Computers and Mathematics with Applications 152 (2023) 143–154

Contents lists available at ScienceDirect

Computers and Mathematics with Applications

journal homepage: www.elsevier.com/locate/camwa

B-methods for the numerical solution of evolution problems with blow-up

solutions part II: Splitting methods
Mélanie Beck a , Martin J. Gander b,∗ , Felix Kwok c
a
Dawson College, Department of Mathematics, H3Z 1A4, Montreal, Canada
b
University of Geneva, Section of Mathematics, 1205, Geneva, Switzerland
c
Université Laval, Département de mathématiques et de statistique, G1V 0A6, Québec, Canada

A R T I C L E I N F O A B S T R A C T

Keywords: B-methods are numerical methods which are especially tailored to solve non-linear partial differential equation
Geometric integration that have blow up solutions. We have presented in Part I a systematic construction of B-methods based on the
Blow-up solutions variation of constants formula. Here, we use splitting methods as a second way to construct B-methods, and we
Non-linear partial differential equations
prove several special properties of such methods. We illustrate our analysis with numerical experiments.
Nonlinear systems of equations
Splitting methods

1. Introduction Very successful are moving mesh methods, see [10]. In [11] self-similar
solution techniques are employed for obtaining a scale invariant adap-
Nonlinear partial differential equations (PDEs) arise in many impor- tive numerical method. A more direct numerical time stepping approach
tant models in science and engineering, and very few of those models can be found in [31], where a numerical method using arclength ingre-
have closed form solutions. One therefore has to resort to numerical dients is constructed and analyzed, and in [49], compactification of base
methods to compute approximations. If the partial differential equation spaces is combined with the validation of Lyapunov functions. Adaptive
has further geometric properties, it is often an advantage for the numer- time stepping is also a very successful technique, where the time step
ical approximation to also have the same geometric property, which led is proportional to the inverse of the norm of the solution, see for exam-
to the research field of geometric numerical integration. Much progress ple [46]. A very recent numerical technique based on a transformation
has been made over the last two decades in this area, see for example which transforms blowup into a zero and thus allows integration right
[29,41,28,15,7] and references therein. through blowup can be found in [16].
In specific applications, the nonlinear PDE models can have solu- Considering blow-up as a geometric property is a more recent area of
tions that blow up in finite time. This is in particular the case for
research, and so far mostly ad hoc constructions have been used to ob-
combustion models [24,18,32,34], turbulent flow [39], nonlinear optics
tain numerical schemes with good blow-up properties, see for example
[37,42,43] and population dynamics [47,27]. This blow-up indicates in
[40]. In a first paper [6], we have shown how one can systematically
general that the model is losing its validity, and it is therefore important
construct B-methods using the technique of variation of the constant.
to understand the precise behavior of the model when the blow-up time
The goal of this paper is to present a second systematic way of obtain-
is approached. Studying such blow-up phenomena is necessarily done
ing B-methods, using splitting techniques.
on a case-by-case basis, see for example [36,50,25,51,26,21,2,20,17],
and the reviews [3,22]. Blow-up can even happen when first integrals
are conserved, see [9] and references therein. The analysis of blow-up 2. B-methods based on splitting
phenomena is an active field of research, and many results have been
obtained over the past two decades, see [8,44,38,45,30,13,14,48] and To fix ideas, we first show the construction of splitting B-methods
references therein. for a quasi-linear parabolic problem. The construction for a few other
The construction of numerical methods to approximate the blow-up scalar or systems of nonlinear partial differential equations can be found
time and rate of such models focuses in general on adaptive techniques. in Section 4.

* Corresponding author.
E-mail addresses: [email protected] (M. Beck), [email protected] (M.J. Gander), [email protected] (F. Kwok).

https://doi.org/10.1016/j.camwa.2023.10.013

Available online 24 October 2023

0898-1221/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

2.1. Model problem and assumptions 𝜑[1]

𝑡 of 𝑢𝑡 = 𝛿𝐹 (𝑢) (note that 𝜑𝑡 does not represent a time deriva-
tive). Indeed, the exact flow of an equation 𝑦𝑡 = 𝑓 (𝑦) is the map de-
We consider the quasi-linear parabolic partial differential equation fined by 𝜑𝑡 (𝑦0 ) = 𝑦(𝑡) if 𝑦(0) = 𝑦0 , so in this case, using (3), we have
𝜑[1]
𝑡 (𝑢𝑛 ) = 𝐺(𝑔(𝑢𝑛 ) − 𝛿𝑡), for 𝑡 < 𝑔(𝑢𝑛 )∕𝛿. Then we can choose any numeri-
𝑢𝑡 = Δ𝑢𝑚 + 𝛿𝐹 (𝑢), for (𝑥, 𝑡) ∈ Ω × (0, 𝑇 ), cal integrator Φ[2] for 𝑢𝑡 = Δ𝑢𝑚 , and by composing the exact flow and the
ℎ
𝑢 = 0, for (𝑥, 𝑡) ∈ 𝜕Ω × (0, 𝑇 ), (1)
numerical integrator, we obtain two new methods for 𝑢𝑡 = Δ𝑢𝑚 + 𝛿𝐹 (𝑢),
𝑢(𝑥, 0) = 𝑢0 (𝑥), for 𝑥 ∈ Ω,

where 𝛿 is a positive constant, Ω is a bounded domain of ℝ𝑑 and 𝑢0 Φℎ = 𝜑[1]

ℎ
◦Φ[2]
ℎ
and Φ∗ℎ = Φ[2]∗
ℎ
◦𝜑[1]
ℎ
, (4)
̄ In our analysis, we need the
is a positive continuous function on Ω.
where Φ[2]∗
ℎ
is the adjoint of Φ[2]
ℎ
(see Section II.3 in [29]). Note that the
following
two original methods Φ[2]
ℎ
and Φ[2]∗
ℎ
are consistent, that is

Assumption 1. The function 𝐹 is assumed to be positive, strictly in- Φ[2] (𝑧0 ) = 𝑧0 +ℎ𝑓 [2] (𝑧0 )+𝑂(ℎ𝑝 ) and Φ[2]∗ (𝑧0 ) = 𝑧0 +ℎ𝑓 [2] (𝑧0 )+𝑂(ℎ𝑝 ),
ℎ ℎ
creasing and strictly convex on (0, ∞), belonging to 𝐶 2 ([0, ∞)) and sat-
isfying with 𝑝 ≥ 2. Moreover, 𝜑[1]
𝑡 is the exact flow of 𝑢𝑡 = 𝛿𝐹 (𝑢), so that its
∞
Taylor expansion is
𝑑𝑠
< ∞, (2) 𝜑[1] (𝑦0 ) = 𝑦(ℎ) = 𝑦0 + ℎ𝑓 [1] (𝑦0 ) + 𝑂(ℎ2 ).
∫ 𝐹 (𝑠) ℎ
𝑏
Therefore, the resulting methods Φℎ and Φ∗ℎ are of first order. This con-
∞ 1
for 𝑏 > 0. Then the function 𝑔(𝑠) = ∫𝑠 𝐹 (𝜎)
𝑑𝜎 is continuous and strictly struction can only lead to methods of first order, however as these two
decreasing on (0, ∞). The function 𝐺 = 𝑔 −1 is continuous and
strictly integrators are adjoint, we can use them as the basis of the composition
decreasing on (0, 𝑀), where 𝑀 = lim𝑠→0 𝑔(𝑠) ≤ ∞. Note also that 𝑔 and method
𝐺 are positive with lim𝑠→∞ 𝑔(𝑠) = 0 and lim𝑠→0 𝐺(𝑠) = ∞.
Φℎ = Φ𝛼𝑠 ℎ ◦Φ∗𝛽 ℎ ◦ … ◦Φ∗𝛽 ℎ ◦Φ𝛼1 ℎ ◦Φ∗𝛽 ℎ ,
𝑠 2 1
In order to be able to construct B-methods, we need to get an explicit to construct methods of any desired order (see [29]). In particular, by
form of 𝑔 and often 𝐺. Examples of functions 𝐹 which satisfy all these choosing 𝛼1 = 𝛽1 = 1∕2 for 𝑠 = 1, we obtain a second-order symmet-
conditions are ric method Ψℎ = Φℎ∕2 ◦Φ∗ℎ∕2 . It is interesting to note that if Φℎ (not
Φ[2] ) is the forward (respectively backward) Euler method, the result-
• 𝐹 (𝑢) = 𝑒𝑢 , 𝑔(𝑢) = 𝑒−𝑢 , 𝐺(𝑢) = − ln 𝑢, ℎ
ing method Ψℎ corresponds to the midpoint (respectively trapezoidal)
• 𝐹 (𝑢) = (𝑢 + 𝛼)𝑝+1 , 𝛼 ≥ 0, 𝑝 > 0, 𝑔(𝑢) = 𝑝(𝑢+𝛼)
1
𝑝 , 𝐺(𝑢) = (𝑝𝑢)
−1∕𝑝 − 𝛼,
( 𝑢 ) rule.
𝑒
• 𝐹 (𝑢) = 𝑒𝑢 − 1, 𝑔(𝑢) = ln 𝑒𝑢 −1 , 𝐺(𝑢) = 𝑢 − ln(𝑒𝑢 − 1),
We saw that the exact flow of 𝑢𝑡 = 𝛿𝐹 (𝑢) is given by 𝜑[1] 𝑡 (𝑢𝑛 ) =
• 𝐹 (𝑢) = (𝑢 + 1)[ln(𝑢 + 1)]𝑝+1 , 𝑝 > 0, 𝑔(𝑢) = 1 −1∕𝑝
, 𝐺(𝑢) = 𝑒(𝑝𝑢) − 1,
𝑝[ln(𝑢+1)]𝑝 𝐺(𝑔(𝑢𝑛 ) − 𝛿𝑡), so we just have to choose a numerical integrator for the
𝜋
• 𝐹 (𝑢) = 𝑢2 + 1, 𝑔(𝑢) = 2
− arctan(𝑢), 𝐺(𝑢) = cot(𝑢). second part 𝑢𝑡 = Δ𝑢𝑚 . For example, even though this problem is stiff,
we start with forward Euler Φ[2]
ℎ
(𝑢𝑛 ) = 𝑢𝑛 + ℎΔ𝑢𝑚
𝑛 , whose adjoint is back-
In problem (1) the nonlinearity in 𝐹 is responsible for the finite- ward Euler Φ[2]∗ (𝑢𝑛 ) = 𝑢𝑛 + ℎΔ𝑢𝑚 . By composing these integrators with
ℎ 𝑛+1
time blow-up and becomes increasingly important as we approach
the exact flow 𝜑[1]
𝑡 , we get two B-methods. The first one is the Splitting
the blow-up time. The conditions imposed on 𝐹 allow us to write
Forward Euler B-Method (SpFE)
explicitly the solution of the nonlinear ordinary differential equation
𝑦(𝑆) 𝑆
𝑦𝑡 = 𝛿𝐹 (𝑦). Indeed we get for any 𝑆 > 0, ∫𝑦(𝑡) 𝐹𝑑𝑠 = ∫𝑡 𝛿𝑑𝑠, and then
(𝑠) Φℎ (𝑢𝑛 ) = 𝜑[1]
ℎ
◦Φ[2]
ℎ
(𝑢𝑛 ), (5)
𝑔(𝑦(𝑡)) = [𝑔(𝑦(𝑆)) + 𝛿𝑆] − 𝛿𝑡, that is
which gives the explicit scheme
𝑦(𝑡) = 𝑦(𝑡, 𝐾) = 𝐺(𝐾 − 𝛿𝑡), (3)
𝑢𝑛+1 = 𝐺(𝑔(𝑢𝑛 + ℎΔ𝑢𝑚
𝑛 ) − 𝛿ℎ), (6)
where 𝐾 is a constant, for all 𝑡 satisfying 𝐾 − 𝛿𝑡 ∈ (0, 𝑀). It is then nat-
ural to seek integrators that exploit this information. In the following and requires the condition 𝑔(𝑢𝑛 + ℎΔ𝑢𝑚 𝑛 )−𝛿ℎ ∈ (0, 𝑀). The second one is
we present a new approach to obtain semi-discretizations in time for the Splitting Forward Euler Adjoint B-Method (SpFE)∗
the semi-linear problem (1) from this exact solution. This approach al-
lows us to derive many new B-methods which are different from the Φ∗ℎ (𝑢𝑛 ) = Φ[2]∗
ℎ
◦𝜑[1]
ℎ
(𝑢𝑛 ), (7)
B-methods obtained using the variation of constants approach in [6]. which gives the implicit scheme

2.2. Construction of B-methods based on splitting 𝑢𝑛+1 = 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ) + ℎΔ𝑢𝑚

𝑛+1
, (8)
and requires the condition 𝑔(𝑢𝑛 ) − 𝛿ℎ ∈ (0, 𝑀). This scheme is studied in
As suggested in Hairer, Lubich and Wanner [29],1 one way to exploit
detail in Section 3 for the special semi-linear case, 𝑚 = 1.
the exact solution of the nonlinear part of the equation is by using split-
Instead of choosing Φ[2]ℎ
to be forward Euler in (4), we could choose
ting methods. We illustrate this construction on the quasi-linear scalar
PDE (1); the construction for a system of semi-linear PDEs is given in it to be backward Euler; then the adjoint Φ[2]∗
ℎ
is forward Euler and the
Section 4.3. resulting schemes are the Splitting Backward Euler B-Method (SpBE)
If we decompose 𝑢𝑡 = Δ𝑢𝑚 + 𝛿𝐹 (𝑢) into 𝑓 [1] (𝑢) = 𝛿𝐹 (𝑢) and 𝑓 [2] (𝑢) =
Φℎ (𝑢𝑛 ) = 𝑢𝑛+1 = 𝐺(𝑔(𝑣) − 𝛿ℎ), with 𝑣 = 𝑢𝑛 + ℎΔ𝑣𝑚 , (9)
Δ𝑢𝑚 , we can make good use of the fact that we know the exact flow
and the Splitting Backward Euler Adjoint B-Method (SpBE)

1
“It may happen that the differential equation 𝑦̇ = 𝑓 (𝑦) can be split according Φ∗ℎ (𝑢𝑛 ) = 𝑢𝑛+1 = 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ) + ℎΔ(𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ))𝑚 . (10)
to 𝑦̇ = 𝑓 [1] (𝑦) +𝑓 [2] (𝑦), such that only the flow of, say, 𝑦̇ = 𝑓 [1] (𝑦) can be computed
exactly. If 𝑓 [1] (𝑦) constitutes the dominant part of the vector field, it is natural Another possibility would be to choose Φ[2]
to be a second-order
ℎ
to search for integrators that exploit this information.”. method, like the symmetric midpoint rule (SpMid) or the trapezoidal

144
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154
( )
ℎ2
rule (SpTrap). However, the scheme becomes more complicated with- 𝜏𝐵 ∶= 𝑢1 − 𝑢(ℎ) = 𝐹 ′ (𝑢0 )Υ(𝑢0 ) − Υ′ (𝑢0 )𝐹 (𝑢0 ) + 𝐸 ′′ (0) + 𝑂(ℎ3 ), (17)
out necessarily bringing more accuracy, as the resulting scheme is only 2
first order. In order to get higher order methods, we need to compose if a first-order standard method is used, and for higher order standard
first order methods. The simplest way to obtain a second-order method methods we get
is thus to construct ( )
ℎ2
𝜏𝐵 = 𝐹 ′ (𝑢0 )Υ(𝑢0 ) − Υ′ (𝑢0 )𝐹 (𝑢0 ) + 𝑂(ℎ3 ).
2
Ψℎ = Φℎ∕2 ◦Φ∗ℎ∕2 = 𝜑[1]
ℎ∕2
◦Φ[2]
ℎ∕2
◦Φ[2]∗
ℎ∕2
◦𝜑[1]
ℎ∕2
, (11)
To construct the adjoint methods, we first use the exact scheme and
where Φ[2]
ℎ
and Φ[2]∗
ℎ
are adjoint first-order methods. then apply numerical methods on the result. In other words, starting
If we choose Φ[2]
ℎ
to be forward Euler, we obtain the Second order with the initial condition 𝑢0 , we define 𝑣0 = 𝜑(ℎ, 𝑢0 ), where 𝜑 satisfies
Splitting Forward Euler B-Method (SoSpFE) condition (14), and we compute 𝑢1 = Φ(ℎ, 𝑣0 ), where Φ is defined by
( ) ( ) (15) (to get a simpler notation, we denote the numerical method by Φ
ℎ 𝛿ℎ ℎ 𝛿ℎ
Ψℎ (𝑢𝑛 ) = 𝐺 𝑔(𝑣 + Δ𝑣𝑚 ) − , with 𝑣 − Δ𝑣𝑚 = 𝐺 𝑔(𝑢𝑛 ) − , instead of Φ∗ ). The definition of Φ implies in particular that for all 𝜉,
2 2 2 2
we have
(12)
and if Φ[2]is chosen to be backward Euler, we get the Second order Φ(0, 𝜉) = 𝜉 + 𝐸(0), Φ𝑡 (0, 𝜉) = Υ(𝜉) + 𝐸 ′ (0), Φ𝑡𝑡 (0, 𝜉) = Υ′ (𝜉)Υ(𝜉) + 𝐸 ′′ (0),
ℎ
Splitting Backward Euler B-Method (SoSpBE)
(18)
𝛿ℎ ℎ
Ψℎ (𝑢𝑛 ) = 𝐺(𝑔(𝑣) − ), with 𝑣 − Δ𝑣𝑚 and
2 2
(13)
𝛿ℎ ℎ 𝛿ℎ 𝑚
= 𝐺(𝑔(𝑢𝑛 ) − ) + Δ(𝐺(𝑔(𝑢𝑛 ) − )) . Φ𝑣 (0, 𝜉) = 1, Φ𝑣𝑣 (0, 𝜉) = 0, and Φ𝑡𝑣 (0, 𝜉) = Υ′ (𝜉). (19)
2 2 2
Similarly we can construct arbitrary high order splitting B-methods. We now expand 𝑢1 = Φ(ℎ, 𝜑(ℎ, 𝑢0 )) in a series of ℎ. The derivatives of 𝑢1
are
2.3. Truncation error analysis
𝑢′1 (ℎ) = Φ𝑡 (ℎ, 𝜑(ℎ, 𝑢0 )) + Φ𝑣 (ℎ, 𝜑(ℎ, 𝑢0 )) ⋅ 𝜑𝑡 (ℎ, 𝑢0 ),
In order to show that B-methods have the potential to be better than 𝑢′′ (ℎ) = Φ𝑡𝑡 (ℎ, 𝜑) + 2Φ𝑡𝑣 (ℎ, 𝜑)𝜑𝑡 (ℎ, 𝜑) + Φ𝑣𝑣 (ℎ, 𝜑)𝜑𝑡 (ℎ, 𝜑)2
1
standard methods, we need to compare the local truncation errors of
both types of methods. To start, we consider the problem 𝑢𝑡 = 𝐹 (𝑢) + + Φ𝑣 (ℎ, 𝜑)𝜑𝑡𝑡 (ℎ, 𝜑).
Υ(𝑢), where Υ can be a function or an operator (like the Laplacian in Noting that 𝑢1 (0) = Φ(0, 𝜑(0, 𝑢0 )) = 𝜑(0, 𝑢0 ) = 𝑢0 , we evaluate 𝑢1 , 𝑢′1 and
our example). We denote by 𝜑 the function that satisfies 𝑢′′ at ℎ = 0 and get
1

𝜑𝑡 (𝑡, 𝑣) = 𝐹 (𝜑(𝑡, 𝑣)), and 𝜑(0, 𝑣) = 𝑣, ∀𝑣. (14) 𝑢′1 (0) = Υ(𝑢0 ) + 𝐸 ′ (0) + 𝐹 (𝑢0 ),
Keeping the notation introduced earlier, we have 𝜑(𝑡, 𝑣) = 𝐺(𝑔(𝑣) − 𝑡). 𝑢′′ (0) = Υ′ (𝑢0 )Υ(𝑢0 ) + 𝐸 ′′ (0) + 2Υ′ (𝑢0 )𝐹 (𝑢0 ) + 𝐹 ′ (𝑢0 )𝐹 (𝑢0 ),
1
We also consider the numerical method Φ applied to 𝑣𝑡 = Υ(𝑣), with
𝑣(0) = 𝑣0 . If 𝑣(𝑡) solves this simplified problem, we have where we used the properties of Φ in (18) and (19) and the definition of
𝜑 given in (14). As the Taylor expansion of the exact solution 𝑢 is given
Φ(ℎ, 𝑣0 ) = 𝑣(ℎ) + 𝐸(ℎ), (15) by (16), the local truncation errors of these B-methods are, as expected,
( )
where 𝐸 represents the local truncation error of the standard method. ℎ2
𝜏𝐵 ∗ = Υ′ (𝑢0 )𝐹 (𝑢0 ) + 𝐸 ′′ (0) − 𝐹 ′ (𝑢0 )Υ(𝑢0 ) + 𝑂(ℎ3 ), (20)
We first consider the B-methods obtained by applying the numerical 2
method first and use the result in the exact scheme (like SpFE or SpBE): for first-order standard methods, and for higher-order standard methods
starting with 𝑢0 , we define 𝑣0 = 𝑢0 and we apply the numerical method we get
Φ to get 𝑣1 = 𝑣(ℎ) + 𝐸(ℎ), then we set 𝑢1 (ℎ) ∶= 𝜑(ℎ, 𝑣1 ) = 𝜑(ℎ, 𝑣(ℎ) + 𝐸(ℎ)). ( )
To expand 𝑢1 as a series of ℎ, we need to compute its derivatives, 𝑢′1 (ℎ) = ℎ2
𝜏𝐵 ∗ = Υ′ (𝑢0 )𝐹 (𝑢0 ) − 𝐹 ′ (𝑢0 )Υ(𝑢0 ) + 𝑂(ℎ3 ).
𝜑𝑡 + 𝜑𝑣 (𝑣′ (ℎ) + 𝐸 ′ (ℎ)) and 2
We now need to show that in case of finite-time blow-up, the local
𝑢′′
1
(ℎ) = 𝜑𝑡𝑡 + 2𝜑𝑡𝑣 (𝑣′ + 𝐸 ′ ) + 𝜑𝑣𝑣 (𝑣′ + 𝐸 ′ )2 + 𝜑𝑣 (𝑣′′ + 𝐸 ′′ ), truncation error of B-methods is smaller than that of the corresponding
where the derivatives of 𝜑 are evaluated at (ℎ, 𝑣(ℎ) + 𝐸(ℎ)). standard methods. We illustrate the difference in truncation errors by
From the definition of 𝜑 given in (14) (or using 𝜑(𝑡, 𝑣) = 𝐺(𝑔(𝑣) − 𝑡)), considering the forward Euler method.
we obtain 𝑢1 (0) = 𝜑(0, 𝑣(0) + 𝐸(0)) = 𝜑(0, 𝑢0 ) = 𝑢0 , 𝜑𝑡 = 𝐹 (𝜑), 𝜑𝑣 (0, 𝑣) = 1, The local truncation error of the forward Euler method applied to
𝜑𝑡𝑡 = 𝐹 ′ (𝜑)𝜑𝑡 , 𝜑𝑡𝑣 = 𝐹 ′ (𝜑)𝜑𝑣 and 𝜑𝑣𝑣 (0, 𝑣) = 0. Moreover we have 𝑣′ (ℎ) = the general equation 𝑦𝑡 = 𝑓 (𝑡, 𝑦) is given by
Υ(𝑣) and 𝑣′′ (ℎ) = Υ′ (𝑣)Υ(𝑣). Hence the derivatives of 𝑢1 evaluated at ℎ2
𝜏 ∶= 𝑦1 − 𝑦(ℎ) = − (𝑓 + 𝑓𝑦 𝑓 ) + 𝑂(ℎ3 ), (21)
ℎ = 0 are 𝑢′1 (0) = 𝐹 (𝑢0 ) + Υ(𝑢0 ) + 𝐸 ′ (0), and 2 𝑡
which means that if we apply this method to 𝑢𝑡 = 𝐹 (𝑢) + Υ(𝑢), we obtain
𝑢′′
1
(0) = 𝐹 ′ (𝑢0 )𝐹 (𝑢0 ) + 2𝐹 ′ (𝑢0 )(Υ(𝑢0 ) + 𝐸 ′ (0)) + Υ′ (𝑢0 )Υ(𝑢0 ) + 𝐸 ′′ (0).
ℎ2 ′
The values of 𝐸 ′ (0) and 𝐸 ′′ (0) depend on the standard method used, 𝜏𝑠 = − (Υ (𝑢0 ) + 𝐹 ′ (𝑢0 ))(Υ(𝑢0 ) + 𝐹 (𝑢0 )) + 𝑂(ℎ3 ). (22)
2
in particular for any consistent method, we have 𝐸 ′ (0) = 0 and if the
On the other hand, if we apply forward Euler to 𝑣𝑡 = Υ(𝑣), we obtain
method is of second or higher order, we also have 𝐸 ′′ (0) = 0. 2 [ ]
𝐸(ℎ) = − ℎ2 Υ′ (𝑣0 )Υ(𝑣0 ) + 𝑂(ℎ3 ), which gives 𝐸 ′′ (0) = −Υ′ (𝑣0 )Υ(𝑣0 ). Go-
The Taylor expansion of the exact solution 𝑢 is
ing back to (17) and (20) we obtain the truncation error of the corre-
ℎ2 ′ sponding B-methods,
𝑢(ℎ) = 𝑢0 + ℎ(Υ(𝑢0 ) + 𝐹 (𝑢0 )) + (Υ (𝑢0 ) + 𝐹 ′ (𝑢0 ))(Υ(𝑢0 ) + 𝐹 (𝑢0 )) + ⋯ ,
2 ( )
ℎ2
(16) 𝜏𝐵 = 𝐹 ′ (𝑢0 )Υ(𝑢0 ) − Υ′ (𝑢0 )𝐹 (𝑢0 ) − Υ′ (𝑢0 )Υ(𝑢0 ) + 𝑂(ℎ3 )
2
where the derivative Υ′ (𝑢0 ) can be an operator, so the local truncation ( )
ℎ2
error of the B-methods is given by 𝜏𝐵 ∗ = − 𝐹 ′ (𝑢0 )Υ(𝑢0 ) − Υ′ (𝑢0 )𝐹 (𝑢0 ) + Υ′ (𝑢0 )Υ(𝑢0 ) + 𝑂(ℎ3 ).
2

145
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

In order for the function 𝐹 to be responsible for the finite-time blow- ̄ satisfying
is the existence of a non-negative function 𝑣 ∈ 𝐶 2+𝛼 (Ω)
up, it needs to be superlinear at infinity, while the remaining part Υ(𝑢)
becomes less important as 𝑢 becomes large. We therefore expect the 𝐴𝑣 ≥ 𝑓 (𝑥, 𝑣) in Ω, 𝑣 ≥ 0 on 𝜕Ω.
term 𝐹 ′ (𝑢0 )𝐹 (𝑢0 ), which is present in 𝜏𝑠 but absent in 𝜏𝐵 and 𝜏𝐵∗ , to be Moreover, if this condition is satisfied, there exist a maximal non-negative
large relative to the other terms. As an example, let us first consider the solution 𝑢̂ ≤ 𝑣 and a minimal non-negative solution 𝑢̄ ≤ 𝑣 in the sense that,
case where Υ(𝑢) is a bounded function of 𝑢. We define 𝐹 (𝑢) ∶= 𝑒𝑢 and for every non-negative solution 𝑢 ≤ 𝑣 of (24), the inequality 𝑢̄ ≤ 𝑢 ≤ 𝑢̂ holds.
Υ(𝑢) ∶= sin(𝑢). The local truncation error can then be written as

ℎ2 ̄ and satisfies
Theorem 2. If the function 𝑢𝑛 is positive in Ω, continuous in Ω,
𝜏𝑠 = − (cos(𝑢0 ) + 𝑒𝑢0 )(sin(𝑢0 ) + 𝑒𝑢0 ) + 𝑂(ℎ3 ),
2
( )
ℎ2 2𝑢0 ‖𝑢𝑛 ‖∞ < 𝐺(𝛿ℎ), (25)
=− 𝑒 + 𝑒𝑢0 (sin(𝑢0 ) + cos(𝑢0 )) + cos(𝑢0 ) sin(𝑢0 ) + 𝑂(ℎ3 )
2 then the scheme (23) has a maximal non-negative solution 𝑢̂ ≤ 𝐶𝑛 =
for the standard method and 𝐺(𝑔(‖𝑢𝑛 ‖∞ ) − 𝛿ℎ), a minimal solution 𝑢̄ ≥ 0, and if 𝑢 is solution, then
( ) ̄ and 𝑢̄ ≤ 𝑢 ≤ 𝑢.
𝑢 ∈ 𝐶 2 (Ω) ̂
ℎ2 𝑢0
𝜏𝐵 = 𝑒 (sin(𝑢0 ) − cos(𝑢0 )) − cos(𝑢0 ) sin(𝑢0 ) + 𝑂(ℎ3 ),
2
Remark 1. We can make the bound in condition (25) on the right-hand
for the specialized SpFE method. We see that the fastest growing term
side as large as desired by choosing ℎ small enough. This condition
(𝑒𝑢0 )2 in 𝜏𝑠 does not appear in 𝜏𝐵 , while the other terms are of similar
is necessary for the scheme (23) to be well-defined; otherwise, there
order. Given the size of this term compared to the remaining terms, 𝜏𝐵
would be points 𝑥 at which 𝑢𝑛 (𝑥) > 𝐺(𝛿ℎ) ⟺ 𝑔(𝑢𝑛 (𝑥)) < 𝛿ℎ, in which
is considerably smaller than 𝜏𝑠 .
case the term 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ) would be undefined at 𝑥.
Going back to the case Υ(𝑢) = Δ𝑢, we observe numerically the same
phenomenon. Indeed, with 𝐹 (𝑢) = 3𝑒𝑢 and Υ(𝑢) = Δ𝑢, the local trunca-
Proof. The constant 𝐶𝑛 is a supersolution of the scheme, if it satisfies
tion errors are
( ) − ℎ1 𝐶𝑛 + ℎ1 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ) ≤ 0 (= 𝐴𝐶𝑛 ), that is 𝐶𝑛 ≥ 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ). Hence
ℎ2 the constant 𝐶𝑛 = 𝐺(𝑔(‖𝑢𝑛 ‖∞ ) − 𝛿ℎ), which is well-defined if condition
𝜏𝑠 = − Δ(Δ𝑢0 + 3𝑒𝑢0 ) + 3𝑒𝑢0 Δ𝑢0 + 9𝑒2𝑢0 + 𝑂(ℎ3 ),
2 (25) is satisfied and positive by definition of 𝐺 (see Assumption 1), is a
( )
ℎ2 supersolution. Moreover, since 𝑓 (𝑥, 0) = ℎ1 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ) > 0, we conclude
𝜏𝐵 = 3𝑒𝑢0 Δ𝑢0 − Δ(3𝑒𝑢0 ) − Δ(Δ𝑢0 ) + 𝑂(ℎ3 )
2 using Theorem 1 from Amann.
for the SpFE method, whereas for the (SpFE)∗ method, we have
( ) Since 𝑢𝑛 ≡ 0 is not a solution of the scheme, this result implies that
ℎ2
𝜏𝐵 ∗ = − 3𝑒𝑢0 Δ𝑢0 − Δ(3𝑒𝑢0 ) + Δ(Δ𝑢0 ) + 𝑂(ℎ3 ). there exists a non-zero nonnegative solution. Moreover the strong max-
2
imum principle applies (see for example [52]) and any nonnegative
In this case, the term 𝑒2𝑢0 of 𝜏𝑠 is also absent from 𝜏𝐵 and 𝜏𝐵∗ , but it solution is positive on Ω. Uniqueness of the positive solution can also
is not obvious that this term is much larger than the remaining terms. be obtained using the following result of Keller [35] with 𝑚 = 0 and
Some numerical experiments using Matlab show that the difference be- 𝑀 = 𝐶𝑛 .
tween 𝑒2𝑢 and the other terms is considerable and increases as 𝑢 gets
larger. Using the built-in adaptive method ode45 we computed the solu- Theorem 3 (Keller). If there exist two constants 𝑚 and 𝑀 such that for all
tion of 𝑢𝑡 = 3𝑒𝑢 + Δ𝑢 on [−1, 1] with 𝑢0 (𝑥) = cos(𝜋𝑥∕2), we then evaluated 𝑥 ∈ Ω and all 𝑢1 , 𝑢2 such that 𝑚 ≤ 𝑢1 < 𝑢2 ≤ 𝑀 , we have 𝑓 (𝑥, 𝑢1 ) ≥ 𝑓 (𝑥, 𝑢2 ),
each of the four terms that appear in 𝜏𝑠 . When 𝑡 = 0.1660 (the blow- then problem (24) has at most one solution 𝑢 ∈ 𝐶 2 satisfying 𝑚 ≤ 𝑢 ≤ 𝑀 .
up occurs approximately at t=0.1664), the norm of the different terms
is ‖Δ(Δ𝑢0 )‖2 = 342 439, ‖Δ(3𝑒𝑢0 )‖2 = 1 466 377, ‖3𝑒𝑢0 (Δ𝑢0 )‖2 = 1 542 768, Since 𝑓 (𝑥, 𝑢) defined in (23) is decreasing in 𝑢, we get the uniqueness
and ‖(3𝑒𝑢0 )2 ‖2 = 16 544 121. So removing this last term from the local of the solution, and we can show the same minimal time of existence as
error greatly improves the results in this example. for the VBE scheme in [6]:

3. Analysis of (SpFE)∗ for semi-linear parabolic problems

Theorem 4. Scheme (23) has a unique positive solution 𝑢𝑛 for 𝑛 such that
𝑡𝑛 = 𝑛ℎ < 𝑇1 , where 𝑇1 = 1𝛿 𝑔(‖𝑢0 ‖∞ ) = ∫‖𝑢 ‖ 𝛿𝐹𝑑𝑠(𝑠) .
∞
We now analyze the properties of the (SpFE)∗ scheme applied to the 0 ∞

model problem (1) for the special case of 𝑚 = 1, i.e., when the problem
is semi-linear. An explicit formula for the method is given by (8). By Since we know from Theorem 2 that
letting 𝐴 ∶= −Δ, the scheme (8) for 𝑚 = 1 can be written in the form
if ‖𝑢𝑛 ‖∞ < 𝐺(𝛿ℎ), then ‖𝑢𝑛+1 ‖∞ ≤ 𝐶𝑛 = 𝐺(𝑔(‖𝑢𝑛 ‖∞ ) − 𝛿ℎ),
1 1
𝐴𝑢𝑛+1 = 𝑓 (𝑥, 𝑢𝑛+1 ) = − 𝑢𝑛+1 + 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ). (23) the proof is exactly the same as the proof of [6, Theorem 3.11].
ℎ ℎ
Finally, we recall that the scheme (23) is linear, so that no special-
3.1. Existence and uniqueness of the solution ized non-linear solver is required to solve for 𝑢𝑛+1 .

Since the scheme (23) is linear, it has a unique solution if and only 3.2. Rate of growth
if 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ) is well defined, i.e., 𝑔(𝑢𝑛 ) ∈ (𝛿ℎ, 𝑀 + 𝛿ℎ). Since 𝑔 is de-
creasing, 𝑀 = lim𝑠→0 𝑔(𝑠) and 𝑢𝑛 > 0, we have 𝑔(𝑢𝑛 ) < 𝑀 + 𝛿ℎ, so the We now prove some growth rate estimates for the scheme (23). Note
only condition is ‖𝑢𝑛 ‖∞ < 𝐺(𝛿ℎ). We will need the following theorem that we will do this on a case-by-case basis for the functions 𝐹 (𝑢) listed
due to Amann [1]: in the introduction, since the estimate depends on the particular func-
tion at hand. We first consider the function 𝐹 (𝑢) = 𝑒𝑢 , before turning our
Theorem 1 (Amann). Let 𝑓 ∈ 𝐶 𝛼 (Ω ̄ × ℝ+ ) be given, with 𝛼 ∈ (0, 1), and attention to the case of 𝐹 (𝑢) = (𝑢 + 𝛼)𝑝+1 .
assume that 𝑓 (𝑥, 0) ≥ 0. Then a necessary and sufficient condition for the
existence of a non-negative solution 𝑢 ∈ 𝐶 2+𝛼 (Ω) of the BVP Theorem 5. Let 𝐶0 be a constant such that

𝐴𝑢 ∶= −Δ𝑢 = 𝑓 (𝑥, 𝑢) in Ω, 𝑢 = 0 on 𝜕Ω, (24) 𝐶0 ≥ 𝛿𝑒‖𝑢0 ‖∞ and 𝐴𝑢0 − 𝛿𝑒𝑢0 + 𝐶0 ≥ 0. (26)

146
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

1 1
If 𝑡𝑛+1 < 𝑇2 ∶= 𝐶0
, the function 𝑢𝑛+1 given by If 𝑡𝑛+1 < 𝑇2 ∶= 𝐶0
, the function 𝑢𝑛+1 given by

𝑢𝑛+1 + ℎ𝐴𝑢𝑛+1 = − ln(𝑒−𝑢𝑛 − 𝛿ℎ) (27) 𝑢𝑛+1 + ℎ𝐴𝑢𝑛+1 = [(𝑢𝑛 + 𝛼)−𝑝 − 𝑝𝛿ℎ]−1∕𝑝 − 𝛼, (34)

satisfies for all 𝑥 satisfies for all 𝑥

( ) ( )
𝑇2 − 𝑡𝑛 𝑇2 − 𝑡𝑛 1∕𝑝
𝑢𝑛+1 (𝑥) ≤ 𝑢𝑛 (𝑥) + ln . 𝑢𝑛+1 + 𝛼 ≤ (𝑢𝑛 + 𝛼).
𝑇2 − 𝑡𝑛+1 𝑇2 − 𝑡𝑛+1

Remark 2. Note that if 𝐴𝑢0 ≥ 0, we can take 𝐶0 = 𝛿𝑒‖𝑢0 ‖∞ , so that 𝑇2 = Proof. Throughout this proof, we will write 𝑤𝑛 = 𝑢𝑛 + 𝛼 for all 𝑛. The
1
𝐶
= 1𝛿 𝑔(‖𝑢0 ‖∞ ) = 𝑇1 , as defined in Theorem 4. recurrence can then be written as
0

𝑤𝑛+1 + ℎ𝐴𝑤𝑛+1 = (𝑤−𝑝

𝑛 − 𝑝𝛿ℎ)
−1∕𝑝
, (35)
Proof. We prove this result by induction, using a supersolution ap-
proach. First, let us prove that if 𝑡1 = ℎ < 𝑇2 , we have where we have 𝐴𝑢𝑛+1 = 𝐴𝑤𝑛+1 , since 𝐴 = −Δ annihilates the constant
( ) 𝑇 −𝑡
𝑇2 𝛼. For the initial step, we want to show that 𝑣1 = ( 𝑇2 −𝑡0 )1∕𝑝 𝑤0 satisfies
𝑢1 ≤ 𝑢0 + ln . (28) 𝑣1 + ℎ𝐴𝑣1 ≥ [𝑤−𝑝 − 𝑝𝛿ℎ]−1∕𝑝 . We calculate
2 1
𝑇2 − ℎ 0
( ) ( ) ( )1∕𝑝 ( )1∕𝑝
𝑇2
The function 𝑢0 + ln 𝑇 −ℎ = 𝑢0 − ln 1 − 𝑇ℎ = 𝑢0 − ln(1 − ℎ𝐶0 ) is a super- 𝑇2 𝑇2
2 2 𝑣1 + ℎ𝐴𝑣1 = (𝑤0 + ℎ𝐴𝑤0 ) ≥ 𝑤0
solution of (27) with 𝑛 = 0 if 𝑢0 − ln(1 − ℎ𝐶0 ) + ℎ𝐴(𝑢0 − ln(1 − ℎ𝐶0 )) ≥ 𝑇2 − ℎ 𝑇2 − ℎ
− ln(𝑒−𝑢0 − 𝛿ℎ), which simplifies to (since 𝐴𝑤0 = 𝐴𝑢0 ≥ 0)
( )1∕𝑝
𝐴𝑢0 ≥
1 1
ln(1 − ℎ𝐶0 ) − ln(1 − 𝛿ℎ𝑒𝑢0 ). (29) 𝑤𝑝0
ℎ ℎ =
1 − ℎ𝐶0
∑ 𝑘
Since ln(1 − 𝑥) = − 𝑘≥1 𝑥𝑘 for 𝑥 smaller than 1, we have
(since 𝑇2 = 1∕𝐶0 )
−1 ∑ (ℎ𝐶0 ) 1 ∑ (𝛿ℎ𝑒𝑢0 )𝑘
∞ 𝑘 ∞
1 1 = (𝑤−𝑝 − ℎ𝐶0 𝑤−𝑝 )−1∕𝑝 ≥ (𝑤−𝑝 − 𝑝𝛿ℎ)−1∕𝑝
𝛽(ℎ) ∶= ln(1 − ℎ𝐶0 ) − ln(1 − 𝛿ℎ𝑒𝑢0 ) = + 0 0 0
ℎ ℎ ℎ 𝑘 ℎ 𝑘=1 𝑘
𝑘=1 (since 𝐶0 𝑤−𝑝
0
≥ 𝑝𝛿).
∑∞
ℎ𝑘
= [(𝛿𝑒𝑢0 )𝑘+1 − 𝐶0𝑘+1 ]. 𝑣1 is therefore a supersolution, so by Theorem 1, there exists a solution
𝑘 +1 𝑇2 1∕𝑝 𝑇 −ℎ
𝑘=0 𝑤1 of (35) that satisfies 𝛼 ≤ 𝑤1 ≤ ( 𝑇 −ℎ ) 𝑤0 and 𝑤−𝑝
1
≥ 2𝑇 𝑤−𝑝 0
≥ (𝑇2 −
2 2
Since 1
= 𝐶0 ≥ 𝛿𝑒‖𝑢0 ‖ ≥ 𝛿𝑒𝑢0 , the bracket is negative and 𝛽 is decreasing ℎ)𝑝𝛿. This completes the base case.
𝑇2
in ℎ so inequality (29) holds for all ℎ ∈ (0, 𝑇2 ) if For the induction step, suppose the solution 𝑤𝑛 satisfies
( ) ( )
1 1 𝑇2 − 𝑡𝑛−1 1∕𝑝
𝐴𝑢0 ≥ lim ln(1 − ℎ𝐶0 ) − ln(1 − 𝛿ℎ𝑒𝑢0 ) = 𝛿𝑒𝑢0 − 𝐶0 , 𝛼 ≤ 𝑤𝑛 ≤ 𝑤𝑛−1 and 𝑤−𝑝
𝑛 ≥ (𝑇2 − 𝑡𝑛 )𝑝𝛿.
ℎ→0 ℎ ℎ 𝑇2 − 𝑡𝑛
which is exactly condition (26), and we get (28). To complete the in- We now need to show that there exists a solution 𝑤𝑛+1 such that
duction we assume that ( )
( ) 𝑇2 − 𝑡𝑛 1∕𝑝
𝑇2 − 𝑡𝑛−1 𝛼 ≤ 𝑤𝑛+1 ≤ 𝑤𝑛 and 𝑤−𝑝
𝑛+1
≥ (𝑇2 − 𝑡𝑛+1 )𝑝𝛿.
𝑢𝑛 ≤ 𝑢𝑛−1 + ln , (30) 𝑇2 − 𝑡𝑛+1
𝑇2 − 𝑡𝑛
( ) 𝑇2 −𝑡𝑛 1∕𝑝
𝑇 −𝑡
We start by showing that 𝑣𝑛+1 ∶= ( 𝑇 ) 𝑤𝑛 is a supersolution, i.e.,
and we show that 𝑢𝑛 + ln 𝑇 2−𝑡 𝑛 is a supersolution of (27), that is 2 −𝑡𝑛+1
2 𝑛+1 we have
( )
𝑇2 − 𝑡𝑛
𝑢𝑛 + ln + ℎ𝐴𝑢𝑛 + ln(𝑒−𝑢𝑛 − 𝛿ℎ) ≥ 0. (31) 𝑣𝑛+1 + ℎ𝐴𝑣𝑛+1 − (𝑤−𝑝
𝑛 − 𝑝𝛿ℎ)
−1∕𝑝
≥ 0. (36)
𝑇2 − 𝑡𝑛+1
It is clear from the definition that 𝑣𝑛+1 ≥ 𝛼 ≥ 0, since 𝑇2 − 𝑡𝑛 > 𝑇2 − 𝑡𝑛+1 .
First, we note that since 1
= 𝐶0 ≥ 𝛿𝑒‖𝑢0 ‖ , we have 𝑇2 ≤ 1
, and 𝑢0 ≤ 𝑇 −𝑡
𝑇2 𝛿𝑒‖𝑢0 ‖ Substituting into (36) gives ( 𝑇 2−𝑡 𝑛 )1∕𝑝 (𝑤𝑛 +ℎ𝐴𝑤𝑛 ) −(𝑤−𝑝
𝑛 −𝑝𝛿ℎ)
−1∕𝑝 ≥ 0.
2 𝑛+1
‖𝑢0 ‖ ≤ ln( 𝛿𝑇1 ), and by induction Since 𝑤𝑛 satisfies 𝑤𝑛 + ℎ𝐴𝑤𝑛 = (𝑤−𝑝 − 𝑝𝛿ℎ)−1∕𝑝 , this criterion is equiva-
2 𝑛−1
( ) 𝑇2 −𝑡𝑛 1∕𝑝
𝑇2 − 𝑡𝑛−1 lent to ( 𝑇 ) (𝑤−𝑝
𝑛−1
− 𝑝𝛿ℎ)−1∕𝑝 − (𝑤−𝑝
𝑛 − 𝑝𝛿ℎ)
−1∕𝑝 ≥ 0. In other words,
2 −𝑡𝑛+1
𝑢𝑛 ≤ 𝑢𝑛−1 + ln 𝑇2 −𝑡𝑛 1∕𝑝
𝑇2 − 𝑡𝑛 we need to show that ( 𝑇 −𝑡 ) (𝑤𝑛−1 − 𝑝𝛿ℎ)−1∕𝑝 ≥ (𝑤−𝑝
−𝑝
𝑛 − 𝑝𝛿ℎ)
−1∕𝑝 ,
( ) ( ) ( ) 2 𝑛+1
1 𝑇2 − 𝑡𝑛−1 1 which is equivalent to showing that
≤ ln + ln = ln . (32) ( )
𝛿(𝑇2 − 𝑡𝑛−1 ) 𝑇2 − 𝑡𝑛 𝛿(𝑇2 − 𝑡𝑛 ) 𝑇2 − 𝑡𝑛
𝑤−𝑝
𝑛−1
− 𝑝𝛿ℎ ≤ 𝑛 − 𝑝𝛿ℎ).
(𝑤−𝑝 (37)
By definition of 𝑢𝑛 , we have 𝑢𝑛 + ℎ𝐴𝑢𝑛 = − ln(𝑒−𝑢𝑛−1 − 𝛿ℎ), and from the 𝑇2 − 𝑡𝑛+1
𝑇 −𝑡
induction hypothesis (30), we obtain − ln(𝑒−𝑢𝑛−1 −𝛿ℎ) > − ln[𝑒−𝑢𝑛 ( 𝑇2 −𝑡𝑛−1 ) − To prove the above inequality, we use the induction hypothesis: we
2 𝑛
𝑇 −𝑡
𝛿ℎ], so that inequality (31) is satisfied if have 𝑤−𝑝 ≤ ( 𝑇2 −𝑡𝑛−1 )𝑤−𝑝
𝑛 which implies
𝑛−1
( ) [ ( ) ] 2 𝑛
𝑇2 − 𝑡𝑛 𝑇2 − 𝑡𝑛−1 ( )
ln − ln 𝑒−𝑢𝑛 − 𝛿ℎ + ln(𝑒−𝑢𝑛 − 𝛿ℎ) ≥ 0, 𝑇2 − 𝑡𝑛−1
𝑇2 − 𝑡𝑛+1 𝑇2 − 𝑡𝑛 𝑤−𝑝
𝑛−1
− 𝑝𝛿ℎ ≤ 𝑛 − 𝑝𝛿ℎ
𝑤−𝑝
𝑇2 − 𝑡𝑛
which simplifies to (𝑇2 − 𝑡𝑛 )𝛿 ≤ 𝑒−𝑢𝑛 , which is exactly (32). ( )[ ]
𝑇2 − 𝑡𝑛 (𝑇2 − 𝑡𝑛−1 )(𝑇2 − 𝑡𝑛+1 ) (𝑇2 − 𝑡𝑛+1 )𝑝𝛿ℎ
= 𝑤−𝑝
𝑛 − .
𝑇2 − 𝑡𝑛+1 (𝑇2 − 𝑡𝑛 )2 𝑇2 − 𝑡𝑛
Next, we consider the (SpFE)∗ scheme for 𝐹 (𝑢) = (𝑢 + 𝛼)𝑝+1 , 𝑝 > 0.
Therefore, the inequality (37) is true if
(𝑇2 − 𝑡𝑛−1 )(𝑇2 − 𝑡𝑛+1 ) (𝑇2 − 𝑡𝑛+1 )𝑝𝛿ℎ
Theorem 6. Let 𝑝 > 0. Suppose there exists a constant 𝐶0 that satisfies 𝑤−𝑝 ≤ 𝑤−𝑝
𝑛 − 𝑇2 − 𝑡𝑛 𝑛 − 𝑝𝛿ℎ,
(𝑇2 − 𝑡𝑛 )2
𝐶0 ≥ 𝑝𝛿(‖𝑢0 ‖∞ + 𝛼)𝑝 and 𝐴𝑢0 ≥ 0. (33) or

147
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154
( ) ( )
𝑇2 − 𝑡𝑛+1 (𝑇2 − 𝑡𝑛−1 )(𝑇2 − 𝑡𝑛+1 )
𝑝𝛿ℎ 1 − ≤ 𝑤−𝑝
𝑛 1 − , 3.3.1. Outline of the proof
𝑇2 − 𝑡𝑛 (𝑇2 − 𝑡𝑛 )2 We need to show that there exists 𝑛∗ < 𝑇 ∗ ∕ℎ such that ‖𝑢𝑛∗ ‖∞ > 𝐾,
which simplifies to (𝑇2 − 𝑡𝑛 )𝑝𝛿 ≤ 𝑤−𝑝 𝑛 , which we know is true by the
where 𝐾 is a fixed large constant. Following the eigenfunction methods,
induction hypothesis. Thus, we have shown that 𝑣𝑛+1 is a supersolu- we introduce the sequence (𝑎𝑛 ), defined by
tion; by Theorem 1, a solution 𝑤𝑛+1 exists and satisfies 𝛼 ≤ 𝑤𝑛+1 ≤
𝑇 −𝑡
( 𝑇 2−𝑡 𝑛 )1∕𝑝 𝑤𝑛 , so that 𝑤−𝑝
𝑇 −𝑡
≥ ( 𝑇2 −𝑡𝑛+1 )𝑤−𝑝 𝑎𝑛 = 𝜑 𝑢𝑛 𝑑𝑥, (40)
𝑛+1 𝑛 ≥ (𝑇2 − 𝑡𝑛+1 )𝑝𝛿, which com- ∫
2 𝑛+1 2 𝑛
Ω
pletes the induction step.
where 𝜑 is the eigenfunction corresponding to the first eigenvalue 𝜆 of
−Δ𝜑 = 𝜆𝜑, 𝜑 = 0 on the boundary, with 𝜆 > 0, 𝜑 ≥ 0 and ∫Ω 𝜑 𝑑𝑥 = 1 (we
3.3. Numerical blow-up
can assume 𝜑 ≥ 0 since by Courant’s theorem, the eigenfunction 𝜑 does
not change sign in Ω). Our approach consists of finding 𝑛∗ such that
In this section, we want to prove that for values of 𝛿 large 𝑎𝑛∗ > 𝐾. Indeed we have
enough, the (SpFE)∗ method will blow up before a certain time
∞ 𝑑𝑠
𝑇 ∗ ≤ ∫0 𝛿𝐹 (𝑠)−𝜆𝑠 < ∞, with 𝜆 being the smallest positive eigenvalue 𝑎𝑛 ≤ 𝜑 ‖𝑢𝑛 ‖∞ 𝑑𝑥 = ‖𝑢𝑛 ‖∞ 𝜑 𝑑𝑥 = ‖𝑢𝑛 ‖∞ .
∫ ∫
of −Δ. The existence of such a blow-up time in the continuous case Ω Ω
has been shown by Kaplan in [33]. Since we already proved that
We divide our proof into the following steps: 1. We prove that (𝑎𝑛 ) is
𝑇 ∗ > 𝑇1 = 1𝛿 (𝑔(‖𝑢0 ‖∞ ), proving this result leads to exactly the same
increasing. 2. We define 𝑎(𝑡), solution of 𝑎′ (𝑡) = 𝛿𝐹 (𝑎(𝑡)) − 𝜆𝑎(𝑡), 𝑎(0) =
bounds as Kaplan for the discrete case. ∞ 𝑑𝑠
𝑎∗ ∈ (0, 𝑎0 ), which blows up in finite time at 𝑇 = ∫𝑎∗ 𝛿𝐹 (𝑠)−𝜆𝑠 if 𝛿 satisfies
To do so, we first need to define what we mean by numerical blow-
condition (38). Defining 𝐷𝑛 = 𝑎𝑛 − 𝑎(𝑛ℎ), we need to bound 𝐷𝑛 from
up time. Suppose we use a numerical method of fixed time step size
below in order to prove that for ℎ small enough, 𝐷𝑛 is positive for all 𝑛
ℎ to integrate the model problem (1). We define the numerical blow-
for which 𝑎𝑛 and 𝑎(𝑡𝑛 ) are well-defined.
up time 𝑇ℎ∗ to be the smallest multiple of ℎ such that the numerical
solution ceases to exist. To estimate 𝑇ℎ∗ , we adapt the approach used by
Kaplan for the continuous problem to our semi-discretization: we show 3.3.2. Growth of the sequence (𝑎𝑛 )
that there exists a finite time 𝑇 ∗ such that for all 𝐾 > 0 and ℎ small To prove that (𝑎𝑛 ) is increasing, we need the following lemma.
enough, there exists 𝑛 < 𝑇 ∗ ∕ℎ such that ‖𝑢𝑛 ‖∞ > 𝐾, so that 𝑇ℎ∗ ≤ 𝑇 ∗ for
all ℎ small enough.2 We now state our main result. Lemma 1. As long as 𝑢𝑛 satisfies ‖𝑢𝑛 ‖∞ < 𝐺(𝛿ℎ), the sequence (𝑎𝑛 ) defined
1
in (40) satisfies 𝑎𝑛+1 ≥ 1+ℎ𝜆 𝐺(𝑔(𝑎𝑛 ) − 𝛿ℎ). The condition is satisfied in par-
𝑔(𝐾)−𝜀
Theorem 7. Suppose that 𝛿 satisfies ticular if ℎ < 𝛿
and ‖𝑢𝑛 ‖∞ ≤ 𝐾.

𝛿𝐹 (𝑢) − 𝜆𝑢 > 0, ∀ 𝑢 ≥ 0, (38) Proof. Since ‖𝑢𝑛 ‖∞ < 𝐺(𝛿ℎ), scheme (39) is well-defined. We multiply
each side by 𝜑 and integrate over Ω to get ∫Ω 𝜑 𝑢𝑛+1 − ℎ𝜑Δ𝑢𝑛+1 𝑑𝑥 =
where 𝜆 is the first eigenvalue of −Δ𝜑 = 𝜆𝜑, 𝜑 = 0 on the boundary. We fix
∫Ω 𝜑 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ)𝑑𝑥. Using the fact that 𝑢𝑛 and 𝜑 vanish on the bound-
some large positive constant 𝐾 and choose 𝜀 ∈ (0, 𝑔(𝐾)). Then there exists
ary, the left-hand side can be rewritten as 𝑎𝑛+1 − ℎ ∫Ω 𝑢𝑛+1 Δ𝜑 𝑑𝑥 =
ℎ∗ > 0 such that for all ℎ < min(ℎ∗ , 𝑔(𝐾)−𝜀
𝛿
), the numerical scheme 1
(1 + ℎ𝜆)𝑎𝑛+1 , and we obtain 𝑎𝑛+1 = 1+ℎ𝜆 ∫Ω 𝜑 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ)𝑑𝑥. We now
𝑢𝑛+1 + ℎ𝐴𝑢𝑛+1 = 𝐺(𝑔(𝑢𝑛 ) − 𝛿ℎ), (39) prove that the function 𝑓 (𝑥) ∶= 𝐺(𝑔(𝑥) − 𝛿ℎ) is convex for 𝑥 ≥ 0. We
have
∞ 𝑑𝑠
has a numerical blow-up time 𝑇∗ ≤ ∫0 𝛿𝐹 (𝑠)−𝜆𝑠 , in the sense that there exists
−1 1
𝑛∗ < 𝑇∗
such that ‖𝑢𝑛∗ ‖∞ > 𝐾. 𝑓 ′ (𝑥) = 𝐺′ (𝑔(𝑥)−𝛿ℎ)𝑔 ′ (𝑥) = −𝐹 (𝐺(𝑔(𝑥)−𝛿ℎ)) = 𝐹 (𝐺(𝑔(𝑥)−𝛿ℎ)),
ℎ 𝐹 (𝑥) 𝐹 (𝑥)
−1
since 𝐺′ (𝑠) = −𝐹 (𝐺(𝑠)) and 𝑔 ′ (𝑠) = 𝐹 (𝑠)
, and then
Note that the proof presented in this section is constructive so that
one can compute an explicit bound ℎ∗ . We suppose thereafter that 𝐾 1 [ ′
𝑓 ′′ (𝑥) = 𝐹 (𝐺(𝑔(𝑥) − 𝛿ℎ))𝐺′ (𝑔(𝑥) − 𝛿ℎ)𝑔 ′ (𝑥)𝐹 (𝑥)
and 𝜀 are fixed. 𝐹 (𝑥)2
]
− 𝐹 ′ (𝑥)𝐹 (𝐺(𝑔(𝑥) − 𝛿ℎ))
Remark 3. The assumption ℎ < 𝑔(𝐾)−𝜀 𝛿
implies that 𝐾 < 𝐺(𝛿ℎ + 𝜀) < 1 [ ′ ]
𝐺(𝛿ℎ) so that as long as ‖𝑢𝑛 ‖∞ ≤ 𝐾, condition (25) is satisfied and = 𝐹 (𝐺(𝑔(𝑥) − 𝛿ℎ))𝐹 (𝐺(𝑔(𝑥) − 𝛿ℎ)) − 𝐹 ′ (𝑥)𝐹 (𝐺(𝑔(𝑥) − 𝛿ℎ))
𝐹 (𝑥)2
scheme (39) has a unique positive solution.
𝐹 (𝐺(𝑔(𝑥) − 𝛿ℎ)) ( ′ )
= 𝐹 (𝐺(𝑔(𝑥) − 𝛿ℎ)) − 𝐹 ′ (𝑥) ,
𝐹 (𝑥)2
Remark 4. Condition (38) imposed on 𝛿 is identical to the one given
which is positive since 𝐹 being strictly convex implies that 𝐹 ′ is in-
by Kaplan in [33]. It cannot be satisfied at 𝑢 = 0 if 𝐹 (0) = 0, however, if
creasing and we have 𝐺(𝑔(𝑥) − 𝛿ℎ) ≥ 𝑥. Hence 𝑓 is convex and we apply
𝐹 (0) > 0, since 𝐹 satisfies (2), we have lim𝑢→0 𝐹 𝑢(𝑢) = 0 and lim𝑢→∞ 𝐹 𝑢(𝑢) =
Jensen’s inequality to get
0, and condition (38) is satisfied for all 𝛿 large enough. For example, if ( )
we consider 𝐹 (𝑢) = 𝑒𝑢 , condition (38) becomes 𝛿 > 𝜆𝑢 𝑒𝑢
, for all 𝑢 ≥ 0, that 𝜑(𝑥)𝑓 (𝑢𝑛 (𝑥)) 𝑑𝑥 ≥ 𝑓 𝜑(𝑥)𝑢𝑛 (𝑥) 𝑑𝑥 = 𝑓 (𝑎𝑛 ),
∫ ∫
is 𝛿 > 𝜆𝑒 . If we consider 𝐹 (𝑢) = (𝑢 + 𝛼)𝑝 , with 𝛼 > 0, since the derivative
(𝑢+𝛼)𝑝 −𝑝(𝑢+𝛼)𝑝−1 𝑢 which completes the proof.
of the function 𝛽(𝑢) ∶= 𝑢∕(𝑢 + 𝛼)𝑝 satisfies 𝛽 ′ (𝑢) = (𝑢+𝛼)2𝑝
>0⟺
𝛼 𝛼 𝛼
𝑢 < 𝑝−1 , we have 𝛽( 𝑝−1 ) = (𝑝−1)(𝛼𝑝)𝑝
, and condition (38) becomes 𝛿 >
𝜆𝛼
Lemma 2. If 𝛿 satisfies condition (38), the sequence (𝑎𝑛 ) defined in (40) is
(𝑝−1)(𝛼𝑝)𝑝
. increasing as long as 𝑢𝑛 satisfies ‖𝑢𝑛 ‖∞ < 𝐺(𝛿ℎ) (this is satisfied in particular
if ℎ < 𝑔(𝐾)−𝜀
𝛿
and ‖𝑢𝑛 ‖∞ ≤ 𝐾).
2
While most of our previous results were following Le Roux’s approach in
[40], we could not use the same method as hers to prove this result. Indeed, Proof. To prove this result, we show that for all 𝑥 ∈ (0, 𝐺(𝛿ℎ)), we have
a key element of Le Roux’s approach is the use of specific functionals, and no
1
equivalent functionals could be found for this scheme. 𝐺(𝑔(𝑥) − 𝛿ℎ) > 𝑥, (41)
1 + ℎ𝜆

148
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

that is 𝑔(𝑥) − 𝑔((1 + ℎ𝜆)𝑥) < 𝛿ℎ. Since 𝑔 is continuously differentiable, 𝜂(0) = 𝑎𝑛 − 𝑎(𝑡𝑛 ) = 𝐷𝑛 ,
we can apply the Mean Value Theorem on the interval (𝑥, (1 + ℎ𝜆)𝑥), so
there exists 𝜉 ∈ (𝑥, (1 + ℎ𝜆)𝑥), such that 𝑔(𝑥) − 𝑔((1 + ℎ𝜆)𝑥) = 𝑔 ′ (𝜉)(𝑥 − (1 + −𝜆𝐺(𝑔(𝑎𝑛 ) − 𝛿ℎ) 𝐹 (𝐺(𝑔(𝑎𝑛 ) − 𝛿ℎ))
𝜂(ℎ) = +𝛿 − 𝑎′ (𝑡𝑛 + ℎ),
ℎ𝜆)𝑥) = −𝑔 ′ (𝜉)ℎ𝜆𝑥, which becomes 𝑔(𝑥) − 𝑔((1 + ℎ𝜆)𝑥) = 𝐹 1(𝜉) ℎ𝜆𝑥. So we (1 + ℎ𝜆)2 1 + ℎ𝜆
1 𝜆𝑥
need 𝐹 (𝜉)
ℎ𝜆𝑥 < 𝛿ℎ, i.e. 𝐹 (𝜉) > 𝛿
, ∀ 𝜉 ∈ (𝑥, (1 + ℎ𝜆)𝑥). Since 𝐹 is in- 𝜂 ′ (0) = 𝛿𝐹 (𝑎𝑛 ) − 𝜆𝑎𝑛 − (𝛿𝐹 (𝑎(𝑡)) − 𝜆𝑎(𝑡)).
𝜆𝑥
creasing and 𝛿 satisfies condition (38), we have 𝐹 (𝜉) > 𝐹 (𝑥) > Hence 𝛿
.
Thus, we have
inequality (41) holds for all 𝑥 ∈ (0, 𝐺(𝛿ℎ)) and Lemma 1 completes the
proof. ℎ2 ′′
𝐷𝑛+1 ≥ 𝐷𝑛 + ℎ(𝜓(𝑎𝑛 ) − 𝜓(𝑎(𝑡𝑛 ))) + 𝜂 (𝜉) (42)
2
3.3.3. Definition of 𝑎(𝑡) and 𝐷𝑛 for some 𝜉 ∈ (0, ℎ), with 𝜓(𝑥) = 𝛿𝐹 (𝑥) − 𝜆𝑥.
From now on, we assume that condition (38) is satisfied and ℎ < Since 𝜂(ℎ) is twice continuously differentiable for all 0 ≤ ℎ ≤ 𝑔(𝐾)−𝜖 ,
𝑔(𝐾)−𝜀 𝛿
and ‖𝑢𝑛 ‖∞ ≤ 𝐾. This implies that 𝑢𝑛+1 is well-defined, thus so are
𝛿 𝜂 (ℎ) is continuous on the same interval, so there exists a (possibly
′′
𝑎𝑛+1 and 𝐷𝑛+1 defined below.
negative) constant 𝐶2 (which depends on 𝛿, 𝐾 and 𝜖, but not on ℎ)
𝑎 −𝑎 such that 𝜂 ′′ (ℎ) ≥ 𝐶2 for all 0 < ℎ < 𝑔(𝐾)−𝜖 . We are now able to prove
Definition of 𝑎(𝑡) From Lemma 1, we have 𝑛+1ℎ 𝑛 ≥ ℎ1 ( 1+ℎ𝜆 1
𝐺(𝑔(𝑎𝑛 ) − 𝛿
Theorem 7.
𝛿ℎ) − 𝑎𝑛 ), hence we will compare (𝑎𝑛 ) with (𝑎(𝑡𝑛 )) where 𝑡𝑛 = 𝑛ℎ and 𝑎(𝑡)
is the solution of
( ) 3.3.4. Proof of Theorem 7
1 1
𝑎′ (𝑡) = lim 𝐺(𝑔(𝑎(𝑡)) − 𝛿ℎ) − 𝑎(𝑡) , 𝑎(0) = 𝑎∗ ,
ℎ→0 ℎ 1 + ℎ𝜆 We suppose that ‖𝑢𝑛 ‖∞ ≤ 𝐾 and 𝐷𝑛 > 0. We now show that 𝐷𝑛+1 > 0.
∞ 𝑑𝑠
where 𝑎∗ can be any fixed number in [0, 𝑎0 ). This limit simplifies to Indeed, since 𝑎(𝑡) blows up at time 𝑇𝑎∗ with 𝑇𝑎∗ ≤ 𝑇0 = ∫0 𝛿𝐹 (𝑠)−𝜆𝑠 , there
( ) exists 𝑛̃ < 𝑇𝑎∗ ∕ℎ, such that 𝑎(𝑡𝑛̃ ) ≤ 𝐾 and either 𝑡𝑛+1
̃ ≥ 𝑇 𝑎 ∗ or 𝑎(𝑡 ̃ )>
𝑛+1
1 1
lim 𝐺(𝑔(𝑎) − 𝛿ℎ) − 𝑎 𝐾. The first case implies that ‖𝑢𝑛 ‖∞ > 𝐾 for some 𝑛 ≤ 𝑛, ̃ and in the
ℎ→0 ℎ 1 + ℎ𝜆
[( ) ] second case, by the positivity of 𝐷𝑛+1 , we have ‖𝑢𝑛+1 ̃ ‖∞ > 𝑎(𝑡𝑛+1̃ )>𝐾
1 1
= lim − 1 𝐺(𝑔(𝑎) − 𝛿ℎ) + 𝐺(𝑔(𝑎) − 𝛿ℎ) − 𝐺(𝑔(𝑎)) with 𝑡𝑛+1
̃ < 𝑇𝑎∗ . Hence there exists 𝑛∗ < 𝑇0 ∕ℎ such that ‖𝑢𝑛∗ ‖∞ > 𝐾.
ℎ→0 ℎ 1 + ℎ𝜆
[( ) ] We assume that 𝐷𝑛 > 0 and we go back to (42) to write
1 −ℎ𝜆 𝐺(𝑔(𝑎) − 𝛿ℎ) − 𝐺(𝑔(𝑎))
= lim 𝐺(𝑔(𝑎) − 𝛿ℎ) − 𝛿
ℎ→0 ℎ 1 + ℎ𝜆 −𝛿
[( ) ] ℎ2
−𝜆 𝐷𝑛+1 ≥ 𝐷𝑛 + ℎ[𝜓(𝑎𝑛 ) − 𝜓(𝑎(𝑡𝑛 ))] + 𝜂(𝜉)
= lim 𝐺(𝑔(𝑎) − 𝛿ℎ) − 𝛿 𝐺 (𝑔(𝑎))
′ 2
ℎ→0 1 + ℎ𝜆
ℎ2
= −𝜆𝐺(𝑔(𝑎)) + 𝛿 𝐹 (𝐺(𝑔(𝑎))) = 𝛿𝐹 (𝑎) − 𝜆𝑎. ≥ 𝐷𝑛 + ℎ[𝜓(𝑎(𝑡𝑛 ) + 𝐷𝑛 ) − 𝜓(𝑎(𝑡𝑛 ))] + 𝐶
2 2
So 𝑎(𝑡) is the solution of ℎ2
≥ 𝐷𝑛 + ℎ𝐷𝑛 𝜓 ′ (𝜁 ) + 𝐶 ,
2 2
𝑎′ (𝑡) = 𝛿𝐹 (𝑎(𝑡)) − 𝜆𝑎(𝑡), 𝑎(0) = 𝑎∗ < 𝑎0 . with 𝜁 ∈ (𝑎(𝑡𝑛 ), 𝑎(𝑡𝑛 ) + 𝐷𝑛 ), by the Mean Value Theorem. The derivative
By integrating this equation, we note that 𝑎(𝑡) is defined on [0, 𝑇𝑎∗ ), 𝜓 ′ (𝑥) = 𝛿𝐹 ′ (𝑥) − 𝜆 is increasing and 𝜁 > 𝑎(𝑡𝑛 ) ≥ 𝑎(0) = 𝑎∗ so we get
∞ 1
where 𝑇𝑎∗ = ∫𝑎∗ 𝛿𝐹 (𝑠)−𝜆𝑠 𝑑𝑠 < ∞, so that 𝑎(𝑡) blows up at finite time 𝑇𝑎∗ .
Our goal is to show that 𝑎𝑛 is larger than 𝑎(𝑡𝑛 ). ℎ2
𝐷𝑛+1 ≥ 𝐷𝑛 (1 + ℎ𝜓 ′ (𝑎∗ )) + 𝐶 . (43)
2 2
∑
By induction, we obtain 𝐷𝑛+1 ≥ (1 + ℎ𝜓 ′ (𝑎∗ ))𝑛+1 𝐷0 + ℎ2 𝐶2 𝑛𝑘=0 (1 +
2
Definition of 𝐷𝑛 For all 𝑛 such that 𝑎𝑛 and 𝑎(𝑡𝑛 ) are well-defined, we
define 𝐷𝑛 ∶= 𝑎𝑛 − 𝑎(𝑡𝑛 ). To prove Theorem 7, we will prove by induction ℎ𝜓 ′ (𝑎∗ ))𝑘 . We assume that 1 + ℎ𝜓 ′ (𝑎∗ ) > 0, so if 𝜓 ′ (𝑎∗ ) < 0, we need ℎ to
that there exists ℎ∗ such that ∀ ℎ ≤ ℎ∗ , ∀𝑛 such that ‖𝑢𝑛 ‖∞ ≤ 𝐾, we have be smaller than 1∕(−𝜓 ′ (𝑎∗ )), that is: if 𝐹 ′ (𝑎∗ ) < 𝜆𝛿 , then ℎ < 𝜆−𝛿𝐹1′ (𝑎∗ ) . If
𝐷𝑛+1 > 0. The initial condition 𝑎∗ was chosen such that 𝐷0 is positive, 𝐶2 is positive, the positivity of 𝐷𝑛+1 follows from (43). We now study
so assuming that 𝐷𝑛 is positive, we prove that 𝐷𝑛+1 is also positive.
the case 𝐶2 < 0. We obtain different bounds on ℎ depending on the sign
First, we need to verify that 𝑎(𝑡𝑛+1 ) exists so that 𝐷𝑛+1 is well-defined
of 𝜓 ′ (𝑎∗ ):
and 𝑡𝑛+1 < 𝑇𝑎∗ .

• if 𝜓 ′ (𝑎∗ ) = 0, we get 𝐷𝑛+1 ≥ 𝐷0 + (𝑛 + 1) ℎ2 𝐶2 , so that since 𝐶2 < 0 and

2

Lemma 3. If 𝐷𝑛 > 0, the function 𝑎(𝑡𝑛 + 𝜉), with 𝜉 ∈ [0, ℎ], is bounded 2𝐷0
above by 𝑎(𝑡𝑛 + 𝜉) < 𝐺(𝜀), where 𝜀 is a fixed number belonging to (0, 𝑔(𝐾)) 𝑡𝑛+1 < 𝑇𝑎∗ , 𝐷𝑛+1 is positive if ℎ < .
(−𝐶2 )𝑇𝑎∗
( )
(see Theorem 7). ℎ2 (1+ℎ𝜓 ′ (𝑎∗ ))𝑛+1 −1
• if 𝜓 ′ (𝑎∗ ) > 0, we get 𝐷𝑛+1 ≥ (1 +ℎ𝜓 ′ (𝑎∗ ))𝑛+1 𝐷 0 + 2 𝐶2 ℎ𝜓 ′ (𝑎∗ )
,
ℎ2 ′
(1+ℎ𝜓 (𝑎 )) ∗ 𝑛+1
Proof. We introduce for 𝑡 ≥ 𝑡𝑛 the function 𝑏(𝑡), solution of so we need 𝐶
2 2
≥ − (1+ℎ𝜓 ′ (𝑎∗ ))𝑛+1 −1 ℎ𝜓 (𝑎 )𝐷0 . The underbraced term
′ ∗

⏟⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏟
2𝜓 ′ (𝑎∗ )𝐷0
𝑏′ (𝑡) = 𝛿𝐹 (𝑏(𝑡)) > 𝛿𝐹 (𝑏(𝑡)) − 𝜆𝑏(𝑡), 𝑏(𝑡𝑛 ) = 𝑎(𝑡𝑛 ). is greater than 1 since 𝜓 ′ (𝑎∗ ) > 0, so we need ℎ < .
(−𝐶2 )
This function can be written explicitly, 𝑏(𝑡) = 𝐺(𝑔(𝑎(𝑡𝑛 )) + 𝛿𝑡𝑛 − 𝛿𝑡), and • if 𝜓 ′ (𝑎∗ ) < 0 we also get
we have 𝑎(𝑡) ≤ 𝑏(𝑡), ∀𝑡 ≥ 𝑡𝑛 . Moreover since 𝛿 satisfies condition (38), 𝑎(𝑡) ( )
is increasing and we have 𝑎(𝑡𝑛 + 𝜉) ≤ 𝑎(𝑡𝑛 + ℎ) ≤ 𝑏(𝑡𝑛+1 ) = 𝐺(𝑔(𝑎(𝑡𝑛 )) − 𝛿ℎ), ℎ2 (1 + ℎ𝜓 ′ (𝑎∗ ))𝑛+1 − 1
𝐷𝑛+1 ≥ (1 + ℎ𝜓 ′ (𝑎∗ ))𝑛+1 𝐷0 + 𝐶 ,
and since 𝑎(𝑡𝑛 ) < 𝑎𝑛 ≤ 𝐾 and ℎ < 𝑔(𝐾)−𝜀
𝛿
, we get as required 𝑎(𝑡𝑛 + 𝜉) ≤ 2 2 ℎ𝜓 ′ (𝑎∗ )
𝐺(𝑔(𝑎(𝑡𝑛 )) − 𝛿ℎ) ≤ 𝐺(𝑔(𝐾) − 𝛿ℎ) < 𝐺(𝜀). ℎ 𝐶2
so we need (1 + ℎ𝜓 ′ (𝑎∗ ))𝑛+1 𝐷0 + 2 𝜓 ′ (𝑎∗ )
[(1 + ℎ𝜓 ′ (𝑎∗ ))𝑛+1 − 1] > 0,
ℎ 2𝐷0
Hence 𝐷𝑛+1 is well-defined and we first bound it using Lemma 1, which simplifies to (1+ℎ𝜓 ′ (𝑎∗ ))𝑛+1 < (−𝐶 ) (−𝜓 ′ (𝑎∗ )) + ℎ. Since ℎ > 0,
2
it is enough to satisfy (1+ℎ𝜓 ′ℎ(𝑎∗ ))𝑛+1 ≤ (−𝐶0 ) (−𝜓 ′ (𝑎∗ )). Since 𝑡𝑛+1 =
2𝐷
1
𝐷𝑛+1 ≥ 𝐺(𝑔(𝑎𝑛 ) − 𝛿ℎ) − 𝑎(𝑡𝑛 + ℎ) =∶ 𝜂(ℎ). 2
1 + ℎ𝜆 (𝑛 + 1)ℎ < 𝑇𝑎∗ , i.e. (𝑛 + 1) < 𝑇𝑎∗ ∕ℎ, and (1 + ℎ𝜓 ′ (𝑎∗ )) ∈ (0, 1), we have
ℎ
We then take a Taylor expansion of the right hand side function 𝜂(ℎ) 𝛽(ℎ) ∶= > (1+ℎ𝜓 ′ℎ(𝑎∗ ))𝑛+1 . To prove that 𝛽(ℎ) is strictly in-
(1+ℎ𝜓 ′ (𝑎∗ ))𝑇𝑎∗ ∕ℎ
around ℎ = 0, creasing for ℎ > 0, we consider

149
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

Fig. 1. Error at 𝑇𝑓 = 0.1660 for first-order methods (left) and second-order methods (right) applied to the semi-linear equation with 𝛿𝐹 (𝑢) = 3𝑒𝑢 , with different values
of ℎ.

1 ℎ
𝛽 ′ (ℎ) = − (1 + ℎ𝜓 ′ (𝑎∗ ))𝑇𝑎∗ ∕ℎ Table 1
(1 + ℎ𝜓 ′ (𝑎∗ ))𝑇𝑎∗ ∕ℎ [(1 + ℎ𝜓 ′ (𝑎∗ ))𝑇𝑎∗ ∕ℎ ]2 Error at 𝑇𝑓 = 0.1660 for first-order methods (top) and second-order methods
[ ]
−𝑇𝑎∗ 𝑇 ∗ 𝜓 ′ (𝑎∗ ) (bottom) applied to the semi-linear equation with 𝛿𝐹 (𝑢) = 3𝑒𝑢 .
⋅ ln(1 + ℎ𝜓 ′ (𝑎∗ )) + 𝑎
ℎ2 ℎ 1 + ℎ𝜓 (𝑎 )
′ ∗
[ ( )] Timestep 5e-005 2.5e-005 1.25e-005 8e-006 5e-006
1 1 𝜓 ′ (𝑎∗ )
= 1 + 𝑇 𝑎∗ ln(1 + ℎ𝜓 ′ (𝑎∗ )) − , FE 0.277 0.152 0.08 0.0522 0.0331
(1 + ℎ𝜓 (𝑎 )) 𝑎
′ ∗ 𝑇 ∗ ∕ℎ ℎ 1 + ℎ𝜓 (𝑎 )
′ ∗
BE 0.468 0.194 0.0904 0.0565 0.0347
𝜓 (𝑎 ) ′ ∗ SpFE 0.0361 0.0183 0.00919 0.00589 0.00369
which is clearly positive if ln(1 + ℎ𝜓 ′ (𝑎∗ )) > ℎ 1+ℎ𝜓 ′ (𝑎∗ ) . Since 𝑥 − SpFEA 0.0379 0.0187 0.0093 0.00594 0.00371
ln 𝑥 > 1 for 𝑥 > 1, and (1 + ℎ𝜓 ′ (𝑎∗ ))−1 ∈ (1, ∞), the above inequal- SpBE 0.00533 0.00269 0.00135 0.000864 0.000541
SpBEA 0.00551 0.00273 0.00136 0.000869 0.000543
ity is satisfied and 𝛽(ℎ) is strictly increasing. Moreover 𝛽(0) = 0
ℎ VCFE 0.019 0.00956 0.0048 0.00307 0.00192
and limℎ→ −1 𝛽(ℎ) = +∞, so that the equation ′ ∗ 𝑇 ∗ ∕ℎ
=
𝜓 ′ (𝑎∗ ) (1+ℎ𝜓 (𝑎 )) 𝑎
2𝐷0 (−𝜓 ′ (𝑎∗ ))
(−𝐶2 )
has exactly one solution ℎ̃ and if ℎ < ℎ̃ we have 𝐷𝑛+1 > 0. Timestep 0.0002 0.000125 0.0001 5e-005 2.5e-005

VCBE 0.0195 0.0097 0.00483 0.00309 0.00193

MR 0.00833 0.00324 0.00207 0.000516 0.000129
The existence and uniqueness results of this section can be general-
TR 0.0407 0.0152 0.00961 0.00237 0.000591
ized to quasi-linear parabolic equations with power-like nonlinearities SoSpFE 0.000305 0.000121 7.75e-005 1.94e-005 4.87e-006
SoSpBE 0.000305 0.000121 7.75e-005 1.94e-005 4.87e-006
𝑢𝑡 = 𝛼𝑢𝑚 + Δ𝑢𝑚 , in Ω × (0, 𝑇 ),
VCMR 0.00033 0.00013 8.36e-005 2.1e-005 5.25e-006
𝑢 = 0, on 𝜕Ω × (0, 𝑇 ), VCTR 0.000733 0.000287 0.000184 4.6e-005 1.15e-005
𝑢(𝑥, 0) = 𝑢0 (𝑥), in Ω,

where Ω is a bounded domain in ℝ𝑑 , 𝑚 > 1 and 𝛼 ≥ 0, see [5], but the

upper bound blow up estimate remains currently open. in Table 1 the errors in the computed solutions up to 𝑇𝑓 = 0.1660 with
different step sizes. We observe that the error of B-methods is approx-
4. Numerical results imately 10 times smaller for first-order methods (and even more for
SpBE and SpBEA) and 30 times smaller for second-order B-methods
We now test the new splitting B-methods on several non-linear par- compared to standard methods. In Fig. 1, we show these results graph-
tial differential equations, and also compare them to the B-methods ically. As expected, the slopes of the lines corresponding to first-order
based on the variation of constants approach from [6] called VCFE, methods are approximately one, whereas the slopes of the lines corre-
VCBE, VCMR and VCTR. Unless otherwise specified, we use a fixed sponding to second-order methods are close to two. In Fig. 2 we show
time-step size ℎ for all experiments, and the reference solution is com- the behavior in time of the methods as blow-up is approached, using
puted using ode45, an adaptive integrator in Matlab that implements ℎ = 0.0001 and computing the solution up to 𝑇𝑓 = 0.1663.
the Dormand–Prince pair. All errors reported correspond to the maxi-
mum distance from this reference solution, taken at a final time 𝑇𝑓 that 4.2. A quasi-linear parabolic equation
will be specified in each case.
Splitting B-methods can also be constructed for more general non-
4.1. A semi-linear parabolic equation linear PDEs. We illustrate this now with the quasi-linear equation with
power-type nonlinearities, 𝑢𝑡 = Δ𝑢𝜎+1 + 𝛼𝑢𝛽+1 , with 𝛽 > 0, 𝜎 > 0 and 𝛼 ≥
For the first example, we study the semi-linear parabolic equation 0. We split the PDE right hand side into 𝑓 [1] (𝑢) = 𝛼𝑢𝛽+1 and 𝑓 [2] (𝑢) =
𝑢𝑡 = Δ𝑢 + 𝛿𝑒𝑢 on the interval Ω = [−1, 1]. We discretize the Laplacian Δ𝑢𝜎+1 , and using that the nonlinear part 𝑦𝑡 = 𝛼𝑦𝛽+1 is solved by 𝑦(𝑡) =
operator in space using a fourth order finite difference method with [ ]−1∕𝛽
1 −𝛽
a five point stencil and a mesh size of Δ𝑥 = 2∕30. We set 𝛿 = 3 and ( 𝐾−𝛼𝛽𝑡 )1∕𝛽 , the exact flow of the first part is 𝜑[1]
𝑡 (𝑢𝑛 ) = 𝑢𝑛 − 𝛼𝛽𝑡 .
𝑢0 (𝑥) = cos(𝜋𝑥∕2), which is concave on the whole interval. Using adap- By choosing Φ[2]
ℎ
to be the forward Euler method, so that Φ[2]∗
ℎ
is the
tive methods, we can estimate the blow-up time at 𝑇𝑏 ≈ 0.1664. We show backward Euler method we obtain the corresponding SpFE method

150
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

Fig. 2. Error for first-order methods (left) and second-order methods (right) applied to the semi-linear equation with 𝛿𝐹 (𝑢) = 3𝑒𝑢 , for time steps close to 𝑇𝑓 = 0.1663.

[ ]−1∕𝛽
Φℎ (𝑢𝑛 ) = 𝑢𝑛+1 = (𝑢𝑛 + ℎΔ𝑢𝜎+1
𝑛 )
−𝛽
− 𝛼𝛽ℎ , (44) Table 2
Error at 𝑇𝑓 = 0.1000 for first-order methods (top) and second-order methods
and its adjoint, the (SpFE)∗ , (bottom) applied to the quasi-linear equation 𝑢𝑡 = Δ𝑢2 + 8𝑢3 .

( )−1∕𝛽 Timestep 0.000125 8e-005 5e-005 2.5e-005 1.25e-005

Φ∗ℎ (𝑢𝑛 ) = 𝑢𝑛+1 = 𝑢−𝛽
𝑛 − 𝛼𝛽ℎ + ℎΔ𝑢𝜎+1
𝑛+1
. (45)
FE 0.0188 0.0121 0.00762 0.00383 0.00192
If we choose Φ[2]
ℎ
to be backward Euler we get SpBE, BE 0.0196 0.0125 0.00774 0.00386 0.00192
SpFE 0.0082 0.00526 0.00329 0.00165 0.000824
SpFEA 0.00829 0.0053 0.00331 0.00165 0.000825
[ ]−1∕𝛽
Φℎ (𝑢𝑛 ) = 𝑣−𝛽 − 𝛼𝛽ℎ , (46) SpBE 0.004 0.00256 0.0016 0.0008 0.0004
SpBEA 0.004 0.00256 0.0016 0.0008 0.0004
where 𝑣 is solution of 𝑣 − ℎΔ(𝑣𝜎+1 ) = 𝑢 𝑛, and its adjoint (SpBE)∗ VCFE 0.00209 0.00134 0.000837 0.000419 0.000209
VCBE 0.0021 0.00134 0.000839 0.000419 0.00021
[ ]−1∕𝛽 ([ ]−(𝜎+1)∕𝛽 )
Φ∗ℎ (𝑢𝑛 ) = 𝑢−𝛽
𝑛 − 𝛼𝛽ℎ + ℎΔ 𝑢−𝛽
𝑛 − 𝛼𝛽ℎ . (47)
Timestep 0.0005 0.00025 0.000125 8e-005 5e-005
The second-order methods obtained by composing these methods are MR 0.000191 4.78e-005 1.19e-005 4.89e-006 1.91e-006
quite simple. If Φ[2]
ℎ
is the forward Euler method, the composed method TR 0.000499 0.000125 3.11e-005 1.28e-005 4.98e-006
SoSpFE 3.72e-005 9.29e-006 2.32e-006 9.52e-007 3.72e-007
is SoSpFE
SoSpBE 5.84e-005 1.46e-005 3.65e-006 1.49e-006 5.84e-007
(( )−𝛽 )−1∕𝛽 VCMR 2.15e-006 5.37e-007 1.34e-007 5.5e-008 2.15e-008
ℎ ℎ VCTR 2.17e-005 5.42e-006 1.35e-006 5.55e-007 2.17e-007
Ψℎ (𝑢𝑛 ) = 𝑣+ Δ(𝑣𝜎+1 ) − 𝛼𝛽 , (48)
2 2

where 𝑣 is the solution of 𝑣 − ℎ2 Δ(𝑣𝜎+1 ) = (𝑢−𝛽 ℎ −1∕𝛽

𝑛 − 𝛼𝛽 2 ) . Similarly, the
second-order method obtained using the backward Euler method for 4.3. A semi-linear system
Φ[2]
ℎ
is SoSpBE, given implicitly by
In [19] and [20], Friedman and Giga considered parabolic sys-
(( )−𝛽 𝛼𝛽ℎ )−1∕𝛽
ℎ tems of the form 𝑢𝑡 − 𝑢𝑥𝑥 = 𝑓 (𝑣), 𝑣𝑡 − 𝑣𝑥𝑥 = 𝑔(𝑢), where 𝑓 and 𝑔 are
Ψℎ (𝑢𝑛 ) = 𝑢𝑛+1 = 𝑣 + Δ(𝑢𝜎+1
𝑛+1
) − , (49)
2 2 positive, increasing and superlinear. They showed that the solutions
𝛼𝛽ℎ −1∕𝛽 ℎ 𝛼𝛽ℎ −(𝜎+1)∕𝛽
exhibit a single-point blow-up. More complex systems of the form
where 𝑣 = [𝑢−𝛽
𝑛 − 2 ] + 2 Δ[(𝑢−𝛽
𝑛 − 2 ) ]. We show a numerical (𝑢𝑖 )𝑡 = Δ𝑢𝑖 + 𝑓𝑖 (𝑢1 , … , 𝑢𝑚 ) were studied by Bebernes and Lacey [4], Gang
example for the quasi-linear equation 𝑢𝑡 = Δ𝑢2 + 8𝑢3 , on Ω = (−1, 1) with and Sleeman [23] and Chen [12]. In this subsection, we derive several
the same initial condition as above: 𝑢0 (𝑥) = cos(𝜋𝑥∕2). The blow-up time specialized methods for the simple case
is approximately 𝑇𝑏 ≈ 0.1128. We list in Table 2 the errors we obtained.
We observe that the B-methods obtained by variation of the constant 𝑢𝑡 = Δ𝑢 + 𝛿𝑒𝑣 , 𝑣𝑡 = Δ𝑣 + 𝛾𝑒𝑢 . (50)
are more accurate than those obtained by splitting methods. Compared
We first solve the nonlinear system of ordinary differential equations
with standard methods, the errors are 10 times smaller for first-order
𝑦′ (𝑡) = 𝛿𝑒𝑧(𝑡) , 𝑧′ (𝑡) = 𝛾𝑒𝑦(𝑡) , to get 𝑦(𝑡) = ln 𝐾 − ln[1 − 𝛿𝑒𝐾𝑡+𝐷 ] − ln 𝛾, 𝑧(𝑡) =
methods of the first type and between 2 and 7 times smaller for first-
ln 𝐾 − ln[1 − 𝛿𝑒𝐾𝑡+𝐷 ] + 𝐾𝑡 + 𝐷, where 𝐾 and 𝐷 are constants of in-
order methods of the second type. Among second-order methods, the
tegration, determined by the initial conditions, 𝐾 = 𝛾𝑒𝑦(0) − 𝛿𝑒𝑧(0) and
method obtained by variation of the constant and the midpoint rule
𝐷 = 𝑧(0) − 𝑦(0) − ln 𝛾. Then for each choice of numerical integrator Φ[2] ℎ
(VCMR) is remarkably better than the others, as its error is more than
applied to 𝑢𝑡 = Δ𝑢, 𝑣𝑡 = Δ𝑣, we obtain two schemes that are adjoint to
fifty times smaller that the error of the standard midpoint rule. In Fig. 3
each other. The forward Euler method leads to the explicit SpFE scheme
we show the corresponding data graphically. The step-by-step errors
are plotted in Fig. 4 up to 𝑇𝑓 = 0.1110, when the solutions are computed ( )
ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] − ln 𝛾
using the timestep ℎ = 0.0001. Φℎ (𝑢𝑛 , 𝑣𝑛 ) = ,
ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] + 𝐷𝑛 + ℎ𝐾𝑛

151
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

Fig. 3. Error at 𝑇𝑓 = 0.1000 for first-order methods (left) and second-order methods (right) applied to the quasi-linear equation 𝑢𝑡 = Δ𝑢2 + 8𝑢3 , with different values
of ℎ.

Fig. 4. Error for first-order methods (left) and second-order methods (right) applied to the quasi-linear equation 𝑢𝑡 = Δ𝑢2 + 8𝑢3 , for timesteps close to 𝑇𝑓 = 0.1110.

where 𝐾𝑛 = 𝛾𝑒𝑢𝑛 +ℎΔ𝑢𝑛 − 𝛿𝑒𝑣𝑛 +ℎΔ𝑣𝑛 , 𝐷𝑛 = 𝑣𝑛 + ℎΔ𝑣𝑛 − 𝑢𝑛 − ℎΔ𝑢𝑛 − ln 𝛾, and We can also compose these methods to construct second-order splitting
the adjoint scheme Φ∗ℎ we call (SpFE)∗ is given by B-methods. For these, we first define 𝐾𝑛 ∶= 𝛾𝑒𝑢𝑛 − 𝛿𝑒𝑣𝑛 , 𝐷𝑛 ∶= 𝑣𝑛 − 𝑢𝑛 −
ln 𝛾 as before. Then, if we choose Φ[2]
ℎ
to be the forward Euler method,
𝑢𝑛+1 = ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] − ln 𝛾 + ℎΔ𝑢𝑛+1 , we define 𝑤1 and 𝑤2 to be the solutions of
𝑣𝑛+1 = ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] + 𝐷𝑛 + ℎ𝐾𝑛 + ℎΔ𝑣𝑛+1 , ℎ
𝑤1 − ℎ2 Δ𝑤1 = ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 + 2 𝐾𝑛 ] − ln 𝛾,
where 𝐾𝑛 = − 𝛾𝑒𝑢𝑛 𝛿𝑒𝑣𝑛 ,
𝐷𝑛 = 𝑣𝑛 − 𝑢𝑛 − ln 𝛾. If we choose instead the ℎ
𝑤2 − ℎ2 Δ𝑤2 = ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 + 2 𝐾𝑛 ] + 𝐷𝑛 + ℎ2 𝐾𝑛 ,
backward Euler method, we obtain the SpBE scheme
( ) and we define 𝐾̃ ∶= 𝛾 exp(𝑤1 + ℎ2 Δ𝑤1 ) − 𝛿 exp(𝑤2 + ℎ2 Δ𝑤2 ), 𝐷̃ ∶= 𝑤2 +
ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] − ln 𝛾 ℎ
Δ𝑤2 − 𝑤1 − ℎ2 Δ𝑤1 − ln 𝛾, to finally get the SoSpFE scheme
Φℎ (𝑢𝑛 , 𝑣𝑛 ) = ,
ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] + 𝐷𝑛 + ℎ𝐾𝑛
2

̃ ℎ ̃
where 𝐾𝑛 = 𝛾𝑒𝑤1 − 𝛿𝑒𝑤2 , 𝐷𝑛 = 𝑤2 − 𝑤1 − ln 𝛾, and 𝑤1 and 𝑤2 are solutions 𝑢𝑛+1 = ln 𝐾̃ − ln[1 − 𝛿𝑒𝐷+ 2 𝐾 ] − ln 𝛾,
̃ ℎ ̃ (51)
of 𝑤1 = 𝑢𝑛 + ℎΔ𝑤1 and 𝑤2 = 𝑣𝑛 + ℎΔ𝑤2 . For its adjoint method (SpBE)∗ , 𝑣𝑛+1 = ln 𝐾̃ − ln[1 − 𝛿𝑒𝐷+ 2 𝐾 ] + 𝐷̃ + ℎ2 𝐾.
̃
we first define 𝐾𝑛 ∶= 𝛾𝑒𝑢𝑛 − 𝛿𝑒𝑣𝑛 , 𝐷𝑛 ∶= 𝑣𝑛 − 𝑢𝑛 − ln 𝛾, and 𝑤1 ∶= ln 𝐾𝑛 −
ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] − ln 𝛾, 𝑤2 ∶= ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 +ℎ𝐾𝑛 ] + 𝐷𝑛 + ℎ𝐾𝑛 . Then If we choose to use the backward Euler method as Φ[2]
ℎ
, we need to first
ℎ ℎ
the (SpBE)∗ scheme can be written as define 𝑢̃ ∶= ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 + 2 𝐾𝑛 ] − ln 𝛾, 𝑣̃ ∶= ln 𝐾𝑛 − ln[1 − 𝛿𝑒𝐷𝑛 + 2 𝐾𝑛 ] +
𝐷𝑛 + ℎ2 𝐾𝑛 , and then 𝑤1 and 𝑤2 are the solutions of
( )
𝑤1 + ℎΔ𝑤1 ℎ ℎ ℎ ℎ
Φ∗ℎ (𝑢𝑛 , 𝑣𝑛 ) = . 𝑤1 − Δ𝑤1 = 𝑢̃ + Δ𝑢,
̃ 𝑤2 − Δ𝑤2 = 𝑣̃ + Δ𝑣,
̃
𝑤2 + ℎΔ𝑤2 2 2 2 2
152
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

Fig. 5. Error at 𝑇𝑓 = 0.1100 for first-order methods (left) and second-order methods (right) applied to the system of semi-linear equations with different values of ℎ.

Table 3 blow-up, while their behavior before blow-up is similar to classical time
Error at 𝑇𝑓 = 0.1100 for first-order methods (top) and second-order methods stepping schemes.
(bottom) applied to the system of semi-linear equations.

Timestep 0.0001 5e-005 2.5e-005 1.25e-005 8e-006 Data availability

FE 0.0146 0.00736 0.00369 0.00185 0.00118
BE 0.015 0.00747 0.00372 0.00186 0.00119 No data was used for the research described in the article.
SpFE 0.00146 0.00073 0.000365 0.000183 0.000117
SpFEA 0.000679 0.000339 0.000169 8.47e-005 5.42e-005 Acknowledgements
SpBE 0.000675 0.000338 0.000169 8.46e-005 5.42e-005
SpBEA 0.00146 0.000731 0.000365 0.000183 0.000117
VCFE 0.00118 0.00059 0.000295 0.000148 9.45e-005 This work was funded by the Swiss National Science Foundation
VCBE 0.00118 0.000591 0.000295 0.000148 9.45e-005 (200020_192064).

Timestep 0.0004 0.0002 0.0001 5e-005 2.5e-005 References

MR 5.91e-005 1.48e-005 3.69e-006 9.23e-007 2.31e-007
TR 0.000339 8.48e-005 2.12e-005 5.3e-006 1.32e-006 [1] H. Amann, On the existence of positive solutions of nonlinear elliptic boundary
SoSpFE 4.85e-006 1.21e-006 3.03e-007 7.57e-008 1.89e-008 value problems, Indiana Univ. Math. J. 21 (1971/1972) 125–146.
SoSpBE 4.85e-006 1.21e-006 3.03e-007 7.57e-008 1.89e-008 [2] R.O. Ayeni, On the blow up problem for semilinear heat equations, SIAM J. Math.
VCMR 5.82e-006 1.46e-006 3.64e-007 9.1e-008 2.28e-008 Anal. 14 (1) (1983) 138–141.
VCTR 6.4e-006 1.6e-006 4e-007 1e-007 2.5e-008 [3] C. Bandle, H. Brunner, Blowup in diffusion equations: a survey, J. Comput. Appl.
Math. 97 (1–2) (1998) 3–22.
[4] J. Bebernes, A. Lacey, Finite-time blowup for a particular parabolic system, SIAM J.
and we define 𝐾̃ ∶= 𝛾 exp(𝑤1 ) − 𝛿 exp(𝑤2 ), 𝐷̃ ∶= 𝑤2 − 𝑤1 − ln 𝛾, to finally Math. Anal. 21 (6) (1990) 1415–1425.
[5] M. Beck, B-methods: Special Time-integrators for Differential Equations with Blow-
get 𝑢𝑛+1 and 𝑣𝑛+1 by (51) for the SoSpBE scheme.
up Solutions, PhD thesis, McGill University, 2009.
We now present the results of numerical experiments for the system [6] M. Beck, M.J. Gander, F. Kwok, B-methods for the numerical solution of evolution
of semi-linear parabolic equations (50) with 𝛿 = 3 and 𝛾 = 5. The initial problems with blow-up solutions part I: variation of the constant, SIAM J. Sci. Com-
conditions are 𝑢0 (𝑥) = cos(𝜋𝑥∕2) and 𝑣0 (𝑥) = cos(𝜋𝑥∕2) on Ω = [−1, 1]. put. 37 (6) (2015) A2998–A3029.
The blow-up time is approximately 𝑇𝑏 ≈ 0.1181. The errors are listed in [7] N. Bou-Rabee, J.M. Sanz-Serna, Geometric integrators and the Hamiltonian Monte
Carlo method, Acta Numer. 27 (2018) 113–206.
Table 3. and shown graphically in Fig. 5. [8] H. Brezis, J.L. Vázquez, Blow-up solutions of some nonlinear elliptic problems, Rev.
In Fig. 6, we show again the evolution of the solution as we approach Mat. Univ. Complut. Madr. 10 (2) (1997) 443–469.
blow up. We used ℎ = 0.0001 and computed the solutions up to 𝑇𝑓 = [9] C.J. Budd, B. Dold, A. Stuart, Blowup in a partial differential equation with con-
0.1170. Further examples, including one for the semilinear hyperbolic served first integral, SIAM J. Appl. Math. 53 (3) (1993) 718–742.
[10] C.J. Budd, W. Huang, R.D. Russell, Moving mesh methods for problems with blow-
equation 𝑢𝑡𝑡 = Δ𝑢 + 5𝑒𝑢 , can be found in [5, Chapter 3 and Appendix A].
up, SIAM J. Sci. Comput. 17 (2) (1996) 305–327.
[11] C.J. Budd, J. Williams, V.A. Galaktionov, Self-similar blow-up in higher-order semi-
5. Conclusions linear parabolic equations, SIAM J. Appl. Math. 64 (5) (2004) 1775–1809.
[12] S. Chen, A sufficient condition for blowup solutions of nonlinear heat equations, J.
We presented in this paper a systematic approach for deriving nu- Math. Anal. Appl. 293 (1) (2004) 227–236.
[13] S. Chen, H. Li, J. Li, M. Mei, K. Zhang, Global and blow-up solutions for compress-
merical integrators which are very accurate for semi- and quasi-linear
ible Euler equations with time-dependent damping, J. Differ. Equ. 268 (9) (2020)
parabolic and hyperbolic partial differential equations exhibiting blow- 5035–5077.
up in finite time. We call this new class of geometric integration meth- [14] G.K. Duong, N. Nouaili, H. Zaag, Refined asymptotic for the blow-up solution of
ods B-methods, where B stands for blow-up. Our construction is com- the complex Ginzburg-Landau equation in the subcritical case, Ann. Inst. Henri
pletely general, and can lead to B-methods for many other nonlinear Poincaré, Anal. Non Linéaire 39 (1) (2021) 41–85.
[15] E. Faou, Geometric Numerical Integration and Schrödinger Equations, vol. 15, Eu-
partial differential equations that were not considered in this paper. ropean Mathematical Society, 2012.
Because of their construction, which takes the blow-up behavior into [16] M. Fasondini, J.R. King, J. Weideman, Blow up in a periodic semilinear heat equa-
account, all these methods will behave substantially better close to tion, Physica D 446 (2023) 133660.

153
M. Beck, M.J. Gander and F. Kwok Computers and Mathematics with Applications 152 (2023) 143–154

Fig. 6. Error for first-order methods (left) and second order methods (right) applied to the system of semi-linear equations, for timesteps close to 𝑇𝑓 = 0.1170.

[17] M.S. Floater, Blow-up at the boundary for degenerate semilinear parabolic equa- [35] H.B. Keller, Elliptic boundary value problems suggested by nonlinear diffusion pro-
tions, Arch. Ration. Mech. Anal. 114 (1) (1991) 57–77. cesses, Arch. Ration. Mech. Anal. 35 (1969) 363–381.
[18] D. Frank-Kamenetskii, Diffusion and Heat Exchange in Chemical Kinetics, Princeton [36] J.B. Keller, On solutions of nonlinear wave equations, Commun. Pure Appl. Math.
University Press, Princeton, N.J., 1955. 10 (1957) 523–530.
[19] A. Friedman, Blow-up of solutions of nonlinear parabolic equations, in: Nonlinear [37] P. Kelley, Self-focusing of optical beams, Phys. Rev. Lett. 15 (26) (1965) 1005–1008.
Diffusion Equations and Their Equilibrium States, I, Berkeley, CA, 1986, in: Math. [38] C.E. Kenig, T. Duyckaerts, F. Merle, Universality of blow-up profile for small radial
Sci. Res. Inst. Publ., vol. 12, Springer, New York, 1988, pp. 301–318. type II blow-up solutions of the energy-critical wave equation, J. Eur. Math. Soc.
[20] A. Friedman, Y. Giga, A single point blow-up for solutions of semilinear parabolic 13 (3) (2011) 533–599.
systems, J. Fac. Sci., Univ. Tokyo, Sect. 1A, Math. 34 (1) (1987) 65–79. [39] A.A. Lacey, The form of blow-up for nonlinear parabolic equations, Proc. R. Soc.
[21] V.A. Galaktionov, A boundary value problem for the nonlinear parabolic equation Edinb., Sect. A 98 (1–2) (1984) 183–202.
𝑢𝑡 = Δ𝑢𝜎+1 + 𝑢𝛽 {English translation: Differ. Equ. 17 (5) (1981) 551–556}, Differ. [40] M.-N. Le Roux, Semidiscretization in time of nonlinear parabolic equations with
Uravn. 17 (5) (1981) 836–842, 956. blowup of the solution, SIAM J. Numer. Anal. 31 (1) (1994) 170–195.
[22] V.A. Galaktionov, J.L. Vázquez, The problem of blow-up in nonlinear parabolic [41] B. Leimkuhler, S. Reich, Simulating Hamiltonian Dynamics, vol. 14, Cambridge Uni-
equations, Discrete Contin. Dyn. Syst. 8 (2) (2002) 399–433. versity Press, 2004.
[23] L. Gang, B. Sleeman, Non Existence of Global Solutions to Systems of Semi Linear [42] B. LeMesurier, G. Papanicolaou, C. Sulem, P.-L. Sulem, The focusing singularity of
Parabolic Equations, University of Dundee. Department of Mathematical Sciences, the nonlinear Schrödinger equation, in: Directions in Partial Differential Equations,
1990. Madison, WI, 1985, in: Publ. Math. Res. Center Univ. Wisconsin, vol. 54, Academic
[24] I.M. Gel’fand, Some problems in the theory of quasilinear equations, Am. Math. Soc. Press, Boston, MA, 1987, pp. 159–201.
Transl. 2 (29) (1963) 295–381. [43] D. McLaughlin, G. Papanicolaou, C. Sulem, P. Sulem, Focusing singularity of the
[25] R.T. Glassey, Blow-up theorems for nonlinear wave equations, Math. Z. 132 (1973) cubic Schrödinger equation, Phys. Rev. A 34 (2) (1986) 1200.
183–203. [44] F. Merle, H. Zaag, Determination of the blow-up rate for a critical semilinear wave
[26] R.T. Glassey, On the blowing up of solutions to the Cauchy problem for nonlinear equation, Math. Ann. 331 (2) (2005) 395–416.
Schrödinger equations, J. Math. Phys. 18 (9) (1977) 1794–1797. [45] F. Merle, H. Zaag, Existence and classification of characteristic points at blow-up for
[27] M.E. Gurtin, R.C. MacCamy, On the diffusion of biological populations, Math. Biosci. a semilinear wave equation in one space dimension, Am. J. Math. 134 (3) (2012)
33 (1) (1977) 35–49. 581–648.
[28] E. Hairer, M. Hochbruck, A. Iserles, C. Lubich, Geometric numerical integration, [46] N. Saito, T. Sasaki, Blow-up of finite-difference solutions to nonlinear wave equa-
Oberwolfach Rep. 3 (1) (2006) 805–882. tions, J. Math. Sci. Univ. Tokyo 23 (1) (2016) 349–380.
[29] E. Hairer, C. Lubich, G. Wanner, Geometric Numerical Integration, Springer-Verlag, [47] P. Souplet, Finite time blow-up for a non-linear parabolic equation with a gradient
Berlin, 2002. term and applications, Math. Methods Appl. Sci. 19 (16) (1996) 1317–1333.
[30] M.A. Hamza, H. Zaag, The blow-up rate for a non-scaling invariant semilinear wave [48] Y. Su, D. Zhang, Construction of minimal mass blow-up solutions to rough nonlinear
equations, J. Math. Anal. Appl. 483 (2) (2020) 123652. Schrödinger equations, J. Funct. Anal. 284 (5) (2023) 109796.
[31] C. Hirota, K. Ozawa, Numerical method of estimating the blow-up time and rate [49] A. Takayasu, K. Matsue, T. Sasaki, K. Tanaka, M. Mizuguchi, S. Oishi, Numerical
of the solution of ordinary differential equations—an application to the blow-up validation of blow-up solutions of ordinary differential equations, J. Comput. Appl.
problems of partial differential equations, J. Comput. Appl. Math. 193 (2) (2006) Math. 314 (2017) 10–29.
614–637. [50] M. Tsutsumi, Existence and nonexistence of global solutions for nonlinear parabolic
[32] A. Kapila, Reactive-diffusive system with Arrhenius kinetics: dynamics of ignition, equations, Publ. Res. Inst. Math. Sci. 8 (1972) 211–229.
SIAM J. Appl. Math. 39 (1) (1980) 21–36. [51] M. Tsutsumi, Existence and nonexistence of global solutions of the first bound-
[33] S. Kaplan, On the growth of solutions of quasi-linear parabolic equations, Commun. ary value problem for a certain quasilinear parabolic equation, Funkc. Ekvacioj 17
Pure Appl. Math. 16 (1963) 305–330. (1974) 13–24.
[34] D. Kassoy, J. Poland, The thermal explosion confined by a constant temperature [52] J.L. Vázquez, A strong maximum principle for some quasilinear elliptic equations,
boundary: I-the induction period solution, SIAM J. Appl. Math. 39 (3) (1980) Appl. Math. Optim. 12 (3) (1984) 191–202.
412–430.

154