A convergent and precise finite element

scheme for Landau–Lifschitz–Gilbert

equation

François Alouges, Evaggelos Kritsikis,

Jutta Steiner & Jean-Christophe
Toussaint

Numerische Mathematik

ISSN 0029-599X

Numer. Math.
DOI 10.1007/s00211-014-0615-3

1 23
Your article is protected by copyright and
all rights are held exclusively by Springer-
Verlag Berlin Heidelberg. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

1 23
 Author's personal copy
Numer. Math. Numerische
DOI 10.1007/s00211-014-0615-3 Mathematik

A convergent and precise finite element scheme

for Landau–Lifschitz–Gilbert equation

François Alouges · Evaggelos Kritsikis ·

Jutta Steiner · Jean-Christophe Toussaint

Received: 16 November 2012 / Revised: 8 November 2013

© Springer-Verlag Berlin Heidelberg 2014

Abstract In this paper, we propose a new scheme for the numerical integration of
the Landau–Lifschitz–Gilbert (LLG) equations in their full complexity, in particular
including stray-field interactions. The scheme is consistent up to order 2 (in time), and
unconditionally stable. It combines a linear inner iteration with a non-linear renor-
malization stage for which a rigorous proof of convergence of the numerical solution
toward a weak solution is given, when both space and time stepsizes tend to 0. A numer-
ical implementation of the scheme shows its performance on physically relevant test
cases. We point out that to the knowledge of the authors this is the first finite element
scheme for the LLG equations which enjoys such theoretical and practical properties.

Mathematics Subject Classification 35K55 · 65M12 · 65M60

1 Introduction

In 1935 Landau and Lifschitz proposed a model for the description of the evolution of
the magnetization in a ferromagnetic material [15]: Suppose that the three dimensional
ferromagnetic sample under consideration occupies some domain Ω ⊂ R3 and let

F. Alouges (B)
CMAP-École Polytechnique, Route de Saclay, 91128 Palaiseau cedex, France
e-mail: [email protected]

E. Kritsikis · J.-C. Toussaint

Institut Néel, 25 avenue des Martyrs, Bâtiment K, BP 166, 38042 Grenoble cedex 9, France

E. Kritsikis · J.-C. Toussaint

Grenoble-INP, 46 Avenue Félix Viallet, 38031 Grenoble Cedex 1, France

J. Steiner
Institute for Applied Mathematics, Mathematik Zentrum, Endenicher Allee 60, 53115 Bonn, Germany

123
 Author's personal copy
F. Alouges et al.

m be the direction of the magnetization. The model, namely the Landau–Lifschitz–

Gilbert (LLG) equations, reads as follows

∂t m − α m × ∂t m = −γ0 Ms m × Heff in Ω,
(1)
∂n m = 0 on ∂Ω.

(As above, vectorial quantities will be denoted in bold letters in the sequel to avoid
ambiguity). The parameters in the equation are the damping parameter α, the gyro-
magnetic constant γ0 and the saturation magnetization Ms . The so-called effective
field Heff , measured in units of Ms , is given by the functional derivative of the micro-
magnetic (free) energy E, more precisely

∂E
Heff (m) = − = d 2 Δm + Hd (m) + Hext + Q (e · m) e (2)
∂m

where the energy E (see [10,13,15]) is given by

⎛
 
1⎝ 2
E(m) = d |∇m|2 dx − Hd (m) · m dx
2
Ω Ω
⎞
 
−2 Hext · m dx − Q (e · m)2 dx⎠ . (3)
Ω Ω

The four contributions to the effective field in (2) and the energy in (3), respec-
tively, correspond to the so-called exchange, stray-field, applied and anisotropy field
or energy, respectively. The material constants in (2) and (3) are the exchange con-
stant d, the anisotropy constant Q and the anisotropy direction e (also called the easy
axis). Furthermore, the vector field Hext models an applied magnetic field. We will
also use the notation Haniso = Q (e · m) e. The stray field Hd (m) is the magnetic
field induced by the magnetization distribution m via the following (subset of) static
Maxwell equations

curl Hd (m) = 0 in R3
(4)
div (Hd (m) + m) = 0 in R3 .

Notice that in the equation above, all fields are measured in units of Ms . Below the Curie
temperature, the magnetization can be described by a directional field that we rescale
to unit length. It is straightforward to check that the magnitude of the magnetization

|m(x, t)| = 1 (5)

is formally conserved by the dynamics (1). Take note that the gyromagnetic term is a
conservative term while the damping term leads to the following energy dissipation
law

123
 Author's personal copy
A convergent and precise finite element scheme


d α
E(m(t)) = − |∂t m|2 dx . (6)
dt γ0 M s
Ω

In the mathematical analysis, most of the physical constants can be normalized to

one by rescaling. For example, rescaling time and redefining α allows to assume that
γ0 Ms = 1. However, in the last section of this paper, where the proposed numerical
scheme is tested in physical situations, it is more appropriate to go back to the original
scaling.
The numerical approximation of solutions to (1) is an important issue in appli-
cations. Nowadays, numerous strategies exist in the literature—though among them
only few reliable ones [11]. Classical schemes are based on finite differences that are,
as usual, well adapted to Cartesian grids. On the other hand, finite element approx-
imations will suit complex geometries and weak solutions, with the drawback that
they are difficult to analyze in practice. In particular, proving the convergence of a
finite element solution towards a solution of (1) as the space and time steps tend to
zero turns out to be quite difficult and has probably been first established in [5]. This
result was further improved in [8] and [1] for the case in which only the exchange
term is present, and in [3] for the general case. Such formulations are now in use in
the physics literature [16]. We hereafter study a further generalization of the scheme
proposed in [1]: a second order (in time) variant. We point out that despite the fact
that LLG equation is highly non-linear, our scheme still uses an inner iteration which
requires a linear and positive definite system to be solved. Numerical tests support the
performance of the method. Notice also that convergent order 2 schemes, based for
instance on midpoint approximations have previously been proposed in [9] with the
drawback of needing to solve a non-linear inner iteration.
Let us start with a brief outline of our paper. In Sect. 2 we first recall the notion of
weak solutions. Section 3 introduces the finite-element spaces. Section 4 restates the
first order scheme proposed in [1]. The non-linearity of the LLG equation calls for
recurrent renormalization of the time-discrete approximation, an issue discussed in
Sect. 5. Section 6 finally provides a derivation of our new scheme, the main con-
vergence result together with the proof. Section 7 finally covers some numerical
examples.

2 Notion of weak solutions to LLG

Let us recall the notion of a weak solution to (1) from [6] and [18].

Definition 1 Consider an initial magnetization, i.e., a vector field m0 ∈ H 1 (Ω)3 that

is a.e. of unit length. A vector field m is called a weak solution to (1) with initial data
m0 if for all times T > 0 there holds

1. m ∈ H 1 (ΩT )3 with ΩT = Ω × (0, T ), and |m| = 1 a.e.

2. For all test functions ∈ H1 (ΩT )3

123
 Author's personal copy
F. Alouges et al.

 
∂t m ·  dx dt − α (m × ∂t m) ·  dx dt
ΩT ΩT


d 
= d2 m × ∂xi m · ∂xi  dx dt (7)
i=1Ω

T

− m × (Hd (m) + Hext + Haniso (m)) ·  dx dt,

ΩT

3. The magnetization initially satisfies m(x, 0) = m0 (x) in the trace sense, and
4. The energy decreases according to

E(m(T )) + α |∂t m|2 dx dt ≤ E(m(0)). (8)
ΩT

3 The finite element scheme

As in [5], our discretization relies on piecewise linear finite elements in space com-
bined with a linear interpolation in time. The domain Ω is discretized by a conformal
triangulation Th of mesh size h with vertices (xih )1≤i≤Nh . Let us denote by (φih )1≤i≤Nh
the set of associated piecewise linear basis functions that satisfy φih (xhj ) = δi, j at the
vertices xhj for 1 ≤ i, j ≤ Nh , where δi, j denotes the Kronecker symbol. This amounts
to a standard P 1 (Th )-discretization. Based on the scalar basis (φih )1≤i≤Nh we construct
the vector-valued finite element space as


Vh = uh = ui φih , s.t. ∀i, ui ∈ R3 .
i

Due to the unit-length constraint (5), the solution to (7) is sought in the subset

Mh = uh ∈ Vh , s.t. ∀i, ui ∈ S2 ⊂ Vh .


Let us also introduce the tangent space to mh = i mi φih ∈ Mh , denoted by


K m = vh = vi φih , s.t. ∀i, vi · mi = 0 .
i

Furthermore, the classical nodal interpolation operator is given by

Ih : C 0 (Ω, R3 ) → Vh

u → u(xih )φih . (9)
i

123
 Author's personal copy
A convergent and precise finite element scheme

To simplify notations, the index h of the ansatz functions will be neglected from now
on most of the times, i.e., we write u, v, etc. instead of uh , vh , respectively, in case
this does not lead to any ambiguities.

4 Revisiting the θ -scheme

The finite element scheme proposed in [5] relies on the observation that the LLG Eq.
(1)—with the notation v = ∂t m—can be rewritten in the following weak form
 
α v ·  dx + α m × v ·  dx
Ω Ω
 
= −d 2
∇m · ∇ dx + (Hd (m) + Hext + Haniso (m)) ·  dx. (10)
Ω Ω

Equation (10) holds for every test function  ∈ H 1 (Ω, R3 ) that satisfies (x)·m(x) =
0 for a.e. x in Ω. The reformulation of (1) in the form of (10) motivated the following
θ -scheme introduced in [1]:

Algorithm 1 Given an initial m0 ∈ Mh choose θ ∈ [0, 1] and a time step size k = T

N
with N ∈ N. For n = 0, 1, . . . , N
⎡
a)Find vn ∈ K mnsuch that for all test functions
  ∈ K mn
⎢
⎢ α v ·  dx + m × v ·  dx = −d
n n n 2
∇(mn + θ kvn ) · ∇ dx
⎢
⎢ Ω Ω Ω
⎢
⎢ (11)
⎢ + (Hd (m ) + Hext + Haniso (m )) ·  dx
n n
⎢
⎢ Ω
⎢ 
⎣ min + kvin
b) Set mn+1 = min+1 φih , where ∀i, min+1 = .
|min + kvin |
i

It is noteworthy that this procedure requires the solution of a linear equation in each
time step only. Moreover, due to the fact that the symmetric part of the underlying
matrix is positive definite, existence and uniqueness of a solution to (11) is guaranteed.
T 
The time discrete solution constructed via algorithm (11) at time-steps N =
k
is interpolated as follows:

Definition 2 In each time interval t ∈ [nk, (n + 1)k) with n ∈ {0, · · · , N } we set

t − nk n+1 (n + 1)k − t n
mh,k (x, t) = m (x) + m (x),
k k
−
mh,k = mn (x), vh,k = vn (x).

123
 Author's personal copy
F. Alouges et al.

−
Our notational convention is thus that mh,k , mh,k and vh,k refer to suitable time
interpolants of the time discrete approximation m and vn . Notice that mh,k is piece-
n
−
wise linear in time whereas mh,k and vh,k are piecewise constant. (The introduction
of the piecewise constant magnetization will be useful in the convergence proof).
Based on this discretization, weak convergence of the constructed approximation was
established in [1]. Both the proof of this result and the proof in case of our new
scheme consist of the following two main “classical” steps: as a first step establish-
ing an energy estimate which guarantees the convergence (sufficiently strong) of the
sequence constructed and then in a second step verifying that the limit indeed satisfies
the equation. As far as the first step is concerned, the following section addresses
the fact that the energy behaves well under renormalization—in principle a strongly
nonlinear modification of the flow.

5 Renormalization decreases the energy

The influence of the renormalization on the exchange energy was for instance inves-
tigated in [2] in the continuous case. More precisely, it was shown that for maps
w ∈ H 1 (Ω, R3 ) with |w(x)| ≥ 1 a.e. x ∈ Ω one has
   
 w 2
∇  dx ≤ |∇w|2 dx. (12)
 |w| 
Ω Ω

Hence, the renormalization step is expected to be energy decreasing—at least as far as

the Dirichlet energy is concerned. Applications more related to finite element approx-
imation of micromagnetic configurations can be found in [4]. The discrete version of
(12) was proved by Bartels in [7]:
Theorem 1 [7] If the basis functions of the P 1 -approximation satisfy

∀i = j, ∇φih · ∇φ hj dx ≤ 0, (13)
Ω

then for all v = i vi φih ∈ Vh such that ∀i ∈ {1, . . . , Nh }, |vi | ≥ 1 it holds that
   2 
 
∇Ih v  dx ≤ |∇v|2 dx. (14)
 |v| 
Ω Ω

In 3d, the condition (13)—and hence (14)—is for instance satisfied provided all
dihedral angles of the tetrahedra of the mesh are smaller than π/2, see [17].

6 The new (almost) order 2-scheme

Consider an exact solution m(nk) at time nk. As remarked in [1], it is a priori not suf-
ficient to choose θ = 21 in (11) to achieve quadratic order. Indeed, the renormalization
stage inherently introduces an order 2 term due to the expansion

123
 Author's personal copy
A convergent and precise finite element scheme

m(nk) + kv k2
mn+1 := = m(nk) + kv − |v|2 m(nk) + O(k 3 ) . (15)
|m(nk) + kv| 2

Hence, in order to have mn+1 = m((n + 1)k) + O(k 3 ), it is necessary to take this
term into account in the variational formulation that defines v. We hence modify the
time-discrete approximation of the magnetization m which we heuristically derive in
the following. It is well known that the midpoint rule is exact up to cubic error, i.e.,
 
1
m((n + 1) k) = m(nk) + kmt (n + ) k + O(k 3 )
2
 
k
= m(nk) + k mt (nk) + mtt (nk) + O(k 3 ) .
2

Now, setting
 
k
v = Pm⊥ mt (nk) + mtt (nk)
2
k
= mt (nk) + Pm⊥ mtt (nk) , (16)
2
where Pm⊥ denotes the projection onto the orthogonal component of m(nk), and
noticing that v · m(nk) = 0, a Taylor expansion up to cubic order reveals

m(nk) + kv k2
= m(nk) + kv − |v|2 m(nk) + O(k 3 )
|m(nk) + kv| 2
k2
= m(nk) + kmt (nk) + Pm⊥ (mtt (nk))
2
k 2
− |mt (nk)|2 m(nk) + O(k 3 )
2
k2
= m(nk) + kmt (nk) + mtt (nk) + O(k 3 )
2
= m((n + 1) k) + O(k 3 ) .

Notice that we have made use in the above computation that, due to the unit length
constraint, one has

mtt (nk) = Pm⊥ (mtt (nk)) + (mtt (nk) · m(nk))m(nk)

= Pm⊥ (mtt (nk)) − |mt (nk)|2 m(nk) .

We therefore propose to modify the original first order scheme by replacing the
tangential update with the higher order approximation (16).
We will use the simplifying shorthand notation m = m(nk) and mt = mt (nk).
Let us proceed with the derivation of the equation that is satisfied by v = mt (nk) +
2 Pm⊥ mtt (nk), i.e. the counterpart to (10). The equation will be inferred from the
k

differentiated LLG equation which we restate as

123
 Author's personal copy
F. Alouges et al.

αmt + m × mt = Heff (m) − (Heff (m) · m) m (17)

by multiplying (1) with m×.

Remark 1 Although the midpoint rule is of order 2, our scheme is only almost of
order 2—as the section’s title suggests and as we will see in the sequel. This is due
to a regularizing term that is introduced in order to obtain the necessary estimates
in the convergence proof. This term prevents the scheme from being of order 2, in
the sense that the consistency error is not of order O(k 3 ) but only O(k 3− ) for any
> 0. On the other hand, this regularization approach allows for unconditional
convergence of the scheme. The approximation in space is still of order 1. Indeed,
only piecewise linear functions (P 1 Lagrange finite elements) are used in order to
ensure condition (13). It is well known that this condition, which for the time being is
mandatory to prove the convergence of the scheme, does not hold for higher order finite
elements.

To begin with, the differentiation of (17) w.r.t. time yields

 
∂Heff ∂Heff
αmtt + m × mtt = (mt ) − (mt ) · m m
∂m ∂m
−(Heff (m) · mt ) m − (Heff (m) · m) mt , (18)

where
∂Heff
(mt ) = d 2 mt + Hd (mt ) + Q (e · mt ) e
∂m

and where we once again used the unit length constraint (5). Now, we compute from
(17) and (18).

k
αv + m × v = αmt + m × mt + α Pm⊥ mtt + m × Pm⊥ mtt
 2 
k
= Pm⊥ αmt + m × mt + (αmtt + m × mtt )
2
 
k ∂Heff k
= Pm⊥ Heff (m) + (mt ) − (Heff (m) · m)mt
2 ∂m 2

that we write in weak form

  
k ∂Heff
αv ·  + m × v ·  dx = Heff (m) ·  dx + (mt ) ·  dx
2 ∂m
Ω Ω Ω

− 2 (Heff (m) · m) mt ·  dx
k

for any test function  with  · m = 0. Observe that mt (nk) = v + O(k), cf. (16).
Therefore up to higher order terms

123
 Author's personal copy
A convergent and precise finite element scheme

   
k k ∂Heff
α + (Heff (m) · m ) v ·  + m × v ·  dx − (v) ·  dx
2 2 ∂m
Ω Ω

= Heff (m) ·  dx + O(k 2 ), (19)
Ω

where we remind that m = m(nk) and mt = mt (nk). Observe that, if we neglect the
residual term, the latter equation is (at first sight surprisingly) linear in v. However, its
well-posedness is non-obvious since both the first and the last contributions on the l.h.s.
of (19) potentially affect the definiteness of the symmetric part of the operator. In order
to guarantee solvability and uniqueness we proceed with higher order modifications
that will finally lead to a well posed formulation. We address the first contribution and
define (see Fig. 1)
⎧
⎪ k
⎪
⎪α + min(x, M) for x ≥ 0,
⎪
⎪ 2
⎨
ϕ̃ M (x) = α (20)
⎪
⎪ for x < 0.
⎪
⎪
⎪
⎩1 + k
min(−x, M)
2α

Notice that ϕ̃ M (x) = α + k

2 min(x, M) + O(k 2 M 2 ). By abuse of notation we define

ϕ M (m) = ϕ̃ M (Heff (m) · m). (21)

As long as Heff (m) · m is uniformly bounded, we derive from (19) by plugging in

(21) that
 
∂Heff
ϕ M (m) v ·  + m × v ·  dx − k
(v) ·  dx
2 ∂m
Ω Ω

= Heff (m) ·  dx + O(k 2 ). (22)
Ω

Fig. 1 The regularizing cut-off function ϕ̃ M (x)

123
 Author's personal copy
F. Alouges et al.

Replacing Heff and ∂H eff

∂m by their very definition, we obtain the counterpart to (10) for
our new second order scheme:
 
k
ϕ M (m) v · +m × v ·  dx+ d 2 ∇v · ∇−Hd (v) ·  − Q(e · v)(e · ) dx
2
Ω Ω

= − d ∇m · ∇ + Hd (m) ·  + Q(e · m)(e · ) + Hext ·  dx.
2
(23)
Ω

We introduce only one further, final modification which implements the strategy delin-
eated in Remark 1: In order to maintain unconditional convergence we additionally
modify the second highest order term on the r.h.s. in the following way
 
k k
d ∇v · ∇ dx
2
 (1 + ρ(k)) d 2 ∇v · ∇ dx,
2 2
Ω Ω

where ρ(k) → 0 as k → 0. Take note that for ρ decreasing at least linearly, quadratic
order is conserved. However, only in case that ρ is slightly sublinear, for example
ρ(k) = k| ln(k)|, do we in fact achieve unconditional convergence.
Adopting Algorithm 1, we arrive at the following scheme:

Algorithm 2 Given an initial m0 ∈ Mh choose a time step size k = T

N with N ∈ N
and appropriate ρ(k) and M, cf. Theorem 2. For n = 0, 1, . . . , N
⎡
a) Find vn ∈ K mn such that for all test functions  ∈ K mn
⎢
⎢
⎢
⎢ ϕ M (mn ) vn ·  + mn × vn ·  dx
⎢
⎢Ω 
⎢
⎢
⎢ + k2 (1 + ρ(k)) d 2 ∇vn · ∇ − Hd (vn ) ·  − Q(e · vn )(e · ) dx
⎢
⎢  Ω (24)
⎢
⎢
⎢ = − d 2 ∇mn · ∇ + Hd (mn ) + Hext + Haniso (mn ) ·  dx.
⎢
⎢ Ω
⎢
⎢
⎢ 
⎣ min + kvin
b) Set mn+1 = min+1 φih , where ∀i, min+1 = .
|min + kvin |
i

The appropriate choice of ρ and M can be inferred from our convergence result, see
Theorem 2.
Let us sum up: the new scheme replaces the search of v as solution to (11) by the
search of v as a solution to (24). Besides this substitution, the algorithm outlined in
Sect. 4 remains as before in the sense that the renormalization and the interpolation
w.r.t. time are left unchanged. Since Eq. (24) is linear in v, our algorithm is very
favorable in practice.

123
 Author's personal copy
A convergent and precise finite element scheme

Before we state our theorem about the convergence let us explicitly make a statement
about its order.

Proposition 1 Consider a smooth (in space and time) solution m to (1) at time t = nk
and a semi-discrete time-approximation to m at time t + k on the basis of (24).
More precisely, given m at time t = nk determine v as a solution to the variational
formulation (24) with ρ(k) = 0 and M(k) sufficiently large and set

m(x, t) + kv(x, t)
m̃(x, t + k) = for all x ∈ Ω.
|m(x, t) + kv(x, t)|

Then m̃(t + k) approximates m(t + k) up to cubic error in k.

Argument for Proposition 1 The proof is a direct consequence of the Taylor expansion
performed in (16).

Remark 2 The smoothness of solutions to (1) has been widely studied during the past
years. In general, the formation of singularities cannot be ruled out and we can usually
not assume that a solution to the initial value problem will be regular. Our statement
about the order of the approximation is thus only a first little step on the way to a proof
of the order of convergence, which is way beyond the scope of this paper. Let us insist
on the fact that, to the knowledge of the authors, there does not yet exist any numerical
scheme for LLG equations for which an order of convergence can be proven.

Let us now turn to the convergence result.

Theorem 2 Let m0 ∈ H 1 (, S 2 ). Suppose m0 → m0 in H 1 () as h → 0. If the

regular sequence of conformal triangulations (Th )h>0 satisfies condition (13), then the
approximation (mh,k ) of the sequence constructed via Algorithm 2 and interpolated
according to Definition 1 converges (up to the extraction of a subsequence) weakly in
H 1 (ΩT ) to a weak solution m of (1) as h and k tend to 0 provided ρ(k) →k→0 0 and
one of the two following conditions hold:
– k −1 ρ(k) →(h,k)→0 ∞ and k M →(h,k)→0 0 or
– ρ ≡ 0 and kh −1 → 0 as (h, k) → 0.

Proof of Theorem 2 As stated before, the proof consists of two main steps: estab-
lishing estimates which guarantee the existence of a sufficiently strong converging
subsequence, then proving that the latter converges indeed to a solution (which sat-
isfies the energy estimate). Several arguments in the proof have already been used in
e.g. [1,3,8], but we restate them for the self-consistency of the paper. We will need
the following classical estimate from elliptic regularity theory, namely

||Hd (m)|| L p (Ω) ≤ C||m|| L p (Ω) , (25)

for all p ∈ (1, +∞) and a positive constant C which depends only on p.

123
 Author's personal copy
F. Alouges et al.

Bounds on the sequence As we have already observed, the variational formulation in

the iteration of (23) has a unique solution vn . We test the equation with  = vn itself
to find that
 
k
ϕ M (mn ) |vn |2 dx + (1 + ρ(k)) d 2 |∇vn |2 − Hd (vn ) · vn − Q(vn · e)2 dx
2
Ω Ω

= − d 2 ∇mn · ∇vn + Hd (mn ) · vn + Q(e · mn )(e · vn ) + Hext · vn dx. (26)
Ω

Since we assume that the triangulation Th satisfies the angle condition (13) we have
that
  2 
   2
∇mn+1  dx ≤ ∇(mn + kvn ) dx
Ω Ω
  
   n 2
≤ ∇mn 2 dx + 2k ∇m · ∇v dx + k
n n 2 ∇v  dx.
Ω Ω Ω

Using (26) we obtain that

  2  
 n+1 
   2
d 2
∇m  dx ≤ d 2  ∇m n 2
dx − 2k ϕ M (mn ) vn  dx
Ω Ω Ω

+k 2
(Hd (v ) · v + Q(e · vn )2 ) dx
n n

Ω

+2k (Hd (mn ) + Haniso (vn ) + Hext ) · vn dx
Ω

 n 2
−k ρ(k) d 2
2 ∇v  dx. (27)
Ω

Before we move on, let us just rewrite the latter estimate as

  2  
     
d 2 ∇mn+1  dx ≤ d 2 ∇mn 2 dx − 2k ϕ M (mn ) vn 2 dx
Ω Ω Ω
 
∂ H̄eff n
+k 2 (v ) · vn dx + 2k H̄eff (mn ) · vn dx
∂m
Ω Ω

 
−k 2 ρ(k) d 2 ∇vn  dx.
2
(28)
Ω

We partially neglect the negative contributions on the r.h.s. of (28)—those which are
quadratic in vn —and use (25) to obtain

123
 Author's personal copy
A convergent and precise finite element scheme

  2  
     
d2 ∇mn+1  dx ≤ d 2 ∇mn 2 dx − 2k ϕ M (m) vn 2 dx
Ω Ω Ω

 n 2
+2k||H̄eff (m )|| L 2 ||v || L 2 − k 2 ρ(k)d 2
n n ∇v  dx
Ω
 
   2
≤ d2 ∇mn 2 dx − 2k ϕ M (m) vn  dx
Ω Ω

 n 2
+Ck||vn || L 2 − k ρ(k) d 2
2 ∇v  dx, (29)
Ω

where the generic constant C depends on Q and |Ω|. Due to Young’s inequality, we
2
have that Ck||vn || L 2 (Ω) ≤ kβ||vn ||2L 2 (Ω) + kC
4β for β > 0. Using the uniform bound

α
ϕ M (m) ≥ β =
1 + k2 M

we find by rewriting (29) that

  2 
   n 2
d2 ∇mn+1  dx + βk||vn ||2L 2 (Ω) + k 2 ρ(k) d 2 ∇v  dx
Ω Ω

 
∇mn 2 dx + kC (Q, |Ω|) .
2
≤ d2 (30)
4β
Ω

Summing up in (30) over the time steps we find that

  2  −1  
 N
N
 n 2  n 2
d2
∇m  dx + βk  v  dx + k ρ(k) d
2 2 ∇v  dx
Ω n=0 Ω Ω
⎛ ⎞

≤ C ⎝T, d 2 |∇m0 |2 dx, β, Q, Haniso ⎠ (31)
Ω

From now on, most of the arguments follow the same line as in [1]. It holds that
 
 mn+1 − mn 
 i i 
  ≤ |vin |, for all n ≤ N , and i ∈ {1, · · · , Nh }.
 k 

Moreover, there exists c > 0 such that for all 1 ≤ p < +∞ and all φh ∈ Vh there
holds
1 p
 p
||φh || L p (Ω) ≤ h d |φh (xih )| p ≤ c||φh || L p (Ω) , (32)
c
i

123
 Author's personal copy
F. Alouges et al.

which implies
 n+1 
m − mn 
  ≤ c2 ||vn || L 2 . (33)
 k 
L2

Hence we obtain from the energy estimate (31) using (33) the following bounds

mh,k is uniformly bounded in H 1 (ΩT ), (34)

vh,k is uniformly bounded in L (ΩT ). 2
(35)

Due to (34) and (35), there exist m̄ ∈ H 1 (ΩT ) and v ∈ L 2 (ΩT ) such that up to the
extraction of subsequences

mh,k (h,k)→0 m̄ weakly in H 1 (ΩT ), (36)

mh,k →(h,k)→0 m̄ strongly in L (ΩT ), 2
(37)
vh,k (h,k)→0 v weakly in L (ΩT ).2
(38)

In addition, we have from (31) that

 −1  T 
N
 n 2  
k ρ(k)
2 ∇v  dx = kρ(k) ∇vh,k 2 dx ≤ C < +∞.
n=0 Ω 0 Ω

If ρ decreases only sublinearly, i.e. k −1 ρ(k) →k→0 +∞, we deduce that

k ||∇v|| L 2 (ΩT ) →(h,k)→0 0. (39)

If ρ decreases linearly or faster we have to resort to the inverse estimate ||∇v|| L 2 (ΩT ) 
−1 → 0 when both (h, k) → 0.
h ||v|| L 2 (ΩT ) so that (39) holds true if kh
1

Preliminary estimates We want to prove that m̄ satisfies (7) and follow the strategy
of [1]. To begin with, we restate some further estimates from [1] and derive some
necessary statements about convergence. Observe that for all n = 0, · · · , J and all
t ∈ [nk, (n + 1)k)
  n+1 
 m (x) − mn (x)   
−
|mh,k (x, t) − mh,k (x, t)| 
= (t − nk) ≤ k ∂t mh,k (x, t) .
k 

Therefore
 
−
||mh,k − mh,k || L 2 (ΩT ) ≤ k ∂t mh,k  L 2 (Ω →(h,k)→0 0,
T)

which entails that

−
mh,k →(h,k)→0 m̄ strongly in L 2 (ΩT ).

123
 Author's personal copy
A convergent and precise finite element scheme

Moreover, on any tetrahedron K of Th , and for any u ∈ Mh , and any vertex xih of K ,
one has
 2
 
|u(x)| − |u(xih )| ≤ Ch 2 |∇u|2 ,

−
(recall that ∇u is constant on K ), from which one deduces (since |mh,k (xih )| = 1)
  2
 −  − 2
1 − |mh,k | dx ≤ Ch 2 ||∇mh,k || L 2 (Ω ) .
T
ΩT

This shows that |m̄(x, t)| = 1 a.e. (x, t) ∈ ΩT .

Eventually, from the fact that at each vertex ∀i ∈ {1, . . . , Nh }

1 2 n2
|min+1 − min − kvin | = |min + kvin | − 1 ≤ k |vi | , (40)
2

we derive
 
 mn+1 − mn  1
 i i n
 − vi  ≤ k|vin |2 .
 k  2

Appealing to (32) the latter entails that

 
∂t mh,k − vh,k  1 ≤ c2 k||vh,k ||2L 2 (Ω →(h,k)→0 0.
L (Ω T) T)

This is sufficient to conclude that v = ∂t m̄ in (38).

General properties of interpolation operator Before we start with the penultimate step
of proving convergence, let us state some general properties of the nodal interpolation
operator which we repeatedly use in the sequel. Up to dimension three, there holds
for any function ϕ ∈ H 2 (Ω) ⊂ C 0 ()¯

||ϕ − Ih (ϕ)|| H 1 (Ω) ≤ Ch||∇ 2 ϕ|| L 2 Ω . (41)

Since the basis functions are linear on each triangle one can deduce from (41) that

− ˜ − Ih (m− × )||
˜ L 2 ([0,T ],H 1 ) ≤ Ch||m− || H 1 (Ω ) ||||W 2,∞ ,
||mh,k × h,k h,k T
(42)

see [1, p.7].

Convergence to a solution of the LLG equation Having established the preliminary

results above, we are now ready to proceed with the proof of convergence. We recall

123
 Author's personal copy
F. Alouges et al.

− ˜ where
that Ih is the nodal interpolation, cf. (9), and test (24) with  = Ih (mh,k × )
˜ ∞
 ∈ C0 (ΩT ) . After suitable integration in time we get
3

 
− − ˜ dx dt + − − ˜ dx dt
ϕ M (mh,k ) vh,k · Ih (mh,k × ) mh,k × vh,k · Ih (mh,k × )
ΩT T

k −
+ (1 + ρ(k)) d 2 ∇vh,k · ∇Ih (mh,k ˜ − Hd (vh,k ) · Ih (m− × )
× ) ˜
h,k
2
T
− ˜ dx dt
−Q(e · vh,k )(e · Ih (mh,k × ))

− − ˜ + Hd (mh,k ) · Ih (m− × )
˜
= − d 2 ∇mh,k · ∇Ih (mh,k × ) h,k
T
− ˜ + Hext · Ih (m− × )
˜ dx dt.
+Q(e · mh,k )(e · Ih (mh,k × )) h,k (43)

Our goal is to pass to the limit (k, h) → 0 in the latter equation (43) to recover the
LLG equation (10). As we shall see, the first and the third term on the l.h.s. and the
first term on the r.h.s. are a little bit subtle and have to be treated with caution. The
remaining contributions behave well under the established convergence; this is par-
ticularly due to the fact that Hd is L 2 -continuous. For the second contribution on the
l.h.s. one further uses that the L ∞ bound on m− improves (37) to strong convergence
in any L p with 1 < p < +∞.
Observing that |ϕ M | is uniformly bounded, and that |ϕ M − α| ≤ k 2M , as long as
−
k M → 0 for both (h, k) → 0 the strong convergence of mh,k is sufficient to conclude
that
 
− − ˜ dx dt →(h,k)→0 α ˜ dx dt. (44)
ϕ M (mh,k ) vh,k · Ih (mh,k × ) v · (m̄ × )
ΩT T

Indeed, using the triangle inequality we find that

 
  
 
 − − ˜ dx dt 
˜ dx dt − α v · (m̄ × )
 ϕ M (mh,k ) vh,k · Ih (mh,k × )
n
 
ΩT T 
 
  
 
 − − ˜ dx dt 
˜ dx dt − α v · (m̄ × )
≤ ϕ M (mh,k ) vh,k · (mh,k × )
 
 ΩT T 
 
 
 
 − − −
˜ − Ih (m × ))˜ dx dt  .
+ ϕ M (mh,k ) vh,k · ((mh,k × ) (45)
 h,k 
 ΩT 

123
 Author's personal copy
A convergent and precise finite element scheme

The first term tends to zero since

ϕ M (m− ) →(h,k)→0 α in L ∞ (Ω),

−
mh,k →(h,k)→0 m̄ in L 2 (T ), and
∂ m̄
vh,k →(h,k)→0 v = in L 2 (T )
∂t
as h, k → 0.
−
Since ϕ M (mh,k ) is uniformly bounded, we can evoke (42) to obtain that the second
contribution tends to zero. This establishes (44).
Let’s turn to the next term in (43). Convergence here essentially relies upon the
estimate (39). In fact, appealing once again to (42) we see that instead of establishing

− ˜ dx →(h,k)→0 0,
k
2 d2 (1 + ρ(k))∇vh,k · ∇Ih (mh,k × ) (46)
T

it suffices to show

− ˜ dx →(h,k)→0 0,
k
2 d2 (1 + ρ(k))∇vh,k · ∇(mh,k × ) (47)
T

which obviously follows from (39) by Young’s inequality.

Finally, the convergence of the last term in (45) follows from the orthogonality
property of the cross product and (36), (37) by once again appealing to (42) since
 
  

k 
− −
 ∇mh,k · ∇Ih (mh,k × ) ˜ dx dt − ∇ m̄ · m̄ × ∇  ˜ dx dt 
2 
T T 
 
 
k  −

− −
 

≤  ˜ − (m × )
∇mh,k · ∇ Ih (mh,k × ) ˜ dx dt 
2 h,k 
 T 
 
 
k  − −


+  ˜ ˜
∇mh,k · ∇(mh,k × ) − ∇ m̄ · m̄ × ∇  dx dt 
2 
 T 
 
 
k    
− − −
˜ − (m × ) ˜ 
=  ∇mh,k · ∇ Ih (mh,k × ) dx dt 
2 h,k 
 T 
 
 
k  − −


+  ˜ ˜
∇mh,k · (mh,k × ∇ ) − ∇ m̄ · (m̄ × ∇ ) dx dt  . (48)
2 
 T 

123
 Author's personal copy
F. Alouges et al.

Energy estimate We finally establish the energy estimate. From (27) we deduce that
∀n ∈ {0, · · · , Nh }
 
 2
E(mn+1 ) − E(mn ) ≤ − k ϕ(mn ) vn  dx + k H̄eff mn · vn dx
Ω Ω
 
k2 ∂ H̄eff n k2  n 2
+ (v ) · vn dx − ρ(k) d 2 ∇v  dx
2 ∂m 2
Ω Ω
  
1
− H̄eff mn+1 + mn · (mn+1 − mn ) dx, (49)
2
Ω

n = H̄
eff (m ) for the remaining effective
cf. (3). Let us introduce the shorthand H̄eff n

field. We consider the contributions in (49) separately and start with the observation
that
 
1 n+1
k H̄eff · v dx −
n n
(H̄eff + H̄eff
n
) · (mn+1 − mn ) dx
2
Ω Ω
 
1 n+1
= H̄eff · (m
n n+1
− mn − kvn ) dx + (H̄eff − H̄eff
n
) · (mn+1 − mn ) dx.
2
Ω Ω

Hence due to (33) and (40) combined with (32)

   
 1 n+1 
n
 k H̄eff · vn dx − (H̄eff + H̄eff
n
) · (mn+1 − mn ) dx
2
Ω Ω
≤ Ck 2 (||vn || L 2 ||vn || L 4 + ||vn ||2L 2 ) (50)

The stray-field contribution can be bounded with the help of (25) with p = 4. The
contributions in the second line on the r.h.s. of (49) are of higher order in k. The first
term can be easily bounded using Young’s inequality:
 
 
 ∂ H̄eff n 
 (v ) · v dx ≤ C||vn ||2L 2 .
n
(51)
 ∂m
 
Ω

Plugging in (50) and (51) into (49) yields that


 2
E(mn+1 ) − E(mn ) + k ϕ(mn ) vn  dx
Ω
≤ Ck 2 (vn  L 2 + ||vn ||2L 2 + ||vn || L 2 ||vn || L 4 + ρ(k) d 2 ||∇vn ||2L 2 )
≤ C  k 2 (vn  L 2 + vn 2L 2 + ||vn || L 2 ||∇vn || L 2 + ρ(k) ||∇vn ||2L 2 ),

123
 Author's personal copy
A convergent and precise finite element scheme

where C denotes a generic constant. Here we made use of the classical Sobolev
embedding

||vn || L 4 ≤ C||∇vn || L 2 .

Summing from n = 0 to N − 1 leads to


−
E(m(N k)) − E(m(0)) + ϕ M (mh,k )|vh,k |2 dx dt
ΩT

≤ Ck(||vh,k || L 2 + ||vh,k ||2L 2 + ||vh,k || L 2 ||∇vh,k || L 2 + ρ(k) ||∇vh,k ||2L 2 ).

We are now ready to pass to the limit. Noticing once again that k ||∇vn || L 2 (ΩT ) is
uniformly bounded from (39) we derive that

 n 2
E(m(N k)) − E(m(0)) + α v  dx dt ≤ 0. (52)
ΩT

7 Numerical results

In this section, we report on two numerical experiments designed to test the scheme.
The first one is intended to confirm the order of the scheme (expected to be nearly
2); the second one investigates the effect of overdamping i.e. taking ρ as suggested in
Theorem 2. In both cases, we refer to numerical test cases that are well-known in the
physics literature of ferromagnetic materials (see for instance [12]).

7.1 Testing the order of the scheme

In submicron Permalloy dots, the two main contributions to the effective field Heff (m)
are the exchange field Hex (m) and the demagnetizing field Hd (m), solution to (4). In
our finite element code “feeLLGood”, Hd (m) is computed by the FMM [14].
The first study mentioned above addresses the dynamical relaxation of the mag-
netization towards equilibrium in a flat cylindrical dot which is 200 nm wide and
20 nm thick. Typical material parameters are A = 1.0 × 10−11 Jm−1 and Ms =
8.0 × 105 Am−1 corresponding to an exchange length

d = (2 A μ−1 −2 1/2
0 Ms ) = 5 nm.

The damping factor α in the Landau–Lifschitz equation is set to 0.05, and we have
taken γ0 = 2.21 × 105 mA−1 s−1 [(cf. (1)].
The initial state is a vortex whose core is displaced in the x-direction. Writing in
complex variables and renormalizing the radius of the dot to 1, the initial magnetization

123
 Author's personal copy
F. Alouges et al.

Fig. 2 Initial vortex distribution

in a unit disk with a = c = 0.5

distribution is then given by

⎧
⎪
⎪
2w
⎨ m x + im y = f (z)
1 + |w|2
where w = (53)
⎪
⎪ 1 − |w|2 max(1, | f (z)|)
⎩ mz =
1 + |w|2

z−a
with the complex function f (z) = ci 1−az̄ and z = x + i y. Parameters a ∈ [0, 1)
and c ∈ R respectively specify the displacement of the vortex core from the sample’s
center and the characteristic size of the vortex. Our simulation uses a = c = 0.5. The
corresponding initial magnetization distribution is depicted in Fig. 2.
Notice that our construction ensures a distribution locally tangent to the circular
boundary. It is well known that the stray-field energy strongly penalizes non-tangential
configurations in thin films and thus, at least over short times, the nucleation of further
vortices. The simulations were performed on a tetrahedral mesh of 46,016 elements
(8,355 nodes) with an in-plane mesh-size h = 4nm < d. The final simulation time is
T = 10 ns, while the bound M [(cf. (20)] is set s.t. k M (2α)−1 = 0.1.
Figure 3 shows the time evolution of the average magnetization components and
monitors the convergence of the scheme as the time step tends to 0. At large time the
vortex turns as expected around the dot’s center at a characteristic frequency of f =
0.62 GHz in agreement with the analytical value given by Guslienko’s ferromagnetic
resonance model [12] within a physically acceptable error of about 10 %.
During its first rotation, the vortex core dilates, dissipating excess energy—
predominantly exchange energy. The rotation and core expansion combined altogether
cause fast magnetization variations, as observed in Fig. 3 at short times.
We performed simulations with a range of time steps to study the order of the
scheme. A log–log plot of the energy gap of the solution at the intermediate time
t = 0.7 ns and an estimation corresponding to k → 0 is presented in Fig. 4. The gap
follows a quadratic power law, confirming the presumed quadratic order.

123
 Author's personal copy
A convergent and precise finite element scheme

Fig. 3 Temporal evolution of the average magnetization

Fig. 4 Energy gap versus time step following almost a quadratic law

7.2 Influence of ρ on stability

We now consider a Permalloy dot with an elliptic 120 nm × 75 nm cross section and
2.5 nm thickness. It is discretized into 23,862 tetrahedral elements (4,732 nodes). The
material parameters are the same as in the preceding experiment.
Starting from a saturated magnetization state along the [110] direction, the system
relaxes to its equilibrium configuration: a symmetric distribution roughly aligned with
the long axis of the ellipse. Again, the magnetization tends to be tangent to the ellipse
boundary, in order to minimize the demagnetizing energy.
In order to test the effect of the (slight) overdamping induced by the parameter ρ,
we performed two simulations with the same time step k = 1 ps but with different
values for ρ:

1. ρ = 0, and
2. ρ(k) = τk | ln τk | ≈ 0.022, corresponding to the characteristic time τ = 250 ps.

123
 Author's personal copy
F. Alouges et al.

Fig. 5 Time evolution of the average magnetization component along Oy with k = 1 ps obtained with
ρ = 0 and ρ ≈ 0.022

Fig. 6 Distribution of the y-component of the magnetization along the y-axis at t = 1ns for ρ = 0 (le f t)
and ρ ≈ 0.022 (right)

In Fig. 5, the evolution of the average magnetization seems almost identical (2 ns),
independent of ρ. A look at Fig. 6, however, reveals that the magnetization distributions
differ significantly, with an instability occurring in the case ρ = 0.
In Fig. 7, which shows the evolution of the energy, one clearly observes the stabiliz-
ing effect of ρ in the numerical scheme. This effect was observed in practice whatever
the time step used. For ρ = 0 it is necessary to drop down the time step (typically
below kc ≈ 0.5 ps in our case) to recover stability.

8 Summary

We started out with an introduction and discussion of the newly proposed tempo-
ral scheme. We proved its order 2 consistency to the equation, its uniform stability
and (weak) convergence. Numerical simulations on physical test cases show strong
evidence of good performance and confirm in practice the theoretical properties of
the algorithm stated in Theorem 2, namely the quadratic convergence and stabilizing
effect of ρ. A rigorous proof of the order 2 convergence is yet missing, though, due

123
 Author's personal copy
A convergent and precise finite element scheme

Fig. 7 Total energy versus time for ρ = 0 (r ed) and ρ ≈ 0.022 (blue)

to the very strong nonlinearities in the equations such a result is clearly out of reach
for the time being.

Acknowledgments FA and JS both acknowledge support from the chair “Mathematical Modelling and
Numerical Simulation, F-EADS-Ecole Polytechnique-INRIA-F-X” and the ANR project ANR-08-BLAN-
0199. JS acknowledges support from the German Science Foundation through the Collaborative Research
Center 611 “Singular Phenomena and Scaling in Mathematical Models”.

References

1. Alouges, F.: A new finite element scheme for Landau-Lifschitz equations. Disc. Cont. Dyn. Syst. Ser.
S 1(2), 187–196 (2008)
2. Alouges, F.: A new algorithm for computing liquid crystal stable configurations: the harmonic mapping
case. SIAM J. Numer. Anal. 34(5), 1708–1726 (1997)
3. Alouges, F., Kritsikis, E., Toussaint, J.-C.: A convergent finite element approximation for the Landau–
Lifschitz–Gilbert equation. Physica B 407 (2012)
4. Alouges, F., Conti, S., DeSimone, A., Pokern, Y.: Energetics and switching of quasi-uniform states in
small ferromagnetic particles. M2AN Math. Model. Numer. Anal. 38(2), 235–248 (2004)
5. Alouges, F., Jaisson, P.: Convergence of a finite element discretization for the Landau–Lifshitz equa-
tions in micromagnetism. Math. Model. Method. Appl. Sci. 16(2), 299–316 (2006)
6. Alouges, F., Soyeur, A.: On global weak solutions for Landau–Lifchitz equations: existence and
nonuniqueness. In: Nonlinear analysis, theory, methods and applications 18(11), 1071–1084 (1992)
7. Bartels, S.: Stability and convergence of finite-element approximation schemes for harmonic maps.
SIAM J. Numer. Anal. 43(1), 220–238 (2005)
8. Bartels, S., Ko, J., Prohl, A.: Numerical analysis of an explicit approximation for the Landau–Lifshitz–
Gilbert equation. Math. Comp. 77, 773–788 (2008)
9. Bartels, S., Prohl, A.: Convergence of an implicit finite element method for the Landau–Lifshitz–Gilbert
equation. SIAM J. Numer. Anal. 44(4), 1405–1419 (2006)
10. Brown, W.F.: Micromagnetics. Wiley, New York (1963)
11. Cimràk, I.: A survey on the numerics and computations for the Landau–Lifshitz equation of micro-
magnetism, pp. 277–309. ACME, Netherlands (2007)
12. Guslienko et al: Eigenfrequencies of vortex state excitations in magnetic submicron-size disks. J. Appl.
Phys. 91, 8037–8040 (2002)
13. Hubert, A., Schäfer, R.: Magnetic domains. Springer, Berlin (1998)

123
 Author's personal copy
F. Alouges et al.

14. Kritsikis, E., Toussaint, J.-C., Fruchart, O., Szambolics, H., Buda-Prejbeanu, L.: Fast computations of
magnetostatic fields by non-uniform fast Fourier transforms. Appl. Phys. Lett. 93, 132508 (2008)
15. Landau, L., Lifschitz, I.: On the theory of the dispersion of magnetic permeability in ferromagnetic
bodies. Phys. Zeitsch. der Sow. 8, 153–169 (1935)
16. Szambolics, H., Toussaint, J.-C., Buda-Prejbeanu, L., Alouges, F., Kritsikis, E., Fruchart, O.: Innovative
weak formulation for the LLG equation. IEEE Trans. Magn. 44(11), 3153–3156 (2008)
17. Vanselow, R.: About Delaunay triangulations and discrete maximum principles for the linear conform-
ing FEM applied to the Poisson equation. Appl. Math. 46(1), 13–28 (2001)
18. Visintin, A.: On Landau–Lishitz equations for ferromagnetism. Jpn. J. Appl. Math. 2, 69–84 (1985)

123