PHYSICAL REVIEW B 92, 014419 (2015)

Breakdown of the adiabatic approach for magnetization damping in metallic ferromagnets

A. T. Costa and R. B. Muniz
Instituto de Fsica, Universidade Federal Fluminense, 24210-346 Niteroi, RJ, Brazil
(Received 11 December 2014; revised manuscript received 9 June 2015; published 22 July 2015)
We present a microscopic theory for magnetization relaxation in metallic ferromagnets of nanoscopic
dimensions that is based on the calculation of the dynamic spin response matrix in the presence of spin-orbit
coupling. Our approach takes into account the collective character of spin excitations in ferromagnetic metals.
We show that calculations of the Gilbert damping based on a mean-field, adiabatic approximation fail to describe
the relaxation of collective excitations in the ballistic regime. Namely, we show that collective spin excitations
such as the ferromagnetic resonance mode have finite damping rates even for perfectly crystalline systems,
whereas existing microscopic approaches, based on adiabatic and mean-field approximations, predict infinite
damping rates. Moreover, we demonstrate that within a finite frequency approach the relaxation properties are
not completely determined by the transverse susceptibility alone, and that the damping rate has a non-negligible
frequency dependence in experimentally relevant situations. Our results indicate that the widely used adiabatic
approach for spin dynamics breaks down for metallic nanostructures in the presence of strong spin-orbit
coupling.
DOI: 10.1103/PhysRevB.92.014419

PACS number(s): 75.30.Ds, 75.40.Gb, 75.70.Tj, 75.78.n

Magnetization relaxation in metals is at the heart of spin

current generation and detection processes currently under
investigation, many of them candidates to play protagonist
roles in innovative spintronic devices. The Landau-LifshitzGilbert (LLG) equation is widely used to describe the spin
dynamic properties of magnetic materials [1,2]. It includes
an important system-dependent parameter, called the Gilbert
damping parameter, usually denoted by G , that regulates the
relaxation of the magnetization towards stability, after it is
driven out of equilibrium. Recently, a lot of effort has been
put into the determination of this damping rate [28], which
characterizes the pumping and absorption of pure spin currents
in nanostructures that are of great interest in the field of
spintronic. In most of them spin-orbit interaction is significant,
and responsible for a desirable interplay between charge spin
and angular momentum excitations.
There is a general agreement between practitioners in
the field that a proper microscopic theory of magnetization
relaxation in metals requires a good description of the
electronic structure of the system including spin-orbit coupling
[38]. The conventional approach is to combine a realistic
electronic structure with some kind of adiabatic approximation
and, within a mean-field treatment of the magnetization
dynamics, to derive expressions that can be directly related
to the Landau-Lifshitz-Gilbert phenomenology. This strategy
has been employed by Kambersky [3] and many others since
[48]. This conventional approach has important limitations.
First and foremost, it adopts the view that, in order to obtain
the damping parameter of a system, it is enough to consider its
mean-field transverse spin response. Furthermore, it assumes
the linewidth of the zero wave vector spin wave resonance
is completely determined by the zero frequency limit of the
mean-field transverse susceptibility. In doing so, it neglects
the coupling between transverse spin, longitudinal spin,
and charge excitations (which is an important consequence
of the spin-orbit coupling). It also incorrectly predicts the
divergence of the damping parameter for a perfectly crystalline
system at zero temperature. Actually, for ferromagnets that
1098-0121/2015/92(1)/014419(5)

display rotation symmetry in spin space, the Goldstone

theorem ensures that any experiment which measures the total
transverse magnetic moment of the sample will produce a
resonant response with zero linewidth [9]. Even in the presence of disorder the uniform mode should still have infinite
lifetime, since this is a consequence of the invariance of the
systems Hamiltonian to global rotations in spin space. In the
presence of spin-orbit interaction, however, this symmetry is
explicitly broken, and the resonant spectrum acquires a finite
linewidth [10].
We put forward a more fundamental microscopic approach
to the calculation of the spin dynamics damping rate that
takes fully into account the effects of SOC on the spectrum
of spin excitations of itinerant systems. Namely, by studying
the magnetization dynamics within a ladder approximation
treatment of the screened Coulomb repulsion, we naturally
incorporate the coupling of transverse spin excitations to
longitudinal spin and charge excitations, induced by the
spin-orbit interaction [10]. We calculate the FMR spectrum
at finite frequencies and arbitrary anisotropy values, without
employing any adiabatic approximation. Thus, the main difference between our approach and previous approaches found
in the literature is the fact that we treat the spin excitations
as collective excitations, whereas the existing approaches
describe the magnetization dynamics as a semiclassical motion
of the magnetization due to a slow change of the mean
exchange field. We will show that treating the spin dynamics as
resulting from collective excitations is an essential ingredient
to correctly describe the magnetization relaxation in very
clean metallic ferromagnets of nanoscopic dimensions at low
temperatures.
Our work is organized as follows: We will present briefly
our formalism, discuss its main features, and present numerical
results for two model systems that illustrate common but
qualitatively different situations. We will also present a version
of the adiabatic mean-field approach that allows us to compare
the effects of those approximations to the predictions of
our approach without abandoning completely our calculation
scheme.

014419-1

2015 American Physical Society

A. T. COSTA AND R. B. MUNIZ

PHYSICAL REVIEW B 92, 014419 (2015)

General formalism. The spectrum of spin excitations of

a ferromagnet can be obtained from the spectral density
associated with the transverse spin susceptibility,

+ (l,l  ; ) = dteit Sl+ (t),Sl (0),
(1)

The system of equations of motion obeyed by these four

susceptibilities can be cast into a form strongly resembling the RPA result by introducing a block vector 
( (1) , (2) , (3) , (4) )T , with (1) + . With an equivalent
definition for the mean-field susceptibilities 0(m) we write
 () = 0 ()  (),

where
Sl+ (t),Sl (0) i (t)[Sl+ (t),Sl (0)],
and
Sl+ =

al al .

(2)

(3)

The operator al creates one electron in the atomic basis state

localized at lattice site l with spin . Here  represents the
frequency of the applied transverse magnetic field, and (t)
is the usual step function which is equal to unit for t > 0
and zero otherwise. Although we are usually interested in
+ (l,l  ; ) as defined above, its equation of motion involves
the orbital-resolved susceptibility,

+


  (l,l ; t) al (t)al (t),al   (0)al   (0).

(4)

One notices that this susceptibility is actually a two-particle

Green function, that describes the propagation of an electron
and a hole of opposite spins. Within the mean-field approximation the electron and the hole propagate independently and
this propagator describes what is known as Stoner excitations.
In order to describe spin waves it is necessary at least to take
into account the Coulomb interaction between the electron
and the hole. The simplest approximation that incorporates
this interaction is a partial summation of Feynman diagrams
of the ladder type, sometimes also called a random phase
approximation (RPA) [11]. In our treatment this interaction is
brought about by a screened exchange term that acts among
electrons occupying d orbitals within the same atomic site
[10]. The ferromagnetic resonance mode, most relevant to the
discussion of Gilbert damping, is actually the zero wave-vector
limit of the spin wave mode.
In the absence of spin-orbit coupling (SOC) and within the
+

RPA, the equation of motion for
  (l,l ; t) is closed and
+

(l,l ; ) can be expressed in the well-known form,

() = [1 +

U 0+ ()]1 0+ (),

where the super-matrix  is proportional to the effective

Coulomb interaction strength and involves convolutions of
single particle Green functions. Explicit forms for its matrix
elements are found in Ref. [10]. The numerical analysis of
the susceptibilities (2) , (3) , and (4) show that their absolute
values are many orders of magnitude smaller than those of
(1) = + . It is, thus, tempting to argue that the transverse
susceptibility is approximately decoupled from (2) , (3) ,
and (4) and that it can be calculated via the usual RPA
expression with the single particle Green functions obtained
with spin-orbit coupling taken into account. This is not a good
approximation in general, since the matrix elements of  that
couple (1) to the other susceptibilities are far from negligible.
Our numerical calculations indicate that they are essential to
determine correctly the features of the FMR mode around
the resonance frequency. Thus, the behavior of (1) in the
presence of spin-orbit coupling cannot be inferred from 0(1) in
the zero-frequency limit alone, as it is usually assumed in the
literature on the calculation of the Gilbert damping parameter
[35,8,12].
Numerical results. We start the discussion by presenting
results for the Gilbert parameter G for unsupported ultrathin
Co films. We define G as the ratio between the FMR
linewidth  and the resonance frequency 0 . First we
turn off spin-orbit coupling to check the consistency of our
approach. Even with SOC turned off we still find a finite
linewidth for the FMR mode. It comes, as we will shortly
demonstrate, from the small imaginary part that is usually
added to the energy in the numerical calculations of the single
particle Green functions, in order to shift their poles from
the real axis. We calculate G for various values of and
extrapolate to 0+ , as shown in Fig. 1. It is clear that
0.05

(5)
0.04

where 0+ () is the mean-field susceptibility. This expression

is schematic and must be understood as a matrix in orbital and
site indices, in real space, or a wave-vector-dependent matrix
in reciprocal space. The crucial point, however, is that, in the
absence of spin-orbit coupling, within the RPA, transverse
spin excitations are decoupled from either longitudinal spin
excitations or charge excitations. This ceases to be true when
SOC is included, as we demonstrated in Ref. [10]: +
becomes coupled to three other susceptibilities, namely

(2)


  (l,l ; t) al (t)al (t),al   (0)al   (0),

(3)


  (l,l ; t) al (t)al (t),al   (0)al   (0),

(4)


  (l,l ; t) al (t)al (t),al   (0)al   (0).

(9)

0.03
0.02
0.01
0
0

(6)
(7)
(8)

10

20
30
(meV)

40

50

FIG. 1. (Color online) Gilbert damping parameter G as a function of the imaginary part added to the real energy, for an ultrathin
film of two atomic layers of Co where SOC has been turned off. It is
clear that G vanishes as 0.

014419-2

BREAKDOWN OF THE ADIABATIC APPROACH FOR . . .

PHYSICAL REVIEW B 92, 014419 (2015)

0.04

0.03

0.02

0.01

0
0

(meV)

10

15

FIG. 2. (Color online) Gilbert damping parameter G as a function of the imaginary part added to the energy, for Co ultrathin films
of various thicknesses: one (circles), two (squares), four (diamonds),
and six (triangles) atomic layers. The strength of the SOC is
= 85 meV. The solid lines are guides to the eye.

0.1

lim0+ G = 0. Thus, our approach correctly predicts that

the Gilbert damping parameter vanishes in the absence of
SOC, as it should. Indeed, it is easy to show [9] that the
FMR mode is a stationary state of the mean-field Hamiltonian
and, as such, has infinite lifetime in the limit 0+ . Now we
discuss the dependence of G on for a fixed, nonzero value of
the spin-orbit coupling strength . We used LCAO parameters
appropriate for bulk Co to describe the electronic structure of
all Co films we investigated. The quantitative details of the
ferromagnetic ground state and excitation spectra are sensitive
to the LCAO parameters used, but their qualitative behavior is
very robust to small changes in the electronic structure. Our
strategy is to use the same set of LCAO parameters for all film
thicknesses to avoid modifications in G coming directly from
changes in the LCAO parameters. This allows us to focus on
geometric effects and on the dependence.
Figure 2 shows the dependence of the Gilbert damping
parameter G on the imaginary part for Co films of various
thicknesses. Clearly G approaches finite values as 0.
This is in sharp contrast with most approaches found in the
literature, for which G is expected to diverge in the 0
limit. In practice this would mean that the FMR mode should be
so strongly damped in very clean samples at low temperatures
that it would be very difficult to observe. To the best of our
knowledge there are no signs of this divergent behavior in
experimental results performed at low temperatures and high
quality crystalline samples.
As we mentioned in the introduction, all existing approaches for calculating the Gilbert damping from the electronic structure are based on the evaluation of the transverse
spin susceptibility within the mean-field approximation, in
the limit of vanishing frequency [13]. We can apply both
approximations to our approach, and also neglect the coupling
of the transverse excitations to the longitudinal and charge
excitations. This provides a direct assessment of the effect of
those approximations without drifting too far away from our
own approach. In Fig. 3 we show the comparison between
the damping parameters obtained within the finite-frequency
RPA and the adiabatic mean-field approaches for an unsup-

0.05

10
(meV)

12

14

FIG. 3. (Color online) Gilbert damping parameter G calculated

within the nonadiabatic RPA approach (black circles and green
triangles), and within the adiabatic mean-field approximation (red
squares and blue diamonds). The calculations are for a single Co
atomic layer with Co = 85 meV (black circles and red squares) and
Co = 128 meV (green triangles and blue diamonds).

ported Co monolayer. Besides the non-negligible quantitative

discrepancy there is also a qualitatively different behavior as
MFAd
decreases; the adiabatic mean-field G
seems to reach a
minimum and starts to increase, whereas the nonadiabatic RPA
RPA
G
continues to decrease monotonically. Still in Fig. 3 we
show the effect of increasing artificially the strength of SOC in
Co. The discrepancy is clearly much bigger, both quantitatively
and qualitatively.
Cobalt has a small spin-orbit coupling constant. We would
like to investigate the effect of a large SOC strength on the
damping rate in a more realistic setting than just tuning the
SOC strength of Co. This can be achieved by considering layers
of Co attached to a nonmagnetic substrate with high SOC parameter and partially filled d bands, such as Pt. This system has
a particularly interesting feature: The magnetization easy axis
is perpendicular to the plane. However, we found that, for the
LCAO parameters we employed, the magnetization in-plane is
also a stable configuration, with a small magnetocrystalline
anisotropy. The damping rate, however, is much larger in the
2Co/2Pt system than in the unsupported Co films. This is a nice
example of how the anisotropy energy is strongly influenced
by the systems symmetry, but the damping rate is relatively
insensitive to it, depending strongly on the intensity of the
spin-orbit coupling. It is also an extremely convenient situation
to test an assumption very frequently found in the literature on
Gilbert damping, although sometimes not explicitly stated, that
the FMR linewidth  is linearly dependent on the resonance
frequency 0 and that  0 as 0 0. This is not an
unreasonable hypothesis, considering the weak static fields
commonly used in FMR experiments and the smallness of
the spin-orbit coupling constant, compared to other energy
scales of a ferromagnet. Our calculations for unsupported Co
films confirm that this relationship approximately holds. In
this case, the Gilbert parameter G may be extracted from the
FMR spectrum obtained by sweeping either the Zeeman field
or excitation frequency, and is practically independent of field
or frequency. However, our results for 2Co/2Pt indicate that

014419-3

A. T. COSTA AND R. B. MUNIZ

PHYSICAL REVIEW B 92, 014419 (2015)

(a)

(b)

2e+06
FMR spectral density

h 0 (meV)

0.8

1.5e+06

0.6
0.4

5e+05

0.2
0
0
0.25

0.1
0.05
_
h (meV)

2
B (T)

0.15
0.1
0.05

-0.4

-0.2

B (T)

0.2

FIG. 5. (Color online) Spectral densities of the FMR mode for

the 2Co/2Pt system plotted as a function of the Zeeman field B at
fixed exciting frequencies: p = 24 GHz (squares) and p = 54 GHz
(circles). The solid curves are Lorentzian fits to the calculated points.

(c)

0.2

1e+06

2
B (T)

FIG. 4. (Color online) FMR frequency 0 for the 2Co/2Pt system

plotted as a function of FMR linewidths  (a), and as a function of
the Zeeman field B (b). In (c) we show the Gilbert damping parameter
G as a function of the applied Zeeman field B calculated with our
nonadiabatic RPA approach (black circles) and the results for the
adiabatic mean-field approximation (red squares). The strengths of
the SOC are Co = 85 meV and Pt = 600 meV.

the FMR linewidth approaches a finite value as 0 0, as

seen in Fig. 4(a). The corresponding resonance frequencies as
a function of the Zeeman field may be seen in Fig. 4(b). This
leads to a significantly frequency-dependent G , as shown in
Fig. 4(c). Thus, it is impossible to characterize the damping
rate of such a system by a single parameter. It is necessary
either to consider the FMR linewidth explicitly as a function
of frequency or to define a frequency-dependent damping
parameter G (0 ).
This discussion is also relevant for the interpretation usually
employed to extract damping parameters from measured FMR
spectra. In experiments the spectra are usually obtained as a
function of Zeeman field intensity, for a fixed frequency of the
exciting field. If the Gilbert parameter G were independent of
frequency then the relationship between frequency sweeping
and Zeeman field sweeping spectra would be straightforward.
With G frequency dependent, however, the relationship is
not trivial. In Fig. 5 we illustrate this issue by plotting the
FMR spectral density as a function of the Zeeman field for
two fixed exciting frequencies, 24 GHz and 54 GHz. The

curves have nice Lorentzian shapes, but the values for the
Gilbert damping parameter G extracted from these curves
depend on the exciting frequency (G = 0.034 for 0 = 0.10
meV and G = 0.042 for 0 = 0.22 meV). Also, they do not
correspond to any of the values shown in Fig. 4(c), although
the Zeeman field values that determine the linewidth in Fig. 5
lie within the range of Zeeman field values shown in Fig. 4(b).
Thus, with G defined as / 0 , its value for a given sample
depends on wether the FMR spectrum is obtained in a fixed
frequency or fixed Zeeman field setup. Our results also imply
that the existing expressions for the damping parameter G
are not valid in general, especially for very clean systems
with large spin-orbit coupling. The conventional approaches
express G as the ratio / 0 in the 0 0 limit. As we
have just shown, this limit may not exist in some cases, since
 approaches a finite value as 0 0.
Once again we can compare our nonadiabatic RPA results
with those obtained from an adiabatic mean-field approximaMFAd
RPA
tion. In Fig. 4(c) we compare G
and G
for the 2Co/2Pt
system as a function of the applied Zeeman field. As expected,
the mean-field adiabatic approximation predicts an essentially
constant damping parameter, whereas the nonadiabatic RPA
approach shows a large variation with the Zeeman field.
In experimental papers [14,15] the FMR linewidth is
assumed to have a zero-frequency offset, just as we described.
This is usually attributed to extrinsic broadening mechanisms,
such as two-magnon scattering [16], due to the combination
between inhomogeneities in the magnetic films and dipolar
interactions. This is certainly the case in systems with small
SOC, such as Fe films deposited on GaAs or Au [14]. However,
we have shown that there can be zero-frequency offset of
intrinsic origin if the SOC is large. The effect of this intrinsic
offset should be easily separated from that of the two-magnon
scattering mechanism, since the latter is not active when the
magnetization is perpendicular to the plane of the film [16].
We would like to remark that Stoner enhancement in Pt also
plays a very important role in the determination of the damping
rate. We had shown previously [17] that, in the absence of
spin-orbit coupling, Stoner enhancement had a very mild

014419-4

BREAKDOWN OF THE ADIABATIC APPROACH FOR . . .

PHYSICAL REVIEW B 92, 014419 (2015)

FMR spectral density

8e+05

6e+05

4e+05

2e+05

0
0

E (meV)

FIG. 6. (Color online) Spectral densities of the FMR mode for

the 2Co/2Pt system with Stoner enhancement in Pt turned on (black
line) and off (red line) as a function of energy E = .

effect on the damping rate in the Co/Pd(001) system. In the

presence of SOC, however, the effect can be very large indeed.
Both magnetocrystalline anisotropy and damping rate are
significantly different in the enhanced and nonenhanced cases,
as shown in Fig. 6. The Gilbert parameter is also very different
nonenh
enh
in the two cases: G
= 0.11, whereas G
= 0.33. Thus,
proper treatment of Stoner enhancement in substrates like
Pd and Pt is essential for the correct determination of spin
relaxation features.
We presented a microscopic scheme for calculating the
Gilbert damping parameter G in ultrathin metallic magnetic
films, and illustrate it by presenting results for unsupported
Co films and for Co/Pt bilayers. Our approach is based on
the evaluation of the dynamic transverse susceptibility in the
presence of spin-orbit coupling, taking into account realistic
electronic structures and the coupling between transverse
spin, longitudinal spin, and charge excitations. The essential

[1] T. Gilbert, Magnetics, IEEE Transactions on 40, 3443 (2004).

[2] Y. Tserkovnyak, A. Brataas, G. Bauer, and B. Halperin, Rev.
Mod. Phys. 77, 1375 (2005).
[3] V. Kambersky, Phys. Rev. B 76, 134416 (2007).
[4] I. Garate and A. MacDonald, Phys. Rev. B 79, 064403 (2009).
[5] I. Garate and A. MacDonald, Phys. Rev. B 79, 064404
(2009).
[6] A. A. Starikov, P. J. Kelly, A. Brataas, Y. Tserkovnyak, and
G. E. W. Bauer, Phys. Rev. Lett. 105, 236601 (2010).
[7] H. Ebert, S. Mankovsky, D. Kodderitzsch, and P. J. Kelly, Phys.
Rev. Lett. 107, 066603 (2011).
[8] E. Barati, M. Cinal, D. M. Edwards, and A. Umerski, Phys. Rev.
B 90, 014420 (2014).
[9] R. B. Muniz and D. L. Mills, Phys. Rev. B 68, 224414 (2003).

difference between our approach and the ones usually found in

the literature is the description of the magnetization dynamics
in terms of collective spin excitations. As a consequence we
find finite values of G in the limit of perfectly crystalline
films, a regime where methods based on the torque correlation
formula find a diverging Gilbert damping parameter. We
showed that the coupling between transverse, longitudinal,
and charge excitations, due to spin-orbit coupling, is of
fundamental importance for the correct determination of
FMR spectra in metallic systems. We have also shown that
the damping rate extracted from the FMR spectrum for a
fixed excitation frequency may differ considerably from that
extracted from the FMR spectrum for a fixed Zeeman field.
In this case the Gilbert damping parameter G becomes
frequency dependent, in contrast to what is assumed in the
standard Landau-Lifshitz-Gilbert phenomenology. Moreover,
we have numerical indications that the Gilbert parameter
may not be well defined in the limit of vanishing resonance
frequency, a fact that is very relevant to calculational schemes
based on the adiabatic approximation. Incidentally, Stoner
enhancement in materials like Pt and Pd also plays an important
role in the determination of FMR frequencies and damping
rates. These results may lead to important modifications of
the interpretation of damping parameters, either calculated
or inferred from experimental results, for systems where
spin-orbit coupling is strong. We believe these issues may be
crucial for the correct description of relaxation in very clean
systems of nanoscopic dimensions, especially in the presence
of relatively weak magnetocrystalline anisotropy.
The authors acknowledge partial financial support from
CNPq and FAPERJ. We are grateful to C. Lewenkopf for a
critical reading of the manuscript and to M. Odashima for
enlightening discussions. The authors acknowledge fruitful
discussions with D. M. Edwards, A. Umerski, A. Liebsch, and
F. Freimuth.

[10] A. T. Costa, R. B. Muniz, S. Lounis, A. B. Klautau, and D. L.

Mills, Phys. Rev. B 82, 014428 (2010).
[11] S. Doniach and E. H. Sondheimer, Greens Functions for Solid
State Physicists (Imperial College Press, London, 1998).
[12] Y. Liu, Z. Yuan, R. J. H. Wesselink, A. A. Starikov, and P. J.
Kelly, Phys. Rev. Lett. 113, 207202 (2014).
[13] A. Sakuma, J. Appl. Phys. 117, 013912 (2015).
[14] R. Urban, G. Woltersdorf, and B. Heinrich, Phys. Rev. Lett. 87,
217204 (2001).
[15] E. Montoya, B. Heinrich, and E. Girt, Phys. Rev. Lett. 113,
136601 (2014).
[16] R. Arias and D. L. Mills, Phys. Rev. B 60, 7395 (1999).
[17] D. L. R. Santos, P. Venezuela, R. B. Muniz, and A. T. Costa,
Phys. Rev. B 88, 054423 (2013).

014419-5