Brigham Young University Brigham Young University BYU ScholarsArchive BYU ScholarsArchive Theses and Dissertations 2020-04-09 Stability of MHD Shock Waves Stability of MHD Shock Waves Bryn Nicole Barker Brigham Young University Follow this and additional works at: https://scholarsarchive.byu.edu/etd Part of the Physical Sciences and Mathematics Commons BYU ScholarsArchive Citation BYU ScholarsArchive Citation Barker, Bryn Nicole, "Stability of MHD Shock Waves" (2020). Theses and Dissertations. 8437. https://scholarsarchive.byu.edu/etd/8437 This Thesis is brought to you for free and open access by BYU ScholarsArchive. It has been accepted for inclusion in Theses and Dissertations by an authorized administrator of BYU ScholarsArchive. For more information, please contact [email protected], [email protected].
77
Embed
Stability of MHD Shock Waves - BYU ScholarsArchive
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Brigham Young University Brigham Young University
BYU ScholarsArchive BYU ScholarsArchive
Theses and Dissertations
2020-04-09
Stability of MHD Shock Waves Stability of MHD Shock Waves
Bryn Nicole Barker Brigham Young University
Follow this and additional works at: https://scholarsarchive.byu.edu/etd
Part of the Physical Sciences and Mathematics Commons
BYU ScholarsArchive Citation BYU ScholarsArchive Citation Barker, Bryn Nicole, "Stability of MHD Shock Waves" (2020). Theses and Dissertations. 8437. https://scholarsarchive.byu.edu/etd/8437
This Thesis is brought to you for free and open access by BYU ScholarsArchive. It has been accepted for inclusion in Theses and Dissertations by an authorized administrator of BYU ScholarsArchive. For more information, please contact [email protected], [email protected].
This thesis focuses on the study of spectral stability of planar shock waves in 2-dimensionalmagnetohydrodynamics. We begin with a numerical approach, computing the Lopatinski de-terminant and Evans function with the goal of determining if there are parameters for whichviscous waves are unstable and the corresponding inviscid waves are stable. We also begindeveloping a method to obtain an explicit, analytical representation of the Evans function.We demonstrate the capabilities of this method with compressible Navier-Stokes and extendour results to 2-D MHD. Finally, using compressible Navier-Stokes again, we derive an energyestimate as a first step in improving the bound on possible roots of the Evans function.
Keywords: MHD, shocks, traveling waves, Lopatinski determinant, Evans function
Acknowledgements
A huge thanks to my advisor Dr. Blake Barker for his dedication to mentoring me
and preparing me for success in mathematical research. I am inspired by his passion for
mathematical discovery and his determination to make discoveries. I am also extremely
grateful for other math professors who have supported and mentored me throughout my
time at BYU, including Dr. Ben Webb, Dr. Emily Evans, and Dr. Wayne Barrett. Without
their encouragement, I would not have the appreciation and love for math that I do today.
In addition, I would like to thank my kids, George and Harris, for ensuring that finishing
this thesis was a real challenge and my family and friends for having confidence in my ability
to succeed. However, I owe the most thanks to my husband DJ for really making this happen.
Note that (2.14) is equivalent to (2.10) but with the retention of the linearization of the
viscous terms.
We recall that to show stability of the viscous shock profile U , we must prove that the
eigenvalue problem (2.14) has no unstable spectra. As mentioned, this is equivalent to
showing that the Evans function D(λ, ξ) is nonzero for all λ such that Re(λ) ≥ 0 excluding
λ = 0. In order to simplify the computations required to check for zeros of the Evans
function, we aim to construct the Evans function in a way that ensures analyticity in λ
which in turn warrants use of the argument principle when searching for zeros.
Our approach is to first formulate (2.14) as the first order system,
W ′ = A(x1;λ, ξ)W, (2.15)
where based on the block structure of Bjk (as mentioned earlier), A is an N × N matrix
where N = 2n− r (this is because we have a second derivative in n− r equations and thus
when made into a first order system we will have 2(n − r) equations from these n − r and
then an additional r hyperbolic equations).
We then define A±(λ, ξ) = limx1→±∞A(x1;λ, ξ) and proceed by building subspaces of
solutions for (2.15). We let {W+1 , ...,W
+k } be an unstable basis for A+ and {W−
k+1, ...,W−N }
be a stable basis for A−. Notice that the basis {W+1 , ...,W
+k } decays as x1 → +∞ and
11
{W−k+1, ...,W
−N } decays as x1 → −∞.
With these bases defined, the Evans function is
D(λ, ξ) = det(W+1 , ...,W
+k ,W
−k+1, ...W
−N )|x1=0. (2.16)
If there exist λ′, ξ′ with Re(λ′) ≥ 0 such that D(λ′, ξ′) = 0 then our two bases are linearly
independent, implying existence of an eigenfunction of (2.15) which in turn implies the λ is
an eigenvalue of (2.15) and the underlying wave is unstable.
Notice that in this case, the system (2.15) is not constant coefficient so we must evolve
our bases to the center of the domain before determining linear independence. This idea is
displayed in Figure 2.1. We begin by determining a basis for each end of our domain, as
shown in Figure 2.1 (a). Then to determine existence of an eigenfunction that spans the
entire domain, we evolve the bases to the center and determine if they are linearly dependent
at their intersection, as shown in Figure 2.1 (b).
(a) (b)
Figure 2.1: Diagrams showing the main idea behind evolving bases from each end of thedomain to determine existence of an eigenfunction for the Evans system.
12
To create the Evans system for 2-D MHD, we start with (2.14), which we can rearrange
to
λA0U + iξA2U + ξ2B22U =(B11U ′ + iξB2U − A1U
)′,
where A2 = A2 + (B21)′ and B2 = B21 + B12.
From this formulation we are able to define our flux variable
f = B11U ′ + iξB2U − A1U,
and we will set Aξ = ξA2, Bξ = ξB2, and Bξξ = ξξB22. We are now ready to cast (2.14) asthe first order system (2.15) with
W =
f
u2
,A =
−(λA0
11 + iAxi11)(A111)−1 0 −λ(A0
11(A111)−1A1
12 − A012)− iAξ11(A1
11)−1A112 + iA
ξ12
−(λA021 + iAxi21)(A1
11)−1 0 −λ(A021(A1
11)−1A112 − A
022)− iAξ21(A1
11)−1A112 + iA
ξ22 + bξξ
−(b11)−1A121(A1
11)−1 (b11)−1 (b11)−1(A122 − A
121(A1
11)−1A112 − b
ξ)
. (2.17)
The full derivation of this system can be found in [2].
The system given by (2.17) is in Eulerian coordinates. It turns out that when computing
the Evans function in Eulerian coordinates the modulus and argument of the output changes
rapidly. This complicates calculations significantly when we use winding number computa-
tions on the values of the Evans function evaluated along a simple contour. We avoid this
issue by instead using pseudo-Lagrangian coordinates for which, as mentioned, the winding
of the Evans function is greatly reduced; for details see [3].
Using the work of [2], we define our 2-D Evans system under this formulation by first
setting r(λ, ξ) = |λ, ξ| and defining
f ] = F/r(λ, ξ), λ] = λ/r(λ, ξ), and ξ] = ξ/r(λ, ξ).
With these changes we arrive at the Evans system in pseudo-Lagrangian coordinates whichis given by W ′ = A](x1;λ, ξ)W , where
W =
f]u2
,A =
−r(λ]A0
11 + iAxi11)(A111)−1 0 −λ](A0
11(A111)−1A1
12 − A012)− iAξ
]
11(A111)−1A1
12 + iAξ]
12
−r(λ]A021 + iAxi21)(A1
11)−1 0 −λ](A021(A1
11)−1A112 − A
022)− iAξ
]
21(A111)−1A1
12 + iAξ]
22 + rbξ]ξ]
−r(b11)−1A121(A1
11)−1 r(b11)−1 (b11)−1(A122 − A
121(A1
11)−1A112 − rb
ξ] )
. (2.18)
13
Following the process explained above, we now build the integrated Evans function. We
define A]±(λ], ξ]) = limx1→∞A](x1;λ], ξ]), with bases {W+
1 , ...,W+k } and {W−
k+1, ...,W−N } as
described previously.
Taking the determinant of these bases evolved to x1 = 0 gives our integrated Evans
function
D(λ], ξ]) = det(W+1 , ...,W
+k ,W
−k+1, ...W
−N )|x1=0. (2.19)
Recall, we evolve the bases to x1 = 0 because due to the complexity of the system, in order
to determine linear independence of the two bases, we must evaluate the bases at the same x1
value. Whereas with the Lopatinski determinant we had a constant coefficient system which
allowed us to determine linear independence without evolving our bases. We also note that
we could alternatively evaluate the determinant at any value of x1. It is somewhat common
to instead evolve the solutions to x1 = −∞ or x1 = +∞.
Chapter 3. Numerical Stability Results
In this section we describe our study of the stability of traveling wave solutions of inviscid and
viscous models for 2-D MHD. More specifically, we seek to determine if there exist unstable
viscous waves with corresponding stable inviscid waves. Computationally, this occurs when
the Evans function has a real root while the Lopatinski determinant does not. We note that
the reverse relationship has already been proven, in other words, we know that instability in
the inviscid model implies instability in the viscous model.
Understanding this relationship between viscous and inviscid stability for 2-D MHD has
meaningful implications in mathematical modeling. At the present, it is significantly easier
to determine stability of an inviscid model than it is for a viscous model. As a result, there
is strong motivation to use the inviscid models rather than the viscous models, even though
we are not fully aware of how stability of these two models differ. In fact, there are some
14
applications that currently use an inviscid model to study a viscous system, justifying this
choice by the assumption that these two models have very similar stability properties and
that the inviscid model is much simpler computationally.
If we find that there exist unstable viscous waves with corresponding stable inviscid waves,
then when using an inviscid model to study a viscous system, in some cases the model will
assume stability where in reality the system in unstable which will lead to inaccurate results.
On the other hand, if unstable viscous waves always correspond to unstable inviscid waves,
then we can use inviscid models to study viscous systems more frequently and with more
confidence in the accuracy of results.
3.1 STABLAB
Throughout this chapter, we will use STABLAB to perform Lopatinski determinant and
Evans function computations. STABLAB is a numerical library available in Matlab and
Python. The library was originally designed for Evans function computation but has since
been expanded to provide additional computational tools for analyzing wave stability [4].
In our use of the code, we provide the profile system, Lopatinski system, and Evans system
as well as a contour bound r. STABLAB then numerically approximates the stability of the
inviscid and viscous models for the system. More specifically, using the contour radius bound
r, STABLAB computes the Lopatinski determinant and numerically approximates the Evans
function for values of λ along the computed contour.
As an example, the normalized contours for the Lopatinski determinant and Evans func-
tion associated with the system given by h1 = 0.75, u1+ = 1.3, Γ = 2/3, µ = 0.1, η = −2/3µ,
κ = 0.1 and cν = 1, are shown in Figure 3.1. In this example, both systems have a winding
number of zero and are thus stable. Note that the points along each contour indicate where
the functions were evaluated.
15
Figure 3.1: Plotted contours for Lopatinski determinant and Evans function as computedby STABLAB.
We strive to determine existence of unstable viscous waves with corresponding stable
inviscid waves in two ways. First, we search for roots of the Evans function and Lopatinski
determinant for a variety of parameter values. Second, we study the convexity of the curve
along which the roots of the Evans function bifurcate.
3.2 Stability Bifurcation Analysis
Recall that in our construction of the Evans system, we were careful to maintain analyticity
in our spectral parameter λ. As a result, we may use the argument principle to identify
unstable spectra in a bounded contour in the right half complex plane. This contour can
take a variety of shapes depending on the system but must be simple and connected.
As mentioned above, we define r as the radius of our bounded contour. We abuse the
word radius in this context as the contour is likely not circular. When we say a contour has
radius r this implies that the contour is contained inside the circle with radius r centered
at the origin. We compute r by curve fitting the Evans function to the known asymptotic
behavior iteratively, terminating when the curve fitting has relative error less than a certain
tolerance (0.2 in our case). Note that the asymptotic behavior is given by C1e√C2λ with
constants C1 and C2. With this method for choosing r, it is possible that our choice of r is
16
far larger than need be to locate roots. However, without a formal proof of a tighter bound
on r, we must use this method to ensure accuracy in our results.
We compute the stability of the inviscid and viscous systems for 50,000 different com-
binations of (h1, u1+, φ) satisfying h1 > 1, u1+ ∈ (0, 1), and φ > 0 where φ is a constant
indicating the viscosity level of the system. (In particular, we set ν = µ = φ.)
More specifically, we use STABLAB to compute the viscous and inviscid stability for
every combination (h1, u1+, φ) where h1 ranges from 1 to 4 incremented by 0.005, u1+ ranges
from 0 to 1 incremented by 0.005, and φ ranges from 0.01 to 0.2 incremented by 0.01. Some
of the results for φ = 0.1 are shown in Figure 3.2.
Figure 3.2: Stability results comparing the inviscid and viscous models as h1 and u1+ vary.
In Figure 3.2, the black line shows where the bifurcation occurs. Through all of our
experiments, we found that this line was exactly the same for the inviscid and viscous
systems for every choice of φ. Based on this result, it is likely that in 2-D MHD, viscous
instability does imply inviscid instability, though still not certainly.
3.3 Convexity of Bifurcation Curve
In the previous section we computed the winding number corresponding to the Evans function
for 50 different values of the Fourier coefficient ξ, ranging from ξ = 1e − 4 to ξ = 0.1, in
order to determine stability of a system characterized by a fixed h1, u1+, and φ. This was
17
necessary because the Evans function is not linear in ξ and thus we must do an exhaustive
search for the ξ value for which the Evans function has a root with positive real part. If we
do not find a root, we suspect the system is stable. As a result, in the previous section, it is
possible that we did not exhaust enough possible ξ values to definitely determine stability.
Note that the roots of the Lopatinski determinant are linear in the Fourier coefficient ξ
and thus we only need to preform computations for a single ξ value to determine stability
or instability, see [6]. In the previous section, we used ξ = 1.0.
Another way to study the relationship between inviscid and viscous stability, is to analyze
the roots of the system as parameterized by the Fourier coefficient ξ. Figure 3.3 gives an
example of what these parameterized curves of roots might look like. We know that initially,
the curve corresponding to the viscous roots will be tangent to that of the inviscid roots, as
is evident in all three figures; see [24].
(a) (b) (c)
Figure 3.3: Diagrams indicating possible relationships between roots of the inviscid systemand roots of the viscous system as parameterized by the Fourier coefficient ξ.
The situation depicted in Figure 3.3 (a) is what we typically expect to happen. In this
case, we see that as ξ increases, the roots of the viscous system stay negative. If we were
to consider a slightly different system where the line corresponding to the inviscid system
was located in the first quadrant, then we would see the roots of the viscous system starting
positive along with the inviscid, and then becoming negative for larger values of ξ.
18
The remaining two plots, (b) and (c) of Figure 3.3, display scenarios in which we might
see viscous instability and corresponding inviscid stability. In both cases, there are values
of ξ for which the viscous system has positive real roots while the inviscid system has only
negative real roots for all ξ. Another alternative is that we have a curve similar to the viscous
case in Figure 3.3 (c), but with the line curving back to the left half plane after crossing the
y-axis.
We seek to determine which of these cases is occurring in 2-D MHD. We do this by
calculating the roots of the Evans function for a fixed u1+ and φ, letting h1 and ξ vary. We
then study the convexity of the parameterized roots. The results are shown in Figure 3.4.
Figure 3.4: The real part of the roots of the viscous system as parameterized by h1 and theFourier coefficient ξ.
From Figure 3.4, it appears that Re(λ) varies nonlinearly in ξ and h1 though it is difficult
to see the exact relation. We can plot the cross sections of Figure 3.4 for fixed h1 and for
fixed ξ to better understand what is happening.
Figure 3.5 shows three different cross sections of Figure 3.4 for fixed values of ξ. By
plotting these cross sections we can better understand the relationship between h1 and
Re(λ). The dashed line plotted alongside the bifurcation curves serves to help determine
19
the convexity of the curves. In this case, we see that there is a quadratic relationship between
these two quantities. It is also relevant to note that the bifurcation curves are convex.
Figure 3.5: The real parts of the roots of the viscous system as parameterized by h1 for fixedvalues of ξ. The dotted line serves as a reference to highlight the convexity of the curves.
Similarly, Figure 3.6 shows three different cross sections of Figure 3.4 but this time for
fixed values of h1. This allows us to more closely examine the relationship between ξ and
Re(λ). We hope to use these results to identify which of the three cases from Figure 3.3 is
occurring in this system.
Figure 3.6: The real parts of the roots of the viscous system as parameterized by ξ for fixedvalues of h1.
As we can see, for each fixed h1, the bifurcation curve is convex and matches the far
left plot from Figure 3.3 This strengthens the possibility that viscous instability does in fact
imply inviscid instability.
20
It is also interesting to notice the linear black line plotted in Figure 3.4. This line
corresponds to where the system bifurcates and the roots transition to having nonzero real
part. The linearity of this line is surprising and suggests that there is an underlying linearity
when Re(λ) is parameterized by h1 and ξ simultaneously.
Chapter 4. Navier-Stokes
As mentioned, MHD waves consist of the Navier-Stokes equations and Maxwell’s equations
coupled together to describe the behavior of conducting fluids. As a result, analyzing the
Navier-Stokes equations can provide meaningful insight into the structure and behavior of
MHD waves. It should also be noted that there is significant relevance to the study of
Navier-Stokes outside of its applications to MHD waves.
In the remainder of this thesis, we describe the development of analytical and computa-
tional tools for determining the stability of traveling waves in the equations of compressible
Navier-Stokes. We do so because carrying out analysis of the compressible Navier-Stokes
equations is a first step to developing analysis for MHD. We now formally introduce the
system.
4.1 Compressible Navier-Stokes
The Navier-Stokes equations for a compressible gas in Lagrangian coordinates are given by,
vt − ux = 0,
ut + (p(v, T ))x =(µuxv
)x,
Et + (up(v, T ))x =(µuux
v
)x
+
(κTxv
)x
,
where E = e + u2/2 is the kinetic energy, e(v, T ) is the internal energy, u is velocity, v
is specific volume, p(v, T ) is the pressure law, and µ and κ are respectively coefficients of
21
viscosity and heat conductivity. The temperature, T , is proportional to internal energy,
e = cνT .
We are interested in analyzing the stability of viscous shock profiles v(x, t) = v(x− st),
u(x, t) = u(x− st), e(x, t) = e(x− st) of (4.1), which correspond to traveling waves moving
with constant speed s. Equivalently, the shock profiles are the stationary solutions of
vt − cvx − ux = 0,
ut − cux + (p(v, T ))x =(µuxv
)x, (4.1)
Et − cEx + (up(v, T ))x =(µuux
v
)x
+
(κTxv
)x
.
We use the scaling of [13], that is, we rescale by
(x, t, v, u, T )→ (−εcx, εc2t, v/ε,−u/(εc), T/(ε2c2))
and choose ε so that v− := limx→−∞ v(x) = 1, which yields
vt + vx − ux = 0
ut + ux + (p(v, T ))x =(µuxv
)x
Et + Ex + (up(v, T ))x =(µuux
v
)x
+
(κTxv
)x
,
where the new energy and pressure functions e and p relate to the old ones e0 and p0 by
e(v, T ) = e0(εv, ε2c2T )/(ε2c2), p(v, T ) = p0(εv, ε2c2T )/(εc2). (4.2)
Under this rescaling, the pressure and energy functions are unchanged up to a rescaling of
parameters for an ideal gas equation of state,
e(v, T ) = cνT, p(v, T ) =RT
v, R = gas constant. (4.3)
22
We solve for the profile system following the same procedure as described in Section 2.2.
The resulting profile solution satisfies the system,
v′ =1
µ[v(v − 1) + Γ(e− νe−)] , (4.4a)
e′ =v
ν
[−(v − 1)2
2+ e− e− + (v − 1)Γe−
]. (4.4b)
We denote the solutions to the profile system by u, v, and e. Then we linearize (4.1)
about the profile solutions to obtain,
λv + v′ − u′ = 0, (4.5a)
λu+ u′ +Γ
ve′ +
Γuxvu+
u
vv′ =
1
vu′′, (4.5b)
λe+ e′ +
[ux −
νuxxv
]u+
[Γe
v− (ν + 1)
uxv
]u′ +
[νexv2
]v′ =
ν
ve′′. (4.5c)
4.2 Energy Estimates
As mentioned, to determine stability, we evaluate the Lopatinski determinant and Evans
function over a contour containing the possible unstable spectra. Methods have been devel-
oped which can be used to bound the contour associated with the system, which reduces the
amount of required computation. Two of these methods are tracking estimates and energy
estimates.
Tracking estimates are based on the idea of analytically bounding the solutions to the
system (4.5) in the large λ limit. This method uses the requirement that any potential
eigenfunction must be tangent to the stable and unstable manifolds as x → ±∞. Using
this fact, one can bound with a cone, the solutions leaving each end state tangent to the
respective stable and unstable manifold. We then evolve these cones from each end state and
require that they intersect, meaning that an evolved solution could potentially live in both
cones. As we increase λ, eventually the cones will no longer intersect, giving an effective
bound on the admissible eigenvalues of the system.
23
Energy estimates, on the other hand, are a method based on the assumption that λ
is an eigenvalue of the system with the corresponding eigenfunction given by the vector
of functions (e.g. [v u e]T for compressible Navier-Stokes). Energy estimates get their
name because they consider the overall energy of the system by taking its integral and then
carefully manipulating the remaining equations to produce a constraint on λ.
For the compressible Navier-Stokes system in the strong shock limit, tracking estimates
where developed in [13], providing an initial bound on the contour, given in Table 4.1.
Γν
0.2 0.5 1.0 2.0 5.0
0.2 398.6 388.8 385.3 733.8 1755.6
0.4 211.7 182.3 175.1 325.0 762.0
0.6 222.5 123.4 111.5 198.4 449.8
0.667 226.8 114.3 100.4 175.3 391.3
0.8 236.5 103.9 85.3 142.6 307.1
1.0 253.8 100.8 73.7 113.8 229.2
1.2 274.3 106.6 69.7 98.7 183.1
1.4 300.8 117.5 70.5 91.6 154.9
1.6 347.4 131.8 74.5 89.8 138.0
1.8 397.3 148.7 80.8 91.8 128.6
2.0 450.3 167.5 88.7 96.7 124.7
Table 4.1: Bounds on Re(λ) + |Im(λ)|, where λ is an admissible eigenvalue, for variousvalues of Γ and ν, as given in [13].
In the following sections, we work toward improving this bound on both the real and
complex parts of λ for compressible Navier-Stokes in Lagrangian coordinates by implement-
ing energy estimates. We implement energy estimates to improve this bound on both the
real and complex parts of λ for compressible Navier-Stokes in Lagrangian coordinates. We
24
use the linearized eigenvalue system for compressible Navier-Stokes as derived in [13] and
given in (4.5). For simplicity, we write (4.5c) as
λe+ e′ + c1u+ c2u′ + c3v
′ =ν
ve′′.
4.3 Bounding Real Valued λ
We will first assume that λ ∈ R.
Lemma 4.1. For v and u satisfying (4.5) with λ ∈ R,
∫|v′|2 ≤
∫|u′|2.
Proof. First multiply (4.5a) by v′ and integrate to obtain
λ
∫vv′ +
∫v′v′ =
∫u′v′. (4.6)
Now take the real part of (4.6) and apply Young’s inequality to the right-hand side, which
yields
∫|v′|2 = Re
(∫u′v′)≤∣∣∣∣∫ u′v′
∣∣∣∣ ≤ ∫ |u′||v′|≤ 1
2
∫|u′|2 +
1
2
∫|v′|2.
Subtracting 12
∫|v′|2 from both sides and multiplying by 2 produces the desired inequality.
This lemma becomes essential in proving the following bound on real-valued λ.
Theorem 4.2. If λ ∈ R and (u, v, e) are an eigenvalue-eigenfunction solution to (4.5), then
λ ≤ max(βu, βe),
25
where
βu =
∣∣∣∣∣∣∣∣α1|c1|+1
2
(1
v
)′′− Γux
v+
Γ
4θ1v− u
4θ2v
∣∣∣∣∣∣∣∣∞,
βe =
∣∣∣∣∣∣∣∣ν2(
1
v
)′′+|c1|4α1
+|c2|4α2
+|c3|4α3
∣∣∣∣∣∣∣∣∞,
and α1, α2, α3, θ1, and θ2 are real-valued, positive constants that satisfy
∫θ1Γ
v|e′|2 ≤
∫ν
v|e′|2 and
∫ [α2|c2|+ α3|c3| −
θ2u
v
]|u′|2 ≤
∫1
v|u′|2.
Proof. The proof can be broken up into three main steps: (1) bounding λ∫|u|2, (2) bounding
λ∫|e|2, and (3) combining the resulting two inequalities.
(1) Start by multiplying (4.5b) by u and integrating to get
λ
∫|u|2 +
∫u′u+
∫Γ
ve′u+
∫Γuxv|u|2 +
∫u
vv′u =
∫1
vu′′u. (4.7)
Take the real part of (4.7)
λ
∫|u|2 +
∫Γuxv|u|2 +Re
(∫Γ
ve′u+
∫u
vv′u
)=
1
2
∫ (1
v
)′′|u|2 −
∫1
v|u′|2,
and apply Young’s inequality twice, using Lemma 4.1 and the fact that −Re(z) ≤ |z|,
λ
∫|u|2 +
∫1
v|u′|2 =
1
2
∫ (1
v
)′′|u|2 −
∫Γuxv|u|2 −Re
(∫Γ
ve′u+
∫u
vv′u
)λ
∫|u|2 +
∫1
v|u′|2 ≤ 1
2
∫ (1
v
)′′|u|2 −
∫Γuxv|u|2 +
∣∣∣∣∫ Γ
ve′u+
∫u
vv′u
∣∣∣∣≤ 1
2
∫ (1
v
)′′|u|2 −
∫Γuxv|u|2 +
∫Γ
v|e′||u|+
∫−uv|v′||u|
≤ 1
2
∫ (1
v
)′′|u|2 −
∫Γuxv|u|2 + θ1
∫Γ
v|e′|2 +
1
4θ1
∫Γ
v|u|2
+ θ2
∫−uv|v′|2 +
1
4θ2
∫−uv|u|2
≤ 1
2
∫ (1
v
)′′|u|2 −
∫Γuxv|u|2 + θ1
∫Γ
v|e′|2 +
1
4θ1
∫Γ
v|u|2
26
+ θ2
∫−uv|u′|2 +
1
4θ2
∫−uv|u|2
= θ2
∫−uv|u′|2 + θ1
∫Γ
v|e′|2 +
∫ [1
2
(1
v
)′′− Γux
v+
Γ
4θ1v− u
4θ2v
]|u|2.
(4.8)
(2) Now multiply (4.5c) by e then integrate and take the real part which yields
λ
∫|e|2 +Re
(∫c1ue+
∫c2u′e+
∫c3v′e
)=ν
2
∫ (1
v
)′′|e|2 −
∫ν
v|e′|2.
Apply Lemma 4.1 and Young’s inequality to get the following,
λ
∫|e|2 +
∫ν
v|e′|2 ≤ν
2
∫ (1
v
)′′|e|2 +
∣∣∣∣∫ c1ue+
∫c2u′e+
∫c3v′e
∣∣∣∣≤ν
2
∫ (1
v
)′′|e|2 +
∫|c1||u||e|+
∫|c2||u′||e|+
∫|c3||v′||e|
≤ν2
∫ (1
v
)′′|e|2 + α1
∫|c1||u|2 +
1
4α1
∫|c1||e|2 + α2
∫|c2||u′|2
+1
4α2
∫|c2||e|2 + α3
∫|c3||v′|2 +
1
4α3
∫|c3||e|2
≤∫
[α2|c2|+ α3|c3|] |u′|2 + α1
∫|c1||u|2
+
∫ [ν
2
(1
v
)′′+|c1|4α1
+|c2|4α2
+|c3|4α3
]|e|2. (4.9)
(3) Combine (4.8) and (4.9) to obtain the resulting inequality:
λ
∫ (|u|2 + |e|2
)+
∫1
v|u′|2 +
∫ν
v|e′|2 ≤
∫ [α2|c2|+ α3|c3| −
θ2u
v
]|u′|2 +
∫θ1Γ
v|e′|2
+
∫ [α1|c1|+
1
2
(1
v
)′′− Γux
v+
Γ
4θ1v− u
4θ2v
]|u|2
+
∫ [ν
2
(1
v
)′′+|c1|4α1
+|c2|4α2
+|c3|4α3
]|e|2. (4.10)
27
We need to specify θ1, θ2, α2, and α3 so that
∫θ1Γ
v|e′|2 ≤
∫ν
v|e′|2 and
∫ [α2|c2|+ α3|c3| −
θ2u
v
]|u′|2 ≤
∫1
v|u′|2.
We first set θ1 = ν/Γ to make the first inequality true. And notice
∫[α2|c2|v + α3|c3|v − θ2u]
1
v|u′|2 ≤ [α2||c2v||∞ + α3||c3v||∞ + θ2||u||∞]
∫1
v|u′|2.
So we choose
α2 =1
3||c2v||∞, α3 =
1
3||c3v||∞, and θ2 =
1
3||u||∞
to satisfy the second inequality. Now let
βu =
∣∣∣∣∣∣∣∣α1|c1|+1
2
(1
v
)′′− Γux
v+
Γ
4θ1v− u
4θ2v
∣∣∣∣∣∣∣∣∞
and
βe =
∣∣∣∣∣∣∣∣ν2(
1
v
)′′+|c1|4α1
+|c2|4α2
+|c3|4α3
∣∣∣∣∣∣∣∣∞.
Substituting our chosen parameter values into (4.10) yields
λ
∫ (|u|2 + |e|2
)≤ max(βu, βe)
∫ (|u|2 + |e|2
),
and thus
λ ≤ max(βu, βe).
The bound on λ for various values of ν and Γ are given in Table 4.2. A comparison
between these bounds and the original bounds developed for this system (Table 4.1) shows
that these bounds are significantly tighter. Though it should be noted that these bounds
apply to a real valued λ where the previous allow λ to be complex valued.
28
Γν
0.2 0.5 1.0 2.0 5.0
0.2 7.3653 7.0340 6.9262 6.8732 6.8402
0.4 4.3233 3.6037 3.3641 3.2449 3.1730
0.6 3.8725 2.7025 2.3126 2.1179 2.0011
0.667 3.9091 2.5759 2.1314 1.9097 1.7764
0.8 4.1387 2.4587 1.8990 1.6192 1.4513
1.0 4.7494 2.4994 1.7497 1.2748 1.1500
1.2 5.5806 2.7010 1.7409 1.2610 0.9732
1.4 6.5796 3.0100 1.8201 1.2251 0.8683
1.6 7.7201 3.4005 1.9606 1.2407 0.8088
1.8 8.9878 3.8582 2.1484 1.2935 0.7806
2.0 10.3741 4.3746 2.3748 1.3749 0.7750
Table 4.2: Numerically approximated bounds on real valued λ for various values of Γ and ν,computed based on Theorem 4.2.
4.4 Bounding Complex Valued λ
We will now get a similar bound on Re(λ) + |Im(λ)|. For simplification, throughout this
section we will define
Λ = Re(λ) + |Im(λ)|.
We start by establishing three necessary lemmas.
Lemma 4.3. If (v, u, e) is an eigenfunction of (4.5), then∫v′v,
∫u′u, and
∫e′e are purely
imaginary.
Proof. Let (v, u, e) be an eigenfunction of (4.5). Without loss of generality we will show that∫v′v is purely imaginary. This proof does not rely on any specific properties of v and thus
29
can also be applied to u and e. We start by noting that we have
∫v′v =
1
2
∫v′v +
1
2
∫v′v
=1
2
∫v′v −
∫vv′
=1
2
(∫v′v −
∫v′v
),
which implies that∫v′v is purely imaginary, as desired.
Lemma 4.4. If (v, u, e) is an eigenfunction of (4.5), then
Re[λ
∫uv′ +
∫u′v′]
=
∫|u′|2 − 2Re(λ)2
∫|v|2. (4.11)
Proof. Let (v, u, e) be an eigenfunction of (4.5). Then we have
λ
∫uv′ +
∫u′v′ = (λ+ λ)
∫uv −
∫u(λv′ + v′′)
= −2Re(λ)
∫(u′v −
∫uu′′
= −2Re(λ)
∫(λv + v′)v +
∫|u′|2.
Taking the real part we get that
Re[λ
∫uv′ +
∫u′v′]
=
∫|u′|2 − 2Re(λ)2
∫|v|2.
Lemma 4.5. If λ ∈ C and (u, v, e) are an eigenvalue-eigenfunction solution to (4.5) and
Re(λ) > δ for δ > 0, with δ, β1, and β2 satisfying δ − β1(Γ + 2)− 2β2 > 0, then
∫|v′|2 ≤ 1
ε1
[∫|u′|2 +
Γ + 2
4β1
∫|e′|2 +
Γ + 2
2Γβ2
∫v|u|2
]. (4.12)
30
Proof. First we multiply (4.5b) by v′ and integrate,
λ
∫uv′ +
∫u′v′ +
∫Γ
ve′v′ +
∫Γuxvuv′ +
∫u
v|v′|2 =
∫1
vu′′v′
=
∫1
v(λv′ + v′′)v′
= λ
∫1
v|v′|2 +
∫1
vv′′v′.
Now take the real part and apply (4.11),
∫|u′|2 − 2Re(λ)2
∫|v|2 +Re
[∫Γ
ve′v′ +
∫Γuxvuv′]
+
∫u
v|v′|2
= Re(λ)
∫1
v|v′|2 +
∫vx2v2|v′|2.
Rearranging yields
∫|u′|2 +Re
[∫Γ
ve′v′ +
∫Γuxvuv′]
=
∫ [Re(λ) +
vx2v− u]
1
v|v′|2 + 2Re(λ)2
∫|v|2.
Notice
vx2v− u =
1
2u+
Γv
v− u =
Γe
v− u
2≥ 0 and
Γ + 2
Γ≥ 1
v≥ 1,
and thus
∫ [Re(λ) +
vx2v− u]
1
v|v′|2 + 2Re(λ)2
∫|v|2 ≥ Re(λ)
∫1
v|v′|2 ≥ Re(λ)
∫|v′|2.
Putting this together and applying Young’s inequality we get
Re(λ)
∫|v′|2 ≤
∫|u′|2 +Re
[∫Γ
ve′v′ +
∫Γuxvuv′]
≤∫|u′|2 +
∣∣∣∣∫ Γ
ve′v′ +
∫Γuxvuv′∣∣∣∣
31
≤∫|u′|2 +
∣∣∣∣∫ Γ
ve′v′∣∣∣∣+
∣∣∣∣∫ Γuxvuv′∣∣∣∣
≤∫|u′|2 +
∫Γ
v|e′||v′|+
∫|Γuxv||u||v′|
≤∫|u′|2 + β1
∫Γ
v|v′|2 +
1
4β1
∫Γ
v|e′|2 + β2
∫|Γuxv||v′|2 +
1
4β2
∫|Γuxv||u|2
≤∫|u′|2 + β1(Γ + 2)
∫|v′|2 +
Γ + 2
4β1
∫|e′|2 + 2β2
∫|v′|2 +
1
2β2
∫|u|2
=
∫|u′|2 + [β1(Γ + 2) + 2β2]
∫|v′|2 +
Γ + 2
4β1
∫|e′|2 +
1
2β2
∫|u|2
=
∫|u′|2 + [β1(Γ + 2) + 2β2]
∫|v′|2 +
Γ + 2
4β1
∫|e′|2 +
1
2β2
∫1
vv|u|2
≤∫|u′|2 + [β1(Γ + 2) + 2β2]
∫|v′|2 +
Γ + 2
4β1
∫|e′|2 +
1
2β2
Γ + 2
Γ
∫v|u|2
=
∫|u′|2 + [β1(Γ + 2) + 2β2]
∫|v′|2 +
Γ + 2
4β1
∫|e′|2 +
Γ + 2
2Γβ2
∫v|u|2.
Now we will require that Re(λ) ≥ δ for some δ > 0 and we will let ε1 = δ−β1(Γ+2)−2β2
and require that β1 and β2 are chosen so that ε1 > 0. With these requirements we get that
∫|v′|2 ≤ 1
ε1
[∫|u′|2 +
Γ + 2
4β1
∫|e′|2 +
Γ + 2
2Γβ2
∫v|u|2
]
With (4.11) and Lemma 4.5, we are ready for the main theorem of this section.
Theorem 4.6. If λ ∈ C and (u, v, e) are an eigenvalue-eigenfunction solution to (4.5) and
Re(λ) > δ for δ > 0, then
Λ ≤ max(uval, eval), (4.13)
where
uval =
∣∣∣∣1− 2Γ
Γ + 2
∣∣∣∣+
√2(Γ + 2)
4θ1
+2√
2
4Γθ2
+1
4θ3
+√
2α1‖c1‖∞ +(Γ + 2)ε2
2Γβ2ε1
,
eval =1
Γ + 2+
√2‖c1‖∞4α1
+
√2(ν + 1)
2α2Γ+
√2
2α3Γ2+
1
4α4
,
32
and δ, β1, β2, θ1, θ2, θ3, α2, α3, and α4 are positive valued constants satisfying
δ > β1(Γ + 2) + 2β2, θ3 +2√
2α2(ν + 1)
Γ + 2+ε2
ε1
≤ 1, and√
2θ1Γ + α4 +(Γ + 2)ε2
4β1ε1
≤ ν.
Proof. This proof can be broken up into three main steps: (1) bounding Λ∫|u|2, (2) bound-
ing Λ∫|e|2, and (3) combining the four resulting inequalities.
(1) First multiply (4.5b) by vu and integrate, applying integration by parts to the right-
hand side which yields
λ
∫v|u|2 +
∫vu′u+
∫Γe′u+
∫Γux|u|2 +
∫uv′u =
∫u′′u = −
∫|u′|2. (4.14)
Now take just the real part of (4.14) to obtain
Re(λ)
∫v|u|2 − 1
2
∫vx|u|2 +Re
[∫Γe′u
]+
∫Γux|u|2 +Re
[∫uv′u
]= −
∫|u′|2,
and simplify, using the fact that ux = vx which yields
Re(λ)
∫v|u|2 +
∫|u′|2 =
(1
2− Γ
)∫vx|u|2 −Re
[∫Γe′u
]−Re
[∫uv′u
]=
(1
2− Γ
)∫vxvv|u|2 −Re
[∫Γe′u
]−Re
[∫uv′u
].
Taking the absolute value of each side produces
∣∣∣∣Re(λ)
∫v|u|2 +
∫|u′|2
∣∣∣∣ =
∣∣∣∣(1
2− Γ
)∫vxvv|u|2 −Re
[∫Γe′u
]−Re
[∫uv′u
]∣∣∣∣Re(λ)
∫v|u|2 +
∫|u′|2 ≤
∣∣∣∣12 − Γ
∣∣∣∣ ∫ ∣∣∣∣ vxv∣∣∣∣ v|u|2 +
∣∣∣∣Re [∫ Γe′u
]∣∣∣∣+
∣∣∣∣Re [∫ uv′u
]∣∣∣∣ .Now we take the imaginary part of (4.14) to obtain
Im(λ)
∫v|u|2 + Im
[∫vu′u
]+ Im
[∫Γe′u
]+ Im
[∫uv′u
]= 0.
33
Again, rearrange and take the absolute value of both sides, producing
|Im(λ)|∫v|u|2 =
∣∣∣∣Im [∫ vu′u
]+ Im
[∫Γe′u
]+ Im
[∫uv′u
]∣∣∣∣≤∣∣∣∣Im [∫ vu′u
]∣∣∣∣+
∣∣∣∣Im [∫ Γe′u
]∣∣∣∣+
∣∣∣∣Im [∫ uv′u
]∣∣∣∣ .Now we can combine the two resulting inequalities and use the fact that for z ∈ C, |Re(z)|+
|Im(z)| ≤√
2|z|, which yields
Λ
∫v|u|2 +
∫|u′|2 ≤
∣∣∣∣12 − Γ
∣∣∣∣ ∫ ∣∣∣∣ vxv∣∣∣∣ v|u|2 +
√2
∣∣∣∣∫ Γe′u
∣∣∣∣+√
2
∣∣∣∣∫ uv′u
∣∣∣∣+
∣∣∣∣∫ vu′u
∣∣∣∣≤∣∣∣∣12 − Γ
∣∣∣∣ ∫ ∣∣∣∣ vxv∣∣∣∣ v|u|2 +
√2
∫Γ|e′||u|+
√2
∫|u||v′||u|+
∫v|u′||u|
Applying Young’s inequality three times gives us
Λ
∫v|u|2 +
∫|u′|2 ≤
∣∣∣∣12 − Γ
∣∣∣∣ ∫ ∣∣∣∣ vxv∣∣∣∣ v|u|2 +
√2θ1
∫Γ|e′|2 +
√2
4θ1
∫Γ|u|2
+√
2θ2
∫|u||v′|2 +
√2
4θ2
∫|u||u|2 + θ3
∫v|u′|2 +
1
4θ3
∫v|u|2
=
∣∣∣∣12 − Γ
∣∣∣∣ ∫ ∣∣∣∣ vxv∣∣∣∣ v|u|2 +
√2θ1
∫Γ|e′|2 +
√2
4θ1
∫Γ
vv|u|2
+√
2θ2
∫|u||v′|2 +
√2
4θ2
∫|u|vv|u|2 + θ3
∫v|u′|2 +
1
4θ3
∫v|u|2
≤∣∣∣∣12 − Γ
∣∣∣∣ 2
Γ + 2
∫v|u|2 +
√2θ1Γ
∫|e′|2 +
√2Γ
4θ1
Γ + 2
Γ
∫v|u|2
+√
2θ22
Γ + 2
∫|v′|2 +
√2
4θ2
2
Γ
∫v|u|2 + θ3
∫|u′|2 +
1
4θ3
∫v|u|2
=
[∣∣∣∣1− 2Γ
Γ + 2
∣∣∣∣+
√2(Γ + 2)
4θ1
+2√
2
4Γθ2
+1
4θ3
]∫v|u|2
+√
2θ1Γ
∫|e′|2 +
2√
2θ2
Γ + 2
∫|v′|2 + θ3
∫|u′|2.
34
(2) Now we multiply (4.5c) by ve and integrate, applying integration by parts to the
right-hand side as before, to obtain
λ
∫v|e|2 +
∫ve′e+
∫c1vue+
∫c2vu
′e+
∫c3vv
′e =
∫νe′′e = −
∫ν|e′|2. (4.15)
Now we will take just the real part of (4.15), which yields
Re(λ)
∫v|e|2− 1
2
∫vx|e|2 +Re
[∫c1vue
]+Re
[∫c2vu
′e
]+Re
[∫c3vv
′e
]= −
∫ν|e′|2,
which we can rearrange and take the absolute value, resulting in
Re(λ)
∫v|e|2 +
∫ν|e′|2 =
1
2
∫vx|e|2 −Re
[∫c1vue
]−Re
[∫c2vu
′e
]−Re
[∫c3vv
′e
]Re(λ)
∫v|e|2 +
∫ν|e′|2 =
∣∣∣∣12∫vx|e|2 −Re
[∫c1vue
]−Re
[∫c2vu
′e
]−Re
[∫c3vv
′e
]∣∣∣∣≤1
2
∫|vx||e|2 +
∣∣∣∣Re [∫ c1vue
]∣∣∣∣+
∣∣∣∣Re [∫ c2vu′e
]∣∣∣∣+
∣∣∣∣Re [∫ c3vv′e
]∣∣∣∣ .
Take the imaginary part of (4.15) provides
Im(λ)
∫v|e|2 + Im
[∫ve′e
]+ Im
[∫c1vue
]+ Im
[∫c2vu
′e
]+ Im
[∫c3vv
′e
]= 0,
which again we can rearrange and take the absolute value of both sides to get
|Im(λ)|∫v|e|2 =
∣∣∣∣Im [∫ ve′e
]+ Im
[∫c1vue
]+ Im
[∫c2vu
′e
]+ Im
[∫c3vv
′e
]∣∣∣∣≤∣∣∣∣Im [∫ ve′e
]∣∣∣∣+
∣∣∣∣Im [∫ c1vue
]∣∣∣∣+
∣∣∣∣Im [∫ c2vu′e
]∣∣∣∣+
∣∣∣∣Im [∫ c3vv′e
]∣∣∣∣ .
35
Now we add the inequalities corresponding to the real and imaginary parts of (4.15) to
Table 4.3: Approximated bounds on Re(λ) + |Im(λ)| for all λ with Re(λ) > 5.7. Thesebounds were computed using Algorithm 1 which is based on Theorem 4.6.
Chapter 5. Analytic Stability
In this section we introduce work on a method to analytically show the stability of traveling
wave solutions using the Evans function. We first motivate this method with a rudimentary
example, then describe the method and its results for compressible Navier-Stokes in the
strong shock limit given in Lagrangian coordinates. We end by applying the method to 2-D
MHD.
41
5.1 Motivating Example
In order to introduce they key ideas for our method, we start with a simple toy problem that
displays the potential success of this method.
Consider an Evans ODE system given by
W ′(x;λ) = A(x;λ)W (x;λ),
where ′ = d/dx, and
A(x;λ) =
3v λ 1
0 −v λ
0 0 −1− v
.
We consider v to be the traveling wave solution for our example system defined by v(x) =
tanh(x) and we define v± = limx→±∞ v(x). We make the coordinate change W (x;λ) =
e−µxV (x;λ) where µ is an eigenvalue of A± := limx→±∞A(x;λ). The resulting system is
given by
V ′(x;λ) = (A(x;λ)− µI)V (x;λ).
We make another coordinate change by defining σ = v − v+ = v − 1. This gives rise to the
new system
(−2σ − σ2)dZ(σ)
dσ= B(σ;λ)Z(σ),
where
B(σ;λ) =
3 + 3σ − µ λ 1
0 −σ − 1− µ λ
0 0 −2− σ − µ
.
We seek to solve this new system by finding a series solution for Z(σ). This is done by
letting Z(σ) =∑∞
n=0Bnσn and then solving the system to find each coefficient vector Bn.
42
After solving this system we get that for a given eigenpair (α, vα), of the matrix A+,
B0 = vα,
B1 =
5− α λ 1
0 1− α λ
0 0 −α
−1−3 0 0
0 1 0
0 0 1
B0,
Bn =
3− α + 2n λ 1
0 2n− 1− α λ
0 0 2n− 2− α
−1−2− n 0 0
0 2− n 0
0 0 2− n
Bn−1,
for n ≥ 2. It is important to mention that we can choose any eigenvector of A+ to define
B0.
To check the relevance of our solution, we compare the norms of each successive coefficient
vector Bn. Figure 5.1 shows the trend of these norms for various values of λ, with the second
most negative eigenvalue of A+ being chosen for α.
Figure 5.1: The convergence of the norms of the coefficient vectors corresponding to theexample system.
As seen in the graph, in all cases the norms converge exponentially as n increases. This
43
is a very positive result because it gives reason to believe that our solution will not blow up
when |σ| ≥ 1.
5.2 Analytic Profile Solution
Notice that in the motivating example, the entire profile system (v(x), e(x)) was expressed
in terms of our new coordinate σ. In order to analytically compute the Evans function for
compressible Navier-Stokes, we will similarly need the profile solutions to be expressed as
functions of σ.
Recall that the profile for the ideal gas case of compressible Navier-Stokes is equivalent
to the profile for MHD (as mentioned in Section 2.2), though it should be noted that the
two systems have different parameters. Additionally, we multiply the profile system by v to
transform to Lagrangian coordinates and further simplify by substituting e− = 0. With this,
our profile system is given by
v′ =1
µ[v(v − 1) + Γe] (5.1a)
e′ =v
ν
[−(v − 1)2
2+ e
], (5.1b)
where v and e are both functions of x. The Jacobian of this system is given by
J =
1µ(2v − 1) Γ
1ν
(e− (v−1)2
2
)− v
ν(v − 1) v
ν
.
This system has two fixed points corresponding to x going to negative and positive infinity.
These are given by (v−, e−) = (1, 0) and (v+, e+) =(
ΓΓ+2
, 2(Γ+2)2
).
Observe the stability of the solutions around the fixed point (v−, e−), found using the
Jacobian,
J(v−, e−) =
1µ
Γ
0 1ν
.
44
This system has eigenvalues λ = 1/µ and λ = 1/ν from which we know this fixed point
is unstable because µ, ν > 0. As a result, we will use the fixed point (v+, e+) to solve the
system.
We seek to find a series solution by making the coordinate change given by
σ = v(x)− v+.
We choose this to be our coordinate change because σ = 0 initializes our solutions at the
stable fixed point. We hope that by starting at this fixed point, we will be able to evolve the
solution in the stable eigendirection across our entire domain to the unstable fixed point.
Notice that with this coordinate change we have
∂e
∂x=∂e
∂σ
∂σ
∂x,
where
∂σ
∂x=∂v
∂x.
This gives us the following
v
ν
[−(v − 1)2
2+ e
]=∂e
∂σ
1
µ[v(v − 1) + Γe] . (5.2)
We can now rewrite v in terms of σ as
v = σ + v+ (5.3)
and we can write e as a series expansion given by
e =∞∑n=0
anσn. (5.4)
45
Substituting (5.3) and (5.4) into (5.2), results in
(σ + v+)ν
[−(σ + v+ − 1)2
2+∞∑n=0
anσn
]=
∞∑n=1
nanσn−1 1
µ
[(σ + v+)(σ + v+ − 1) + Γ
∞∑n=0
anσn
]. (5.5)
We solve (5.5) to find the values of the coefficients an,
a0 = e+
a1 =1
2A(−B −
√B2 − 4AC)
a2 =a1µ− a1ν − 3
2µv+ + µ
3Γa1ν − µv+ + 4νv+ − 2ν(5.6)
a3 =−2Γa2
2ν + a2µ− 2a2ν − 12µ
4Γa1ν − µv+ + 6νv+ − 3ν
an =(µ− (n− 1)ν)an−1 − νΓ
(∑n−1r=2 raran−r+1
)nν(Γ−2)−µΓ
Γ+2+ (n+ 1)νΓa1
,
where A = νΓ, B = 2νv+ − ν − µv+, and C = −e+µ+ 32µv2
+ − 2µv+ + 12µ.
This gives us a series solution for e(σ). We check our accuracy by comparing this plot to
the numerical solution found using STABLAB. The results are shown below.
Figure 5.2: The series solution of e(σ) with 100 terms plotted against σ.
46
Figure 5.3: The series solution of e(σ) with five terms plotted against σ.
Notice that even with just five terms in the series, this approximation does very well
against the numerical solution. We can verify that the solution converges by creating a plot
of the series coefficients analogous to Figure 5.1. Figure 5.4 shows the trend in coefficients for
the analytic profile solution. Unlike the example system, these coefficients do not converge
to zero. Fortunately σ ∈ [0, 1) on our domain and the coefficients grow slower than σn for
all σ ∈ [0, 1). As a result, our series solution does not diverge and evolves to the unstable
fixed point.
Figure 5.4: The absolute values of the coefficients |an| of e(σ).
.
It is important to address why this coordinate change does not affect the behavior of the
profile system. This is justified in the following theorem.
47
Theorem 5.1. [11] Let v(x) and e(x) be the solutions to the profile system (2.7). The
function v(x) is monotone in x.
Proof. Recall our profile system (5.1) is given by
v′ =1
µ[v(v − 1) + Γe]
e′ =v
ν
[−(v − 1)2
2+ e
],
with corresponding nullclines e = −v(v−1)Γ
(corresponding to v′ = 0) and e = 12(v − 1)2
(corresponding to e′ = 0). We have already found that these nullclines intersect at the fixed
points (v+, e+) =(
ΓΓ+2
, 2(Γ+2)2
)and (v−, e−) = (1, 0). Now to show that v(x) is monotone,
it suffices to show that the solution v(x) remains strictly below the nullcline given by v′ = 0
from x = −∞ to x = +∞. This can be done by showing that the upper nullcline creates
a trapping region for solutions which implies that solutions that start below the nullcline
cannot cross the nullcline; see [11] for details.
Figure 5.5: Nullclines for the profile system (5.1) corresponding to v′ = 0 and e′ = 0, whichintersect at the fixed points of the system.
To do this, we will look at e′ along the line e = −v(v−1)Γ
. So we have
e′ =v
ν
(−(v − 1)2
2− v(v − 1)
Γ
).
48
Or equivalently
e′ =−v(v − 1)
ν
(v − 1
2+v
Γ
).
Notice that for v ∈ (v+, v−) we have v < 1 and v > ΓΓ+2
. So
v <Γ
Γ + 2
=⇒ Γv + 2v < Γ
=⇒ Γv + 2v − Γ < 0
=⇒ Γv + 2v − Γ
2Γ< 0
=⇒ Γ(v − 1) + 2v
2Γ< 0
=⇒ v − 1
2+v
Γ< 0
Thus v > 0, v − 1 < 0 and v−12
+ vΓ< 0. Which implies that along this curve, e′ < 0 as
desired. Thus our solution lies strictly below this nullcline and thus v′ < 0 along our solution
which implies that v(x) is monotone decreasing.
5.3 Analytic Evans System
We can express (4.5) as the first order system
W ′(x;λ) = A(x;λ)W (x;λ), (5.7)
with
A(x;λ) =
0 1 0 0 0
λν−1v ν−1v ν−1vux − uxx λg(U) g(U)− h(U)
0 0 0 λ 1
0 0 0 0 1
0 Γ λv + Γux λv f(U)− λ
, (5.8)
49
W = (e, e′, u, v, v′)T , ′ = d/dx. (5.9)
We note that u = v − 1 and vx = ux, where vx and ex are given by (5.1) and we define
g(U) = ν−1(Γe− (ν + 1)ux),
f(U) =ux − Γe
v+ v = 2v − 1− Γe−,
h(U) = − exv
= −ν−1
(−(v − 1)2
2+ (e− e−) + (v − 1)Γe−
).
The quantities v and u are the traveling wave solutions of the Navier-Stokes equations for
compressible gas. As before, we now define v± := limx→±∞ v(x) and e± := limx→±∞ e(x). In
this case we have
v+ =Γ
Γ + 2, e+ =
2
(Γ + 2)2, v− = 1, and e− = 0.
Similarly we compute A± := limx→±∞A(x;λ).
A+(λ) =
0 1 0 0 0
λν−1v+ ν−1v+ 0 λν−1Γe+ ν−1Γe+
0 0 0 λ 1
0 0 0 0 1
0 Γ λv+ λv+ 2v+ − 1− λ
, (5.10)
A−(λ) =
0 1 0 0 0
λν−1 ν−1 0 0 0
0 0 0 λ 1
0 0 0 0 1
0 Γ λ λ 1− λ
. (5.11)
50
Following the same procedure as in the example, we make the coordinate changeW (x;λ) =
eαxV (x;λ) where α is an eigenvalue of A± := limx→±∞A(x;λ).
V ′(x;λ) = (A(x;λ)− αI)V (x;λ). (5.12)
We also make the coordinate change σ = v − v+. This gives rise to the new system
1
µ((σ + v+)(σ + v+ − 1) + Γe(σ))
dZ(σ)
dσ= B(σ;λ)Z(σ), (5.13)
with
B(σ;λ) =
−α 1 0 0 0
λ(σ+v+)ν
(σ+v+)ν− α (σ+v+)uσ
ν− uσσ λg(σ) g(σ)− h(σ)
0 0 −α λ 1
0 0 0 −α 1
0 Γ λ(σ + v+) + Γuσ λ(σ + v+) f(σ)− λ− α
, (5.14)
where
e(σ) =∞∑n=0
anσn as defined in Section 5.2,
uσ =1
µ((σ + v+)(σ + v+ + 1) + Γe(σ)),
eσ =σ + v+
ν
(−(σ + v+ − 1)2
2+ e(σ)
),
uσσ =1
µ((2(σ + v+)− 1)uσ + Γeσ) ,
g(σ) =1
ν(Γe(σ)− (ν + 1)uσ),
f(σ) = 2(σ + v+)− 1,
h(σ) =−1
ν
(−(σ + v+ − 1)2
2+ e(σ)
).
51
We seek to solve this new system to find a series solution to Z(σ). This is done by letting
Z(σ) =∑∞
n=0 Bnσn and then solving the system to find each coefficient vector Bn. Using
SymPy, we have developed a generalized linear solver which computes the coefficient vectors
Bn, the code for this solver can be found in Appendix A.
After using the solver to compute the coefficient vectors, we can look at our results for
various values of λ. We chose λ = 10. With this choice, we find the eigenvalues of A+ to
be 9.692,−10.565,−5.119, 2.2208,−1.775. For any of the negative eigenvalues, α, of A+, we
can initialize our series solution with the eigenpair (α, vα) where vα is the eigenvalue of A+
associated with α. Initializing the solution with an eigenpair is equivalent to setting B0 = vα
and using α as our scale when defining B(σ;λ) as in (5.14).
When solving this system with α = −10.565, we get a stable solution all the way from
x = −∞ to x = +∞ or equivalently from σ = v− − v+ to σ = v+ − v+ = 0. The solution is
shown in Figure 5.6, where the asterisks correspond to the eigenvector vα.
Figure 5.6: The fifteen-term series solution to the Evans system with λ = 10, scaled out bythe most negative eigenvalue of A+.
Our fifteen-term series solution with this eigenpair is very similar to the numerical solution
generated by STABLAB, which is quite promising. However, if we consider the relevance by
examining the norms of each successive coefficient vector Bn, as was done in Figure 5.1, we
see that the coefficients grow quite quickly, shown in Figure 5.7.
52
Figure 5.7: The trend in the norms of the coefficient vector for the series solution Z(σ).
From Figure 5.7, it is not surprising that as more terms are added to the series, the
solution diverges from the actual Evans system for x in the unstable manifold. As a result,
the solution can only be used from x = +∞ to x = 0.
We tried the same approach to solve the system scaled out by the other two negative
eigenvalues with less positive results, as shown below. In both cases, the solutions blew up
before reaching the fixed point at x = −∞. But we were able to get a system that stayed
reasonably bounded for part of the domain from x = +∞ to x = 0 for both additional
negative eigenvalues. Figures 5.8 and 5.9 show the evolved solutions from x = −∞ to x = 0
scaling the system by the second and third most negative eigenvalues respectively. Notice
for the third most negative eigenvalue, shown in Figure 5.9, the solution is unstable before
evolving to x = 0.
53
Figure 5.8: The fifteen-term series solution to the Evans system with λ = 10, scaled out bythe second most negative eigenvalue of A+.
Figure 5.9: The fifteen-term series solution to the Evans system with λ = 10, scaled out bythe third most negative eigenvalue of A+.
Recall that in order to compute the Evans function, we must take the determinant of the
stable eigenvectors from both the unstable and stable manifolds, evaluated at x = 0. With
the results above, we will not be able to evolve the entire basis to the center of our domain.
In the next section, we introduce an alternative method that will only require one solution
be evolved across the domain.
5.4 Compound Matrix Method
Because our previous approach was unsuccessful in evolving the entire basis from the stable
manifold to the center of our domain, we try a new approach. Instead of finding five basis
54
elements for the Evans function, we lift the system to ten dimensions and find one stable
solution from the stable manifold to the unstable manifold. Recall that our goal is to
determine if there is an eigenfunction for a particular potentially unstable eigenvalue. When
we lift our system using the complex matrix method, one single exterior product encodes
the information for the entire stable basis for the non-lifted stable manifold. As a result, if
we are able to evolve one exterior product across our entire domain, we can then evaluate
the Wronskian at x = −∞ to determine linear independence of our bases which in turn
determines the existence of an eigenfunction; see [1] for details.
We denote the matrix A(x;λ) lifted into∧2(C4) by A(2)(x;λ). Using properties of the
exterior product, we can find the leading eigenvalues and eigenvectors for this matrix at
x = +∞ and x = −∞. Let α1, α2, α3 be the three most negative eigenvalues of A+ with
corresponding eigenvectors v1, v2, v3. Then α1 + α2 + α3 is the most negative eigenvalue for
A(2)+ with corresponding eigenvector v1
∧v2
∧v3. We can use this same technique to find the
most negative eigenvalue and eigenvector for the lifted system at negative infinity A(2)− .
Now we will introduce a different coordinate change to solve for solution using the lifted
system A(2). The hope is that we can find a solution that is stable all the way from x = +∞
to x = −∞. If we are able to make it all the way from the stable manifold to the unstable
manifold, we avoid the issue of not being able to get a stable solution starting at x = −∞.
We make the coordinate change W (x;λ) = eg(x)V (x;λ) for some function g(x). This
gives rise to the new system.
V ′(x;λ) = (A(2) − g′(x))V (x;λ) (5.15)
Next, we make the familiar coordinate change σ = v − v+, which gives us
1
µ((σ + v+)(σ + v+ − 1) + Γe(σ))
dZ(σ)
dσ= B(2)(σ;λ)Z(σ), (5.16)
where B(2)(σ;λ) = A(2)(σ;λ)− f ′(σ) and f ′(σ) = g′(x).
55
To control our solution we want to choose f ′(σ) so that at both x = +∞ and x = −∞
(or equivalently at σ = 0 and σ = v− − v+) the solution stays stable. One way to do this is
by choosing f ′(σ) so that at x = ±∞, f ′(σ) scales the solutions out by the corresponding
dominating eigenvalues. Because we are going from positive infinity to negative infinity, the
dominating eigenvalue will be the most negative in both cases. Let α+ be the most negative
eigenvalue for A(2)+ and α− be the most negative eigenvalue for A
(2)− . Then set
f ′(σ) = α+ + (α− − α+)σ
v− − v+
.
Notice f ′(0) = α+ and f ′(v− − v+) = α− as desired.
Using this choice of f ′(σ) we define Z(σ) =∑∞
n=0Bnσn and solve for each coefficient
vector Bn with positive results. Figure 5.10 shows the evolved fifteen-term lifted system
solution corresponding to λ = 10. As you can see, the solution stays stable across the entire
domain and thus can be used to compute the Evans function. Note that we initialize our
solution with the eigenvector of A(2)+ corresponding the most negative eigenvector. That is,
we set B0 = vα, where vα is the eigenvector of A(2)+ corresponding to α which is the most
negative eigenvalue of A(2)+ .
Figure 5.10: The fifteen-term series solution to the lifted Evans system with λ = 10, scaledby f ′(σ).
56
We now want to test if this method is stable for all relevant values of λ on the contour
bounding unstable spectra. Using Algorithm 1, we can bound Λ for this system. With
δ = 0.001, meaning for all λ with Re(λ) > 0.001, we compute that Λ ≤ 54.27.
Figure 5.11 demonstrates the stability of the system for λ satisfying Λ = 54.27 and
Re(λ) ≥ 0.001. The figure on the left shows the complex values of λ for which the system is
evaluated. The figure on the right shows the value of the evolved eigenvector at the unstable
manifold for values of λ along the contour. Notice that the contour shown on the left is
parameterized by θ such that
tan(θ) =Re(λ)
Im(λ). (5.17)
Figure 5.11: The analytic solution to the Evans system evaluated on the unstable manifoldfor various λ values satisfying Λ = 54.27, where λ is parameterized by θ according to (5.17).
From Figure 5.11, we can conclude that the series solution is stable across the entire
domain for all λ values in the bounded unstable contour. With this result, we can now
use the series solution to evaluate the Evans function. We further confirm this by ensuring
the series solution converges all along the specified contour. This is done by comparing the
relative error between consecutive series solutions evaluated at the unstable manifold. More
precisely, comparing
||Zm(v− − v+)− Zm−1(v− − v+)||2||Zm(v− − v+)||2
,
as m → ∞ where Zm(v− − v+) =∑m
n=0Bn(v− − v+)n. The resulting norms for various λ
57
values satisfying Λ = 54.27 are shown in Figure 5.12.As you can see, the relative error is
order O(1/n), where n is the number of terms in the series.
Figure 5.12: Successive relative errors for finite series solution of the lifted Evans system forvarious values of λ, satisfying Re(λ) > 0.001 and Λ = 54.27. Note that λ is parameterizedby θ according to (5.17). The dotted line, included for reference, plots y = 1/n.
With the convergence of relative error and stability for λ all along our contour bounded
by Theorem 4.6, we have shown that this method for analytically computing the Evans
function has significant potential, especially when applied to the Navier-Stokes equations.
In the next section, we study this method for 2-D MHD.
5.5 Extending to MHD
We now extend our method to 2-D MHD. Again we start by introducing the coordinate
change σ = u1 − u1+, with u1 = σ + u1+ and e =∑∞
n=0 anσn. Note that due to the
equivalence of the profile systems of Navier-Stokes and MHD, we have the same fixed points,
(u1−, e−) = (1, 0) and (u1+, e+) = ( ΓΓ+2
, 2(Γ+2)2 ). And again, from the Jacobian we determine
that (u1−, e−) is unstable and (u1+, e+) is stable, which is why we center our coordinate
change around the stable fixed point.
Using the relation 5.2 we again seek to solve
cν(σ + u1+)
κ
[∞∑n=0
anσn − e− −
(σ + u1+ − 1)2
2+ (σ + u1+ − 1)Γe−
]=
58
∞∑n=1
nanσn−1 1
2µ+ η
[(σ + u1+)(σ + u1+ − 1) + Γ
(∞∑n=0
anσn − e−
)](5.18)
for the coefficients an. Notice that we have transformed our system into pseudo-Lagrangian-
coordinates by multiplying by u1. Solving for the coefficients results in a recurrence relation
nearly identical to (5.6). It is no surprise that this series solution for the MHD profile
converges and agrees well with the numerical solution with very few terms. Figure 5.13
shows the analytic series compared with the numerical solution for the MHD profile.
Figure 5.13: The fifteen-term series solution of the MHD profile e(σ) plotted against σ.
Using this profile, we are now prepared to build our Evans system. We will use the Evans
system in pseudo-Lagrangian coordinates given in Section 2.4 and repeated below
W ′ = A](x1;λ, ξ)W ,
where
A]
=
−r(λ]A0
11 + iAxi11)(A111)−1 0 −λ](A0
11(A111)−1A1
12 − A012)− iAξ
]
11(A111)−1A1
12 + iAξ]
12
−r(λ]A021 + iAxi21)(A1
11)−1 0 −λ](A021(A1
11)−1A112 − A
022)− iAξ
]
21(A111)−1A1
12 + iAξ]
22 + rbξ]ξ]
−r(b11)−1A121(A1
11)−1 r(b11)−1 (b11)−1(A122 − A
121(A1
11)−1A112 − rb
ξ] )
, (5.19)
W = (f ], u2)T . (5.20)
Following the same procedure as before and using the special scaling function introduced
in Section 5.4, we make the coordinate change W (x;λ) = eg(x)V (x;λ) for some function g(x),
59
resulting in,
V ′(x;λ) = (A](x;λ)− g′(x)I)V (x;λ). (5.21)
We also make the coordinate change σ = u1 − u1+. This gives rise to the new system
1
2µ+ η((σ + u1+)(σ + u1+ − 1) + Γe(σ))
dZ(σ)
dσ= B](σ, λ)Z(σ), (5.22)
where B](σ;λ) = A](σ;λ)− f ′(σ)I and f ′(σ) = g′(x).
Again we want to choose f ′(σ) that will appropriately scale A] on the stable and unstable
manifold as we evolve toward the unstable manifold. As before, we choose
f ′(σ) = α+ + (α− − α+)σ
u1− − u1+
,
where α± is the most negative eigenvalue of A]± = limx1→±∞A].
If we are able to get solutions across the entire domain, corresponding to the six stable
eigenvectors of the system, then we will be able to use this method to compute the Evans
function. Unfortunately, this is not the case. None of the solutions remain stable across the
entire domain and only the solution corresponding to the first eigenvector remains stable
halfway across the domain before diverging. Figure 5.14 shows the solution corresponding
the first eigenvector on the stable half of the domain.
Figure 5.14: The fifteen-term series solution to the Evans system for 2-D MHD with λ = 10,scaled by f ′(σ).
60
The stability of this method for the non-lifted 2-D MHD system gives reason to believe
that we might have success using the lifted system as we did with compressible Navier-Stokes.
In this case, we start with an 11×11 system and so the resulting lifted system will be 55×55.
Our resulting Evans system is given by
1
2µ+ η((σ + u1+)(σ + u1+ − 1) + Γe(σ))
dZ(σ)
dσ= (B])(2)(σ, λ)Z(σ), (5.23)
where (B])(2)(σ;λ) = (A])(2)(σ;λ) − f ′(σ)I, (A])(2)(σ;λ) is the lifted system in pseudo-
Lagrangian coordinates, and f ′(σ) = g′(x).
We use the same choice for f ′(σ) as before, scaling by the dominating eigenvalues of
(A])(2) at both the stable and unstable manifold. Our results are shown in Figure 5.15.
Figure 5.15: The five-term series solution to the lifted Evans system for 2-D MHD withλ = 10, scaled by f ′(σ).
While these results seem promising with just five terms in the series, unfortunately when
we add more terms, our solutions diverge on the right. From this we see that we will
need further modifications to our method developed for compressible Navier-Stokes. These
modifications could involve adjusting f ′(σ) to have a stronger scale on the unstable manifold,
or even increasing the order of the function from linear to quadratic or cubic.
61
Appendix A. Analytic Evans System Solver
This appendix contains the generalized Python class developed for solving for the coefficient
vectors of the analytic solution to the Evans system. This code was used in Sections 5.1,
5.3, 5.4, and 5.5.
import numpy as np
import sympy as sp
class SystemSolve():
"""
This class analytically solves for the solution to the Evans
system £Z(\sigma)£ using the coordinate change £\sigma = v_+£.
Input:
dim (int) : dimension of the system
terms (int) : number of terms in the series solution £Z(\sigma)£
A (array) : the matrix dictating the unscaled Evans system
deriv (func) : the derivative £d\sigma/dx£ associated with the coordinate change
scale (func) : scaling function (typically called £f'(\sigma)£), can be a float
e_vec (vector): the eigenvector with which to initialize the system
display (bool): determines whether output is printed as system is solved
copmlex (bool): determines whether or not the system is complex valued
vec = [complex(vec[i].subs(self.B_tups)) for i in range(self.dim)]
mat = [[complex(mat[i,j].subs(self.B_tups)) for j in range(self.dim)] \
for i in range(self.dim)]
vec = np.array(vec)
mat = np.array(mat)
# Solve the system
if self.comp:
soln = np.linalg.solve(mat,vec)
64
else:
soln = mat.LUsolve(vec)
# Iteratively solve for B_n
for i in range(self.dim):
# Determine value of B_i
if self.comp:
val = soln[i]
val = complex(val)
else:
val = soln[i].subs(self.B_tups)
# Store value
if j != 0:
self.B_tups.append((self.B[j][i],val))
# Display progress (optional)
if self.p:
val = str(self.B[j][i]) +' = '+str(val)
print(val,'\n\n')
def evalZ(self, sig_val):
""" Evaluate £Z(\sigma)£ """
to_sub = self.B_tups + [(sigma,sig_val)]
Z_vals = []
# Substitute sigma into each dimension of solution
for i in range(self.dim):
val = self.Z[i].subs(to_sub)
if self.comp: val = complex(val)
Z_vals.append(val)
# Return £Z(\sigma)£
return Z_vals
def checkBNorms(self):
""" Compare ||B_n|| for each n """
norms = []
# Iterate through each coefficient vector B_n
for b in self.B:
# Compute ||B_n}}
65
norm = sp.sympify(0)
for i in b:
norm = norm + sp.Abs(i)
norm = (norm)**(1/2)
norm = norm.subs(self.B_tups)
# Store resulting norm
norms.append(norm)
return norms
66
Bibliography
[1] Blake Barker. Evans function computation, 2009. Master’s Thesis.
[2] Blake Barker, Jeffrey Humpherys, Gregory Lyng, and Kevin Zumbrun. Balancedflux formulations for multidimensional Evans-function computations for viscous shocks.Quarterly of Applied Mathematics, 76(3):531–545, Oct 2017.
[3] Blake Barker, Jeffrey Humpherys, Gregory Lyng, and Kevin Zumbrun. Euler versusLagrange: The role of coordinates in practical Evans-function computations. SIAMJournal on Applied Dynamical Systems, 17(2):1766–1785, Jan 2018.
[4] Blake Barker, Jeffrey Humpherys, Joshua Lytle, and Kevin Zumbrun. STA-BLAB: A MATLAB-based numerical library for Evans function computation.https://github.com/nonlinear-waves/stablab.git.
[5] Blake Barker, Jeffrey Humpherys, Keith Rudd, and Kevin Zumbrun. Stability of viscousshocks in isentropic gas dynamics. Comm. Math. Phys., 281(1):231–249, 2008.
[6] Blake Barker, Rafael Monteiro, and Kevin Zumbrun. Transverse bifurcation of viscousslow MHD shocks. http://arxiv.org/pdf/1901.09153.pdf.
[7] G. K. Batchelor. An introduction to fluid dynamics. Cambridge Mathematical Library.Cambridge University Press, Cambridge, paperback edition, 1999.
[8] H. Cabannes. Theoretical magnetofluiddynamics. Academic Press, New York, 1970.
[9] Constantine M Dafermos. Hyperbolic conservation laws in continuum physics, vol-ume 325 of Grundlehren der mathematischen wissenschaften [Fundamental principlesof mathematical sciences], 2010.
[10] Heinrich Freistuhler and Yuri Trakhinin. On the viscous and inviscid stability of magne-tohydrodynamic shock waves;. Physica D: Nonlinear Phenomena, 237(23):3030–3037,2008.
[11] David Gilbarg. The existence and limit behavior of the one-dimensional shock layer.Amer. J. Math., 73:256–274, 1951.
[12] Peter Howard and Kevin Zumbrun. The Evans function and stability criteria for de-generate viscous shock waves. Discrete Contin. Dyn. Syst., 10(4):837–855, 2004.
[13] J. Humpherys, G. Lyng, and K. Zumbrun. Spectral stability of ideal-gas shock layers.Arch. Ration. Mech. Anal., 2009.
[14] Jeffrey Humpherys, Gregory Lyng, and Kevin Zumbrun. Multidimensional stability oflarge-amplitude Navier–Stokes shocks. Archive for Rational Mechanics and Analysis,226(3):923–973, Jul 2017.
[15] Alan Jeffrey. Magnetohydrodynamics. University Mathematical Texts, No. 33. Oliver &Boyd, Edinburgh, 1966.
[16] Helge Kristian Jenssen and Gregory Lyng. Evaluation of the Lopatinski determinantfor gas dynamics. arXiv:0904.0268. Appendix to Z2.
[17] Shuichi Kawashima. Systems of a hyperbolic–parabolic composite type, with applicationsto the equations of magnetohydrodynamics. PhD thesis, Kyoto University, 1983.
[18] Corrado Mascia and Kevin Zumbrun. Pointwise Green function bounds for shock profilesof systems with real viscosity. Arch. Ration. Mech. Anal., 169(3):177–263, 2003.
[19] Corrado Mascia and Kevin Zumbrun. Stability of large-amplitude viscous shock profilesof hyperbolic-parabolic systems. Arch. Ration. Mech. Anal., 172(1):93–131, 2004.
[20] Guy Metivier and Kevin Zumbrun. Large viscous boundary layers for noncharacteristicnonlinear hyperbolic problems. Mem. Amer. Math. Soc., 175(826):vi+107, 2005.
[21] D. H. Sattinger. On the stability of waves of nonlinear parabolic systems. Advances inMath., 22(3):312–355, 1976.
[22] Andreas Schafelner. Introduction to magnetohydrodynamics., 2016. Bachelor’s Thesis.
[23] Benjamin Texier and Kevin Zumbrun. Hopf bifurcation of viscous shock waves in com-pressible gas dynamics and MHD. Arch. Ration. Mech. Anal., 190(1):107–140, 2008.
[24] Kevin Zumbrun. Multidimensional stability of planar viscous shock waves. In Advancesin the theory of shock waves, volume 47 of Progr. Nonlinear Differential Equations Appl.,pages 307–516. Birkhauser Boston, Boston, MA, 2001.
[25] Kevin Zumbrun. Planar stability criteria for viscous shock waves of systems with realviscosity. In Hyperbolic systems of balance laws, volume 1911 of Lecture Notes in Math.,pages 229–326. Springer, Berlin, 2007.
[26] Kevin Zumbrun. Numerical error analysis for Evans function computations: a numer-ical gap lemma, centered-coordinate methods, and the unreasonable effectiveness ofcontinuous orthogonalization. arXiv.org:0904.0268, 2009.
[27] Kevin Zumbrun. The refined inviscid stability condition and cellular instability of vis-cous shock waves. Phys. D, 239(13):1180–1187, 2010.
[28] Kevin Zumbrun and D. Serre. Viscous and inviscid stability of multidimensional planarshock fronts. Indiana Univ. Math. J., 48(3):937–992, 1999.