Mathematical Modelling and Numerical Analysis Will be set by the publisher

Modelisation Mathematique et Analyse Numerique

AN EASILY COMPUTABLE ERROR ESTIMATOR IN SPACE AND

TIME FOR THE WAVE EQUATION
arXiv:1710.08410v1 [math.NA] 22 Oct 2017

O. Gorynina 1 , A. Lozinski 1 and M. Picasso 2

Abstract. We propose a cheaper version of a posteriori error estimator from [7] for
the linear second-order wave equation discretized by the Newmark scheme in time and
by the finite element method in space. The new estimator preserves all the properties of
the previous one (reliability, optimality on smooth solutions and quasi-uniform meshes)
but no longer requires an extra computation of the Laplacian of the discrete solution on
each time step.

1991 Mathematics Subject Classification. 65M15, 65M50, 65M60.

...

Introduction
In this paper, we are interested in a posteriori time-space error estimates for finite element
discretisations of the wave equation. Such estimates were designed, for instance, in [2] for the
case of implicit Euler discretization in time, in [5], [6] for the case of second order discretizations
in time by Cosine (or, equivalently, Newmark) scheme, and in [7] for a particular variant of the
Newmark scheme ( = 1/4, = 1/2). In both [6] and [7], the error is measured in a physically
natural norm: H 1 in space, L in time. Another common feature of these two papers is that
the time error estimators proposed there contain the Laplacian of the discrete solution which
should be computed via auxiliary finite element problems at each time step. This requires thus
a non-negligible extra work in comparison with computing the discrete solution itself. In the
present paper, we propose an alternative time error estimator for the particular Newmark scheme
considered in [7] that avoids these additional computations.
In deriving our a posteriori estimates, we follow first the approach of [7]. First of all, we
recognize that the Newmark method can be reinterpreted as the Crank-Nicolson discretization
of the reformulation of the governing equation as the first-order system, as in [1]. We then use
the techniques stemming from a posteriori error analysis for the Crank-Nicolson discretization of
the heat equation in [8], based on a piecewise quadratic polynomial in time reconstruction of the
numerical solution. Finally, in a departure from [7], we replace the second derivatives in space
(Laplacian of the discrete solution) in the error estimate with the forth derivatives in time by
reusing the governing equation. This leads to the new a posteriori error estimate in time and also
allows us to easily recover the error estimates in space that turn out to be the same as those of [7].
The resulting estimate is referred to as the 5-point estimator since it contains the fourth order
finite differences in time and thus involves the discrete solution at 5 points in time at each time

Keywords and phrases: a posteriori error bounds in time and space and wave equation and Newmark scheme
1 Laboratoire de Mathematiques de Besancon, CNRS UMR 6623, Univ. Bourgogne Franche-Comte, 16 route de
Gray, 25030 Besancon Cedex, France, e-mail: olga.gorynina@univ-fcomte.fr & alexei.lozinski@univ-fcomte.fr
2 Institute of Mathematics, Ecole Polytechnique Federale de Lausanne, Station 8, CH 1015, Lausanne, Switzer-

land, e-mail: marco.picasso@epfl.ch

c EDP Sciences, SMAI 1999
2 TITLE WILL BE SET BY THE PUBLISHER

step. On the other hand, the estimate [7] involves only 3 points in time at each time step and will
be thus referred to as the 3-point estimator.
Like in the case of the 3-point estimator, we are able to prove that the new 5-point estimator
is reliable on general regular meshes in space and non-uniform meshes in time (with constants
depending on the regularity of meshes in both space and time). Moreover, the 5-point estimator
is proved to be of optimal order at least on sufficiently smooth solutions, quasi-uniform meshes in
space and uniform meshes in time, again reproducing the results known for the 3-point estimator.
Numerical experiments demonstrate that 3-point and 5-point error estimators produce very similar
results in the majority of test cases. Both turn out to be of optimal order in space and time, even
in situations not accessible to the current theory (non quasi-uniform meshes, not constant time
steps). It should be therefore possible to use the new estimator for mesh adaptation in space and
time. In fact, the best strategy in practice may be to combine both estimators to take benefit from
the strengths of each of them: the relative cheapness of the 5-point one, and the better numerical
behavior of the 3-point estimator under abrupt changes of the mesh.
The outline of the paper is as follows. We present the governing equations and the discretization
in Section 1. Since our work based on techniques from [7], Section 2 is devoted to reminding the
a posteriori bounds in time and space from there. In Section 3, the 5-point a posteriori error
estimator for the fully discrete wave problem is derived. Numerical experiments on several test
cases are presented in Section 4.

1. The Newmark scheme for the wave equation

We consider initial-boundary-value problem for the wave equation. Let be a bounded domain
in R2 with boundary and T > 0 be a given final time. Let u = u(x, t) : [0, T ] R be the
solution to
2
u
u = f, in ]0, T ] ,


t2




u = 0, on ]0, T ] ,
(1.1)


u(, 0) = u0 , in ,
u(, 0) = v0 ,


in ,

t
where f, u0 , v0 are given functions. Note that if we introduce the auxiliary unknown v = u
t then
model (1.1) can be rewritten as the following first-order in time system

u
v = 0, in ]0, T ] ,




t
v

u = f, in ]0, T ] , (1.2)
t
u = v = 0, on ]0, T ] ,





u(, 0) = u0 , v(, 0) = v0 , in .

The above problem (1.1) has the following weak formulation [4]: for given
f L2 (0, T ; L2 ()), u0 H01 () and v0 L2 () find a function

u
2u
u L2 0, T ; H01 () , L2 0, T ; L2 () , L2 0, T ; H 1 ()


2
(1.3)
t t

u
such that u(x, 0) = u0 in H01 (), (x, 0) = v0 in L2 () and
t
* +
2u
, + (u, ) = (f, ) , H01 (), (1.4)
t2
 TITLE WILL BE SET BY THE PUBLISHER 3

where h, i denotes the duality pairing between H 1 () and H01 () and the parentheses (, ) stand
for the inner product in L2 (). Following Chap. 7, Sect. 2, Theorem 5 of [4], we observe that in
fact

u
2u
u C 0 0, T ; H01 () , C 0 0, T ; L2 () , C 0 0, T ; H 1 () .


t t2
Higher regularity results with more regular data are also available in [4].
Let us now discretize (1.1) or, equivalently, (1.2) in space using the finite element method and
in time using an appropriate marching scheme. We thus introduce a regular mesh Th on with
triangles K, diam K = hK , h = maxKTh hK , internal edges E Eh , where Eh represents the
internal edges of the mesh Th and the standard finite element space Vh H01 ():

Vh = vh C() : vh |K P1 K Th and vh | = 0 .

Let us also introduce a subdivision of the time interval [0, T ]

0 = t0 < t1 < < tN = T

with non-uniform time steps n = tn+1 tn for n = 0, . . . , N 1 and = max n .

0nN 1
The Newmark scheme [9, 10] with coefficients = 1/4, = 1/2 as applied to the wave equation
(1.1): given approximations u0h , vh0 Vh of u0 , v0 compute u1h Vh from

u1h u0h 0 (u1h + u0h )

     0 
, h + , h = vh0 + (f 1 + f 0 ), h ,
0 4 4
h Vh (1.5)

and then compute un+1

h Vh for n = 1, . . . , N 1 from equation

un+1 unh un un1 n (un+1 + unh ) + n1 (unh + un1

   
)
h
h h
, h + h h
, h
n n1 4
n+1
+ f ) + n1 (f + f n1 )
n n
 
n (f
= , h , h Vh . (1.6)
4

where f n is an abbreviation for f (, tn ).

Following [1] and [7], we observe that this scheme is equivalent to the Crank-Nicolson discretiza-
tion of the governing equation written in the form (1.2): taking u0h , vh0 Vh as some approximations
to u0 , v0 compute unh , vhn Vh for n = 0, . . . , N 1 from the system

un+1 unh v n + vhn+1

h
h = 0, (1.7)
n 2
 n+1
vh vhn un+1 + unh
 n+1
+ fn
   
h f
, h + , h = , h , h Vh . (1.8)
n 2 2

Note that the additional unknowns vnh are the approximations are not present in the Newmark
scheme (1.5)(1.6). If needed, they can be recovered on each time step by the following easy
computation
un+1 unh
vhn+1 = 2 h vhn . (1.9)
n
From now on, we shall use the following notations

n+1/2 un+1
h + unh un+1 unh un+1 un1
uh := , n+1/2 uh := h , n uh := h h
(1.10)
2 n n + n1
 n+1
uh unh un un1

1 n + n1
n2 uh := h h
with n1/2 :=
n1/2 n n1 2
4 TITLE WILL BE SET BY THE PUBLISHER

We apply this notations to all quantities indexed by a superscript, so that, for example, f n+1/2 =
(f n+1 + f n )/2.

We also denote u(x, tn ), v(x, tn ) by un , v n so that, for example, un+1/2 =
n+1 n
u + u /2 = (u(x, tn+1 ) + u(x, tn )) /2.
We shall measure the error in the following norm
1/2
u 2

2
u 7 max (t) + |u(t)|H 1 () . (1.11)

t[0,T ] t 2
L ()

u
Here and in what follows, we use the notations u(t) and (t) as a shorthand for, respectively,
t
u
u(, t) and (, t). The norms and semi-norms in Sobolev spaces H k () are denoted, respectively,
t
by k kH k () and | |H k () . We call (1.11) the energy norm referring to the underlying physics of
the studied phenomenon. Indeed, the first term in (1.11) may be assimilated to the kinetic energy
and the second one to the potential energy.
Let us introduce enu = unh h un and env = vhn Ih v n where h : H01 () Vh is the H01 -
orthogonal projection operator, i.e.

(h v, h ) = (v, h ) , v H01 (), h Vh (1.12)

and Ih : H01 () Vh is a Clement-type interpolation operator which is also a projection, i.e.

Ih Vh = Vh [3, 11].
Let us recall, for future reference, the well known properties of these operators (see [3]): for
every sufficiently smooth function v the following inequalities hold

|h v|H 1 () |v|H 1 () , |v h v|H 1 () Ch|v|H 2 () (1.13)

with a constant C > 0 which depends only on the regularity of the mesh. Moreover, for all K Th
and E Eh we have
1/2
kv Ih vkL2 (K) ChK |v|H 1 (K ) and kv Ih vkL2 (E) ChE |v|H 1 (E ) (1.14)

Here K (resp. E ) represents the set of triangles of Th having a common vertex with triangle K
(resp. edge E) and the constant C > 0 depends only on the regularity of the mesh.

2. The 3-point time error estimator

The aim of this section is to recall a posteriori bounds in time and space from [7] for the error
measured in the norm (1.11). Their derivation is based on the following piecewise quadratic (in
time) 3-point reconstruction of the discrete solution.
Definition 2.1. Let unh be the discrete solution given by the scheme (1.6). Then, the piecewise
quadratic reconstruction uh (t) : [0, T ] Vh is constructed as the continuous in time function that
is equal on [tn , tn+1 ], n 1, to the quadratic polynomial in t that coincides with un+1h (respectively
unh , un1
h ) at time tn+1 (respectively t n , t n1 ). Moreover, uh (t) is defined on [t 0 , t1 ] as the
quadratic polynomial in t that coincides with u2h (respectively u1h , u0h ) at time t2 (respectively t1 ,
t0 ). Similarly, we introduce piecewise quadratic reconstruction vh (t) : [0, T ] Vh based on vhn
defined by (1.9) and f (t) : [0, T ] L2 () based on f (tn , ).
The quadratic reconstructions uh , vh are thus based on three points in time (normally looking
backwards in time, with the exemption of the initial time slab [t0 , t1 ]). This is also the case for
the time error estimator (2.3), recalled in the following Theorem and therefore referred to as the
3-point estimator.
 TITLE WILL BE SET BY THE PUBLISHER 5

Theorem 2.2. The following a posteriori error estimate holds between the solution u of the wave
equation (1.1) and the discrete solution unh given by (1.5)(1.6) for all tn , 0 n N with vhn
given by (1.9):

2 1/2
u
2
vhn

(tn )

+ |unh u(tn )|H 1 ()
t 2
L ()
 2 2 1/2
vh0 v0 L2 () + u0h u0 H 1 ()


n
X Z tn
+ S (tn ) + k1 T (tk1 ) + kf f kL2 () dt (2.1)
k=1 0

where the space indicator is defined by

" 2
X vh
h2K

S (tn ) = C1 max u h f (2.2)
06t6tn t 2
L (K)
KTh
#1/2
X 2
+ hE |[n uh ]|L2 (E)
EEh
n1
" 2
Z tm+1 2
vh
X X uh f
+ C2 h2K
t2 t t 2

m=0 t m KTh L (K)
  2 # 1/2
X uh
+ hE n dt,
t L2 (E)
EEh

here C1 , C2 are constants depending only on the mesh regularity, [] stands for a jump on an edge
E Eh , and uh , vh are given by Definition 2.1.
The error indicator in time for k = 1, . . . , N 1 is
  1/2
1 2 1  2 
k fh zhk 2 2
2
T (tk ) = k + k1 k k vh  1
H ()
+ L ()
(2.3)
12 8

where zhk is such that

zhk , h = (k2 uh , h ),


h Vh (2.4)
and   1/2
5 2 1  2 
1 fh zh1 2 2
2
T (t0 ) = 0 + 1 0 1 vh  1
H ()
+ L ()
(2.5)
12 2
We also recall the optimality result for the 3-point time error estimator.
3u 1 2u
Theorem 2.3. Let u be the solution of wave equation (1.1) and (0) H (), (0) H 2 (),
t3 t2
2f 3f
2
(t) L (0, T ; L2 ()), (t) L2 (0, T ; L2 ()). Suppose that mesh Th is quasi-uniform, the
t t3
mesh in time is uniform (tk = k ), and the initial approximations are chosen as

u0h = h u0 , vh0 = h v0 . (2.6)

Then, the 3-point time error estimator T (tk ) defined by (2.3, 2.5) is of order 2 , i.e.

T (tk ) C 2 .

with a positive constant C depending only on u, f , and the regularity of mesh Th .

6 TITLE WILL BE SET BY THE PUBLISHER

Remark 2.4. Note that the approximation of initial conditions (2.6) is crucial for optimality
of our time and space error estimators. As explained theoretically and demonstrated numerically
in [7], this is due to the boundedness of higher order discrete derivatives that contains in time and
space error estimators (2.3, 2.5, 2.2).

3. The 5-point a posteriori error estimator

As already mentioned in the Introduction, the time error estimator (2.3) contains a finite
element approximation to the Laplacian of ukh , i.e. zhk given by (2.4). This is unfortunate because
zhk should be computed by solving an additional finite element problem that implies additional
computational effort. Having in mind that the term n2 fh zhn in (2.3) is a discretization of
2 f /t2 + u = 4 u/t4 at time tn our goal now is to avoid the second derivatives in space in
the error estimates and replace them with the forth derivatives in time.
We introduce a fourth order finite difference in time n4 for any whn Vh , k = 0, 1, . . . as

n4 wh = n2 2 wh (3.1)

where

whn whn1 wn1 whn2

 
2
n2 wh = h , (3.2)
(tn tn2 ) tn tn1 tn1 tn2
tn+1 + tn1
tn =
2

and 2 wh stands for the collection of k2 wh with k = 1, 2, . . ., so that the notation n2 2 wh means
that one should substitute k2 wh for whk , k = n, n1, n2, in the definition of n2 wh . In particular,
n4 wh is indeed the standard fourth order finite difference in time at tn1 of wh in the case of
constant time steps n = :

whn+1 4whn + 6whn1 4whn2 + whn3

n4 wh =
4

However, n4 is not necessarily consistent with the fourth time derivative in the general case
n 6= const.
We denote
n (whn+1 + whn ) + n1 (whn + whn1 )
whn = . (3.3)
4n1/2
Consistently with the conventions above, wh will stand for the collection of whk , k = 0, 1, . . .
We shall need a connection between second order discrete derivatives n2 and n2 which we
establish in the following technical lemma.
Lemma 3.1. For all integer n = 3, . . . N 1 there exist coefficients k , k = n 2, n 1, n such
that for all {whn }
Xn
n2 wh = k k2 wh (3.4)
k=n2

Moreover
n
X
|k | c, for k = n 2, n 1, n, and k C
k=n2

where c and C are positive

 constants depending only on the mesh regularity in time, i.e. on
maxk0 k+1k + k
k+1 .
 TITLE WILL BE SET BY THE PUBLISHER 7

Proof. We first note that relation (3.4) does not contain any derivatives in space and thus it should
hold at any point x . Consequently, it is sufficient to prove this Lemma assuming that whn ,
n2 wh , etc. are real numbers, i.e. replacing Vh by R. This is the assumption adopted in this proof.
We shall thus drop the sub-indexes h everywhere. Furthermore, it will be convenient to reinterpret
wn in (3.1), (3.2) and (3.3) as the values of a real valued function w(t) at t = tn . We shall also use
the notations like wn , n2 w, and so on, where w is a continuous function on R, always assuming
wn = w(tn ).
Observe that n2 w is a linear combination of 5 numbers {wn3 , . . . , wn+1 }. Thus, it is enough
to check equality (3.4) on any 5 continuous functions (k) (t), k = n 3, . . . , n + 1, such that the
vector of values of (k) at times tl , l = n 3, . . . , n + 1, form a basis of R5 . For fixed n, let us
choose these functions as

t tk1 , if t < tk ,


(k) (t) = k1 k = n 3, . . . , n + 1. (3.5)
t k+1 t

, if t tk ,
k

First we notice that for every linear function u(t) on [tn3 , tn+1 ] we have n2 u = n2 u = 0.
Thus, we get immediately n2 (n3) = n2 (n3) = 0 and n2 (n+1) = n2 (n+1) = 0 so that (3.4) is
fulfilled on functions (n3) , (n+1) with any coefficients k , k = n 2, n 1, n. Now we want to
provide coefficients k , k = n 2, n 1, n for which (3.4) is fulfilled on functions (n2) , (n1)
and (n) . For brevity, we demonstrate the idea only for function (n) (t). Function (n) (t) is linear
on [tn3 , tn ] and thus

2 2
n2 (n) = 0, n1 (n) = 0.

From direct computations it is easy to show that

1 1
n2 (n) 2
, (n) 1, n2 (n) 2
n n

where hides some factors that can be bounded by constants depending only on the mesh regu-
n2 (n)
larity. Thus we are able to establish expression for coefficient n = 2 C. Similar reasoning
n (n)
n2 (n1) n2 (n2)
for function n1
n and n2
n shows that n1 = 2 C and n2 = 2 C.
n1 (n1) n2 (n2)
n
X
The next step is to show boundedness from below of k . We will show it by applying
k=n2
2
t
equality (3.4) to second order polynomial function s(t) = . Using a Taylor expansion of s(t)
2
around tn in the definition of sn gives

1 1
n (t2n + tn n1 + (n2 + n1
2
)) + n1 (t2n tn n + (n2 + n1
2
))
n
s = 4 4
2(n + n1 )
2
t 1 2
= n+ 2


+ n1
2 8 n
8 TITLE WILL BE SET BY THE PUBLISHER

Substituting this into the definition of n2 s we obtain

sn sn1 sn1 sn2

2 2 2 2
n n2 n1 n3

2 2
t tn1 tn1 tn2 1 n + n2 n1 + n3
n2 s = n
=1+ 1

tn tn2 /2 8
4 (n + n1 + n2 + n3 )

n n1 n2 + n3
=1+
n + n1 + n2 + n3

Using (3.4) and the fact that k2 s = 1 for k = n 2, n 1, n we note that

n
n n1 n2 + n3 X
1+ = k
n + n1 + n2 + n3
k=n2

n
X
This implies k C. 
k=n2

Lemma 3.2. Let whn , snh Vh be such that

whn+1 whn sn + sn+1

h h
= 0, n 0. (3.6)
n 2

For all n 3 there exist coefficients k , k = n 2, n 1, n such that

n n
! n
X X X
k k2 wh = k n2 wh n k k2 sh (3.7)
k=n2 k=n2 k=n2

where coefficients k , k = n 2, n 1, n are introduced in Lemma 3.1. Moreover

|k | C, k = n 2, n 1, n
 
k+1 k
where C is a positive constant depending only on the mesh regularity in time, i.e. on maxk0 k + k+1 .

Proof. As in proof of Lemma 3.1, we assume Vh = R, drop the sub-indexes h and interpret wn , sn
as the values of continuous real valued functions w(t), s(t) at t = tn . Using (3.6) and notations
(1.10) implies k2 w = k s. Now, we are able to rewrite (3.7) in terms of sn only

n n
! n
X X X
k k s = k n s n k k2 s (3.8)
k=n2 k=n2 k=n2

As in the proof of Lemma 3.1 we take into account the fact that equation (3.8) should hold for
every 5 numbers {sn3 , . . . , sn+1 } and therefore its enough to check equality (3.8) on 5 linearly
independent piecewise linear functions (k) introduced by (3.5). Using the reasoning as in Lemma
3.1 leads to desired result (3.7). 

We can now prove an a posteriori error estimate involving n4 uh . Since the latter is computed
through 5 points in time {tn3 , . . . , tn+1 }, we shall refer to this approach as the 5-point estimator.
For the same reason, this estimator is only applicable from time t4 . The error at first 3 time steps
should be thus measured differently, for example using the 3-point estimator from Theorem 2.2.
The deriving of 5-point time error estimator is based on main ideas from the proof of Theorem
2.2 using Lemma 3.1.
 TITLE WILL BE SET BY THE PUBLISHER 9

Theorem 3.3. The following a posteriori error estimate holds between the solution u of the wave
equation (1.1) and the discrete solution unh given by (1.5)(1.6) for all tn , 4 n N with vhn
given by (1.9):

2 1/2

n u
2
vh (tn ) + |unh u(tn )|H 1 ()

t 2
L ()
2 1/2
3 u
2
+ u3h u(t3 )

vh (t3 )

t 2 H 1 ()
L ()
N
X 1
+ S (tN ) + k T (tk ) + h.o.t. (3.9)
k=3

where the space indicator is defined by (2.2) and the error indicator in time is
 
1 2 1  2  4 
T (tk ) = k + k1 k k vh  1
H ()
+ k uh 2
L ()
(3.10)
12 8
 
k+1 k
with a constant C > 0 depending only on the mesh regularity in time, i.e. on maxk0 k + k+1 .
3
Finally. h.o.t. (higher order terms) in (3.9) hide terms of order O( ).

Proof. Rewrite scheme (1.6) as

n2 uh + Ah unh = fhn (3.11)
for n = 0, . . . , N 1 where fhn
= Ph f (tn , ) and operator Ah defined in (3.20). Taking a linear
combination of instances of (3.11) at steps n, n 1, n 2 with appropriate coefficients gives

n4 uh + Ah n2 uh = n2 fh (3.12)

Using the definition of operator n2 and re-introducing vhn by (1.7) leads to

n n
! n
X X X
n2 uh = k k2 uh = k n2 uh n k k2 vh
k=n2 k=n2 k=n2

withcoefficients k , k introduced in Lemma 3.1 and Lemma 3.2. Moreover, by Lemma 3.1 =
Pn
1
k=n2 k is positive and bounded so that

n
X
n2 uh = n2 uh + n k k2 vh
k=n2

with k = k that are all uniformly bounded on regular meshes in time. Similarly,

n
X
n2 fh = n2 fh + n k k2 fh
k=n2

with fhn satisfying

fhn+1 fhn fn + fhn+1
= h
n 2
10 TITLE WILL BE SET BY THE PUBLISHER

Thus,
n
X  
n2 fh Ah n2 uh = n2 fh Ah n2 uh + n k k2 fh Ah k2 vh
k=n2
n
X  
= n4 uh + n k k2 fh Ah k2 vh
k=n2

We can now  reproduce

 the proof of Theorem 2.2. In the following, we adopt the vector notation
u(t, x)
U (t, x) = where v = u/t. Note that the first equation in (1.2) implies that
v(t, x)
!
u
, (v, ) = 0, H01 ()
t

by taking its gradient, multiplying it by and integrating over . Thus, system (1.2) can be
rewritten in the vector notations as
!
U
b , + (AU, ) = b(F, ), (H01 ())2 (3.13)
t
   
0 1 0
where A = ,F = and
1 0 f
   
u
b(U, ) = b , := (u, ) + (v, )
v

Similarly, Newmark scheme (1.7)(1.8) can be rewritten as

Uhn+1 Uhn Uhn+1 + Uhn

     
b , h + A , h = b F n+1/2 , h , h Vh2 (3.14)
n 2
 n  
n uh n+1/2 0
where Uh = and F = .
vhn f n+1/2
The a posteriori analysis
 relies on an appropriate residual equation for the quadratic recon-
uh
struction Uh = . We have thus for t [tn , tn+1 ], n = 1, . . . , N 1
vh

1
Uh (t) = Uhn+1 + (t tn+1 )n+1/2 Uh + (t tn+1 )(t tn )n2 Uh (3.15)
2
so that, after some simplifications,
!
Uh  
b , h + (AUh , h ) = b (t tn+1/2 )n2 Uh + F n+1/2 , h
t
 
1
+ (t tn+1/2 )An+1/2 Uh + (t tn+1 )(t tn )An2 Uh , h (3.16)
2

Consider now (3.14) at time steps n and n 1. Subtracting one from another and dividing by
n1/2 yields
b n2 Uh , h + (An Uh , h ) = b (n F, h )

or
  n1 2  
b n2 Uh , h + A n+1/2 Uh n Uh , h = b (n F, h )
2
 TITLE WILL BE SET BY THE PUBLISHER 11

so that (3.16) simplifies to

!
Uh  
b , h + AUh , h
t
 
= pn An2 Uh , h + b t tn+1/2 n F + F n+1/2 , h



 
= pn An2 Uh , h + b F pn n2 F, h


(3.17)

where

n1 1
pn = (t tn+1/2 ) + (t tn+1 )(t tn ),
2 2
1
F (t) = Fhn+1 + (t tn+1 )n+1/2 F + (t tn+1 )(t tn )n2 F.
2

Introduce the error between reconstruction Uh and solution U to problem (3.13) :

E = Uh U (3.18)

or, component-wise
   
Eu uh u
E= =
Ev vh v

Taking the difference between (3.17) and (3.13) we obtain the residual differential equation for the
error valid for t [tn , tn+1 ], n = 1, . . . , N 1
!
U h  
b(t E, ) + (AE, ) = b F, h + AU h , ( h ) (3.19)
t
 
+ pn An2 Uh , h + b F F pn n2 F, h , h Vh2

 
h E u
Now we take = E, h = where h : H01 () Vh is the H01 -orthogonal projection
Ih Ev
operator (1.12) and Ih : H01 () Vh is a Clement-type interpolation operator which is also a
projection [3, 11]. Noting that (AE, E) = 0 and
!
uh
, (Eu h Eu ) = (vh , (Eu h Eu )) = 0
t

Introducing operator Ah : Vh Vh such that

(Ah wh , h ) = (wh , h ), h Vh (3.20)

we get
! ! !
Ev Eu v h   
, Ev + Eu , = f, Ev h Ev + u h , Ev Ih Ev
t t t
 
 
+ pn Ah n2 uh n2 fh , Ih Ev pn n2 vh , Eu + f f, Ih Ev .


12 TITLE WILL BE SET BY THE PUBLISHER

Note that equation similar to (3.19) also holds for t [t0 , t1 ]

!
U h  
b(t E, ) + (AE, ) = b F, h + AU h , ( h ) (3.21)
t
 
+ p1 A12 Uh , h + b F F p1 12 F, h .

That follows from the definition of the piecewise quadratic reconstruction uh (t) for t [t0 , t1 ].
Integrating (3.19) and (3.21) in time from 0 to some t t1 yields

1 
|Eu |2H 1 () + kEv k2L2 () (t )
2
1 
= |Eu |2H 1 () + kEv k2L2 () (0)
2
Z t ! Z t 
v h 
+ f, Ev Ih Ev dt + u h , (Ev Ih Ev ) dt
0 t 0
Z t h 
 i
pn Ah n2 uh n2 fh , Ih Ev pn n2 vh , Eu + f f, Ih Ev dt


+
t1
Zt1 h 
 i
p1 Ah 12 uh 12 fh , Ih Ev p1 12 vh , Eu + f f, Ih Ev dt


+
0
:= I + II + III + IV.
(3.22)

Let q
Z(t) = |Eu |2H 1 () + kEv k2L2 ()
and assume that t is the point in time where Z attains its maximum and t (tn , tn+1 ] for some
n. For the first and second terms in (3.22) we have

" 2
v
X h
I + II C1 h2K uh f

t

2
KTh L (K)
#1/2
X 2
+ hE k[nuh ]kL2 (E) (t )|Eu |H 1 () (t )
EEh
" 2
v
X h
+ C1 h2K uh f

t

2
KTh L (K)
#1/2
X 2
+ hE k[nuh ]kL2 (E) (0)|Eu |H 1 () (0)
EEh
n
" #1/2
X m1 X 2
h2K
2 2
+ C2 m vh m1 vh L2 (K) |Eu |H 1 () (tm )
m=1
2
KTh
2
min(tm+1 ,t )
n Z
"
2 v u f
h h
X X
+ C3 h2K

t2 t t 2

m=0 tm KTh L (K)
" # 2 #1/2
X u h
+ hE n (t)|Eu |H 1 () (t)dt.

t 2


EEh L (E)
 TITLE WILL BE SET BY THE PUBLISHER 13

Indeed, it follows from integration by parts with respect to time, see the proof of Theorem 3.2
in [7]. The third and the fourth term in (3.22) are responsible for the time estimator. We neglect
the error at first 3 time steps, in other words we forget about the fourth term in (3.22) and do
not take into account terms at time steps before t3 . Thus the third term can be now written as

1 Z min(tm+1 ,t )
N
" m
! !
X X  
III = pm 4
m uh + m k k2 fh Ah k2 vh , Ih Ev
n=3 tm k=m2
#
2
(pm m vh , Eu )
+ (f f, Ih Ev ) dt. (3.23)

We now observe for t [tm , tm+1 ], m = 1, . . . , N 1

Eu 2
Ev = pm m vh
t

and integrate by parts in time in the terms in (3.23) involving Ah k2 vh

Z min(tm+1 ,t )  
2
m pm Ah m vh , Ih Ev dt
tm
Z min(tm+1 ,t )  
Eu
= m pm m 2
vh , Ih 2
pm m vh dt
tm t
h imin(tm+1 ,t ) Z min(tm+1 ,t )  
2
= m pm (k vh , Ih Eu ) m p0m m
2
vh , Ih Eu dt
tm tm
Z min(tm+1 ,t )
m p2m m
2 2


vh , m vh dt.
tm
 
The first two terms above will result in a contribution to the error estimator n2 n2 vh H 1 ()
(note that n p0n n2 ) while the remaining term is of a higher order. Thus the order of term h.o.t
in error estimator (3.9) is

N
X 1  
n4 n2 fh + n6 n2 vh H 1 () .
 
h.o.t.

L2 ()
n=1

Thus noting the fact

Z tm+1
1 3 1 2
|pm |dt + m1 m
tm 12 m 8
we obtain (3.10). 

Finally we show the upper bound for 5-point time estimator at least for sufficiently smooth
solutions, quasi-uniform meshes in space and uniform meshes in time.
3u 1 2u
Theorem 3.4. Let u be the solution of wave equation (1.1) and (0) H (), (0) H 2 (),
t3 t2
2f 2 3f
(t) L (0, T ; L ()), (t) L2 (0, T ; L2 ()). Suppose that mesh Th is quasi-uniform and
t2 t3
the mesh in time is uniform (tk = k ). Then, the 5-point time error estimator T (tk ) defined by
(3.10) is of order 2 , i.e.
T (tk ) C 2 .
with a positive constant C depending only on u, f , and the mesh regularity.
14 TITLE WILL BE SET BY THE PUBLISHER

Proof. The result follows from Theorem 2.3 by using (3.12) and Lemma 3.1
n
X
4 2 2 uh k k2 fh Ah k2 uh


n uh 2 = f A = .

L ()
n h h n
L2 ()
k=n2 L2 ()


Remark 3.5. Note, that as in the case for 3-point error estimator, the approximation of ini-
tial conditions and right-hand-side function is crucial for optimality of our time and space error
estimators.

4. Numerical results
4.1. A toy model: a second order ordinary differential equation
Let us consider first the following ordinary differential equation
2
d u(t)
+ Au(t) = f (t), t [0; T ]



dt2

u(0) = u0 , (4.1)


du
(0) = v0

dt
with a constant A > 0. This problem serves as simplification of the wave equation in which we
get rid of the space variable. The Newmark scheme reduces in this case to

un+1 un un un1 n (un+1 + un ) + n1 (un + un1 )

+A =
n n1 4
n (f n+1 + f n ) + n1 (f n + f n1 )
= , 1nN 1 (4.2)
4
u1 u0 0 0
= v0 A(u1 + u0 ) + (f 1 + f 0 ),
0 4 4
0
u = u0
 1/2
2 2
and the error become e = max |v n u0 (tn )| + A |un u(tn )| . 3-point and 5-point a
0nN
posteriori error estimates n : 0 n N simplify to this form:

n1  q
X 5 2 1
e k T (tk ) = 0 + 0 1 A(12 v)2 + (12 f A12 u)2 (4.3)
12 0 2
k=0
n1
X 1 q
2 1
+ k + k1 k A(k2 v)2 + (k2 f Ak2 u)2 ,
12 k 8
k=1
n1 n1
X 1 q
X 1
e T (tk ) = k k2 + k1 k A(k2 v)2 + (k4 u)2 (4.4)
12 8
k=3 k=3

We define the following effectivity indexes in order to measure the quality of our estimators T
and T
T T
eiT = , eiT = .
e e

We present in Table 1 the results for equation (4.1) setting f = 0, the exact solution u = cos( At),
final time T = 1, and using constant time steps = T /N . We observe that 3-point and 5-point
 TITLE WILL BE SET BY THE PUBLISHER 15

A N T T e eiT T
ei
100 100 .21 .203 .085 2.47 2.39
100 1000 .0021 .0021 8.34e-04 2.5 2.49
100 10000 2.08e-05 2.08e-05 8.35e-06 2.5 2.5
1000 100 20.51 19.47 8.35 2.46 2.33
1000 1000 .209 .208 .084 2.5 2.49
1000 10000 .0021 .0021 8.33e-04 2.5 2.5
10000 100 1.68e+03 1.4e+03 200 8.38 6.98
10000 1000 20.8 20.7 8.34 2.5 2.49
10000 10000 .208 .208 .083 2.5 2.5
Table 1. Effective indices for constant time steps and f = 0.

estimators are divided by about 100 when the time step is divided by 10. The true error e also
behaves as O( 2 ) and hence both time error estimators behave as the true error.
In order to check behaviour of time error estimators for variable time step (see Table 2) we take
the previous example with time step n : 0 n N
(
0.1 , mod(n, 2) = 0
n = (4.5)
, mod(n, 2) = 1

where is a given fixed value. As in the case of constant time step we have the equivalence
between thePtrue error and both estimated errors. We have plotted on Fig 1 evolution in time of
n1 Pn1
the values k=0 T (tk ) and k=3 T (tk ) compared to e.
Table 3 contains the results for even more non-uniform time step n : 0 n N
(
0.01 , mod(n, 2) = 0
n = (4.6)
, mod(n, 2) = 1

on otherwise the same test case. Note that in case when A = 100 and N = 19800 5-points error
estimator T blows up, while 3-point estimator behaves as the true error. This effect is consistent
with Theorem 2.2. Indeed, the constants in the bounds of this Lemma may depend on the meshes
regularity in time.
Our conclusion is thus that for toy model classic and alternative a posteriori error estimators
are sharp on both constant and variable time grids.

A N T T e eiT T
ei
100 180 .09 .087 .077 1.17 1.13
100 1816 8.85e-04 8.82 e-04 7.59e-04 1.17 1.16
100 18180 8.83e-06 8.83e-06 7.6e-06 1.16 1.16
1000 180 8.91 8.52 7.6 1.17 1.13
1000 1816 .089 .088 .076 1.17 1.16
1000 18180 8.84e-04 8.83e-04 7.59e-04 1.16 1.16
10000 180 802.84 725.1 200 4.01 3.63
10000 1816 8.84 8.8 7.58 1.17 1.16
10000 18180 .088 .088 .076 1.16 1.16
Table 2. Effective indices for variable time step (4.5) and f = 0.
16 TITLE WILL BE SET BY THE PUBLISHER

Figure 1. Evolution in time of 3-point and 5-point time estimators for variable
time step (4.5), A = 100, N = 180, T = 1

A N T T e eiT T
ei
100 196 .086 .083 .084 1.02 0.98
100 1978 8.39e-04 8.36 e-04 8.26e-04 1.02 1.01
100 19800 8.38e-06 1.82e-05 8.1e-06 1.03 2.24
1000 196 8.47 8.1 8.26 1.02 0.98
1000 1978 .083 .084 .0827 1.02 1.01
1000 19800 8.37e-04 8.37e-04 8.26e-04 1.01 1.01
10000 196 764.2 691.7 200 3.82 3.46
10000 1978 8.39 8.35 8.25 1.02 1.01
10000 19800 .084 .084 .083 1.01 1.01
Table 3. Effective indices for variable time step (4.6) and f = 0.

4.2. The error estimator for the wave equation on Delaunay mesh
We now report numerical results for initial boundary-value problem for wave equation with
non-uniform time steps when using 3-point time error estimator (2.3, 2.5) and 5-point time error
estimator(3.10). We compute two parts of the space estimator (2.2) in practice as follows:

" #1/2
(1)
X 2
X
S (tN ) = max h2K kn vh fhn kL2 (K) + hE k[n unh ]k2L2 (E) ,
1nN 1
KTh EEh
1
N
"
(2)
X X 2
h2K n2 vh n fh L2 (K)

S (tN ) = n
n=1 KTh
#1/2
X 2
+ hE k[n n uh ]kL2 (E) . (4.7)
EEh
 TITLE WILL BE SET BY THE PUBLISHER 17

The quality of our error estimators in space and time is determined by following effectivity
indices:
T + S T + S
ei = , ei =
e e
where the first index contains 3-point time error estimator and space estimator, while the last
measures the grade of 5-point time error estimator and space estimator. The true error is
2 1/2

n u
2
e = max vh (tn ) + |unh u(tn )|H 1 ()

06n6N t 2
L ()

Consider the problem (1.1) with = (0, 1) (0, 1), T = 1 and the exact solution u given by
2
u(x, y, t) = e100r (x,y,t)
, (4.8)

where
r2 (x, y, t) = (x 0.3 0.4t2 )2 + (y 0.3 0.4t2 )2 . (4.9)
Thus, u is a Gaussian function, whose center moves from point (0.3, 0.3) at t = 0 to point (0.7, 0.7)
at t = 1. The transport velocity 0.8t(1, 1)T is peaking at t = 1. We choose non-uniform time
1
step as n = 0 1/2 for n = 1, . . . , N 1. We interpolate initial conditions and right-hand-side
tn
function with the orthogonal projections as it noted in Remarks 2.4, 3.5

u0h = h u0 , vh0 = h v 0 , fhn = Ph f n , 0 n N. (4.10)

Unstructured Delaunay meshes in space are used in all the experiments. Numerical results are
reported in Table 4. Note that this case is chosen so that the non-uniform time step is required,
see Fig.2.

Table 4. Results for case (4.8), non-uniform time step.

h 0 ei
ei T T S F Nts e
.05 .01 4.85 4.83 .096 .088 2.55 .0063 105 .58
.025 .0071 5.39 5.38 .054 .051 1.39 .0045 149 .27
.0125 .005 5.94 5.93 .028 .026 .72 .0032 210 .13
.00625 .0035 5.94 5.94 .014 .013 .36 .0022 297 .065
.003125 .0025 5.94 5.94 .0067 .0065 .18 .0016 421 .032

Referring to Table 4, we observe that when setting initial time step as 2 O(h) the error is
divided by 2 each time h is divided by 2, consistent with e O( 2 + h). The space error estimator
and the two time error estimators behave similarly and thus provide a good representation of the
true error. Both effectivity indices tend to a constant value.
We therefore conclude that our space and time error estimators are sharp in the regime of
non-uniform time steps and Delaunay space meshes. They separate well the two sources of the
error and can be thus used for the mesh adaptation in space and time. In particularly, 3-point
and 5-point time estimators become more and more close to each other when h and tend to 0.

References
[1] Baker, G. A. Error estimates for finite element methods for second order hyperbolic equations. SIAM J.
Numer. Anal. 13, 4 (1976), 564576.
[2] Bernardi, C., and Suli, E. Time and space adaptivity for the second-order wave equation. Math. Models
Methods Appl. Sci. 15, 2 (2005), 199225.
[3] Ern, A., and Guermond, J.-L. Theory and practice of finite elements. Springer, 2004.
18 TITLE WILL BE SET BY THE PUBLISHER

Figure 2. Solution (4.8). From up to down: time t = 0, 0.5, 1.

[4] Evans, L. C. Partial differential equations, vol. 19 of Graduate Studies in Mathematics. American Mathemat-
ical Society, Providence, RI, 1998.
[5] Georgoulis, E. H., Lakkis, O., and Makridakis, C. A posteriori L (L2 )-error bounds for finite element
approximations to the wave equation. IMA J. Numer. Anal. 33, 4 (2013), 12451264.
[6] Georgoulis, E. H., Lakkis, O., Makridakis, C. G., and Virtanen, J. M. A Posteriori Error Estimates for
Leap-Frog and Cosine Methods for Second Order Evolution Problems. SIAM J. Numer. Anal. 54, 1 (2016),
120136.
[7] Gorynina, O., Lozinski, A., and Picasso, M. Time and space adaptivity of the wave equation discretized in
time by a second order scheme. Submitted (2017), available on arXiv:1707.00057.
 TITLE WILL BE SET BY THE PUBLISHER 19

[8] Lozinski, A., Picasso, M., and Prachittham, V. An anisotropic error estimator for the Crank-Nicolson
method: application to a parabolic problem. SIAM J. Sci. Comput. 31, 4 (2009), 27572783.
[9] Newmark, N. M. A method of computation for structural dynamics. Journal of the engineering mechanics
division 85, 3 (1959), 6794.
[10] Raviart, P.-A., and Thomas, J.-M. Introduction a lanalyse numerique des equations aux derivees partielles.
Collection Mathematiques Appliquees pour la Matrise. [Collection of Applied Mathematics for the Masters
Degree]. Masson, Paris, 1983.
[11] Scott, L. R., and Zhang, S. Finite element interpolation of nonsmooth functions satisfying boundary condi-
tions. Math. Comp. 54, 190 (1990), 483493.
-