Journal of Computational and Applied Mathematics 470 (2025) 116691

Contents lists available at ScienceDirect

Journal of Computational and Applied Mathematics

journal homepage: [Link]/locate/cam

A nodally bound-preserving finite element method for

time-dependent convection–diffusion equations
Abdolreza Amiri a ,∗, Gabriel R. Barrenechea a , Tristan Pryer b
a
Department of Mathematics and Statistics, University of Strathclyde, 26 Richmond Street, G1 1XH, Glasgow, Scotland, United Kingdom
b
Department of Mathematical Sciences, University of Bath, Claverton down, Bath, BA2 7AY, UK

ARTICLE INFO ABSTRACT

Keywords: This paper presents a new method to approximate the time-dependent convection–diffusion
Time-dependent convection–diffusion equation equations using conforming finite element methods, ensuring that the discrete solution respects
Stabilised finite-element approximation the physical bounds imposed by the differential equation. The method is built by defining,
Positivity preservation
at each time step, a convex set of admissible finite element functions (that is, the ones that
Variational inequality
satisfy the global bounds at their degrees of freedom) and seeks for a discrete solution in this
admissible set. A family of 𝜃-schemes is used as time integrators, and well-posedness of the
discrete schemes is proven for the whole family, but stability and optimal-order error estimates
are proven for the implicit Euler scheme. Nevertheless, our numerical experiments show that
the method also provides stable and optimally-convergent solutions when the Crank–Nicolson
method is used.

1. Introduction

In many applications the simulation of time-dependent convection–diffusion equations is required. The solution of this equation
provides a theory for the transport-controlled reaction rate of two molecules in a shear flow, also it models the processes of the
chemical reaction in a flow field. Such simulations are obtained by a nonlinear system of time dependent convection–diffusion
equations for the concentrations of the reactants and the products. An inaccuracy in one of the equations in the system of nonlinear
equations affects all the concentrations. Usually in applications the convection field is dominant over the diffusion by several orders
of magnitude, thus making the numerical approximation of such problems very unstable and prone to spurious oscillations. For
example, it is well-known that the standard Galerkin finite element method for space discretisation is not appropriate for convection-
dominated problems. The preferred way to deal with this fact over the last few decades has been to replace the standard Galerkin
finite element method by a stabilised finite element method, that is, a method that adds an extra term to the formulation in such a
way that stability is enhanced. The first stabilised methods for convection–diffusion problems, e.g., the SUPG method [1] and the
Galerkin Least-Squares method [2] were of residual type, i.e., a weighted term based on the residual was added to the formulation
to add numerical diffusion. In the time dependent context the residual character of the SUPG method requires the time derivative
of the solution to be included in the stabilisation term, thus creating a somewhat artificial coupling between the time step and
the stabilisation parameter. This coupling has been extensively studied in the literature, see, e.g.,[3] where stability is analysed
for different choices of time discretisation. In fact, if a standard stability analysis is performed, the addition of SUPG stabilisation
suggests a loss of stability if the mesh size and time step are not correctly balanced (although some numerical evidence contradicting
this claim can be found, e.g., in [4]). Due to this, in later years the possibility of using non-residual, symmetric stabilisations for the

∗ Corresponding author.
E-mail address: [Link]@[Link] (A. Amiri).

[Link]
Received 12 September 2024; Received in revised form 3 April 2025
Available online 17 April 2025
0377-0427/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license
([Link]
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

time-dependent convection–diffusion has been studied more intensively. Examples of symmetric stabilisations include the subgrid
viscosity method [5], the orthogonal subscale method [6], and the continuous interior penalty (CIP) method [7]. In the work [8] a
detailed analysis on the stability analysis and advantages of using symmetric stabilisation can be found.
Now, none of the methods mentioned in the previous paragraph can be proven to satisfy the so-called Discrete Maximum
Principle (DMP), or even the weaker property of respecting the bounds satisfied by the continuous problem. In fact, this property
has been derived for only a selection of methods, usually under geometric conditions on the computational mesh. For example, in
the pioneering work [9] it is shown that for the finite element discretisation of a diffusion equation in two space dimensions using
piecewise linear elements a sufficient condition for the satisfaction of the DMP is that the mesh satisfies the Delaunay condition.
This condition was then proven necessary for the matrix to have the right sign pattern in the later work [10]. The situation is more
dramatic for convection–diffusion equations, where the mesh needs to be acute, and sufficiently fine, for the matrix to have the right
sign pattern, and thus the finite element method to satisfy the DMP. For convection–diffusion equations, the more straightforward
remedy is to add isotropic artificial diffusion to the problem, in such a way that the dominating matrix in the linear system is the
diffusion one. This strategy does lead to a method that satisfies the DMP under less stringent conditions on the fineness of the mesh.
Unfortunately, the consistency error introduced by adding artificial diffusion is too large, and the results obtained by this method
tend to smear the layers excessively (see, e.g., [11, Section 5.2] for a discussion and further references). In the quest of adding
diffusion so the DMP is satisfied and layers are not excessively smeared, the idea of localising the diffusion has been proposed as a
way to make the problem locally diffusion-dominated in areas near extrema and layers, thus preventing spurious oscillations and
local violations of the discrete maximum principle. This leads to nonlinear (shock-capturing related) discretisations. In the last few
decades numerous nonlinear discretisations for the steady-state convection–diffusion equation have been proposed, e.g., [10,12–15],
and [11] for a review.
In many instances, the numerical solution does not need to be completely free of local spurious oscillations; it only needs to
satisfy global bounds to lead to a stable discretisation. One way to achieve a solution that respects this bound is to simply cut the
solution at the bound, like it is done in, e.g., the cut-off finite element method [16]. However, this process results in a solution that
is not an element of the finite element space, and thus analysing its stability and convergence is, in general, a challenging affair.
In the recent works [17,18] an alternative path is followed. It uses the following property: There is a correspondence between the
imposition of bounds on the numerical solution and searching for the numerical solution on a convex subset of the finite element
space of the steady state convection–diffusion equation consisting of the discrete functions satisfying those bounds at their nodal
values. So, in those works a finite element method that seeks directly for a finite element solution that satisfies the global bounds
is proposed, and analysed in different contexts. In this work we build on those previous works and extend that method to the time
dependent convection–diffusion equation. The basic idea can be summarised as follows: for each time step we define a set 𝑉+ , of
admissible finite element functions as those satisfying the global bound at at their degrees of freedom (nodal values in the case of
Lagrangian elements); then, introduce an algebraic projection onto the admissible set, denote by 𝑢+ ℎ
the projection of 𝑢ℎ onto 𝑉+
and write a finite element problem for the projected object. To eliminate the non-trivial kernel (and so to remove the singularity)
generated by this process, a stabilisation term is incorporated into the discretised equations at each time step to remove this kernel
(that is, the part of the solution which has been removed). In our numerical experience we have observed that more stable results
are obtained if we add a linear stabilised term to the problem for 𝑢+ ℎ
, and thus we present the method including CIP stabilisation.
The remainder of the paper is organised as follows: In Section 2 we introduce the notation, the model problem, and all the
preliminary material for the setup of the method. In Section 3 we present the finite element method and show its well-posedness.
The stability and error analysis is carried out in Section 4, and in Section 5 we test the performance of the method via different
numerical experiments. Finally, some conclusions are drawn in Section 6.

2. General setting and the model problem

We will adopt standard notations for Sobolev spaces in line with, e.g., [19]. For 𝐷 ⊆ R𝑑 , 𝑑 = 1, 2, 3, we denote by ‖ ⋅ ‖0,𝑝,𝐷 the
𝐿𝑝 (𝐷)-norm; when 𝑝 = 2 the subscript 𝑝 will be omitted and we only write ‖ ⋅ ‖0,𝐷 . The inner product in 𝐿𝑝 (𝐷) is also denoted by
(⋅, ⋅)𝐷 . In addition, for 𝑠 ≥ 0, 𝑝 ∈ [1, ∞], we denote by ‖ ⋅ ‖𝑠,𝑝,𝐷 (| ⋅ |𝑠,𝑝,𝐷 ) the norm (seminorm) in 𝑊 𝑠,𝑝 (𝐷); when 𝑝 = 2, we define
𝐻 𝑠 (𝐷) = 𝑊 𝑠,2 (𝐷), and again omit the subscript 𝑝 and only write ‖ ⋅ ‖𝑠,𝐷 (| ⋅ |𝑠,𝐷 ). The following space will also be used repeatedly
within the text
{ }
𝐻01 (𝐷) = 𝑣 ∈ 𝐻 1 (𝐷) ∶ 𝑣 = 0 on 𝜕𝐷 . (1)

For 1 ≤ 𝑝 ≤ +∞, 𝐿𝑝 ((0, 𝑇 ); 𝑊 𝑠,𝑝 (𝐷)) is the space defined by

{ }
𝐿𝑝 ((0, 𝑇 ); 𝑊 𝑠,𝑝 (𝐷)) = 𝑢(𝑡, ⋅) ∈ 𝑊 𝑠,𝑝 (𝐷) for almost all 𝑡 ∈ [0, 𝑇 ] ∶ ‖𝑢‖𝑠,𝑝,𝐷 ∈ 𝐿𝑝 (0, 𝑇 ) ,

and it is a Banach space for the norm

( )1
⎧ 𝑇 𝑝 𝑝
⎪ ∫0 ‖𝑢‖𝑠,𝑝,𝐷 d𝑡 if 1 ≤ 𝑝 < ∞,
‖𝑢‖𝐿𝑝 ((0,𝑇 );𝑊 𝑠,𝑝 (𝐷)) =⎨
⎪
⎩ ess sup𝑡∈(0,𝑇 ) ‖𝑢‖𝑠,𝑝,𝐷 if 𝑝 = ∞.

2
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

2.1. The model problem

Let 𝛺 be an open bounded Lipschitz domain in R𝑑 (𝑑 = 2, 3) with polyhedral boundary 𝜕𝛺, and 𝑇 > 0. For a given
𝑓 ∈ 𝐿2 ((0, 𝑇 ); 𝐿2 (𝛺)), we consider the following convection–diffusion problem:
⎧𝜕𝑡 𝑢 − 𝜀 𝛥𝑢 + 𝜷 ⋅ ∇𝑢 + 𝜇𝑢 = 𝑓 in (0, 𝑇 ] × 𝛺,
⎪
⎨ 𝑢(𝒙, 𝑡) = 0 on (0, 𝑇 ] × 𝜕𝛺, (2)
⎪
⎩ 𝑢(⋅, 0) = 𝑢0 in 𝛺,

where 𝜀 ∈ R+ , 𝜷 = (𝛽𝑖 )𝑑𝑖=1 ∈ 𝐿∞ ((0, 𝑇 ); 𝑊 1,∞ (𝛺))𝑑 , and 𝜇 ∈ R+

0
, respectively, are the diffusion coefficient, the convective field, and
the reaction coefficient. We will assume that div𝜷 = 0 in 𝛺 × [0, 𝑇 ].
The standard weak formulation of (2) reads as follows: Find 𝑢 ∈ 𝐿∞ ((0, 𝑇 ), 𝐻01 (𝛺)) ∩ 𝐻 1 ((0, 𝑇 ), 𝐻 −1 (𝛺)) such that, for almost all
𝑡 ∈ (0, 𝑇 ) the following holds
{
(𝜕𝑡 𝑢, 𝑣)𝛺 + 𝑎(𝑢, 𝑣) = (𝑓 , 𝑣)𝛺 ∀𝑣 ∈ 𝐻01 (𝛺),
(3)
𝑢(⋅, 0) = 𝑢0 .
Here, the bilinear form 𝑎(⋅, ⋅) is defined by

𝑎(𝑤, 𝑣) ∶= 𝜀 (∇𝑤, ∇𝑣)𝛺 + (𝜷 ⋅ ∇𝑤, 𝑣)𝛺 + 𝜇(𝑤, 𝑣)𝛺 ∀𝑣 ∈ 𝐻01 (𝛺), 𝑡 ∈ (0, 𝑇 ). (4)

In the above definition we have slightly abused the notation, as the convective term 𝜷 might depend on 𝑡, but unless the context
requires it, we will always denote this bilinear form by 𝑎(⋅, ⋅). Since we have supposed that 𝜷 is solenoidal, for each 𝑡 ∈ (0, 𝑇 ) the
bilinear form 𝑎(⋅, ⋅) induces the following ‘‘energy’’ norm in 𝐻01 (𝛺)
√
‖𝑣‖𝑎 = 𝑎(𝑣, 𝑣) 𝑡 ∈ [0, 𝑇 ].

The well-posedness of (3) is a well-studied problem. In fact, this is a consequence of Lions’ Theorem (see, e.g., [20, Theo-
rem 10.9]). Moreover, as a consequence of the maximum principle for parabolic partial differential equations (see, e.g., [21, Theorem
[ ] [ ]
12, Section 7.1]), the solution of (3) reaches its extrema on 𝛺 × {0} ∪ 𝜕𝛺 × (0, 𝑇 ] if 𝑓 = 0, and its extrema depend also on the
values of 𝑓 otherwise. Motivated by this, we make the following assumption on 𝑢, solution of (3).

Assumption (A1): We will suppose that the weak solution of (3) satisfies

0 ≤ 𝑢(𝒙, 𝑡) ≤ 𝜅(𝑡) for almost all (𝒙, 𝑡) ∈ 𝛺 × [0, 𝑇 ], (5)

where 𝜅(𝑡) is a known positive function dependent on 𝑡.

In the above assumption, the lower bound in (5) is not required to be zero; however, we set it to zero for the sake of clarity in
the explanation. Additionally, the results presented in this work will remain valid if 𝜅(𝑡) is replaced by a positive function 𝜅(𝒙, 𝑡).

2.2. Space discretisation

As it is usual in the discretisation of parabolic partial differential equations, we first discretise (3) only in space. To build a finite
element space of 𝐻01 (𝛺), let  be a conforming, shape-regular partition of 𝛺 into simplices (or affine quadrilaterals/hexahedra).
Over , and for 𝑘 ≥ 1, we define the finite element space

𝑉 ∶= {𝑣ℎ ∈ 𝐶 0 (𝛺) ∶ 𝑣ℎ |𝐾 ∈ R(𝐾) ∀𝐾 ∈ } ∩ 𝐻01 (𝛺), (6)

where
{
P𝑘 (𝐾), if 𝐾 is a simplex,
R(𝐾) = (7)
Q𝑘 (𝐾), if 𝐾 is an affine quadrilateral/hexahedron,
where P𝑘 (𝐾) denotes the polynomials of total degree 𝑘 on 𝐾 and Q𝑘 (𝐾) denotes the mapped space of polynomial of degree of at
most 𝑘 in each variable.
For a mesh , the following notations are used:

• let {𝒙1 , 𝒙2 , … , 𝒙𝑁 } denote the set of internal nodes; the usual Lagrangian basis functions associated to these nodes, spanning
the space 𝑉 , are denoted by 𝜙1 , … , 𝜙𝑁 ;
• let 𝐼 denote the set of internal facets, 𝜕 denote the set of boundary facets of ; ℎ = 𝐼 ∪ 𝜕 denote the set of all facets of
; for an element 𝐾 ∈  the set of its facets is denoted by 𝐾 ;
• for a facet 𝐹 ∈ 𝐼 , [[⋅]] denotes the jump of a function across 𝐹 .

The diameter of a set 𝐺 ⊂ R𝑑 is denoted by ℎ𝐺 , and the mesh size is defined as ℎ = max{ℎ𝐾 ∶ 𝐾 ∈ }. We also define the mesh
function h as a continuous, element-wise linear function that represents a local average of element diameters, commonly used in

3
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

finite element analysis [22]. To construct this, we introduce the set of vertices of the mesh, 𝒗1 , … , 𝒗𝑀 , and define h as the piecewise
linear function specified by the nodal values
∑
𝐾∶𝒗𝑖 ∈𝐾 ℎ𝐾
h(𝒗𝑖 ) = . (8)
#{𝐾 ∶ 𝒗𝑖 ∈ 𝐾}
2
In the construction of the method, and its analysis, the following mass-lumped 𝐿 -inner product will be of importance: for every
𝑣ℎ , 𝑤ℎ ∈ 𝑉 , we define

∑
𝑀
(𝑣ℎ , 𝑤ℎ )ℎ = h(𝒙𝑖 )𝑑 𝑣ℎ (𝒙𝑖 )𝑤ℎ (𝒙𝑖 ), (9)
𝑖=1
1
which induces the norm |𝑣ℎ |ℎ ∶= (𝑣ℎ , 𝑣ℎ )ℎ2 in 𝑉 . This norm is, in fact, equivalent to the standard 𝐿2 (𝛺)-norm. More precisely, the
following result, whose proof can be found in [23, Propositions 28.5, 28.6], will be used repeatedly in our analysis below: There
exist 𝐶, 𝑐 > 0, independent of ℎ, such that
∑ ∑
𝑐 ℎ𝑑𝐾 𝑣2ℎ (𝒙𝑖 ) ≤ ‖𝑣ℎ ‖20,𝐾 ≤ 𝐶 ℎ𝑑𝐾 𝑣2ℎ (𝒙𝑖 ) ∀ 𝐾 ∈ , ∀𝑣ℎ ∈ 𝑉 , (10)
𝑖∶𝒙𝑖 ∈𝐾 𝑖∶𝒙𝑖 ∈𝐾

and thus, as a consequence of the shape-regularity of the mesh, the following holds:

𝑐 |𝑣ℎ |2ℎ ≤ ‖𝑣ℎ ‖20,𝛺 ≤ 𝐶 |𝑣ℎ |2ℎ , (11)

for all 𝑣ℎ ∈ 𝑉 .
Next, we recall that the Lagrange interpolation operator is defined by (see, e.g., [19, Chapter 11])
𝑖ℎ ∶  0 (𝛺) ∩ 𝐻01 (𝛺) ⟶ 𝑉 ,
∑
𝑁
𝑣 ⟼ 𝑖ℎ 𝑣 = 𝑣(𝑥𝑗 )𝜙𝑗 . (12)
𝑗=1
With the above ingredients, we now state some inequalities and properties that will be useful in what follows:

(a) Inverse inequality:([19, Lemma 12.1]) For all 𝑚, 𝓁 ∈ N0 , 0 ≤ 𝑚 ≤ 𝓁 and all 𝑝, 𝑞 ∈ [1, ∞], there exists a constant 𝐶,
independent of ℎ, such that
( )
𝑚−𝓁+𝑑 1𝑝 − 1𝑞
|𝑣ℎ |𝓁,𝑝,𝐾 ≤ 𝐶ℎ𝐾 |𝑣ℎ |𝑚,𝑞,𝐾 ∀ 𝑣ℎ ∈ 𝑉  . (13)

(b) Discrete Trace inequality: ([19, Lemma 2.15]) There exists 𝐶 > 0 independent of ℎ such that, for every 𝑣 ∈ 𝐻 1 (𝐾) the
following holds
( )
2
‖𝑣‖20,𝜕𝐾 ⩽ 𝐶 ℎ−1 2
𝐾 ‖𝑣‖0,𝐾 + ℎ𝐾 |𝑣|1,𝐾 . (14)

(c) Approximation property of the Lagrange interpolant: ([19, Proposition 1.12]) Let 1 ≤ 𝓁 ≤ 𝑘 and 𝑖ℎ be the Lagrange
interpolant. Then, there exists 𝐶 > 0, independent of ℎ, such that for all ℎ and 𝑣 ∈ 𝐻 𝓁+1 (𝛺) ∩ 𝐻01 (𝛺) the following holds:

‖𝑣 − 𝑖ℎ 𝑣‖0,𝐾 + ℎ𝐾 |𝑣 − 𝑖ℎ 𝑣|1,𝐾 ≤ 𝐶ℎ𝓁+1

𝐾
|𝑣|𝓁+1,𝐾 . (15)

The standard Galerkin semi-discretisation of (3) using the finite element space (6) reads:
⎧ For almost all 𝑡 ∈ (0, 𝑇 ), find 𝑢ℎ ∈ 𝑉 such that
⎪
⎨ (𝜕𝑡 𝑢ℎ , 𝑣ℎ )𝛺 + 𝑎(𝑢ℎ , 𝑣ℎ ) = (𝑓 , 𝑣ℎ )𝛺 ∀𝑣ℎ ∈ 𝑉 , (16)
⎪ 𝑢ℎ (⋅, 0) = 𝑖ℎ 𝑢0 .
⎩
It is a well-known that the standard Galerkin finite element method for (3) in the convective-dominated regime leads to discrete
solutions that are polluted by global spurious oscillations, and thus, in particular, violate Assumption (A1) (see, e.g. [24] for a
comprehensive review). The usual way of enhancing the stability is to add a linear stabilising term aimed at dampening the
oscillations caused by the dominating convection. Several alternatives are available, with stabilisations based on the addition of
symmetric semi-positive-definite being among the most popular for time-dependent problems. In this work we have chosen to use
the continuous interior penalty (CIP) method originally proposed in [6] and analysed in detail for the time-dependent problem
in [8]. The CIP method adds the following stabilising term to the Galerkin scheme (16):
∑
𝐽 (𝑢ℎ , 𝑣ℎ ) = 𝛾 ‖𝜷‖0,∞,𝐹 ℎ2𝐹 [[𝛁𝑢ℎ ]] ⋅ [[𝛁𝑣ℎ ]] d𝑠 , (17)
∫𝐹
𝐹 ∈𝐼

where 𝛾 ≥ 0 is a non-dimensional constant, and thus it reads as follows:

⎧ For almost all 𝑡 ∈ (0, 𝑇 ), find 𝑢ℎ ∈ 𝑉 such that
⎪
⎨ (𝜕𝑡 𝑢ℎ , 𝑣ℎ )𝛺 + 𝑎𝐽 (𝑢ℎ , 𝑣ℎ ) = (𝑓 , 𝑣ℎ )𝛺 ∀𝑣ℎ ∈ 𝑉 , (18)
⎪ 𝑢 (⋅, 0) = 𝑖 𝑢 0.
⎩ ℎ ℎ

4
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

where

𝑎𝐽 (𝑢ℎ , 𝑣ℎ ) ∶= 𝑎(𝑢ℎ , 𝑣ℎ ) + 𝐽 (𝑢ℎ , 𝑣ℎ ) ∀𝑣ℎ ∈ 𝑉 . (19)

It is worth mentioning that, even if in this work we have chosen to add CIP stabilisation, the results presented herein remain valid if
we choose any symmetric stabilisation for the convective term, in particular, they are valid if we use any of the stabilised methods
analysed in [8].
Although (19) does help remove spurious oscillations, and provides a stable solution, its discrete solution does not preserve the
physical bounds (5). In the next section we introduce the main ingredients to build a finite element method that enforces the bound
(5) on its solution.

2.3. The admissible set

Assumption (A1) motivates the introduction of the following admissible set, that is, the set of finite element functions that satisfy
the bound (5) at their degrees of freedom:

𝑉+ ∶= {𝑣ℎ ∈ 𝑉 ∶ 𝑣ℎ (𝒙𝑖 ) ∈ [0, 𝜅(𝑡)] for all 𝑖 = 1, … , 𝑁}. (20)

Every element 𝑣ℎ ∈ 𝑉 can be split as the sum 𝑣ℎ = 𝑣+

ℎ
+ 𝑣−
ℎ
, where 𝑣+
ℎ
and 𝑣−
ℎ
are given by

∑
𝑀 { }
𝑣+
ℎ
= max 0, min{𝑣ℎ (𝒙𝑖 ), 𝜅(𝑡)} 𝜙𝑖 for 𝑡 ∈ (0, 𝑇 ], (21)
𝑖=1

and

𝑣− +
ℎ = 𝑣ℎ − 𝑣ℎ . (22)

We refer to 𝑣+
ℎ
and 𝑣−
ℎ
as the constrained and complementary parts of 𝑣ℎ , respectively. Using this decomposition we define the following
algebraic projection

(⋅)+ ∶ 𝑉 → 𝑉+ , 𝑣 ℎ → 𝑣+
ℎ
. (23)

Remark 2.1. Strictly speaking (⋅)+ should be denoted by (⋅)+,𝑡 , as 𝜅(𝑡) depends on 𝑡. To lighten the notation, we will simply use (⋅)+
unless it is necessary to specify the time.

The following result, whose proof is identical to that of [18, Lemma 3.2], will be useful in the analysis presented below.

Lemma 2.2. Let the operator (⋅)+ be defined in (23). There exists a constant 𝐶 > 0, independent of ℎ, such that for all 𝑡 ∈ [0, 𝑇 ] the
following holds

‖𝑤+
ℎ
− 𝑣+ ‖
ℎ 0,𝛺
≤ 𝐶‖𝑤ℎ − 𝑣ℎ ‖0,𝛺 , (24)
‖𝑣+ ‖
ℎ 0,𝛺
≤ 𝐶𝜅(𝑡), (25)

for all 𝑤ℎ , 𝑣ℎ ∈ 𝑉 .

3. The finite element method

In this section, we propose a 𝜃-scheme time-space discretisation of (18). Let 𝑁 > 0 be a given positive integer. In what follows,
we consider a partition of the time interval [0, 𝑇 ] as 𝑡0 = 0 < 𝑡1 < 𝑡2 < ⋯ < 𝑡𝑁 = 𝑇 with the time step size 𝛥𝑡𝑛 ∶= 𝑡𝑛 − 𝑡𝑛−1 . To simplify
𝑇
the notation we assume that the time step size is uniform i.e., 𝛥𝑡𝑛 = 𝛥𝑡 = 𝑁 . In addition, the discrete value 𝑢𝑛ℎ ∈ 𝑉 stands for the
[ ]
𝑛 1
approximation of 𝑢 = 𝑢(𝑡𝑛 ) in 𝑉 for 0 ≤ 𝑛 ≤ 𝑁. For 𝜃 ∈ 2 , 1 , we denote

𝑢𝑛ℎ − 𝑢𝑛−1
ℎ
𝛿𝑢𝑛ℎ ∶= ,
𝛥𝑡
𝑡𝑛−1+𝜃 = 𝜃𝑡𝑛 + (1 − 𝜃)𝑡𝑛−1 , 𝑢𝑛−1+𝜃
ℎ
∶= 𝜃𝑢𝑛ℎ + (1 − 𝜃)𝑢𝑛−1
ℎ .

With these notations, the finite element method used in this work reads as follows:
⎧ For 1 ≤ 𝑛 ≤ 𝑁, find 𝑢𝑛 ∈ 𝑉 such that
⎪ ℎ
⎨ (𝛿(𝑢𝑛ℎ )+ , 𝑣ℎ )𝛺 + 𝑎ℎ (𝑢𝑛−1+𝜃
ℎ
; 𝑣ℎ ) = (𝑓 𝑛−1+𝜃 , 𝑣ℎ )𝛺 ∀𝑣ℎ ∈ 𝑉 , (26)
⎪ 𝑢0ℎ = 𝑖ℎ 𝑢0 .
⎩
Here, 𝑎ℎ (⋅; ⋅) is defined as

𝑎ℎ (𝑢𝑛−1+𝜃
ℎ
; 𝑣ℎ ) ∶= 𝜃𝑎𝐽 ((𝑢𝑛ℎ )+ , 𝑣ℎ ) + (1 − 𝜃)𝑎𝐽 ((𝑢𝑛−1 + 𝑛 −
ℎ ) , 𝑣ℎ ) + 𝑠((𝑢ℎ ) , 𝑣ℎ ) , (27)

5
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

where the stabilisation term 𝑠(⋅, ⋅) is defined as

𝑀 (
∑ ) ( )
1
𝑠(𝑣ℎ , 𝑤ℎ ) ∶= 𝛼 𝜀h(𝒙𝑖 )𝑑−2 + ‖𝜷(𝒙, 𝑡𝑛 )‖0,∞,𝜔𝑖 h(𝒙𝑖 )𝑑−1 +
+ 𝜇 h(𝒙𝑖 )𝑑 𝑣ℎ (𝒙𝑖 )𝑤ℎ (𝒙𝑖 ). (28)
𝑖=1
𝛥𝑡
Setting 𝜀̃ = 𝛥𝑡𝜃𝜀, 𝜷̃ = 𝛥𝑡𝜃𝜷, 𝜇̃ = (𝜇𝛥𝑡𝜃 + 1) and 𝐽̃(⋅, ⋅) = 𝛥𝑡𝜃𝐽 (⋅, ⋅), we can define the following norm at each time step 𝑡𝑛 :
( )1
2 2
‖𝑣ℎ ‖ℎ,𝜃𝛥𝑡 ∶= 𝜀‖∇𝑣
̃ ̃ ℎ ‖20,𝛺 + 𝐽̃(𝑣ℎ , 𝑣ℎ ) .
ℎ ‖0,𝛺 + 𝜇‖𝑣 (29)

In addition, the stabilising form 𝑠(⋅, ⋅) for 0 ≤ 𝑛 ≤ 𝑁 induces the following norm on 𝑉
√
‖𝑣ℎ ‖𝑠 ∶= 𝑠(𝑣ℎ , 𝑣ℎ ) . (30)

The following result is a direct consequence of (10) (see [18, Lemma 3.1] for the proof of a very similar result).

Lemma 3.1. There exists a constant 𝐶equiv > 0, depending only on the shape regularity of , such that

𝐶equiv
‖𝑣ℎ ‖2ℎ,𝜃𝛥𝑡 ≤ 𝛥𝑡 ‖𝑣ℎ ‖2𝑠 ∀𝑣ℎ ∈ 𝑉 . (31)
𝛼

Remark 3.2. At each time-level 0 ≤ 𝑛 ≤ 𝑁, the finite element method (26) is a particular case of the finite element method proposed
in [18]. In fact, at each time step 1 ≤ 𝑛 ≤ 𝑁, (26) can be written as

((𝑢𝑛ℎ )+ , 𝑣ℎ )𝛺 + 𝛥𝑡𝜃𝑎𝐽 ((𝑢𝑛ℎ )+ , 𝑣ℎ ) + 𝛥𝑡𝑠((𝑢𝑛ℎ )− , 𝑣ℎ ) = 𝐹 𝑛 (𝑣ℎ ) ∀𝑣ℎ ∈ 𝑉 , (32)

where 𝑎𝐽 (⋅, ⋅) is defined in (19), and

( +
) ( 𝑛−1 +
)
𝐹 𝑛 (𝑣ℎ ) ∶= 𝛥𝑡(𝑓 𝑛−1+𝜃 , 𝑣ℎ )𝛺 − 𝛥𝑡(1 − 𝜃)𝜀 ∇(𝑢𝑛−1
ℎ ) , ∇𝑣ℎ 𝛺 − 𝛥𝑡(1 − 𝜃) 𝜷 ⋅ ∇(𝑢ℎ ) , 𝑣ℎ 𝛺
( +
) +
− (𝜇𝛥𝑡(1 − 𝜃) − 1) (𝑢𝑛−1ℎ ) , 𝑤ℎ − 𝛥𝑡(1 − 𝜃)𝐽 ((𝑢𝑛−1
ℎ ) , 𝑣ℎ ) ∀𝑣ℎ ∈ 𝑉 . (33)
𝛺
The realisation that at each time step the method (26) is related to the method proposed in [18] will be instrumental in the
well-posedness result presented in the next section. □

3.1. Well-posedness

In this section, we analyse the well-posedness of (26). The first step is given by the following monotonicity result, whose the
proof is similar to that of [17, Lemma 3.1].

Lemma 3.3. The bilinear form 𝑠(⋅, ⋅) defined in (28) for 0 ≤ 𝑛 ≤ 𝑁 satisfies the following inequalities:

𝑠(𝑣−
ℎ − 𝑤− + +
ℎ , 𝑣 ℎ − 𝑤ℎ ) ≥ 0 , (34)
𝑠(𝑣− +
ℎ , 𝑤ℎ − 𝑣+
ℎ
) ≤0, (35)

for every 𝑣ℎ , 𝑤ℎ ∈ 𝑉 .

The well-posedness of (26) is addressed now. For this, the connection pointed out at the end of the last section is exploited.
Namely, we use an approach very similar to the one used in [18, Theorem 3.1] to prove that, for each 𝑛 = 1, … , 𝑁 the problem
(32) has a unique solution, which imples that the problem (26) is well-posed.

Theorem 3.4. Let 𝑛 = 1, … , 𝑁, then,

a There exists 𝑢𝑛ℎ ∈ 𝑉 that solves (32).

b (𝑢𝑛ℎ )+ ∈ 𝑉+ satisfies

((𝑢𝑛ℎ )+ , 𝑣ℎ − (𝑢𝑛ℎ )+ )𝛺 + 𝛥𝑡𝜃𝑎𝐽 ((𝑢𝑛ℎ )+ , 𝑣ℎ − (𝑢𝑛ℎ )+ ) ≥ 𝐹 𝑛 (𝑣ℎ − (𝑢𝑛ℎ )+ ) ∀𝑣ℎ ∈ 𝑉+ . (36)

c (𝑢𝑛ℎ )− is the unique solution of

𝛥𝑡𝑠((𝑢𝑛ℎ )− , 𝑣ℎ ) = 𝐹 𝑛 (𝑣ℎ ) − ((𝑢𝑛ℎ )+ , 𝑣ℎ )𝛺 + 𝛥𝑡𝜃𝑎𝐽 ((𝑢𝑛ℎ )+ , 𝑣ℎ ) ∀𝑣ℎ ∈ 𝑉 . (37)

d The solution of (32) is unique.

Proof. As was mentioned earlier, the proof of this result has many points in common with that of [18, Theorem 3.1]. So, we will
skip many of the technical details and will just present the main arguments.
a. We begin by defining the following bilinear form
( )
𝐵(𝑣ℎ , 𝑤ℎ ) ∶= 𝛥𝑡𝜃𝜀 ∇𝑣ℎ , ∇𝑤ℎ 𝛺 + (𝜇𝛥𝑡𝜃 + 1)(𝑣ℎ , 𝑤ℎ )𝛺 + 𝛥𝑡𝜃𝐽 (𝑣ℎ , 𝑤ℎ ) ∀𝑣ℎ , 𝑤ℎ ∈ 𝑉 ,

6
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

and the mapping

𝑇 ∶𝑉 ⟶ 𝑉 ,
𝑢̂ 𝑛ℎ ⟶ 𝑢𝑛ℎ = 𝑇 (𝑢̂ 𝑛ℎ ), 𝑛 = 1, … , 𝑁,
where 𝑢𝑛ℎ = 𝑇 (𝑢̂ 𝑛ℎ ) solves the following equation

𝐵((𝑢𝑛ℎ )+ , 𝑣ℎ ) + 𝛥𝑡𝑠((𝑢𝑛ℎ )− , 𝑣ℎ ) = 𝐹 𝑛 (𝑣ℎ ) − 𝛥𝑡𝜃𝜷 ⋅ ∇((𝑢̂ 𝑛ℎ )+ , 𝑣ℎ )𝛺 , (38)

at each time-level 1 ≤ 𝑛 ≤ 𝑁 and 𝐹 𝑛 (⋅) is defined in (33). We observe that 𝑢𝑛ℎ
solves (32) if and only if = So, the proof will 𝑇 (𝑢𝑛ℎ ) 𝑢𝑛ℎ .
consist on proving that 𝑇 satisfies the hypotheses of Brouwer’s fixed point Theorem [25, Theorem 10.41].
(i) T is well-defined: To prove that 𝑇 is well-defined, we see that (38) is a particular example of the method proposed in [17].
So, using [17, Theorem 3.2], there exists a unique solution 𝑢𝑛ℎ ∈ 𝑉 of (38), and thus 𝑇 is well-defined.
(ii) T is continuous: Using the monotonicity result proven in [17, Theorem 3.3] we obtain, for all 𝑣ℎ , 𝑤ℎ ∈ 𝑉
𝐵(𝑣+
ℎ
− 𝑤+ , 𝑣 − 𝑤ℎ ) + 𝛥𝑡𝑠(𝑣−
ℎ ℎ
− 2
ℎ − 𝑤ℎ , 𝑣ℎ − 𝑤ℎ ) ≥ 𝐶‖𝑣ℎ − 𝑤ℎ ‖ℎ,𝜃𝛥𝑡 .

Next, let 𝑣̂ ℎ , 𝑤̂ ℎ ∈ 𝑉 and let 𝑣ℎ = 𝑇 (𝑣̂ ℎ ) and 𝑤ℎ = 𝑇 (𝑤̂ ℎ ). Then, a lengthy calculation using the last result, (38), integration by parts,
Hölder’s inequality, Lemma 2.2, and (29), leads to the following Lipschitz continuity of the operator 𝑇 :
1
‖𝑇 (𝑣̂ ℎ ) − 𝑇 (𝑤̂ ℎ )‖ℎ,𝜃𝛥𝑡 ≤ 𝐶𝜃𝛥 𝑡 𝜀̃− 2 ‖𝜷‖0,∞,𝛺 ‖𝑣̂ ℎ − 𝑤̂ ℎ ‖0,𝛺 .
(iii) There exists 𝑅 > 0, such that 𝑇 (𝐵(0, 𝑅)) ⊆ 𝐵(0, 𝑅) ∶ Let 𝑧̂ ℎ ∈ 𝑉 be arbitrary and 𝑧ℎ = 𝑇 (𝑧̂ ℎ ). By using 𝑣ℎ = 𝑧+
ℎ
in (38), we
get

𝐵(𝑧+ , 𝑧+ ) + 𝛥𝑡𝑠(𝑧−
ℎ ℎ
+ + + + +
ℎ , 𝑧ℎ ) = 𝐹 (𝑧ℎ ) − 𝜃𝛥𝑡(𝜷 ⋅ ∇𝑧̂ ℎ , 𝑧ℎ )𝛺 ≤ 𝑀‖𝑧ℎ ‖ℎ,𝜃𝛥𝑡 . (39)
⏟⏞⏞⏞⏟⏞⏞⏞⏟
≥0

In fact, using the Cauchy–Schwarz and Young inequalities, the fact that 𝜇̃ = 1 + 𝜇𝜃𝛥𝑡 ≠ 0 and 𝜃 ≥ 12 , we obtain
( 1 1 1
|𝐹 𝑛 (𝑧+
ℎ
)| ≤ 𝐶 𝛥𝑡𝜇̃ − 2 ∥ 𝑓 𝑛−1+𝜃 ∥0,𝛺 +𝛥𝑡𝜃‖𝜷‖0,∞,𝛺 𝜀̃− 2 𝜇̃ − 2 ‖(𝑢𝑛−1 +
ℎ ) ‖ℎ,𝜃𝛥𝑡
)
+ +
ℎ ) ‖ℎ,𝜃𝛥𝑡 ‖𝑧ℎ ‖ℎ,𝜃𝛥𝑡 .
+‖(𝑢𝑛−1

Moreover, integration by parts and the Hölder inequality yield

1
𝜃𝛥𝑡(𝜷 ⋅ ∇𝑧̂ + , 𝑧+ ) ≤ 𝜃𝛥𝑡𝜀̃− 2 ‖𝑧̂ +
ℎ ℎ 𝛺
‖ ‖𝜷‖0,∞,𝛺 ∥ 𝑧+
ℎ 0,𝛺
∥
ℎ ℎ,𝜃𝛥𝑡
.
So, setting
( 1 1 1
𝑀 ∶= 𝐶 𝛥𝑡𝜇̃ − 2 ∥ 𝑓 𝑛−1+𝜃 ∥0,𝛺 +𝛥𝑡𝜃‖𝜷‖0,∞,𝛺 𝜀̃− 2 𝜇̃ − 2 ‖(𝑢𝑛−1 +
ℎ ) ‖ℎ,𝜃𝛥𝑡
)
+ − 12
ℎ ) ‖ℎ,𝜃𝛥𝑡 + 𝜃𝛥𝑡𝜀̃
+‖(𝑢𝑛−1 𝜅(𝑡𝑛 )‖𝜷‖0,∞,𝛺 , (40)

and using the ellipticity of 𝐵(⋅, ⋅), (25), and (39) we arrive at the following bound for 𝑧+
ℎ
:
‖𝑧+ ‖
ℎ ℎ,𝜃𝛥𝑡
≤𝑀.
Next, we take 𝑣ℎ = 𝑧−ℎ
in (38). Using analogous arguments, that is, integration by parts, Hölder’s inequality, (25), and Lemma
3.1 we derive the following bound for 𝑧−ℎ
:
‖𝑧−
ℎ ‖ℎ,𝜃𝛥𝑡 ≤ 𝐶2 (𝑓
𝑛−1+𝜃 𝑛−1
, 𝑢ℎ , 𝜀, ̃ 𝜷, 𝜅(𝑡𝑛 ), ℎ, 𝛥𝑡),
̃ 𝜇,
where 𝐶2 (𝑓 𝑛−1+𝜃 , 𝑢𝑛−1
ℎ
, 𝜀, ̃ 𝜷, 𝜅(𝑡𝑛 ), ℎ, 𝛥𝑡) is independent of 𝑧ℎ . Hence, 𝑧ℎ = 𝑇 (𝑧̂ ℎ ) satisfies the following (uniform) bound
̃ 𝜇,
‖𝑧ℎ ‖ℎ,𝜃𝛥𝑡 ≤ ‖𝑧− +
ℎ ‖ℎ,𝜃𝛥𝑡 + ‖𝑧ℎ ‖ℎ,𝜃𝛥𝑡 ≤ 𝑀 + 𝐶2 (𝑓
𝑛−1+𝜃 𝑛−1
, 𝑢ℎ , 𝜀, ̃ 𝜷, 𝜅(𝑡𝑛 ), ℎ, 𝛥𝑡) =∶ 𝑅.
̃ 𝜇,
Therefore, 𝑧ℎ = 𝑇 (𝑧̂ ℎ ) ∈ 𝐵(0, 𝑅), for every 𝑧̂ ℎ ∈ 𝑉 , which shows that 𝑇 (𝐵(0, 𝑅)) ⊆ 𝐵(0, 𝑅). Hence, using Brouwer’s fixed point
theorem, there exists one 𝑢𝑛ℎ ∈ 𝑉 such that 𝑇 (𝑢𝑛ℎ ) = 𝑢𝑛ℎ . In other words, problem (32) has at least one solution.
The proofs of (b) and (c) are identical to those of [18, Lemma 3.3]. Finally, the proof of (d) is identical to that of [18,
Corollary 3.1]. □

Remark 3.5. It is worth mentioning that, due to the equivalence between (26) and the variational inequality (36), and the well-
posedness of the latter, (𝑢𝑛ℎ )+ is independent of the choice of the stabilisation, as long as it satisfies (31) and (34). In particular, all
the mentioned methods proven in this work are remained valid if 𝑠(⋅, ⋅) defined in (28) is replaced by
𝑀 (
∑ )
𝑠(𝑣ℎ , 𝑤ℎ ) ∶= 𝛼 𝜀h(𝒙𝑖 )𝑑−2 + ‖𝜷(𝒙, 𝑡𝑛 )‖0,∞,𝜔𝑖 h(𝒙𝑖 )𝑑−1 + 𝜇h(𝒙𝑖 )𝑑 𝑣ℎ (𝒙𝑖 )𝑤ℎ (𝒙𝑖 ).
𝑖=1

In this case, the solution (𝑢𝑛ℎ )+ remains unchanged, as it still satisfies (36), meaning the overall analysis remains the same.
1
The inclusion of the factor 𝛥𝑡 in the time derivative was primarily motivated by the performance of the nonlinear solver. Without
this factor, the nonlinear solver exhibited significantly slower convergence.

7
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

4. Stability and error analysis

This section is devoted to proving a stability result and derive optimal error estimates for the method (26) in the particular case
𝜃 = 1, that is, in the case the time discretisation is carried out using the implicit Euler method. An important tool that will be used
throughout is the following discrete Gronwall Lemma, proved originally in [26, Lemma 5.1].

Lemma 4.1. Let 𝑘, 𝐵, 𝑎𝑛 , 𝑏𝑛 , 𝑐𝑛 , 𝛾𝑛 , 𝑛 = 0, … , 𝑚, be non-negative numbers such that

∑
𝑚 ∑
𝑚 ∑
𝑚
𝑎𝑛 + 𝑘 𝑏𝑛 ≤ 𝑘 𝛾𝑛 𝑎𝑛 + 𝑘 𝑐𝑛 + 𝐵 for 𝑚 ≥ 0.
𝑛=0 𝑛=0 𝑛=0

Suppose 𝑘𝛾𝑛 ≤ 1 for every 𝑗, and set 𝜎𝑛 = (1 − 𝑘𝛾𝑛 )−1 . Then

( 𝑚 )( 𝑚 )
∑𝑚 ∑ ∑
𝑎𝑚 + 𝑘 𝑏𝑛 ≤ exp 𝑘 𝜎𝑛 𝛾𝑛 𝑘 𝑐𝑛 + 𝐵 for 𝑚 ≥ 0. (41)
𝑛=0 𝑛=0 𝑛=0

For the case 𝜃 = 1, the method (26) reads:

⎧ For 1 ≤ 𝑛 ≤ 𝑁, find 𝑢 ∈ 𝑉 such that
⎪ ℎ 
𝑛 + 𝑛 𝑛 (42)
⎨ (𝛿(𝑢ℎ ) , 𝑣ℎ )𝛺 + 𝑎ℎ (𝑢ℎ ; 𝑣ℎ ) = (𝑓 , 𝑣ℎ )𝛺 ∀𝑣ℎ ∈ 𝑉 ,
⎪ 𝑢 0 = 𝑖 𝑢0 .
⎩ ℎ ℎ
To prove the stability, we use the test function 𝑣ℎ = (𝑢𝑛ℎ )+ in (42), and obtain

(𝛿(𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ )𝛺 + 𝜀(∇(𝑢𝑛ℎ )+ , ∇(𝑢𝑛ℎ )+ )𝛺 + (𝜷 ⋅ ∇(𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ )𝛺

+ 𝜇((𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ )𝛺 + 𝐽 ((𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ ) + 𝑠((𝑢𝑛ℎ )− , (𝑢𝑛ℎ )+ ) = (𝑓 𝑛 , (𝑢𝑛ℎ )+ )𝛺 ,
( )
or, equivalently, using that 𝜷 ⋅ ∇(𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ 𝛺 = 0,
{ }
𝑛 + 2
((𝑢𝑛ℎ )+ − (𝑢𝑛−1 + 𝑛 + 𝑛 + 2
ℎ ) , (𝑢ℎ ) )𝛺 + 𝛥𝑡 𝜀 |(𝑢ℎ ) |1,𝛺 + 𝜇 ‖(𝑢ℎ ) ‖0,𝛺 + 𝐽 ((𝑢ℎ ) , (𝑢ℎ ) )
𝑛 + 𝑛 +

+ 𝛥𝑡𝑠((𝑢𝑛ℎ )− , (𝑢𝑛ℎ )+ ) = 𝛥𝑡(𝑓 𝑛 , (𝑢𝑛ℎ )+ )𝛺 .

The relation 2𝑝(𝑝 − 𝑞) = 𝑝2 + (𝑝 − 𝑞)2 − 𝑞 2 , the Cauchy–Schwarz inequality, and the fact that 𝑠((𝑢𝑛ℎ )− , (𝑢𝑛ℎ )+ ) ≥ 0 (by Lemma 3.3) lead
to
{
∥ (𝑢𝑛ℎ )+ ∥20,𝛺 − ∥ (𝑢𝑛−1 + 2 𝑛 + 𝑛−1 + 2
ℎ ) ∥0,𝛺 + ∥ (𝑢ℎ ) − (𝑢ℎ ) ∥0,𝛺 +2𝛥𝑡 𝜀|(𝑢ℎ ) |1,𝛺
𝑛 + 2

}
+𝜇 ∥ (𝑢𝑛ℎ )+ ∥20,𝛺 +𝐽 ((𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ ) ≤ 2𝛥𝑡 ∥ 𝑓 𝑛 ∥0,𝛺 ∥ (𝑢𝑛ℎ )+ ∥0,𝛺 . (43)

Using Young’s inequality for the right hand side and then summing through 𝑛, 𝑛 = 0, … , 𝑚, we get that
∑
𝑚 ∑
𝑚 {
+ 2
∥ (𝑢𝑚
ℎ ) ∥0,𝛺 + ∥ (𝑢𝑛ℎ )+ − (𝑢𝑛−1 + 2
ℎ ) ∥0,𝛺 +2 𝛥𝑡 𝜀|(𝑢𝑛ℎ )+ |21,𝛺 + 𝜇 ∥ (𝑢𝑛ℎ )+ ∥20,𝛺
𝑛=0 𝑛=0
} ∑𝑚 ( )
1
+𝐽 ((𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ ) ≤∥ 𝑢0ℎ ∥20,𝛺 + 𝛥𝑡 𝑇 ∥ 𝑓 𝑛 ∥20,𝛺 + ∥ (𝑢𝑛ℎ )+ ∥20,𝛺 .
𝑛=0
𝑇
( )−1
𝛥𝑡 𝛥𝑡
If we set 𝑎𝑛 =∥ (𝑢𝑛ℎ )+ ∥20,𝛺 , 𝐵 =∥ 𝑢0ℎ ∥20,𝛺 , 𝑘 = 1, 𝛾𝑛 = 𝑇
and 𝜎𝑛 = 1 − 𝑇
, 𝑐𝑛 = 𝛥𝑡𝑇 ∥ 𝑓 𝑛 ∥20,𝛺 and
( )
𝑏𝑛 = 2𝛥𝑡 𝜀|(𝑢𝑛ℎ )+ |21,𝛺 + 𝜇 ∥ (𝑢𝑛ℎ )+ ∥20,𝛺 +𝐽 ((𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ ) ,

then using the Grönwall’s inequality Lemma 4.1, we get

𝑚 (
∑ )
+ 2
∥ (𝑢𝑚
ℎ ) ∥0,𝛺 + 2𝛥𝑡 𝜀|(𝑢𝑛ℎ )+ |21,𝛺 + 𝜇 ∥ (𝑢𝑛ℎ )+ ∥20,𝛺 +𝐽 ((𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ )
𝑛=0
(𝑚 )( )
∑ 𝛥𝑡 ( )
𝛥𝑡 −1 ∑𝑚
≤ exp 1− ∥ 𝑢0ℎ ∥20,𝛺 +𝛥𝑡𝑇 ∥ 𝑓 𝑛 ∥20,𝛺 .
𝑛=0
𝑇 𝑇 𝑛=0

In this way we have proved the following stability result for the scheme (42).

Lemma 4.2. Let 𝑢𝑛ℎ ∈ 𝑉 , for 𝑛 = 1, … , 𝑁 solve (42). Then the following stability estimates holds true:
𝑁 (
∑ )
+ 2
max ∥ (𝑢𝑚
ℎ ) ∥0,𝛺 + 2𝛥𝑡 𝜀|(𝑢𝑛ℎ )+ |21,𝛺 + 𝜇 ∥ (𝑢𝑛ℎ )+ ∥20,𝛺 +𝐽 ((𝑢𝑛ℎ )+ , (𝑢𝑛ℎ )+ )
1≤𝑚≤𝑁
𝑛=0
( )
∑
𝑁
≤ 𝑒2 ∥ 𝑢0ℎ ∥20,𝛺 +𝛥𝑡𝑇 ∥ 𝑓 𝑛 ∥20,𝛺 .
𝑛=0

8
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Remark 4.3. In the case particular case 𝑓 = 0, then (43) implies that
+ 0
∥ (𝑢𝑚
ℎ ) ∥0,𝛺 ≤∥ 𝑢ℎ ∥0,𝛺 ,

and then (42) is strongly stability preserving. □

The next result states optimal order error estimates for the method (42).

Theorem 4.4. Let 𝑢0 ∈ 𝐻 𝑘+1 (𝛺), 𝑢 ∈ 𝐿∞ ((0, 𝑇 ); 𝐻 𝑘+1 (𝛺))∩𝐻 1 ((0, 𝑇 ); 𝐻 𝑘+1 (𝛺))∩𝐻 2 ((0, 𝑇 ); 𝐿2 (𝛺)) be the solution of (2), 𝑢(⋅, 𝑡) ∈ 𝐻01 (𝛺)
for almost all 𝑡 ∈ [0, 𝑇 ], and 𝑢𝑛ℎ ∈ 𝑉 be the solution of (42) at the time step 𝑛. Then, defining 𝐸 𝑚 = (𝑢𝑚 ℎ
)+ − 𝑢𝑚 , there exists a constant
𝐶 > 0, independent of ℎ, 𝛥𝑡 and any physical parameter, such that
𝑁 (
∑
)
max ∥ 𝐸 𝑚 ∥20,𝛺 +𝛥𝑡 𝜀|𝐸ℎ𝑛 |21,𝛺 + 𝜇 ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 )
1≤𝑚≤𝑁
𝑛=1
[ { ( )
∑𝑁
≤ 𝐶𝑒 2
ℎ 2𝑘
𝛥𝑡 𝜀 + 𝑇 ‖𝜷‖20,∞,𝛺 + ℎ𝛾‖𝜷‖0,∞,𝛺 + ℎ 𝑇 𝜇 2 2
|𝑢𝑛 |2𝑘+1,𝛺
𝑛=0
} ]
𝑇 𝑇
+ℎ 2
max |𝑢𝑚 |2𝑘+1,𝛺 + 𝑇ℎ 2
|𝜕𝑡 𝑢(𝑡)|2𝑘+1,𝛺 d𝑡 + 𝛥𝑡 𝑇 2
∥ 𝜕𝑡𝑡 𝑢(𝑡) ∥20,𝛺 d𝑡 . (44)
1≤𝑚≤𝑁 ∫0 ∫0

Proof. For 𝑛 = 1, … , 𝑁 we decompose 𝐸 𝑛 = (𝑢𝑛ℎ )+ − 𝑢𝑛 as

( ) ( )
𝐸 𝑛 = (𝑢𝑛ℎ )+ − 𝑢𝑛 = (𝑢𝑛ℎ )+ − 𝑖ℎ 𝑢𝑛 + 𝑖ℎ 𝑢𝑛 − 𝑢𝑛 =∶ 𝐸ℎ𝑛 + 𝜂ℎ𝑛 , (45)

𝑢𝑛
where we recall that 𝑖ℎ is the Lagrange interpolant (12) of 𝑢𝑛 .
Subtracting (4) from the method (42) we arrive at the following
error equation
( 𝑛 + ) ( ) ( )
𝛿(𝑢ℎ ) − 𝜕𝑡 𝑢𝑛 , 𝑣ℎ 𝛺 + 𝜀 ∇((𝑢𝑛ℎ )+ − 𝑢𝑛 ), ∇𝑣ℎ 𝛺 + 𝜷 ⋅ ∇((𝑢𝑛ℎ )+ − 𝑢𝑛 ), 𝑣ℎ 𝛺
( )
+ 𝜇 (𝑢𝑛ℎ )+ − 𝑢𝑛 , 𝑣ℎ 𝛺 + 𝐽 ((𝑢𝑛ℎ )+ , 𝑣ℎ ) + 𝑠((𝑢𝑛ℎ )− , 𝑣ℎ ) = 0. (46)

Rearranging and using (45),we get

( )
(𝛿𝐸ℎ𝑛 , 𝑣ℎ )𝛺 + 𝜀 ∇𝐸ℎ𝑛 , ∇𝑣ℎ 𝛺 + (𝜷 ⋅ ∇𝐸ℎ𝑛 , 𝑣ℎ )𝛺 + 𝜇(𝐸ℎ𝑛 , 𝑣ℎ )𝛺 + 𝐽 ((𝑢𝑛ℎ )+ , 𝑣ℎ ) + 𝑠((𝑢𝑛ℎ )− , 𝑣ℎ )
= −(𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 , 𝑣ℎ )𝛺 − 𝜀(∇𝜂ℎ𝑛 , ∇𝑣ℎ )𝛺 − (𝜷 ⋅ ∇𝜂ℎ𝑛 , 𝑣ℎ )𝛺 − 𝜇(𝜂ℎ𝑛 , 𝑣ℎ )𝛺 . (47)

Since 𝑢𝑛 ∈ 𝐻 2 (𝛺), then 𝐽 (𝑢𝑛 , 𝑣) = 0, and so we deduce that

𝐽 ((𝑢𝑛ℎ )+ , 𝑣) = 𝐽 ((𝑢𝑛ℎ )+ − 𝑢 , 𝑣) = 𝐽 (𝐸ℎ𝑛 , 𝑣) + 𝐽 (𝜂ℎ𝑛 , 𝑣).

𝑛

Using the test function 𝑣ℎ = 𝐸ℎ𝑛 in (47) and (𝜷 ⋅ ∇𝐸ℎ𝑛 , 𝐸ℎ𝑛 )𝛺 = 0, we get

(𝛿𝐸ℎ𝑛 , 𝐸ℎ𝑛 )𝛺 + 𝜀 |𝐸ℎ𝑛 |21,𝛺 + 𝜇 ‖𝐸ℎ𝑛 ‖20,𝛺 + 𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 ) + 𝑠((𝑢𝑛ℎ )− , 𝐸ℎ𝑛 )
( )
= −(𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 , 𝐸ℎ𝑛 )𝛺 − 𝜀 ∇𝜂ℎ𝑛 , ∇𝐸ℎ𝑛 𝛺 − (𝜷 ⋅ ∇𝜂ℎ𝑛 , 𝐸ℎ𝑛 )𝛺 − 𝜇(𝜂ℎ𝑛 , 𝐸ℎ𝑛 )𝛺 − 𝐽 (𝜂ℎ𝑛 , 𝐸ℎ𝑛 ). (48)

Since (𝑖ℎ 𝑢𝑛 )− = 0, the monotonicity inequality (35) yields

𝑠((𝑢𝑛ℎ )− , 𝐸ℎ𝑛 ) = 𝑠((𝑢𝑛ℎ )− , (𝑢𝑛ℎ )+ − 𝑖ℎ 𝑢𝑛 ) = 𝑠((𝑢𝑛ℎ )− − (𝑖ℎ 𝑢𝑛 )− , (𝑢𝑛ℎ )+ − (𝑖ℎ 𝑢𝑛 )+ ) ≥ 0. (49)

So, using the relation 2𝑝(𝑝−𝑞) = 𝑝2 +(𝑝−𝑞)2 −𝑞 2

for the first term of (48), applying the inequality (49), next using the Cauchy–Schwarz
inequality and then Young’s inequality for the terms on the right hand side, we get
( )
∥ 𝐸ℎ𝑛 ∥20,𝛺 − ∥ 𝐸ℎ𝑛−1 ∥20,𝛺 + ∥ 𝐸ℎ𝑛 − 𝐸ℎ𝑛−1 ∥20,𝛺 +2𝛥𝑡 𝜀|𝐸ℎ𝑛 |21,𝛺 + 𝜇 ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 )
(
≤ 2𝛥𝑡 ∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥0,𝛺 ∥ 𝐸ℎ𝑛 ∥0,𝛺 +𝜀|𝜂ℎ𝑛 |1,𝛺 |𝐸ℎ𝑛 |1,𝛺 + ∥ 𝜷 ⋅ ∇𝜂ℎ𝑛 ∥0,𝛺 ∥ 𝐸ℎ𝑛 ∥0,𝛺
1 1)
+ 𝜇 ∥ 𝜂ℎ𝑛 ∥0,𝛺 ∥ 𝐸ℎ𝑛 ∥0,𝛺 +𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 ) 2 𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 ) 2
( ( )2
≤ 𝛥𝑡 𝜀|𝜂ℎ𝑛 |21,𝛺 + 𝜀|𝐸ℎ𝑛 |21,𝛺 + 𝑇 ∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥0,𝛺 + ∥ 𝜷 ⋅ ∇𝜂ℎ𝑛 ∥0,𝛺 +𝜇 ∥ 𝜂ℎ𝑛 ∥0,𝛺
)
1
+ ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 ) + 𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 ) .
𝑇
Rearranging terms on the both sides of the inequality yields
( )
∥ 𝐸ℎ𝑛 ∥20,𝛺 − ∥ 𝐸ℎ𝑛−1 ∥20,𝛺 + ∥ 𝐸ℎ𝑛 − 𝐸ℎ𝑛−1 ∥20,𝛺 +𝛥𝑡 𝜀|𝐸ℎ𝑛 |21,𝛺 + 2𝜇 ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 )
( ( )2
≤ 𝛥𝑡 𝜀|𝜂ℎ𝑛 |21,𝛺 + 𝑇 ∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥0,𝛺 + ∥ 𝜷 ∥0,∞,𝛺 |𝜂ℎ𝑛 |1,𝛺 + 𝜇 ∥ 𝜂ℎ𝑛 ∥0,𝛺
)
1
+ ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 )
𝑇

9
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691
(( )
≤ 𝐶𝛥𝑡 𝜀 + 𝑇 ‖𝜷‖20,∞,𝛺 |𝜂ℎ𝑛 |21,𝛺 + 𝑇 ∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥20,𝛺 +𝑇 𝜇 2 ∥ 𝜂ℎ𝑛 ∥20,𝛺
)
1
+ ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 ) .
𝑇
Summing for 𝑛 = 0 to 𝑚 and using 𝐸ℎ0 = 0 leads to
𝑚 (
∑ )
∥ 𝐸ℎ𝑚 ∥20,𝛺 + 𝛥𝑡 𝜀|𝐸ℎ𝑛 |21,𝛺 + 2𝜇 ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 )
𝑛=1
𝑚 {(
∑ )
≤ 𝐶 𝛥𝑡 𝜀 + 𝑇 ‖𝜷‖20,∞,𝛺 |𝜂ℎ𝑛 |21,𝛺 + 𝑇 ∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥20,𝛺
𝑛=0
}
1
+𝑇 𝜇 2 ∥ 𝜂ℎ𝑛 ∥20,𝛺 +
∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 ) .
𝑇
We are now ready to use Grönwall’s Lemma 4.1 with 𝑘 = 1, 𝛾𝑛 = 𝛥𝑡 𝑇
, 𝜎𝑛 = (1 − 𝛥𝑡 𝑇
)−1 , 𝑎𝑛 =∥ 𝐸ℎ𝑛 ∥0,𝛺 , 𝐵 = 0 and
(( ) )
𝑐𝑛 = 𝛥𝑡 𝜀 + 𝑇 ‖𝜷‖20,∞,𝛺 |𝜂ℎ𝑛 |21,𝛺 + 𝑇 ∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥20,𝛺 +𝑇 𝜇2 ∥ 𝜂ℎ𝑛 ∥20,𝛺 +𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 ) ,

which gives
𝑚 (
∑ )
∥ 𝐸ℎ𝑚 ∥20,𝛺 + 𝛥𝑡 𝜀|𝐸ℎ𝑛 |21,𝛺 + 2𝜇 ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 )
𝑛=1
[ {( )
∑
𝑚
≤ 𝐶𝑒2 𝛥𝑡 𝜀 + 𝑇 ‖𝜷‖20,∞,𝛺 |𝜂ℎ𝑛 |21,𝛺 + 𝑇 ∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥20,𝛺 (50)
𝑛=0
}]
+ 𝑇 𝜇 2 ∥ 𝜂ℎ𝑛 ∥20,𝛺 +𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 ) .

Next, to reach the error estimate (44) we bound each term of the right hand side of (50). First, using the triangle inequality,
yields

∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝜕𝑡 𝑢𝑛 ∥0,𝛺 ≤∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝛿𝑢𝑛 ∥0,𝛺 + ∥ 𝛿𝑢𝑛 − 𝜕𝑡 𝑢𝑛 ∥0,𝛺 .

For the first term in the above inequality, using (15), the Taylor’s Theorem and the Cauchy–Schwarz inequality, we have

∥ 𝛿(𝑖ℎ 𝑢𝑛 ) − 𝛿𝑢𝑛 ∥20,𝛺 ≤ 𝐶ℎ2𝑘+2 |𝛿𝑢𝑛 |2𝑘+1,𝛺

|1 𝑡𝑛 |2
| |
≤ 𝐶ℎ2𝑘+2 | 𝜕𝑡 𝑢(𝑡)d𝑡|
| 𝛥𝑡 ∫𝑡 |
| 𝑛−1 |𝑘+1,𝛺
| ( ) (
1 ) 1 |2
| 𝑡𝑛 2 𝑡𝑛 2|
2𝑘+2 | 1 2 |
≤ 𝐶ℎ | d𝑡 |𝜕𝑡 𝑢(𝑡)| d𝑡 |
| 𝛥𝑡 ∫𝑡 ∫ |
| 𝑛−1 𝑡𝑛−1 |
| |𝑘+1,𝛺
𝑡
ℎ2𝑘+2 𝑛
≤𝐶 |𝜕 𝑢(𝑡)|2𝑘+1,𝛺 d𝑡.
𝛥𝑡 ∫𝑡𝑛−1 𝑡
For the second term, one further use of Taylor’s Theorem and the Cauchy–Schwarz inequality gives

∥ 𝛿𝑢𝑛 − 𝜕𝑡 𝑢𝑛 ∥20,𝛺 = (𝛿𝑢𝑛 − 𝜕𝑡 𝑢𝑛 )2 d𝒙

∫𝛺
( )2
𝑡𝑛
≤ |𝜕𝑡𝑡 𝑢(𝑡)|d𝑡 d𝒙
∫𝛺 ∫𝑡𝑛−1
𝑡𝑛 𝑡𝑛
≤ d𝑡 |𝜕𝑡𝑡 𝑢(𝑡)|2 d𝑡 d𝒙
∫𝛺 ∫𝑡𝑛−1 ∫𝑡𝑛−1
𝑡𝑛
= 𝛥𝑡 ∥ 𝜕𝑡𝑡 𝑢(𝑡) ∥20,𝛺 d𝑡.
∫𝑡𝑛−1
Next, using (15), we have

|𝜂ℎ𝑛 |21,𝛺 ≤ 𝐶ℎ2𝑘 |𝑢𝑛 |2𝑘+1,𝛺 , ‖𝜂ℎ𝑛 ‖20,𝛺 ≤ 𝐶 ℎ2𝑘+2 |𝑢𝑛 |2𝑘+1,𝛺 . (51)

Furthermore, using the inverse inequality (13) and the approximation inequality (15), we have
∑
𝐽 (𝜂ℎ𝑛 , 𝜂ℎ𝑛 ) = 𝛾 ‖𝜷‖0,∞,𝐹 ℎ2𝐹 [[∇𝜂ℎ𝑛 ]] ⋅ [[∇𝜂ℎ𝑛 ]]d𝑠 ≤ 𝐶𝛾ℎ2𝑘+1 ‖𝜷‖0,∞,𝛺 |𝑢𝑛 |2𝑘+1,𝛺 . (52)
∫𝐹
𝐹 ∈𝐼
Gathering all the above bounds, we arrive at
𝑚 (
∑
)
∥ 𝐸ℎ𝑚 ∥20,𝛺 + 𝛥𝑡 𝜀|𝐸ℎ𝑛 |21,𝛺 + 2𝜇 ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 )
𝑛=1

10
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Fig. 1. Three coarse level indicative meshes used in the experiments all with 𝑁 = 5.

[ { ( )
∑
𝑚
≤ 𝐶𝑒 2
ℎ 2𝑘
𝛥𝑡 𝜀 + 𝑇 ‖𝜷‖20,∞,𝛺 + ℎ𝛾‖𝜷‖0,∞,𝛺 + ℎ 𝑇 𝜇 2 2
|𝑢𝑛 |2𝑘+1,𝛺
𝑛=0
} ]
∑
𝑚 𝑡𝑛 ∑
𝑚 𝑡𝑛
+ 𝑇ℎ 2
|𝜕𝑡 𝑢(𝑡)|2𝑘+1,𝛺 d𝑡 + 𝑇 𝛥𝑡 2
∥ 𝜕𝑡𝑡 𝑢(𝑡) ∥20,𝛺 d𝑡 .
∫𝑡𝑛−1 ∫𝑡𝑛−1
𝑛=0 𝑛=0

Finally, using the triangle inequality and (15) once again, we arrive at the final estimate
∑𝑁 ( )
max ∥ 𝐸 𝑚 ∥20,𝛺 +𝛥𝑡 𝜀|𝐸ℎ𝑛 |21,𝛺 + 𝜇 ∥ 𝐸ℎ𝑛 ∥20,𝛺 +𝐽 (𝐸ℎ𝑛 , 𝐸ℎ𝑛 )
1≤𝑚≤𝑁
𝑛=1
[ { ( )
∑𝑁
≤ 𝐶𝑒 2
ℎ 2𝑘
𝛥𝑡 𝜀 + 𝑇 ‖𝜷‖20,∞,𝛺 + ℎ𝛾‖𝜷‖0,∞,𝛺 + ℎ 𝑇 𝜇 2 2
|𝑢𝑛 |2𝑘+1,𝛺
𝑛=0
} ]
𝑇 𝑇
+ℎ 2
max |𝑢𝑚 |2𝑘+1,𝛺 + 𝑇ℎ 2
|𝜕𝑡 𝑢(𝑡)|2𝑘+1,𝛺 d𝑡 + 𝑇 𝛥𝑡 2
∥ 𝜕𝑡𝑡 𝑢(𝑡) ∥20,𝛺 d𝑡 ,
1≤𝑚≤𝑁 ∫0 ∫0

which proves the result. □

5. Numerical experiments

In this section we present two experiments to test the numerical performance of (26). In these experiments we have used
𝛺 = (0, 1)2 , and the value 𝛼 = 1 in the stabilising bilinear form 𝑠(⋅, ⋅). Except for the very last numerical result in this section,
we have used two types of meshes, a three-directional triangular mesh and a regular quadrilateral one. The coarsest level of each
is depicted in Fig. 1.
To solve the nonlinear problem (26) at each discrete time 𝑡𝑛 , 𝑛 = 1, 2, … , 𝑁, first we set 𝑢̃ 0 ∶= 𝑢𝑛−1
ℎ
. Next, by choosing an
appropriate damping parameter 𝜔 ∈ (0, 1] the following fixed point Richardson-like iterative method is used to find 𝑢̃ 𝑚+1 ∈ 𝑉 , such
that
( 𝑚+1 ) ( ) ( ) ( )
𝑢̃ , 𝑣ℎ 𝛺 + 𝛥𝑡𝜃𝑎𝐽 𝑢̃ 𝑚+1 , 𝑣ℎ = 𝑢̃ 𝑚 , 𝑣ℎ 𝛺 + 𝛥𝑡𝜃𝑎𝐽 𝑢̃ 𝑚 , 𝑣ℎ (53)
{ 𝑛 [( 𝑚 + ) ( 𝑚+ ) ( )]}
+ 𝜔 𝐹 (𝑣ℎ ) − (𝑢̃ ) , 𝑣ℎ 𝛺 + 𝛥𝑡𝜃𝑎𝐽 (𝑢̃ ) , 𝑣ℎ + 𝑠 (𝑢̃ 𝑚 )− , 𝑣ℎ ∀𝑣ℎ ∈ 𝑉 ,

for 𝑚 = 1, 2, … , 𝑁max , or until the following stopping criterion is achieved

∥ 𝑢̃ 𝑚+1 − 𝑢̃ 𝑚 ∥0,𝛺 ≤ 10−8 . (54)

Finally, for 𝑚0 which satisfies in (54), we set = 𝑢𝑛ℎ 𝑢̃ 𝑚0 +1 .

In all figures, 𝑃 − 1 indicates the number of divisions in the 𝑥 and 𝑦 directions, resulting in a total of 𝑃 2 vertices, including the
boundary. We evaluate the method’s asymptotic performance in the ∥ ⋅ ∥0,𝛺 -norm at the final step, i.e., ∥ 𝑒𝑁 ℎ
∥0,𝛺 , and to verify the
result from Theorem 4.4 we examine the asymptotic behaviour of the error by the following norm
∑
𝑁 ( )
∥ 𝑒𝑁 ∥2ℎ ∶=∥ 𝑒𝑁 ∥20,𝛺 + 𝛥𝑡 𝜀 ∥ ∇𝑒𝑛ℎ ∥20,𝛺 +𝜇 ∥ 𝑒𝑛ℎ ∥20,𝛺 +𝐽 (𝑒𝑛ℎ , 𝑒𝑛ℎ ) . (55)
𝑛=0

We have used P1 and P2 elements in the triangular meshes, and Q1 elements in the quadrilateral mesh. In the numerical experiments
we use the bound preserving Euler (BP-Euler) (42) and the bound preserving Crank–Nicholson (BP-CN), i.e., the method (26) with
𝜃 = 12 , even though stability and error estimates for BP-CN has not been proven.

11
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Fig. 2. Comparison of the error of the approximated solution by the BP-Euler method and BP-CN method with the exact solution in ∥ ⋅ ∥0,𝛺 -norm (using mesh
1(a)).

Example 1 (A problem with a smooth solution). We consider 𝜇 = 1, 𝜀 = 10−6 , 𝜷 = (2, 1), and set 𝑓 and 𝑢0 such that the function
𝑢(𝑥, 𝑦, 𝑡) = exp(𝑡) sin(𝜋𝑥) sin(𝜋𝑦) 𝛺 = (0, 1)2 ,
is the analytical solution of (2). Notice that 𝑢(𝑥, 𝑦, 𝑡) ∈ [0, exp(𝑡)], and thus we set 𝜅(𝑡) = exp(𝑡) as the upper bound at time 𝑡. The CIP
stabilisation parameter 𝛾 = 0.05 has been used in (17) and we set 𝜔 = 0.1 for the damping parameter in all the time steps.
Fig. 2 illustrates the asymptotic behaviour of the error ∥ 𝑒𝑁 ℎ
∥0,𝛺 using P1 and P2 elements. These results align with the theoretical
findings we established in Theorem 4.4. By fixing 𝛥𝑡 = 4 × 10−4 and decreasing the mesh size as depicted in Figs. 2(a) and 2(c), we
observe second- and third-order convergence when using P1 and P2 elements, respectively, for both BP-Euler and BP-CN. Also, by
fixing the mesh size ℎ = 5 × 10−3 and varying the length of the time-step 𝛥𝑡, as shown in Figs. 2(b) and 2(d), we obtained first-order
convergence for the Euler method and second-order convergence for Crank–Nicholson for both P1 and P2 elements, as expected.
Fig. 3 depicts the asymptotic behaviour of the error ∥ 𝑒𝑁 ∥2ℎ using P1 and P2 elements. These results also corroborate the
theoretical results we proved in Theorem 4.4. By fixing the time step and decreasing the mesh size as shown in Figs. 3(a) and
3(c), we observed second- and third-order convergence when using P1 and P2 elements, respectively for the BP-Euler method. This
extra order of convergence is, most likely, due to the fact that the small value of 𝜀 makes that the ‖ ⋅ ‖ℎ norm is dominated by the
𝐿2 (𝛺)-norm. Additionally, we achieved first-order convergence for the BP-Euler method for both P1 and P2 elements when the size
of the time step is decreased. Figs. 3(b) and 3(d) show the asymptotic behaviour with respect to time for the BP-Euler method.
To assess the computational cost of the nonlinear algorithm at each time step, we depicted in Fig. 4 the average number of
iterations per step over 1000 time steps for a sequence of meshes with decreasing mesh size. The results indicate that there is no
significant increase in the average number of iterations, which remains relatively low regardless of the mesh size.

Example 2 (Three body rotation). This example is a modified version of the three body rotation transport problem from [27]. We
used 𝜷 = (0.5 − 𝑦, 𝑥 − 0.5), 𝜀 = 10−12 and 𝜇 = 𝑓 = 0. The initial setup involves three separate bodies, as depicted in Fig. 5. Each

12
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Fig. 3. Using norm (55) for the comparison of the error of the approximated solution by the BP-Euler method with the exact solution (using mesh 1(a)).

Fig. 4. The average number of the Richardson iterations of 1000 time steps (𝑇 = 1) needed to reach convergence using P1 and Q1 elements and BP-Euler and
BP-CN methods and the meshes 1(a) and 1(b).

13
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Fig. 5. Initial data 𝑢0 for rotating body problem.

body’s position is defined by its centre at coordinates (𝑥0 , 𝑦0 ). Every body is contained within a circle of radius 𝑟0 = 0.15 cantered
at (𝑥0 , 𝑦0 ). Outside these three bodies, the initial condition is zero.
Let
√
1
𝑟(𝑥, 𝑦) = (𝑥 − 𝑥0 )2 + (𝑦 − 𝑦0 )2 . (56)
𝑟0
The centre of the slotted cylinder is in (𝑥0 , 𝑦0 ) = (0.5, 0.75) and its geometry is given by
{
1 if 𝑟(𝑥, 𝑦) ≤ 1, |𝑥 − 𝑥0 | ≥ 0.0225 or𝑦 ≥ 0.85;
𝑢(0; 𝑥, 𝑦) = (57)
0 else,
The conical body at the bottom side is described by (𝑥0 , 𝑦0 ) = (0.5, 0.25) and

𝑢(0; 𝑥, 𝑦) = 1 − 𝑟(𝑥, 𝑦). (58)

Finally, the hump at the left hand side is given by (𝑥0 , 𝑦0 ) = (0.25, 0.5) and
1
𝑢(0; 𝑥, 𝑦) =(1 + cos(𝜋 min{𝑟(𝑥, 𝑦), 1})). (59)
4
The rotation of the bodies occurs counter-clockwise. A full revolution takes 𝑡 = 2𝜋. We use 𝑃 = 130 so a regular grid consisting of
130 × 130 mesh cells for P1 , P2 , and Q1 elements.

The simulations were performed with the final time 𝑇 = 2𝜋 and the time step 𝛥𝑡 = 10−3 . Fig. 6 depicts the approximation solution
for the BP-Euler method and BP-CN method using P1 , P2 and Q1 elements. In both methods the CIP term (17) was used with the
parameter 𝛾 = 0.001. Our numerical experiments show that the optimal value of 𝜔 (relative to the number of iterations needed to
reach convergence) is approximately 0.07 when using the quadrilateral mesh, while for P1 and P2 elements, it is around 0.12. So,
we report the results using these values.
For comparison purposes, we also approximated the same problem with CIP-Euler and CIP-Crank–Nicolson (CIP-CN) (the method
that only use the CIP term i.e., the full time-space discretisation 𝜃 scheme of the method (18)) with the same value for the parameter
𝛾. To compare the numerical solution of different methods, a cross section along the line 𝑦 = 0.75 was taken of 𝑢0 , BP-Euler, BP-CN,
CIP-Euler and CIP-CN methods. The results have been shown in Fig. 7. Since we perform a full rotation, we can compare the solution
from each method with the initial condition to assess the diffusive properties of each method. As shown in Fig. 7, among all the
methods, BP-CN demonstrates the best performance regardless of the type of elements used.
Two final experiments are presented. The first aims at assessing the effect of adding CIP stabilisation to the method (26). For
this, we set 𝛾 = 0 in 𝐽 (⋅, ⋅) and the results are shown in Fig. 8 for the BP-CN method using P1 elements. In Fig. 8 we can observe the
solution 𝑢+
ℎ
, while respecting the bounds of the exact solution, exhibits spurious oscillations near the layers. This justifies the need
for CIP stabilisation in (26).
To test the performance of the method in the case when the mesh used is not Delaunay, we have approximated this example
also in the non-Delaunay mesh depicted in Fig. 1(c), using 𝑃 = 130. The same cross-sections of the approximate solutions for the
BP-Euler and BP-CN methods are depicted in Fig. 9, alongside the cross-sections for the CIP stabilised finite element method. In

14
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Fig. 6. The approximation of the solution of Example 2 for BP-Euler method and BP-CN method at 𝑇 = 6.28 (𝛾 = 0.001, 𝑃 = 130).

15
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Fig. 7. Cross sections were taken along the line 𝑦 = 0.75 of Initial data 𝑢0 , BP-Euler, BP-CN, CIP-Euler and CIP-CN methods at 𝑇 = 6.28 (𝛾 = 0.001, 𝑃 = 130).
For plotting these cross-sections, when P2 elements are used 10,000 equidistant points were chosen along the line 𝑦 = 0.75, and the values of the approximated
solution have been plotted at these points.

16
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

Fig. 8. Left: The approximation of the solution of Example 2 for BP-CN method without CIP term (𝛾 = 0) using P1 elements and mesh 1(a) (𝑃 = 130) Right:
Cross sections were taken along the line 𝑦 = 0.75 of Initial data 𝑢0 and BP-CN method without CIP term at 𝑇 = 6.28.

Fig. 9. Cross sections were taken along the line 𝑦 = 0.75 of Initial data 𝑢0 , BP-Euler, BP-CN, CIP-Euler and CIP-CN methods at 𝑇 = 6.28 (𝛾 = 0.001, 𝑃 = 130) on
the non-Delaunay mesh 1(c).

Fig. 10. The evolution of mass over time employing the BP-Euler and BP-CN schemes. These methods have been implemented with P1 and P2 elements on
Mesh 1(a), and Q1 elements on Mesh 1(b).

17
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

both cases 𝛾 = 0.001 has been used in the simulation. From the results we can observe, once again, that 𝑢+ ℎ
respects the bounds of
Assumption (A1), while the CIP solution presents noticeable over and undershoots.
Finally, to study mass conservation we use the relative mass, i.e. the ratio of the mass at time 𝑡 to the initial mass defined by
𝑀(𝑡)
𝑀𝑟 (𝑡) = ,
𝑀(0)
where 𝑀(𝑡) is the total mass at time 𝑡, and is defined as

𝑀(𝑡) = 𝑢(𝒙, 𝑡) d𝒙.

∫𝛺
The evolution of mass over time for BP-Euler and BP-CN methods is presented in Fig. 10. The plot depicts the evolution of the
relative mass. In there, we observe that, despite the fact that the scheme does not preserve mass, the mass loss/gain remains low
throughout the simulation.

6. Conclusions and outlook

In this work we have build-up on the works [17,18] and extended that framework to the time-dependent convection–diffusion
equation. The analysis of stability and error estimates has been restricted to the implicit Euler scheme, but the method itself has
also been tested in the context of the Crank–Nicolson time discretisation. For both options, the numerical experiments show a good
performance, with the solution respecting the physical bounds, while at the same time the layers in the solution do not seem to be
excessively smeared.
It is important to mention that more economical alternatives, such as linearised flux-corrected transport (FCT) methods (see,
e.g., [28]), are also available to ensure bound preservation. Nevertheless, several factors should be considered. One aspect is the
CFL condition, as most linear flux-corrected transport methods require such a condition to guarantee bound preservation, whereas
our approach does not impose this restriction. Another consideration is the applicability to higher-order elements. FCT methods
have primarily been developed for linear finite elements, and bound preservation is not guaranteed for higher-order elements, as the
necessary analysis has not yet been carried out. Additionally, improvements to the nonlinear solver itself can enhance performance.
In this paper, we used a simple Richardson type solver to highlight the simplicity of the scheme, but more efficient nonlinear solvers,
such as localised Newton methods (see, e.g., [29]), or active set methods [30], can significantly improve convergence speed. Our
preliminary results indicate that these alternatives provide much faster convergence.
Several problems remain open at this point. The extension of the stability and error analysis to higher-order time discretisation
is, at the moment, an open problem. In addition, the extension of this framework to the transport equation is also of interest. A
parallel development is the extension of this methodology to discontinuous Galerkin scheme in space, which is the topic of the
companion paper [31]. These, and other topics will be the subject of future research.

Acknowledgements

The work of AA, GRB, and TP has been partially supported by the Leverhulme Trust Research Project, United Kingdom Grant
No. RPG-2021-238. TP is also partially supported by EPRSC grants EP/W026899/2, EP/X017206/1 and EP/X030067/1.

Data availability

No data was used for the research described in the article.

References

[1] A.N. Brooks, T.J.R. Hughes, Streamline upwind/Petrov-Galerkin formulations for convection dominated flows with particular emphasis on the incompressible
Navier-Stokes equations, Comput. Methods Appl. Mech. Engrg. 32 (1–3) (1982) 199–259, [Link] FENOMECH
’’81, Part I (Stuttgart, 1981).
[2] T.J. Hughes, L.P. Franca, G.M. Hulbert, A new finite element formulation for computational fluid dynamics: VIII. The galerkin/least-squares method
for advective-diffusive equations, Comput. Methods Appl. Mech. Engrg. 73 (2) (1989) 173–189, [Link] URL
[Link]
[3] E. Burman, Consistent SUPG–method for transient transport problems: Stability and convergence, Comput. Methods Appl. Mech. Engrg. 199 (17) (2010)
1114–1123, [Link] URL [Link]
[4] P.B. Bochev, M.D. Gunzburger, J.N. Shadid, Stability of the SUPG finite element method for transient advection–diffusion problems, Comput. Methods
Appl. Mech. Engrg. 193 (23) (2004) 2301–2323, [Link] URL [Link]
S0045782504000830.
[5] J.-L. Guermond, Stabilization of Galerkin approximations of transport equations by subgrid modeling, M2AN Math. Model. Numer. Anal. 33 (6) (1999)
1293–1316, [Link]
[6] E. Burman, P. Hansbo, Edge stabilization for Galerkin approximations of convection-diffusion-reaction problems, Comput. Methods Appl. Mech. Engrg. 193
(15–16) (2004) 1437–1453, [Link]
[7] R. Becker, M. Braack, A two-level stabilization scheme for the Navier-Stokes equations, in: Numerical Mathematics and Advanced Applications, Springer,
Berlin, 2004, pp. 123–130.
[8] E. Burman, M.A. Fernández, Finite element methods with symmetric stabilization for the transient convection–diffusion–reaction equation, Comput. Methods
Appl. Mech. Engrg. 198 (33) (2009) 2508–2519, [Link] URL [Link]
S0045782509000851.

18
A. Amiri et al. Journal of Computational and Applied Mathematics 470 (2025) 116691

[9] P.G. Ciarlet, P.-A. Raviart, Maximum principle and uniform convergence for the finite element method, Comput. Methods Appl. Mech. Engrg. 2 (1973)
17–31.
[10] J. Xu, L. Zikatanov, A monotone finite element scheme for convection-diffusion equations, Math. Comp. 68 (228) (1999) 1429–1446, [Link]
10.1090/S0025-5718-99-01148-5.
[11] G.R. Barrenechea, V. John, P. Knobloch, Finite element methods respecting the discrete maximum principle for convection-diffusion equations, SIAM Rev.
66 (1) (2024) 3–88, [Link]
[12] A. Mizukami, T.J.R. Hughes, A Petrov-Galerkin finite element method for convection-dominated flows: an accurate upwinding technique for satisfying the
maximum principle, Comput. Methods Appl. Mech. Engrg. 50 (2) (1985) 181–193, [Link]
[13] E. Burman, A. Ern, Stabilized Galerkin approximation of convection-diffusion-reaction equations: discrete maximum principle and convergence, Math.
Comp. 74 (252) (2005) 1637–1652, [Link] (electronic).
[14] D. Kuzmin, Algebraic flux correction for finite element discretizations of coupled systems, in: M. Papadrakakis, E. Oñate, B. Schrefler (Eds.), Proceedings
of the Int. Conf. on Computational Methods for Coupled Problems in Science and Engineering, CIMNE, Barcelona, 2007, pp. 1–5.
[15] G.R. Barrenechea, E. Burman, F. Karakatsani, Edge-based nonlinear diffusion for finite element approximations of convection-diffusion equations and its
relation to algebraic flux-correction schemes, Numer. Math. 135 (2) (2017) 521–545.
[16] C. Lu, W. Huang, E.S. Van Vleck, The cutoff method for the numerical computation of nonnegative solutions of parabolic PDEs with application
to anisotropic diffusion and lubrication-type equations, J. Comput. Phys. 242 (2013) 24–36, [Link] URL https:
//[Link]/science/article/pii/S0021999113001307.
[17] G.R. Barrenechea, E.H. Georgoulis, T. Pryer, A. Veeser, A nodally bound-preserving finite element method, IMA J. Numer. Anal. 44 (4) (2024) 2198–2219,
[Link]
[18] A. Amiri, G.R. Barrenechea, T. Pryer, A nodally bound-preserving finite element method for reaction–convection–diffusion equations, Math. Models Methods
Appl. Sci. 34 (8) (2024) 1533–1565, [Link]
[19] A. Ern, J.-L. Guermond, Finite Elements I, Springer, 2021.
[20] H. Brezis, Functional Analysis, Sobolev Spaces and Partial Differential Equations, in: Universitext, Springer, New York, 2011, p. xiv+599.
[21] L.C. Evans, Partial Differential Equations, second ed., in: Graduate Studies in Mathematics, vol. 19, American Mathematical Society, Providence, RI, 2010.
[22] C.G. Makridakis, On the Babuška-Osborn approach to finite element analysis: 𝐿2 estimates for unstructured meshes, Numer. Math. 139 (4) (2018) 831–844,
[Link]
[23] A. Ern, J.-L. Guermond, Finite Elements II, Springer, 2021.
[24] H.-G. Ross, M. Stynes, L. Tobiska, Robust Numerical Methods for Singularly Perturbed Differential Equations, SSCM, vol. 24, Springer, 2008.
[25] M. Renardy, R.C. Rogers, An Introduction to Partial Differential Equations, second ed., in: Texts in Applied Mathematics, vol. 13, Springer-Verlag, New
York, 2004, p. xiv+434.
[26] J.G. Heywood, R. Rannacher, Finite-element approximation of the nonstationary Navier–Stokes problem. Part IV: error analysis for second-order time
discretization, SIAM J. Numer. Anal. 27 (2) (1990) 353–384.
[27] R.J. Leveque, High-resolution conservative algorithms for advection in incompressible flow, SIAM J. Numer. Anal. 33 (2) (1996) 627–665, http:
//[Link]/10.1137/0733033.
[28] D. Kuzmin, Explicit and implicit FEM-FCT algorithms with flux linearization, J. Comput. Phys. 228 (7) (2009) 2517–2534, [Link]
jcp.2008.12.011.
[29] I.K. Argyros, On Newton’s method and nondiscrete mathematical induction, Bull. Aust. Math. Soc. 38 (1) (1988) 131–140.
[30] B.S. Ashby, A. Hamdan, T. Pryer, A nodally bound-preserving finite element method for hyperbolic convection-reaction problems, 2025, arXiv:2501.11042.
[31] G.R. Barrenechea, T. Pryer, A. Trenam, A nodally bound-preserving discontinuous Galerkin method for the drift-diffusion equation, 2024, arXiv preprint
arXiv:2410.05040.

19