Non-Adiabatic Spin Transfer Torque in Real Materials

Ion Garate1 , K. Gilmore2,3 , M. D. Stiles2 , and A.H. MacDonald1

1
Department of Physics, The University of Texas at Austin, Austin, TX 78712
2
Center for Nanoscale Science and Technology, National Institute
of Standards and Technology, Gaithersburg, MD 20899-8412 and
3
Maryland NanoCenter, University of Maryland, College Park, MD, 20742
(Dated: December 15, 2008)
The motion of simple domain walls and of more complex magnetic textures in the presence of a
transport current is described by the Landau-Lifshitz-Slonczewski (LLS) equations. Predictions of
the LLS equations depend sensitively on the ratio between the dimensionless material parameter
β which characterizes non-adiabatic spin-transfer torques and the Gilbert damping parameter α.
This ratio has been variously estimated to be close to 0, close to 1, and large compared to 1. By
identifying β as the influence of a transport current on α, we derive a concise, explicit and relatively
simple expression which relates β to the band structure and Bloch state lifetimes of a magnetic
metal. Using this expression we demonstrate that intrinsic spin-orbit interactions lead to intra-
band contributions to β which are often dominant and can be (i) estimated with some confidence
and (ii) interpreted using the “breathing Fermi surface” model.

PACS numbers:

I. INTRODUCTION βs0 × (vs · ∇)s0 , acts in the perpendicular direction and

is frequently referred to as the non-adiabatic torque.9
Unfortunately, the name is a misnomer in the present
An electric current can influence the magnetic state of
context. There are two contributions that have the pre-
a ferromagnet by exerting a spin transfer torque(STT) on
ceding form. The first is truly non-adiabatic and occurs
the magnetization.1–3 This effect occurs whenever cur-
in systems in which the magnetization varies too rapidly
rents travel through non-collinear magnetic systems and
in space for the spins of the transport electrons to fol-
is therefore promising for magnetoelectronic applications.
low the local magnetization direction as they traverse
Indeed, STT’s have already been exploited in a number of
the magnetization texture. For wide domain walls, these
technological devices.4 Partly for this reason and partly
effects are expected to be small.10 The contribution of
because the quantitative description of order parameter
interest in this paper is a dissipative contribution that
manipulation by out-of-equilibrium quasiparticles poses
occurs in the adiabatic limit. The adiabatic torque dis-
great theoretical challenges, the study of the STT effect
cussed above is the reactive contribution in this limit. As
has developed into a major research subfield of spintron-
we discuss below, processes that contribute to magnetic
ics.
damping, whether they derive from spin-orbit coupling
Spin transfer torques are important in both magnetic or spin-dependent scattering, also give a spin-transfer
multilayers, where the magnetization changes abruptly,5 torque parameterized by β as above. In this paper, we
and in magnetic nanowires, where the magnetization follow the common convention and refer to this torque as
changes smoothly.6 Theories of the STT in systems with non-adiabatic. However, it should be understood that it
smooth magnetic textures identify two different types of is a dissipative spin transfer torque that is present in the
spin transfer. On one hand, the adiabatic or Slonczewski3 adiabatic limit.
torque results when quasiparticle spins follow the under- The non-adiabatic torque plays a key role in current-
lying magnetic landscape adiabatically. It can be math- driven domain wall dynamics, where the ratio between
ematically expressed as (vs · ∇)s0 , where s0 stands for β and the Gilbert parameter α can determine the veloc-
the magnetization and vs is the “spin velocity”, which ity of domain walls under the influence of a transport
is proportional to the charge drift velocity, and hence to current. There is no consensus on its magnitude of the
the current and the applied electric field. The micro- parameter β.6,11 Although there have a few theoretical
scopic physics of the Slonczewski spin-torque is thought studies12–14 of the STT in toy models, the relationship
to be well understood5–7 , at least8 in systems with weak between toy model STT’s and STT’s in either transition
spin-orbit coupling. A simple angular momentum conser- metal ferromagnets or ferromagnetic semiconductors is
vation argument argues that in the absence of spin-orbit far from clear. As we will discuss the toy models most
coupling vs = σs E/es0 , where s0 is the magnetization, σs often studied neglect spin-orbit interactions in the band-
is the spin conductivity and E is the electric field. How- structure of the perfect crystal, intrinsic spin-orbit inter-
ever, spin-orbit coupling plays an essential role in real actions, which can alter STT physics qualitatively.
magnetic materials and hence the validity of this sim- The main objectives of this paper are (i) to shed new
ple expression for vs needs to be tested by more rigorous light on the physical meaning of the non-adiabatic STT
calculations. by relating it to the change in magnetization damping
The second spin transfer torque in continuous media, due to a transport current, (ii) to derive a concise for-
 2

mula that can be used to evaluate β in real materials in approximately the same way as the ratio between the
from first principles and (iii) to demonstrate that α and itinerant spin density and the total spin density. We also
β have the same qualitative dependence on disorder (or suggest that α and β may have the opposite signs in sys-
temperature), even though their ratio depends on the tems with both hole-like and electron-like carriers. We
details of the band structure. As a byproduct of our the- present concrete results for (Ga,Mn)As, where we obtain
oretical study, we find that the expression for vs in terms α/β ' 0.1. This is reasonable in view of the weak spin
of the spin conductivity may not always be accurate in polarization and the strong spin-orbit coupling of valence
materials with strong spin-orbit coupling. band holes in this material.
We begin in Section II by reviewing and expanding Section VII describes the generalization of the torque-
on microscopic theories of α, β and vs . In short, our correlation formula employed in ab-initio calculations of
microscopic approach quantifies how the micromagnetic the Gilbert damping to the case of the non-adiabatic
energy of an inhomogeneous ferromagnet is altered in spin-transfer torque. The torque correlation formula in-
response to external rf fields and dc transport currents corporates scattering of quasiparticles simply by intro-
which drive the magnetization direction away from lo- ducing a phenomenological lifetime for the Bloch states
cal equilibrium. These effects are captured by the spin and assumes that the most important electronic transi-
transfer torques, damping torques, and effective magnetic tions occur between states near the Fermi surface in the
fields that appear in the LLS equation. By relating mag- same band. Our ability to make quantitative predictions
netization dynamics to effective magnetic fields, we de- based on this formula is limited mainly by an incomplete
rive explicit expressions for α,β and vs in terms of mi- understanding of Bloch state scattering processes in real
croscopic parameters. Important contributions to these ferromagnetic materials. These simplifications give rise
materials parameters can be understood in clear physical to ambiguities and inaccuracies that we dissect in Section
terms using the breathing Fermi surface model.15 Read- VII. Our assessment indicates that the torque correlation
ers mainly interested in a qualitative explanation for our formula for β is most accurate at low disorder and rela-
findings may skip directly to Section VIII where we dis- tively weak spin-orbit interactions.
cuss of our main results in that framework. Regardless Section VIII restates and complements the effective
of the approach, the non-adiabatic STT can be under- field calculation explained in Section II. Within the adi-
stood as the change in the Gilbert damping contribution abatic approximation, the instantaneous energy of a fer-
to magnetization dynamics when the Fermi sea quasi- romagnet may be written in terms of the instantaneous
particle distribution function is altered by the transport occupation factors of quasiparticle states. We determine
electric field. The outcome of this insight is a concise an- the effect of the external perturbations on the occupation
alytical formula for β which is simple enough that it can factors by combining the relaxation time approximation
be conveniently combined with first-principles electronic and the master equation. In this way we recover the re-
structure calculations to predict β-values in particular sults of Section II and are able to interpret the intra-band
materials.16 contributions to β in terms of a generalized breathing
In Sections III, IV and V we apply our expression for β Fermi surface picture.
to model ferromagnets. In Section III we perform a nec- Section IX contains a brief summary which concludes
essary reality check by applying our theory of β to the this work.
parabolic band Stoner ferromagnet, the only model for
which more rigorous fully microscopic calculations13,14
of β have been completed. Section IV is devoted to
II. MICROSCOPIC THEORY OF α, β AND vs
the study of a two-dimensional electron-gas ferromag-
net with Rashba spin-orbit interactions. Studies of this
model provide a qualitative indication of the influence The Gilbert damping parameter α, the non-adiabatic
of intrinsic spin-orbit interactions on the non-adiabatic spin transfer torque coefficient β and the “spin velocity”
STT. We find that, as in the microscopic theory17,18 vs appear in the generalized Landau-Lifshitz-Gilbert ex-
for α, spin-orbit interactions induce intra-band contri- pression for collective magnetization dynamics of a fer-
butions to β which are proportional to the quasiparticle romagnet under the influence of an electric current:
lifetimes. These considerations carry over to the more
sophisticated 4-band spherical model that we analyze in (∂t + vs · ∇) Ω̂ − Ω̂ × Heff = −αΩ̂ × ∂t Ω̂ − β Ω̂ × (vs · ∇)Ω̂.
Section V; there our calculation is tailored to (Ga,Mn)As. (1)
We show that intra-band (conductivity-like) contribu- In Eq. (1) Heff is an effective magnetic field which we
tions are prominent in the 4-band model for experimen- elaborate on below and Ω̂ = s0 /s0 ' (Ωx , Ωy , 1 − (Ω2x +
tally relevant scattering rates. Ω2y )/2) is the direction of the magnetization.19 Eq. (1) de-
Section VI discusses the phenomenologically important scribes the slow dynamics of smooth magnetization tex-
α/β ratio for real materials. Using our analytical results tures in the presence of a weak electric field which induces
derived in Section II (or Section VIII) we are able to re- transport currents. It explicitly neglects the dynamics of
produce and extrapolate trends expected from toy mod- the magnetization magnitude which is implicitly assumed
els which indicate that α/β should vary across materials to be negligible. For small deviations from the easy di-
 3

rection (which we take to be the ẑ-direction), it reads α and β we refer to Section VIII. The transverse spin re-
sponse function studied here describes the change in the
Heff,x = (∂t + vs · ∇) Ωy + (α∂t + βvs · ∇) Ωx micromagnetic energy due to the departure of the magne-
Heff,y = − (∂t + vs · ∇) Ωx + (α∂t + βvs · ∇) Ωy (2) tization away from its equilibrium direction, where equi-
librium is characterized by the absence of currents and
The gyromagnetic ratio has been absorbed into the units external rf fields. This change in energy defines an ef-
of the field Heff so that this quantity has inverse time fective magnetic field which may then be identified with
units. We set ~ = 1 throughout. Eq. (2), thereby allowing us to microscopically determine
In this section we relate the α, β and vs parameters α,β and vs . To first order in frequency ω, wave vector q
to microscopic features of the ferromagnet by consider- and electric field, the transverse spin response function
ing the transverse total spin response function. For a is given by
technically more accessible (yet less rigorous) theory of

X X h (0) (1) (2)

i
S0 Ω̂a = χa,b Hext,b ' χa,b + ωχa,b + (vs · q)χa,b Hext,b (3)
b b

where a, b ∈ {x, y}, Hext is the external magnetic field with frequency ω and wave vector q, S0 = s0 V is the total
spin of the ferromagnet (V is the sample volume), and χ is the transverse spin-spin response function in the presence
of a uniform time-independent electric field:
Z ∞ Z
£ ¤
χa,b (q, ω; vs ) = i dt dr exp(iωt − iq · r)h S a (r, t), S b (0, 0) i. (4)
0

In Eq. (3), χ(0) = χ(q = 0, ω = 0; E = 0) de- −(1/S0 )∂Eint /∂ Ω̂a = −S0 [χ(0) ]−1
a,b Ω̂b , where Hint is the
scribes the spin response to a constant, uniform ex- internal energy contribution to the effective magnetic
ternal magnetic field in absence of a current, χ(1) = field. Multiplying Eq. (3) on the left by [χ(0) ]−1 and
limω→0 χ(q = 0, ω; E = 0)/ω characterizes the spin re- using Heff = Hint + Hext we obtain a formal equation for
sponse to a time-dependent, uniform external mag- Heff :
netic field in absence of a current, and χ(2) = X h (1) i
limq,vs →0 χ(q, ω = 0; E)/q · vs represents the spin re- (2)
Heff,a = La,b ∂t + La,b (vs · ∇) Ω̂b , (5)
sponse to a constant, non-uniform external magnetic field b
combined with a constant, uniform electric field E . Note
that first order terms in q are allowed by symmetry in where
presence of an electric field. In addition, hi is a ther-
mal and quantum mechanical average over states that L(1) = −iS0 [χ(0) ]−1 χ(1) [χ(0) ]−1
describe a uniformly magnetized, current carrying ferro- L(2) = iS0 [χ(0) ]−1 χ(2) [χ(0) ]−1 . (6)
magnet.
The approach underlying Eq. (3) comprises a linear Identifying of Eqs. (5) and (2) results in concise micro-
response theory with respect to an inhomogeneous mag- scopic expressions for α and β and vs :
netic field followed by a linear response theory with re-
spect to an electric field. Alternatively, one may treat the α = L(1) (1)
x,x = Ly,y
electric and magnetic perturbations on an equal footing β = L(2) (2)
x,x = Ly,y
without predetermined ordering; for further considera- h i
tions on this matter we refer to Appendix A. 1 = L(2)
x,y =⇒ vs · q = iS0 (χ
(0) −1
) χ(χ(0) )−1 (. 7)
x,y
In the following we emulate and appropriately gen-
eralize a procedure outlined elsewhere.17 First, we rec- In the third line of Eq. (7) we have combined the second
ognize that in the static limit and in absence of a cur- line of Eq. (6) with χ(2) = χ/(vs · q).
rent the transverse magnetization responds to the exter- When applying Eq. (7) to realistic conducting fer-
nal magnetic field by adjusting its orientation to min- romagnets, one must invariably adopt a self-consistent
imize the total energy including the internal energy mean-field (Stoner) theory description of the magnetic
Eint and the energy due to coupling with the exter- state derived within a spin-density-functional theory
nal magnetic field, Eext = −S0 Ω̂ · Hext . It follows (SDFT) framework.20,21 In SDFT the transverse spin
(0)
that χa,b = S02 [∂ 2 Eint /∂ Ω̂a ∂ Ω̂b ]−1 and thus Hint,a = response function is expressed in terms of Kohn-Sham
 4

quasiparticle response to both external and induced mag- current. This argument is based on the assumption that
netic fields; this allows us to transform17 Eq. (7) into the (transverse) angular momentum lost by spin polar-
ized electrons traversing an inhomogeneous ferromagnet
QP
1 Im[χ̃+,− (q = 0, ω, E = 0)] is transferred to the magnetization. However, this as-
α = lim sumption fails when spin angular momentum is not con-
S0 ω→0 ω
served as it is not in the presence of spin-orbit coupling.
1 Im[χ̃QP
+,− (q, ω = 0, E)] In general, part of the transverse spin polarization lost
β = − lim
S0 vs ,q→0 q · vs by the current carrying quasiparticles is transferred to
1 QP the lattice rather than to collective magnetic degrees of
vs · q = − Re[χ̃+,− (q, ω = 0, E)], (8)
S0 freedom8 when spin-orbit interactions are present. It is
often stated that the physics of spin non-conservation is
(0) ¯ and
where we have used22 χa,b = δa,b S0 /∆ captured by the non-adiabatic STT; however, the non-
adiabatic STT per se is limited to dissipative processes
1X fj − fi and cannot describe the changes in the reactive spin
χ̃QP
+,− (q, ω; E) = torque due to spin-flip events. Our expression in terms
2 i,j ²i − ²j − ω − iη
of the transverse spin response function does not rely on
hj|S + ∆0 (r)eiq·r |iihi|S − ∆0 (r)e−iq·r |ji spin conservation, and while it agrees with the conven-
(9) tional picture24 in simplest cases (see below), it departs
from it when e.g. intrinsic spin-orbit interactions are
is the quasiparticle response to changes in the direction strong.
of the exchange-correlation effective magnetic field.23 To
In this paper we incorporate the influence of an electric
estimate β this response function should be evaluated
field by simply shifting the Kohn-Sham orbital occupa-
in the presence of an electric current. In the derivation
(1) tion factors to account for the energy deviation of the
of Eq. (8) we have made use of the fact that χx,x and distribution function in a drifting Fermi sea:
(2) (2)
χx,x are purely imaginary, whereas χx,y is purely real;
this can be verified mathematically through S ± = Sx ±
iSy . Physically, “Im” and “Re” indicate that the Gilbert fi ' f (0) (²i + Vi ) ' f (0) (²i ) + Vi ∂f (0) /∂²i (10)
damping and the non-adiabatic STT are dissipative while
the adiabatic STT is reactive. Furthermore, in the third
line it is implicit that we expand Re[χ̃QP ] to first order where Vi is the effective energy shift for the i-th eigenen-
in q and E. ergy due to acceleration between scattering events by an
In Eq. (9), S ± is the spin-rising/lowering operator, |ii, electric field and f (0) is the equilibrium Fermi factor.
²i and fi are the Kohn-Sham eigenstates, eigenenergies This approximation to the steady-state induced by an
and Fermi factors in presence of spin-dependent disorder, external electric field is known to be reasonably accurate
and ∆0 (r) is the difference in the magnetic ground state in many circumstances, for example in theories of electri-
between the majority spin and minority spin exchange- cal transport properties, and it can be used24 to provide
correlation potential - the spin-splitting potential. This a microscopic derivation of the adiabatic spin-transfer
quantity is always spatially inhomogeneous at the atomic torque. As we discuss below, this ansatz provides a result
scale and is typically larger in atomic regions than in for β which is sufficiently simple that it can be combined
interstitial regions. Although the spatial dependence of with realistic ab initio electronic structure calculations
∆0 (r) plays a crucial role in realistic ferromagnets, we to estimate β values in particular magnetic metals. We
replace it by a phenomenological constant ∆0 in the toy support this ansatz by demonstrating that it agrees with
models we discuss below. full non-linear response calculations in the case of toy
Our expression of vs in terms of the transverse spin models for which results are available.
response function may be unfamiliar to readers familiar
with the argument given in the introduction of this pa- Using the Cauchy identity, 1/(x − iη) = 1/x + iπδ(x),
per in which vs is determined by the divergence in spin and ∂f (0) /∂² ' −δ(²) we obtain

πX
Im[χ̃QP
+,− ] ' [ω − Vj,i ] |hj|S + ∆0 (r) eiq·r |ii|2 δ(²i − ²F ) δ(²j − ²F )
2 i,j
1X Vj δ(²j − ²F ) − Vi δ(²i − ²F )
Re[χ̃QP
+,− ] ' − |hj|S + ∆0 (r)eiq·r |ii|2 (11)
2 i,j ²i − ²j

where we have defined the difference in transport devia- tion energies by

Vj,i ≡ Vj − Vi . (12)
 5

In the first line of Eq. (11), the two terms within the part selects scattering processes that relax the spin of
square brackets correspond to the energy of particle- the particle-hole pairs mediated either by phonons or by
hole excitations induced by radio frequency magnetic magnetic impurities.25 Substituting Eq. (11) into Eq. (8)
and static electric fields, respectively. The imaginary we can readily extract α, β and vs :

π X
α = |hj|S + ∆0 (r)|ii|2 δ(²i − ²F )δ(²j − ²F )
2S0 i,j
π X
β = lim |hj|S + ∆0 (r) eiq·r |ii|2 Vj,i δ(²i − ²F )δ(²j − ²F )
q,vs →0 2S0 q · vs
i,j
1 X Vj δ(²j − ²F ) − Vi δ(²i − ²F )
vs · q = |hj|S + ∆0 (r)eiq·r |ii|2 (13)
2S0 i,j ²i − ²j

where we have assumed a uniform precession mode for able solution of the Kohn-Sham equations with arbitrary
the Gilbert damping. impurities. An approximate yet more tractable treat-
Eq. (13) and Eq. (11) identify the non-adiabatic STT ment of disorder consists of the following steps: (i) re-
as a correction to the Gilbert damping in the presence place the actual eigenstates of the disordered system by
of an electric current; in other words, the magnetiza- Bloch eigenstates corresponding to a pure crystal, e.g.
tion damping at finite current is given by the sum of |ii → |k, ai, where k is the crystal momentum and a
the Gilbert damping and the non-adiabatic STT. We feel is the band index of the perfect crystal; (ii) switch Vi to
that this simple interpretation of the non-adiabatic spin- Va = τk,a vk,a ·eE, where τ is the Bloch state lifetime and
transfer torque has not received sufficient emphasis in the vk,a = ∂²k,a /∂k is the quasiparticle group velocity, (iii)
literature. substitute the δ(²k,a − ²F ) spectral function of a Bloch
Strictly speaking the influence of a transport current state by a broadened spectral function evaluated at the
on magnetization dynamics should be calculated by con- Fermi energy: δ(²k,a − ²F ) → Aa (²F , k)/(2π), where
sidering non-linear response of transverse spin to both
effective magnetic fields and the external electric field
which drives the transport current. Our approach, in Γk,a
Aa (²F , k) = Γ2k,a
(14)
which we simply alter the occupation probabilities which (²F − ²k,a )2 + 4
appear in the transverse spin response function is admit-
tedly somewhat heuristic. We demonstrate below that
it gives approximately the same result as the complete and Γa,k = 1/τa,k is the inverse of the quasiparticle
calculation for the case of the very simplistic model for lifetime. This minimal prescription can be augmented
which that complete calculation has been carried out. by introducing impurity vertex corrections in one of the
In Eq. (13), the eigenstates indexed by i are not Bloch spin-flip operators, which restores an exact treatment of
states of a periodic potential but instead the eigenstates disorder in the limit of dilute impurities. This task is
of the Hamiltonian that includes all of the static dis- for the most part beyond the scope of this paper (see
order. Although Eq. (13) provides compact expressions next section, however). The expression for α in Eq. (13)
valid for arbitrary metallic ferromagnets, its practical- has already been discussed in a previous paper;17 hence
ity is hampered by the fact that the characterization of from here on we shall concentrate on the expression for
disorder is normally not precise enough to permit a reli- β which now reads

Z
1 X (vk+q,b τk+q,b − vk,a τk,a ) · eE
β (0) = lim |hk + q, b|S + ∆0 (r)|k, ai|2 Aa (²F , k)Ab (²F , k + q) (15)
q,vs →0 8πs0 k q · vs
a,b

P R R
where we have used k →V dD k/(2π)D ≡ V k
with D as the dimensionality, V as the volume and
Z
1 X evk+q,b τk+q,b δ(²F − ²k+q,b ) − evk,a τk,a δ(²F − ²k,a )
q · vs = |hk + q, b|S + ∆0 (r)|k, ai|2 . (16)
2s0 k ²k,a − ²k+q,b
a,b
 6

In Eq. (15) the superscript “0” is to remind of the absence of impurity vertex corrections; . In addition, we recall that
s0 = S0 /V is the magnetization of the ferromagnet and |aki is a band eigenstate of the ferromagnet without disorder.
It is straightforward to show that Eq. (16) reduces to the usual expression vs = σs E/(es0 ) for vanishing intrinsic
spin-orbit coupling. However, we find that in presence of spin-orbit interaction Eq. (16) is no longer connected to
the spin conductivity. Determining the precise way in which Eq. (16) departs from the conventional formula in real
materials is an open problem that may have fundamental and practical repercussions. Expanding the integrand in
Eq. (15) to first order in q and rearranging the result we arrive at

1 XZ £ ¤
β (0) = − |ha, k|S + ∆0 (r)|b, ki|2 + |ha, k|S − ∆0 (r)|b, ki|2 Aa (²F , k)A0b (²F , k)(vk,a · eE)(vk,b · q)τa
8πs0 q · vs
a,b k
X Z
1 £
− Re hb, k|S − ∆0 (r) |a, kiha, k|S + ∆0 (r) q · ∂k |b, ki
4πs0 q · vs
a,b k
¤
+ (S + ↔ S − ) Aa (²F , k)Ab (²F , k)(vk,a · eE)τa
(17)

b,k; ωn+ω ent simplified models of ferromagnets. For a simpler-to-

ω ω implement approximate version of Eq. (15) or Eq. (17)
(a) S+ S− we refer to Section VI.
III. NON-ADIABATIC STT FOR THE
a,k; ωn PARABOLIC TWO-BAND FERROMAGNET

The model described in this section bears little resem-

b, k+q; ω+
n eVa,b blance to any real ferromagnet. Yet, it is the only model
eVa,b eVa,b in which rigorous microscopic results for β are presently
(b) S+ S− available, thus providing a valuable test bed for Eq. (17).
The mean-field Hamiltonian for itinerant carriers in a
a,k; ωn two-band Stoner model with parabolic bands is simply

FIG. 1: Feynman diagrams for (a) α and (b) β(q · vs ), the k2

latter with a heuristic consideration of the electric field (for a H (k) = − ∆0 S z (18)
2m
more rigorous treatment see Appendix A). Solid lines corre-
spond to Green’s functions of the band quasiparticles in the z
Born approximation, dashed lines stand for the magnon of fre- where ∆0 is the exchange field and Sa,b = δa,b sgn(a).
quency ω and wavevector q, ωn is the Matsubara frequency In this model the eigenstates have no momentum depen-
and eVa,b is the difference in the transport deviation energies. dence and hence Eq. (17) simplifies to
Z
(0) ∆20 X
(vs · q)β = − Aa (²F , k)A0−a (²F , k)
£ ¤2 2πs0 a k
where A0 (²F , k) ≡ 2(²F − ²k,a )Γa / (²F − ²k,a )2 + Γ2a /4
stands for the derivative of the spectral function and we k · q k · eE
τk,a , (19)
have neglected ∂Γ/∂k. Eq. (17) (or Eq. (15)) is the cen- m m
tral result of this work and it provides a gateway to eval-
uate the non-adiabatic STT in materials with complex where a = +(−) for majority (minority) spins, vk,± =
band structures;16 for a diagrammatic interpretation see k/m, and S ± = S x ± iS y with Sa,b
x
= δa,b . Also, from
Fig. (1). An alternative formula with a similar aspiration here on repeated indexes will imply a sum. Taking ∆0 ≤
has been proposed recently,26 yet that formula ignores EF and ∆0 >> 1/τ , the momentum integral in Eq. (19)
intrinsic spin-orbit interactions and relies on a detailed is performed in the complex energy plane using a keyhole
knowledge of the disorder scattering mechanisms. In the contour around the branch cut that stems from the 3D
following three sections we apply Eq. (17) to three differ- density of states:
 7

Z
(0) ∆20 2eE · q ∞
(vs · q)β = − ν(²)Aa (²F,a − ²)A0−a (²F,−a − ²)²τk,a
2πs0 3m 0
eE · q
' sgn(a)νa ²F,a τa Γ−a
6m∆0 s0
eE · q
= (n↑ τ↑ γ↓ − n↓ τ↓ γ↑ ) (20)
2m∆0 s0

where ²F,a = ²F + sgn(a)∆0 , νa is the spin-dependent the contribution from the leading order vertex correction
density of states at the Fermi surface, na = 2νa ²F,a /3 (β (1) ), we shall recover the results obtained previously
is the corresponding number density, and γa ≡ Γa /2. for this model by a full calculation of the transverse spin
The factor 1/3 on the first line of Eq. (20) comes from response function. As it turns out, β (1) is qualitatively
the angular integration. In the second line of Eq. (20) important because it ensures that only spin-dependent
we have neglected a term that is smaller than the one impurities contribute to the non-adiabatic STT in the ab-
retained by a factor of ∆20 /(12²2F ); such extra term (which sence of an intrinsic spin-orbit interaction. In Appendix
would have been absent in a two-dimensional version of B we derive the following result:
the model) appears to be missing in previous work.13,14
The simplicity of this model enables a partial incorpo-
ration of impurity vertex corrections. By adding to β (0)

e∆20
Z h i A (² , k) · A (² , k + q) Ab0 (²F , k0 + q)
¸
+ − a F b F
(vs · q)β (1) = ui Re Sa,b i
Sb,b b ,a
i
0 S 0 0 Sa0 ,a V b,a + V b ,a ,
0 (21)
4πs0 k,k0 (²F − ²k0 ,a0 ) (²F − ²k0 +q,b0 ) (²F − ²k+q,b )

where ui ≡ ni wi2 (i = 0, x, y, z), ni is the density of scatterers, wi is the Fourier transform of the scattering potential
and the overline denotes an average over different disorder configurations.13 Also, Va,b = (τb vk+q,b − τa vk,a ) · eE.
Expanding Eq. (21) to first order in q, we arrive at
Z · ¸
∆2 Aa (²F , k) A0−a (²F , k) A−a (²F , k0 ) k · q k · eE
(vs · q)β (1) = − 0 (u0 − uz ) + τk,a (22)
2πs0 k,k0 ²F − ²k0 ,a ²F − ²k0 ,−a (²F − ²k,−a )2 m m
In theh derivation of Eq. i (22) we have used S ± = S x ± iS y and assumed that ux = uy ≡ ux,y , so that
¡ ¢ R
ui Re Sa,b
x i
Sb,b x i
0 Sb0 ,a0 Sa0 ,a = u0 − uz δa,a0 δb,b0 δa,−b . In addition, we have used k,k0 F (|k|, |k0 |)ki kj0 = 0. The first
term inside the square brackets of Eq. (22) can be ignored in the weak disorder regime because its contribution is
linear in the scattering rate, as opposed to the second term, which contributes at zeroth order. Then,
Z
(1) ∆20 0 z Aa (²F , k)A−a (²F , k0 ) k · q k · eE
(vs · q)β = − (u − u ) 2 m
τk,a
πs0 k,k0 (²F − ²k0 ,a )(²F − ²k,−a ) m
Z
∆2 2eE · q ∞ Aa (²F,a − ²)A−a (²F,−a − ²0 )
' − 0 (u0 − uz ) d²d²0 ν(²)ν(²0 ) ²τa
πs0 3m −∞ (²F − ²0a )(²F − ²−a )2
eE · q
' −π(u0 − uz ) sign(a)na τa ν−a (23)
2m∆0 s0
Combining this with Eq. (20), we get
(vs · q)β ' (vs · q)β (0) + (vs · q)β (1)
eE · q £ ¤
= n↑ τ↑ γ↓ − n↓ τ↓ γ↑ − π(u0 − uz )(n↑ τ↑ ν↓ − n↓ τ↓ ν↑ )
2ms0 ∆0
eE · q
= π [n↑ τ↑ (uz ν↓ + ux,y ν↑ ) − n↓ τ↓ (uz ν↑ + ux,y ν↓ )] (24)
ms0 ∆0

£ ¤
where we have used γa = π (u0 + uz )νa + 2ux,y ν−a . In ically, whereupon Eq. (24) agrees with the results pub-
this model it is simple to solve Eq. (16) for vs analyt-
 8

lished by other authors in Refs.[ 13,14] from full non- model for this enterprise is the two-dimensional electron-
linear response function calculations. However, we reit- gas ferromagnet with Rashba spin-orbit interaction, rep-
erate that in order to reach such agreement we had to resented by
neglect a term of order ∆20 /²2F in Eq. (20). This extra
term is insignificant in all but nearly half metallic ferro- k2
magnets. H (k) = − b · S, (25)
2m

where b = (λky , −λkx , ∆0 ), λ is the Rashba spin-orbit

IV. NON-ADIABATIC STT FOR A coupling strength and ∆0 is the exchange field.
MAGNETIZED TWO-DIMENSIONAL The eigenspinors of this model are |+, ki =
ELECTRON GAS
(cos(θ/2), −i exp(iφ) sin(θ/2)) and |−, ki =
(sin(θ/2), i exp(iφ) cos(θ/2)), where the pspinor an-
The model studied in the previous section misses the gles are defined through cos θ = ∆0 / λ2 k 2 + ∆20
intrinsic spin-orbit interaction that is inevitably present and tan φ = ky /kx . The corresponding
in the band structure of actual ferromagnets. Further- p eigenen-
ergies are Ek± = k 2 /(2m) ∓ ∆20 + λ2 k 2 .
more, since intrinsic spin-orbit interaction is instrumental
Therefore, ³ the band velocities´ are given by
for the Gilbert damping at low temperatures, a similarly p
prominent role may be expected in regards to the non- vk± = k 1/m ∓ λ2 / λ2 k 2 + ∆20 = k/m± . Dis-
adiabatic spin transfer torque. Hence, the present section regarding the vertex corrections, the non-adiabatic
is devoted to investigate the relatively unexplored26,27 ef- spin-torque of this model may be evaluated analytically
fect of intrinsic spin-orbit interaction on β. The minimal starting from Eq. (17). We find that (see Appendix C):

· µ ¶ ¸
∆20 eE · q m2 ∆2 1 1 λ2 kF2 ∆20
(vs · q)β (0) ' 1 + 20 +
8πs0 4m+ m− b b2 4 b6
· µ ¶ µ ¶ ¸
∆20 eE · q 1 m2 λ2 kF2 δm+ ∆20 2 1 m2 λ2 kF2 δm− ∆20
+ 1− τ + 1− τ2 (26)
8πs0 2 m2+ b2 m b2 2 m2− b2 m b2

p √
where b = λ2 kF2 + ∆20 (kF = 2m²F ), and δm± = spherical model for the host semiconductor band struc-
m − m± . As we explain in the Appendix, Eq. (26) ap- ture, Ga1−x Mnx As may be described by
plies for λkF , ∆0 , 1/τ << ²F ; for a more general analysis,
Eq. (17) must be solved numerically (e.g. see Fig. (2)).
Eq. (26) reveals that intrinsic spin-orbit interaction en-
ables intra-band contributions to β, whose signature is ·µ ¶ ¸
the O(τ 2 ) dependence on the second line. In contrast, 1 5
H (k) = γ1 + γ2 k 2 − 2γ3 (k · S)2 + ∆0 Sz ,
the inter-band contributions appear as O(τ 0 ). Since vs 2m 2
itself is linear in the scattering time, it follows that β (27)
is proportional to the electrical conductivity in the clean where S is the spin operator projected onto the J=3/2 to-
regime and the resistivity in the disordered regime, much tal angular momentum subspace at the top of the valence
like the Gilbert damping α. We expect this qualitative band and {γ1 = 6.98, γ2 = γ3 = 2.5} are the Luttinger
feature to be model-independent and applicable to real parameters for the spherical approximation to the valence
ferromagnets. bands of GaAs. In addition, ∆0 = Jpd sNMn = Jpd s0 is
the exchange field, Jpd = 55 meVnm3 is the p-d exchange
coupling, s = 5/2 is the spin of Mn ions, NMn = 4x/a3
V. NON-ADIABATIC STT FOR (Ga,Mn)As is the density of Mn ions and a = 0.565 nm is the lattice
constant of GaAs. We solve Eq. (27) numerically and
input the outcome in Eqs. (16), (17).
In this section we shall apply Eq. (17) to a more sophis-
ticated model which provides a reasonable description of
(III,Mn)V magnetic semiconductors.28 Since the orbitals The results are summarized in Fig. (3). We find that
at the Fermi energy are very similar to the states near the intra-band contribution dominates as a consequence
the top of the valence band of the host (III,V) semicon- of the strong intrinsic spin-orbit interaction, much like for
ductor, the electronic structure of (III,Mn)V ferromag- the Gilbert damping;18 . Incidentally, β barely changes
nets is remarkably simple. Using a p-d mean field theory regardless of whether the applied electric field is along
model for the ferromagnetic ground state and a four-band the easy axis of the magnetization or perpendicular to it.
 9

quantitative relationship between α and β in realistic fer-

∆0=0.5εF ; λkF=0.05εF romagnets, which always have intrinsic spin-orbit inter-
actions. We begin by recollecting the expression for the
0.50
Gilbert damping coefficient derived elsewhere:17
intra−band
0.40 inter−band
total
Z
1 X
α= |hb, k|S + ∆0 |a, ki|2 Aa (²F , k)Ab (²F , k)
0.30
8πs0 k a,b
β

(28)
0.20
where we have ignored disorder vertex corrections. This
expression is to be compared with Eq. (15); for peda-
0.10 gogical purposes we discuss intra-band and inter-band
contributions separately.
0.00
0.00 0.05 0.10 0.15 0.20 Starting from Eq. (15) and expanding the integrand to
1/(εFτ)
first order in q we obtain
FIG. 2: M2DEG: inter-band contribution, intra-band con-
tribution and the total non-adiabatic STT for a magnetized
two-dimensional electron gas (M2DEG). In this figure the ex- Z
change field dominates over the spin-orbit splitting. At higher 1
βintra = |ha, k|S + ∆0 |a, ki|2 Aa (²F , k)2
disorder the inter-band part (proportional to resistivity) dom- 8πs0 k
inates, while at low disorder the inter-band part (proportional j
to conductivity) overtakes. For simplicity, the scattering time eτa q i ∂ki vk,a Ej
(29)
τ is taken to be the same for all sub-bands. q · vs

x=0.08 ; p=0.4 nm
−3 where we have neglected the momentum dependence of
the scattering lifetime and a sum over repeated indices
is implied. Remarkably, only matrix elements that are
0.30
diagonal in momentum space contribute to βintra ; the
intra−band implications of this will be highlighted in the next section.
i
inter−band Recognizing that ∂kj vk,a = (1/m)i,ja , where (1/m)a is
0.20
total the inverse effective mass tensor corresponding to band
a, Eq. (29) can be rewritten as
β

0.10 Z
1 q · vd,a
βintra = |ha, k|S + ∆0 |a, ki|2 Aa (²F , k)2 ,
8πs0 k q · vs
(30)
0.00
0.00 0.10 0.20 0.30
where
1/(εFτ)
FIG. 3: GaMnAs: β (0) for E perpendicular to the easy axis
of magnetization (ẑ). x and p are the Mn fraction and the i
vd,a = eτa (m−1 )i,j j
a E (31)
hole density, respectively. The intra-band contribution is con-
siderably larger than the inter-band contribution, due to the
strong intrinsic spin-orbit interaction. Since the 4-band model
typically overestimates the influence of intrinsic spin-orbit in- is the “drift velocity” corresponding to the quasiparticles
teraction, it is likely that the dominion of intra-band con- in band a. For Galilean invariant systems33 vd,a = vs
tributions be reduced in the more accurate 6-band model. for any (k, a) and consequently βintra = αintra . At first
By evaluating β for E||ẑ (not shown) we infer that it does
glance, it might appear that vs , which (at least in ab-
not depend significantly on the relative direction between the
magnetic easy axis and the electric field.
sence of spin-orbit interaction) is determined by the spin
current, must be different than vd,a . However, recall that
vs is determined by the ratio of the spin current to the
VI. α/β IN REAL MATERIALS
magnetization. If the same electrons contribute to the
transport as to the magnetization, vs = vd,a provided
The preceding three sections have been focused on test- the scattering rates and the masses are the same for all
ing and analyzing Eq. (17) for specific models of ferro- states. These conditions are the conditions for an elec-
magnets. In this section we return to more general con- tron system to be Galilean invariant.
siderations and survey the phenomenologically important
 10

The interband contribution can be simplified by noting that

i i i i i i
τb vk+q,b − τa vk,a = (τb vk+q,b − τa vk+q,a ) + (τa vk+q,a − τa vk,a ). (32)

The second term on the right hand side of Eq.( 32) can then be manipulated exactly as in the intra-band case to
arrive at
1 X Z q · vd,a
βinter = |hb, k|S + ∆0 |a, ki|2 Aa (²F , k)Ab (²F , k) + δβinter (33)
8πs0 k q · vs
a,b(a6=b)

where
1 X Z (τb vk,b − τa vk,a ) · E
δβinter = |ha, k − q|S + ∆0 |b, ki|2 Aa (²F , k − q)Ab (²F , k) . (34)
8πs0 k q · vs
a,b(a6=b)

When Galilean invariance is preserved the quasiparticle

velocity and scattering times are the same for all bands, FIG. 4: Comparison of α and β in (Ga,Mn)As for x = 0.08
which implies that δβ = 0 and hence that βinter = αinter . and p = 0.4nm−3 . It follows that β/α ' 8, with a weak
Although realistic materials are not Galilean invariant, dependence on the scattering rate off impurities. If we use the
δβ is nevertheless probably not significant because the torque correlation formula (Section VII), we obtain β/α ' 10.
term between parenthesis in Eq. (34) has an oscillatory
behavior prone to cancellation. The degree of such can-
cellation must ultimately be determined by realistic cal-
culations for particular materials. As an illustration of the foregoing discussion, in
With this proviso, we estimate that Fig. (4) we evaluate α/β for (Ga,Mn)As. We find β to
be about an order of magnitude larger than α, which is
Z
1 reasonable because (i) the local moment magnetization is
β ' |hb, k|S + ∆0 |a, ki|2 Aa (²F , k)Ab (²F , k) larger than the valence band hole magnetization, and (ii)
8πs0 k
q · vd,a the spin-orbit coupling in the valence band decreases the
. (35) transport spin polarization. Accordingly β is of the order
q · vs
of unity, in qualitative agreement with recent theoretical
As long as δβ ' 0 is justified, the simplicity of Eq. (35) work30 .
in comparison to Eq. (15) or (17) makes of the former the
preferred starting point for electronic structure calcula-
tions. Even when δβ 6= 0 Eq. (35) may be an adequate VII. TORQUE-CORRELATION FORMULA FOR
platform for ab-initio studies on weakly disordered transi- THE NON-ADIABATIC STT
tion metal ferromagnets and strongly spin-orbit coupled
ferromagnetic semiconductors,29 where β is largely de- Thus far we have evaluated non-adiabatic STT us-
termined by the intra-band contribution. Furthermore, ing the bare vertex ha, k|S + |b, k+qi. In this section,
a direct comparison between Eq. (28) and Eq. (35) leads we shall analyze an alternative matrix element denoted
to the following observations. First, for nearly parabolic ha, k|K|b, k+qi (see below for an explicit expression),
bands with nearly identical curvature, where the “drift which may be better suited to realistic electronic struc-
velocity” is weakly dependent on momentum or the band ture calculations.16,31 We begin by making the ap-
index, we obtain β ' (vd /vs )α and thus β/α is roughly proximation that the exchange splitting can be writ-
proportional to the ratio of the total spin density to the ten as a constant spin-dependent shift Hex = ∆0 S z .
itinerant spin density, in concordance with predictions Then, the mean-field quasiparticle Hamiltonian H (k) =
from toy models.12 Second, if α/β > 0 for a system with (k) (k)
Hkin + Hso + Hex can be written as the sum of a spin-
purely electron-like carriers, then α/β > 0 for the same (k)
independent part Hkin , the exchange term, and the spin-
system with purely hole-like carriers because for a fixed (k)
carrier polarization vda and vs reverse their signs under orbit coupling Hso . With this approximation, we have
m → −m. However, if both hole-like and electron-like the identity:
carriers coexist at the Fermi energy, then the integrand
in Eq. (35) is positive for some values of a and negative ha, k|S + |b, k + qi
1 h i
for others. In such situation it is conceivable that α/β be = ha, k| H (k) , S + |b, k + qi
either positive or negative. A negative value of β implies ∆0
a decrease in magnetization damping due to an applied 1 h i
(k)
current. − ha, k| Hso , S + |b, k + qi. (36)
∆0
 11

−3
∆0=0.5εF ; λkF=0.05εF x=0.08 ; p=0.4 nm

0.50
+
S 0.40

0.40 K +
S
K
0.30
0.30
β

β
0.20
0.20

0.10
0.10

0.00 0.00
0.00 0.05 0.10 0.15 0.20 0.00 0.10 0.20 0.30
1/(εFτ) 1/(εFτ)
FIG. 5: M2DEG: comparing S and K matrix element ex- FIG. 7: GaMnAs: comparison between S and K matrix
pressions for the non-adiabatic STT formula in the weakly element expressions for E ⊥ ẑ. The disagreement between
spin-orbit coupled regime. Both formulations agree in the both formulations stems from inter-band transitions, which
clean limit, where the intra-band contribution is dominant. are less important as τ increases. Little changes when E k ẑ.
In more disordered samples inter-band contributions become
more visible and S and K begin to differ; the latter is known
to be more accurate in the weakly spin-orbit coupled regime.

is trivially satisfied for intra-band transitions but less

so for inter-band transitions.18 For q 6= 0 the agreement
∆0=0.5εF ; λkF=0.8εF between intra-band matrix elements is no longer obvi-
0.40
ous and is affected by the momentum dependence of
the band eigenstates. At any rate, Eq. (29) demon-
strates that only q = 0 matrix elements contribute to
0.30 βintra ; therefore βintra has the same value for S and K
S
+
matrix elements. The disparity between the two formu-
K lations is restricted to βinter , and may be significant if
the most prominent inter-band matrix elements connect
β

0.20
states that are not close in energy. When they disagree,
it is generally unclear32 whether S or K matrix elements
0.10 will yield a better estimate of βinter . The weak spin-orbit
limit is a possible exception, in which the use of K ap-
pears to offer a practical advantage over S. In this regime
0.00 S generates a spurious inter-band contribution in the ab-
0.00 0.05 0.10 0.15 0.20
1/(εFτ) sence of magnetic impurities (recall Section III) and it is
FIG. 6: M2DEG: In the strongly spin-orbit coupled limit only after the inclusion of the leading order vertex correc-
the intra-band contribution reigns over the inter-band contri- tion that such deficiency gets remedied. In contrast, K
bution and accordingly S and K matrix element expressions vanishes identically in absence of spin-orbit interactions,
display a good (excellent in this figure) agreement. Neverthe- thus bypassing the pertinent problem without having to
less, this agreement does not guarantee quantitative reliabil- introduce vertex corrections.
ity, because for strong spin-orbit interactions impurity vertex
corrections may play an important role. Figs. (5)- (7) display a quantitative comparison be-
tween the non-adiabatic STT obtained from K and S,
both for the M2DEG and (Ga,Mn)As. Fig. (5) reflects
the aforementioned overestimation of S in the inter-band
The last term in the right hand side of Eq. (36) is the dominated regime of weakly spin-orbit coupled ferromag-
generalization of the torque matrix element used in ab- nets. In the strong spin-orbit limit, where intra-band
initio calculations of the Gilbert damping: contributions dominate in the disorder range of interest,
1 h i K and S agree fairly well (Figs. (6) and (7)). Summing
(k)
ha, k|K|b, k + qi ≡ ha, k| Hso , S + |b, k + qi (37) up, insofar as impurity vertex corrections play a minor
∆0
role and the dominant contribution to β stems from intra-
Eq. (36) implies that at q = 0 hb, k|S + |a, ki ' band transitions the torque-correlation formula will pro-
hb, k|K|a, ki provided that (Ek,a − Ek,b ) << ∆0 , which vide a reliable estimate of β.
 12

VIII. CONNECTION TO THE EFFECTIVE shape of the Fermi surface, provided that there is an in-
FIELD MODEL trinsic spin-orbit interaction. Consequently, empty (full)
states appear below (above) the Fermi energy, giving rise
As explained in Section II we view the non-adiabatic to an out-of-equilibrium quasiparticle distribution. This
STT as the change in magnetization damping due to a configuration tends to relax back to equilibrium, but re-
transport current. The present section is designed to population requires a time τ . Due to the time delay,
complement that understanding from a different perspec- the quasiparticle distribution lags behind the dynamical
tive based on an effective field formulation, which pro- configuration of the Fermi surface, effectively creating a
vides a simple physical interpretation for both intra-band friction (damping) force on the magnetization. From a
and inter-band contributions to β. quantitative standpoint, the preceding discussion means
An effective field Heff may be expressed as the varia- that the quasiparticle energies ²k,a follow the magnetiza-
tion of the system energy with respect to the magnetiza- tion adiabatically, whereas the occupation numbers nk,a
tion direction Hieff = −(1/s0 )∂E/∂Ωi . Here we approxi- deviate from the instantaneous equilibrium distribution
mate the energy with the Kohn-Sham eigenvalue sum fk,a via
X µ ¶
E= nk,a ²k,a . (38) ∂fk,a ∂fk,a ∂fk,a
k,a nk,a = fk,a − τk,a + r˙a · + k̇ · ,
∂t ∂r ∂k
(40)
The variation of this energy with respect to the magne-
tization direction yields where we have used the relaxation time approximation.
· ¸ As we explain below, the last two terms in Eq. (40) do
1 X ∂²k,a ∂nk,a not contribute to damping in the absence of an electric
Hieff = − nk,a + ²k,a . (39) field and have thus been ignored by prior applications of
s0 ∂Ωi ∂Ωi
k,a
the breathing Fermi surface model, which concentrate on
It has previously been shown that, in the absence of cur- Gilbert damping. It is customary to associate intra-band
rent, the first term in the sum leads to intra-band Gilbert magnetization damping with the torque exerted by the
damping15,35 while the second term produces inter-band part of the effective field
damping.34 In the following, we generalize these results
by allowing the flow of an electrical current. α and β may eff 1 X ∂²k,a
Hintra =− nk,a (41)
be extracted by identifying the the dissipative part of the s0 ∂ Ω̂
k,a
effective field with −α∂ Ω̂/∂t − βvs · ∇Ω̂ that appears in
the LLS equation.
Intra-band terms: We begin by recognizing that as the that is lagging behind the instantaneous magnetization.
direction of magnetization changes in time, so does the Plugging Eq. (40) in Eq. (41) we obtain

· ¸
eff 1 X ∂²k,a ∂fk,a ∂²k,a ∂²k,a ∂Ωj ∂fk,a ∂²k,a ∂²k,a ∂Ωj ∂fk,a ∂²k,a ∂²k,a
Hintra,i = −fk,a + τa + τa ṙal + τa k̇ j (42)
s0 ∂Ωi ∂²k,a ∂Ωi ∂Ωj ∂t ∂²k,a ∂Ωi ∂Ωj ∂rl ∂²k,a ∂kj ∂Ωi
k,a

where a sum is implied over repeated Latin indices. The ing linear order in perturbation theory insufficient (see
first term in Eq. (42) is a contribution to the anisotropy Appendix A). In order to account for the relevant non-
field; it evolves in synchrony with the dynamical Fermi linearity we use ṙ = ∂²(k−ev·Eτ )/∂k in Eq.( 42), where
eff
surface and is thus the reactive component of the effective v = ∂²(k)/∂k. The dissipative part of Hintra then reads
field. The remaining terms, which describe the time lag
of the effective field due to a nonzero relaxation time, are
responsible for intra-band damping. The last term van-
· ¸
ishes in crystals with inversion symmetry because k̇ = eE eff,damp 1 X ∂f²k ,a ∂²k,a ∂²k,a ∂Ωj l ∂Ωj
and ∂²/∂k is an odd function of momentum. Similarly, Hintra,i = τk,a + vd,a ,
s0 ∂²k,a ∂Ωi ∂Ωj ∂t ∂rl
if we take ṙ = ∂²(k)/∂k the second to last term ought to k,a
vanish as well. This leaves us with the first two terms in (43)
Eq. ( 42), which capture the intra-band Gilbert damping
i j
but not the non-adiabatic STT. This is not surprising as where vd,a = eτa (m−1 )i,j
a E is the “drift velocity” cor-
the latter involves the coupled response to spatial varia- responding to band a. Eq. (43) may now be identified
tions of magnetization and a weak electric field, render- with −αintra ∂ Ω̂/∂t − βintra vs · ∇Ω̂ that appears in the
 13

LLS equation. For an isotropic system this results in Multiplying Eq. (46) with ∂ Ω̂/∂t we get
µ ¶2
1 X ∂fk,a ∂²k,a h i
αintra = − τk,a eff,damp 1 X
s0 ∂²k,a ∂Ωi Hinter · ∂t Ω̂ = − ²k,a ∂na,k /∂ Ω̂ · ∂ Ω̂/∂t
k,a,i s0
µ ¶2 k,a
1 X ∂fk,a ∂²k,a q · vd,a 1 X
βintra = − τk,a . (44) = − ²k,a ∂na,k /∂t . (47)
s0 ∂²k,a ∂Ωi q · vs s0
k,a,i
k,a

Since h[S x , Hso ]i = ∂φ hexp(iS x φ)Hso exp(−iS x φ)i = The rate of change of the populations of the Kohn-
∂²/∂φ for an infinitesimal angle of rotation φ around Sham states can be approximated by the following master
the instantaneous magnetization, β in Eq. (44) may be equation
rewritten as
∆20 X ∂fk,a q · vd,a ∂na,k X
βintra = τk,a |hk, a|K|k, ai|2 (45) =− Wa,b (nk,a − nk0 ,b ), (48)
2s0 ∂²k,a q · vs ∂t
k,a 0 b,k

where K = [S + , Hso ]/∆0 is the spin-torque operator in- where

troduced in Eq. ( 37) and we have claimed spin rota-
tional invariance via |h[S x , Hso ]i|2 = |h[S y , Hso ]i|2 . Using 2
Wa,b = 2π |hb, k0 |∆0 S x |a, ki| δk0 ,k+q δ(²b,k0 − ²a,k − ω)
∂f /∂² ' −δ(² − ²F ) and recalling from Section VII that (49)
+
Ka,a = Sa,a , Eq. (45) is equivalent to Eq. (30); note that is the spin-flip inter-band transition probability as dic-
the product of spectral functions in the latter yields a tated by Fermi’s golden rule. Eqs. (48) and (49)
factor of 4πτ upon momentum integration. These obser- rely on the principle of microscopic reversibility36 and
vations prove that βintra describes the contribution from are rather ad hoc because they circumvent a rigorous
a transport current to the “breathing Fermi surface” type analysis of the quasiparticle-magnon scattering, which
of damping. Furthermore, Eq. (44) highlights the impor- would for instance require keeping track of magnon occu-
tance of the ratio between the two characteristic veloci- pation number. Furthermore, quasiparticle-phonon and
ties of a current carrying ferromagnet, namely vs and vd . quasiparticle-impurity scattering are allowed for simply
As explained in Section VI these two velocities coincide by broadening the Kohn-Sham eigenenergies (see below).
in models with Galilean invariance. Only in these arti- The right hand side of Eq. (48) is now closely related
ficial models, which never apply to real materials, does to inter-band magnetization damping because it agrees37
α = β hold. with the net decay rate of magnons into particle-hole
Inter-band terms: The Kohn-Sham orbitals are effec- excitations, where the particle and hole are in different
tive eigenstates of a mean-field Hamiltonian where the bands. Combining Eq. (47) and (48) and rearranging
spins are aligned in the equilibrium direction. As spins terms we arrive at
precess in response to external rf fields and dc trans-
port currents, the time-dependent part of the mean-field eff 1 X
Hinter · ∂t Ω̂ = Wa,b (nk,a − nk0 ,b )(²k,a − ²k0 ,b ).
Hamiltonian drives transitions between the ground-state 2s0 0
k,k ,a,b
Kohn-Sham orbitals. These processes lead to the second (50)
term in the effective field and produce the inter-band For the derivation of αinter it is sufficient to approximate
contribution to damping. nk,a as a Fermi distribution in Eq. (50); here we ac-
We thus concentrate on the second term in Eq. (39), count for a transport current by shifting the Fermi seas as
1 X ∂nk,a nk,a → nk,a − evk,a · Eτk,a ∂nk,a /∂²k,a , which to leading
eff
Hinter =− ²k,a . (46) order yields
s0 ∂ Ω̂
k,a

πω X ¯¯ ¯2 ∂nk,a
eff
Hinter · ∂t Ω̂ = − hb, k + q|∆0 S + |a, ki¯ δ(²b,k+q − ²a,k − ω) (−ω + eVb,a )
2s0 ∂²k,a
k,a,b
ω X ¯¯ ¯2
= hb, k + q|∆0 S + |a, ki¯ Aa (k, ²F )Ab (k + q, ²F )(−ω + eVb,a ) (51)
8πs0
k,a,b

where we have used S x = (S + + S − )/2 and defined Vb,a = evk+q,b · Eτk+q,b − evk,a · Eτk,a . In the second line of
Eq.( 51) we have assumed low temperatures, and have introduced a finite quasiparticle lifetime by broadening the
spectral functions of the Bloch states into Lorentzians with the convention outlined in Eq. (14): δ(x) → A(x)/(2π).
 14

Identifying Eq.( 51) with (−αinter ∂t Ω̂ − βinter (vs · ∇)Ω̂) · ∂t Ω̂ = −αinter ω 2 + βinter ω(q · vs ) we arrive at
1 X X ¯¯ ¯2
αinter = hb, k + q|∆0 S + |a, ki¯ Aa (k, ²F )Ab (k + q, ²F )
8πs0
a,b6=a k,a,b
1 X X¯ ¯
βinter = ¯hb, k + q|∆0 S + |a, ki¯2 Aa (k, ²F )Ab (k + q, ²F )Vb,a (52)
8πs0 q · vs
a,b6=a k,a,b

in agreement with our results of Section II.

IX. SUMMARY AND CONCLUSIONS the inter-band contribution the spin-torque formula of-
fers a physically sensible result in the weak spin-orbit
Starting from the Gilbert damping α and including the limit but its quantitative reliability is questionable un-
influence of an electric field in the transport orbitals semi- less the prominent inter-band transitions connect states
classically, we have proposed a concise formula for the that are close in energy. Finally, we have extended the
non-adiabatic spin transfer torque coefficient β that can breathing Fermi surface model for the Gilbert damping
be applied to real materials with arbitrary band struc- to current carrying ferromagnets and have accordingly
tures. Our formula for β reproduces results obtained found a complementary physical interpretation for the
by more rigorous non-linear response theory calculations intra-band contribution to β; similarly, we have applied
when applied to simple toy models. By applying this ex- the master equation in order to offer an alternative inter-
pression to a two-dimensional electron-gas ferromagnet pretation for the inter-band contribution to β. Possible
with Rashba spin-orbit interaction, we have found that avenues for future research consist of carefully analyzing
it implies a conductivity-like contribution to β, related to the importance of higher order vertex corrections in β,
the corresponding contribution to the Gilbert damping α, better understanding the disparities between the differ-
which is proportional to scattering time rather than scat- ent approaches to vs , and finding real materials where
tering rate and arises from intra-band transitions. Our α/β is negative.
subsequent calculations using a four-band model have Acknowledgements
shown that intra-band contributions dominate in ferro-
magnetic semiconductors such as (Ga,Mn)As. We have We acknowledge informative correspondence with
then discussed the α/β ratio in realistic materials and Rembert Duine and Hiroshi Kohno. In addition, I.G.
have confirmed trends expected from toy models, in ad- is grateful to Paul Haney for interesting discussions and
dition to suggesting that α and β can have the oppo- generous hospitality during his stay in the National In-
site sign in systems where both hole-like and electron-like stitute of Standards and Technology. This work was sup-
bands coexist at the Fermi surface. Afterwards, we have ported in part by the Welch Foundation, by the National
analyzed the spin-torque formalism suitable to ab-initio Science Foundation under grant DMR-0606489, and by
calculations, and have concluded that it may provide a the NIST-CNST/UMD-NanoCenter Cooperative Agree-
reliable estimate of the intra-band contribution to β; for ment.

APPENDIX A: QUADRATIC SPIN RESPONSE TO AN ELECTRIC AND MAGNETIC FIELD

Consider a system that is perturbed from equilibrium by a time-dependent perturbation V(t). The change in the
expectation value of an operator O(t) under the influence of V(t) can be formally expressed as

δhO(t)i = hΨ0 |U † (t)O(t)U (t)|Ψ0 i − hΨ0 |O(t)|Ψ0 i (A1)

where |Ψ0 i is the unperturbed state of the system,

· Z t ¸
0 0
U (t) = T exp −i V(t )dt (A2)
−∞

is the time-evolution operator in the interaction representation and T stands for time ordering. Expanding the
exponentials up to second order in V we arrive at
Z t Z
1 t
δhO(t)i = i dt0 h[O(t), V(t0 )]i − dt0 dt00 h[[O(t), V(t0 )] , V(t00 )]i. (A3)
−∞ 2 −∞

For the present work, O(t) → S a (a = x, y, z) and

 15
Z Z
V(t) = − drj · A(r, t) + drS · Hext (r, t), (A4)

where A is the vector potential, Hext is the external magnetic field, and j is the current operator. Plugging Eq. (A4)
into Eq. (A3) and neglecting O(A2 ), O(Hext2
) terms we obtain
X Z XZ XZ
δS a (x) = dx0 χa,b
S,j A b 0
(x ) + dx0 a,b
χ H b
S,S ext (x0
) + dx0 dx00 χa,b,c b 0 c 00
S,S,j A (x )Hext (x ), (A5)
b b b,c
R R
R∞
where x ≡ (r, t) and dx0 ≡ −∞ dt0 dr0 . The linear and quadratic response functions introduced above are defined
as
£ a ¤
χa,b 0 b 0
S,j (x, x ) = ih S (x), j (x ) Θ(t − t )
0
£ ¤
χa,b 0 a b 0
S,S (x, x ) = ih S (x), S (x ) Θ(t − t )
0
££ a ¤ c 00 ¤
χa,b,c 0 00 b 0 0
S,S,j (x, x , x ) = h S (x), j (x ) , S (x ) Θ(t − t )Θ(t − t )
0 00
££ a ¤ ¤
+h S (x), S b (x00 ) , j c (x0 ) Θ(t − t00 )Θ(t00 − t0 ) (A6)
0 0 00 0 00 00 0 00 0
where we have used T [F (t)G(t )] = F (t )G(t )Θ(t − t ) + G(t )F (t )Θ(t − t ), Θ being the step function. χS,j is the
spin density induced by an electric field in a uniform ferromagnet, and it vanishes unless there is intrinsic spin-orbit
interaction. χS,S is the spin density induced by an external magnetic field. χS,S,j is the spin density induced by the
combined action of an electric and magnetic field (see Fig. (8) for a diagrammatic representation); this quantity is
closely related to (vs · q)χ(2) , introduced in Section II.

APPENDIX B: FIRST ORDER IMPURITY VERTEX CORRECTION

The aim of this Appendix is to describe the derivation of Eq. (21). We shall begin by evaluating the leading order
vertex correction to the Gilbert damping. From there, we shall obtain the counterpart quantity for the non-adiabatic
STT by shifting the Fermi occupation factors to first order in the electric field. The analytical expression for the
transverse spin response with one vertex correction is (see Fig. (9))
Z
QP,(1) ∆2 X 0 0
χ̃+,− = −V 0 T +
ui Ga (iωn , k)Sa,b i
Gb (iωn + iω, k + q)Sa,b − i
0 Gb0 (iωn + iω, k + q)S 0 0 Ga0 (iωn , k )Sa0 ,a . (B1)
b ,a
2 ω k,k 0
n

where V is the volume of the system and the minus sign originates from fermionic statistics. Using the Lehmann
representation of the Green’s functions G and performing the Matsubara sum we get
Z h i Z ∞ d² d²0 d² d²0
QP,(1) ∆2 1 1 2 2
χ̃+,− = −V 0 ui 2 Re Sa,b + i
Sb,b 0 S −
S i
b0 ,a0 a0 ,a 4
Aa (²1 , k)Aa0 (²01 , k0 )
2 k,k0 −∞ (2π)
· µ ¶¸
0 0 f (²1 ) ²1 ↔ ²2 , ²01 ↔ ²02 ,
×Ab (²2 , k + q)Ab0 (²2 , k + q) + (B2)
(²1 − ²01 )(iω + ²1 − ²2 )(iω + ²1 − ²02 ) ω ↔ −ω
where twice the real part arose after absorbing two of the terms coming from the Matsubara sum. Next, we apply
iω → ω + i0+ and take the imaginary part:
Z h i Z ∞ d² d²0 d² d²0
QP,(1) ∆2 1 1 2 2
χ̃+,− = V 0 2π +
ui Re Sa,b i
Sb,b − i
0 S 0 0 Sa0 ,a
b ,a Aa (²1 , k)Aa0 (²01 , k0 )Ab (²2 , k + q)Ab0 (²02 , k0 + q)
2 k,k0 −∞ (2π)4
· µ ¶¸
f (²1 ) δ(ω + ²1 − ²2 ) δ(ω + ²1 − ²02 ) ω → −ω,
× + − (B3)
²1 − ²01 ω + ²1 − ²02 ω + ²1 − ²2 q → −q

+ −
S S
v.A

FIG. 8: Feynman diagram for χS,S,j . The dashed lines correspond to magnons, whereas the wavy line represents a photon.
 16

S+ S−

FIG. 9: Feynman diagram for the first order vertex correction. The dotted line with a cross represents the particle-hole
correlation mediated by impurity scattering.

where we used 1/(x − iη) = P V (1/x) + iπδ(x), and invoked spin-rotational invariance to claim that terms with
x i y i
Sa,b Sb,b 0 Sb0 ,a0 Sa0 ,a will vanish. Integrating the delta functions we arrive at

Z Z ∞ 0
QP,(1) ∆20 d²1 d²2 d²02 f (²2 )Aa (²2 , k)Aa0 (²01 , k0 )
χ̃+,− = V ui Re [...]
2 k,k0 −∞ (2π)3 (²2 − ²02 )(²2 − ²01 )
µ ¶
£ ¤ ω → −ω,
× Ab (²2 + ω, k + q)Ab0 (²02 + ω, k0 + q) + Ab (²02 + ω, k + q)Ab0 (²2 + ω, k0 + q) − (B4)
q → −q

The next step is to do the ²01 and ²02 integrals, taking advantage of the fact that for weak disorder the spectral
functions are sharply peaked Lorentzians ( in fact at the present order of approximation one can take regard them as
Dirac delta functions). The result reads
Z Z ∞ · ¸
QP,(1) ∆2 d²2 f (²2 )Aa (²2 , k) Ab (²2 + ω, k + q) Ab0 (²2 + ω, k0 + q)
χ̃+,− = V 0 ui Re [...] +
2 k,k0 −∞ 2π ²2 − ²k0 ,a0 ²2 + ω − ²k0 +q,b0 ²2 + ω − ²k+q,b
− (ω → −ω, q → −q) (B5)

By making further changes of variables, this equation can be rewritten as

Z Z ∞ · ¸
QP,(1) ∆2 d²2 (f (²2 ) − f (²2 + ω)) Aa (²2 , k) Ab (²2 + ω, k + q) Ab0 (²2 + ω, k0 + q)
χ̃+,− = V 0 ui Re [...] + (B6)
2 k,k0 −∞ 2π ²2 − ²k0 ,a0 ²2 + ω − ²k0 +q,b0 ²2 + ω − ²k+q,b

This is the first order vertex correction for the Gilbert damping. In order to obtain an analogous correction for the
non-adiabatic STT, it suffices to shift the Fermi factors in Eq. (B6) as indicated in the main text. This immediately
results in Eq. (21).

APPENDIX C: DERIVATION OF EQ. (26)

Let us first focus on the first term of Eq. (17), namely

Z
£ ¤ j
Ei qj |ha, k|S + |b, ki|2 + |ha, k|S − |b, ki|2 Aa A0b vk,a
i
vk,b τk,a (C1)
k

We shall start with the azimuthal integral. It is easy to show that the entire angle dependence comes from v i v j ∝ ki kj ,
from which the azimuthal integral vanishes unless i = j.
Regarding the |k| integral, we assume that λkF , ∆0 , 1/τ << ²F ; otherwise R the analytical
R∞ calculation is complicated
and must be tackled numerically. Such assumption allows us to use k → N2D −∞ d². For inter-band transitions
(a 6= b), Aa A0b contributes mainly thru the pole at ²F,a , thus all the slowly varying factors in the integrand may be
set at the Fermi energy. For intra-band transitions (a = b), Aa A0a has no peak at the Fermi energy; hence it is best
to keep the slowly varying factors inside the integrand.
The above observations lead straightforwardly to the following result:
Z
£ ¤ j
Ei qj |ha, k|S + |b, ki|2 + |ha, k|S − |b, ki|2 Aa A0b vk,a
i
vk,b τk,a
k
µ ¶
m2 ∆2 (²F,− τ− Γ+ − ²F,+ τ+ Γ− )
' E·q 1 + 20
8m+ m− b b3
· 2 µ ¶ µ ¶ ¸
m 1 λ2 kF2 ∆20 2 m2 1 λ2 kF2 ∆20 2
− E·q 1 + τ+ + 1 + τ− (C2)
m2+ 2 b2 b2 m2− 2 b2 b2
 17

The second and third line in Eq. (C2) come from inter-band and intra-band transitions, respectively. The latter
vanishes in absence of spin-orbit interaction, leading to a 2D version of Eq. (20). Since the band-splitting is much
smaller than the Fermi energy, one can further simplify the above equation via τ+ ' τ− → τ .
Let us now move on the second term of Eq. (17), namely
Z
£ ¤
Ei qj Re hb, k|S − |a, kiha, k|S + ∂kj |b, ki + (S + ↔ S − ) Aa Ab vk,a
i
τk,a (C3)
k

Most of the observations made above apply for this case as well. For instance, the azimuthal integral vanishes
unless i = j. This follows from a careful evaluation of the derivatives of the eigenstates with respect to momentum;
∂kj θ = sin(θ) cos(θ)kj /k 2 (0 ≤ θ ≤ π/2) is a useful relation in this regards, while ∂kj φ plays no role. As for the |k|
integral, we no longer have the derivative of a spectral function, but rather a product of two spectral functions; the
resulting integrals may be easily evaluated using the method of residues. The final result reads
Z
£ ¤
Ei qj Re hb, k|S − |a, kiha, k|S + ∂kj |b, ki + (S + ↔ S − ) Aa Ab vk,a
i
τk,a
k
· µ ¶ µ ¶¸
m λ2 kF2 ∆20 τ− m λ2 kF2 ∆20 τ+
' −E · q 1 + + 1 +
32m− b6 τ+ 32m+ b6 τ−
· ¸
m λ2 kF2 ∆20 2 m λ2 kF2 ∆20 2
+ E·q τ+ + τ− (C4)
4m+ b4 4m− b4

The first line in Eq. (C4) stems from inter-band transitions, whereas the second comes from intra-band transitions;
both vanish in absence of SO. Once again we can take τ+ ' τ− → τ . Combining Eqs. (C2) and (C4) one can
immediately reach Eq. (26).

1
L. Berger, J. Appl. Phys. 3, 2156 (1978); ibid. 3, 2137 73, 054428 (2006).
11
(1979). M. Yamanouchi, D. Chiba, F. Matsukura and H. Ohno,
2
L. Berger, Phys. Rev. B 54, 9353 (1996). Phys. Rev. Lett. 96, 96601 (2006).
3 12
J.C. Slonczewski, J. Magn. Magn. Mater. 159, L1 (1996). Y. Tserkovnyak, H.J. Skadsem, A. Brataas and G.E.W
4
H. Kubota, A. Fukushima, Y. Ootani, S. Yuasa, K. Bauer, Phys. Rev. B 74, 144405 (2006).
13
Ando, H. Maehara, K. Tsunekawa, D.D. Djayaprawira, N. H. Kohno, G. Tatara and J. Shibata, J. Phys. Soc. Japan
Watanabe and Y. Suzuki, Jap. J. of Appl. Phys. 44, L1237 75, 113707 (2006).
14
(2005); J. Hayakawa, S. Ikeda, Y.M. Lee, R. Sasaki, T. Me- R.A. Duine, A.S. Nunez, J. Sinova and A.H. MacDonald,
guro, F. Matsukura, H. Takahashi and H. Ohno, Jap. J. Phys. Rev. B 75, 214420 (2007).
15
of Appl. Phys. 44, L1267 (2005); J. A. Katine and E. E. See for instance J. Kunes and V. Kambersky, Phys. Rev.
Fullerton, J. Magn. Magn. Mater. 320, 1217 (2007). B 65 212411 (2002) and references therein.
5 16
For reviews of spin transfer torque in magnetic multilayers K. Gilmore, I. Garate, P.M. Haney, A.H. MacDonald and
see D.C. Ralph and M.D. Stiles, J. Magn. Magn. Mater. M.D. Stiles (in preparation).
17
320, 1190 (2007); M. D. Stiles and J. Miltat, Top. Appl. I. Garate and A.H. MacDonald, arXiv:0808.1373.
18
Phys. 101, 225 (2006). I. Garate and A.H. MacDonald, arXiv:0808.3923.
6 19
For reviews of spin transfer torque in continuously varying Here we assume that the dependence of energy on mag-
magnetizations see P.M. Haney, R.A. Duine, A.S. Nunez netization direction which determines Heff is specified as
and A.H. MacDonald, J. Magn. Magn. Mater. 320, 1300 a function of Ωx and Ωy only with Ωz implicitly fixed by
(2007); Y. Tserkovnyak, A. Brataas and G.E.W. Bauer, the constraint Ωz = [1 − Ω2x − Ω2y ]1/2 . If the free energy
J. Magn. Magn. Mater. 320, 1282 (2007); G. Tatara, H. was expressed in a form with explicit Ωz dependence we
Kohno and J. Shibata, arXiv:0807.2894 (accepted to Phys. would find Heff,x = −∂F/∂Ωx − (∂F/∂Ωz )(∂Ωz /∂Ωx ) =
Rep.). −∂F/∂Ωx +(∂F/∂Ωz )Ωx , where F is the free energy of the
7
M.D. Stiles and A. Zangwill, Phys. Rev. B 66, ferromagnet. Similarly we would find Heff,y = −∂F/∂Ωy +
14407(2002); A. Shapiro, P. M. Levy, and S. Zhang, Phys. (∂F/∂Ωz )Ωy . The terms which arise from the Ωz depen-
Rev. B, 67, 104430 (2003); J. Xiao, A. Zangwill, and M. dence of the free energy would more commonly be regarded
D. Stiles, Phys. Rev. B 70, 172405 (2004); A. Brataas, G. as contributions to Hef f,z . The difference is purely a mat-
E. W. Bauer, and P. J. Kelly, Phys. Rep. 427, 157 (2006). ter of convention since both results would give the same
8
A. S. Nunez and A. H. MacDonald, Solid State. Comm. value for Ω̂ × Heff .
20
139, 31 (2006). Z. Qian and G. Vignale, Phys. Rev. Lett. 88, 056404
9
S. Zhang and Z. Li, Phys. Rev. Lett. 93, 127204 (2004). (2002).
10 21
J. Q. Xiao, A. Zangwill, and M. D. Stiles, Phys. Rev. B O. Gunnarsson, J. Phys. F 6, 587 (1976).
 18

22 30
We assume that magnetic anisotropy and the external K.M.D. Hals, A.K. Nguyen and A. Brataas,
magnetic fields are weak compared to the exchange- arXiv:0811.2235.
correlation splitting of the ferromagnet. ∆ ¯ is the spin- 31
V. Kambersky, Phys. Rev. B 76, 134416 (2007); K.
density weighted average of ∆(r) (see Ref. [17]). Gilmore, Y.U. Idzerda and M.D. Stiles, Phys. Rev. Lett.
23
For convenience in Eq. (8) we use hS + S − i response func- 99, 27204 (2007).
tions instead of hS x S x i and hS y S y i. They are related via 32
In order to gauge the accuracy of either matrix element,
S x = (S + + S − )/2 and S y = (S + − S − )/2i. one must obtain an exact evaluation of the non-adiabatic
24
J. Fernandez-Rossier, M. Braun, A. S. Nunez, A. H. Mac- STT, which entails a ladder-sum renormalization18 of S ± .
Donald, Phys. Rev. B 69, 174412 (2004). This is beyond the scope of the present work.
25 33
J.A.C. Bland and B. Heinrich (Eds.), Ultrathin Mag- S.E. Barnes and S. Maekawa, Phys. Rev. Lett. 95, 107204
netic Structures III: Fundamentals of Nanomagnetism (2005).
34
(Springer-Verlag, New York, 2005). K. Gilmore, Y.U. Idzerda and M.D. Stiles, J. Appl. Phys.
26
G. Tatara and P. Entel, Phys. Rev. B 78, 064429 (2008). 103, 07D303 (2008).
27 35
For a theoretical study on how Rashba spin-orbit interac- D. Steiauf and M. Fahnle, Phys. Rev. B 72, 064450 (2005);
tion affects domain wall dynamics see K. Obata and G. D. Steiauf, J. Seib and M. Fahnle, Phys. Rev. B 78,
Tatara, Phys. Rev. B 77, 214429 (2008). 02410(R) (2008).
28 36
T. Jungwirth, J. Sinova, J. Masek, J. Kucera and A.H. This principle states that Wa,b = Wb,a exp((²a − ²b )/T ).
MacDonald, Rev. Mod. Phys. 78, 809 (2006). Since the magnon energy is much smaller than the un-
29
For actual ab-initio calculations it may be more con- certainty in the quasiparticle energies, we approximate
venient to substitute |ha, k|∆0 S + |b, ki|2 in Eq. (35) by Wa,b ' Wb,a .
|ha, k|K|b, ki|2 , where K is the spin-torque operator dis- 37
For an analogous observation in the context of electron-
cussed in Section VII. In either case we are disregarding phonon interaction see e.g. D. Pines, Elementary Excita-
impurity vertex corrections, which may become significant tions in Solids (Benjamin, 1963).
in disordered and/or strongly spin-orbit coupled systems.