 18 ] . Because both transverse vertices are kept, we treat noncollinear and collinear states on the same footing, without assuming a Heisenberg model.
For distinct sites,
∂ 2 H ^ ∂ θ i a ∂ θ j b = 0 , i ≠ j (15) \frac{\partial^2 \hat{H}}{\partial\theta_i^a\partial\theta_j^b}=0,\qquad i\ne j \tag*{(15)} ∂ θ i a ∂ θ j b ∂ 2 H ^ = 0 , i = j (15)
and the off-site pair Hessian is therefore
H i j a b = 1 2 ∑ n m f n − f m ε n − ε m [ ⟨ n ∣ V i a ∣ m ⟩ ⟨ m ∣ V j b ∣ n ⟩ + ⟨ n ∣ V j b ∣ m ⟩ ⟨ m ∣ V i a ∣ n ⟩ ] , i ≠ j . (16) H_{ij}^{ab} = \frac{1}{2}\sum_{nm}\frac{f_n-f_m}{\varepsilon_n-\varepsilon_m}\left[\langle n|V_i^a|m\rangle\langle m|V_j^b|n\rangle+ \langle n|V_j^b|m\rangle\langle m|V_i^a|n\rangle\right],\qquad i\ne j. \tag*{(16)} H ij ab = 2 1 nm ∑ ε n − ε m f n − f m [ ⟨ n ∣ V i a ∣ m ⟩ ⟨ m ∣ V j b ∣ n ⟩ + ⟨ n ∣ V j b ∣ m ⟩ ⟨ m ∣ V i a ∣ n ⟩ ] , i = j . (16)
For rotations of the specified one-electron Hamiltonian, this is the exact curvature of the one-electron grand potential at fixed chemical potential; the LFT is unnecessary to derive it, and it does not include self-consistent moment relaxation. The same pair curvature can be expressed through the retarded one-electron Green function G R ( ε ) = [ ε + i 0 + − H r e f ] − 1 G^{R}(\varepsilon)=[\varepsilon+i0^{+}-H_{\mathrm{ref}}]^{-1} G R ( ε ) = [ ε + i 0 + − H ref ] − 1 , where H r e f H_{\mathrm{ref}} H ref is the complete spin-dependent reference Hamiltonian, including its exchange splitting (not merely the spin-independent part). At zero temperature,
H i j a b = − 1 2 π Im ∫ E F d ε Tr [ V i a G i j V j b G j i + V j b G j i V i a G i j ] , i ≠ j (17) H_{ij}^{ab}=-\frac{1}{2\pi}\operatorname{Im}\int^{E_F}d\varepsilon\,\operatorname{Tr}\left[V_i^aG_{ij}V_j^bG_{ji}+V_j^bG_{ji}V_i^aG_{ij}\right],\qquad i\ne j \tag*{(17)} H ij ab = − 2 π 1 Im ∫ E F d ε Tr [ V i a G ij V j b G j i + V j b G j i V i a G ij ] , i = j (17)
(17) expresses the intersite response as two rotational vertices connected by electronic propagation. In a noncollinear state, both local transverse directions enter.
The spectral sum and Green-function integral give the same static response when they use the same vertices and full reference propagator. (17) explicitly symmetrizes the two vertex orderings and therefore carries a prefactor of 1 / ( 2 π ) 1/(2\pi) 1/ ( 2 π ) ; the equivalent unsymmetrized full-Hessian trace has prefactor 1 / π 1/\pi 1/ π . There is no additional factor of two. For metals, degenerate states and Fermi-surface terms require the proper static limit. The integral shown is for zero temperature.
Equations (16) and (17) give the prescribed-field pair response at finite splitting and arbitrary wave vector; the relation to constrained-moment interactions is discussed in Sec. II.C and made explicit for ferromagnets in Sec. V.
Local transverse axes are arbitrary; comparisons with LFT calculations require matching rotation coordinates and pair-counting conventions.
To compare the electronic rotational curvatures with a spin model, one can use the isotropic bilinear Heisenberg Hamiltonian
H H = − 1 2 ∑ i ≠ j J i j e i ⋅ e j , ∣ e i ∣ = 1. (18) \mathcal{H}_{H}=-\frac{1}{2}\sum_{i\ne j}J_{ij}\,\mathbf{e}_i\cdot\mathbf{e}_j,\qquad|\mathbf{e}_i|=1. \tag*{(18)} H H = − 2 1 i = j ∑ J ij e i ⋅ e j , ∣ e i ∣ = 1. (18)
Here J i j = J j i J_{ij}=J_{ji} J ij = J j i multiplies unit moment directions rather than full moment amplitudes, J i j > 0 J_{ij}>0 J ij > 0 favors parallel alignment, and the half-sum avoids double counting. This model predicts a particular mixed curvature under two local rotations. For a general bilinear Cartesian interaction tensor J i j α β \mathcal{J}_{ij}^{\alpha\beta} J ij α β in the same half-sum convention, the off-site spin-model Hessian, the spin-model counterpart of the electronic rotational Hessian, is its projection onto the local tangent planes:
H i j a b = − ∑ α β t i α a J i j α β t j β b . (19) H_{ij}^{ab}=-\sum_{\alpha\beta}t_{i\alpha}^{a}\mathcal{J}_{ij}^{\alpha\beta}t_{j\beta}^{b}. \tag*{(19)} H ij ab = − α β ∑ t i α a J ij α β t j β b . (19)
(19) is a spin-model relation, not an identity derived from electronic perturbation theory. The two Hessians can be compared only in matching angular coordinates and pair conventions. The isotropic component of a specified bilinear Cartesian tensor is J i j i s o = 1 3 ∑ α J i j α α J_{ij}^{\mathrm{iso}}=\frac{1}{3}\sum_{\alpha}\mathcal{J}_{ij}^{\alpha\alpha} J ij iso = 3 1 ∑ α J ij α α ; its remaining symmetric traceless and antisymmetric components describe anisotropic bilinear couplings. This tensor decomposition does not determine a unique Cartesian interaction tensor from the transverse electronic Hessian.
For the isotropic Heisenberg model, J i j α β = J i j δ α β \mathcal{J}_{ij}^{\alpha\beta}=J_{ij}\delta_{\alpha\beta} J ij α β = J ij δ α β , and this projection reduces to
H i j a b = − J i j P i j a b , P i j a b = t i a ⋅ t j b . (20) H_{ij}^{ab}=-J_{ij}P_{ij}^{ab},\qquad P_{ij}^{ab}=t_i^a\cdot t_j^b. \tag*{(20)} H ij ab = − J ij P ij ab , P ij ab = t i a ⋅ t j b . (20)
Sec. IV applies this test to spin spirals.
The local second-order contribution is required by global spin-rotation invariance. In the absence of spin-orbit coupling, a single infinitesimal laboratory-frame rotation δ Θ \delta\boldsymbol{\Theta} δ Θ changes each reference direction according to
δ e i = δ Θ × e i (21) \delta\mathbf{e}_i = \delta\boldsymbol{\Theta} \times\mathbf{e}_i \tag*{(21)} δ e i = δ Θ × e i (21)
The vector δ Θ \delta\boldsymbol{\Theta} δ Θ is common in the laboratory frame, but its transverse components Θ i a \Theta_i^a Θ i a in the local tangent basis depend on the site. The deviations can therefore be written
δ e i = ∑ a Θ i a t i a (22) \delta\mathbf{e}_i = \sum_a \Theta_i^a \mathbf{t}_i^a \tag*{(22)} δ e i = a ∑ Θ i a t i a (22)
At a stationary reference, the corresponding global-rotation vectors are zero modes of the complete Hessian,
∑ j b H i j a b Θ j b = 0 (23) \sum_{jb} H_{ij}^{ab}\Theta_j^b = 0 \tag*{(23)} j b ∑ H ij ab Θ j b = 0 (23)
More generally, one can write τ A = ∂ E / ∂ θ A \tau_A = \partial E/\partial\theta_A τ A = ∂ E / ∂ θ A and H B A = ∂ 2 E / ( ∂ θ B ∂ θ A ) H_{BA} = \partial^2 E/(\partial\theta_B\partial\theta_A) H B A = ∂ 2 E / ( ∂ θ B ∂ θ A ) , with A = ( i , a ) A=(i,a) A = ( i , a ) , and let v A ( μ ) ( θ ) v_A^{(\mu)}(\theta) v A ( μ ) ( θ ) be the angular-coordinate vector field generated by a global laboratory spin rotation about axis μ \mu μ . Global rotational invariance implies ∑ A τ A v A ( μ ) = 0 \sum_A \tau_A v_A^{(\mu)}=0 ∑ A τ A v A ( μ ) = 0 ; differentiating gives the full Ward identity
∑ A H B A v A ( μ ) = − ∑ A τ A ∂ v A ( μ ) ∂ θ B . (24) \sum_A H_{BA}v_A^{(\mu)} = -\sum_A \tau_A\frac{\partial v_A^{(\mu)}}{\partial\theta_B}. \tag*{(24)} A ∑ H B A v A ( μ ) = − A ∑ τ A ∂ θ B ∂ v A ( μ ) . (24)
Only when the energy whose Hessian is being tested is stationary in these angular coordinates (τ A = 0 \tau_A=0 τ A = 0 ) does this reduce to Eq. (23) . In a collinear state with a common transverse basis it becomes the usual row-sum rule; in a noncollinear state the site-dependent global-rotation components must be retained.
The global-rotation identity can also be verified directly within PT, without invoking the local force theorem. For a spin-rotation-invariant one-electron Hamiltonian and a uniform spin rotation generated by a Hermitian operator S S S , let U ( λ ) = e − i λ S U(\lambda)=e^{-i\lambda S} U ( λ ) = e − iλ S and H ^ ( λ ) = U ( λ ) H ^ 0 U † ( λ ) \hat{H}(\lambda)=U(\lambda)\hat{H}_0U^\dagger(\lambda) H ^ ( λ ) = U ( λ ) H ^ 0 U † ( λ ) . The first and second derivatives of this Hamiltonian are V = − i [ S , H ^ 0 ] V=-i[S,\hat{H}_0] V = − i [ S , H ^ 0 ] and W = − [ S , [ S , H ^ 0 ] ] W=-[S,[S,\hat{H}_0]] W = − [ S , [ S , H ^ 0 ]] . In the eigenbasis of H ^ 0 \hat{H}_0 H ^ 0 , the two terms in the second-order energy derivative of an eigenstate obey
⟨ n ∣ W ∣ n ⟩ = − 2 ∑ m ( ε n − ε m ) ∣ S n m ∣ 2 , \langle n|W|n\rangle=-2\sum_m(\varepsilon_n-\varepsilon_m)|S_{nm}|^2, ⟨ n ∣ W ∣ n ⟩ = − 2 m ∑ ( ε n − ε m ) ∣ S nm ∣ 2 ,
2 ∑ m ≠ n ∣ V n m ∣ 2 ε n − ε m = + 2 ∑ m ≠ n ( ε n − ε m ) ∣ S n m ∣ 2 . (25) 2\sum_{m\ne n}\frac{|V_{nm}|^2}{\varepsilon_n-\varepsilon_m}
=+2\sum_{m\ne n}(\varepsilon_n-\varepsilon_m)|S_{nm}|^2. \tag*{(25)} 2 m = n ∑ ε n − ε m ∣ V nm ∣ 2 = + 2 m = n ∑ ( ε n − ε m ) ∣ S nm ∣ 2 . (25)
With a complete eigenstate basis and both rotational vertices, the terms cancel for each eigenstate and hence in the occupied-state sum. This verifies the uniform-orbit curvature directly in PT; the stationary Hessian zero-mode condition and its nonstationary torque correction follow separately from Eq. (24) . Nonlocal operators require their complete rotated vertices and contact terms, while relaxed or correlated responses require the corresponding induced electronic response.
The indices a , b a,b a , b refer to local transverse bases and therefore do not by themselves define basis-independent scalar quantities. Under a local change of transverse coordinates,
θ i ′ a = ∑ c ( R i ) a c θ i c (26) \theta_i^{\prime a}=\sum_c(R_i)_{ac}\theta_i^c \tag*{(26)} θ i ′ a = c ∑ ( R i ) a c θ i c (26)
The first-order vertices transform as V i a ′ = ∑ c ( R i ) a c V i c V_{ia}'=\sum_c(R_i)_{ac}V_{ic} V ia ′ = ∑ c ( R i ) a c V i c , and the direct second-order vertex transforms with one factor of R R R per angular index. Thus both terms in Eq. (14) give the covariance law
H i j ′ = R i H i j R j T (27) H_{ij}'=R_iH_{ij}R_j^T \tag*{(27)} H ij ′ = R i H ij R j T (27)
The quadratic energy in Eq. (13) is consequently independent of the chosen transverse frames. This is coordinate-independent covariance, not a symmetry under arbitrary site-dependent physical spin rotations.
Rotations of the magnetization
Unlike a prescribed field rotation, a moment rotation fixes the magnetization direction and determines the conjugate field needed to produce it; the constrained total-energy Hessian then follows from that response.
For a perturbation H ( λ ) = H 0 + λ V + λ 2 W / 2 + ⋯ H(\lambda)=H_0+\lambda V+\lambda^2W/2+\cdots H ( λ ) = H 0 + λV + λ 2 W /2 + ⋯ , ordinary nondegenerate many-electron PT gives the first-order wave-function correction ∣ 0 ( 1 ) ⟩ = ∑ n ≠ 0 ∣ n ⟩ ⟨ n ∣ V ∣ 0 ⟩ / ( E 0 − E n ) |0^{(1)}\rangle=\sum_{n\ne0}|n\rangle\langle n|V|0\rangle/(E_0-E_n) ∣ 0 ( 1 ) ⟩ = ∑ n = 0 ∣ n ⟩ ⟨ n ∣ V ∣0 ⟩ / ( E 0 − E n ) . Its magnetization response is
M α ( 1 ) ( r ) = 2 Re ∑ n ≠ 0 ⟨ 0 ∣ M ^ α ( r ) ∣ n ⟩ ⟨ n ∣ V ∣ 0 ⟩ E 0 − E n (28) M_\alpha^{(1)}(\mathbf{r})=2\operatorname{Re}\sum_{n\ne0}\frac{\langle0|\widehat{M}_\alpha(\mathbf{r})|n\rangle\langle n|V|0\rangle}{E_0-E_n} \tag*{(28)} M α ( 1 ) ( r ) = 2 Re n = 0 ∑ E 0 − E n ⟨ 0∣ M α ( r ) ∣ n ⟩ ⟨ n ∣ V ∣0 ⟩ (28)
Thus, the specified magnetization change is not itself a Hamiltonian vertex: it specifies the electronic response. For the same local directions e i \mathbf{e}_i e i and rotation coordinates θ i a \theta_i^a θ i a used above, one can consider a path M i ( θ ) = M i e i ( θ ) \mathbf{M}_i(\theta)=M_i\mathbf{e}_i(\theta) M i ( θ ) = M i e i ( θ ) with fixed local magnitudes M i M_i M i . Let t i a t_i^a t i a and s i , a b s_{i,ab} s i , ab be the first and second derivatives of e i \mathbf{e}_i e i in the same rotational parametrization.
∂ M i ∂ θ j a = δ i j M i t i a , ∂ 2 M i ∂ θ j a ∂ θ k b = δ i j δ i k M i s i , a b (29) \frac{\partial M_i}{\partial\theta_j^a}=\delta_{ij}M_it_i^a,\qquad
\frac{\partial^2M_i}{\partial\theta_j^a\partial\theta_k^b}
=\delta_{ij}\delta_{ik}M_is_{i,ab} \tag*{(29)} ∂ θ j a ∂ M i = δ ij M i t i a , ∂ θ j a ∂ θ k b ∂ 2 M i = δ ij δ ik M i s i , ab (29)
Following Ref. [14 ] , we introduce auxiliary constraining fields B i c \mathbf{B}_i^c B i c conjugate to the local moments, with V c = − ∑ i B i c ⋅ M ^ i V^c=-\sum_i\mathbf{B}_i^c\cdot\widehat{\mathbf{M}}_i V c = − ∑ i B i c ⋅ M i . For a constrained nonstationary reference state, the reference field is part of the unperturbed auxiliary Hamiltonian; PT concerns its variation. The first-order response is
δ M i α = ∑ j β X i α , j β δ B j β c , δ B c = X − 1 δ M (30) \delta M_{i\alpha}=\sum_{j\beta}X_{i\alpha,j\beta}\delta B^{\mathrm c}_{j\beta},\qquad\delta\mathbf{B}^{\mathrm c}=X^{-1}\delta\mathbf{M} \tag*{(30)} δ M i α = j β ∑ X i α , j β δ B j β c , δ B c = X − 1 δ M (30)
Here X X X is the static susceptibility of the specified constrained moments to their conjugate fields, with all electronic degrees of freedom allowed to relax in the chosen physical functional. The many-electron spectral representation follows from Eq. (28) with the auxiliary-field variation as the perturbation; an independent-particle bubble alone is not generally the fully self-consistent response. When local moments are projections of the full magnetization, X X X must first be constructed in that projected moment space and only then inverted. The inverse is understood on a locally invertible constrained subspace, or with a specified regularization for singular symmetry channels.
We denote the ground-state energy of the auxiliary Hamiltonian, including its field coupling, by E ( B c ) \mathcal{E}(\mathbf{B}^{\mathrm c}) E ( B c ) . The Hellmann–Feynman theorem gives ∂ E / ∂ B i α c = − M i α \partial\mathcal{E}/\partial B^{\mathrm c}_{i\alpha}=-M_{i\alpha} ∂ E / ∂ B i α c = − M i α . Since this Hamiltonian is linear in the varied auxiliary field, ordinary second-order PT gives the second derivative in Eq. (31) . The quadratic energy change itself contains an additional factor of 1 / 2 1/2 1/2 : Δ E ( 2 ) = δ 2 E / 2 \Delta\mathcal{E}^{(2)}=\delta^2\mathcal{E}/2 Δ E ( 2 ) = δ 2 E /2 .
δ 2 E = − ∑ i α , j β δ B i α c X i α , j β δ B j β c (31) \delta^2\mathcal{E}=-\sum_{i\alpha,j\beta}\delta B^{\mathrm c}_{i\alpha}X_{i\alpha,j\beta}\delta B^{\mathrm c}_{j\beta} \tag*{(31)} δ 2 E = − i α , j β ∑ δ B i α c X i α , j β δ B j β c (31)
Here E \mathcal{E} E is the auxiliary-field ground-state energy, whereas E E E is the constrained physical energy after removal of the auxiliary coupling. The Legendre transform of E \mathcal{E} E and the first differential of E E E are
E ( M ) = E ( B c ) + ∑ i B i c ⋅ M i , d E = ∑ i B i c ⋅ d M i (32) E(\mathbf{M})=\mathcal{E}(\mathbf{B}^{\mathrm c})+\sum_i\mathbf{B}^{\mathrm c}_i\cdot\mathbf{M}_i,\qquad dE=\sum_i\mathbf{B}^{\mathrm c}_i\cdot d\mathbf{M}_i \tag*{(32)} E ( M ) = E ( B c ) + i ∑ B i c ⋅ M i , d E = i ∑ B i c ⋅ d M i (32)
The Hessian of the constrained moment energy with respect to moment components is therefore the inverse response:
∂ 2 E ∂ M i α ∂ M j β = ( X − 1 ) i α , j β (33) \frac{\partial^2E}{\partial M_{i\alpha}\partial M_{j\beta}}=(X^{-1})_{i\alpha,j\beta} \tag*{(33)} ∂ M i α ∂ M j β ∂ 2 E = ( X − 1 ) i α , j β (33)
Applying the chain rule to the angular coordinates gives H M H^M H M , the Hessian of the relaxed, constrained moment energy:
H i a , j b M = M i M j t i a ⋅ ( X − 1 ) i j t j b + δ i j M i B i c ⋅ s i , a b (34) H^M_{ia,jb}=M_iM_jt_i^a\cdot(X^{-1})_{ij}t_j^b+\delta_{ij}\mathbf{M}_i\mathbf{B}^{\mathrm c}_i\cdot\mathbf{s}_{i,ab} \tag*{(34)} H ia , j b M = M i M j t i a ⋅ ( X − 1 ) ij t j b + δ ij M i B i c ⋅ s i , ab (34)
The first term in Eq. (34) is the inverse-response contribution for the fixed-length moment path; the second is the geometric chain-rule contribution from its second angular derivative. At an unconstrained stationary reference with B c = 0 \mathbf{B}^{\mathrm c}=0 B c = 0 , the geometric term vanishes. Angular stationarity alone does not imply this: a longitudinal constraining field can give zero torque while contributing to the geometric term. The inverse response must act in the same moment space as the Legendre transform.
In contrast, Eq. (14) differentiates the grand potential of a prescribed rotating one-electron Hamiltonian. Neither the Legendre transform nor ordinary second-order PT identifies these two energy curvatures without further assumptions about the functional, the angular variables, and the permitted electronic relaxation.
The effect of internal electronic relaxation can be illustrated by two magnetic angles θ 1 \theta_1 θ 1 and θ 2 \theta_2 θ 2 coupled to an internal transverse electronic response coordinate x x x . Near a stationary reference, one can consider the quadratic energy:
Δ E = ( A / 2 ) ( θ 1 − θ 2 ) 2 + g ( θ 1 − θ 2 ) x + ( C / 2 ) x 2 , C > 0. (35) \Delta E=(A/2)(\theta_1-\theta_2)^2+g(\theta_1-\theta_2)x+(C/2)x^2,\qquad C>0. \tag*{(35)} Δ E = ( A /2 ) ( θ 1 − θ 2 ) 2 + g ( θ 1 − θ 2 ) x + ( C /2 ) x 2 , C > 0. (35)
Holding x x x fixed gives the relative-angle curvature A A A . Minimizing the same energy over x x x gives the stationary electronic displacement and the reduced curvature:
x r e l = − ( g / C ) ( θ 1 − θ 2 ) , H r e l = A − g 2 / C . (36) x_{\mathrm{rel}}=-(g/C)(\theta_1-\theta_2),\qquad H_{\mathrm{rel}}=A-g^2/C. \tag*{(36)} x rel = − ( g / C ) ( θ 1 − θ 2 ) , H rel = A − g 2 / C . (36)
The uniform-rotation mode remains unchanged, while the relative-rotation mode softens. For A > 0 A>0 A > 0 and C > 0 C>0 C > 0 , stability against relative rotation and internal-coordinate fluctuations requires A C > g 2 AC>g^2 A C > g 2 ; the global-rotation zero mode remains. More generally, for magnetic coordinates θ \theta θ and stable internal coordinates x x x , with quadratic blocks A A A , G G G , and C C C , elimination of x x x yields the Schur complement:
H r e l = A − G C − 1 G T . (37) H_{\mathrm{rel}}=A-GC^{-1}G^T. \tag*{(37)} H rel = A − G C − 1 G T . (37)
The Schur complement compares unrelaxed and relaxed curvatures of the same functional and angular coordinates. Identifying its blocks with H B H^B H B and H M H^M H M would additionally require matching the magnetic variables, energy functional, and electronic relaxation prescription.
Ref. [22 ] further developed the inverse-susceptibility approach to magnetic exchange. Equations (33) and (34) link it to the moment-component and angular Hessians of the constrained moments, respectively.
Without spin–orbit coupling (SOC), simultaneous global rotation of all moments leaves the energy unchanged. At a stationary configuration, the associated angular directions are zero modes of H M H^M H M . At a nonstationary constrained configuration, rotational invariance instead yields a torque identity and its derivative, which also contains the first energy derivatives; the ordinary angular Hessian need not have global-rotation zero modes. For a magnetically ordered reference state, an infinitesimal symmetry-breaking field can select the magnetic orientation when applying perturbation theory. We evaluate the inverse static response before removing this regularization, so its Goldstone-channel eigenvalue tends to zero at stationarity.
Thus, prescribed-field and constrained-moment rotations define different energy curvatures. The many-electron spectral identities provide the formal basis; the self-consistent Kohn–Sham response supplies the implementation within density-functional theory (DFT). Neither Hessian alone establishes exact individual Heisenberg exchange parameters.
Magnetic torques and constraining fields
The first-order rotational vertex V i a a V_{ia}^{a} V ia a introduced above gives the field-rotation torque directly from electronic PT:
τ i a B = − ∂ E B ∂ θ i a = − ⟨ V i a ⟩ (38) \tau_{ia}^{B}=-\frac{\partial E^{B}}{\partial\theta_{i}^{a}}=-\langle V_{i}^{a}\rangle \tag*{(38)} τ ia B = − ∂ θ i a ∂ E B = − ⟨ V i a ⟩ (38)
(38) gives the first derivative of the prescribed-field energy. It need not equal the torque derived from the fully relaxed moment energy.
For the magnetization, the constraining field introduced in (30) is conjugate to the constrained local moment. The Legendre transform in (32) gives d E = ∑ i B i c ⋅ d M i dE=\sum_i\mathbf{B}_i^c\cdot d\mathbf{M}_i d E = ∑ i B i c ⋅ d M i . Under an infinitesimal rigid rotation δ M i = δ θ i × M i \delta\mathbf{M}_i=\delta\boldsymbol{\theta}_i\times\mathbf{M}_i δ M i = δ θ i × M i , defining δ E = − ∑ i τ i M ⋅ δ θ i \delta E=-\sum_i\tau_i^M\cdot\delta\boldsymbol{\theta}_i δ E = − ∑ i τ i M ⋅ δ θ i yields
τ i M = − M i × B i c ; τ i a M = − M i ⋅ B i c × t i a (39) \boldsymbol{\tau}_i^M=-\mathbf{M}_i\times\mathbf{B}_i^c;\qquad\tau_{ia}^M=-\mathbf{M}_i\cdot\mathbf{B}_i^c\times\mathbf{t}_i^a \tag*{(39)} τ i M = − M i × B i c ; τ ia M = − M i ⋅ B i c × t i a (39)
H i a , j b M = − ∂ τ i a M ∂ θ j b (40) H_{ia,jb}^{M}=-\frac{\partial\tau_{ia}^{M}}{\partial\theta_j^b} \tag*{(40)} H ia , j b M = − ∂ θ j b ∂ τ ia M (40)
The auxiliary field holds the constrained moment against the energy-gradient torque. Only its transverse component contributes. Here τ i a M \tau_{ia}^{M} τ ia M is the angular-coordinate component of the torque, obtained by projection onto the reference rotation axis u i a \mathbf{u}_i^a u i a ; it is not an arbitrary Cartesian component. Differentiating the angular-coordinate torque with respect to the same rotation coordinates yields (40) , which includes both the inverse-response and geometric terms in (34) . At an unconstrained stationary configuration, the rotational torques vanish.
In the local spin-density approximation (LSDA) without spin–orbit coupling, the local exchange-correlation field is parallel to the magnetization and has no local torque. The absence of a local torque does not remove intersite interactions: the spin-continuity equation balances spin-current transport against field torques, including the constraining-field torque. This balance and the derivative of the constrained energy describe the same physics and must not be counted twice.
The adiabatic approximation in spin dynamics [23 , 24 ] separates the slow collective rotation of the magnetic moments from the electronic response. In a two-time response function, t 2 t_2 t 2 is the time of the perturbation and t 1 t_1 t 1 is the time at which the induced response is evaluated. The center time t ˉ = ( t 1 + t 2 ) / 2 \bar{t}=(t_1+t_2)/2 t ˉ = ( t 1 + t 2 ) /2 labels the evolving magnetic configuration, while the relative time Δ t = t 1 − t 2 \Delta t=t_1-t_2 Δ t = t 1 − t 2 describes the electronic response interval. Writing M i ( t ) = M i ( t ) e i ( t ) \mathbf{M}_i(t)=M_i(t)\mathbf{e}_i(t) M i ( t ) = M i ( t ) e i ( t ) , with ∣ e i ( t ) ∣ = 1 |\mathbf{e}_i(t)|=1 ∣ e i ( t ) ∣ = 1 , the separation requires the direction to change little over relative times that contribute appreciably to the response:
e i ( t ˉ + Δ t / 2 ) = e i ( t ˉ ) + Δ t 2 e ˙ i ( t ˉ ) + O ( Δ t 2 ) , ∣ Δ t e ˙ i ( t ˉ ) ∣ ≪ 1. (41) \mathbf{e}_i(\bar{t}+\Delta t/2)=\mathbf{e}_i(\bar{t})+\frac{\Delta t}{2}\dot{\mathbf{e}}_i(\bar{t})+O(\Delta t^2),\qquad|\Delta t\dot{\mathbf{e}}_i(\bar{t})|\ll1. \tag*{(41)} e i ( t ˉ + Δ t /2 ) = e i ( t ˉ ) + 2 Δ t e ˙ i ( t ˉ ) + O ( Δ t 2 ) , ∣Δ t e ˙ i ( t ˉ ) ∣ ≪ 1. (41)
If longitudinal fluctuations can also be neglected, M i ( t ˉ ) ≃ M i ( 0 ) M_i(\bar{t})\simeq M_i^{(0)} M i ( t ˉ ) ≃ M i ( 0 ) and hence M i ( t ˉ ) ≃ M i ( 0 ) e i ( t ˉ ) \mathbf{M}_i(\bar{t})\simeq M_i^{(0)}\mathbf{e}_i(\bar{t}) M i ( t ˉ ) ≃ M i ( 0 ) e i ( t ˉ ) . The rigidity of the moment amplitude is an additional assumption, not a consequence of time-scale separation. At leading adiabatic order, one can solve stationary DFT–Pauli equations for successive configurations stabilized by constraining fields. (39) then gives the instantaneous static torque, whereas precession requires a dynamical coefficient: the static inverse susceptibility determines the moment Hessian, and its frequency-linear term determines the leading low-frequency motion. This coefficient need not equal that of the prescribed-field response in Sec. II.E.
Near a stationary state, small oscillations can alternatively be obtained directly from the dynamical transverse susceptibility, without prescribing successive moment directions. The low-frequency expansion must be regular; electronic memory and damping effects can limit the adiabatic description.
Time-dependent rotational PT
Extending the time-dependent treatments of Refs. [9 , 10 , 19 ] , we now allow the exchange-field rotations to oscillate in time. Their first-order perturbation is
V ( 1 ) ( t ) = ∑ i a θ i a ( t ) V i a , θ i a ( t ) = θ i a ( ω ) e − i ω t (42) V^{(1)}(t)=\sum_{ia}\theta_i^a(t)V_i^a,\qquad\theta_i^a(t)=\theta_i^a(\omega)e^{-i\omega t} \tag*{(42)} V ( 1 ) ( t ) = ia ∑ θ i a ( t ) V i a , θ i a ( t ) = θ i a ( ω ) e − iω t (42)
The response then follows from the same first-order vertices, with the full-Hessian normalization of (14) :
K i a , j b ( ω ) = ⟨ W i a , j b ⟩ + Π i a , j b ( ω ) (43) \mathcal{K}_{ia,jb}(\omega)=\langle W_{ia,jb}\rangle+\Pi_{ia,jb}(\omega) \tag*{(43)} K ia , j b ( ω ) = ⟨ W ia , j b ⟩ + Π ia , j b ( ω ) (43)
Π i a , j b ( ω ) = ∑ n m ( f n − f m ) ( V i a ) n m ( V j b ) m n ϵ n − ϵ m + ℏ ω + i 0 + (44) \Pi_{ia,jb}(\omega)=\sum_{nm}\frac{(f_n-f_m)(V_i^a)_{nm}(V_j^b)_{mn}}{\epsilon_n-\epsilon_m+\hbar\omega+i0^+} \tag*{(44)} Π ia , j b ( ω ) = nm ∑ ϵ n − ϵ m + ℏ ω + i 0 + ( f n − f m ) ( V i a ) nm ( V j b ) mn (44)
The occupations and state normalization are the same as in the static calculation. At finite frequency, the two vertex orderings differ; the regular zero-frequency limit recovers the static Hessian.
K i a , j b ( 0 ) = H i a , j b (strictly static limit) (45) \mathcal{K}_{ia,jb}(0)=H_{ia,jb}\qquad\text{(strictly static limit)} \tag*{(45)} K ia , j b ( 0 ) = H ia , j b (strictly static limit) (45)
Equation (45) gives the prescribed-field Hessian in this limit. In a metal, the zero-frequency and zero-wave-vector limits can differ, and a pole of the prescribed-field response need not be a collective spin-wave mode.
When the electronic response is regular at low frequency, it admits the expansion
K ( q , ω ) = H ( q ) + ℏ ω Λ ( q ) + O ( ( ℏ ω ) 2 ) (46) \mathcal{K}(\mathbf{q},\omega)=H(\mathbf{q})+\hbar\omega\Lambda(\mathbf{q})+O\left((\hbar\omega)^2\right) \tag*{(46)} K ( q , ω ) = H ( q ) + ℏ ω Λ ( q ) + O ( ( ℏ ω ) 2 ) (46)
Electron–hole excitations can damp SWs and spoil a regular low-frequency expansion. Collective modes must then be found from the full self-consistent transverse susceptibility.
Frequency-dependent intersite interactions
The off-site elements of Eq. (43) provide a frequency-dependent extension of the static pair curvature, without assuming an instantaneous spin Hamiltonian. For i ≠ j i\ne j i = j , we define the exchange-like transverse tensor, in the same full-angular-Hessian and pair-counting convention as Eq. (18) , by
J i j a b ( ω ) ≡ − K i a , j b R ( ω ) , i ≠ j . (47) \mathcal{J}_{ij}^{ab}(\omega)\equiv-\mathcal{K}_{ia,jb}^{R}(\omega),\qquad i\ne j. \tag*{(47)} J ij ab ( ω ) ≡ − K ia , j b R ( ω ) , i = j . (47)
The superscript R R R denotes retarded response. For a collinear reference with consistently chosen local transverse frames, if the static prescribed-field curvature admits an isotropic bilinear representation, then J i j a b ( 0 ) = J i j B δ a b \mathcal{J}_{ij}^{ab}(0)=J_{ij}^{B}\delta_{ab} J ij ab ( 0 ) = J ij B δ ab . Here J i j B J_{ij}^{B} J ij B is the prescribed-field exchange-like coefficient and is not, in general, the constrained-moment coefficient J i j M J_{ij}^{M} J ij M . Outside this scalar limit, the response is a tensor. Its frequency dependence does not define a frequency-dependent Heisenberg Hamiltonian.
Writing z = ℏ ω z=\hbar\omega z = ℏ ω , a regular low-frequency limit of Eq. (44) yields
K i a , j b R ( z ) = H i a , j b + z Λ i a , j b R + o ( z ) , (48) \mathcal{K}_{ia,jb}^{R}(z)=H_{ia,jb}+z\Lambda_{ia,jb}^{R}+o(z), \tag*{(48)} K ia , j b R ( z ) = H ia , j b + z Λ ia , j b R + o ( z ) , (48)
Λ i a , j b R = ∂ Π i a , j b R ( z ) ∂ z ∣ z = 0 = − lim z → 0 ∑ n m ( f n − f m ) ( V i a a ) n m ( V j b ) m n ( ϵ n − ϵ m + z + i 0 + ) 2 . (49) \Lambda_{ia,jb}^{R}=\left.\frac{\partial\Pi_{ia,jb}^{R}(z)}{\partial z}\right|_{z=0}
=-\lim_{z\to0}\sum_{nm}\frac{(f_n-f_m)(V_{ia}^{a})_{nm}(V_{j}^{b})_{mn}}{(\epsilon_n-\epsilon_m+z+i0^+)^2}. \tag*{(49)} Λ ia , j b R = ∂ z ∂ Π ia , j b R ( z ) z = 0 = − z → 0 lim nm ∑ ( ϵ n − ϵ m + z + i 0 + ) 2 ( f n − f m ) ( V ia a ) nm ( V j b ) mn . (49)
For prescribed rotations of a static one-electron Hamiltonian, the second-order vertex W i a , j b W_{ia,jb} W ia , j b has no explicit frequency dependence. When the frequency derivative of the response is well defined and dissipation is absent, the additional energy denominator in Eq. (48) gives a spatial dependence different from that of the static exchange. In metals, however, the low-frequency retarded response may be nonanalytic, and the zero-frequency limit cannot generally be interchanged with the electronic-state summation.
For the collinear case considered here, the transverse response takes the form
K i j R ( z ) = − J i j B I + i z g i j ϵ + i z η i j I + o ( z ) , i ≠ j , (50) \mathcal{K}_{ij}^{R}(z)=-J_{ij}^{B}I+izg_{ij}\epsilon+iz\eta_{ij}I+o(z),\qquad i\ne j, \tag*{(50)} K ij R ( z ) = − J ij B I + i z g ij ϵ + i z η ij I + o ( z ) , i = j , (50)
ϵ = ( 0 1 − 1 0 ) . \epsilon=\begin{pmatrix}0&1\\-1&0\end{pmatrix}. ϵ = ( 0 − 1 1 0 ) .
The coefficient g i j g_{ij} g ij describes precession, while η i j \eta_{ij} η ij describes the frequency-linear dissipative response when this expansion exists. These terms have different physical meanings even though both contain i ω i\omega iω in Cartesian notation. The full absorption matrix, not an individual off-site element, must satisfy the passivity condition − I m H K i j R ( ω ) ≥ 0 -\operatorname{Im_H}\mathcal{K}_{ij}^{R}(\omega)\geq0 − I m H K ij R ( ω ) ≥ 0 for ω > 0 \omega>0 ω > 0 . In the present sign convention this gives − η ≥ 0 -\eta\geq0 − η ≥ 0 for the complete matrix. A particular off-site η i j \eta_{ij} η ij can have either sign.
To fix the circular labels, we take e ± = ( 1 , ± i ) T / 2 \mathbf{e}_{\pm}=(1,\pm i)^{T}/\sqrt{2} e ± = ( 1 , ± i ) T / 2 , so that ϵ e ± = ± i e ± \epsilon\mathbf{e}_{\pm}=\pm i\mathbf{e}_{\pm} ϵ e ± = ± i e ± . The eigenvalues of Eq. (50) and the corresponding exchange-like projections are then
K i j , ± R ( z ) = − J i j B + z g i j ± i z η i j + o ( z ) , \mathcal{K}_{ij,\pm}^{R}(z)=-J_{ij}^{B}+zg_{ij}\pm iz\eta_{ij}+o(z), K ij , ± R ( z ) = − J ij B + z g ij ± i z η ij + o ( z ) ,
J i j , ± ( ω ) = − K i j , ± R ( ω ) = J i j B ± ℏ ω g i j − i ℏ ω η i j + o ( ω ) . (51) J_{ij,\pm}(\omega)=-\mathcal{K}_{ij,\pm}^{R}(\omega)
=J_{ij}^{B}\pm\hbar\omega g_{ij}-i\hbar\omega\eta_{ij}+o(\omega). \tag*{(51)} J ij , ± ( ω ) = − K ij , ± R ( ω ) = J ij B ± ℏ ω g ij − i ℏ ω η ij + o ( ω ) . (51)
These ± \pm ± labels specify the Cartesian vectors, not independently the positive-frequency magnon branch; the latter must be identified using the pole and Fourier conventions of the dynamical response.
An elementary normalization test is the one-electron Hamiltonian H = − b σ z H=-b\sigma_z H = − b σ z (b > 0 b>0 b > 0 ), with its lower spin state occupied. For V x = − b σ x V^x=-b\sigma_x V x = − b σ x and V y = − b σ y V^y=-b\sigma_y V y = − b σ y , the contact term ⟨ W x x ⟩ = 0 \langle W_{xx}\rangle=0 ⟨ W xx ⟩ = 0 cancels the zero-frequency transverse bubble Π x x R ( 0 ) = − b \Pi_{xx}^{R}(0)=-b Π xx R ( 0 ) = − b , while ∂ z Π x y R ( 0 ) = − i / 2 \partial_z\Pi_{xy}^{R}(0)=-i/2 ∂ z Π x y R ( 0 ) = − i /2 . Hence K x x R ( 0 ) = 0 \mathcal{K}_{xx}^{R}(0)=0 K xx R ( 0 ) = 0 and Λ x y R = − i / 2 \Lambda_{xy}^{R}=-i/2 Λ x y R = − i /2 , with no low-frequency absorption below the spin-flip gap. This checks both the static contact-term cancellation and the reactive frequency-linear sign.
The uniform reactive coefficient provides a normalization check directly from Eq. (44) . For a spin-rotation-invariant collinear reference with local static magnetic potential − b σ z -b\sigma_z − b σ z , the complete uniform vertices are V x = − b σ x V^x=-b\sigma_x V x = − b σ x and V y = − b σ y V^y=-b\sigma_y V y = − b σ y , while H ↓ − H ↑ = 2 b H^\downarrow-H^\uparrow=2b H ↓ − H ↑ = 2 b . Between spin-up and spin-down eigenstates,
⟨ n ↑ ∣ b ∣ m ↓ ⟩ = ϵ m ↓ − ϵ n ↑ 2 ⟨ n ↑ ∣ m ↓ ⟩ . (52) \langle n^{\uparrow}|b|m^{\downarrow}\rangle=\frac{\epsilon_m^{\downarrow}-\epsilon_n^{\uparrow}}{2}\langle n^{\uparrow}|m^{\downarrow}\rangle. \tag*{(52)} ⟨ n ↑ ∣ b ∣ m ↓ ⟩ = 2 ϵ m ↓ − ϵ n ↑ ⟨ n ↑ ∣ m ↓ ⟩ . (52)
The two spin-flip orderings in Eq. (44) , followed by completeness of both spin eigenstate sets, give
∑ i j Λ i x , j y R = − i 2 ( N ↑ − N ↓ ) = − i M 2 , (53) \sum_{ij}\Lambda_{ix,jy}^{R}=-\frac{i}{2}(N_{\uparrow}-N_{\downarrow})=-\frac{iM}{2}, \tag*{(53)} ij ∑ Λ i x , j y R = − 2 i ( N ↑ − N ↓ ) = − 2 i M , (53)
∑ i j Λ i y , j x R = + i M 2 , ∑ i j Λ i x , j x R = 0. \sum_{ij}\Lambda_{iy,jx}^{R}=+\frac{iM}{2},\qquad\sum_{ij}\Lambda_{ix,jx}^{R}=0. ij ∑ Λ i y , j x R = + 2 i M , ij ∑ Λ i x , j x R = 0.
Thus, with g i j = Λ i x , j y R / i g_{ij}=\Lambda_{ix,jy}^{R}/i g ij = Λ i x , j y R / i as in Eq. (50) , the full site sum satisfies ∑ i j g i j = − M / 2 \sum_{ij}g_{ij}=-M/2 ∑ ij g ij = − M /2 . For a translationally invariant ferromagnet with one magnetic site per cell, this becomes
g o n + ∑ n ≥ 1 z n g n = − m 2 , (54) g_{\mathrm{on}}+\sum_{n\ge1}z_n g_n=-\frac{m}{2}, \tag*{(54)} g on + n ≥ 1 ∑ z n g n = − 2 m , (54)
where z n z_n z n is the number of sites in coordination shell n n n , and m m m is the spin polarization per cell. The sum includes the on-site coefficient, although the exchange-like definition in Eq. (47) applies only off site. It fixes the total precessional normalization, not any individual shell coefficient. In particular, g n g_n g n need not be proportional to the corresponding static J n J_n J n .
The uniform sum rule in Eqs. (52) –(54) assumes a local, static spin splitting. Its derivation uses the relation between the splitting and the spin-flip energy differences. For a nonlocal magnetic potential, the full covariant vertices and contact terms must be included. A frequency-dependent self-energy also requires its induced response and the corresponding Ward identity. The local formula does not establish the corresponding result in either case.
MAGNETIC INTERACTIONS FROM ROTATIONAL PT
Microscopic noncollinear pair interaction
We now calculate how the energy changes when two moments are rotated in a fixed magnetic state. The reference Green functions describe that state, which may be noncollinear. Only the added rotations are small. For a local magnetic potential h i = − b i e i ⋅ σ h_i=-b_i\mathbf{e}_i\cdot\boldsymbol{\sigma} h i = − b i e i ⋅ σ , with b i = Δ i / 2 b_i=\Delta_i/2 b i = Δ i /2 , the first-order rotational vertex is V i = − b i t i a ⋅ σ V_i=-b_i\mathbf{t}_i^a\cdot\boldsymbol{\sigma} V i = − b i t i a ⋅ σ , where t i a = u i a × e i \mathbf{t}_i^a=\mathbf{u}_i^a\times\mathbf{e}_i t i a = u i a × e i is a local tangent vector.
For rotations at two distinct sites, the electronic energy responds through the two-vertex term:
H i j a b = − 1 π Im ∫ E F d ε Tr [ V i a G i j ( ε ) V j b G j i ( ε ) ] (55) H_{ij}^{ab}=-\frac{1}{\pi}\operatorname{Im}\int^{E_F}d\varepsilon\operatorname{Tr}\left[V_i^aG_{ij}(\varepsilon)V_j^bG_{ji}(\varepsilon)\right] \tag*{(55)} H ij ab = − π 1 Im ∫ E F d ε Tr [ V i a G ij ( ε ) V j b G j i ( ε ) ] (55)
Equation (54) gives the off-site mixed curvature. The full rotational Hessian also includes the on-site second-order vertex of Eq. (11) , as incorporated in Eq. (14) ; this local term does not contribute to Eq. (55) for distinct sites. Comparing the independent transverse channels of Eq. (55) with the isotropic Heisenberg-model Hessian, Eq. (20) , gives candidate coefficients
J i j ( a b ) = 1 π Im ∫ E F d ε Tr [ V i a G i j V j b G j i ] t i a ⋅ t j b (56) J_{ij}^{(ab)}=\frac{1}{\pi}\frac{\operatorname{Im}\int^{E_F}d\varepsilon\operatorname{Tr}\left[V_i^aG_{ij}V_j^bG_{ji}\right]}{\mathbf{t}_i^a\cdot\mathbf{t}_j^b} \tag*{(56)} J ij ( ab ) = π 1 t i a ⋅ t j b Im ∫ E F d ε Tr [ V i a G ij V j b G j i ] (56)
Equation (55) applies only when the geometrical denominator is nonzero. Unequal values from independent transverse channels rule out a scalar Heisenberg representation of this local curvature.
The Hessian-to-exchange conversion must retain the energy functional being differentiated: below, H i j H_{ij} H ij denotes either the prescribed-field block H i j B H_{ij}^{B} H ij B calculated from Eq. (55) , or the corresponding relaxed constrained-moment block H i j M H_{ij}^{M} H ij M , with these alternatives interpreted separately. For practical extraction, choose orthonormal local rotation axes u i a \mathbf{u}_i^a u i a perpendicular to each reference direction e i \mathbf{e}_i e i , and form the transverse displacement vectors t i a = u i a × e i \mathbf{t}_i^a=\mathbf{u}_i^a\times\mathbf{e}_i t i a = u i a × e i . The 2 × 2 2\times2 2 × 2 geometrical matrix P i j a b = t i a ⋅ t j b P_{ij}^{ab}=\mathbf{t}_i^a\cdot\mathbf{t}_j^b P ij ab = t i a ⋅ t j b must use exactly the same local coordinates as the calculated off-site Hessian. Equation (55) gives an exact componentwise exchange only when the Hessian has the Heisenberg form. In general, a single coefficient can instead be defined by minimizing ∑ a b ( H i j a b + J P i j a b ) 2 \sum_{ab}(H_{ij}^{ab}+J P_{ij}^{ab})^2 ∑ ab ( H ij ab + J P ij ab ) 2 , giving
J i j p r o j = − ∑ a b H i j a b P i j a b 1 + ( e i ⋅ e j ) 2 , (57) J_{ij}^{\mathrm{proj}}=-\frac{\sum_{ab}H_{ij}^{ab}P_{ij}^{ab}}{1+(\mathbf{e}_i\cdot\mathbf{e}_j)^2}, \tag*{(57)} J ij proj = − 1 + ( e i ⋅ e j ) 2 ∑ ab H ij ab P ij ab , (57)
since ∑ a b ( P i j a b ) 2 = 1 + ( e i ⋅ e j ) 2 \sum_{ab}(P_{ij}^{ab})^2=1+(\mathbf{e}_i\cdot\mathbf{e}_j)^2 ∑ ab ( P ij ab ) 2 = 1 + ( e i ⋅ e j ) 2 . This expression is invariant under consistent changes of the local transverse bases and avoids division by vanishing individual geometrical factors. The Hessian must be the full mixed second derivative of the energy, with the pair-counting and rotation conventions of Eqs. (5) –(20) .
For illustration, let c = cos ϕ c=\cos\phi c = cos ϕ for two reference moments separated by angle ϕ \phi ϕ . In tangent frames where the geometrical matrix is diagonal, consider a general calculated pair block:
P i j = ( c 0 0 1 ) , H i j = ( h 11 h 12 h 21 h 22 ) . (58) P_{ij}=\begin{pmatrix}c&0\\0&1\end{pmatrix},\qquad H_{ij}=\begin{pmatrix}h_{11}&h_{12}\\h_{21}&h_{22}\end{pmatrix}. \tag*{(58)} P ij = ( c 0 0 1 ) , H ij = ( h 11 h 21 h 12 h 22 ) . (58)
Equation (56) gives
J i j p r o j = − h 11 c + h 22 1 + c 2 . (59) J_{ij}^{\mathrm{proj}}=-\frac{h_{11}c+h_{22}}{1+c^2}. \tag*{(59)} J ij proj = − 1 + c 2 h 11 c + h 22 . (59)
The difference between the calculated block and its reconstructed Heisenberg form is
Δ H i j = H i j + J i j p r o j P i j = ( d i j h 12 h 21 − c d i j ) , d i j = h 11 − c h 22 1 + c 2 . (60) \Delta H_{ij}=H_{ij}+J_{ij}^{\mathrm{proj}}P_{ij}
=\begin{pmatrix}d_{ij}&h_{12}\\h_{21}&-c\,d_{ij}\end{pmatrix},\qquad d_{ij}=\frac{h_{11}-c h_{22}}{1+c^2}. \tag*{(60)} Δ H ij = H ij + J ij proj P ij = ( d ij h 21 h 12 − c d ij ) , d ij = 1 + c 2 h 11 − c h 22 . (60)
Thus the pair block is exactly Heisenberg-like if and only if h 11 = h 22 c h_{11}=h_{22}c h 11 = h 22 c and h 12 = h 21 = 0 h_{12}=h_{21}=0 h 12 = h 21 = 0 . The off-diagonal components do not contribute to the projected scalar in these particular frames, but remain in the residual and cannot be discarded. For parallel moments (c = 1 c=1 c = 1 ), J i j p r o j = − ( h 11 + h 22 ) / 2 J_{ij}^{\mathrm{proj}}=-(h_{11}+h_{22})/2 J ij proj = − ( h 11 + h 22 ) /2 ; exact compatibility additionally requires h 11 = h 22 h_{11}=h_{22} h 11 = h 22 and h 12 = h 21 = 0 h_{12}=h_{21}=0 h 12 = h 21 = 0 . For perpendicular moments (c = 0 c=0 c = 0 ), J i j p r o j = − h 22 J_{ij}^{\mathrm{proj}}=-h_{22} J ij proj = − h 22 , while h 11 h_{11} h 11 , h 12 h_{12} h 12 , and h 21 h_{21} h 21 remain outside the scalar Heisenberg representation. The projected coefficient J i j p r o j J_{ij}^{\mathrm{proj}} J ij proj is a least-squares Heisenberg representation of the chosen off-site Hessian block, not a unique microscopic pair interaction when non-Heisenberg terms are present. It is obtained from H i j H_{ij} H ij and the Heisenberg matrix P i j P_{ij} P ij , expressed in the same transverse coordinates. The residual Δ H i j = H i j + J i j p r o j P i j \Delta H_{ij}=H_{ij}+J_{ij}^{\mathrm{proj}}P_{ij} Δ H ij = H ij + J ij proj P ij measures the deviation from the Heisenberg model. For H i j ≠ 0 H_{ij}\ne0 H ij = 0 , the relative residual r i j = r_{ij}= r ij = ∥ Δ H i j ∥ F / ∥ H i j ∥ F \lVert\Delta H_{ij}\rVert_F/\lVert H_{ij}\rVert_F ∥ Δ H ij ∥ F / ∥ H ij ∥ F lies between zero and one, with r i j 2 r_{ij}^{2} r ij 2 giving the fraction of the squared Hessian norm not reproduced by the model. There is no universal cutoff for r i j r_{ij} r ij , and the absolute residual is more useful for small H i j H_{ij} H ij . Even a zero residual establishes agreement only for the off-site Hessian block, not for on-site curvature, magnetic torques, or finite-angle energies. The projected exchange parameter may therefore differ from one obtained by fitting finite-angle energies or from another magnetic configuration. Finally, projecting the field-rotation Hessian H B H^{B} H B gives J i j B , p r o j J_{ij}^{B,\mathrm{proj}} J ij B , proj , whereas the constrained-moment parameter J i j M , p r o j J_{ij}^{M,\mathrm{proj}} J ij M , proj must be obtained from H M H^{M} H M .
In collinear calculations, the sign of individual Hessian components depends on whether local rotation axes or common transverse displacement axes are used; the geometrical projection accounts for this convention.
A noncollinear pair can have a well-defined rotational response without having a unique bilinear exchange tensor. The following example shows why.
One can consider the mechanism discussed by dos Santos Dias et al. [25 ] . For two unit-length spin directions, the isotropic bilinear–biquadratic energy is
E i j = − J i j e i ⋅ e j + Q i j ( e i ⋅ e j ) 2 . (61) E_{ij}=-J_{ij}\,\mathbf{e}_i\cdot\mathbf{e}_j+Q_{ij}\left(\mathbf{e}_i\cdot\mathbf{e}_j\right)^2 . \tag*{(61)} E ij = − J ij e i ⋅ e j + Q ij ( e i ⋅ e j ) 2 . (61)
The first term is bilinear exchange; the second is biquadratic exchange. Both depend only on the angle between the spins. The model has no spin–orbit coupling or DMI.
This model also provides an explicit test of the projection in Eq. (57) , rather than merely an example of an apparent chiral coupling.
We choose a reference state in which the two spins make an angle ϕ \phi ϕ . This angle need not minimize the energy of the isolated pair: neighboring spins or external constraints may hold it fixed.
For infinitesimal transverse variations, e i ⋅ δ e i = e j ⋅ δ e j = 0 \mathbf{e}_i\cdot\delta\mathbf{e}_i=\mathbf{e}_j\cdot\delta\mathbf{e}_j=0 e i ⋅ δ e i = e j ⋅ δ e j = 0 , the mixed second variation (one variation at each site) is
δ i δ j E i j = ( − J i j + 2 Q i j cos ϕ ) δ e i ⋅ δ e j + 2 Q i j ( δ e i ⋅ e j ) ( e i ⋅ δ e j ) . (62) \delta_i\delta_j E_{ij}=\left(-J_{ij}+2Q_{ij}\cos\phi\right)\delta\mathbf{e}_i\cdot\delta\mathbf{e}_j+2Q_{ij}\left(\delta\mathbf{e}_i\cdot\mathbf{e}_j\right)\left(\mathbf{e}_i\cdot\delta\mathbf{e}_j\right). \tag*{(62)} δ i δ j E ij = ( − J ij + 2 Q ij cos ϕ ) δ e i ⋅ δ e j + 2 Q ij ( δ e i ⋅ e j ) ( e i ⋅ δ e j ) . (62)
The standard vector identity gives, for these transverse variations,
( e i × e j ) ⋅ ( δ e i × δ e j ) = − ( e i ⋅ δ e j ) ( e j ⋅ δ e i ) . (63) \left(\mathbf{e}_i\times\mathbf{e}_j\right)\cdot\left(\delta\mathbf{e}_i\times\delta\mathbf{e}_j\right)=-\left(\mathbf{e}_i\cdot\delta\mathbf{e}_j\right)\left(\mathbf{e}_j\cdot\delta\mathbf{e}_i\right). \tag*{(63)} ( e i × e j ) ⋅ ( δ e i × δ e j ) = − ( e i ⋅ δ e j ) ( e j ⋅ δ e i ) . (63)
Consequently,
δ i δ j E i j = ( − J i j + 2 Q i j cos ϕ ) δ e i ⋅ δ e j − 2 Q i j ( e i × e j ) ⋅ ( δ e i × δ e j ) . (64) \delta_i\delta_j E_{ij}=\left(-J_{ij}+2Q_{ij}\cos\phi\right)\delta\mathbf{e}_i\cdot\delta\mathbf{e}_j-2Q_{ij}\left(\mathbf{e}_i\times\mathbf{e}_j\right)\cdot\left(\delta\mathbf{e}_i\times\delta\mathbf{e}_j\right). \tag*{(64)} δ i δ j E ij = ( − J ij + 2 Q ij cos ϕ ) δ e i ⋅ δ e j − 2 Q ij ( e i × e j ) ⋅ ( δ e i × δ e j ) . (64)
The same Hessian can be analyzed without assigning its individual terms to separate interactions. In the transverse frames of the preceding projection example, with c = cos ϕ c=\cos\phi c = cos ϕ and P i j = diag ( c , 1 ) P_{ij}=\operatorname{diag}(c,1) P ij = diag ( c , 1 ) , its full mixed block is
H i j = ( − J i j c + 2 Q i j ( 2 c 2 − 1 ) 0 0 − J i j + 2 Q i j c ) . (65) H_{ij}=
\begin{pmatrix}
-J_{ij}c+2Q_{ij}(2c^2-1) & 0\\
0 & -J_{ij}+2Q_{ij}c
\end{pmatrix}. \tag*{(65)} H ij = ( − J ij c + 2 Q ij ( 2 c 2 − 1 ) 0 0 − J ij + 2 Q ij c ) . (65)
The full-block Heisenberg projection and its remainder are
J i j p r o j = J i j − 4 Q i j c 3 1 + c 2 , (66) J_{ij}^{\mathrm{proj}}=J_{ij}-\frac{4Q_{ij}c^3}{1+c^2}, \tag*{(66)} J ij proj = J ij − 1 + c 2 4 Q ij c 3 , (66)
Δ H i j = − 2 Q i j ( 1 − c 2 ) 1 + c 2 ( 1 0 0 − c ) . (67) \Delta H_{ij}=-\frac{2Q_{ij}(1-c^2)}{1+c^2}
\begin{pmatrix}
1 & 0\\
0 & -c
\end{pmatrix}. \tag*{(67)} Δ H ij = − 1 + c 2 2 Q ij ( 1 − c 2 ) ( 1 0 0 − c ) . (67)
Thus the residual is generally nonzero in a noncollinear reference state. Yet it vanishes at c = ± 1 c=\pm1 c = ± 1 , where the same biquadratic interaction is locally absorbed into the distinct effective coefficients J i j − 2 Q i j J_{ij}-2Q_{ij} J ij − 2 Q ij (parallel) and J i j + 2 Q i j J_{ij}+2Q_{ij} J ij + 2 Q ij (antiparallel). At c = 0 c=0 c = 0 the projection returns J i j J_{ij} J ij but leaves a finite residual for Q i j ≠ 0 Q_{ij}\ne0 Q ij = 0 . Therefore a vanishing pair-block residual does not establish the absence of higher-order interactions or the transferability of the extracted exchange.
The second term looks like the mixed variation of a DMI energy D i j ⋅ ( e i × e j ) \mathbf{D}_{ij}\cdot(\mathbf{e}_i\times\mathbf{e}_j) D ij ⋅ ( e i × e j ) . If interpreted in isolation, it would suggest the apparent , reference-dependent coefficient
D i j a p p = − 2 Q i j ( e i × e j ) , (68) \mathbf{D}_{ij}^{\mathrm{app}}=-2Q_{ij}\left(\mathbf{e}_i\times\mathbf{e}_j\right), \tag*{(68)} D ij app = − 2 Q ij ( e i × e j ) , (68)
with magnitude 2 ∣ Q i j sin ϕ ∣ 2\lvert Q_{ij}\sin\phi\rvert 2 ∣ Q ij sin ϕ ∣ . This apparent DMI depends on the reference angle. It is not a separate chiral interaction: the same biquadratic term also changes the coefficient of δ e i ⋅ δ e j \delta\mathbf{e}_i\cdot\delta\mathbf{e}_j δ e i ⋅ δ e j . Both contributions belong to one Hessian.
We can check that the model has no DMI by reversing the relative angle. For coplanar spins,
E i j ( ϕ ) = − J i j cos ϕ + Q i j cos 2 ϕ = E i j ( − ϕ ) . (69) E_{ij}(\phi)=-J_{ij}\cos\phi+Q_{ij}\cos^2\phi=E_{ij}(-\phi). \tag*{(69)} E ij ( ϕ ) = − J ij cos ϕ + Q ij cos 2 ϕ = E ij ( − ϕ ) . (69)
Opposite chiralities have the same energy, although the apparent DMI coefficient changes sign. The antisymmetric-looking Hessian term therefore does not reveal a physical DMI. This issue is separate from the difference between field and moment rotations discussed in Sec. II.C.
Infinitesimal rotations probe only directions perpendicular to each reference spin. They cannot determine longitudinal components of a Cartesian interaction tensor. Thus, several full tensors can produce the same transverse Hessian. The Hessian is unambiguous; its division into bilinear, symmetric, and DMI-like couplings need not be.
A three-spin example shows how the apparent interaction between two spins can depend on a third. We consider the rotationally invariant energy
E = − J ( e 1 ⋅ e 2 + e 2 ⋅ e 3 ) − K 3 s ( e 1 ⋅ e 2 ) ( e 2 ⋅ e 3 ) . (70) E=-J(\mathbf{e}_1\cdot\mathbf{e}_2+\mathbf{e}_2\cdot\mathbf{e}_3)-K_{3s}(\mathbf{e}_1\cdot\mathbf{e}_2)(\mathbf{e}_2\cdot\mathbf{e}_3). \tag*{(70)} E = − J ( e 1 ⋅ e 2 + e 2 ⋅ e 3 ) − K 3 s ( e 1 ⋅ e 2 ) ( e 2 ⋅ e 3 ) . (70)
For coplanar spins, define α = θ 1 − θ 2 \alpha=\theta_1-\theta_2 α = θ 1 − θ 2 and β = θ 2 − θ 3 \beta=\theta_2-\theta_3 β = θ 2 − θ 3 . The mixed curvature between the end spins, despite the absence of a bilinear ( e 1 ⋅ e 3 ) (\mathbf{e}_1\cdot\mathbf{e}_3) ( e 1 ⋅ e 3 ) term, is
∂ 2 E ∂ θ 1 ∂ θ 3 = K 3 s sin α sin β . (71) \frac{\partial^2 E}{\partial\theta_1\partial\theta_3}=K_{3s}\sin\alpha\sin\beta. \tag*{(71)} ∂ θ 1 ∂ θ 3 ∂ 2 E = K 3 s sin α sin β . (71)
For a fixed orientation of spin 3, the coefficient of ( e 1 ⋅ e 2 ) (\mathbf{e}_1\cdot\mathbf{e}_2) ( e 1 ⋅ e 2 ) becomes J + K 3 s cos β J+K_{3s}\cos\beta J + K 3 s cos β , demonstrating its dependence on the orientation of the third spin. The three-spin interaction also contributes directly to the transverse coupling between spins 1 and 3. For coplanar reference directions e i = ( cos θ i , sin θ i , 0 ) \mathbf{e}_i=(\cos\theta_i,\sin\theta_i,0) e i = ( cos θ i , sin θ i , 0 ) , the local transverse vectors are t i 1 = ( − sin θ i , cos θ i , 0 ) \mathbf{t}_i^1=(-\sin\theta_i,\cos\theta_i,0) t i 1 = ( − sin θ i , cos θ i , 0 ) and t i 2 = ( 0 , 0 , 1 ) \mathbf{t}_i^2=(0,0,1) t i 2 = ( 0 , 0 , 1 ) . The mixed second variation is
δ 1 δ 3 E = − K 3 s ( δ e 1 ⋅ e 2 ) ( e 2 ⋅ δ e 3 ) , (72) \delta_1\delta_3E=-K_{3s}(\delta\mathbf{e}_1\cdot\mathbf{e}_2)(\mathbf{e}_2\cdot\delta\mathbf{e}_3), \tag*{(72)} δ 1 δ 3 E = − K 3 s ( δ e 1 ⋅ e 2 ) ( e 2 ⋅ δ e 3 ) , (72)
where the bilinear terms make no contribution to the 1–3 block. Writing s = sin α sin β s=\sin\alpha\sin\beta s = sin α sin β and c = e 1 ⋅ e 3 = cos ( α + β ) c=\mathbf{e}_1\cdot\mathbf{e}_3=\cos(\alpha+\beta) c = e 1 ⋅ e 3 = cos ( α + β ) gives
H 13 = ( K s 0 0 0 ) , P 13 = ( c 0 0 1 ) . (73) H_{13}=\begin{pmatrix}K_s&0\\0&0\end{pmatrix},\qquad P_{13}=\begin{pmatrix}c&0\\0&1\end{pmatrix}. \tag*{(73)} H 13 = ( K s 0 0 0 ) , P 13 = ( c 0 0 1 ) . (73)
Consequently, Eq. (57) yields
J 13 p r o j = − K c s 1 + c 2 , Δ H 13 = K s 1 + c 2 ( 1 0 0 − c ) . (74) J_{13}^{\mathrm{proj}}=-\frac{K_cs}{1+c^2},\qquad\Delta H_{13}=\frac{K_s}{1+c^2}\begin{pmatrix}1&0\\0&-c\end{pmatrix}. \tag*{(74)} J 13 proj = − 1 + c 2 K c s , Δ H 13 = 1 + c 2 K s ( 1 0 0 − c ) . (74)
For K s ≠ 0 K_s\ne0 K s = 0 , the relative Frobenius residual is
r 13 = 1 1 + c 2 , 1 2 ≤ r 13 ≤ 1. (75) r_{13}=\frac{1}{\sqrt{1+c^2}},\qquad\frac{1}{\sqrt{2}}\le r_{13}\le1. \tag*{(75)} r 13 = 1 + c 2 1 , 2 1 ≤ r 13 ≤ 1. (75)
Thus the induced pair curvature is never exactly scalar-Heisenberg-like when nonzero: at least half of its squared Frobenius norm remains outside the projected bilinear form. At c = 0 c=0 c = 0 the fitted coefficient vanishes even though H 13 H_{13} H 13 can remain finite; at s = 0 s=0 s = 0 the entire 1–3 mixed block vanishes and its relative residual is undefined. This illustrates both an apparent 1–3 pair coefficient without a microscopic bilinear 1–3 bond and the independent diagnostic supplied by the full-block residual. The calculation concerns local curvature at a prescribed reference state, not its stability.
To compare with other noncollinear calculations, we first write the response in Cartesian spin components:
A i j α β = − 1 π Im ∫ E F d ε Tr [ b i σ α G i j b j σ β G j i ] (76) A_{ij}^{\alpha\beta}=-\frac{1}{\pi}\operatorname{Im}\int^{E_F}d\varepsilon\operatorname{Tr}\left[b_i\sigma_\alpha G_{ij}b_j\sigma_\beta G_{ji}\right] \tag*{(76)} A ij α β = − π 1 Im ∫ E F d ε Tr [ b i σ α G ij b j σ β G j i ] (76)
H i j α β = t i α a A i j α β t j β b (77) H_{ij}^{\alpha\beta}=t_{i\alpha}^{a}A_{ij}^{\alpha\beta}t_{j\beta}^{b} \tag*{(77)} H ij α β = t i α a A ij α β t j β b (77)
Reference [19 ] obtained the energy of prescribed rotations in a general magnetic configuration using scattering matrices. Its formulation introduces Cartesian electronic coefficients and a noncollinear interaction tensor, together with local contributions and rotational constraints. In the notation of Ref. [19 ] , the Cartesian spin-energy combination takes the form J A , i j α β = A i j 00 δ α β − A i j α β J_{A,ij}^{\alpha\beta}=A_{ij}^{00}\delta_{\alpha\beta}-A_{ij}^{\alpha\beta} J A , ij α β = A ij 00 δ α β − A ij α β , with its original scattering and pair-counting conventions. The object to compare with Eq. (77) is its angular second derivative, including the orientation dependence and local terms, not an isolated Cartesian entry.
The Green-function product in Eq. (55) can be compared with a scattering-path expression only after matching the vertices, propagators, and pair-counting conventions.
Comparison with Ref. ref26
The coefficients A i j μ ν A_{ij}^{\mu\nu} A ij μν used by Ref. [26 ] are constructed from scalar and spin-vector components of the scattering-path operator and must not be identified with A i j α β A_{ij}^{\alpha\beta} A ij α β of Eq. (76) . Their connected two-site energy variation contains the symmetric Cartesian combination
K S , i j ( α β ) = ( A i j 00 − ∑ μ = x , y , z A i j μ μ ) δ α β + 2 A i j ( α β ) (78) K_{S,ij}^{(\alpha\beta)}=\left(A_{ij}^{00}-\sum_{\mu=x,y,z}A_{ij}^{\mu\mu}\right)\delta_{\alpha\beta}+2A_{ij}^{(\alpha\beta)} \tag*{(78)} K S , ij ( α β ) = ( A ij 00 − μ = x , y , z ∑ A ij μμ ) δ α β + 2 A ij ( α β ) (78)
Here the coefficient K S K_S K S denotes the term inside the published two-site energy variation, before applying the overall prefactor and pair-counting convention. In contrast, the spin-energy construction involves A i j 00 δ α β − A i j α β A_{ij}^{00}\delta_{\alpha\beta}-A_{ij}^{\alpha\beta} A ij 00 δ α β − A ij α β , where A i j 00 A_{ij}^{00} A ij 00 and A i j α β A_{ij}^{\alpha\beta} A ij α β denote the spin-scalar and spin-dependent components, respectively. These expressions are not interchangeable: the connected two-site variation and a proposed global bilinear energy must also reproduce the one-site variation. The comparison also involves the antisymmetric components and the rotational sum rule; comparing the two-site expressions alone does not establish these. We therefore make no claim of general tensor equality.
Collinear limit
For moments along z z z , Eq. (55) reduces to the exchange formula of Refs. [13 , 16 , 17 , 18 ] . The propagators are then spin-diagonal, and the transverse vertices are spin-flip operators, giving two opposite-spin propagation products. For the half-sum pair convention specified above, the x–x response gives
J i j = 1 π Im ∫ E F d ε Tr L [ b i G i j ↑ b j G j i ↓ + b i G i j ↓ b j G j i ↑ ] (79) J_{ij}=\frac{1}{\pi}\operatorname{Im}\int^{E_F}d\varepsilon\operatorname{Tr}_{L}\left[b_iG_{ij}^{\uparrow}b_jG_{ji}^{\downarrow}+b_iG_{ij}^{\downarrow}b_jG_{ji}^{\uparrow}\right] \tag*{(79)} J ij = π 1 Im ∫ E F d ε Tr L [ b i G ij ↑ b j G j i ↓ + b i G ij ↓ b j G j i ↑ ] (79)
The preceding collinear equation follows by tracing the two Pauli spin-flip channels and using H i j x x = − J i j H_{ij}^{xx}=-J_{ij} H ij xx = − J ij . The prefactor 1 / π 1/\pi 1/ π follows from the energy-Hessian normalization of the pair-response formula in Sec. III.A. In the scattering notation of Ref. [26 ] , the corresponding collinear combination is J i j = A i j 00 − A i j z z J_{ij}=A_{ij}^{00}-A_{ij}^{zz} J ij = A ij 00 − A ij z z , with their normalization.
Collinear exchange with a nonlocal magnetic potential
(79) assumes a site-local magnetic potential. A nonlocal potential can also connect orbitals on different sites. We first calculate its response to rotations of the magnetic potential and then consider rotations of the magnetic moments.
We write the spin-diagonal reference Hamiltonian as H 0 s = h 0 + Σ s H_0^s=h_0+\Sigma^s H 0 s = h 0 + Σ s (s = ↑ , ↓ s=\uparrow,\downarrow s =↑ , ↓ ), where Σ s \Sigma^s Σ s is a matrix in the orbital and site indices. We choose h 0 h_0 h 0 to contain the spin-independent part of the reference operator, including any spin-independent nonlocal terms; Σ s \Sigma^s Σ s then denotes the magnetic part in this prescribed-rotation convention. This separation fixes which operator is rotated, but does not make site-resolved rotations of a nonlocal magnetic potential unique. We let P i P_i P i project onto the orbitals whose spin direction is rotated at site i i i . A prescribed local rotation of the entire nonlocal magnetic operator is defined by Σ ( θ ) = U ( θ ) Σ 0 U † ( θ ) \Sigma(\theta)=U(\theta)\Sigma^0U^\dagger(\theta) Σ ( θ ) = U ( θ ) Σ 0 U † ( θ ) , where U = exp [ − i ∑ i θ i P i σ y / 2 ] U=\exp[-i\sum_i\theta_iP_i\sigma_y/2] U = exp [ − i ∑ i θ i P i σ y /2 ] . Its first-order transverse vertex has the spin-flip blocks
V i ↑ ↓ = 1 2 ( Σ ↑ P i − P i Σ ↓ ) , V_i^{\uparrow\downarrow}=\frac{1}{2}\left(\Sigma^\uparrow P_i-P_i\Sigma^\downarrow\right), V i ↑↓ = 2 1 ( Σ ↑ P i − P i Σ ↓ ) ,
V i ↓ ↑ = 1 2 ( P i Σ ↑ − Σ ↓ P i ) . (80) V_i^{\downarrow\uparrow}=\frac{1}{2}\left(P_i\Sigma^\uparrow-\Sigma^\downarrow P_i\right). \tag*{(80)} V i ↓↑ = 2 1 ( P i Σ ↑ − Σ ↓ P i ) . (80)
This unitary rotates the specified nonlocal magnetic operator, including its spin-dependent bond matrix elements. Rotating the moment at a site does not uniquely specify how those bond elements should change. Thus J i j B J_{ij}^{B} J ij B depends on the chosen operator rotation. We use mutually orthogonal, complete site projectors, P i P j = δ i j P i P_iP_j=\delta_{ij}P_i P i P j = δ ij P i and ∑ i P i = I \sum_iP_i=I ∑ i P i = I .
For a local splitting Σ ↑ = − ∑ i b i P i \Sigma^\uparrow=-\sum_i b_iP_i Σ ↑ = − ∑ i b i P i and Σ ↓ = + ∑ i b i P i \Sigma^\downarrow=+\sum_i b_iP_i Σ ↓ = + ∑ i b i P i , both blocks reduce to − b i P i -b_iP_i − b i P i , as in (79) . For a nonlocal operator, however, V i V_i V i contains matrix elements connecting site i i i with other sites. Replacing b i b_i b i by a matrix element of the nonlocal splitting in (79) would miss these terms.
The corresponding prescribed-operator pair coefficient is the direct Green-function extension of (79) :
J i j B = 1 π Im ∫ E F d ε Tr [ V i ↑ ↓ G ↓ V j ↓ ↑ G ↑ + V i ↓ ↑ G ↑ V j ↑ ↓ G ↓ ] − Tr ( γ 0 W i j ) , i ≠ j . (81) J_{ij}^{B}=\frac{1}{\pi}\operatorname{Im}\int^{E_F}d\varepsilon\ \operatorname{Tr}\left[V_i^{\uparrow\downarrow}G^\downarrow V_j^{\downarrow\uparrow}G^\uparrow+V_i^{\downarrow\uparrow}G^\uparrow V_j^{\uparrow\downarrow}G^\downarrow\right]-\operatorname{Tr}(\gamma_0W_{ij}),\qquad i\ne j. \tag*{(81)} J ij B = π 1 Im ∫ E F d ε Tr [ V i ↑↓ G ↓ V j ↓↑ G ↑ + V i ↓↑ G ↑ V j ↑↓ G ↓ ] − Tr ( γ 0 W ij ) , i = j . (81)
Here the trace is over all orbital and site indices, G s = ( ε − H 0 s + i 0 + ) − 1 G^s=(\varepsilon-H_0^s+i0^+)^{-1} G s = ( ε − H 0 s + i 0 + ) − 1 , and W i j = ∂ θ i ∂ θ j Σ ( θ ) ∣ 0 W_{ij}=\left.\partial_{\theta_i}\partial_{\theta_j}\Sigma(\theta)\right|_0 W ij = ∂ θ i ∂ θ j Σ ( θ ) 0 is the mixed second-order vertex. In the rotation prescription above,
W i j = − 1 4 [ P i σ y , [ P j σ y , Σ 0 ] ] , i ≠ j . (82) W_{ij}=-\frac{1}{4}[P_i\sigma_y,[P_j\sigma_y,\Sigma_0]],\qquad i\ne j. \tag*{(82)} W ij = − 4 1 [ P i σ y , [ P j σ y , Σ 0 ]] , i = j . (82)
The contact term vanishes for a site-diagonal magnetic potential at distinct sites; it need not vanish for a nonlocal potential. (81) therefore reduces to (79) in the local limit (with the same pair-counting convention). It remains a curvature of the rotated one-electron operator.
The contact contribution, including its mixed nonlocal term, is required for the cancellation of the uniform-rotation energy curvature. For a uniform rotation generated by S = σ y / 2 S=\sigma_y/2 S = σ y /2 , with the spin-independent Hamiltonian commuting with S S S , the reference Hamiltonian transforms as H ( Θ ) = e − i Θ S H r e f e i Θ S H(\Theta)=e^{-i\Theta S}H_{\rm ref}e^{i\Theta S} H ( Θ ) = e − i Θ S H ref e i Θ S . The corresponding perturbations are V = − i [ S , H r e f ] V=-i[S,H_{\rm ref}] V = − i [ S , H ref ] and W = − [ S , [ S , H r e f ] ] W=-[S,[S,H_{\rm ref}]] W = − [ S , [ S , H ref ]] . In the eigenbasis of H r e f H_{\rm ref} H ref , the contact and bubble contributions satisfy
⟨ W ⟩ = 2 ∑ h ∈ o c c , p ∈ e m p ( ε p − ε h ) ∣ S p h ∣ 2 = − H b u b b l e u n i f . (83) \langle W\rangle=2\sum_{h\in\mathrm{occ},\,p\in\mathrm{emp}}(\varepsilon_p-\varepsilon_h)|S_{ph}|^2=-H_{\mathrm{bubble}}^{\mathrm{unif}}. \tag*{(83)} ⟨ W ⟩ = 2 h ∈ occ , p ∈ emp ∑ ( ε p − ε h ) ∣ S p h ∣ 2 = − H bubble unif . (83)
The total uniform-rotation curvature therefore vanishes, as required by spin-rotational symmetry. This cancellation involves all site contributions, not individual pair interactions J i j B J_{ij}^{B} J ij B , and requires the complete first- and second-order perturbations generated by the same rotation. Omitting the mixed nonlocal contact term can violate this condition.
For constrained-moment rotations, the curvature is instead obtained from the stationary total energy E [ γ ] = Tr ( h 0 γ ) + Φ s t a t [ γ ] E[\gamma]=\operatorname{Tr}(h_0\gamma)+\Phi_{\mathrm{stat}}[\gamma] E [ γ ] = Tr ( h 0 γ ) + Φ stat [ γ ] . The interaction functional Φ \Phi Φ retains all interaction contributions, including nonlocal terms, even when these terms do not appear explicitly in the prescribed one-electron rotation vertices. The reference potential and the induced potential are, respectively,
Σ 0 = δ Φ s t a t δ γ ∣ γ 0 , δ Σ = I ( δ γ ) , (84) \Sigma_0=\left.\frac{\delta\Phi_{\mathrm{stat}}}{\delta\gamma}\right|_{\gamma_0},\qquad\delta\Sigma=\mathcal{I}(\delta\gamma), \tag*{(84)} Σ 0 = δ γ δ Φ stat γ 0 , δ Σ = I ( δ γ ) , (84)
where I \mathcal{I} I is the second functional derivative of Φ s t a t \Phi_{\mathrm{stat}} Φ stat . For an infinitesimal rotation of the constrained moment at site i i i , we denote the self-consistent first-order density-matrix change by δ i γ \delta_i\gamma δ i γ . In the occupied–empty eigenstate basis of H 0 H_0 H 0 , we write δ i γ = ∑ p h ( z p h ( i ) ∣ p ⟩ ⟨ h ∣ + h . c . ) \delta_i\gamma=\sum_{ph}(z_{ph}^{(i)}|p\rangle\langle h|+\mathrm{h.c.}) δ i γ = ∑ p h ( z p h ( i ) ∣ p ⟩ ⟨ h ∣ + h.c. ) . The stationary mixed curvature follows directly from the second-order energy:
J i j M = − 2 Re ∑ p h ( ε p − ε h ) z p h ( i ) ∗ z p h ( j ) − Tr [ δ i γ I ( δ j γ ) ] , i ≠ j . (85) J_{ij}^{M}=-2\operatorname{Re}\sum_{ph}(\varepsilon_p-\varepsilon_h)z_{ph}^{(i)*}z_{ph}^{(j)}-\operatorname{Tr}\left[\delta_i\gamma\,\mathcal{I}(\delta_j\gamma)\right],\qquad i\ne j. \tag*{(85)} J ij M = − 2 Re p h ∑ ( ε p − ε h ) z p h ( i ) ∗ z p h ( j ) − Tr [ δ i γ I ( δ j γ ) ] , i = j . (85)
We determine the amplitudes z p h ( i ) z_{ph}^{(i)} z p h ( i ) from stationarity of the same quadratic energy subject to the prescribed transverse moments, Tr ( P k σ x δ i γ ) = M k δ i k \operatorname{Tr}(P_k\sigma_x\delta_i\gamma)=M_k\delta_{ik} Tr ( P k σ x δ i γ ) = M k δ ik , with the sign of M k M_k M k understood in the chosen local transverse coordinates. This displays the x x x -channel constraint; the y y y channel follows by spin-rotation symmetry. All other electronic amplitudes are allowed to relax. Equation (85) is the stationary-PT counterpart of Eq. (79) : it retains the ordinary particle–hole contribution and adds the explicit change of the nonlocal interaction energy. The second term is therefore not, in general, represented by a site-only exchange splitting or an independent-particle Green-function product.
The amplitudes in Eq. (85) must minimize the quadratic energy while satisfying the moment constraints. All other allowed occupied–empty channels must be free to relax. Freezing those channels or omitting the induced kernel I \mathcal{I} I gives a different curvature, not the stated J i j M J_{ij}^{M} J ij M .
The stationary functional has its own, distinct Goldstone cancellation. If h 0 h_0 h 0 is spin independent and Φ [ R γ R † ] = Φ s t a t [ γ ] \Phi[R\gamma R^\dagger]=\Phi_{\mathrm{stat}}[\gamma] Φ [ R γ R † ] = Φ stat [ γ ] under a global spin rotation, the uniform tangent δ γ u n i f = − i [ S , γ 0 ] \delta\gamma_{\mathrm{unif}}=-i[S,\gamma_0] δ γ unif = − i [ S , γ 0 ] is an exact symmetry direction. Evaluating the stationary second variation along this direction gives
2 ∑ p h ( ϵ p − ϵ h ) ∣ z p h u n i f ∣ 2 + Tr [ δ γ u n i f I ( δ γ u n i f ) ] = 0. (86) 2\sum_{ph}(\epsilon_p-\epsilon_h)\lvert z_{ph}^{\mathrm{unif}}\rvert^2+\operatorname{Tr}[\delta\gamma_{\mathrm{unif}}\mathcal{I}(\delta\gamma_{\mathrm{unif}})]=0. \tag*{(86)} 2 p h ∑ ( ϵ p − ϵ h ) ∣ z p h unif ∣ 2 + Tr [ δ γ unif I ( δ γ unif )] = 0. (86)
Here z p h u n i f z_{ph}^{\mathrm{unif}} z p h unif denotes the occupied–empty amplitude of the global rotation. The induced functional-kernel contribution cancels the positive particle–hole term; the cancellation is not obtained in general by keeping only the equilibrium nonlocal potential. This is the stationary-total-energy identity, not the prescribed-operator contact-term identity (82).
The two expressions describe different rotations. Equation (81) extends the prescribed-potential Green-function formula to a nonlocal operator; Eq. (85) gives the constrained-moment exchange from the stationary total energy. Their difference can reflect induced electronic relaxation and interaction-kernel contributions, as well as the distinct constraints and energy functionals discussed in Sec. II.C; it is not, without a common-functional derivation, uniquely attributable to one relaxation correction. The two-electron example in Appendix A evaluates the stationary result explicitly and finds
J A B M = − 2 t 2 2 U − K n l , 2 U − K n l > 4 t . (87) J_{AB}^{M}=-\frac{2t^2}{2U-K_{\mathrm{nl}}},\qquad2U-K_{\mathrm{nl}}>4t. \tag*{(87)} J A B M = − 2 U − K nl 2 t 2 , 2 U − K nl > 4 t . (87)
Here K n l K_{\mathrm{nl}} K nl multiplies an orbital-nonlocal density-matrix interaction, not a site-only Stoner coupling. This model result illustrates why the full nonlocal functional must be retained; it is not a general replacement U ↦ U − K n l / 2 U\mapsto U-K_{\mathrm{nl}}/2 U ↦ U − K nl /2 .
The distinction between a nonlocal one-electron magnetic potential and a nonlocal magnetic response kernel can be made explicit within a generalized Stoner functional,
F [ γ ] = F 0 [ γ ] − 1 4 ∑ i j I i j M i ⋅ M j , (88) F[\gamma]=F_0[\gamma]-\frac{1}{4}\sum_{ij}I_{ij}\mathbf{M}_i\cdot\mathbf{M}_j, \tag*{(88)} F [ γ ] = F 0 [ γ ] − 4 1 ij ∑ I ij M i ⋅ M j , (88)
which generates a site-local magnetic field B i = 1 2 ∑ j I i j M j \mathbf{B}_i=\frac{1}{2}\sum_j I_{ij}\mathbf{M}_j B i = 2 1 ∑ j I ij M j , even when I i j I_{ij} I ij is nonlocal. For a translationally invariant collinear ferromagnet, B = I ( 0 ) M / 2 B=I(0)M/2 B = I ( 0 ) M /2 .
In terms of the independent-electron static transverse susceptibility χ 0 ( q ) \chi_0(q) χ 0 ( q ) , defined by δ M ⊥ ( q ) = χ 0 ( q ) δ B ⊥ ( q ) \delta M_\perp(q)=\chi_0(q)\delta B_\perp(q) δ M ⊥ ( q ) = χ 0 ( q ) δ B ⊥ ( q ) and satisfying χ 0 ( 0 ) = M / B \chi_0(0)=M/B χ 0 ( 0 ) = M / B , the difference between the constrained-moment and prescribed-field rotational curvatures takes the form
H M ( q ) − H B ( q ) = [ M − B χ 0 ( q ) ] 2 χ 0 ( q ) + M 2 2 [ I ( 0 ) − I ( q ) ] . (89) H^M(q)-H^B(q)=\frac{[M-B\chi_0(q)]^2}{\chi_0(q)}+\frac{M^2}{2}[I(0)-I(q)]. \tag*{(89)} H M ( q ) − H B ( q ) = χ 0 ( q ) [ M − B χ 0 ( q ) ] 2 + 2 M 2 [ I ( 0 ) − I ( q )] . (89)
Here H B H^B H B is the curvature of the frozen prescribed-field electronic free energy, whereas H M H^M H M is the stationary constrained-moment curvature of Eq. (87); the expression assumes a regular transverse Legendre transform. In the Pauli-spin response convention used here, the Stoner functional generates δ B ( q ) = [ I ( q ) / 2 ] δ M ( q ) \delta B(q)=[I(q)/2]\delta M(q) δ B ( q ) = [ I ( q ) /2 ] δ M ( q ) , so the interaction kernel in the inverse transverse susceptibility is I r e s p ( q ) = I ( q ) / 2 I_{\mathrm{resp}}(q)=I(q)/2 I resp ( q ) = I ( q ) /2 . For an analytic, inversion-symmetric response, the first term begins at order q 4 q^4 q 4 , whereas the second can contribute already at order q 2 q^2 q 2 . Consequently, a nonlocal magnetic response kernel can modify the leading exchange stiffness in the constrained-moment representation even when the equilibrium one-electron magnetic potential is local. The equilibrium magnetic splitting alone is therefore insufficient to determine the stationary magnetic exchange interactions.
MAGNETIC PAIR INTERACTIONS IN SPIN SPIRALS
For planar and conical spirals, we evaluate the transverse electronic pair response and ask whether a single field-based scalar coefficient reproduces its angular dependence. Here ϑ \vartheta ϑ denotes the cone angle, Q \mathbf{Q} Q the spiral wave vector, and R i \mathbf{R}_i R i a site position; Eq. (91) specifies the site phase.
e i = ( sin ϑ cos ϕ i , sin ϑ sin ϕ i , cos ϑ ) (90) \mathbf{e}_i=(\sin\vartheta\cos\phi_i,\sin\vartheta\sin\phi_i,\cos\vartheta) \tag*{(90)} e i = ( sin ϑ cos ϕ i , sin ϑ sin ϕ i , cos ϑ ) (90)
with
ϕ i = Q ⋅ R i (91) \phi_i=\mathbf{Q}\cdot\mathbf{R}_i \tag*{(91)} ϕ i = Q ⋅ R i (91)
A convenient orthonormal tangent basis is
t i 1 = ( cos ϑ cos ϕ i , cos ϑ sin ϕ i , − sin ϑ ) (92) \mathbf{t}_i^1=(\cos\vartheta\cos\phi_i,\cos\vartheta\sin\phi_i,-\sin\vartheta) \tag*{(92)} t i 1 = ( cos ϑ cos ϕ i , cos ϑ sin ϕ i , − sin ϑ ) (92)
and
t i 2 = ( − sin ϕ i , cos ϕ i , 0 ) (93) \mathbf{t}_i^2=(-\sin\phi_i,\cos\phi_i,0) \tag*{(93)} t i 2 = ( − sin ϕ i , cos ϕ i , 0 ) (93)
The labels 1 and 2 denote normalized transverse tilts. A change of spiral phase instead moves the moment by
∂ e i ∂ ϕ i = sin ϑ t i 2 (94) \frac{\partial\mathbf{e}_i}{\partial\phi_i}=\sin\vartheta\,\mathbf{t}_i^2 \tag*{(94)} ∂ ϕ i ∂ e i = sin ϑ t i 2 (94)
The coordinate θ i 2 \theta_i^2 θ i 2 used below is normalized by the transverse rotation rather than by the phase change; this matters for a conical spiral with sin ϑ ≠ 1 \sin\vartheta\ne1 sin ϑ = 1 .
For the spiral tangent vectors of Eqs. (92) , (93) , the first-order vertices are given by Eq. (9) , and the off-site electronic pair Hessian by Eq. (17) .
Conical spiral: derivation of scalar pair interactions
The isotropic Heisenberg pair Hessian of Eq. (20) gives the geometrical projection for a conical spiral.
Writing ϕ i j = ϕ i − ϕ j = Q ⋅ ( R i − R j ) \phi_{ij}=\phi_i-\phi_j=\mathbf{Q}\cdot(\mathbf{R}_i-\mathbf{R}_j) ϕ ij = ϕ i − ϕ j = Q ⋅ ( R i − R j ) , direct evaluation for the conical spiral gives
P i j = ( sin 2 ϑ + cos 2 ϑ cos ϕ i j cos ϑ sin ϕ i j − cos ϑ sin ϕ i j cos ϕ i j ) (95) P_{ij}=\begin{pmatrix}
\sin^2\vartheta+\cos^2\vartheta\cos\phi_{ij} & \cos\vartheta\sin\phi_{ij}\\
-\cos\vartheta\sin\phi_{ij} & \cos\phi_{ij}
\end{pmatrix} \tag*{(95)} P ij = ( sin 2 ϑ + cos 2 ϑ cos ϕ ij − cos ϑ sin ϕ ij cos ϑ sin ϕ ij cos ϕ ij ) (95)
A scalar bilinear interaction can reproduce the electronic rotational response only if
H i j a b = − J i j P i j a b (96) H_{ij}^{ab}=-J_{ij}P_{ij}^{ab} \tag*{(96)} H ij ab = − J ij P ij ab (96)
Equation (95) tests whether the electronic response has the scalar-exchange form.
The diagonal channels yield two candidate exchanges wherever their geometric factors are nonzero:
J i j ( 1 ) = − H i j 11 sin 2 ϑ + cos 2 ϑ cos ϕ i j (97) J_{ij}^{(1)}=-\frac{H_{ij}^{11}}{\sin^2\vartheta+\cos^2\vartheta\cos\phi_{ij}} \tag*{(97)} J ij ( 1 ) = − sin 2 ϑ + cos 2 ϑ cos ϕ ij H ij 11 (97)
and
J i j ( 2 ) = − H i j 22 cos ϕ i j (98) J_{ij}^{(2)}=-\frac{H_{ij}^{22}}{\cos\phi_{ij}} \tag*{(98)} J ij ( 2 ) = − cos ϕ ij H ij 22 (98)
Matching the two diagonal values alone does not establish bilinear compatibility. In a conical spiral, the mixed channels also obey
H i j 12 = − J i j cos ϑ sin ϕ i j (99) H_{ij}^{12}=-J_{ij}\cos\vartheta\sin\phi_{ij} \tag*{(99)} H ij 12 = − J ij cos ϑ sin ϕ ij (99)
and
H i j 21 = J i j cos ϑ sin ϕ i j (100) H_{ij}^{21}=J_{ij}\cos\vartheta\sin\phi_{ij} \tag*{(100)} H ij 21 = J ij cos ϑ sin ϕ ij (100)
The opposite signs are consistent with the symmetry of the complete Hessian,
H i j a b = H j i b a (101) H_{ij}^{ab}=H_{ji}^{ba} \tag*{(101)} H ij ab = H j i ba (101)
Equations (96)–(100) test diagonal and mixed channels. Changing the cone angle or pitch then tests transferability beyond a single spiral.
Planar spin spiral
In a planar spiral, the tangent geometry simplifies:
ϑ = π 2 (102) \vartheta=\frac{\pi}{2} \tag*{(102)} ϑ = 2 π (102)
and the tangent vectors become
t i 1 = ( 0 , 0 , − 1 ) , t i 2 = ( − sin ϕ i , cos ϕ i , 0 ) (103) \mathbf{t}_i^1=(0,0,-1),\qquad\mathbf{t}_i^2=(-\sin\phi_i,\cos\phi_i,0) \tag*{(103)} t i 1 = ( 0 , 0 , − 1 ) , t i 2 = ( − sin ϕ i , cos ϕ i , 0 ) (103)
The geometrical matrix is then diagonal,
P i j = ( 1 0 0 cos ϕ i j ) (104) P_{ij}=\begin{pmatrix}
1 & 0\\
0 & \cos\phi_{ij}
\end{pmatrix} \tag*{(104)} P ij = ( 1 0 0 cos ϕ ij ) (104)
Here channel 1 denotes the out-of-plane tilt, whereas channel 1 in the local frame of Sec. III.A denotes the in-plane tilt; the channel labels are local conventions, not different physical responses. The out-of-plane tilt is especially simple. A scalar exchange would require
H i j 11 = − J i j (105) H_{ij}^{11}=-J_{ij} \tag*{(105)} H ij 11 = − J ij (105)
and hence
J i j = − H i j 11 (106) J_{ij}=-H_{ij}^{11} \tag*{(106)} J ij = − H ij 11 (106)
This channel needs no division by an angle-dependent factor and remains regular at every relative angle.
The out-of-plane electronic vertex follows from Eq. (9) using the tangent convention of Eq. (103) :
V i 1 = b i σ z (107) V_i^1=b_i\sigma_z \tag*{(107)} V i 1 = b i σ z (107)
Substitution into Eq. (17) gives
H i j 11 = − b i b j 2 π Im ∫ E F d ε Tr [ σ z G i j σ z G j i + σ z G j i σ z G i j ] (108) H_{ij}^{11}=-\frac{b_i b_j}{2\pi}\operatorname{Im}\int^{E_F}d\varepsilon\operatorname{Tr}\left[\sigma_zG_{ij}\sigma_zG_{ji}+\sigma_zG_{ji}\sigma_zG_{ij}\right] \tag*{(108)} H ij 11 = − 2 π b i b j Im ∫ E F d ε Tr [ σ z G ij σ z G j i + σ z G j i σ z G ij ] (108)
and therefore
J i j = b i b j 2 π Im ∫ E F d ε Tr [ σ z G i j σ z G j i + σ z G j i σ z G i j ] (109) J_{ij}=\frac{b_i b_j}{2\pi}\operatorname{Im}\int^{E_F}d\varepsilon\operatorname{Tr}\left[\sigma_zG_{ij}\sigma_zG_{ji}+\sigma_zG_{ji}\sigma_zG_{ij}\right] \tag*{(109)} J ij = 2 π b i b j Im ∫ E F d ε Tr [ σ z G ij σ z G j i + σ z G j i σ z G ij ] (109)
The in-plane and mixed channels provide independent checks. For the in-plane channel, Eq. (96) requires
H i j 22 = − J i j cos ϕ i j (110) H_{ij}^{22}=-J_{ij}\cos\phi_{ij} \tag*{(110)} H ij 22 = − J ij cos ϕ ij (110)
or, equivalently,
H i j 22 = H i j 11 cos ϕ i j (111) H_{ij}^{22}=H_{ij}^{11}\cos\phi_{ij} \tag*{(111)} H ij 22 = H ij 11 cos ϕ ij (111)
Unlike an extraction based on division by cos ϕ i j \cos\phi_{ij} cos ϕ ij , Eq. (111) remains regular when cos ϕ i j = 0 \cos\phi_{ij}=0 cos ϕ ij = 0 .
For a scalar isotropic pair interaction, the mixed components must satisfy
H i j 12 = H i j 21 = 0 (112) H_{ij}^{12}=H_{ij}^{21}=0 \tag*{(112)} H ij 12 = H ij 21 = 0 (112)
Equations (105) and (108) define the scalar coefficient through the out-of-plane channel; Eqs. (110) and (111) are then the independent tests, and both must hold. In particular, planar geometry alone does not make the mixed electronic responses vanish.
For the isotropic bilinear–biquadratic pair of Sec. III.A, the planar-spiral transverse curvatures provide an explicit test:
H i j ∥ = ∂ 2 E i j ∂ α i ∂ α j = − J i j cos ϕ + 2 Q i j cos ( 2 ϕ ) . (113) H_{ij}^{\parallel}=\frac{\partial^{2}E_{ij}}{\partial\alpha_i\partial\alpha_j}=-J_{ij}\cos\phi+2Q_{ij}\cos(2\phi). \tag*{(113)} H ij ∥ = ∂ α i ∂ α j ∂ 2 E ij = − J ij cos ϕ + 2 Q ij cos ( 2 ϕ ) . (113)
For small out-of-plane tilts β i \beta_i β i and β j \beta_j β j about the same planar reference, it is
H i j ⊥ = ∂ 2 E i j ∂ β i ∂ β j = − J i j + 2 Q i j cos ϕ . (114) H_{ij}^{\perp}=\frac{\partial^{2}E_{ij}}{\partial\beta_i\partial\beta_j}=-J_{ij}+2Q_{ij}\cos\phi. \tag*{(114)} H ij ⊥ = ∂ β i ∂ β j ∂ 2 E ij = − J ij + 2 Q ij cos ϕ . (114)
A purely bilinear isotropic Heisenberg pair with scalar coefficient J i j e f f J_{ij}^{\mathrm{eff}} J ij eff would instead give H i j ∥ , H = − J i j e f f cos ϕ H_{ij}^{\parallel,\mathrm{H}}=-J_{ij}^{\mathrm{eff}}\cos\phi H ij ∥ , H = − J ij eff cos ϕ and H i j ⊥ , H = − J i j e f f H_{ij}^{\perp,\mathrm{H}}=-J_{ij}^{\mathrm{eff}} H ij ⊥ , H = − J ij eff . Thus the two independent extractions give
J i j e f f , ∥ = J i j − 2 Q i j cos ( 2 ϕ ) cos ϕ , cos ϕ ≠ 0 , (115) J_{ij}^{\mathrm{eff},\parallel}=J_{ij}-2Q_{ij}\frac{\cos(2\phi)}{\cos\phi},\qquad\cos\phi\ne0, \tag*{(115)} J ij eff , ∥ = J ij − 2 Q ij cos ϕ cos ( 2 ϕ ) , cos ϕ = 0 , (115)
J i j e f f , ⊥ = J i j − 2 Q i j cos ϕ , (116) J_{ij}^{\mathrm{eff},\perp}=J_{ij}-2Q_{ij}\cos\phi, \tag*{(116)} J ij eff , ⊥ = J ij − 2 Q ij cos ϕ , (116)
and consequently
J i j e f f , ∥ − J i j e f f , ⊥ = 2 Q i j sin 2 ϕ cos ϕ . (117) J_{ij}^{\mathrm{eff},\parallel}-J_{ij}^{\mathrm{eff},\perp}=\frac{2Q_{ij}\sin^{2}\phi}{\cos\phi}. \tag*{(117)} J ij eff , ∥ − J ij eff , ⊥ = cos ϕ 2 Q ij sin 2 ϕ . (117)
Equation (116) shows that for a generic noncollinear angle and nonzero Q i j Q_{ij} Q ij , the two candidates differ. At ϕ = π / 2 \phi=\pi/2 ϕ = π /2 , the in-plane Heisenberg curvature vanishes for any finite scalar J i j e f f J_{ij}^{\mathrm{eff}} J ij eff , whereas the biquadratic model gives H i j ∥ = − 2 Q i j H_{ij}^{\parallel}=-2Q_{ij} H ij ∥ = − 2 Q ij . Thus, a single scalar bilinear interaction cannot reproduce both channels at this reference configuration. The apparent singularity of Eq. (115) at cos ϕ = 0 \cos\phi=0 cos ϕ = 0 reflects the vanishing bilinear geometrical factor, not a divergence of the original Hessian.
No longitudinal tensor components enter this comparison.
For this coplanar bilinear–biquadratic model, the mixed in-plane/out-of-plane curvatures vanish by symmetry. This special property does not apply to the general electronic pair response, which is subject to the independent conditions in Eqs. (110) and (111), with the coefficient defined by Eq. (108).
Collinear limits
The planar spiral contains the usual collinear ferromagnetic (FM) and two-sublattice antiferromagnetic (AFM) limits. Their pair responses are worth checking before turning to long-wavelength modes.
Q = 0 (118) Q=0 \tag*{(118)} Q = 0 (118)
all moments are parallel,
e i = e j (119) \mathbf{e}_i=\mathbf{e}_j \tag*{(119)} e i = e j (119)
For the nonrelativistic isotropic problem, the two transverse channels are equivalent, and Eq. (96) becomes
H i j 11 = H i j 22 = − J i j , H i j 12 = H i j 21 = 0 (120) H_{ij}^{11}=H_{ij}^{22}=-J_{ij},\qquad H_{ij}^{12}=H_{ij}^{21}=0 \tag*{(120)} H ij 11 = H ij 22 = − J ij , H ij 12 = H ij 21 = 0 (120)
This gives the familiar FM pair response; its long-wavelength limit follows in Sec. V.
A collinear two-sublattice AFM is obtained from a commensurate spiral vector Q A F \mathbf{Q}_{\mathrm{AF}} Q AF satisfying
Q A F ⋅ ( R i − R j ) = π ( m o d 2 π ) (121) \mathbf{Q}_{\mathrm{AF}}\cdot(\mathbf{R}_i-\mathbf{R}_j)=\pi\pmod{2\pi} \tag*{(121)} Q AF ⋅ ( R i − R j ) = π ( mod 2 π ) (121)
for opposite sublattices
For the planar representation,
ϕ i j = π (122) \phi_{ij}=\pi \tag*{(122)} ϕ ij = π (122)
for an opposite-sublattice pair. Since
cos ϕ i j = − 1 , sin ϕ i j = 0 (123) \cos\phi_{ij}=-1,\qquad\sin\phi_{ij}=0 \tag*{(123)} cos ϕ ij = − 1 , sin ϕ ij = 0 (123)
the geometrical factors become
P i j = ( 1 0 0 − 1 ) (124) P_{ij}=\begin{pmatrix}1&0\\0&-1\end{pmatrix} \tag*{(124)} P ij = ( 1 0 0 − 1 ) (124)
A scalar bilinear interaction therefore gives
H i j 11 = − J i j (125) H_{ij}^{11}=-J_{ij} \tag*{(125)} H ij 11 = − J ij (125)
and
H i j 22 = + J i j (126) H_{ij}^{22}=+J_{ij} \tag*{(126)} H ij 22 = + J ij (126)
The two AFM transverse channels have opposite signs because the local axes point in opposite directions. Both come from the same electronic pair response.
The FM and AFM limits of these spirals lead to the long-wavelength energy costs calculated next. We treat their dynamics separately.
FERROMAGNETIC LIMIT AND SPIN-WAVE STIFFNESS
We compare the static curvatures for field and moment rotations in a ferromagnet and examine the conditions for extracting the physical spin-wave stiffness.
The two transverse directions are equivalent. The two-vertex term gives the intersite curvature, and the rotational sum rule fixes the local term. We denote the off-site transverse Hessian elements by H 0 R H_{0R} H 0 R .
H 0 R = − 1 2 π Im ∫ E F d ε Tr [ V 0 G 0 R V R G R 0 + V R G R 0 V 0 G 0 R ] , R ≠ 0 (127) H_{0R}=-\frac{1}{2\pi}\operatorname{Im}\int^{E_F}d\varepsilon\operatorname{Tr}\left[V_0G_{0R}V_RG_{R0}+V_RG_{R0}V_0G_{0R}\right],\qquad R\ne0 \tag*{(127)} H 0 R = − 2 π 1 Im ∫ E F d ε Tr [ V 0 G 0 R V R G R 0 + V R G R 0 V 0 G 0 R ] , R = 0 (127)
Expanding at small wave vector gives
H ( q ) = ∑ R ≠ 0 H 0 R ( e i q ⋅ R − 1 ) = A α β q α q β + O ( q 4 ) (128) H(\mathbf{q})=\sum_{R\ne0}H_{0R}\left(e^{i\mathbf{q}\cdot\mathbf{R}}-1\right)=A_{\alpha\beta}q_\alpha q_\beta+O(q^4) \tag*{(128)} H ( q ) = R = 0 ∑ H 0 R ( e i q ⋅ R − 1 ) = A α β q α q β + O ( q 4 ) (128)
A α β = − 1 2 ∑ R ≠ 0 H 0 R R α R β (129) A_{\alpha\beta}=-\frac{1}{2}\sum_{R\ne0}H_{0R}R_\alpha R_\beta \tag*{(129)} A α β = − 2 1 R = 0 ∑ H 0 R R α R β (129)
Since H 0 R = H 0 , − R H_{0R}=H_{0,-R} H 0 R = H 0 , − R , odd powers of q q q vanish. The coefficient A A A in Eq. (129) is a long-wavelength Hessian coefficient, distinct from the Cartesian electronic tensor A A A in Eq. (76) . We denote the equilibrium spin angular momentum per magnetic site, in units of ℏ \hbar ℏ , by S S S . If the prescribed-field curvature is combined with an assumed adiabatic precession coefficient − S -S − S , it gives the provisional stiffness estimate
D α β = A α β S = 1 4 π S Im ∫ E F d ε ∑ R ≠ 0 R α R β × Tr [ V 0 G 0 R V R G R 0 + V R G R 0 V 0 G 0 R ] (130) \begin{aligned}
D_{\alpha\beta}=\frac{A_{\alpha\beta}}{S}=\frac{1}{4\pi S}\operatorname{Im}\int^{E_F}d\varepsilon\sum_{R\ne0}R_\alpha R_\beta\\
&\times\operatorname{Tr}\left[V_0G_{0R}V_RG_{R0}+V_RG_{R0}V_0G_{0R}\right] \tag*{(130)}
\end{aligned} D α β = S A α β = 4 π S 1 Im ∫ E F d ε R = 0 ∑ R α R β × Tr [ V 0 G 0 R V R G R 0 + V R G R 0 V 0 G 0 R ] (130)
(130) is therefore a prescribed-field estimate, not yet an identification with the physical spin-wave stiffness. We next compare the static curvatures and then examine the frequency-dependent response.
The Pauli-spin convention of Eq. (2) is used here: the magnetic Hamiltonian is − b σ ⋅ e -b\boldsymbol{\sigma}\cdot\mathbf{e} − b σ ⋅ e , with b = Δ 2 b=\frac{\Delta}{2} b = 2 Δ and m = ⟨ σ z ⟩ m=\langle\sigma_z\rangle m = ⟨ σ z ⟩ per site. The bare transverse response X 0 ( q ) X_0(q) X 0 ( q ) is defined by δ ⟨ σ x ⟩ = X 0 ( q ) δ b x \delta\langle\sigma_x\rangle=X_0(q)\delta b_x δ ⟨ σ x ⟩ = X 0 ( q ) δ b x for the perturbation − δ b x σ x -\delta b_x\sigma_x − δ b x σ x , with the same site projection and Fourier convention as the rotational vertices. The local second-order rotational term contributes + b m +bm + bm to the full angular Hessian, whereas the two-vertex response contributes − b 2 X 0 ( q ) -b^2X_0(q) − b 2 X 0 ( q ) . Their cancellation for a uniform rotation gives m = b X 0 ( 0 ) m=bX_0(0) m = b X 0 ( 0 ) . Thus the complete field Hessian, not merely its off-site pair term, is
H B ( q ) = b m − b 2 X 0 ( q ) = b 2 [ X 0 ( 0 ) − X 0 ( q ) ] . (131) H^B(q)=bm-b^2X_0(q)=b^2[X_0(0)-X_0(q)]. \tag*{(131)} H B ( q ) = bm − b 2 X 0 ( q ) = b 2 [ X 0 ( 0 ) − X 0 ( q )] . (131)
For the relaxed moment, Eqs. (30) –(34) give H M ( q ) = m 2 X ( q ) − 1 H^M(q)=m^2X(q)^{-1} H M ( q ) = m 2 X ( q ) − 1 when the conjugate field and moment use the same Pauli-spin normalization. In a scalar transverse channel with a wave-vector-independent response kernel I r e s p I_{\mathrm{resp}} I resp , X ( q ) − 1 = X 0 ( q ) − 1 − I r e s p X(q)^{-1}=X_0(q)^{-1}-I_{\mathrm{resp}} X ( q ) − 1 = X 0 ( q ) − 1 − I resp . The Goldstone condition fixes I r e s p = X 0 ( 0 ) − 1 I_{\mathrm{resp}}=X_0(0)^{-1} I resp = X 0 ( 0 ) − 1 . In the generalized Stoner convention of Eq. (88) , I r e s p = I ( 0 ) / 2 I_{\mathrm{resp}}=I(0)/2 I resp = I ( 0 ) /2 ; the factor of two follows from the prefactor − 1 / 4 -1/4 − 1/4 in that functional and is not an additional physical interaction. Combining this response identity with m = b X 0 ( 0 ) m=bX_0(0) m = b X 0 ( 0 ) gives
H M ( q ) = m 2 [ X 0 ( q ) − 1 − X 0 ( 0 ) − 1 ] = X 0 ( 0 ) X 0 ( q ) H B ( q ) (132) \begin{aligned}
H^M(q)=m^2\left[X_0(q)^{-1}-X_0(0)^{-1}\right] \\
&=\frac{X_0(0)}{X_0(q)}H^B(q) \tag*{(132)}
\end{aligned} H M ( q ) = m 2 [ X 0 ( q ) − 1 − X 0 ( 0 ) − 1 ] = X 0 ( q ) X 0 ( 0 ) H B ( q ) (132)
For the relaxed moment, Eqs. (30) –(34) give H M ( q ) = m 2 X ( q ) − 1 H^M(q)=m^2X(q)^{-1} H M ( q ) = m 2 X ( q ) − 1 when the conjugate field and moment use the same Pauli-spin normalization. In a scalar transverse channel with a wave-vector-independent response kernel I r e s p I_{\mathrm{resp}} I resp , X ( q ) − 1 = X 0 ( q ) − 1 − I r e s p X(q)^{-1}=X_0(q)^{-1}-I_{\mathrm{resp}} X ( q ) − 1 = X 0 ( q ) − 1 − I resp . The Goldstone condition fixes I r e s p = X 0 ( 0 ) − 1 I_{\mathrm{resp}}=X_0(0)^{-1} I resp = X 0 ( 0 ) − 1 . In the generalized Stoner convention of Eq. (88) , I r e s p = I ( 0 ) / 2 I_{\mathrm{resp}}=I(0)/2 I resp = I ( 0 ) /2 ; the factor of two follows from the prefactor − 1 / 4 -1/4 − 1/4 in that functional and is not an additional physical interaction. Combining this response identity with m = b X 0 ( 0 ) m=bX_0(0) m = b X 0 ( 0 ) gives
If X 0 ( 0 ) X_0(0) X 0 ( 0 ) is finite and X 0 ( q ) − X 0 ( 0 ) = O ( q 2 ) X_0(q)-X_0(0)=O(q^2) X 0 ( q ) − X 0 ( 0 ) = O ( q 2 ) , Eqs. (131) and (132) give H M ( q ) − H B ( q ) = O ( q 4 ) H^M(q)-H^B(q)=O(q^4) H M ( q ) − H B ( q ) = O ( q 4 ) . Thus the two methods give the same leading q 2 q^2 q 2 coefficient, but not necessarily the same interactions at finite q q q . This agreement requires fixed moment lengths and a local, q q q -independent interaction kernel. It need not hold for nonlocal kernels, coupled transverse modes, or nonanalytic response.
The energy cost of a static rotation does not by itself fix the spin-wave frequency. The spin-wave frequency also depends on the precessional dynamics. For the circular vectors e ± = ( 1 , ± i ) T / 2 \mathbf{e}_{\pm}=(1,\pm i)^T/\sqrt{2} e ± = ( 1 , ± i ) T / 2 defined in Sec. II, the uniform local-splitting Ward identity gives g u n i f = − m / 2 = − S g_{\mathrm{unif}}=-m/2=-S g unif = − m /2 = − S per cell, where m = N ↑ − N ↓ m=N_{\uparrow}-N_{\downarrow} m = N ↑ − N ↓ per cell and S = m / 2 S=m/2 S = m /2 is the equilibrium spin angular momentum in units of ℏ \hbar ℏ (one magnetic site per cell). Indeed, i z g u n i f e i z g_{\mathrm{unif}}\mathbf{e} i z g unif e has circular eigenvalues ∓ z g u n i f = ± z S \mp z g_{\mathrm{unif}}=\pm zS ∓ z g unif = ± z S ; thus the e − \mathbf{e}_{-} e − branch has the coefficient Λ − = − S \Lambda_{-}=-S Λ − = − S and, for positive restoring curvature, the positive-frequency pole. The e + \mathbf{e}_{+} e + branch has Λ + = + S \Lambda_{+}=+S Λ + = + S and the opposite-frequency pole. These signs follow from the chosen circular basis and the e − i ω t e^{-i\omega t} e − iω t time convention. The corresponding coefficient for the relaxed response is not established by this local-splitting identity.
For the physical spin-wave mode, the inverse response has the low-energy form λ − ( q , ω ) = A α β p h y s q α q β − S ℏ ω + O ( q 4 , q 2 ω , ω 2 ) \lambda_{-}(\mathbf{q},\omega)=A^{\mathrm{phys}}_{\alpha\beta}q_\alpha q_\beta-S\hbar\omega+O(q^4,q^2\omega,\omega^2) λ − ( q , ω ) = A α β phys q α q β − S ℏ ω + O ( q 4 , q 2 ω , ω 2 ) . Setting it to zero gives ℏ ω q = D α β q α q β \hbar\omega_{\mathbf{q}}=D_{\alpha\beta}q_\alpha q_\beta ℏ ω q = D α β q α q β , where D α β = A α β p h y s / S D_{\alpha\beta}=A^{\mathrm{phys}}_{\alpha\beta}/S D α β = A α β phys / S . Here A p h y s A^{\mathrm{phys}} A phys denotes the leading coefficient of the relaxed constrained-moment curvature, A p h y s = A M A^{\mathrm{phys}}=A^M A phys = A M , which is the relevant static coefficient when the collective mode is described by constrained-moment adiabatic precession. (130) therefore gives the physical stiffness only if the field-rotation coefficient A B A^B A B equals A p h y s A^{\mathrm{phys}} A phys and the precession coefficient is − S -S − S . The conditions for the first equality were given above; the second follows from the dynamics, not the static energy.
As emphasized by Antropov [22 ] , the term linear in ω \omega ω has the same long-wavelength order as q 2 q^2 q 2 for quadratic ferromagnetic spin waves. This frequency dependence must be included when calculating the dispersion directly from the dynamical susceptibility. However, the spin-wave stiffness can also be obtained from static exchange interactions, as in the Liechtenstein–Katsnelson approach [16 ] , using the magnetic moments and the standard equations of spin precession. In this case, an explicit calculation of the linear-frequency coefficient is not required when the precession coefficient is fixed at − S -S − S by the local-splitting Ward identity; this conclusion does not automatically extend to relaxed, nonlocal, or frequency-dependent responses.
ANTIFERROMAGNETIC LIMIT AND SPIN-WAVE VELOCITY
In an antiferromagnet, the dynamics involves coupled rotations of the two sublattices. The spin-wave velocity depends on their individual restoring forces and on their coupling.
We consider a collinear antiferromagnet with two symmetry-equivalent sublattices and equal spin magnitudes S S S (in units of ℏ \hbar ℏ ). We use the same diagonal Hessian element for both sublattices because symmetry relates them; equal moment sizes alone are not enough. Because the moments point in opposite directions, their precession terms have opposite signs. In common laboratory axes, the frequency term is − ℏ ω S σ z -\hbar\omega S\sigma_z − ℏ ω S σ z (with a corresponding choice of circular polarization). We use the same precession convention as in the ferromagnetic case. We assume adiabatic Landau–Lifshitz dynamics; delayed electronic response can change the frequencies and cause damping.
With opposite equilibrium moments on the two sublattices, let d 0 μ , R ν d_{0\mu,R\nu} d 0 μ , R ν denote the displacement from site ( 0 , μ ) (0,\mu) ( 0 , μ ) to site ( R , ν ) (R,\nu) ( R , ν ) , including the basis-site positions, and write
H μ ν ( q ) = H μ ν l o c + ∑ R H 0 μ , R ν V G e i q ⋅ d 0 μ , R ν (133) H_{\mu\nu}(\mathbf{q})=H_{\mu\nu}^{\mathrm{loc}}+\sum_R H_{0\mu,R\nu}^{\mathrm{VG}}e^{i\mathbf{q}\cdot d_{0\mu,R\nu}} \tag*{(133)} H μν ( q ) = H μν loc + R ∑ H 0 μ , R ν VG e i q ⋅ d 0 μ , R ν (133)
H 0 μ , R ν V G = − 1 2 π Im ∫ E F d ε Tr [ V 0 μ G 0 μ , R ν V R ν G R ν , 0 μ + V R ν G R ν , 0 μ V 0 μ G 0 μ , R ν ] (134) H_{0\mu,R\nu}^{\mathrm{VG}}=-\frac{1}{2\pi}\operatorname{Im}\int^{E_F}d\varepsilon\ \operatorname{Tr}\left[V_{0\mu}G_{0\mu,R\nu}V_{R\nu}G_{R\nu,0\mu}+V_{R\nu}G_{R\nu,0\mu}V_{0\mu}G_{0\mu,R\nu}\right] \tag*{(134)} H 0 μ , R ν VG = − 2 π 1 Im ∫ E F d ε Tr [ V 0 μ G 0 μ , R ν V R ν G R ν , 0 μ + V R ν G R ν , 0 μ V 0 μ G 0 μ , R ν ] (134)
Equations (133) and (134) calculate the field-rotation Hessian H B H^B H B . The algebra below applies to either static Hessian. A physical spin-wave velocity, however, requires the relaxed-moment Hessian H M H^M H M and its corresponding precession dynamics. Equation (134) does not include this relaxation.
For the two oppositely magnetized sublattices, the restoring-force matrix has diagonal and intersublattice elements a ( q ) a(\mathbf{q}) a ( q ) and c ( q ) c(\mathbf{q}) c ( q ) , expressed in common laboratory axes. In the absence of anisotropy, rotational symmetry requires a ( 0 ) = − c ( 0 ) = κ a(0)=-c(0)=\kappa a ( 0 ) = − c ( 0 ) = κ , where κ \kappa κ is the uniform canting coefficient. The static Hessian is
H ( q ) = ( a ( q ) c ( q ) c ∗ ( q ) a ( q ) ) , a ( 0 ) = κ , c ( 0 ) = − κ (135) H(\mathbf{q})=\begin{pmatrix}a(\mathbf{q})&c(\mathbf{q})\\c^*(\mathbf{q})&a(\mathbf{q})\end{pmatrix},\qquad a(0)=\kappa,\qquad c(0)=-\kappa \tag*{(135)} H ( q ) = ( a ( q ) c ∗ ( q ) c ( q ) a ( q ) ) , a ( 0 ) = κ , c ( 0 ) = − κ (135)
The two AFM sublattices precess in opposite directions, giving opposite signs in their frequency terms. The following dynamical determinant describes one circular-polarization sector in common laboratory axes; the conjugate sector gives the opposite-frequency partners. The coupled Landau–Lifshitz equations give det [ H ( q ) − ℏ ω S σ z ] = 0 \det[H(\mathbf{q})-\hbar\omega S\sigma_z]=0 det [ H ( q ) − ℏ ω S σ z ] = 0 , with δ E = 1 2 θ † H θ \delta E=\frac{1}{2}\theta^\dagger H\theta δ E = 2 1 θ † H θ . The spin-wave frequency therefore obeys
( ℏ ω ) 2 = a ( q ) 2 − ∣ c ( q ) ∣ 2 S 2 (136) (\hbar\omega)^2=\frac{a(\mathbf{q})^2-\lvert c(\mathbf{q})\rvert^2}{S^2} \tag*{(136)} ( ℏ ω ) 2 = S 2 a ( q ) 2 − ∣ c ( q ) ∣ 2 (136)
We expand both matrix elements at small q \mathbf{q} q . The intersublattice element c ( q ) c(\mathbf{q}) c ( q ) may contain a term linear in q \mathbf{q} q and imaginary:
a ( q ) = κ + a α β q α q β + O ( q 4 ) , a(\mathbf{q})=\kappa+a_{\alpha\beta}q_\alpha q_\beta+O(q^4), a ( q ) = κ + a α β q α q β + O ( q 4 ) ,
c ( q ) = − κ + i L α q α + c α β q α q β + O ( q 3 ) (137) c(\mathbf{q})=-\kappa+iL_\alpha q_\alpha+c_{\alpha\beta}q_\alpha q_\beta+O(q^3) \tag*{(137)} c ( q ) = − κ + i L α q α + c α β q α q β + O ( q 3 ) (137)
The coefficients are sums of Hessian elements weighted by intersite displacements:
a α β = − 1 2 ∑ R H 0 A , R A V G d α d β , L α = ∑ R H 0 A , R B V G d α , a_{\alpha\beta}=-\frac{1}{2}\sum_R H_{0A,R A}^{\mathrm{VG}}d_\alpha d_\beta,\qquad L_\alpha=\sum_R H_{0A,R B}^{\mathrm{VG}}d_\alpha, a α β = − 2 1 R ∑ H 0 A , R A VG d α d β , L α = R ∑ H 0 A , R B VG d α ,
c α β = − 1 2 ∑ R H 0 A , R B V G d α d β (138) c_{\alpha\beta}=-\frac{1}{2}\sum_R H_{0A,R B}^{\mathrm{VG}}d_\alpha d_\beta \tag*{(138)} c α β = − 2 1 R ∑ H 0 A , R B VG d α d β (138)
The linear imaginary term enters the long-wavelength curvature through its square; with Eq. (136) , the squared velocity becomes
v 2 ( q ^ ) = [ 2 κ ( a α β + Re c α β ) − L α L β ] q ^ α q ^ β ℏ 2 S 2 (139) v^2(\hat{\mathbf{q}})=\frac{\left[2\kappa\left(a_{\alpha\beta}+\operatorname{Re}c_{\alpha\beta}\right)-L_\alpha L_\beta\right]\hat{q}_\alpha\hat{q}_\beta}{\hbar^2S^2} \tag*{(139)} v 2 ( q ^ ) = ℏ 2 S 2 [ 2 κ ( a α β + Re c α β ) − L α L β ] q ^ α q ^ β (139)
The role of the L α L β L_\alpha L_\beta L α L β term is clearest under a change of Fourier origin. Shifting the Fourier origin of sublattice B by δ \delta δ multiplies c ( q ) c(\mathbf{q}) c ( q ) by exp ( i q ⋅ δ ) \exp(i\mathbf{q}\cdot\delta) exp ( i q ⋅ δ ) without changing ∣ c ( q ) ∣ \lvert c(\mathbf{q})\rvert ∣ c ( q )∣ . The individual coefficients L α L_\alpha L α and c α β c_{\alpha\beta} c α β then change, but the combination entering Eq. (139) is invariant. The corresponding coefficient transformations are
L α ′ = L α − κ δ α , L'_\alpha=L_\alpha-\kappa\delta_\alpha, L α ′ = L α − κ δ α ,
c α β ′ = c α β − L α δ β + L β δ α 2 + κ δ α δ β 2 (140) c'_{\alpha\beta}=c_{\alpha\beta}-\frac{L_\alpha\delta_\beta+L_\beta\delta_\alpha}{2}+\frac{\kappa\delta_\alpha\delta_\beta}{2} \tag*{(140)} c α β ′ = c α β − 2 L α δ β + L β δ α + 2 κ δ α δ β (140)
2 κ Re c α β ′ − L α ′ L β ′ = 2 κ Re c α β − L α L β . (141) 2\kappa\operatorname{Re}c'_{\alpha\beta}-L'_\alpha L'_\beta=2\kappa\operatorname{Re}c_{\alpha\beta}-L_\alpha L_\beta. \tag*{(141)} 2 κ Re c α β ′ − L α ′ L β ′ = 2 κ Re c α β − L α L β . (141)
We define C α β = 2 κ ( a α β + Re c α β ) − L α L β C_{\alpha\beta}=2\kappa(a_{\alpha\beta}+\operatorname{Re}c_{\alpha\beta})-L_\alpha L_\beta C α β = 2 κ ( a α β + Re c α β ) − L α L β . A real long-wavelength velocity requires C α β q ^ α q ^ β ≥ 0 C_{\alpha\beta}\hat{q}_\alpha\hat{q}_\beta\geq0 C α β q ^ α q ^ β ≥ 0 in every propagation direction, assuming S > 0 S>0 S > 0 and κ > 0 \kappa>0 κ > 0 . A positive value gives a nonzero linear velocity. This condition does not establish stability at larger wave vectors. In particular, the L α L β L_\alpha L_\beta L α L β term in Eq. (139) cannot be omitted. The formulas assume equal sublattice angular momenta and the energy convention δ E = 1 2 θ † H θ \delta E = \frac{1}{2}\theta^\dagger H\theta δ E = 2 1 θ † H θ . Unequal moments or significant electronic retardation require a different dynamical matrix.
If symmetry makes the intersublattice element even to linear order, L = 0 L = 0 L = 0 and the result reduces to the familiar second-moment form, for which A α β = a α β + Re c α β A_{\alpha\beta} = a_{\alpha\beta} + \operatorname{Re} c_{\alpha\beta} A α β = a α β + Re c α β :
A α β = − 1 2 ∑ R ( H 0 R A A + H 0 R A B ) d α d β (142) A_{\alpha\beta} = -\frac{1}{2}\sum_R \left(H_{0R}^{AA} + H_{0R}^{AB}\right)d_\alpha d_\beta \tag*{(142)} A α β = − 2 1 R ∑ ( H 0 R AA + H 0 R A B ) d α d β (142)
v 2 ( q ^ ) = 2 κ A α β q ^ α q ^ β ℏ 2 S 2 v^2(\widehat{\boldsymbol{q}}) = \frac{2\kappa A_{\alpha\beta}\widehat{q}_\alpha\widehat{q}_\beta}{\hbar^2 S^2} v 2 ( q ) = ℏ 2 S 2 2 κ A α β q α q β
We now consider interacting electrons without spin-orbit coupling. Their self-energy may depend on frequency, but we take it to be local in space. We denote the bare propagator by G 0 G_0 G 0 and the dressed propagator by G G G . They satisfy the Dyson equation
G − 1 ( i ω n ) = G 0 − 1 ( i ω n ) − Σ ( i ω n ) (143) G^{-1}(i\omega_n) = G_0^{-1}(i\omega_n) - \Sigma(i\omega_n) \tag*{(143)} G − 1 ( i ω n ) = G 0 − 1 ( i ω n ) − Σ ( i ω n ) (143)
The self-energy contains spin-independent and magnetic parts. For a collinear reference, Σ i ( z ) = Σ i c ( z ) + Σ i s ( z ) σ z \Sigma_i(z) = \Sigma_i^c(z) + \Sigma_i^s(z)\sigma_z Σ i ( z ) = Σ i c ( z ) + Σ i s ( z ) σ z , where Σ i s ( z ) = [ Σ i ↑ ( z ) − Σ i ↓ ( z ) ] / 2 \Sigma_i^s(z) = [\Sigma_i^\uparrow(z) - \Sigma_i^\downarrow(z)]/2 Σ i s ( z ) = [ Σ i ↑ ( z ) − Σ i ↓ ( z )] /2 . A prescribed rotation is Σ i ( θ i ) = U i Σ i U i † \Sigma_i(\theta_i) = U_i\Sigma_i U_i^\dagger Σ i ( θ i ) = U i Σ i U i † , with U i = exp [ − i θ i a σ a / 2 ] U_i = \exp[-i\theta_i^a\sigma_a/2] U i = exp [ − i θ i a σ a /2 ] . Differentiation yields the rotational vertices:
V i a = − i 2 [ σ a , Σ i ] , V_i^a = -\frac{i}{2}[\sigma_a,\Sigma_i], V i a = − 2 i [ σ a , Σ i ] ,
W i a , i b = − 1 8 ( [ σ a , [ σ b , Σ i ] ] + [ σ b , [ σ a , Σ i ] ] ) (144) W_{ia,ib} = -\frac{1}{8}\left([\sigma_a,[\sigma_b,\Sigma_i]] + [\sigma_b,[\sigma_a,\Sigma_i]]\right) \tag*{(144)} W ia , ib = − 8 1 ( [ σ a , [ σ b , Σ i ]] + [ σ b , [ σ a , Σ i ]] ) (144)
These vertices rotate the specified self-energy, not the self-consistent magnetization. Their frequency dependence comes from Σ \Sigma Σ ; relaxation of the interacting electrons must be calculated separately.
Static pair coefficients and the collinear limit
Expanding the electronic trace logarithm about the equilibrium dressed propagator, with G − 1 ( θ ) = G − 1 − δ V G^{-1}(\theta) = G^{-1} - \delta V G − 1 ( θ ) = G − 1 − δ V and δ V \delta V δ V representing the rotated self-energy vertices, gives a quadratic two-vertex term and a term containing W W W . For rotations on distinct sites, the latter has no mixed local derivative, so the off-site prescribed-rotation contribution is
H i a , j b ( 0 ) = T ∑ n Tr L σ [ V i a ( i ω n ) G i j ( i ω n ) × V j b ( i ω n ) G j i ( i ω n ) ] , i ≠ j (145) H_{ia,jb}^{(0)} = T\sum_n \operatorname{Tr}_{L\sigma}\left[V_i^a(i\omega_n)G_{ij}(i\omega_n)\times V_j^b(i\omega_n)G_{ji}(i\omega_n)\right],\qquad i\ne j \tag*{(145)} H ia , j b ( 0 ) = T n ∑ Tr L σ [ V i a ( i ω n ) G ij ( i ω n ) × V j b ( i ω n ) G j i ( i ω n ) ] , i = j (145)
Equation (145) uses dressed propagators and frequency-dependent vertices. It gives only the off-site two-vertex term. The local second-order vertex is still needed for the full Hessian and its rotational sum rule.
Transverse rotations connect opposite spin projections. The spin trace gives
H i j , y y ( 0 ) = T ∑ n Tr L [ Σ i s G i j ↑ Σ j s G j i ↓ + Σ i s G i j ↓ Σ j s G j i ↑ ] (146) H_{ij,yy}^{(0)} = T\sum_n \operatorname{Tr}_L\left[\Sigma_i^s G_{ij}^{\uparrow}\Sigma_j^s G_{ji}^{\downarrow} + \Sigma_i^s G_{ij}^{\downarrow}\Sigma_j^s G_{ji}^{\uparrow}\right] \tag*{(146)} H ij , y y ( 0 ) = T n ∑ Tr L [ Σ i s G ij ↑ Σ j s G j i ↓ + Σ i s G ij ↓ Σ j s G j i ↑ ] (146)
Equation (146) contains both spin-flip channels. Their traces are equal if the propagators obey transpose reciprocity and the local spin self-energy is symmetric. We keep both terms so the formula does not rely on those additional conditions. The scalar coefficient is
J i j ( 0 ) = − T ∑ n Tr L [ Σ i s G i j ↑ Σ j s G j i ↓ + Σ i s G i j ↓ Σ j s G j i ↑ ] (147) J_{ij}^{(0)} = -T\sum_n \operatorname{Tr}_L\left[\Sigma_i^s G_{ij}^{\uparrow}\Sigma_j^s G_{ji}^{\downarrow} + \Sigma_i^s G_{ij}^{\downarrow}\Sigma_j^s G_{ji}^{\uparrow}\right] \tag*{(147)} J ij ( 0 ) = − T n ∑ Tr L [ Σ i s G ij ↑ Σ j s G j i ↓ + Σ i s G ij ↓ Σ j s G j i ↑ ] (147)
The scalar coefficient takes the usual form [21 ] .
Rotations of the magnetization
A rigid moment rotates geometrically in the same way whether the electrons are correlated or not:
δ M i ( 1 ) = θ i × M i , δ M i ( 2 ) = 1 2 θ i × ( θ i × M i ) . (148) \delta M_i^{(1)} = \theta_i \times M_i,\qquad\delta M_i^{(2)} = \frac{1}{2}\theta_i \times(\theta_i \times M_i). \tag*{(148)} δ M i ( 1 ) = θ i × M i , δ M i ( 2 ) = 2 1 θ i × ( θ i × M i ) . (148)
To hold a moment at a chosen angle, the constraining field changes with the electronic state. In a stationary Luttinger–Ward treatment with functional Φ L W [ G ] \Phi_{\mathrm{LW}}[G] Φ LW [ G ] , the self-energy responds through the irreducible vertex
Γ = δ Σ δ G = δ 2 Φ L W δ G δ G . (149) \Gamma= \frac{\delta\Sigma}{\delta G} = \frac{\delta^2\Phi_{\mathrm{LW}}}{\delta G\,\delta G}. \tag*{(149)} Γ = δ G δ Σ = δ G δ G δ 2 Φ LW . (149)
For the source coupling − ∑ j β B j c ⋅ M ^ j -\sum_{j\beta} B_j^c\cdot\widehat{M}_j − ∑ j β B j c ⋅ M j , the linearized Dyson equation and projected moment response are
δ G = G ( − ∑ j β M ^ j β δ B j β + Γ ⋅ δ G ) G , \delta G = G\left(-\sum_{j\beta}\widehat{M}_{j\beta}\delta B_j^\beta+ \Gamma\cdot\delta G\right)G, δ G = G − j β ∑ M j β δ B j β + Γ ⋅ δ G G ,
δ M i α = Tr ( M ^ i α δ G ) = ∑ j β X i α , j β δ B j β . (150) \delta M_{i\alpha} = \operatorname{Tr}(\widehat{M}_{i\alpha}\delta G) = \sum_{j\beta}X_{i\alpha,j\beta}\delta B_j^\beta. \tag*{(150)} δ M i α = Tr ( M i α δ G ) = j β ∑ X i α , j β δ B j β . (150)
Here Γ \Gamma Γ acts in the full orbital, site, spin, and internal-frequency space. The Matsubara trace and source sign follow the convention of Eq. (30) ; X X X includes the induced self-energy rather than only a dressed-propagator bubble.
The stationary Legendre transformation and inverse-response identity were established in Eqs. (32) –(34) . They apply here with the fully interacting response X X X generated by Eq. (150) ; no additional exchange mapping is assumed.
To expose the correlated electronic content of X X X , define the two-propagator operator L [ A ] = G A G \mathcal{L}[A] = GAG L [ A ] = G A G and a source-to-electron coupling Q [ δ B c ] = − ∑ j β M ^ j β δ B j β c Q[\delta B^c] = -\sum_{j\beta}\widehat{M}_{j\beta}\delta B^c_{j\beta} Q [ δ B c ] = − ∑ j β M j β δ B j β c . Solving Eq. (150) in the full electronic space and writing M [ δ G ] i α = Tr ( M ^ i α δ G ) \mathcal{M}[\delta G]_{i\alpha} = \operatorname{Tr}(\widehat{M}_{i\alpha}\delta G) M [ δ G ] i α = Tr ( M i α δ G ) gives the projected response
X = M ( 1 − L Γ ) − 1 L Q . (151) X = \mathcal{M}(1-\mathcal{L}\Gamma)^{-1}\mathcal{L}Q. \tag*{(151)} X = M ( 1 − L Γ ) − 1 L Q . (151)
The electronic response is solved before projection, and the resulting moment-space X X X is then inverted. The dressed two-vertex contribution in Eq. (145) is therefore not substituted for the relaxed moment curvature: the induced self-energy and the electronic relaxation enter through the same stationary response.
The absence of an explicit second-order δ ( 2 ) G \delta^{(2)}G δ ( 2 ) G term follows from stationarity rather than from neglecting electronic relaxation. Writing the stationary functional as F [ G , B ] \mathcal{F}[G,B] F [ G , B ] , differentiation of F G [ G ∗ ( B ) , B ] = 0 \mathcal{F}_{G}[G_{*}(B),B] = 0 F G [ G ∗ ( B ) , B ] = 0 gives G ∗ , B = − F G G − 1 F G B G_{*,B} = -\mathcal{F}_{GG}^{-1}\mathcal{F}_{GB} G ∗ , B = − F GG − 1 F GB and hence
F B B = F B B − F B G F G G − 1 F G B . (152) \mathcal{F}_{BB} = \mathcal{F}_{BB} - \mathcal{F}_{BG}\mathcal{F}_{GG}^{-1}\mathcal{F}_{GB}. \tag*{(152)} F B B = F B B − F B G F GG − 1 F GB . (152)
Here the subscripts denote functional derivatives evaluated at the stationary solution, and the inverse acts in the allowed electronic relaxation space. This stationary Schur complement retains the first-order induced self-energy through Γ = δ Σ / δ G \Gamma= \delta\Sigma/\delta G Γ = δ Σ/ δ G ; second-order electronic variations multiply the vanishing first derivative of F \mathcal{F} F . For a spontaneously spin-rotation-broken state, the source and response are defined with a suitable symmetry-breaking regulator, or inverses are restricted to nonsingular transverse sectors before taking the zero-field limit. These prescriptions should not be interchanged without checking the limit. The static density-matrix functional used for orbital-nonlocal interactions earlier in the paper is a distinct specialization; its instantaneous kernel is not automatically identical to the frequency-dependent Luttinger–Ward vertex.
The quadratic constrained energy contains both the constraining-field response and the geometric second-order change of the moment:
δ E ( 2 ) = 1 2 ∑ i δ B i c ( 1 ) ⋅ δ M i ( 1 ) + ∑ i B i c ( 0 ) ⋅ δ M i ( 2 ) . (153) \delta E^{(2)} = \frac{1}{2}\sum_i \delta B_i^{c(1)}\cdot\delta M_i^{(1)}+\sum_i B_i^{c(0)}\cdot\delta M_i^{(2)}. \tag*{(153)} δ E ( 2 ) = 2 1 i ∑ δ B i c ( 1 ) ⋅ δ M i ( 1 ) + i ∑ B i c ( 0 ) ⋅ δ M i ( 2 ) . (153)
Using the local tangent vectors t i a t_i^a t i a and second angular derivatives s i , a b s_{i,ab} s i , ab , as in Eq. (29) , together with δ B c = X − 1 δ M \delta B^c = X^{-1}\delta M δ B c = X − 1 δ M from Eq. (30) , gives the angular Hessian:
H i a , j b M = M i M j t i a ⋅ ( X − 1 ) i j ⊥ t j b + δ i j M i B i c ⋅ s i , a b (154) H^M_{ia,jb}=M_iM_j t_i^a\cdot(X^{-1})_{ij}^{\perp}t_j^b+\delta_{ij}M_i B_i^c\cdot s_{i,ab} \tag*{(154)} H ia , j b M = M i M j t i a ⋅ ( X − 1 ) ij ⊥ t j b + δ ij M i B i c ⋅ s i , ab (154)
Equation (153) specializes Eq. (34) to the correlated response in Eq. (151) . With the pair-counting convention of Eq. (18) , the off-site moment Hessian is minus the exchange. Thus
J i j M = − M i M j ( X − 1 ) i j ⊥ , i ≠ j (155) J^M_{ij}=-M_iM_j(X^{-1})_{ij}^{\perp},\qquad i\ne j \tag*{(155)} J ij M = − M i M j ( X − 1 ) ij ⊥ , i = j (155)
Dynamical magnetization fluctuations {#sec:D}
We next consider time-dependent moment fluctuations. The self-energy depends on the electron energy z z z ; the magnetic response depends on the excitation energy Ω \Omega Ω .
We couple a time-dependent source h h h to the same projected moment operators as in Sec. II.C. We denote the dimensionless Legendre effective action by S M [ M ] S_M[M] S M [ M ] and the connected magnetic response to h h h by X X X . With β = 1 / ( k B T ) \beta=1/(k_{\mathrm{B}}T) β = 1/ ( k B T ) and ν n = 2 π n / ( β ℏ ) \nu_n=2\pi n/(\beta\hbar) ν n = 2 π n / ( β ℏ ) , we define M i α ( n ) = 1 β ℏ ∫ 0 β ℏ d τ e i ν n τ M i α ( τ ) M_{i\alpha}(n)=\frac{1}{\beta\hbar}\int_0^{\beta\hbar}d\tau\,e^{i\nu_n\tau}M_{i\alpha}(\tau) M i α ( n ) = β ℏ 1 ∫ 0 β ℏ d τ e i ν n τ M i α ( τ ) . The inverse-Hessian identity is
1 β δ 2 S M δ M i α ( − n ) δ M j β ( n ) = X i α , j β − 1 ( i ℏ ν n ) . (156) \frac{1}{\beta}\frac{\delta^2 S_M}{\delta M_{i\alpha}(-n)\delta M_{j\beta}(n)}=X^{-1}_{i\alpha,j\beta}(i\hbar\nu_n). \tag*{(156)} β 1 δ M i α ( − n ) δ M j β ( n ) δ 2 S M = X i α , j β − 1 ( i ℏ ν n ) . (156)
The full X X X follows from the finite-frequency Dyson variation. With fermionic frequencies ω m = ( 2 m + 1 ) π / ( β ℏ ) \omega_m=(2m+1)\pi/(\beta\hbar) ω m = ( 2 m + 1 ) π / ( β ℏ ) and bosonic frequencies ν n = 2 π n / ( β ℏ ) \nu_n=2\pi n/(\beta\hbar) ν n = 2 π n / ( β ℏ ) , set Ω = i ℏ ν n \Omega=i\hbar\nu_n Ω = i ℏ ν n and let z + = i ℏ ω m + Ω z_+=i\hbar\omega_m+\Omega z + = i ℏ ω m + Ω , z − = i ℏ ω m z_-=i\hbar\omega_m z − = i ℏ ω m , and denote the external one-electron perturbation by δ h ( z + , z − ) \delta h(z_+,z_-) δ h ( z + , z − ) . For the constraining-field coupling in Eq. (150) , δ h = − ∑ j β M ^ j β δ B j β c \delta h=-\sum_{j\beta}\widehat{M}_{j\beta}\delta B^c_{j\beta} δ h = − ∑ j β M j β δ B j β c ; thus the positive sign multiplying δ h \delta h δ h below agrees with the negative sign in Q Q Q . The interacting perturbation equation is
δ G + − = G + ( δ h + − + δ Σ + − ) G − , δ Σ + − = Γ Ω ⋅ δ G + − , G ± = G ( z ± ) . (157) \delta G_{+-}=G_+(\delta h_{+-}+\delta\Sigma_{+-})G_-,
\qquad
\delta\Sigma_{+-}=\Gamma_{\Omega}\cdot\delta G_{+-},
\qquad
G_{\pm}=G(z_{\pm}). \tag*{(157)} δ G +− = G + ( δ h +− + δ Σ +− ) G − , δ Σ +− = Γ Ω ⋅ δ G +− , G ± = G ( z ± ) . (157)
Here Γ Ω = δ Σ / δ G \Gamma_{\Omega}=\delta\Sigma/\delta G Γ Ω = δ Σ/ δ G acts on the full orbital, spin, site, and internal-frequency structure; its frequency arguments are not replaced by a single scalar function of Ω \Omega Ω . With L Ω [ A ] + − = G + A + − G − \mathcal{L}_{\Omega}[A]_{+-}=G_+A_{+-}G_- L Ω [ A ] +− = G + A +− G − , this gives
δ G = ( 1 − L Ω Γ Ω ) − 1 L Ω [ δ h ] , (158) \delta G=(1-\mathcal{L}_{\Omega}\Gamma_{\Omega})^{-1}\mathcal{L}_{\Omega}[\delta h], \tag*{(158)} δ G = ( 1 − L Ω Γ Ω ) − 1 L Ω [ δ h ] , (158)
where the inverse denotes the solution of the integral equation in electronic space. Projection onto the moment operators defines X ( Ω ) X(\Omega) X ( Ω ) ; only then is its inverse taken to obtain the moment kernel. In the regular static limit Ω → 0 \Omega\to0 Ω → 0 , G + → G − G_+\to G_- G + → G − , and Eq. (157) reduces to Eq. (150) . The zero-frequency response therefore gives Eq. (154) , with the same treatment of rotational zero modes.
The uniform-rotation identity follows within a spin-rotation-invariant stationary functional Φ L W [ G ] \Phi_{\mathrm{LW}}[G] Φ LW [ G ] . For a
(155) assumes a stationary collinear reference with finite, sufficiently rigid moments.
In Eq. (155) , X X X is obtained from Eq. (150) , rather than by rotating the equilibrium self-energy alone. Thus Eq. (155) need not coincide with Eq. (147) , which describes rotations of constant spin rotation U = e − i ϵ S U=e^{-i\epsilon S} U = e − i ϵ S , covariance requires Σ [ U G U † ] = U Σ [ G ] U † \Sigma[UGU^\dagger]=U\Sigma[G]U^\dagger Σ [ U G U † ] = U Σ [ G ] U † . Differentiation gives
δ S G = − i [ S , G ] , \delta_S G=-i[S,G], δ S G = − i [ S , G ] ,
Γ 0 ⋅ δ S G = − i [ S , Σ ] = δ S Σ . (159) \Gamma_0\cdot\delta_S G=-i[S,\Sigma]=\delta_S\Sigma. \tag*{(159)} Γ 0 ⋅ δ S G = − i [ S , Σ ] = δ S Σ. (159)
When the spin-independent bare inverse propagator commutes with S S S , Dyson’s identity yields δ S G = G ( δ S Σ ) G \delta_S G=G(\delta_S\Sigma)G δ S G = G ( δ S Σ ) G . Hence δ S G \delta_S G δ S G is a nonzero solution of the homogeneous equation ( 1 − L 0 Γ 0 ) δ S G = 0 (1-\mathcal{L}_0\Gamma_0)\delta_S G=0 ( 1 − L 0 Γ 0 ) δ S G = 0 in a spontaneously magnetized state. After projection, the corresponding uniform transverse rotation is a zero mode of the inverse moment response, X − 1 ( q = 0 , Ω = 0 ) δ S M = 0 X^{-1}(q=0,\Omega=0)\delta_S M=0 X − 1 ( q = 0 , Ω = 0 ) δ S M = 0 . This is the correlated Goldstone identity, provided the self-energy and vertex are obtained consistently from the same spin-invariant functional. It does not follow from the prescribed two-vertex term in Eq. (145) alone; spin–orbit coupling or explicit symmetry breaking removes the premise. At nonzero Ω \Omega Ω , the full frequency-dependent vertex and the response in Eq. (157) are retained rather than replacing them by the static Ward identity. Symmetry zero modes require the same restricted-inverse treatment as in the static case. Sec. VII.E uses Ω \Omega Ω for energy; the retarded response follows by analytic continuation to Ω + i 0 + \Omega+i0^+ Ω + i 0 + .
Writing M i α ( τ ) = M i α ( 0 ) + δ M i α ( τ ) M_{i\alpha}(\tau)=M_{i\alpha}^{(0)}+\delta M_{i\alpha}(\tau) M i α ( τ ) = M i α ( 0 ) + δ M i α ( τ ) , the quadratic dimensionless effective action is
S M ( 2 ) = β 2 ∑ n , i , j , α , β δ M i α ( − n ) X i α , j β − 1 ( i ℏ ν n ) δ M j β ( n ) . (160) S_M^{(2)}=\frac{\beta}{2}\sum_{n,i,j,\alpha,\beta}\delta M_{i\alpha}(-n)X^{-1}_{i\alpha,j\beta}(i\hbar\nu_n)\delta M_{j\beta}(n). \tag*{(160)} S M ( 2 ) = 2 β n , i , j , α , β ∑ δ M i α ( − n ) X i α , j β − 1 ( i ℏ ν n ) δ M j β ( n ) . (160)
The factor β \beta β in Eq. (160) comes from the Fourier convention. At zero frequency, the quadratic action reduces to the static constrained-energy curvature multiplied by β \beta β . At finite frequency, the response retains memory of earlier times and cannot be treated as an instantaneous Heisenberg exchange. For finite stationary moments, δ M i = M i ( 0 ) δ e i \delta M_i=M_i^{(0)}\delta e_i δ M i = M i ( 0 ) δ e i recovers the inverse-response contribution to Eq. (154) ; nonstationary constrained references also require the geometric term.
At zero equilibrium magnetization, M i ( 0 ) = 0 M_i^{(0)}=0 M i ( 0 ) = 0 , the angular rotation of the expectation-value moment is undefined, although ⟨ M ^ i 2 ⟩ \langle\hat{M}_i^2\rangle ⟨ M ^ i 2 ⟩ may remain finite. Thus Eq. (155) cannot be continued as an angular exchange mapping to this limit. The dynamical magnetic susceptibility remains meaningful wherever its inverse is defined. Non-Gaussian fluctuations require higher derivatives of the magnetic effective action, treated consistently with the quadratic kernel to avoid double counting.
A frequency-dependent self-energy links two different times. A prescribed rotation must therefore act at both ends of this link:
Σ θ ( τ 1 , τ 2 ) = U ( τ 1 ) Σ ( τ 1 − τ 2 ) U † ( τ 2 ) (161) \Sigma^\theta(\tau_1,\tau_2)=U(\tau_1)\Sigma(\tau_1-\tau_2)U^\dagger(\tau_2) \tag*{(161)} Σ θ ( τ 1 , τ 2 ) = U ( τ 1 ) Σ ( τ 1 − τ 2 ) U † ( τ 2 ) (161)
The first-order vertex then connects two electronic frequencies:
V a ( 1 ) ( z + , z − ) = − i 2 [ σ a Σ ( z − ) − Σ ( z + ) σ a ] , V_a^{(1)}(z_+,z_-)=-\frac{i}{2}\left[\sigma_a\Sigma(z_-)-\Sigma(z_+)\sigma_a\right], V a ( 1 ) ( z + , z − ) = − 2 i [ σ a Σ ( z − ) − Σ ( z + ) σ a ] ,
z + − z − = Ω (162) z_+-z_-=\Omega \tag*{(162)} z + − z − = Ω (162)
In Eqs. (162) –(164) , Ω = z + − z − \Omega=z_+-z_- Ω = z + − z − is an energy, not an angular frequency; Ω = ℏ ω \Omega=\hbar\omega Ω = ℏ ω . With z + = z + Ω / 2 z_+=z+\Omega/2 z + = z + Ω/2 and z − = z − Ω / 2 z_-=z-\Omega/2 z − = z − Ω/2 , expansion in Ω \Omega Ω gives
V a ( 1 ) ( z + , z − ) = − i 2 [ σ a , Σ ( z ) ] + i Ω 4 { σ a , ∂ Σ ( z ) ∂ z } + O ( Ω 2 ) (163) V_a^{(1)}(z_+,z_-)=-\frac{i}{2}[\sigma_a,\Sigma(z)]
+\frac{i\Omega}{4}\left\{\sigma_a,\frac{\partial\Sigma(z)}{\partial z}\right\}+O(\Omega^2) \tag*{(163)} V a ( 1 ) ( z + , z − ) = − 2 i [ σ a , Σ ( z )] + 4 i Ω { σ a , ∂ z ∂ Σ ( z ) } + O ( Ω 2 ) (163)
At zero frequency, Eq. (163) reduces to the static vertex in Eq. (144) . The term proportional to Ω ∂ Σ / ∂ z \Omega\partial\Sigma/\partial z Ω ∂ Σ/ ∂ z arises because the self-energy depends on time; it vanishes for a frequency-independent self-energy. The two-first-order-vertex contribution to the quadratic electronic response has the structure
W 2 v ( 2 ) = 1 2 Tr [ G V ( 1 ) G V ( 1 ) ] (164) W_{2v}^{(2)}=\frac{1}{2}\operatorname{Tr}[G V^{(1)}G V^{(1)}] \tag*{(164)} W 2 v ( 2 ) = 2 1 Tr [ G V ( 1 ) G V ( 1 ) ] (164)
Equation (163) contains only the two-first-order-vertex contribution. The second-order vertex, induced self-energy, and spin-frame terms must also be included. The spin-wave frequency follows from that complete response, not from Eq. (164) alone.
DISCUSSION
The symmetry checks in Sec. II.B establish global rotational consistency and local-frame covariance of the complete PT Hessian without invoking the local force theorem. They do not imply equality of prescribed-field and relaxed constrained-moment curvatures.
The off-site-block projection in Sec. III.A provides a scalar Heisenberg fit and its unreconstructed transverse curvature. The quadratic example shows that a higher-order, rotationally invariant interaction can produce reference-dependent apparent exchange and a DMI-like Hessian contribution without genuine chiral coupling, consistent with the issue discussed by dos Santos Dias et al. [25 ] ; its residual can even vanish in a collinear reference. The three-spin example instead produces a nonzero end-spin mixed curvature without a direct bilinear end-spin bond, and a nonzero block cannot be represented exactly by a scalar Heisenberg coupling. These examples demonstrate why neither a vanishing local residual nor an apparent pair coefficient establishes a transferable microscopic bilinear interaction. The numerical residual bound and its scope are given in Sec. III.A; the reference-state and local-to-global issues are also relevant to Cardias et al. [27 ] .
Finite-angle total energies and comparisons across varied reference states can help constrain a specified bilinear, biquadratic, or multisite interaction basis, subject to its completeness and the independence of the tested configurations. Such validation is particularly important for calculating Curie temperatures and magnetic phase diagrams, which depend on magnetic interactions beyond the leading small-q q q expansion. For a Curie-temperature calculation using an effective Heisenberg model, exchange parameters evaluated consistently in a disordered-local-moment (DLM) reference can be preferable to transferring ordered-ferromagnet parameters, provided that the reference retains meaningful local moments and the mapping is validated. A configuration-independent bilinear mapping is most plausible for localized, sufficiently rigid moments. In itinerant magnets, moment amplitudes, electronic propagation, and effective interactions can change with configuration. The inverse static susceptibility determines the local constrained-moment curvature, while total-energy differences provide complementary information over finite rotations.
Beyond the local Hessian-to-spin-model mapping, the electronic interaction itself may be nonlocal. The stationary nonlocal extension in Sec. III E requires the interaction kernel as well as the equilibrium nonlocal potential; Appendix A demonstrates the distinction in an analytically tractable two-site model. The two-electron example also shows that the reference-state bond operator can be spin scalar even though the orbital-nonlocal functional generates spin-dependent induced vertices under nonuniform rotations. This distinction is also important when interpreting the exchange extracted from the two-site example: the calculated J A B M J^{M}_{AB} J A B M is an AF reference-state curvature, not an effective exchange inferred from the FM–AF total-energy difference; the two coincide in this model but need not coincide in general.
The distinction between the equilibrium nonlocal potential and its induced interaction kernel also matters in the long-wavelength limit, where spatial dispersion can affect the leading rotational curvature. For a local, wave-vector-independent response kernel, the FM analysis in § 0.5 gives equal leading static wave-vector coefficients under the assumptions stated there. The spatially dispersive kernel in Eq. (89) shows why this result is not general. Neither statement establishes equality of finite-wave-vector pair interactions. Dynamical normalization remains essential for spin-wave stiffness.
For antiferromagnets, the problem is different because long-wavelength dynamics involve coupled sublattice rotations. The Fourier-origin-invariant velocity formula of Sec. VI follows from the specified sublattice Hessian and adiabatic precessional dynamics. Its identification with the fully relaxed spin-wave velocity requires the corresponding constrained-moment response.
sponding constrained-moment response.
More generally, the wave-vector and frequency dependence of the transverse magnetic response also contains contributions associated with spin-current transport. Describing this response through electronic susceptibility does not discard spin-current effects: we include their contribution to the extent captured by the electronic Hamiltonian and the response approximation employed. An explicit spin-current formulation with different constraints or electronic responses can double-count contributions. We therefore derive the static interactions and leading spin-wave coefficients from the magnetic response without introducing an additional independent spin-current variable; this does not establish equivalence to a separately truncated spin-current formulation.
Relativistic Dzyaloshinskii–Moriya interactions require separate assessment, since the present calculations exclude SOC. In such relativistic calculations, an SOC-induced odd-in-q q q response can represent a genuine chiral contribution. Identifying an exchange-field rotation coefficient with the DMI of constrained moments would still require the relativistic response and field-to-moment mapping; the resulting errors are not quantified here. The favorable nonrelativistic stiffness result therefore cannot be extended to DMI without further analysis.
A separate extension concerns correlated electronic systems with frequency-dependent self-energies. Rotating a specified self-energy at two times in Sec. VII defines an electronic vertex, not the full response of relaxed moments. Determining collective spin dynamics requires the induced self-energy and interaction vertex, consistent transformation of the electronic action, and spin-conservation identities. Accordingly, DFT-derived long-wavelength error estimates cannot be transferred to correlated systems without a separate analysis.
An angular description requires a finite local magnetic amplitude on the relevant time scale, not necessarily a nonzero ensemble-averaged magnetization. Our rotational Hessians refer to specified finite-moment configurations; applying them to disordered states requires an appropriate reference ensemble, rather than setting the ordered moment to zero. When longitudinal and transverse fluctuations strongly mix, a fixed-length spin model need not reproduce the full response. If the moment amplitude vanishes, its orientation is undefined and the dynamical magnetic susceptibility provides the appropriate quadratic description. Higher-order fluctuations then require terms beyond the Gaussian action.
CONCLUSIONS
We calculated the energy change caused by small rotations of the exchange field using second-order perturbation theory at the full exchange splitting. The calculation includes the local second-order contribution and the intersite response described by Green functions. Rotations of the magnetic moments instead give an energy curvature determined by the inverse static susceptibility, including geometric terms where needed. The two curvatures generally differ beyond the leading long-wavelength order.
For ferromagnets, the leading static terms proportional to q 2 q^2 q 2 agree under the assumptions of a local, wave-vector-independent response kernel, but this agreement can be lost when the interaction kernel depends on wave vector. The spin-wave stiffness also depends on the frequency-linear part of the magnetic response. For two-sublattice antiferromagnets, we derived the long-wavelength spin-wave velocity from the sublattice energy curvature and assumed adiabatic precessional dynamics. The contribution from the relative positions of the magnetic sublattices makes the result independent of the choice of Fourier origin. A general equivalence between exchange-field and constrained-moment AFM coefficients has not been established.
We also developed a practical method for obtaining effective Heisenberg exchange parameters from the transverse energy curvature, together with a measure of how well the Heisenberg model reproduces that curvature. Examples with biquadratic and three-spin interactions show that higher-order magnetic interactions can either reproduce the local behavior of a Heisenberg model or generate apparent pair interactions that cannot be fully described by such a model. The resulting exchange parameters can depend on the magnetic reference state and need not reproduce finite-angle energies. Planar and conical spin spirals provide additional tests of the Heisenberg description. Apparent DMI-like terms obtained without spin–orbit coupling are distinct from genuine relativistic DMI, which is outside the present treatment.
For nonlocal exchange functionals, the induced interaction kernel contributes to the constrained-moment curvature, as illustrated by the two-site model in Appendix A. In correlated systems, the stationary moment response includes the self-energy variation and its irreducible vertex; the inverse of this projected response determines the moment curvature. These results extend rotational PT beyond a fixed electronic potential without assuming a configuration-independent Heisenberg mapping.
Two-electron two-site test of stationary PT with nonlocal exchange
Functional and self-consistent antiferromagnetic state
We test the stationary nonlocal theory using two sites A A A , B B B , one spatial orbital per site, and two electrons. In the spin basis ( A ↑ , A ↓ , B ↑ , B ↓ ) (A \uparrow, A \downarrow, B \uparrow, B \downarrow) ( A ↑ , A ↓ , B ↑ , B ↓ ) , we define
E [ γ ] = Tr ( h 0 γ ) − U 4 ( ∣ M A ∣ 2 + ∣ M B ∣ 2 ) − K n l 2 tr s ( γ A B γ B A ) , (165) E[\gamma] = \operatorname{Tr}(h_0\gamma) - \frac{U}{4}\left(\lvert\mathbf{M}_A\rvert^2+\lvert\mathbf{M}_B\rvert^2\right) - \frac{K_{\mathrm{nl}}}{2}\operatorname{tr}_{\mathrm{s}}(\gamma_{AB}\gamma_{BA}), \tag*{(165)} E [ γ ] = Tr ( h 0 γ ) − 4 U ( ∣ M A ∣ 2 + ∣ M B ∣ 2 ) − 2 K nl tr s ( γ A B γ B A ) , (165)
h 0 = − t ∑ s ( ∣ A s ⟩ ⟨ B s ∣ + ∣ B s ⟩ ⟨ A s ∣ ) , M i = tr s ( σ γ i i ) . h_0=-t\sum_s\left(\lvert As\rangle\langle Bs\rvert+\lvert Bs\rangle\langle As\rvert\right),
\qquad
\mathbf{M}_i=\operatorname{tr}_{\mathrm{s}}(\sigma\gamma_{ii}). h 0 = − t s ∑ ( ∣ A s ⟩ ⟨ B s ∣ + ∣ B s ⟩ ⟨ A s ∣ ) , M i = tr s ( σ γ ii ) .
The density matrix γ \gamma γ describes two occupied one-electron states and satisfies γ 2 = γ \gamma^2=\gamma γ 2 = γ . We use this rotationally invariant model to illustrate nonlocal exchange, without identifying it with a particular exchange-correlation approximation. The resulting one-electron Hamiltonian contains an exchange term connecting sites A A A and B B B :
h i i = − U 2 M i ⋅ σ , h_{ii}=-\frac{U}{2}\mathbf{M}_i\cdot\boldsymbol{\sigma}, h ii = − 2 U M i ⋅ σ ,
h A B = − t 1 − K n l 2 γ A B , h B A = h A B † . (166) h_{AB}=-t\mathbf{1}-\frac{K_{\mathrm{nl}}}{2}\gamma_{AB},
\qquad
h_{BA}=h_{AB}^{\dagger}. \tag*{(166)} h A B = − t 1 − 2 K nl γ A B , h B A = h A B † . (166)
For t > 0 t>0 t > 0 and d ≡ 2 U − K n l > 4 t d\equiv2U-K_{\mathrm{nl}}>4t d ≡ 2 U − K nl > 4 t , we consider the collinear AF determinant with occupied orbitals
∣ h 1 ⟩ = c ∣ A ↑ ⟩ + a ∣ B ↑ ⟩ , ∣ h 2 ⟩ = a ∣ A ↓ ⟩ + c ∣ B ↓ ⟩ , a 2 + c 2 = 1. (167) \lvert h_1\rangle=c\lvert A\uparrow\rangle+a\lvert B\uparrow\rangle,
\qquad
\lvert h_2\rangle=a\lvert A\downarrow\rangle+c\lvert B\downarrow\rangle,
\qquad
a^2+c^2=1. \tag*{(167)} ∣ h 1 ⟩ = c ∣ A ↑ ⟩ + a ∣ B ↑ ⟩ , ∣ h 2 ⟩ = a ∣ A ↓ ⟩ + c ∣ B ↓ ⟩ , a 2 + c 2 = 1. (167)
Its local moments are M A = m z ^ \mathbf{M}_A=m\hat{\mathbf{z}} M A = m z ^ and M B = − m z ^ \mathbf{M}_B=-m\hat{\mathbf{z}} M B = − m z ^ , where m = c 2 − a 2 > 0 m=c^2-a^2>0 m = c 2 − a 2 > 0 . In this AF state, the intersite density matrix is γ A B = a c 1 \gamma_{AB}=ac\mathbf{1} γ A B = a c 1 . The nonlocal exchange term therefore changes the effective hopping by − K n l a c / 2 -K_{\mathrm{nl}}ac/2 − K nl a c /2 . It is spin independent in the collinear reference state, but its induced response to nonuniform spin rotations is spin dependent. The total energy is
E A F ( a ) = − 4 t a c − U 2 ( c 2 − a 2 ) 2 − K a 2 c 2 . (168) E_{\mathrm{AF}}(a)=-4tac-\frac{U}{2}(c^2-a^2)^2-Ka^2c^2. \tag*{(168)} E AF ( a ) = − 4 t a c − 2 U ( c 2 − a 2 ) 2 − K a 2 c 2 . (168)
Setting x = 2 a c x=2ac x = 2 a c reduces this expression to
E A F ( x ) = − U 2 − 2 t x + d 4 x 2 , E_{\mathrm{AF}}(x)=-\frac{U}{2}-2tx+\frac{d}{4}x^2, E AF ( x ) = − 2 U − 2 t x + 4 d x 2 ,
x ∗ = 4 t d , m = 1 − 16 t 2 d 2 . (169) x_*=\frac{4t}{d},
\qquad
m=\sqrt{1-\frac{16t^2}{d^2}}. \tag*{(169)} x ∗ = d 4 t , m = 1 − d 2 16 t 2 . (169)
At the self-consistent minimum, the AF energy is E A F = − U / 2 − 4 t 2 / d E_{\mathrm{AF}}=-U/2-4t^2/d E AF = − U /2 − 4 t 2 / d . The fully polarized FM determinant has E F M = − U / 2 E_{\mathrm{FM}}=-U/2 E FM = − U /2 , so their difference is − 4 t 2 / d -4t^2/d − 4 t 2 / d . This finite-angle energy difference is not used to define the local pair exchange. In this particular model the two definitions nevertheless coincide: a bilinear pair energy − J A B e A ⋅ e B -J_{AB}\mathbf{e}_A\cdot\mathbf{e}_B − J A B e A ⋅ e B gives E F M − E A F = − 2 J A B E_{\mathrm{FM}}-E_{\mathrm{AF}}=-2J_{AB} E FM − E AF = − 2 J A B , and the difference 4 t 2 / d 4t^2/d 4 t 2 / d yields the same J A B = − 2 t 2 / d J_{AB}=-2t^2/d J A B = − 2 t 2 / d obtained below from the local curvature. The model therefore illustrates the stationary nonlocal formalism but does not discriminate between the two definitions.
Transverse particle–hole Hessian and constrained exchange
The unoccupied eigenstates can be chosen as
∣ p 1 ⟩ = − a ∣ A ↑ ⟩ + c ∣ B ↑ ⟩ , ∣ p 2 ⟩ = c ∣ A ↓ ⟩ − a ∣ B ↓ ⟩ . (170) |p_1\rangle= -a|A\uparrow\rangle+ c|B\uparrow\rangle,\qquad|p_2\rangle= c|A\downarrow\rangle- a|B\downarrow\rangle. \tag*{(170)} ∣ p 1 ⟩ = − a ∣ A ↑ ⟩ + c ∣ B ↑ ⟩ , ∣ p 2 ⟩ = c ∣ A ↓ ⟩ − a ∣ B ↓ ⟩ . (170)
A small transverse rotation mixes occupied and empty states. For one real transverse direction, we write the first-order change in the density matrix as
δ γ = z 1 ( ∣ p 2 ⟩ ⟨ h 1 ∣ + h . c . ) + z 2 ( ∣ p 1 ⟩ ⟨ h 2 ∣ + h . c . ) . (171) \delta\gamma= z_1(|p_2\rangle\langle h_1|+\mathrm{h.c.}) + z_2(|p_1\rangle\langle h_2|+\mathrm{h.c.}). \tag*{(171)} δ γ = z 1 ( ∣ p 2 ⟩ ⟨ h 1 ∣ + h.c. ) + z 2 ( ∣ p 1 ⟩ ⟨ h 2 ∣ + h.c. ) . (171)
The projected transverse moments are
δ M A x = 2 ( c 2 z 1 − a 2 z 2 ) , δ M B x = 2 ( − a 2 z 1 + c 2 z 2 ) . (172) \delta M_A^x = 2(c^2z_1-a^2z_2),\qquad\delta M_B^x = 2(-a^2z_1+c^2z_2). \tag*{(172)} δ M A x = 2 ( c 2 z 1 − a 2 z 2 ) , δ M B x = 2 ( − a 2 z 1 + c 2 z 2 ) . (172)
Expanding the total energy to second order, including the occupied–empty excitation energies and the complete local and intersite interaction responses, gives
E ⊥ ( 2 ) = d a 2 c 2 ( z 1 + z 2 ) 2 = 1 2 ( z 1 z 2 ) 2 d a 2 c 2 ( 1 1 1 1 ) ⏟ A ⊥ ( z 1 z 2 ) . (173) E_\perp^{(2)}=da^2c^2(z_1+z_2)^2=\frac{1}{2}\begin{pmatrix}z_1 & z_2\end{pmatrix}\underbrace{2da^2c^2\begin{pmatrix}1&1\\1&1\end{pmatrix}}_{A_\perp}\begin{pmatrix}z_1\\z_2\end{pmatrix}. \tag*{(173)} E ⊥ ( 2 ) = d a 2 c 2 ( z 1 + z 2 ) 2 = 2 1 ( z 1 z 2 ) A ⊥ 2 d a 2 c 2 ( 1 1 1 1 ) ( z 1 z 2 ) . (173)
This follows by expanding γ = e Ω γ 0 e − Ω \gamma=e^\Omega\gamma_0e^{-\Omega} γ = e Ω γ 0 e − Ω , where Ω \Omega Ω is anti-Hermitian and its occupied–empty components are set by ( z 1 , z 2 ) (z_1,z_2) ( z 1 , z 2 ) . The second-order orbital mixing gives the excitation-energy term, while the interaction energy supplies both response kernels; no frozen-band-energy approximation is used. A uniform spin rotation costs no energy: the Hessian has a zero mode ( z 1 , z 2 ) ∝ ( 1 , − 1 ) (z_1,z_2)\propto(1,-1) ( z 1 , z 2 ) ∝ ( 1 , − 1 ) and a second eigenvalue 4 d a 2 c 2 4da^2c^2 4 d a 2 c 2 . The other transverse polarization has the corresponding zero mode.
For the AF unit-vector angular coordinates, δ M A x = m θ A \delta M_A^x=m\theta_A δ M A x = m θ A and δ M B x = − m θ B \delta M_B^x=-m\theta_B δ M B x = − m θ B . Equation (A8) implies
z 1 + z 2 = θ A − θ B 2 , E c o n s t r ( 2 ) = t 2 d ( θ A − θ B ) 2 . (174) z_1+z_2=\frac{\theta_A-\theta_B}{2},\qquad E_{\mathrm{constr}}^{(2)}=\frac{t^2}{d}(\theta_A-\theta_B)^2. \tag*{(174)} z 1 + z 2 = 2 θ A − θ B , E constr ( 2 ) = d t 2 ( θ A − θ B ) 2 . (174)
The expansion of − J A B e A ⋅ e B -J_{AB}\mathbf{e}_A\cdot\mathbf{e}_B − J A B e A ⋅ e B around antiparallel moments is − J A B ( θ A − θ B ) 2 / 2 -J_{AB}(\theta_A-\theta_B)^2/2 − J A B ( θ A − θ B ) 2 /2 , giving
J A B M = − 2 t 2 2 U − K n l , J A B M ( K n l = 0 ) = − t 2 U . (175) J_{AB}^{M}=-\frac{2t^2}{2U-K_{\mathrm{nl}}},\qquad J_{AB}^{M}(K_{\mathrm{nl}}=0)=-\frac{t^2}{U}. \tag*{(175)} J A B M = − 2 U − K nl 2 t 2 , J A B M ( K nl = 0 ) = − U t 2 . (175)
This is a local harmonic exchange coefficient of the stationary AF state, not an assertion of a globally transferable Heisenberg interaction.
Independent finite-angle and stability checks
We check the result independently by rotating the occupied orbitals in Eq. (167) so that the two AF moments turn toward one another by a relative angle δ \delta δ . The local moment magnitudes and the bond invariant tr s ( γ A B γ B A ) \operatorname{tr}_s(\gamma_{AB}\gamma_{BA}) tr s ( γ A B γ B A ) remain fixed under this path. The hopping energy changes according to
E p a t h ( δ ) = − 4 t a c cos ( δ / 2 ) − U 2 m 2 − K a 2 c 2 , E_{\mathrm{path}}(\delta)=-4tac\cos(\delta/2)-\frac{U}{2}m^2-Ka^2c^2, E path ( δ ) = − 4 t a c cos ( δ /2 ) − 2 U m 2 − K a 2 c 2 ,
E p a t h ( δ ) − E A F = 4 t a c [ 1 − cos ( δ / 2 ) ] = t 2 d δ 2 + O ( δ 4 ) . (176) E_{\mathrm{path}}(\delta)-E_{\mathrm{AF}}=4tac[1-\cos(\delta/2)]=\frac{t^2}{d}\delta^2+O(\delta^4). \tag*{(176)} E path ( δ ) − E AF = 4 t a c [ 1 − cos ( δ /2 )] = d t 2 δ 2 + O ( δ 4 ) . (176)
This rotated state is a variational path, not necessarily the lowest constrained-energy state at finite δ \delta δ . For small rotations, however, the two transverse amplitudes in Eq. (171) are fixed by the moment directions. The other particle–hole channels describe longitudinal or charge changes to first order and do not alter the quadratic transverse energy when the AF reference is stable. The finite-angle result thus checks the constrained local curvature in Eq. (174) .
At ( t , U , K n l ) = ( 1 , 4 , 1 ) (t,U,K_{\mathrm{nl}})=(1,4,1) ( t , U , K nl ) = ( 1 , 4 , 1 ) , numerical differentiation of the full eight-dimensional real particle–hole Hessian gives two Goldstone zero modes and six positive eigenvalues, confirming local AF stability at this parameter set, but not throughout the existence region d > 4 t d>4t d > 4 t .
The model gives a K n l K_{\mathrm{nl}} K nl -dependent constrained exchange curvature. Stationary PT agrees with the finite-angle expansion and passes the Goldstone and stability tests. Its two transverse amplitudes are fixed by the moment constraints, so it does not independently validate relaxation among extra unconstrained transverse channels; that would require additional orbitals or sites. In particular, Eq. (175) must not be extrapolated to a general nonlocal functional or interpreted as the prescribed-field pair coefficient of Sec. III.E [Eq. (80)].
Acknowledgments
Work at Ames National Laboratory was supported by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division. Ames National Laboratory is operated for the U.S. DOE by Iowa State University under Contract No. DE-AC02-07CH11358.