J. Nonlinear Sci. Vol. 14: pp.

429–451 (2004)
DOI: 10.1007/s00332-004-0634-9

© 2005 Springer Science+Business Media, Inc.

Epitaxial Growth Without Slope Selection:

Energetics, Coarsening, and Dynamic Scaling
Bo Li1 and Jian-Guo Liu2
1
Department of Mathematics, University of California at San Diego, 9500 Gilman Drive, Mail
Code 0112, La Jolla, CA 92093-0112, USA.
e-mail: bli@math.ucsd.edu
2
Department of Mathematics and Institute for Physical Science and Technology, University
of Maryland, College Park, MD 20742-4015, USA.
e-mail: jliu@math.umd.edu

Received March 1, 2004; accepted September 14, 2004

Online publication January 7, 2005
Communicated by R. Kohn

Summary. We study a continuum model for epitaxial growth of thin films in which the
slope of mound structure of film surface increases. This model is a diffusion equation for
the surface height profile h which is assumed to satisfy the periodic boundary condition.
The equation happens to possess a Liapunov or “free-energy” functional. This functional
consists of the term |
h|2 , which represents the surface diffusion, and − log(1 + |∇h|2 ),
which describes the effect of kinetic asymmetry in the adatom attachment-detachment.
We first prove for large time t that the interface width—the standard deviation of the
height profile—is bounded above by O(t 1/2 ), the averaged gradient is bounded above
by O(t 1/4 ), and the averaged energy is bounded below by O(− log t). We then consider
a small coefficient ε 2 of |
h|2 with ε = 1/L and L the linear size of the underlying
system, and study the energy asymptotics in the large system limit ε → 0. We show that
global minimizers of the free-energy functional exist for each ε > 0, the L 2 -norm of
the gradient of any global minimizer scales as O(1/ε), and the global minimum energy
scales as O(log ε). The existence of global energy minimizers and a scaling argument
are used to construct a sequence of equilibrium solutions with different wavelengths.
Finally, we apply our minimum energy estimates to derive bounds in terms of the linear
system size L for the saturation interface width and the corresponding saturation time.

Key words. epitaxial growth, coarsening, interface width, dynamic scaling, energy min-
imization, equilibrium

MSC numbers. 35K55, 35Q99, 74K35, 82D25

PAC numbers. 68.35.Ct; 68.43.Jk; 81.15.Aa

430 B. Li and J.-G. Liu

1. Introduction

In a typical layer-by-layer epitaxial growth that begins with a flat substrate, surface
morphological instabilities often occur as the film thickness reaches a critical value.
These instabilities manifest themselves as a sort of spinodal decomposition. As a result,
the nucleation of islands starts and many nuclei appear on the film surface. Such nuclei
evolve into mounds, and the mound structure coarsens. During the coarsening process,
the number of mounds decreases. Experiments and numerical simulations suggest that
the well-characterized lateral size of mounds, λ(t), increases as λ(t) ∝ t n , where t is the
time variable and n > 0 is a constant called the coarsening exponent. The interface width
w(t), which is the standard deviation of the height profile and measures the roughness
of the surface, also increases as w(t) ∝ t β , where β > 0 is a constant called the growth
exponent. When the finite size of the underlying system becomes effective, the interface
width saturates, and the saturation value ws = ws (L) satisfies that ws (L) ∝ L α , where
L is the linear size of the underlying system and α > 0 is a constant called the roughness
exponent. The corresponding saturation time ts = ts (L) satisfies the dynamic scaling
law ts (L) ∝ L z , where z is a constant called the dynamic exponent. In general, z = α/β.
See Figure 1. These scaling laws are often experimentally measurable, and contain much
microscopic information. See [2], [3], [6], [8], [12], [16], [21], [23], [25]–[28], [32]–[37]
and the references therein.
To understand these interesting phenomena, we consider in this work the diffusion
equation


∇h
∂t h = −∇ · −
2 h (1.1)
1 + |∇h|2
for the height profile h, measured in a co-moving frame, of a thin film in epitaxial growth.
We assume that the height profile h is spatially -periodic, where = (0, L)d ⊂ Rd is
a d-dimensional, open cube with d ≥ 1 an integer and L > 0 the linear size of the cube,
and is the closure of , i.e., = [0, L]d .
For epitaxial growth, d = 2. In this case, the biharmonic term in (1.1) describes the
surface diffusion [10], [22]. The nonlinear, lower-order term in (1.1) was first proposed

ws

β
t

0 ts t
Fig. 1. Scaling laws in epitaxial growth.
Epitaxial Growth without Slope Selection 431

phenomenologically in [12] to model the Ehrlich-Schwoebel effect: Adatoms (adsorbed

atoms) must overcome a higher energy barrier to stick to a step from an upper rather than
from a lower terrace [5], [30], [31]. See also [17], [36] for issues related to the model
derivation. The equation describes a physical process that is far from equilibrium, and
is thus not associated with the free energy of the system in the usual sense.
However, (1.1) happens to possess the following Liapunov or “free-energy” func-
tional,
  
1   1
E(h) = − − log 1 + |∇h| + |
h| d x,
2 2
(1.2)
2 2
where
 
1
− u dx = u dx
| |
is the averaged integral over of an integrable function u : → R, and | | = L d is
the d-dimensional volume of . This means that (1.1) is the gradient flow associated
with the energy functional (1.2) with a suitable constant mobility. Note that we consider
here, for convenience, the averaged integral in (1.2).
With the periodical boundary condition, any solution h of (1.1) satisfies

d
h(x, t) d x = 0,
dt
i.e., the mass is conserved. For simplicity, we thus assume that the constant, spatial
mean-value of h over is zero. In this case, the interface width of the profile h becomes

wh (t) = − |h(x, t)|2 d x.

If h is smooth and -periodic, then we have by integration by parts that

  d 2 d 
|
h| d x =
2
∂xi xi h dx = ∂xi xi h ∂x j x j h d x
i=1 i, j=1
 
d d
 2
= − ∂xi xi x j h ∂x j h d x = ∂xi x j h d x. (1.3)
i, j=1 i, j=1

Thus, the term |
h|2 in the energy functional (1.2) can be replaced by i,d j=1 |∂xi x j h|2 ,
which has all the second-order derivatives. This property follows from the periodicity
along the coordinate directions, and does not imply that the surface diffusion is isotropic.
Using the identities


∇h 1 − |∇h|2 2|∇h|3
∇· =
h + κ
1 + |∇h|2 (1 + |∇h|2 )2 (1 + |∇h|2 )2
and

h = ∂ 2 h + |∇h|κ,
432 B. Li and J.-G. Liu
 ∇h  ∇h
where κ = ∇ · |∇h| is the mean curvature of level curves h = constants and ∂ = |∇h| ·∇
denotes the derivative tangential to the surface, we can rewrite our underlying growth
equation (1.1) as

|∇h|2 − 1 2 |∇h|
ht = ∂ h− κ −
2 h. (1.4)
(1 + |∇h|2 )2 1 + |∇h|2

Thus, the nonlinear term in the equation describes the anisotropic diffusion in epitaxial
growth: The curvature term represents the diffusion in the transverse direction and the
∂ 2 h term represents the diffusion in the direction of the surface gradient.
The linearization of (1.4) around a flat surface h 0 (x) = m · x for a given vector
m = (m 1 , m 2 ) ∈ R2 leads to the linearized equation of the growth equation (1.1) [18],
[29]
|m|2 − 1 2 1
∂t h = ∂ h− ∂ 2 h −
2 h,
(1 + |m|2 )2 m 1 + |m|2 ⊥m
where ∂ m = m · ∇/|m| and ∂⊥m = m ⊥ · ∇/|m| with m ⊥ = (−m 2 , m 1 ) denote the
derivatives parallel and perpendicular to the tilt vector m, respectively. Clearly, a singular
surface (m = 0) is linearly unstable. The change of the sign in the coefficient of ∂ 2m h
at |m| = 1 indicates the transition to step-flow growth. Since the coefficient of ∂⊥m 2
h
is always negative, the step-flow growth is linearly unstable with respect to transverse
fluctuations [4], [18].
Heuristic scaling analysis with an assumption of the existence of scaling laws and
large-scale numerical simulations of this model suggest that the interface width w(t)
and the lateral size of mounds λ(t) grow as O(t 1/2 ) and O(t 1/4 ), respectively [8], [11],
[24], [28], [33]. Thus, the characteristic slope of mounds scales as O(t 1/4 ) and becomes
unbounded. Moreover, this no-slope-selection model predicts that the saturation inter-
face width ws (L) ∝ L 2 . Consequently, the predicted growth, roughness, and dynamic
exponents are
1
β= , α = 2, z = 4. (1.5)
2
In contrast to this model without slope selection, the model with slope selection is
governed by the (scaled) free energy
  
1 2 1
Ẽ(h) = |∇h|2 − 1 + |
h|2 d x, (1.6)
4 2

where the first term selects the (scaled) mound slope 1. This model predicts the exponent
β = 1/3 [15], [16], [21], [23], [28]. Recently, Kohn and Yan [15] rigorously proved an
averaged version of a one-sided bound for this one-third law.
For both of the models, we showed in [20] a nonlinear morphological instability in the
rough-smooth-rough pattern that is experimentally observed [9] and the well-posedness
of the corresponding initial-boundary-value problems.
Setting ε = 1/L, we can rescale the “energy” to get

E(ĥ) = E ε (h), with h(x) = ĥ(x̂)/L and x̂ = L x, (1.7)

Epitaxial Growth without Slope Selection 433

where
  
1   ε2
E ε (h) = − log 1 + |∇h|2 + |
h|2 d x, (1.8)
1
2 2
and 1 = (0, 1)d is the unit cube in Rd . The related gradient flow, which is the same as
the scaled (1.1), is


∇h
∂t h = −∇ · − ε 2
2 h. (1.9)
1 + |∇h|2
Our goal of this work is to understand the energetics, coarsening, and dynamic scaling
of the interfacial dynamics in epitaxial growth without the slope selection, and to justify
rigorously the scaling laws predicted by the underlying model.
Our main results are as follows:
(1) For any solution h of (1.1), we show for large time t that

wh (t) ≤ O(t 1/2 ), (1.10)

 t  1/2
− − |∇h(x, τ )| d x dτ
2
≤ O(t 1/4 ), (1.11)
t0
 t
− E(h(τ )) dτ ≥ O(− log t). (1.12)
t0

See Section 2. All the bounds are independent of the dimension d and the system
size L. They are only one-sided. A two-sided bound is often not universally true.
For instance, an upper bound for the energy like E(h(t)) ≤ O(− log t) will not be
true for a steady-state solution h.
Note that our basic bounds lead to the O(−t 1/2 log t) lower bound for E(h(t))wh (t).
This is different from a constant lower bound for the same quantity in the slope-
selection model, cf. [15].
(2) For any ε > 0, we show that global minimizers of the scaled energy E ε defined
in (1.8) exist. Here, a global minimizer is a height profile that has the least energy
among the class of smooth and periodic height profiles. For small ε > 0, we also
show that


1
∇ hε L2( 1) = O
m
, (m = 0, 1, 2),
ε
for any energy minimizer of E ε and that

min E ε (h) ∼ log ε.

h

By a scaling argument, we can construct for each integer j ≥ 1 an equilibrium

solution h j of (1.1) with wavelength proportional to L/ j; cf. Section 4.
To better understand the variational properties of the model, we present in Section 3
some heuristic calculations in a one-dimensional setting of the rescaled energy (1.8)
for a trial profile and of the local shape of an equilibrium solution of the rescaled
equation (1.9).
434 B. Li and J.-G. Liu

(3) For any solution h of (1.1), any ξ ∈ R with 1/L < ξ < 1, and tξ > 0 such that
E(h(tξ )) = − log(ξ L), we show for large t that

 t 1/2

− [wh (τ )]2 dτ ≥ O(ξ 2 L 2 ).

tξ

Moreover, we show for any t > tξ that in fact

4(σ −1)

t ≥ O (ξ L) σ ,

where σ = t/tξ . See Section 5.

Our approach is different from that in Kohn and Otto [14] and Kohn and Yan [15]: We
do not need an isoperimetric inequality, since the O(t 1/2 ) upper bound on the interface
width can easily be obtained for the underlying model. Our analysis on the variational
problem of minimizing the energy helps us understand why the slope of mounds can
grow unboundedly. It also helps us determine a time scale for bounding the saturation
value of the interface width and the corresponding saturation time.
There are several important issues that we have not been able to address and resolve
in this work but that we wish to study further:
(1) An upper bound for the characteristic lateral size of mounds λ(t). For the slop-
selection model, this size λ(t) is of the same order as that of the height—the interface
width. Thus, there is no need to do extra work to bound λ(t). In general, a precise
mathematical concept that describes the lateral size λ(t) is needed.
(2) The optimality of bounds. For the slope-selection model, Ortiz, Repetto, and Si [23]
constructed a solution for the reduced dynamics that achieves the optimal bound. Can
one have a similar construction for the underlying model without slope selection?
The difficulty seems to lie in the fact that the energy (1.2) is not bounded below as
the system becomes larger and larger.
(3) The limiting dynamics as ε → 0. This is a nontrivial problem that is related to the
singular perturbation, or regularization, of a conservation law, cf. (1.9). From the
viewpoint of energy minimization, one may try to calculate the related -limit and
gradient flow of such a limit to obtain the reduced dynamic law, as is done in [23]
for the slope-selection model. But, again, the difficulty is the unboundedness of the
energy.
(4) The mathematical analysis and interpretation of the phase-ordering method. This
method is used to predict the underlying scaling laws (1.5) assuming a priori scaling
laws with certain exponents [8]. Mathematically, we understand little about such a
method. It would be interesting to explore such a method with the rigorous analysis
presented in Kohn and Otto [14], Kohn and Yan [15], and this work.
(5) Rigorous analysis of evolution equation models for epitaxial growth that do not pos-
sess any Liapunov functionals. Some good and important examples are the Kardar-
Parisi-Zhang and Lai-Das Sarma-Villain equations [13], [19], [36]. We expect that
such analysis can be more difficult and that a new framework is needed.
Epitaxial Growth without Slope Selection 435

2. Bounds on the Interface Width, Gradient, and Energy

∞
Let C per ( ) be the set of all restrictions onto of all real-valued, -periodic, C ∞ -
∞
functions on Rd . For any integer m ≥ 0, let Hperm
( ) be the closure of C per ( ) in the
usual Sobolev space W ( ) [1], [7]. Let
m,2

  
H( ) = h ∈ Hper 2
( ): − h dx = 0 .

It is clear that H( ) is a closed subspace of Hper

2
( ). Throughout the paper, we denote by
· the L -norm for an underlying domain. We also write a function u : ×[0, T ] → R,
2

which is in a function space X for each t as a mapping u = u(t) : [0, T ] → X .

Let T > 0 and h : [0, T ] → L 2 ( ). The interface width for h is defined for any
t ∈ [0, T ] by
 
wh (t) = − |h(x, t) − h̄(t)|2 d x, with h̄(t) = − h(x, t) d x. (2.1)

In particular,

wh (t) = − |h(x, t)|2 d x, ∀ h ∈ H( ).

Theorem 2.1 (bounds on the interface width, gradient, and energy). Let h(·): [0,∞) →
H( ) be a weak solution of (1.1) on (0, T ) for any T > 0 [20]. Let t0 ≥ 0.
(1) An upper bound on the interface width. We have

wh (t) ≤ 2(t − t0 ) + [wh (t0 )]2 , ∀ t ≥ t0 . (2.2)

(2) Upper bounds on the gradients. We have

 t
[wh (t0 )]2
− − |
h(x, τ )|2 d xdτ ≤ 1 + , ∀ t > t0 , (2.3)
t0 2(t − t0 )

and
 t
1/2
[wh (t0 )]2  1/2
− − |∇h(x, τ )|2 d xdτ ≤ 1 + t + t0 + [wh (t0 )]2 , ∀ t > t0 .
t0 2(t − t0 )
(2.4)
(3) A lower bound on the energy. We have
 t √ 
1
− E(h(τ ))dt ≥ − log 1 + 3t , ∀ t > t0 + [wh (t0 )]2 . (2.5)
t0 2

Note that, for large t, the bounds in (2.2), (2.4), and (2.5) are exactly those in (1.10),
(1.11), and (1.12), respectively.
436 B. Li and J.-G. Liu

Proof of Theorem 2.1. (1) By the definition of a weak solution [20], we have (see [28])
 

d |∇h|2
[wh (t)] = 2− hh t d x = 2−
2
− |
h| d x ≤ 2,
2
∀ t > 0. (2.6)
dt 1 + |∇h|2

Thus, integrating from t0 to t > t0 and then taking the square root, we obtain (2.2).
(2) It follows from (2.6) that
 
1 d |∇h|2
[wh (t)] + − (
h) d x = −
2 2
d x ≤ 1.
2 dt 1 + |∇h|2

Thus, we have for any t > t0 that

 t

1 1 1
− − |
h(x, τ )|2 d xdτ ≤ 1 + [wh (t0 )] − [wh (t)] ,
2 2
t0 t − t0 2 2

leading to (2.3).
Now, it follows from an integration by parts, the Cauchy-Schwarz inequality, (2.2),
and (2.3) that for any t > t0 ,
 t  t
− − |∇h(x, τ )|2 d xdτ = − − [−h(x, τ )]
h(x, τ ) d xdτ
t0 t0

 t  1/2
 t  1/2
≤ − − |h(x, τ )| d xdτ
2
− − |
h(x, τ )| d xdτ
2
t0 t0

 t 1/2
1/2
  [wh (t0 )]2
≤ − 2(τ − t0 ) + [wh (t0 )]2 dτ 1+
t0 2(t − t0 )


 1/2 [wh (t0 )]2
1/2
≤ t + t0 + [wh (t0 )]2 1+ .
2(t − t0 )

This proves (2.4).

(3) If t > t0 + [wh (t0 )]2 , then

[wh (t0 )]2 1

≤ and t + t0 + [wh (t0 )]2 ≤ 2t. (2.7)
2(t − t0 ) 2

Since − log is a convex function, we obtain by Jensen’s inequality, (2.4), and (2.7) that
 t
 t 
1
− E(h(τ )) dτ ≥ − log 1 + − − |∇h(x, τ )|2 d xdτ
t0 2 t0

1/2
1 [wh (t0 )]2  1/2
≥ − log 1 + 1 + t + t0 + [wh (t0 )]2
2 2(t − t0 )
1 √ 
≥ − log 1 + 3t ,
2
proving (2.5).
Epitaxial Growth without Slope Selection 437

3. Heuristic Calculations of Energetics and Equilibria

In this section, we assume the space dimension is d = 1 and consider the energy
functional (1.8) and the related gradient flow (1.9). We first calculate simple height
profiles that have low energies. We then determine the local shape of equilibria. Detailed
one-dimensional analysis on the scaling laws can be found in [24].

3.1. Energy of a Trial Profile

Let j ≥ 1 be an integer. Divide the interval [0, 1] into 2 j small intervals of the same
length 1/2 j. Let k > 0 and δ > 0 be real numbers with 2δ < 1/2 j. Define a trial function
h ∈ C 1 [0, 1] by


kx, if 0 ≤ x < 1
− δ,

 2 
4j



 − k
x − 1
+ k 1
− δ
, if 1
−δ ≤ x < 1
+ δ,



2δ 4j
 4j 2 4j 4j

h(x) = −k x − 2 j , +δ ≤ x < − δ,
1 1 3
if

 2 
4j 4j




k
x − 3
− k 1
− δ
, if 3
−δ ≤ x < 3
+ δ,

 2δ

4j 4j 2 4j 4j


k x − 1 , if 3
+ δ ≤ x ≤ 1j ,
j 4j

and



1 (i − 1) i
h x+ = h(x) ∀x ∈ , , i = 2, . . . , j.
j j j

This function is quadratic in each “transition region” (ci − δ, ci + δ) with ci = 12

[(i − 1)/2 j + i/2 j] (i = 1, . . . , 2 j) and linear with the slope k or −k elsewhere. See
Figure 2.

k -k
O 1/4j- δ 1/4j+ δ x
1/j 1

Fig. 2. A trial height profile.

438 B. Li and J.-G. Liu

Straightforward calculations lead to

 1 
1 2 ε 2 2
E ε (h) = − log(1 + h ) + h dx
0 2 2
   
1 ε2
= + − log(1+h 2 )+ h 2 d x
nontransition regions transition regions 2 2
1 2 jε 2 k 2
= − log(1 + k 2 ) + 2 jδ f (k) + ,
2 δ
where
 1
f (k) = log(1 + k ) −
2
log(1 + k 2 s 2 ) ds
0
 1
 

1 + k2 1 k 1 + k2
= log ds = log ds.
0 1 + k2s2 k 0 1 + s2
It is not difficult to see that f increases on (0, ∞) from 0 to 2 and f (k) ∼ 2k 2 /3 as
k → 0+ . √
Fix k. The energy E ε (h) is minimized at δ = εk/ f (k). At this value of δ, the energy
becomes
1 
E ε (h) = − log(1 + k 2 ) + 4 jεk f (k).
2
With varying k, this is minimized at k = g(k)/( jε), where
√
k 2 f (k)
g(k) = > 0.
2(1 + k 2 )[2 f (k) + k f  (k)]
√
We have g(k) > 0, since f (k) > 0 and f  (k) > 0. Moreover, g(k) → 2/8 as k → ∞,
since
 1
2k 2 1 − s2
k f  (k) = ds → 0, as k → ∞.
1 + k2 0 1 + k2s2
Thus, if 0 < jε  1, the optimal value of slope k has the asymptotics


1
k=O , for k  1. (3.1)
jε
With this k, the size of each transition region is


g(k) 1
δ= √ =O , for k  1, (3.2)
j f (k) j
and the minimum energy is
E ε (h) ∼ log( jε), as ε → 0+ . (3.3)
Our calculations indicate that the size of each transition region is of the same order
as that of the base of a mound, and in particular, it is independent of ε. Moreover, in
the large system limit ε → 0, the mound slope of a global minimizer is proportional to
the linear size of the underlying system. Finally, the minimum energy scales as log ε for
small ε > 0. All these properties are quite different from those of the slope selection
model.
Epitaxial Growth without Slope Selection 439

-a O a x

Fig. 3. The local shape of an equilibrium solution.

3.2. The Local Shape of an Equilibrium

We now consider the one-dimensional steady-state equation


h
+ ε 2 h (4) = 0, on (0, 1), (3.4)
1 + h 2
with the periodic boundary condition, cf. (1.9). We assume an equilibrium solution h is a
profile that consists of hills and valleys, similar to that shown in Figure 2. To understand
the local shape of such an equilibrium solution, we assume without loss of generality
that h is an even and convex function on [−a, a] for some real number a > 0. We also
assume that h  (±a) = ±k for some constant k > 0 and h  (±a) = 0. See Figure 3.
Set g = h  and integrate both sides of (3.4) to get
g
+ ε 2 g  = c1 ,
1 + g2
for some constant c1 . Clearly, c1 = 0, since g(0) = g  (0) = 0. Thus,
g
+ ε 2 g  = 0.
1 + g2
Multiply both sides of this equation by 2g  and integrate to get

log(1 + g 2 ) + ε 2 g 2 = c2 ,

where c2 is a constant. Since g(a) = k and g  (a) = 0, we obtain that c2 = log(1 + k 2 ).

Consequently,


1 + g2
log + ε 2 g 2 = 0. (3.5)
1 + k2
Now solving (3.5) with g  = dg/d x and only considering x ∈ [0, a], we get
 g
dg
x = x(g) = ε  .
0 1+k 2
log 1+g 2
440 B. Li and J.-G. Liu

Let 0 < σ < 1. With the change of variables g = kz, we obtain that
 σk  σ
dg kdz
x(σ k) = ε  =ε  .
0 1+k 2 0 1+k 2
log 1+g2 log 1+k 2 z2

Here, σ represents the ratio of the profile slope at the position x and the far-field slope
k. Clearly,
x(σ k) → 0, as σ → 0+ . (3.6)
Moreover, since (1 + k 2 )/(1 + k 2 z 2 ) ≤ 1/z 2 for z ∈ (0, 1), we get by changing the
variable y = 1/z that
 σ  ∞  ∞
dz dy dy
x(σ k) ≥ kε  = kε 2
√ → kε 2
√ , as σ → 1.(3.7)
0 log z12 1/σ y 2 log y 1 y 2 log y

The last integral on [1, ∞) is finite and independent of ε and k. It follows from (3.6) and
(3.7) that the size of the transition region is O(kε). This agrees with (3.1) and (3.2).
Multiplying both sides of (3.4) by h and integrating by parts, we get
 
1 1
h 2
ε 2 h 2 d x = d x ≤ 1.
0 0 1 + h 2

This, together with (3.5) which is generally satisfied by g = h  everywhere in (0, 1),
leads to
 1 
1 ε 2 2
E ε (h) = − log(1 + g ) + g
2
dx
0 2 2
 1 
1
= − log(1 + k 2 ) + ε 2 g 2 d x
0 2
∼ − log k, as k → ∞.

This agrees with our previous calculations of the energetics; cf. (3.1) and (3.3).

4. Energy Minimization

In this section, we study the variational problem of minimizing the scaled energy (1.8)
defined on 1 = (0, 1)d . Our main result is the following theorem, which is a rigorous
version of the basic properties we obtained from our heuristic calculations:

Theorem 4.1 (energy minimization).

(1) For any ε > 0, there exists h ε ∈ H( 1) such that

E  (h ε ) = min E ε (h).
h∈H( 1)
Epitaxial Growth without Slope Selection 441

(2) Denote eε = minh∈H( 1)

E ε (h). There exists a constant C1 > 0 such that

eε ≤ log ε + C1 , ∀ ε > 0. (4.1)

Moreover,
eε ∼ log ε, as ε → 0. (4.2)
√
(3) There exist two constants C2 > 0 and C3 > 0 such that for any ε ∈ (0, e−C1 / 2)
and any global minimizer h ε ∈ H( 1 ) of E ε : H( 1 ) → R,
C2 C3
≤ ∇ m hε ≤ , m = 0, 1, 2, (4.3)
ε ε
where ∇ 0 h = h, ∇ 1 h = ∇h, ∇ 2 h =
h, and · is the L 2 ( 1 )-norm.

A direct consequence of the existence of global minimizers is the following result

of the existence of a sequence of equilibrium solutions h j ( j = 1, . . .) of the original
equation (1.1) over = (0, L)d with each h j being L/ j -periodic, where L/ j =
(0, L/ j) :
d

Corollary 4.1. For any integer j ≥ 1, there exists h j ∈ H( ) that satisfies the following
properties:
(1) The function h j is L/ j -periodic. Moreover, if h ∈ H( ) is L/ j -periodic, then

E(h j ) ≤ E(h). (4.4)

In particular, for any integer I ≥ 0, we have

E(h 2 I ) ≥ · · · ≥ E(h 2i ) ≥ E(h 2i−1 ) · · · ≥ E(h 1 ) = min E(h). (4.5)

h∈H( )

(2) The function h j is an equilibrium solution of (1.1), i.e.,


∇h j
∇· +
2 h j = 0, in . (4.6)
1 + |∇h j |2

Proof. Fix j. Let h̃ j ∈ H( ) be a global energy minimizer of E j : H( ) → R, where

   
1 j2
E j (h̃) = − − log 1 + |∇ h̃| + |
h̃| d x, ∀ h̃ ∈ H( ).
2 2
2 2

Now, define h j (x) = (1/ j)h̃ j ( j x) for any x ∈ Rd . One easily verifies that h j is L/ j -
periodic, and that (4.4) holds true, since for any h ∈ H( ) that is L/ j -periodic,
1 1
E(h) = E j (h̃) ≥ d E j (h̃ j ) = E(h j ),
jd j

where h(x) = (1/ j)h̃( j x) for any x ∈ Rd . This proves (4.4), which in turn implies (4.5)
directly. Part (1) is proved.
442 B. Li and J.-G. Liu

Since h̃ j ∈ H( ) is a global energy minimizer of E j : H( ) → R, it is a critical

point, i.e., a weak solution of


∇h j
∇· + j 2
2 h j = 0, in . (4.7)
1 + |∇h j |2

Equivalently,


∇h j
j
(
h j ) = −∇ ·
2
, in .
1 + |∇h j |2
Thus, by a simple argument using the regularity theory of elliptic equations and the fact
that h j is periodic, we see that h j is smooth and satisfies (4.7) pointwise. Consequently,
(4.6) follows from the scaling h j (x) = (1/ j)h̃ j ( j x).

To prove Theorem 4.1, we need several lemmas. For the first lemma, see Figure 4.

Lemma 4.1. (1) If µ ≥ 1, then

log(1 + s) < µs, ∀ s > 0. (4.8)

(2) If µ ∈ (0, 1), then there exists a unique sµ ∈ (0, ∞) such that

log(1 + sµ ) = µsµ , (4.9)

and that
log(1 + s) > µs, if s ∈ (0, sµ ),
(4.10)
log(1 + s) < µs, if s ∈ (sµ , ∞).
(3) As a function of µ, sµ defined above increases from 0 to ∞ as µ decreases from 1
to 0. Moreover,
log(1 + sµ )
lim = 1. (4.11)
µ→0+ log µ1

Proof. (1) This follows from the fact that

1 + s < es ≤ eµs ∀ s > 0.

(2) Fix µ ∈ (0, 1). Let f (s) = log(1+s)−µs (s > −1) and s0 = 1/µ−1 > 0. Then,
f  (s0 ) = 0. We have that f (s) > 0 on (0, s0 ), since f  (s) > 0 on (0, s0 ) and f (0) = 0.
Similarly, f  (s) < 0 on (s0 , ∞), and f (s0 ) = µ−log µ−1. This is a decreasing function
of µ ∈ (0, 1), since the derivative with respect to µ ∈ (0, 1) is negative. Moreover, its
value at µ = 1 is 0. Thus, f (s0 ) > 0. Also, f (s) → −∞ as s → ∞. Thus, there exists
a unique sµ ∈ (s0 , ∞) that satisfies (4.9) and (4.10).
(3) Notice that µ = s −1 log(1 + s) defines a continuously differentiable function for
s ∈ (0, ∞) with µ (s) < 0, and µ(s) → 1 as s → 0+ and µ(s) → 0 as s → ∞. Thus,
by the inverse function theorem, s = s(µ) defines a function of µ ∈ (0, 1). Obviously,
Epitaxial Growth without Slope Selection 443

y = µs

y = log(1+s)

O sµ s
Fig. 4. The meaning of sµ in Lemma 4.1.

s(µ) = sµ as defined in (4.9). Moreover, since s  (µ) = 1/µ (s) < 0, sµ increases from 0
to ∞ as µ decreases from 1 to 0. Finally, taking the logarithmic functions of both sides
of (4.9) and dividing them further by log sµ , we get
log log(1 + sµ ) log µ log µ log(1 + sµ )
= +1= · + 1. (4.12)
log sµ log sµ log(1 + sµ ) log sµ
Since sµ → ∞ as µ → 0+ ,
log log(1 + sµ ) log log(1 + sµ ) log(1 + sµ )
= · → 0, as µ → 0+ . (4.13)
log sµ log(1 + sµ ) log sµ
Now, (4.11) follows from (4.12) and (4.12).

We recall that
h is exactly the seminorm |h| H 2 ( 1)

for any h ∈ Hper
2
( 1 ), i.e.,
d

h 2
= ∂xi x j h 2 , ∀ h ∈ Hper
2
( 1 ), (4.14)
i, j=1

cf. (1.3). We also recall the following Poincaré inequalities for the unit cell 1 =
(0, 1)d [7]:

h ≤ C4 ∇h , ∀ h ∈ H( 1 ), (4.15)
∇h ≤ C5
h , ∀ h ∈ H( 1 ), (4.16)

where C4 > 0 and C5 > 0 are constants. The first inequality follows from the fact that
any h ∈ H( 1 ) has zero mean. The second inequality follows from the fact that the
mean of any first-order partial derivative of h ∈ H( 1 ) over 1 vanishes.

Lemma 4.2 (lower bound). Let ε > 0 and µ = ε2 /(2C52 ) > 0. Let s(ε) = 0 if µ ≥ 1
and s(ε) = sµ ∈ (0, ∞), as defined in Lemma 4.1, if 0 < µ < 1. We have

1 ε2
E ε (h) ≥ − log(1 + s(ε)) + (
h)2 d x, ∀ h ∈ H( 1 ).
2 4 1
444 B. Li and J.-G. Liu

Proof. Fix h ∈ H( 1 ).
We have by Lemma 4.1 and the Poincaré inequality (4.16) that

1
E ε (h) = − log(1 + |∇h|2 ) d x
2 {x∈ 1 :|∇h|2 ≤s(ε)}
 
1 ε2
− log(1 + |∇h|2 ) d x + (
h)2 d x
2 {x∈ 1 :|∇h|2 >s(ε)} 2 1
 
1 ε2 ε2
≥ − log(1 + s(ε)) − |∇h| 2
d x + (
h)2 d x
2 4C52 1 2 1

1 ε2
≥ − log(1 + s(ε)) + (
h)2 d x,
2 4 1

as desired.

Lemma 4.3 (upper bound). For each ε > 0, there exists h̃ ε ∈ H( 1) such that
E ε (h̃ ε ) ≤ log ε + C6 , ∀ε > 0,
where C6 > 0 is a constant independent of ε.

Proof. To better illustrate the idea, we prove the result for the case d = 2. The general
case can be treated similarly.
Define as before a 1-periodic, C 1 -function φε : R → R by


 ks, if 0 ≤ s < 18 ,

  
 2
−4k s − 14 + 3k , if 18 ≤ s < 38 ,
16
φε (s) = −k s − 2 , 1
if 38 ≤ s < 58 ,

  2 3k

 4k s − 4 − 16 ,
3
if 58 ≤ s < 78 ,


k(s − 1), if 78 ≤ s ≤ 1,
√ √
where k = 1/ ε. Note that we choose the slope k to be proportional to 1/ ε not 1/ε,
cf. (3.1). This is because the constructed profile will be a product of two such one-
dimensional trial functions. Note also that we choose the size of a “transition region” to
be δ = 1/8. It is easy to verify that
 1
φ(s) ds = 0,
0
   
k 1 3 5 7
|φε (s)| ≥ , ∀s ∈ , ∪ , , (4.17)
8 8 8 8 8
and
3k
|φε (s)| ≤ , |φε (s)| ≤ k, |φε (s)| ≤ 8k, a.e. s ∈ R. (4.18)
16
Define
h̃ ε (x) = φε (x1 )φε (x2 ), ∀ x = (x1 , x2 ) ∈ R2 .
Clearly, h̃ ε ∈ H( 1 ).
Epitaxial Growth without Slope Selection 445

We now calculate and estimate the energy

  
1 ε2
E ε (h̃ ε ) = − log(1 + |∇ h̃ ε | ) + (
h̃ ε ) d x.
2 2

1
2 2
The second
√ term in the energy E ε (h̃ ε ) is easy to bound by (4.18) and the fact that
k = 1/ ε:
 
ε2 ε2   2
(
h̃ ε )2 d x = φε (x1 )φε (x2 ) + φε (x1 )φε (x2 ) d x
2 2
1  1
  
≤ ε 2
|φε (x1 )|2 |φε (x2 )|2 + |φε (x1 )|2 |φε (x2 )|2 d x
1
9
≤ . (4.19)
2
For the first term in the energy E ε (h̃ ε ), we have by the symmetry and (4.17) that

1
− log(1 + |∇ h̃ ε |2 ) d x
2
1 
= −8 log(1 + |∇ h̃ ε |2 ) d x
(0,1/4)×(0,1/4)

 
= −8 log 1 + |φε (x1 )|2 |φε (x2 )|2 + |φε (x1 )|2 |φε (x2 )|2 d x
(0,1/4)×(0,1/4)
 
= −8 log 1 + k 4 (x12 + x22 ) d x
(0,1/8)×(0,1/8)
 
−8 log 1 + |φε (x1 )|2 |φε (x2 )|2 + k 2 |φε (x1 )|2 d x
(1/8,1/4)×(0,1/8)
 
−8 log 1 + k 2 |φε (x2 )|2 + |φε (x1 )|2 |φε (x2 )|2 d x
(0,1/8)×(1/8,1/4)
 
−8 log 1 + |φε (x1 )|2 |φε (x2 )|2 + |φε (x1 )|2 |φε (x2 )|2 d x
(1/8,1/4)×(1/8,1/4)
  
≤ −8 log k 4 x12 + x22 d x
(0,1/8)×(0,1/8)
 
−8 log k 2 |φε (x1 )|2 d x
(1/8,1/4)×(0,1/8)
 
−8 log k 2 |φε (x2 )|2 d x
(0,1/8)×(1/8,1/4)
 
−8 log |φε (x1 )|2 |φε (x2 )|2 + |φε (x1 )|2 |φε (x2 )|2 d x
(1/8,1/4)×(1/8,1/4)

1  
≤ − log k − 8 4
log x12 + x22 d x
8 (0,1/8)×(0,1/8)

4
4
1 k 1 k
− log − log
8 64 8 64
 


2 
 
1 2 k 2 k 1 2
−8 log 8k x1 − + 8k x2 − dx
(1/8,1/4)×(1/8,1/4) 4 8 8 4
= log ε + C7 ,
446 B. Li and J.-G. Liu

where

3  
C7 = log 2 − 8 log x12 + x22 d x
2 (0,1/8)×(0,1/8)
 
2
2 
1 1
−8 log x1 − + x2 − dx
(1/8,1/4)×(1/8,1/4) 4 4

is a finite number. This, together with (4.19), leads to the desired estimate with C6 =
C7 + 9/2.

Lemma 4.4 (equidistribution of energy). If h ε ∈ Hper

2
( 1) is a critical point of E ε :
Hper ( 1 ) → R, then
2

 
|∇h ε |2
ε (
h ε ) d x =
2 2
d x.
1 + |∇h ε |2

Proof. If h ε ∈ Hper
2
( 1)
2
is a critical point of E ε : Hper ( 1) → R, then


∇h ε · ∇g
δ E ε (h ε )(g) = − + ε 2

h ε
g d x = 0, ∀ g ∈ Hper
2
( 1 ).
1 + |∇h ε |2

Choosing g = h ε , we obtain the desired identity.

Proof of Theorem 4.1. (1) Fix ε > 0. By Lemma 4.2, we have eε = infh∈H( 1 ) E ε (h) >
−∞. Let {h j }∞ j=1 be an infimizing sequence of E ε : H( 1 ) → R. It follows from
Lemma 4.2, the Poincaré inequalities (4.15) and (4.16), and the identity (4.14) that
{h j }∞j=1 is bounded in Hper ( 1 ). Thus, up to a subsequence, h j  h ε in H ( 1 ) for
2 2

some h ε ∈ Hper ( 1 ). In particular,
h j 
h ε in L ( 1 ) and, up to a further subse-

2 2

quence if necessary, h j → h ε in H 1 ( 1 ) as j → ∞. Thus, h ε ∈ H( 1 ). Moreover, by

(4.8) with µ = 1 and the Cauchy-Schwarz inequality,
  
 
     |∇h j |2 − |∇h ε |2 
  
log(1 + |∇h j | ) − log(1 + |∇h ε | ) d x  = 
2 2
log 1 + d x 
 1 + |∇h | 2
ε
1

1

 
 |∇h j |2 − |∇h ε |2 
≤ log 1 +    dx
1 + |∇h | 2 
ε
  
1

 |∇h j |2 − |∇h ε |2 
≤  
 1 + |∇h |2  d x
1 ε
≤ ( ∇h j + ∇h ε ) ∇h j − ∇h ε
→ 0, as j → ∞.

Further, since for each j ≥ 1, (
h j )2 + (
h ε )2 ≥ 2
h j
h ε in 1 , we have

     
lim inf (
h j ) d x ≥ lim inf 2
2

h j
h ε d x − (
h ε ) d x =
2
(
h ε )2 d x.
j→∞ j→∞
1 1 1 1
Epitaxial Growth without Slope Selection 447

Therefore,
  
1 ε2
eε = lim inf E ε (h j ) ≥ − log(1 + |∇h ε |2 ) + (
h ε )2 d x = E ε (h ε ) ≥ eε .
j→∞
1
2 2
This implies that h ε is a global minimizer of E ε in H( 1 ).
(2) The estimate (4.1) follows from Lemma 4.3 with C1 = C6 . It follows from
Lemma 4.2, the definition of s(ε) (cf. Lemma 4.2), and (4.11) in Lemma 4.1 that
eε E ε (h ε ) − 12 log(1 + s(ε))
lim inf = lim inf ≥ lim inf = 1. (4.20)
ε→0 + log ε ε→0+ log ε ε→0+ log ε
By Lemma 4.3,
eε log ε + C6
lim sup ≤ lim sup = 1. (4.21)
ε→0+ log ε ε→0+ log ε
Now, the desired asymptotics (4.2) follows from (4.20) and (4.21).
(3) Let h ε ∈ H( 1 ) be a global minimizer of E ε : H( 1 ) → R. Clearly, h ε is also a
2
global minimizer of E ε : Hper ( 1 ) → R. Thus, h ε is a critical point of E ε : Hper
2
( 1) →
R. Consequently, we have by Lemma 4.4 that
 
|∇h ε |2
ε2 (
h ε )2 d x = ≤ 1. (4.22)
1 2
1 + |∇h ε |2
Noting that the negative logarithmic function is convex, we have by Jensen’s inequality
that

 
1
eε = E ε (h ε ) ≥ − log 1 + |∇h ε |2 d x .
2 1

This, together with the upper bound (4.1), leads to

 
1
log ε + C1 ≥ − log 1 + |∇h ε | d x .
2
2 1

Hence,


e−2C1 1 −C1
|∇h ε | d x ≥2
, if ε ∈ 0, √ e . (4.23)
1
2ε 2 2
From (4.23), we have by an integration by parts, the Cauchy-Schwarz inequality, and
(4.22) that
 
e−2C1
≤ |∇h ε |2
d x = (−h ε )
h ε d x
2ε 2 1 1

 1/2
 1/2
 1/2
1
≤ |h ε | d x
2
|
h ε | d x
2
≤ |h ε | d x
2
,
1 1
ε 1

leading to

e−4C1
|h ε |2 d x ≥
. (4.24)
1
4ε 2
Now all the estimates in (4.3) follow from (4.22), (4.24), the Poincaré inequalities
(4.15) and (4.16), and the equivalence of norms (4.14).
448 B. Li and J.-G. Liu

5. Bounds on the Saturation Interface Width and Saturation Time

We now consider the free energy (1.2) that is defined with = (0, L)d . By Theorem 4.1
and the change of variables in (1.7), we have that

E L := min E(h) ∼ − log L , as L → ∞,

h∈H( )

√ C
and for L ≥ 2e 1 that

C2 L 4−2m
≤ − |∇ m h L |2 d x ≤ C3 L 4−2m , m = 0, 1, 2, (5.1)

where h L ∈ H( ) is any minimizer of E : H( ) → R, and C2 and C3 are the same

constants as in (4.3).
It is reasonable to think that the profile will be near a global minimizer after the
saturation of the interface width. Recall from Theorem 2.1 that the interface width is
bounded above by O(t 1/2 ). This, together with (5.1) with m = 0, then sets the saturation
time ts = O(L 4 ), and hence the saturation interface width ws (L) = O(ts1/2 ) = O(L 2 ).
These are exactly the predicted scaling laws, cf. (1.5). The following result is a rigorous
justification of some forms of these scaling laws.

√ h(·) : [0, ∞)
Theorem 5.1. Let √ → H( ) be a weak solution of (1.1) on (0, T ) for any
T > 0. Let L > 2 and ξ ∈ ( 2/L , 1). Let tξ > 0 be such that

E(h(tξ )) = − log(ξ L). (5.2)

(1) If t ≥ 2tξ + (1/2)[wh (0)]2 , then

 t 1/2
ξ2
− [wh (τ )]2 dτ ≥ √ L 2. (5.3)
tξ 8

(2) If t > max(tξ , 1/3, [wh (0)]2 ), then in fact

1 − 4 E(h(0)) 4(σ −1)

t≥ e σ (ξ L) σ , (5.4)
12
where σ = t/tξ .

We remark that the time tξ approximates the saturation time. In general, tξ is larger
when the parameter ξ is closer to 1. The inequality (5.3) gives a lower bound, in an
averaged version, for the saturation interface width. The inequality (5.4) is an asymptotic,
lower bound for the saturation time: As ξ is closer to 1 and σ becomes very large, the
bound becomes O(L 4 ). The assumption for this lower bound is that the time t is larger
than tξ , which is the time at which the energy of a profile h reaches nearly the minimum
value.
Epitaxial Growth without Slope Selection 449

Proof of Theorem 5.1. (1) It is easy to verify from the energy (1.2) and (1.1) that

d
E(h(t)) = −− h 2t d x ≤ 0, ∀ t > 0. (5.5)
dt
Thus, the energy decays. Consequently, we have by (5.2) that
E(h(t)) ≤ E(h(tξ )) = − log(ξ L), ∀ t ≥ tξ . (5.6)
This and Jensen’s inequality imply that
 t
− log(ξ L) = − E(h(tξ )) dτ
tξ
 t
≥ − E(h(τ )) dτ
tξ
 t  
1  
≥ − − − log 1 + |∇h(x, τ )|2 d xdτ
tξ 2
  t 
1
≥ − log 1 + − − |∇h(x, τ )| d xdτ .
2
2 tξ

Therefore, applying an integration by parts and the Cauchy-Schwarz inequality, we

obtain by (2.3) that
 t
(ξ L)2 ≤ 1 + − − |∇h(x, τ )|2 d xdτ
tξ
 t
= 1 + − − [−h(x, τ )]
h(x, τ ) d xdτ
tξ
 t 1/2  t 1/2

≤ 1 + − − |h(x, τ )|2 d xdτ − − |
h(x, τ )|2 d xdτ

tξ tξ
 t 1/2
1/2
[wh (tξ )]2
≤ 1 + − [wh (τ )]2 dτ 1+ . (5.7)
tξ 2(t − tξ )

Now, if t ≥ 2tξ + (1/2)[wh (0)]2 , then by (2.2),

[wh (tξ )]2 2tξ + [wh (0)]2
1+ ≤1+ = 2. (5.8)
2(t − tξ ) 2(tξ + 12 [wh (0)]2 )
√
Combining (5.7), (5.8), and the assumption that ξ L ≥ 2, we obtain (5.3).
(2) Setting t0 = 0 in (2.5), by (5.6) and (5.2), we have for any t > [wh (0)]2 that
√   t
1
− log 1 + 3t ≤ − E(h(τ )) dτ
2 0
 
1 tξ 1 t
= E(h(τ )) dτ + E(h(τ )) dτ
t 0 t tξ
tξ t − tξ
≤ E(h(0)) − log(ξ L).
t t
450 B. Li and J.-G. Liu

Consequently, if t = σ tξ with σ > 1, then

1 √  1 σ −1
− log 1 + 3t ≤ E(h(0)) − log(ξ L).
2 σ σ
Thus, for t ≥ 1/3,
√ √ 2(σ −1)
2 3t ≥ 1 + 3t ≥ e− σ E(h(0)) (ξ L) σ .
2

This leads to (5.4).

Acknowledgments

This work was partially supported by the NSF through grant DMS-0072958 and DMS-
0107218. The authors thank Dr. Robert L. Pego for many interesting discussions. They
also thank the referee for the insightful and suggestive comments.

References

[1] R. Adams. Sobolev Spaces. Academic Press, New York, 1975.

[2] J. G. Amar and F. Family. Critical temperature for mound formation in molecular-beam-
epitaxy. Phys. Rev. B, 54(19):14071–14076, 1996.
[3] J. G. Amar and F. Family. Effects of crystalline microstructure on epitaxial growth. Phys.
Rev. B, 54(20):14742–14753, 1996.
[4] G. S. Bales and A. Zangwill. Morphological instability of a terrace edge during step-flow
growth. Phys. Rev. B, 41(9):5500–5508, 1990.
[5] G. Ehrlich and F. G. Hudda. Atomic view of surface diffusion: Tungsten on tungsten. J.
Chem. Phys., 44:1036, 1966.
[6] H.-J. Ernst, F. Fabre, R. Folkerts, and J. Lapujoulade. Observation of a growth instability
during low temperature molecular beam epitaxy. Phys. Rev. Lett., 72(1):112–115, 1994.
[7] L. C. Evans. Partial Differential Equations, volume 19 of Graduate Studies in Mathematics.
AMS, Providence, 1998.
[8] L. Golubović. Interfacial coarsening in epitaxial growth models without slope selection.
Phys. Rev. Lett., 78(1):90–93, 1997.
[9] M. F. Gyure, J. J. Zinck, C. Ratsch, and D. D. Vvedensky. Unstable growth on rough surfaces.
Phys. Rev. Lett., 81(22):4931–4934, 1998.
[10] C. Herring. Surface tension as a motivation for sintering. In W. E. Kingston, editor, The
Physics of Powder Metallurgy, pages 143–179. McGraw-Hill, 1951.
[11] A. W. Hunt, C. Orme, D. R. M. Williams, B. G. Orr, and L. M. Sander. Instabilities in MBE
growth. Europhys. Lett., 27:611, 1994.
[12] M. D. Johnson, C. Orme, A. W. Hunt, D. Graff, J. Sudijono, L. M. Sander, and B. G. Orr.
Stable and unstable growth in molecular beam epitaxy. Phys. Rev. Lett., 72(1):116–119,
1994.
[13] M. Kardar, G. Parisi, and Y. C. Zhang. Dynamic scaling of growing interfaces. Phys. Rev.
Lett., 56:889–892, 1986.
[14] R. V. Kohn and F. Otto. Upper bounds on coarsening rates. Commun. Math. Phys., 229:375–
395, 2002.
[15] R. V. Kohn and X. Yan. Upper bounds on the coarsening rate for an epitaxial growth model.
Commun. Pure Appl. Math., 56(11):1549–1564, 2003.
[16] J. Krug. Origins of scale invariance in growth processes. Adv. Phys., 46:139–282, 1997.
[17] J. Krug, M. Plischke, and M. Siegert. Surface diffusion currents and the universality classes
of growth. Phys. Rev. Lett., 70(21):3271–3274, 1993.
Epitaxial Growth without Slope Selection 451

[18] J. Krug and M. Schimschak. Metastability of step flow growth in 1 + 1 dimensions. J. Phys.
I France, 5:1065–1086, 1995.
[19] Z.-W. Lai and S. Das Sarma. Kinetic growth with surface relaxation: Continuum versus
atomistic models. Phys. Rev. Lett., 66:2348–2351, 1991.
[20] B. Li and J.-G. Liu. Thin film epitaxy with or without slope selection. European J. Applied
Math., 14(6):713–743, 2003.
[21] D. Moldovan and L. Golubović. Interfacial coarsening dynamics in epitaxial growth with
slope selection. Phys. Rev. E, 61(6):6190–6214, 2000.
[22] W. W. Mullins. Theory of thermal grooving. J. Appl. Phys., 28:333–339, 1957.
[23] M. Ortiz, E. Repetto, and H. Si. A continuum model of kinetic roughening and coarsening
in thin films. J. Mech. Phys. Solids, 47:697–730, 1999.
[24] P. Politi and A. Torcini. Coarsening in surface growth models without slope selection. J.
Phys. A: Math. Gen., 33:L77–L82, 2000.
[25] P. Politi and J. Villain. Ehrlich-Schwoebel instability in molecular-beam epitaxy: A minimal
model. Phys. Rev. B, 54(7):5114–5129, 1996.
[26] I. Raistrick and M. Hawley. Scanning tunneling and atomic force microscope studies of thin
sputtered films of YBa2 Cu3 O7 . In S. L. Shinde and D. A. Rudman, editors, Interfaces in
High-Tc Superconducting Systems, pages 28–70. Springer-Verlag, New York, 1994.
[27] A. Roshko, F. Stork, D. Rudman, D. Aldrich, and P. M. Hotsenpiller. Comparison of het-
eroepitaxial YBa2 Cu3 O7−δ and TiO thin film growth. J. Crystal Growth, 174:398–408, 1997.
[28] M. Rost and J. Krug. Coasening of surface structures in unstable epitaxial growth. Phys.
Rev. E, 55(4):3952–3957, 1997.
[29] M. Rost, P. Šmilauer, and J. Krug. Unstable epitaxy on vicinal surfaces. Surface Sci.,
369:393–402, 1996.
[30] R. L. Schwoebel. Step motion on crystal surfaces II. J. Appl. Phys., 40:614, 1969.
[31] R. L. Schwoebel and E. J. Shipsey. Step motion on crystal surfaces. J. Appl. Phys., 37:3682,
1966.
[32] M. Siegert. Coarsening dynamics in crystalline thin films. Phys. Rev. Lett., 81(25):5481–
5484, 1998.
[33] M. Siegert and M. Plischke. Slope selection and coarsening in molecular beam epitaxy. Phys.
Rev. Lett., 73(11):1517–1520, 1994.
[34] J. A. Stroscio, D. T. Pierce, M. D. Stiles, and A. Zangwill. Coarsening of unstable surface
features during Fe(001) homoepitaxy. Phys. Rev. Lett., 75(23):4246–4249, 1995.
[35] K. Thürmer, R. Koch, M. Weber, and K. H. Rieder. Dynamic evolution of pyramid structures
during growth of epitaxial Fe(001) films. Phys. Rev. Lett., 75(9):1767–1770, 1995.
[36] J. Villain. Continuum models of crystal growth from atomic beams with and without des-
orption. J. de Phys. I, 1:19–42, 1991.
[37] D. D. Vvedensky, A. Zangwill, C. N. Luse, and M. R. Wilby. Stochastic equations of motion
for epitaxial growth. Phys. Rev. E, 48:852–862, 1993.