An optimization-based approach for high-order accurate discretization of

conservation laws with discontinuous solutions

M. J. Zahra,b,2,∗, P.-O. Perssona,c,3

a Mathematics Group, Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA 94720, United States
b Department of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN, 46556, United States
c Department of Mathematics, University of California, Berkeley, Berkeley, CA 94720, United States
arXiv:1712.03445v2 [math.NA] 20 Mar 2018

Abstract
This work introduces a novel discontinuity-tracking framework for resolving discontinuous solutions of con-
servation laws with high-order numerical discretizations that support inter-element solution discontinuities,
such as discontinuous Galerkin or finite volume methods. The proposed method aims to align inter-element
boundaries with discontinuities in the solution by deforming the computational mesh. A discontinuity-aligned
mesh ensures the discontinuity is represented through inter-element jumps while smooth basis functions in-
terior to elements are only used to approximate smooth regions of the solution, thereby avoiding Gibbs’
phenomena that create well-known stability issues. Therefore, very coarse high-order discretizations accu-
rately resolve the piecewise smooth solution throughout the domain, provided the discontinuity is tracked.
Central to the proposed discontinuity-tracking framework is a discrete PDE-constrained optimization for-
mulation that simultaneously aligns the computational mesh with discontinuities in the solution and solves
the discretized conservation law on this mesh. The optimization objective is taken as a combination of the
the deviation of the finite-dimensional solution from its element-wise average and a mesh distortion metric
to simultaneously penalize Gibbs’ phenomena and distorted meshes. It will be shown that our objective
function satisfies two critical properties that are required for this discontinuity-tracking framework to be
practical: (1) possesses a local minima at a discontinuity-aligned mesh and (2) decreases monotonically to
this minimum in a neighborhood of approximately O(h), whereas other popular discontinuity indicators fail
to satisfy the latter. Another important contribution of this work is the observation that traditional reduced
space PDE-constrained optimization solvers that repeatedly solve the conservation law at various mesh con-
figurations are not viable in this context since severe overshoot and undershoot in the solution, i.e., Gibbs’
phenomena, may make it impossible to solve the discrete conservation law on non-aligned meshes. There-
fore, we advocate a gradient-based, full space solver where the mesh and conservation law solution converge
to their optimal values simultaneously and therefore never require the solution of the discrete conservation
law on a non-aligned mesh. The merit of the proposed method is demonstrated on a number of one- and
two-dimensional model problems including the L2 projection of discontinuous functions, Burgers’ equation
with a discontinuous source term, transonic flow through a nozzle, and supersonic flow around a bluff body.
We demonstrate optimal O(hp+1 ) convergence rates in the L1 norm for up to polynomial order p = 6 and
show that accurate solutions can be obtained on extremely coarse meshes.
Keywords: r-adaptivity, shock tracking, high-order methods, discontinuous Galerkin, full space
PDE-constrained optimization, transonic and supersonic flow

∗ Corresponding author
Email addresses: [email protected] (M. J. Zahr), [email protected] (P.-O. Persson)
1 Luis W. Alvarez Postdoctoral Fellow, Computational Research Division, Lawrence Berkeley National Laboratory
2 Assistant Professor, Department of Aerospace and Mechanical Engineering, University of Notre Dame
3 Associate Professor, Department of Mathematics, University of California, Berkeley

Preprint submitted to Elsevier March 21, 2018

1. Introduction

Even with the continued advancement of computational hardware and algorithms, it is apparent that
important developments are still required to improve the predictive capability of the computational sim-
ulation tools that are used for design. For example, it is widely believed that higher fidelity is required
for problems with propagating waves, turbulent fluid flow, nonlinear interactions, and multiple scales [51].
This has resulted in a significant interest in high-order accurate methods, such as discontinuous Galerkin
(DG) methods [15, 25], which have the potential to produce more accurate solutions on coarser meshes than
traditional compatible discretizations, such as mimetic finite differences and finite volumes.
One of the remaining problems with the DG methods is their sensitivity to under-resolved features, in
particular for non-linear problems where the spurious oscillations often cause a break-down of the numerical
solvers. This is exacerbated for problems with shocks, where the natural dissipative mechanisms introduced
by DG methods through jump terms is insufficient to stabilize the solution for high-order approximations.
Since shocks are present in many important problems in fields such as aerospace, astrophysics, and combus-
tion, the lack of efficient ways to handle them is a fundamental challenge for the wide adoption of these new
numerical methods.
One of the most straight-forward approaches for stabilizing high-order methods in the presence of shocks is
to identify the elements close to the shocks using a sensor, and reduce the corresponding approximation orders
[7, 12]. This increases the numerical dissipation which helps stabilize the solution, and in the extreme case of
piecewise constant solutions will completely avoid any oscillations. An obvious drawback with this strategy
is that it will typically result in a first-order scheme in the affected elements. This can partially be addressed
by using h-adaptivity and remeshing [16, 27]; however, generating efficient anisotropic meshes in 3D is still
a difficult task and the resulting meshes typically have a very large number of degrees of freedom around
the shocks. Further issues with these methods include moving shocks, which require constant remeshing and
solution transfer, and a difficulty for implicit solvers with large timesteps since order reduction strategies
typically cannot be differentiated for a nonlinear Newton solver. A related class of methods use limiters
based on the ENO/WENO schemes [22, 34, 29] to limit the high-order representation of the solution inside
the elements [13, 42, 57, 32, 55]. While these schemes can be remarkably robust and stable for very strong
shocks, they suffer from spurious oscillations, have difficulties with fully unstructured meshes, particularly
in 3D, and are typically only used for explicit timestepping.
Another popular way to stabilize problems with shocks is to explicitly add artificial viscosity to the
governing equations [50]. In [6, 7, 24], this was applied to the DG method using element-wise viscosity based
on the residual of the Euler equations. To address issues with the consistency of the resulting scheme, a
proper discretization of the artificial viscosity was used in [41] together with a resolution-based indicator
based on the decay-rate of the high-order terms. The resulting scheme could smoothly resolve shocks using
subcell resolution, and the continuous nature of both the sensor and the viscosity made it possible to get fully
converged steady-state solutions and accurate time-stepping using implicit methods. Further developments
of this approach include some refinements of the parameter selection [30], the solution of an auxiliary PDE
for the viscosity field [5], and the extension to implicit time-stepping for transient problems with smooth
viscosity fields [39]. An alternative approach for the shock sensor is to somehow incorporate the physics
of the problem [28, 9, 35]. Overall, these artificial viscosity approaches are widely used, but like limiting
they require anisotropic h-adaptivity to be competitive [2, 53], which again is difficult in 3D, particularly for
moving shocks, and produces large meshes.
In this work, we propose a fundamentally different approach based on the idea of shock tracking or
shock fitting [45, 46, 8, 23, 49, 48, 56, 47, 4, 44, 20, 37] using r-adaptivity. For generality, we will use the
term discontinuity tracking in this work, although our motivating applications involve shock waves. Since
the solution basis in DG methods (as well as the finite volume methods) naturally support discontinuities
between each element, it is in principle possible to keep an unmodified high-order DG discretization for
problems with strong discontinuities, provided the element faces can be aligned with the discontinuities of
the problem. However, this is a very challenging problem since the unknown discontinuity surfaces can
have complex geometries and if the elements are even just slightly misaligned, the discrete equations cannot
be solved due to spurious oscillations. Instead, we employ a new optimization formulation that aims to

2
align faces of the computational mesh with discontinuities and solve the discrete conservation law. The
optimization formulation is classified as fully discrete r-adaptive PDE-constrained optimization since the
discretized conservation law defines nonlinear equality constraints and the state vector and nodal positions
of the computational mesh are optimization variables. In this framework, the objective function must
possess a local minima at a discontinuity-aligned mesh and must monotonically approach this minima in a
neighborhood of radius approximately h/2 for a solver to successfully locate such a minima. Here, we propose
a simple robustness-based discontinuity indicator that will be experimentally shown to satisfy both criteria,
while a number of intuitive physics- and error-based discontinuity indicators fail to satisfy the monotonicity
property. Finally, we use a full space approach to solve the PDE-constrained optimization problem whereby
the discrete solution of the conservation law and the nodal positions of the mesh simultaneously converge
to their optimal values without ever requiring the solution of the discrete conservation law on a non-aligned
mesh. The popular alternative, the reduced space approach, is not a viable solver as it requires the solution
of the discrete conservation law on non-aligned meshes, which may be impossible due to Gibbs’ instabilities.
We demonstrate our method can produce high-order accurate solutions on remarkably coarse meshes, for a
range of problems in 1D and 2D.
The remainder of the paper is organized as follows. Section 2 introduces the governing system of steady,
inviscid conservation laws on a parametrized domain and its discretization using a discontinuous Galerkin
method. Since deformation of the computational domain is a key ingredient in this work, a parametrized
domain deformation is introduced at the continuous level and the conservation laws are recast on a fixed
reference domain. In the end, Section 2 reduces the conservation law to a discrete nonlinear system of
equations that depend on the discrete state vector and computational mesh, which is the point of departure
for the proposed discontinuity-tracking framework, introduced in Section 3. The proposed PDE-constrained,
r-adaptive formulation is presented in Section 3.1. The objective function used in this work, introduced
in Section 3.2, is a combination of new discontinuity indicator and a standard mesh distortion indicator.
In this section, we show the indicator possesses a minima at a discontinuity-aligned mesh and decreases
monotonically to this minima, whereas the latter condition fails for some existing indicators. Section 3.3
discusses the full space solver required for the proposed discontinuity-tracking and a detailed discussion
on why a traditional reduced space PDE-constrained optimization approach is not sufficient. Section 3.4
discusses a number of practical details including implementation, initialization of the nonlinear optimization
solver, and robustness and efficiency of the method. Finally, Section 4 presents a number of one- and two-
dimensional test problems that demonstrate optimal O(hp+1 ) convergence up to p = 6 and accurate solutions
are obtained on extremely coarse meshes.

2. Governing equations and high-order numerical discretization

Consider a general system of Nc conservation laws, defined on the physical domain Ω ⊂ Rd ,

∇ · F(U ) = 0 in Ω, (1)

where U (x) ∈ RNc is the solution of the system of conservation laws at x ∈ Ω and F(U ) ∈ RNc ×d is the
physical flux. While we solely consider steady conservation laws, i.e., Ω is a d-dimensional spatial domain
and the solution is independent of time, the conservation law in (1) and the proposed numerical method
encapsulate the unsteady case where Ω is the d-dimensional space-time domain and the spatial domain is
(d − 1)-dimensional. Furthermore, we assume the solution U contains discontinuities, in which case the
conservation law (1) holds away from these discontinuities.
The proposed optimization-based method for tracking discontinuities, introduced in Section 3, is built
upon existing numerical discretizations that possess the following properties: 1) represents a stable and con-
vergent discretization of the conservation law in (1), 2) allows for deformation of the computational domain,
and 3) employs a solution basis that supports discontinuities between computational cells or elements. While
the proposed method will be valid for any numerical discretization that satisfies these requirements, such as
finite volume or discontinuous Galerkin methods, we focus on high-order DG methods given their potential
to deliver accurate solutions on very coarse discretizations, provided the discontinuities are tracked.

3
 n da

N dA x = G(X, µ)
Ω
x2

Ω0

X2 x1

X1

Figure 1: Mapping between reference and physical domains.

The remainder of this section will detail the discretization of the conservation law (1) using DG such
that it reduces to the discrete form
r(u, x) = 0 (2)
where u ∈ RNu is the discrete representation of the conservation law state U , x ∈ RNx is the discrete
representation of the conservation law domain Ω, and r is the discretized conservation law. A standard
nodal, isoparametric DG method is used for the discretization, with special attention given to treatment of
the domain deformation and the numerical fluxes, which must be selected carefully in a tracking method since
the inter-element jumps do not tend to zero on the discontinuity surface. We will discuss the formulation
of the proposed optimization-based tracking method in the context of the general discrete conservation law
(15). As a result, our developments will be applicable to a broad range of admissible discretizations with
only minor modifications. Section 3 will mention the modifications required to use a finite volume method
as the underlying discretization.

2.1. Transformed conservation law from deformation of physical domain

Before introducing a discretization of (1) it is convenient to explicitly treat deformations to the domain
of the conservation law Ω. Given the divergence-form of the conservation law in (1), i.e., no explicit time
derivative term, the domain deformation is handled directly. That is, the physical domain can be taken as
the result of a parametrized diffeomorphism applied to a reference domain (Figure 1)

Ω = G(Ω0 , µ), (3)

where Ω0 ⊂ Rd is a fixed reference domain, µ ∈ RNµ is a vector of parameters, and G : Rd × RNµ → Rd

is the parametrized diffeomorphism defining the domain mapping. Under the domain mapping (3), the
conservation law becomes
∇ · F(U ) = 0 in G(Ω0 , µ). (4)
For convenience, the conservation law on the physical domain Ω is transformed to a conservation law on
the reference domain Ω0 using the procedure in [40] to yield

∇X · F (u, µ) = 0 in Ω0 (5)

where ∇X denotes spatial derivatives with respect to the reference domain Ω0 with coordinates X. The
transformed state vector, u, and flux, F , take the form

u = gµ U, F (u, µ) = gµ F(gµ−1 u)G−T

µ (6)
∂
where Gµ (X) = ∂X G(X, µ) is the deformation gradient of the domain mapping and gµ (X) = det Gµ (X)
is the Jacobian. The transformed conservation law in (5)-(6) is, by construction, identical to the Arbitrary

4
Lagrangian-Eulerian formulation of the steady conservation law in (4). For details regarding the derivation
of the transformed equations, the reader is referred to [40].
While a numerical scheme and implementation can be designed around (4) or (5), we chose the formulation
on the reference domain in (5) as it provides a convenient framework to handle domain deformation. In this
case, the governing equation depends on the domain deformation solely through the transformed flux function
and definition of the state vector. This detail leads to a clean implementation, particularly when computing
required derivative quantities, as discussed in Section 3.4.

2.2. Discontinuous Galerkin discretization of transformed conservation law

In this section, a standard nodal discontinuous Galerkin method [15, 25] is used to discretize the trans-
formed conservation law (5) with special attention given to the numerical fluxes. Let Eh,p represent a
discretization of the reference domain Ω0 into non-overlapping, potentially curved, computational elements,
where h is a mesh element size parameter and p is the polynomial order associated with the curved elements.
The DG construction begins by considering the element-wise weak form of the governing equation in (5) that
results from multiplication by a test function ψ, integration over a single element K ∈ Eh,p , and application
of the divergence theorem
Z Z Z
∇X · F (u, µ) · ψ dV = ψ · F (u, µ)N dA − F (u, µ) : ∇X ψ dV = 0, (7)
K ∂K K

where N is the outward normal to the surface ∂K. The global weak form, upon which DG methods are
built, arises from the summation of the local residuals over all elements in Eh,p
X Z Z
ψ · F (u, µ)N dA − F (u, µ) : ∇X ψ dV = 0. (8)
K∈Eh,p ∂K Ω0

At this point, a numerical flux, F ∗ , is introduced in the first term of (8)

X Z Z
∗
ψ · F (u, µ, N ) dA − F (u, µ) : ∇X ψ dV = 0 (9)
K∈Eh,p ∂K Ω0

to ensure the flux is single-valued along ∂K where u is multi-valued. In DG methods, the numerical flux
must be chosen such that it is consistent to ensure local conservation [14], i.e., F ∗ (ū, µ, N ) = F (ū, µ)N
where ū(X) is single-valued, and leads to a stable discretization. Given that inter-element jumps tend to zero
under refinement, the particular choice of numerical flux is less important, provided it is stable and consistent
[15, 52]. However, recall that the goal of this work is to track discontinuities with element faces, which implies
inter-element jumps at the discontinuity will not tend to zero under refinement and care must be taken when
choosing a numerical flux such that it is consistent with the governing equation at discontinuities. In this
work, Roe’s method [43] with the Harten-Hyman entropy fix [23] is used for the numerical fluxes at interior
faces and the appropriate boundary conditions determine the numerical fluxes on faces that intersect ∂Ω.
The Harten-Hyman entropy fix is used to ensure non-physical rarefaction shocks do not appear.
To establish the finite-dimensional form of (9), we introduce the isoparametric finite element space of
piecewise polynomial functions associated with the mesh Eh,p :

Vh,p = v ∈ [L2 (Ω0 )]Nc | v|K ◦ TK ∈ [P(K0 )]Nc ∀K ∈ Eh,p


n o
V̂h,p (µ) = v ∈ [L2 (Ω)]Nc | v|G(K, µ) ◦ TG(K, µ) ∈ [P(K0 )]Nc ∀K ∈ Eh,p

where Pp (K0 ) is the space of polynomial functions of degree at most p ≥ 1 on the parent element K0 and
K = TK (K0 ) defines a mapping from the parent element to element K ∈ Eh,p . For notational brevity,
we assume all elements map from a single parent element. The parametrized space V̂h,p (µ) will be used
in later sections to define an L2 projection onto a piecewise polynomial space attached to the physical

5
domain Ω = G(Ω0 , µ). The Petrov-Galerkin weak form in (9) becomes: find uh,p ∈ Vh,p such that for all
ψh0 ,p0 ∈ Vh0 ,p0

h0 ,p0
X Z Z
rh,p (uh,p , µ) := ψh0 ,p0 · F ∗ (uh,p , µ, N ) dA − F (uh,p , µ) : ∇X ψh0 ,p0 dV = 0, (10)
K∈Eh,p ∂K Ω0

0 0
h ,p
where rh,p is the finite-dimensional residual of (9) corresponding to the trial space Vh,p and test space Vh0 ,p0 .
In this work, these spaces are taken to be the same and we have the Galerkin weak form: find uh,p ∈ Vh,p
such that for all ψh,p ∈ Vh,p
X Z Z
ψh,p · F ∗ (uh,p , µ, N ) dA − F (uh,p , µ) : ∇X ψh,p dV = 0. (11)
K∈Eh,p ∂K Ω0

The Petrov-Galerkin residual in (10) will be used in Section 3.2.2 to define an accuracy-based discontinuity
Np
indicator. To obtain the equivalent algebraic representation of (11), we introduce a basis {ϕi }i=1 for Pp (K0 ),
where Np is the number of basis functions. While this can be any valid basis, a flexibility afforded by the DG
framework, it is convenient for the proposed shock tracking method to work with a nodal basis. Define a set
Np
of nodes {ξj }j=1 within the parent element K0 and let ϕi be the Lagrange polynomial associated with nodes
ξi over element K0 with the property ϕi (ξj ) = δij . For brevity, we introduce notation for the Lagrange basis
−1
defined on the reference element K ∈ Eh,p : ϕK i (X) = ϕi (TK (X)). The finite-dimensional solution, uh,p , in
each element K ∈ Eh,p is written in terms of its discrete expansion coefficients as
Np
X
uh,p (X)|K = uK K
i ϕi (X), (12)
i=1

where uKi ∈ R
Nc
is the solution at node i of element K. Under the isoparametric assumption, the physical
coordinates are expanded in the nodal basis as
Np
X
xh,p (X)|K = xK K
i ϕi (X), (13)
i=1

where xK d
i ∈ R is the coordinate of node i of element K. It will also be convenient in later sections to
introduce the corresponding nodal coordinates in the reference domain, XiK ∈ Rd . From this expansion, the
deformation gradient, G, and Jacobian, g, required to define the transformed state vector and fluxes are
Np
X ∂ϕK
i
Gh,p (X)|K = xK
i (X)
i=1
∂X (14)
gh,p (X)|K = det Gh,p (X)|K .
Therefore, in the isoparametric setting, the mapping to the physical domain is completely determined from
the nodal positions of each element xKi . To ensure the domain mapping is continuous, nodal positions co-
located in the reference domain are required to be co-located in the physical domain. This is accomplished
by using a single parameter to define the position of all nodes co-located in the reference domain, which
is equivalent to using the nodal positions of the continuous high-order mesh corresponding to Eh,p as the
parameter set. Let x ∈ RNx denote these positions, then xK i is reconstructed from x by selecting the ap-
propriate entry based on the position of the corresponding node in the reference domain. The corresponding
nodal coordinates, denoted X ∈ RNx , in the reference domain are defined similarly. With this notation, the
continuous domain mapping is parametrized by µ = x, i.e., Ω = G(Ω0 , x). The remainder of the document
will use x to parametrize the domain deformation instead of µ. Finally, the integrals in (11) are evaluated
using high-order Gaussian quadrature rules to yield the discrete form of the governing equations
r(u, x) = 0, (15)

6
where u ∈ RNu (Nu = Nc Np |Eh,p |) is the solution vector comprised of the coefficients uK
i for all elements.
The discrete form of the governing equations in (15) will be the point of departure for the remainder of
the document. The proposed shock tracking method presented in the next section will be applicable to any
discretization of the governing equations that assumes the general form (15) and satisfies the requirements
outlined in the beginning of this section, such as finite volume methods.

3. High-order discontinuity tracking via optimization-based r-adaptivity

In this section, we present a novel method designed to align discontinuous features in a finite-dimensional
solution basis with features in the solution itself, with a driving application being transonic and supersonic
compressible flow where shock waves are present. In the discretization setting outlined in Section 2, e.g.,
discontinuous Galerkin or finite volumes, this amounts to aligning element faces, where discontinuities are
supported, with discontinuities in the solution. With the discontinuous features tracked with element faces,
very coarse high-order discretizations effectively resolve the solution throughout the domain. However, in
general, the location of the discontinuities is not known a-priori and therefore the face alignment cannot be
done explicitly. To circumvent this difficulty that has plagued previous attempts at discontinuity tracking,
the proposed method is built around three critical contributions:
1) a PDE-constrained optimization formulation with optimization variables taken as the nodal positions of
the high-order mesh, i.e., an optimization-based r-adaptive framework, that allows elements to curve to
accurately track the location of the discontinuity,
2) an objective function that attains a (local) minimum at a discontinuity-aligned mesh and monotonically
decreases to this minimum in a neighborhood of radius approximately h/2,

3) a PDE-constrained optimization solver that does not require solution of the discrete PDE (15) on meshes
not aligned with the discontinuity as this cannot be done reliably in a numerical scheme due to Gibbs’
phenomena around the discontinuity (see Section 1 for the complete discussion).

The remainder of this section will detail these contributions and summarize the proposed high-order discontinuity-
tracking algorithm in its entirety.

3.1. Optimization formulation for r-adaptivity

The proposed method for high-order resolution of discontinuous solutions of conservation laws refor-
mulates the discrete nonlinear system in (15) as a PDE-constrained optimization problem. The original
nonlinear system formulation searches for a discrete solution u on a given and fixed mesh x, whereas the
proposed optimization formulation searches for the discrete solution and mesh that minimize some objective
function, f : RNu × RNx → R, and satisfy the discrete PDE

minimize f (u, x)
u∈RNu , x∈RNx
(16)
subject to r(u, x) = 0.

The objective function, discussed in detail in Section 3.2, is constructed such that it is locally minimized at
a discontinuity-aligned mesh. The proposed method falls into the category of r-adaptive methods given that
it moves nodes to improve approximation quality; however, it has stronger requirements than existing r-
adaptive methods as we aim to align mesh faces with the discontinuity rather than just nodes. See Figure 5
(top left) for a case where several mesh nodes are aligned with a discontinuity without any faces being
aligned. Such situations are avoided through the construction of the objective function in Section 3.2 that
is minimized when faces are aligned with discontinuities.
While the general formulation of the optimization problem in (16) is desirable in that it does not require
any a-priori knowledge of the location of the discontinuity, it can be cumbersome due to the large number
of optimization variables and difficult to avoid ill-conditioning and non-uniqueness in the domain mapping,

7
i.e., severe mesh distortion or inversion. Ill-conditioning and non-uniqueness of the domain mapping will be
addressed in the construction of the objective function in Section 3.2. To reduce the number of optimization
variables, we re-parametrize the optimization problem in (16). Let φ ∈ RNφ be a vector of parameters such
that Nφ ≤ Nx and A : RNφ → RNx be an injective mapping that maps φ 7→ x, i.e., x = A(φ). For clarity,
we use x(φ) in place of A(φ). Then the optimization problem in (16) becomes

minimize f (u, x(φ))

u∈RNu , φ∈RNφ (17)
subject to r(u, x(φ)) = 0.

In addition to reducing the number of optimization parameters, the mapping A can incorporate mesh
smoothing operations to improve mesh quality. However, despite the advantages of re-parametrization, the
fact that A is not necessarily bijective limits the reachable mesh configurations and care must be taken to
ensure the parametrization is capable of aligning with discontinuities. In this document, three different mesh
parametrizations will be considered: the entire mesh, a single node parametrization with piecewise linear
mesh smoothing, and a n-node parametrization with linear elastic smoothing in d dimensions. In the most
general case where all mesh nodes are used as optimization variables, the re-parametrization map is the
identity map, A(φ) = φ, or any other bijection, and (17) is equivalent to the original optimization problem.
The other mesh parametrizations used in this work will be discussed in Sections 3.2 and 4.4.

3.2. Discontinuity tracking objective function

A critical contribution of this work is the specific form of the objective function in the PDE-constrained
optimization problem in (17). For the proposed framework to successfully align faces of the computational
mesh with discontinuities in the solution the objective function must 1) attain a local minimum at some
discontinuity-aligned mesh and 2) monotonically decrease to such a minima is a neighborhood of radius
approximately h/2. The first requirement ensures a local minima of (17) leads to a desired mesh and the
second requirement gives gradient-based numerical optimization methods hope to locate it even in the worst-
case scenario where element faces are initially h/2 away from discontinuities. Thus, we seek an objective
function that is an effective discontinuity indicator that monotonically decreases from large values at non-
aligned meshes to smaller values at aligned meshes.
A host of discontinuity indicators exist in the literature [4, 41] as sensing discontinuous features is a
backbone of many numerical methods for either capturing or tracking discontinuities. Two common classes
of discontinuity indicators are physics-based and error-based. Physics-based indicators typically involve the
residual associated with the integral form of the conservation law, i.e., the Rankine-Hugoniot conditions
[4]. Error-based indicators use the magnitude of an approximate error measure such as adjoint-based error
estimators [17] or the residual of the solution in a richer test space [1]. In this work, we propose a departure
from these traditional discontinuity indicators and consider an indicator driven purely by the robustness
of the numerics [41]. Recall from Section 1 that high-order discretizations of conservation laws fail when
attempting to resolve discontinuous solutions due to severe spurious oscillations. Therefore we propose an
objective function that penalizes these oscillations by considering the element-wise distance of the numerical
solution from its mean value, i.e.,
X Z 2
−2
fshk (u, x) = h0 uh,p − ūK
h,p W dV (18)
K∈Eh,p G(K, x)

where the dependence of the finite dimensional solution uh,p on the discrete representation is implied, ūK
h,p
is the mean value of uh,p over element K, W ∈ RNc ×Nc is the symmetric positive semi-definite matrix that
defines the local semi-norm, and h0 is the length scale of the reference mesh Eh,p
Z Z  Z 1/d
1 1
ūK
h,p = uh,p dV, |G(K, x)| = dV, h0 = dV . (19)
|G(K, x)| G(K, x) G(K, x) |Eh,p | Ω0

8
The h−20 factor ensures the indicator scales as O(1) instead of O(h2 ). This objective function does not
explicitly depend on the governing equation, only the approximation space Vh,p , nor does it take into account
the accuracy of the finite-dimensional solution, which marks a departure from physics-based and error-based
discontinuity indicators.
Remark. The discontinuity indicator in (18) will be identically zero, and therefore a useless objective func-
tion, if a piecewise constant polynomial approximation space is used (p = 0), which is equivalent to a finite
volume scheme. In this case, the pointwise values of the finite dimensional solution uh,p in (18) can be
replaced by its non-limited reconstruction using the solution in neighboring elements [33].
The discontinuity indicator in (18) is not well-suited as the objective function in the discontinuity-tracking
optimization setting (17) in its current form because it is agnostic to a poor quality or inverted mesh that
may arise from certain choices of x. Therefore we construct the objective function as a weighted combination
of (18) and a function fmsh : RNx → R that penalizes mesh distortion

f (u, x; α) = fshk (u, x) + αfmsh (x). (20)

In the remainder of the document the dependence on α will be dropped unless explicitly required. In this
work, we use the following mesh distortion measures

h0
 X

 h0 −1 d=1
|G(K, x)|


 K∈E
h,p
fmsh (x) = 2
!r (21)
kGh,p kF
Z
 d
X 1
h0
 otherwise,
|G(K, x) G(K, x) (det Gh,p )2/d


K∈Eh,p +

where r = 2 is used in this work. For the general case of d > 1, the quantity in (21) is widely accepted in
the high-order meshing community as reasonable metric to penalize mesh distortion and is well-suited for
optimization [31, 18]. However, this mesh distortion measure is not useful for the case d = 1. Therefore we
treat the case d = 1 separately and define mesh distortion as the deviation from a uniform mesh, modified
to heavily penalize zero-volume elements. The hd0 factors are again included to ensure the fmsh scales
independently of h. Since both fshk and fmsh are independent of h and p, the value of α that appropriately
balances discontinuity tracking and mesh regularization should be mesh independent, i.e., independent of h
and p.

3.2.1. Behavior of objective function: L2 projection of discontinuous function

In this section, we use a simple one- and two-dimensional model problem to demonstrate the behavior of
the proposed discontinuity indicator (18) near a discontinuity-aligned mesh and verify it possesses the desired
properties suggested at the beginning of Section 3.2, namely attains a local minimum at a discontinuity-
aligned mesh and monotonically approaches such a minima. Additionally, we show the impact of the mesh
regularization term (21) on the combined objective function (20) as a function of the weighting parameter
α. Unlike the physics- and error-based indicator, the proposed indicator is independent of the conservation
law, which provides the opportunity to study the indicator without the complication of a partial differential
equation. Instead, we consider the L2 projection of a discontinuous function onto the piecewise polynomial
space V̂h,p (x).
A complete description of the L2 projection in the finite-dimensional setting of Section 2.2 is provided in
Section 4.1. Here, we let uH (x) denote the discrete representation of the L2 projection of the discontinuous
function H : Ω → R onto the space V̂h,p (x) and investigate the behavior of

jα (φ) = f (uH (x(φ)), x(φ); α), (22)

where x(φ) denotes a parametrization of the computational mesh. In order to graphically study the objective
function, it is important to use a single degree of freedom to parametrize the mesh. This is accomplished in

9
 x

X
xl xr

Figure 2: Single degree of freedom 1d mesh parametrization in (23) corresponding to a function with a discontinuity at ( ).
The choice of x̄ ( ) is chosen to align with a mesh node ( ) near the position of the discontinuity and the parameter φ
reduces to the position of this node. The remaining nodes ( ) are determined by uniformly distributing them in the intervals
(xl , x̄) and (x̄, xr ). The first configuration (bottom) shows the reference domain with a uniform nodal distribution and the
remaining configurations corresponds to the deformation with φ = x∗ , x∗ − 0.75h, x∗ + 0.75h, respectively. This range of φ is
the same range used to perform the sweep of the objective function (22) in Section 3.2.1 and Figure 3.

one spatial dimension (d = 1) by using φ to define the mapping of a single point x̄ in the reference domain
Ω0 and mapping the remaining points piecewise linearly between the fixed boundaries and φ
φ − xl

xl + (X − xl )
 X ≤ x̄
xh,p (X; φ) = x̄ − xl (23)
x − φ
xr − (xr − X) r
 X > x̄
xr − x̄
where Ω0 = (xl , xr ) ⊂ R. In the discrete setting, the mapping in (23) is applied to the nodal positions of
the reference domain X to define the nodal positions of the physical domain x as
φ − xl

xl + (Xi − xl )
 Xi ≤ x̄
xi (φ) = x̄ − xl (24)
x −φ
xr − (xr − Xi ) r
 Xi > x̄
xr − x̄
for j ∈ {1, . . . , Nx }. This single degree of freedom mesh parametrization possesses the obvious property
xi (x̄) = Xi , which implies that the mesh quality will be perfect in terms of the distortion metric fmsh in (21)
at φ = x̄ if the reference nodes are uniformly distributed. Additionally, if the reference nodes are uniformly
distributed, the parametrization in (24) inherently incorporates ideal smoothing since the resulting nodes
will be uniformly spaced in the intervals [xl , x̄] and (x̄, xr ]. A depiction of the single degree of freedom mesh
parametrization is provided in Figure 2. This figure also shows the range that φ will sweep to study the
behavior of jα (φ) next.
With this single degree of freedom parametrization, we consider the graph of jα (φ) for various values
of α in Figure 3, where uH corresponds to the L2 projection of the piecewise sinusoidal function in (50)
with k = 2 onto a piecewise polynomial basis with 17 elements and polynomial orders p ∈ {1, 2, 3, 4}.
The reference nodes are uniformly distributed and therefore fmsh possesses a minimum at φ = x̄. For the
case with no mesh regularization, the proposed objective function satisfies the two desired properties, i.e.,
possesses a local minima when the mesh is aligned with the discontinuity and monotonically approaches this
minima in a neighborhood of radius h/2. As the penalty parameter is increased to α = 1, the shape of the
graph changes in that values near x̄ decrease, but monotonicity is not broken. As α increases to 10, the
monotonicity property breaks down and a local minima is introduced at x̄, which becomes stronger as α is
increased to 20. Once α is sufficiently large (α = 1000), the mesh regularization term dominates causing
the local minima at the shock location to vanish. Similar trends were observed when using coarser and finer
discretization into piecewise polynomials. This simple test shows that the discontinuity indicator proposed
in (20) is well-suited as the objective function in our discontinuity-tracking framework, provided the penalty
parameter is chosen to balance the scale of the discontinuity indicator and mesh regularization terms.
Given the promising performance of the indicator, we aim to study it in the situation where the solution
oscillates within an element and the basis may be underresolved. To this end, we repeat the previous

10
 3 3
jα (φ), α = 0.0

jα (φ), α = 1
2 2

1 1

0 0
0.46 0.48 0.5 0.52 0.54 0.46 0.48 0.5 0.52 0.54

4 6
jα (φ), α = 10

jα (φ), α = 20
3
4
2
2
1

0 0
0.46 0.48 0.5 0.52 0.54 0.46 0.48 0.5 0.52 0.54

150
8
jα (φ), α = 1000
jα (φ), α = 40

6 100

4
50
2

0 0
0.46 0.48 0.5 0.52 0.54 0.46 0.48 0.5 0.52 0.54
φ (position of node closest to shock) φ (position of node closest to shock)

Figure 3: The objective function jα (φ) corresponding to the L2 projection of the discontinuous function in (50) with k = 2 onto
a piecewise polynomial basis of degree p = 1 ( ), p = 2 ( ), p = 3 ( ), p = 4 ( ) with |Eh,p | = 17 elements for
various values of the mesh regularization parameter α under the mesh parameterization in (23) and Figure 2. The parameter φ
is swept over the range [x∗ − 0.75h, x∗ + 0.75h], where x∗ is the location of the discontinuity ( ). The position of x̄ in (23) is
indicated with ( ); when φ = x̄ nodes are uniformly distributed and fmsh takes its minimum values. The red shaded region
identifies the neighborhood of radius h/2 about x∗ . The objective function with no mesh regularization (α = 0) in the top left
plot shows the two desired properties discussed in Section 3.2, namely, a minimum at x∗ and monotonically decreases to this
minimum in a neighborhood centered at x∗ with radius h/2. The remaining plots show increasing values of the regularization
parameter α. The minimum does not noticeably move until α is sufficiently large (α = 40) at which point is moves from x∗ to
x̄, i.e., the objective function transitions from prioritizing discontinuity tracking to mesh regularization.

11
 jα (φ), α = 0.0 0.15 0.4

0.3
0.1
0.2
0.05
0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

0.4 0.6
jα (φ), α = 0.0

0.3
0.4
0.2
0.2
0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
φ (position of node closest to shock) φ (position of node closest to shock)

Figure 4: The objective function jα (φ) with no mesh regularization (α = 0.0) corresponding to the L2 projection of the periodic
discontinuous function in (50) of varying frequency (top left: k = 1, top right: k = 3, bottom left: k = 5, bottom right: k = 7)
onto a piecewise polynomial basis of degree p = 1 ( ), p = 2 ( ), p = 3 ( ), p = 4 ( ) with |Eh,p | = 2 elements
under the mesh parameterization in (23) and Figure 2. The parameter φ is swept over the range [x∗ − 0.75h, x∗ + 0.75h], where
x∗ is the location of the discontinuity ( ). The red shaded region identifies the neighborhood of radius h/2 about x∗ .

numerical experiment without mesh regularization on a piecewise polynomial basis with two elements and
polynomial orders p ∈ {1, 2, 3, 4} for increasingly oscillatory discontinuous functions: (50) with k = 1,
k = 3, k = 5, k = 7; see Figure 4. As expected, for all functions and polynomial orders, the objective
function obtains its minima when the mesh is aligned with the shock. As the frequency of the oscillations
in the discontinuous function increases, the graphs of the objective function become more oscillatory yet
remain mostly monotonic within a radius of h/2 of the shock location. Monotonicity does break down for
the most extreme case (k = 7) for the lowest polynomial orders p = 1 and p = 2, which is an artifact of an
extremely underresolved basis.
Since the mesh distortion metric is defined separately for d = 1 and d > 1, it is important to repeat
the investigation into jα (φ) in d = 2 dimensions. The same mesh parametrization is applied to the first
coordinate in the two-dimensional setting with the second coordinate held fixed
φ − xl

xl + (Xi · e1 − xl )
 Xi ≤ x̄
xi (φ) · e1 = x̄ − xl
x −φ (25)
xr − (xr − Xi · e1 ) r
 Xi > x̄
xr − x̄
xi (φ) · e2 = Xi · e2
A depiction of the single degree of freedom mesh parametrization in d = 2 dimensions is provided in Figure 5.
This figure also shows the range that φ will be swept to study the behavior of jα (φ) next.
Figure 6 shows the graph of jα (φ) for various values of α corresponding to this single degree of freedom
parametrization, where uH corresponds to the L2 projection of the piecewise sinusoidal function
(
0.25 sin(4πx) − 0.5, x < 0
η(x, y) = (26)
0.25 sin(4πx) + 0.5, x > 0,

12
 1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 5: Single degree of freedom 2d mesh parametrization in (23) and problem setup for investigation into the proposed
objective function in (22) as applied to the L2 projection of the discontinuous function in (26). The choice of x̄ ( ) is chosen
to align with a line of mesh nodes ( ) near the position of the discontinuity ( ) and the parameter φ reduces to the position
of this node. The remaining nodes (x-coordinates only; y-coordinates are frozen) are determined by uniformly distributing
them in the intervals (xl , x̄) and (x̄, xr ). The first configuration (top left) shows the reference domain with a uniform nodal
distribution and the remaining configurations corresponds to the deformation with φ = x∗ , x∗ − 0.75h, x∗ + 0.75h. This range
of φ is the same range used to perform the sweep of the objective function (22) in Section 3.2.1 and Figure 6.

13
 15
10
jα (φ), α = 0.0

jα (φ), α = 1
10

5
5

0 0
0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.35 0.4 0.45 0.5 0.55 0.6 0.65

35 2,400

jα (φ), α = 1000
jα (φ), α = 10

2,300
30
2,200
25
2,100

20 2,000
0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.35 0.4 0.45 0.5 0.55 0.6 0.65

Figure 6: The objective function jα (φ) corresponding to the L2 projection of the discontinuous function in (26) onto a piecewise
polynomial basis of degree p = 1 ( ), p = 2 ( ), p = 3 ( ) with |Eh,p | = 50 elements for various values of the mesh
regularization parameter α under the mesh parameterization in (23) and Figure 5. The parameter φ is swept over the range
[x∗ − 0.75h, x∗ + 0.75h], where x∗ is the location of the discontinuity ( ). The position of x̄ in (23) is indicated with
( ); when φ = x̄ nodes are uniformly distributed and fmsh takes its minimum values. The red shaded region identifies the
neighborhood of radius h/2 about x∗ . The objective function with no mesh regularization (α = 0) in the top left plot shows
the two desired properties discussed in Section 3.2, namely, a minimum at x∗ and monotonically decreases to this minimum in
a neighborhood centered at x∗ with radius h/2. The remaining plots show increasing values of the regularization parameter α.
The minimum does not noticeably move until α is sufficiently large at which point is moves from x∗ to x̄, i.e., the objective
function transitions from prioritizing discontinuity tracking to mesh regularization.

(x, y) ∈ Ω = (−1, 1) × (−1, 1), onto a piecewise polynomial basis with 50 elements and polynomial orders
p ∈ {1, 2, 3}. Similar to the one-dimensional case, numerical studies indicate that the proposed objective
satisfies the desired properties for a broad range of problem configurations, i.e., possesses a local minima
when the mesh is aligned with the discontinuity and monotonically approaches this minima in a neighborhood
of radius about h/2. Also, as the penalty parameter is increased the mesh regularization is preferred and
the local minima at the discontinuity disappears. Similar trends were observed when using coarser and finer
discretization into piecewise polynomials.

3.2.2. Comparison with physics-based and error-based indicators: inviscid Burgers’ equation
With the merit of the proposed discontinuity indicator verified in the previous section, we compare it
with other metrics that have been used as discontinuity or refinement indicators. The first is an error-based
indicator that measures the extent that a Galerkin solution corresponding to a test and trial space Vh,p
0
satisfies the residual in an enriched test space Vhp0
0 0 0 0
h ,p h ,p
fh,p (u, x) = rh,p (uh,p , x) (27)
0 0
h ,p
where rh,p is the finite-dimensional Petrov-Galerkin residual corresponding to the trial space Vh,p and
0
test space Vhp0 defined in (10). The rationale behind this type of indicator is there will be no contribution
from regions of the domain away from the discontinuity that are well-resolved and near the discontinuity
the oscillations will be penalized by the additional test functions. We will consider two instances of this

14
class of indicators: h-refinement where h0 = h/2, p0 = p and p-refinement where h0 = h, p0 = p + 1. We
also consider a physics-based indicator based on the Rankine-Hugoniot conditions [4]. Recall the Rankine-
Hugoniot conditions for a steady state conservation law reduce to

F (u+ ) = F (u− ) on Γs , (28)

where Γs is any surface in the domain and u+ , u− are the values of the solution u on either side of Γs . This
will hold in smooth regions of the domain or in the case Γs is the shock surface; however, it will be violated
in regions in the domain with non-physical jumps that arise from non-aligned meshes.
The condition in (28) is converted to a discontinuity indicator by integrating the residual over all inter-
element faces
X Z 2
frh (u, x) = F (u− +
h,p ) − F (uh,p ) ds, (29)
∂K W
∂K∈∂Eh,p

where W ∈ RNc ×Nc is a symmetric positive semi-definite matrix that defines the local semi-norm. In this
work, we take W = e1 eT1 , where e1 ∈ Rd is the first canonical unit vector.
To study the indicators in (27)-(29) we consider the modified inviscid Burgers’ equation with a discontin-
uous source term (54) whose solution is given by (56). A complete description of this problem is deferred to
Section 4.2. Here, we let u denote the discrete representation of the numerical solution of the inviscid Burg-
ers’ equation in (54) using the test and trial space Vh,p and investigate the behavior of various indicators:
1) the proposed discontinuity tracking indicator in (18), 2) a residual-based indicator under p-refinement
(p0 = p + 1), 3) a residual-based indicator under h-refinement (h = h/2), 4) and the Rankine-Hugoniot
indicator

jshk (φ) = fshk (u(x(φ)), x(φ)) (30)

h,p+1 h,p+1
jh,p (φ) = fh,p (u(x(φ)), x(φ)) (31)
h/2,p h/2,p
jh,p (φ) = fh,p (u(x(φ)), x(φ)) (32)
jrh (φ) = frh (u(x(φ)), x(φ)). (33)

To study these indicators in isolation, the mesh regularization term is excluded. We consider the same single
degree of freedom parametrization from the previous section, i.e., x(φ) is given by (24). The graphs of these
functions over the range φ ∈ [x∗ − 0.75h, x∗ + 0.75h] where x∗ = 0.0 is the discontinuity location, are shown
in Figure 7 for a discretization with 40 elements of orders p = 1, 2, 3. It is clear from this figure that all of
the functions considered possess a local minimum when the mesh is perfectly aligned with the discontinuity
(φ = x∗ ). However, the proposed objective function is the only one that monotonically approaches this
minimum in a neighborhood of radius h/2. The other indicators are extremely oscillatory and in many cases
are only monotonic in a neighborhood of radius h/10 about the minimum. It is interesting to note that as
the polynomial order of the discretization increases, the physics- and error-based indicators becomes more
oscillatory, while the proposed indicator appears to be agnostic to the polynomial order. This implies the
physics- and error-based indicators are not suitable candidates for the objective function in the discontinuity-
tracking framework since the mesh faces would have to be initialized very close to the unknown discontinuity
for any hope of convergence. On the other hand, the monotonicity of the proposed indicator around the
discontinuity suggests that a gradient-based optimization solver will be capable of aligning the mesh with
discontinuities even in the worst case where a face is O(h/2) away from the discontinuity. In the remainder
of this document, we will solely consider the proposed discontinuity indicator with mesh regularization term
(20).

3.3. Full space optimization solver

The final ingredient to complete the description of the proposed discontinuity tracking method is the
numerical solver for the PDE-constrained optimization problem in (17) with the objective function in (20)
and constraints in (15). The first, and most common, PDE-constrained optimization solver is known as

15
 80

20 60

(φ)
jshk (φ)

h,p+1
40

jh,p
10
20

0 0
−0.06 −0.03 0 0.03 0.06 −0.06 −0.03 0 0.03 0.06

40 6

30
jh,p (φ)

4
jrh (φ)
h/2,p

20
2
10

0 0
−0.06 −0.03 0 0.03 0.06 −0.06 −0.03 0 0.03 0.06
φ (position of node closest to shock) φ (position of node closest to shock)

Figure 7: A side-by-side comparison of the behavior of the proposed objective function with standard physics- and error-based
objective functions (top left: deviation from mean (30), top right: residual-based p-refinement (31), bottom left: residual-based
h-refinement (32), bottom right: Rankine-Hugoniot conditions (33)) corresponding to the solution of the inviscid Burgers’
equation with a discontinuous source term in (54) using a piecewise polynomial basis of degree p = 1 ( ), p = 2 ( ),
p = 3 ( ) with |Eh,p | = 40 elements without mesh regularization α = 0. The single degree mesh parameterization from
(23) and Figure 2 is used and the parameter φ is swept over the range [x∗ − 0.75h, x∗ + 0.75h], where x∗ is the location of
the discontinuity ( ). The red shaded region identifies the neighborhood of radius h/2 about x∗ . Even with the added
complexity of the conservation law, the proposed objective function attains a minimum when the mesh is aligned with the
discontinuity and approaches this monotonically in a neighborhood of radius h/2. The other discontinuity indicators possess
the correct minimum, but clearly do not satisfy monotonicity.

16
the reduced space approach [26] where the PDE constraint is explicitly enforced by defining the implicit
function u(x) as the solution of r(u, x) = 0. The PDE-constrained optimization problem reduces to the
unconstrained optimization problem
minimize f (u(φ), x(φ)) (34)
φ∈RNφ

where the state variable u is no longer an optimization variable as it can be determined given φ. Standard
unconstrained optimization solvers such as quasi-Newton methods [36] can be applied to solve (34). The
reduced space approach is popular because the number of parameters is greatly reduced compared to the
full space formulation in (17) due to the elimination of u and sophisticated PDE solvers can be employed as
a black box to query the implicit function u(x) and its gradient using the sensitivity or adjoint method [26].
Despite these advantages, the reduced space approach cannot, in general, be used to solve the discontinuity-
tracking PDE-constrained optimization problem in (17). The reduced space formulation implicitly assumes
the PDE constraint (15) can be solved given x(φ) for any φ in the feasible set (taken as RNφ ). Recall
from the discussion in Section 1 that the PDE constraint cannot be solved unless x = x∗ , where x∗ is the
discrete representation of any discontinuity-aligned mesh, or enough artificial viscosity is added to stabilize
the solution procedure. Since the mesh will not align with discontinuities until convergence, the reduced
space approach requires repeated evaluation of the PDE solution at meshes not aligned with the discontinuity
and can only be used if sufficient viscosity, or Laplacian-based diffusion, is added during intermediate iter-
ations. Such an approach would lead to a heuristic homotopy method and certainly impact, if not prevent,
convergence of the optimization procedure.
To overcome the numerous issues associated with the reduced space approach to PDE-constrained opti-
mization for the discontinuity-tracking problem in (17), we propose using the full space approach [21, 10]. In
this setting, the PDE state vector u and mesh parameters φ are taken as independent optimization variables,
as suggested in the formulation in (17), and converged to their optimal values simultaneously. This implies
the solution of the PDE is never required away from a discontinuity-aligned mesh x∗ and overcomes the
fundamental difficulty with the reduced space formulation. Since u and φ are varying independently, the
PDE constraint cannot be eliminated and a nonlinearly constrained optimization solver is required. In this
work, we use the linesearch-based Sequential Quadratic Programming (SQP) method [11] with augmented
Lagrangian merit function implemented in the well-known optimization software SNOPT [19, 38] as our
constrained optimization solver. Any SQP method will reduce the nonlinear optimization problem in (17)
to a sequence of quadratic programs of the form
minimize q(∆u, ∆φ)
∆u∈RNu , ∆φ∈RNφ
∂r ∂r (35)
subject to r(uk , x(φk )) + (uk , x(φk ))∆u + (uk , x(φk ))∆φ = 0
∂u ∂φ
where uk and φk correspond to the kth iterate of the optimization variables and q : RNu × RNφ → R is the
quadratic Taylor approximation of the objective function where the Hessian is replaced by the Hessian of
the Lagrangian (or a symmetric positive definite approximation)
L(u, φ, λ) = f (u, φ) − λT r(u, x(φ))
with Lagrange multiplier λ ∈ RNu . The solution of this quadratic program provides search directions that
are used to update the optimization variables as
uk+1 = uk + τ ∆u
(36)
φk+1 = φk + τ ∆φ,
where τ is determined via a linesearch4 on a merit function [36]. It can be shown that the KKT conditions
of (35) are equivalent to the linearized KKT conditions of (17) [11, 36] and therefore SQP methods are

4 Trust-region SQP methods could be used in place of a linesearch-based SQP method. In this case, the trust-region radius

is determined at the beginning of the iteration and the quadratic program is restricted to the trust region.

17
equivalent to applying damped quasi-Newton methods to the KKT system of the nonlinear optimization
problem (17).
An optimization setting such as the linesearch-based SQP method in (35)-(36) demonstrates the supe-
riority of the full space approach compared to the reduced space approach for the discontinuity-tracking
problem. In the reduced space formulation, the complete nonlinear solution of r(u, x(φ)) = 0 for a given φ
is required, which causes severe robustness issues if x(φ) is not aligned with discontinuities and will likely
not be possible since oscillations will cause the solver to terminate at non-physical values. However, the full
space approach only requires the solution of the quadratic program in (35) that only involves the linearized
governing equation
∂r ∂r
r(uk , x(φk )) + (uk , x(φk ))∆u + (uk , x(φk ))∆φ = 0.
∂u ∂φ
The solution uk and the mesh x(φk ) are updated simultaneously based on the result of the linesearch.
Therefore, at every optimization iteration the mesh is updated alongside the solution and the linesearch (or
trust-region) prevents steps that cause non-physical values as they would adversely affect the merit function.
While we use the general nonlinear optimization software SNOPT in this work to demonstrate the
merit of the proposed discontinuity-tracking framework, there is a considerable opportunity for efficiency
to develop a specialized constrained optimization solver that takes advantage of the considerable structure
∂r
present in (17), namely, nonlinear equality constraints with an invertible partial Jacobian . Furthermore,
∂u
a specialized optimizer will be able to re-use data structures and operations from the PDE implementation
and incorporate homotopy measures (viscosity or temporal relaxation) into the discrete PDE for particularly
difficult problems. Such a solver is the subject of ongoing work.

3.4. Practical considerations: implementation, initialization, robustness, and efficiency

In this section we discuss a number of practical considerations important in the application of the pro-
posed method to interesting problems. These include relevant implementation details, initialization of the
optimization problem in (17), and the expected efficiency compared to standard steady state solvers or a
reduced space optimization formulation.

Implementation
The implementation of the proposed method consists of three main components: the residual r(u, x) and
objective f (u, x) functions, the definition of the mapping from optimization parameters to nodal coordinates
of the high-order mesh x = A(φ), and the nonlinear optimization solver for (17),
The high-dimensional nature of the optimization problem in (17) with Nu + Nφ optimization variables,
requires a gradient-based optimization solver as derivative-free methods are impractical in many dimensions.
Any such solver will require objective function and gradient evaluations as well as the constraint function
and gradient evaluations, i.e.,
∂f ∂f
f (u, x(φ)), (u, x(φ)), (u, x(φ)),
∂u ∂φ
(37)
∂r ∂r
r(u, x(φ)), (u, x(φ)), (u, x(φ)).
∂u ∂φ
These are the same terms that would be required if a reduced space optimization approach was used; however,
in that setting, the derivative terms would have to be combined using the sensitivity or adjoint method to
compute the total derivative of the objective function with respect to φ. The partial derivatives with respect
to φ are computed via a simple application of the chain rule
∂f ∂f ∂x
(u, x(φ)) = (u, x(φ)) (φ)
∂φ ∂x ∂φ
(38)
∂r ∂r ∂x
(u, x(φ)) = (u, x(φ)) (φ)
∂φ ∂x ∂φ

18
 ∂x ∂A
where = .
∂φ ∂φ
The terms involving the objective and residual function, i.e.,
∂f ∂f
f (u, x), (u, x), (u, x),
∂u ∂x (39)
∂r ∂r
r(u, x), (u, x), (u, x)
∂u ∂x
are solely dependent on the discretization of the conservation law and can be completely separated from
the mesh parametrization and optimization solver. The implementation of the objective function and its
derivatives are straightforward due to the element-wise integral form of (18) that can be expressed in terms
of element mass matrices. The residual function and its derivative with respect to the state variable u are
standard terms required by an implicit steady state solver. The derivative of the residual function with
respect to the nodal coordinates x is not standard and usually only required for the purposes of gradient-
based shape optimization [54]. However, the formulation of the conservation law on a reference domain
in (9) implies the elements over which the integrals are performed K ∈ Eh,p are independent of x and
the dependence on x is restricted to the pointwise definition of the transformed fluxes. This implies the
∂r ∂r
implementation and data structures used to compute will mimic those for .
∂x ∂u
The mesh parametrization can be treated independently of the PDE discretization and must implement
the mapping from φ to the mesh coordinates x = A(φ) as well as the Jacobian of this mapping, i.e.,

x(φ) = A(φ)
∂x ∂A (40)
(φ) = (φ).
∂φ ∂φ
The most convenient and general mesh parametrization uses the entire mesh as the optimization variables,
i.e.,
φ=x
A(φ) = φ
(41)
∂A
(φ) = I.
∂φ
However, this burdens the optimization solver due to the large number of optimization variables and likely
introduces local minima that do not correspond to discontinuity-aligned meshes. A general single-node
mesh parametrization for one-dimensional problems is presented in Section 3.2.1 and used in Sections 3.2.1-
3.2.2, 4.1-4.3. A general approach in arbitrary dimensions that effectively reduces the number of optimization
variables selects φ to represent coordinates of nodes in the vicinity of discontinuities and smoothing based on
linear elasticity is used to determine the positions of the remaining nodes. This assumes a-priori knowledge of
the shock region, which can easily be found by introducing sufficient viscosity and finding the corresponding
steady state using traditional methods. Regardless of the choice of mesh parametrization, the terms in (39)
and (40) can easily be combined to yield the required derivative information for gradient-based optimization
solvers in (37).
The full space optimization solver is the final component required to realize the proposed discontinuity-
tracking method in practice, which was discussed in Section 3.3.

Initialization
The initial guess for the optimization variables u, φ is a particularly important consideration in gradient-
based optimization with nonlinear constraints due to the presence of local minima at meshes not aligned
with discontinuities. The objective function in (20) was required to mostly eliminate local minima due
to the monotonicity constraint, but cannot be guaranteed. In this work we use an initial guess for the
mesh parameters φ0 such that the resulting mesh is identical to the reference mesh Eh,p . This simple mesh

19
initialization strategy will be shown to be quite robust in the numerical examples in Section 4. Unlike the
reduced space approach to optimization that only requires an initial guess for φ, the full space approach
requires initialization of u. Naturally, the solution of r(u, x(φ0 )) = 0 is a desirable initial guess since u0 ,
φ0 would satisfy the PDE constraint. However, this is not possible since x(φ0 ) corresponds to a mesh that
is not aligned with discontinuities. Instead, define

rν (u, x) = 0 (42)

as a discontinuous Galerkin discretization of a viscous conservation law with viscosity parameter ν

∇ · F(U ) + ∇ · Fν (U, ∇U ) = 0 in G(Ω0 , µ), (43)

where Fν becomes the zero map as ν → 0. The reference [3] provides an overview of various methods
available for the treatment of viscous terms in a DG setting. Since we are only looking for an initial guess
for the optimizer, the viscous terms in (43) can either correspond to physical viscosity, e.g., the viscous
terms in the Navier-Stokes equations, or non-physical Laplacian smoothing. Let ν̄ be a value of the viscosity
parameter that is as close to zero as possible but sufficiently large that the solution of

rν̄ (u, x(φ0 )) = 0 (44)

can be found reliably. Then the solution of (44) is taken as the initial guess for the state variable.
These heuristics provide high quality initial guesses, particularly at low polynomial orders such as p = 1,
but issues arise for high-order elements due to the introduction of local minima not necessarily sensed by
the proposed indicator (18). Therefore, we propose a homotopy method on the polynomial order since even
the p = 1 solution finds a quality approximation of the shock location and solution. The transfer of the
PDE solution u and mesh x to a higher order mesh is a trivial injection operation and the corresponding
optimization parameters φ are determined to exactly, if possible, reconstruct the high-order mesh. Collec-
tively, these initialization and homotopy strategies and our particular choice of objective function seem to,
in practice, eliminate potential local minima that do not align with discontinuities. Expections arise only in
rather pedantic cases such as a discontinuity positioned exactly in the center of an element in one-dimension.

Robustness
Recall from Section 1 and 3.2 that robustness is a key consideration in the design of the proposed
method. In particular, the objective function aligns the faces of the mesh with the discontinuities to avoid
severe oscillations in the solution and a full space optimization approach is used to ensure the PDE solution
is never required at non-aligned meshes. These measures effectively solve the robustness issue of traditional
steady-state solvers provided the optimization problem is properly initialized as discussed in the previous
section.

Efficiency
A natural question regarding efficiency of the proposed method arises given the shift from a traditional
nonlinear system of equations in Nu variables to a nonlinearly constrained optimization problem in Nu + Nφ
variables and Nu constraints. In fact, the proposed discontinuity-tracking method is expected to perform
similarly to a traditional steady state solver on a given mesh. This comes from the observation that per
iteration an implicit steady state solver will require evaluation of the residual and its Jacobian with respect
to u and a linear solve with this Jacobian. A full space solver will also require these operations as well as
the additional terms in (37), i.e., the objective function and its derivatives and the Jacobian of the residual
with respect to x. The evaluation of the objective function and its derivatives are inexpensive operations:
they require a single pass over the mesh, have obvious data structures, and a parallel implementation is
straightforward. The evaluation of the Jacobian of the residual with respect to x will incur a cost similar to
the evaluation of the Jacobian with respect to u.
Despite a similar per iteration performance on a given mesh, the proposed method that tracks disconti-
nuities in the solution is expected to achieve orders of magnitude reduction in degrees of freedom compared

20
to traditional shock capturing methods such as limiting and artificial viscosity, even when combined with
adaptive mesh refinement. This reduction is expected because the discontinuity is perfectly represented by
the solution basis given there is sufficient resolution for the isoparametric elements to track the geometry
of the shock. Thus mesh resolution is driven by the geometry of the shock and the smooth features in the
solution away from the shock, which allows for very coarse discretizations on high-order meshes as will be
seen in Section 4.

3.5. Summary of proposed discontinuity-tracking method

Sections 3.1-3.4 provide a complete description of the proposed optimization-based discontinuity-tracking
method. The key ingredients of the proposed method are:
• a high-order discretization of steady conservation laws in (1) that supports discontinuities along element
faces and is parametrized by the positions of the high-order mesh nodes; in this work we use a high-order
discontinuous Galerkin method with isoparametric elements
• the PDE-constrained optimization formulation in (16)
• a parametrization of the mesh as x = A(φ) where φ is a reduced set of optimization parameters and A
maps them to the nodes of the high-order mesh, possibly including mesh smoothing; the most general
choice is the identity map, with other options presented in Section 3.2.1 and 4.4
• the objective function in (20) that penalizes spurious oscillations and mesh distortion; numerical ex-
periments on one-dimensional test cases in Section 3.2.1- 3.2.2 show this indicator has a local minima
at discontinuity-aligned mesh and decreases monotonically to this minima in a neighborhood of radius
approximately h/2

• the full space PDE-constrained optimization solver that simultaneously converges the solution and
mesh to their optimal values without ever requiring the PDE solution on a non-aligned mesh and
• the practical issue of the initialization of the non-convex optimization problem in (16) with viscosity
solutions of the conservation law or the result of the discontinuity tracking method at a lower polynomial
order.

4. Numerical experiments

4.1. L2 projection of discontinuous function onto piecewise polynomial basis

In this section, the proposed framework is investigated without the complication of a conservation law.
Instead, an L2 projection of a given discontinuous function onto a piecewise polynomial space V̂h,p (x) induced
by the discretization Eh,p of the reference domain Ω0 is considered. The purpose of this section is three-fold:
1) demonstrate the merit of the proposed framework across a range of mesh sizes and polynomial orders,
2) study the convergence behavior of the full space approach, and 3) verify optimal p + 1 convergence rates
for the proposed discontinuity-tracking framework and compare to alternative methods.
The L2 projection of a function η ∈ L2 (Ω) onto V̂h,p (x) is defined as the function u ∈ V̂h,p (x) such that
Z Z
ψu dV = ψη dV, ∀ψ ∈ V̂h,p (x). (45)
Ω Ω

Expansion of the solution and test function in the basis {ϕi } and restriction to element K ∈ Eh,p leads to
the discrete problem
Z ! Z
ϕi ϕj dV uK
K K
i = ϕK
i η dV, i ∈ {1, . . . , Np } (46)
G(K, x) G(K, x)

21
that can be written compactly as
M K (x) · uK = η K (x), (47)
where the element mass matrix and contribution of the function η are
Z Z
K K K K
Mij (x) = ϕi ϕj dV, η i (x) = ϕKi η dV, i, j ∈ {1, . . . , Np }. (48)
G(K, x) G(K, x)

Therefore, the discrete residual for the case of an L2 projection takes the form

r(u, x) = M (x) · u − η(x) = 0. (49)

In general, discontinuities present in η will not align with element faces and care must be taken to ensure
the right-hand side of (46) is integrated accurately using quadrature. In this work, for any element K ∈ Eh,p
such that G(K, x) contains discontinuities of η, we compute special quadrature rules that do not traverse
these discontinuities. This special case of a L2 projection does not require the full space solver advocated for
in Section 3.3 since the governing equation in (49) can be solved given any x that does not invert the mesh,
i.e., the implicit function u(x) is well-defined, and the reduced space approach is a valid option. However,
we use the full space approach to test the proposed method in this simple scenario. The mesh x is initialized
as a uniform mesh and u is initialized as the solution of (49) on the uniform mesh, regardless of polynomial
order.
We begin by considering the discontinuous, piecewise sinusoidal function in one spatial dimension:
(
0.25 sin(2πkx) − 0.5, x<0
ηk (x) = (50)
0.25 sin(2πkx) + 0.5, x>0

for all x ∈ Ω = (−1, 1). The one-dimensional mesh Eh,p is parametrized with one degree of freedom corre-
sponding to each non-boundary node in the continuous, high-order mesh. The results of the discontinuity-
tracking method, applied to the L2 projection of (50) with k = 2, are shown in Figures 8-9 for a range of
mesh sizes and polynomial orders. For an appropriate choice of α (α = 1 in this case), the discontinuity
is tracked and the solution is well-resolved with as few as five elements of degree p = 3. If α is too small,
mesh regularization is ignored and the element containing the discontinuity is crushed. If α is too large,
discontinuity-tracking is ignored and the mesh remains uniform.
The convergence history of the full space optimization solver corresponding to the mesh with 17 p = 1
elements for α = 1 is provided in Figure 10. It shows the solution of the discretized PDE is only required
at convergence, i.e., once the objective is minimized and the grid is aligned with the discontinuities. In
fact, the PDE constraint violation is relatively large until the objective function has been reduced to nearly
its final value and the discontinuity is closely tracked. This is a feature of the full space approach and
provides robustness not possible with the reduced space approach since a PDE solution is never required on
non-aligned meshes.
The convergence of the L2 projection of (50) with k = 2 onto a sequence of increasingly refined meshes
with polynomial orders p = 1, 2, 3, 4, 5, 6 to the exact function is provided in Figure 11. To ensure the
mesh regularization term does not impede convergence, the regularization parameter is set to zero (α = 0)
and the single degree of freedom parametrization of (2) is used. The L1 error
Z
eL1 (u, uh,p ) = |u − uh,p | dV (51)
Ω

is used to measure convergence of the finite-dimensional solution uh,p to the analytical solution u given by
(50) with k = 2. Despite the L1 norm being the weakest norm, it is appropriate to study problems with
discontinuities as it encapsulates the error in the smooth region of the solution and discontinuity location
and is expected to converge at the appropriate rate of O(hp+1 ) if the discontinuity location converges at
this rate. Figure 11 shows the expected convergence rate for polynomial orders up to p = 6 and a highly
accurate solution is possible with as few as four p = 6 elements.

22
 1
0.5
0
−0.5
−1
1
0.5
0
−0.5
−1
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x x x

Figure 8: Results of discontinuity-tracking framework using the objective function defined in (20) with α = 1 applied to the L2
projection of the discontinuous function in (50) with k = 2 onto discretizations with |Eh,p | = 5 (left column), |Eh,p | = 9 (middle
column), |Eh,p | = 17 (right column) elements and polynomial order p = 1 (top row ) and p = 3 (bottom row ). The p = 2 results
are excluded for brevity as they are visually identical to the p = 3 results. Legend: piecewise sinusoidal function in (50) with
k=2( ), mesh ( ) and solution ( ) obtained from discontinuity-tracking framework.

1
0.5
0
−0.5
−1
1
0.5
0
−0.5
−1
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
x x x

Figure 9: Results of discontinuity-tracking framework using the objective function defined in (20) with α = 0 (top row ) and
α = 10 (bottom row ) applied to the L2 projection of the discontinuous function in (50) with k = 2 onto discretizations with
|Eh,p | = 5 (left column), |Eh,p | = 9 (middle column), |Eh,p | = 17 (right column) elements and polynomial order p = 1. Legend:
piecewise sinuosity function in (50) with k = 2 ( ), mesh ( ) and solution ( ) obtained from discontinuity-tracking
framework.

23
 10−1

)
10−2

)
Error in discontinuity location (
10−3
)

1.5

Residual norm, kr(u, x)k∞ (

10−4
Objective function (

10−5
1
10−6
10−7
0.5 10−8
10−9
10−10
0
0 20 40 60 80 100 120 140
Iteration

Figure 10: Convergence history of full space optimization solver (SNOPT [19]) applied to discontinuity-tracking optimization
problem in (17) with objective function in (20) (α = 1) corresponding to the L2 projection of the discontinuous function in
(50) with k = 2 onto a mesh with |Eh,p | = 17, p = 1 elements. Legend: Value of the objective function ( , left axis), the
residual norm kr(u, x)k∞ ( , right axis), and error in the location of the discontinuity ( , right axis).

10−2

10−4
L1 error

10−6

10−8

10−10

10−12
101 102 103
Number of elements

Figure 11: Convergence of discontinuity-tracking method using the single degree of freedom parametrization in (24) and no
mesh regularization (α = 0) applied to the L2 projection of the discontinuous function in (50) with k = 2 for polynomial orders
p=1( ), p = 2 ( ), p = 3 ( ), p = 4 ( ), p = 5 ( ), p = 6 ( ). The expected convergence rates of p + 1
are obtained in all cases.

24
 100

10−3
L1 error

10−6

10−9

10−12 0
10 101 102 103 100 101 102 103 100 101 102 103
Number of elements Number of elements Number of elements

Figure 12: Comparison of convergence of discontinuity-tracking method (single degree of freedom parametrization in (24) and
no mesh regularization, α = 0) to uniform and adaptive mesh refinement at fixed polynomial orders for the L2 projection of
the discontinuous function in (50) with k = 2. The discontinuity-tracking method and adaptive mesh refinement achieve the
expected p + 1 convergence rate in the asymptotic regime while uniform refinement is limited to first order due to poor shock
resolution. Legend: discontinuity-tracking with p = 1 ( ), p = 3 ( ), p = 5 ( ); uniform refinement with p = 1
( ), p = 3 ( ), p = 5 ( ); and adaptive mesh refinement with p = 1 ( ), p = 3 ( ), p = 5 ( ).

With the convergence rates of the proposed method established, we compare its performance to alternative
methods: 1) uniform refinement and 2) adaptive mesh refinement. The mesh adaptation algorithm uses a
straight-forward refinement strategy where elements with large values of the residual in an enriched (h-
refined) test space are split. The comparison is provided in Figure 12 for p = 1, 3, 5. As expected, uniform
refinement is only first-order O(h) accurate and the adaptive mesh refinement achieves the optimal order
of accuracy O(hp+1 ) eventually. Our discontinuity-tracking provides better accuracy for a given number of
elements across a range of mesh sizes and polynomial orders since the discontinuity is accurately tracked
and the basis functions are optimally utilized in regions of the domain where the solution is smooth.
To close this section, we consider the L2 projection of the discontinuous, piecewise constant function in
two spatial dimensions where the discontinuity surface is a circle
(
2, x2 + y 2 < r2
η(x) = (52)
1, x2 + y 2 > r2

for all (x, y) ∈ Ω = (0, 1) × (0, 1). Due to the difficulty associated with producing quadrature rules that do
not cross discontinuities in two-dimensions on curved meshes, the discontinuous function in (52) is replaced
with a smooth approximation
  q 
1
ην (x) = 1 + 1 − tanh ν −1 (x − cx )2 + (y − cy )2 − r (53)
2

where ν is a viscosity-like parameter that controls the steepness of the smoothed discontinuity. In this work,
ν = 10−3 is used. To ensure the term on the right-hand-side of (46) is integrated accurately, a Gaussian
quadrature rule with 20 quadrature points is used. The results of the discontinuity-tracking framework
applied to the L2 projection of (53) onto a coarse mesh of only 50 elements with polynomial orders p = 1, 2, 3
is shown in Figure 13. The mesh is parametrized using all continuous, high-order, non-boundary mesh nodes
for the most general parametrization. The L2 projection onto a uniform, non-aligned mesh is clearly a poor
approximation to the piecewise constant function where the discontinuity surface is a circle. Despite being
a poor approximation to the function in (52), this mesh and solution are used as the initial guess for the

25
discontinuity-tracking method with the p = 1 mesh. Subsequently, the mesh and solution produced by our
discontinuity-tracking method using polynomial order p = p̄ is used to initialize the method with polynomial
order p = p̄ + 1. The coarse mesh with p = 2 and p = 3 track the discontinuity very well, sufficiently
regularize the mesh, and accurately approximate the piecewise constant function in (52).

4.2. Inviscid Burgers’ equation with discontinuous source term

In this section, the proposed discontinuity-tracking framework is applied to the inviscid, modified one-
dimensional Burgers’ equation with a discontinuous source term from [5]
 
∂ 1 2
u = βu + f (x), for x ∈ Ω ⊂ R, (54)
∂x 2
where β = −0.1 and (
(2 + sin( πx π πx
2 ))( 2 cos( 2 ) − β), x<0
f (x) = (55)
πx
(2 + sin( 2 ))( 2 cos( πx
π
2 ) + β), x > 0.
It can be shown that the solution of (54) is
(
2 + sin( πx
2 ), x<0
u(x) = πx
(56)
−2 − sin( 2 ), x > 0.

The domain is taken as Ω = (−2, 2) with Dirichlet boundary conditions specified at both ends corresponding
to the exact solution, i.e., u(−2) = 2 and u(2) = −2. The discretization of (54) proceeds according to the
formulation in Section 2 and the numerical flux is taken as the solution of the exact Riemann problem
corresponding to the time-dependent version of (54) without source terms (β = 0 and f (x) = 0).
The one-dimensional mesh Eh,p is parametrized with one degree of freedom corresponding to each non-
boundary node in the continuous, high-order mesh. The results of the discontinuity-tracking method, applied
to the solution of the modified, inviscid Burgers’ equation are shown in Figure 14 for a range of mesh sizes
and polynomial orders. For an appropriate choice of α (α = 0.1 in this case), the discontinuity is tracked
and the solution is well-resolved with as few as five p = 3 elements.
Convergence of the proposed discontinuity-tracking method to the exact solution in (56) using a sequence
of increasingly refined meshes with polynomial orders p = 1, 2, 3, 4, 5, 6 is provided in Figure 15. Similarly
to the previous section, the L1 error is used to quantify the error in the solution and the single degree of
freedom mesh parametrization (2) without regularization (α = 0) is used to ensure mesh regularization does
not impede convergence. Figure 15 shows the expected convergence rate for most polynomial orders up to
p = 6 and a highly accurate solution is possible with as few as four p = 6 elements. Figure 16 compares the
performance of the discontinuity-tracking method to uniform and adaptive mesh refinement. As expected,
uniform refinement is only first-order O(h) accurate and adaptive mesh refinement nearly achieves the optimal
order of accuracy O(hp+1 ). For this test case involving a conservation law, the improved accuracy provided
by tracking the discontinuity is more pronounced than the previous case of an L2 projection of an analytical
function, e.g., for the p = 3 case, an error of 10−4 is possible with only 10 elements when the discontinuity
is tracked while adaptive mesh refinement requires over 700 elements.

4.3. Transonic, inviscid flow through nozzle

Our final one-dimensional example considers the relevant situation of transonic, inviscid flow through
a converging-diverging nozzle. The quasi-one-dimensional Euler equations are used to model the inviscid,
compressible flow in a variable-area stream tube A(x)
∂
(Aρu) = 0,
∂x
∂ p ∂A
(A[ρu2 + p]) = , (57)
∂x A ∂x
∂
(A[ρE + p]u) = 0
∂x

26
 1 1

0.8 0.8

0.6 0.6
y

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

1 1

0.8 0.8

0.6 0.6
y

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x

Figure 13: Results of discontinuity-tracking framework using the objective function defined in (20) with α = 1 applied to the
L2 projection of the function in (53) with steep gradients onto discretizations with |Eh,p | = 50 and polynomial order p = 1 (top
row ), p = 2 (bottom left), and p = 3 (bottom right). Top left: L2 projection of (53) onto uniform non-aligned mesh with 50
p = 1 elements. Top right: result of the discontinuity-tracking framework with p = 1 elements using the non-aligned mesh and
solution in the top left as the initial guess. Bottom left: result of the discontinuity-tracking framework with p = 2 elements
using the p = 1 aligned mesh and solution in the top right as the initial guess. Bottom right: result of discontinuity-tracking
framework with p = 3 elements using the p = 2 aligned mesh and solution in the bottom left as the initial guess.

27
 2

−2

−2
−2 −1 0 1 2 −2 −1 0 1 2 −2 −1 0 1 2
x x x

Figure 14: Results of discontinuity-tracking framework (objective function defined in (20) with α = 0.1) applied to the solution
of the inviscid Burgers’ equation (54) onto discretizations with |Eh,p | = 5 (left column), |Eh,p | = 9 (middle column), |Eh,p | = 17
(right column) elements and polynomial order p = 1 (top row ) and p = 3 (bottom row ). The p = 2 results are excluded for
brevity as they are visually identical to the p = 3 results. Legend: exact solution ( ) of the modified inviscid Burgers’
equation in (54), mesh ( ) and solution ( ) obtained from discontinuity-tracking framework.

100

10−1

10−2

10−3
L1 error

10−4

10−5

10−6

10−7

10−8
100 101 102 103
Number of elements

Figure 15: Convergence of discontinuity-tracking method using the single degree of freedom parametrization in (24) and no
mesh regularization (α = 0) applied to the modified inviscid Burgers’ equation in (54) for polynomial orders p = 1 ( ), p = 2
( ), p = 3 ( ), p = 4 ( ), p = 5 ( ), p = 6 ( ). The expected convergence rates of p + 1 are obtained in most cases. The slopes of
the best-fit lines to the data points in the asymptotic regime are: ∠ − 1.95 ( ), ∠ − 3.13 ( ), ∠ − 3.85 ( ), ∠ − 5.47
( ), ∠ − 4.36 ( ), ∠ − 8.67 ( ).

28
 10−1
L1 error

10−4

10−7 0
10 101 102 103 100 101 102 103 100 101 102 103
Number of elements Number of elements Number of elements

Figure 16: Comparison of convergence of discontinuity-tracking method (single degree of freedom parametrization in (24) and
no mesh regularization, α = 0) to uniform and adaptive mesh refinement at fixed polynomial orders for modified inviscid
Burgers’ equation. The discontinuity-tracking method achieves the expected p + 1 convergence rate in the asymptotic regime
while uniform refinement is limited to first order and adaptive mesh refinement falls short of optimal convergence rates due to
poor shock resolution. Legend: discontinuity-tracking with p = 1 ( ), p = 2 ( ), p = 3 ( ); uniform refinement with
p=1( ), p = 2 ( ), p = 3 ( ); and adaptive mesh refinement with p = 1 ( ), p = 2 ( ), p = 3 ( ).

where ρ is the fluid density, u is the fluid velocity, p is the thermodynamic pressure, and
1
ρE = ρe + ρu2 (58)
2
is the total energy. The pressure is related to ρE by the equation of state
 
1 2
p = (γ − 1) ρE − ρu (59)
2
for a perfect gas with ratio of specific heats γ = 1.4. The domain is taken as Ω = (0, 1) and the nozzle
profile takes the form
(
1 − (1 − T ) cos(π(x − 0.5)/0.8)2 x ∈ [0.1, 0.9]
A(x) = (60)
1 otherwise
where T = 0.8 is the height of the nozzle throat. The boundary conditions weakly impose the farfield
conditions ρi = 1.0, ui = 1.0, Mi = 0.40 at the inflow and ρo = 1.0, uo = 1.0, Mo = 0.45 at the outflow using
Roe’s approximate Riemann solver. A schematic of the domain, including boundary conditions, is provided
in Figure 17. The discretization of the governing equations proceed according to the formulation in Section 2
and Roe’s approximate Riemann solver corresponding to the time-dependent version of (57) is used for the
numerical flux.
The one-dimensional mesh Eh,p is parametrized with one degree of freedom corresponding to each non-
boundary node in the continuous, high-order mesh. The results of the discontinuity-tracking method, applied
to the solution of the quasi-1d Euler equations are shown in Figure 18 for a range of mesh sizes and polynomial
orders. For an appropriate choice of α (α = 0.1 in this case), the discontinuity is tracked and the solution is
well-resolved with as few as eight p = 2 elements.
Convergence of the proposed discontinuity-tracking method to a reference solution using a sequence of
increasingly refined meshes with polynomial orders p = 1, 2 is provided in Figure 19. The reference solution
is computed using DG on a highly adapted mesh near the discontinuity consisting of over 80000 p = 1
elements. Similar to the previous sections, the L1 error is used to quantify the error in the solution. To
ensure the mesh regularization term does not impede convergence, the regularization parameter is set to
zero (α = 0) and the single degree of freedom parametrization of (2) is used. Figure 19 shows the expected
convergence rate for polynomial orders p = 1, 2 and a highly accurate solution is possible with as few as 20
p = 2 elements. The discontinuity-tracking method is shown to outperform the adaptive mesh refinement
method with p = 1, at least in terms of number of elements for a required accuracy, which is limited to
second-order accuracy.

29
 A(x)

1
0.8

0.5 0.5

Figure 17: Geometry and boundary conditions for nozzle flow. Boundary conditions: inviscid wall ( ), inflow ( ), outflow
( ).

1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x x x

Figure 18: Results of discontinuity-tracking framework (objective function defined in (20) with α = 0.1) applied to the quasi-
1d Euler equations in (57) with |Eh,p | = 8 (left column), |Eh,p | = 16 (middle column), |Eh,p | = 32 (right column) elements
and polynomial order p = 1 (top row ) and p = 2 (bottom row ). Legend: reference solution to the quasi-1d Euler equations
determined from aggressive adaptive mesh refinement with over 80000 p = 1 elements ( ), mesh ( ) and solution ( )
obtained from discontinuity-tracking framework applied to quasi-1d Euler equations.

30
 100

10−1

10−2
L1 error

10−3

10−4

10−5
100 101 102 103
Number of elements

Figure 19: Convergence of discontinuity-tracking method using the single degree of freedom parametrization in (24) and no mesh
regularization (α = 0) applied to the quasi-1d Euler equations (57) for polynomial orders p = 1 ( ) and p = 2 ( ). The expected
convergence rates of p + 1 are obtained in both cases. The slopes of the best-fit lines to the data points in the asymptotic
regime are: ∠ − 1.97 ( ) and ∠ − 2.70 ( ). A reference second-order method that uses adaptive mesh refinement with
p = 1 elements ( ).

4.4. Supersonic, inviscid flow around cylinder

With the merit of the proposed discontinuity tracking framework established in one spatial dimension,
we turn our focus to its performance on two-dimensional conservation laws, specifically the two-dimensional
steady Euler equations
∂
(ρuj ) = 0,
∂xj
∂
(ρui uj + p) = 0 for i = 1, 2, (61)
∂xj
∂
(uj (ρE + p)) = 0,
∂xj
where ρ is the fluid density, u1 , u2 are the velocity components, and E is the total energy. For an ideal gas,
the pressure p has the form  
1
p = (γ − 1)ρ E − uk uk , (62)
2
where γ is the adiabatic gas constant. The discretization of the governing equations proceed according to
the formulation in Section 2 and Roe’s approximate Riemann solver corresponding to the time-dependent
version of (61) is used for the numerical flux. All farfield and inviscid wall boundary conditions are weakly
imposed using the Roe solver.
The two-dimensional problem considered is supersonic flow around a circle at Mach 2. Due to symmetry,
only one quarter of the domain is modeled; see Figure 20 for a schematic of the domain and boundary
conditions. For simplicity, we use a simple mesh parametrization that does not include the position of all
the nodes of the continuous high-order mesh, but rather a well-chosen subset of these nodes. The remainder
of the nodes are determined through a linear operator that incorporates mesh smoothing. In this work, we
use linear elasticity with prescribed displacements at the parametrized nodes and in the normal direction
along domain boundaries. For this problem, we use a mesh with only |Eh,p | = 48 elements and polynomial

31
 8

3 1

Figure 20: Geometry and boundary conditions for supersonic flow around cylinder. Boundary conditions: inviscid
wall/symmetry condition ( ) and farfield ( ).

Figure 21: Reference domain for and mesh for supersonic flow around cylinder with |Eh,p | = 48 elements and polynomial orders
p = 1 (left), p = 2 (middle left), p = 3 (middle right), and p = 4 (right). The blue circles identify parametrized nodes, i.e.,
nodal positions that whose displacements compose the optimization variables φ. Only the displacement normal to the boundary
are taken as optimization variables for the two parametrized nodes that lie on boundaries.

degree up to p = 4. Figure 21 shows these meshes and identifies the parametrized nodes whose displacements
compose φ and correspond to Nφ = 12, 24, 36, 48 for the p = 1, 2, 3, 4 meshes, respectively.
As a non-convex optimization problem underlies the discontinuity-tracking framework, the performance
of the full space solver relies on a quality initial guess. For the case of p = 1, the initial guess for the mesh
is taken as the reference mesh, shown in Figure 21, while for the p > 1 case, the mesh is initialized with
the solution of the discontinuity-tracking problem on the p − 1 mesh. The initial state vector, regardless of
polynomial degree, is taken as a viscosity solution of (61) with a global viscosity parameter ν = 0.05, i.e., a
solution of the Navier-Stokes equations, on the initial mesh, as discussed in Section 3.4.
The result of the discontinuity-tracking method using the above initialization strategy and the reference
domain in Figure 21 is provided in Figures 22-23. Figure 22 shows the high-order computational mesh super-
imposed on the fluid density, element-wise discontinuity indicator (fshk ), and element-wise mesh distortion
indicator (fmsh ) for the viscosity solution of the p = 1, non-aligned mesh (e.g., initial guess for the p = 1
discontinuity-tracking problem) and the inviscid solution using the discontinuity-tracking framework for the
p = 1, 2, 3, 4 meshes (mesh regularization parameter is set to α = 0.1). The non-aligned mesh is clearly
underresolved and the shock is smeared out. Several elements contain non-trivial under- and over-shoot,

32
which is identified by the discontinuity indicator that takes large values in elements near the discontinu-
ity. The mesh distortion indicator is identically zero since the reference mesh is used. It is clear that the
discontinuity-tracking framework, even for p = 1, considerably reduces the overshoot in the solution near
the shock at the price of mesh distortion. The p = 1 solution is still clearly underresolved and the shock
surface is faceted since the element cannot bend in the piecewise linear setting. However, the p = 2, 3, 4
reference meshes deliver quite accurate and smooth solutions away from the shock and track the smooth,
curved shock surface very well. Since the search space is enriched in the cases relative to the p = 1 case, the
discontinuity indicator and mesh distortion indicator also improve in most elements.
The bottom row of Figure 22 shows the mesh quality marginally suffers from aligning the mesh with
discontinuities, particularly in the p = 1 case. However, as mentioned above, the mesh quality improves as
the polynomial order increases due to the enriched search space and the presence of the fmsh term in the
objective function. This term also prevents intermediate optimization iterations from inverting or severely
distorting the mesh since this would cause a substantial increase in the objective function and be rejected
by the linesearch. This is significant as an inverted or distorted mesh would cause the condition number of
the Jacobian matrix to degrade, which would in turn lead to a failure of the linear solver.
Figure 23 shows the fluid density, entropy, and enthalpy corresponding to the non-aligned, viscosity solver
and the inviscid, tracking framework without the mesh edges to emphasize the smooth, high-quality solution
obtained away from the shock.
To quantify the quality of the solution obtained from the discontinuity-tracking framework, we consider
two metrics from inviscid fluid mechanics: the total enthalpy H = (ρE + p)/ρ, and the stagnation pressure.
In a steady, inviscid flow the total enthalpy is constant and the stagnation pressure, p0 , is given by
γ
2
(γ + 1)2 M∞
2
  γ−1
1 − γ + 2γM∞
p0 = p∞ 2
, (63)
γ+1 4γM∞ − 2(γ − 1)

where p∞ and M∞ are the freestream pressure and Mach number, respectively. Therefore we use the
following error metrics to quantify the performance of the discontinuity tracking framework without requiring
a reference solution
ep (u, x) = |ps (u, x) − p0 |
sR
(H(u, x) − H0 )2 dv (64)
Ω
eH (u, x) = R ,
Ω
dv
where ps (u, x) is the pressure at the stagnation point, p0 is the analytical stagnation pressure (63), H(u, x)
is the total enthalpy, and H0 is the farfield total enthalpy. These quantities are summarized in Table 1 for the
discontinuity-tracking framework at all polynomial orders considered. The table shows the discontinuity-

Polynomial order (p) 1 2 3 4

Degrees of freedom (Nu ) 576 1152 1920 2880
Enthalpy error (eH ) 0.0106 0.000462 0.00151 0.000885
Stagnation pressure error (ep ) 0.0711 0.00479 0.0112 0.000616
Number of optimization iterations 396 283 103 121

Table 1: Discontinuity-tracking performance summary, including the number of degrees of freedom and enthalpy and stagnation
pressure error for the reference mesh in Figure 21 with 48 elements and polynomials orders p = 1, p = 2, p = 3, p = 4 with mesh
regularization α = 0.1. The number of optimization iterations are reported to quantify the computational expense of the full
space solver. The solution and mesh are considered converged when the first-order optimality conditions of (17) are satisfied
within a tolerance of 10−6 .

tracking framework provides accurate solutions on discretizations with very few degrees of freedom, e.g.,
Nu ∼ O(103 ) with errors on the order of eH ∼ O(10−3 ). However, the table also shows counter-intuitive
behavior in that the enthalpy and stagnation pressure errors do not decrease monotonically as the polynomial
order is increased. The p = 2 errors are exceptionally low and without this column the error decreases
monotonically. While monotonicity is expected, it is not guaranteed by the finite element convergence theory

33
Figure 22: Discontinuity-tracking results corresponding to the reference meshes in Figure 21 with mesh edges superimposed.
Rows: fluid density (top), element-wise discontinuity indicator (middle), element-wise mesh distortion indicator (bottom).
Columns: viscosity solution on the non-aligned p = 1 mesh (left) and the inviscid, discontinuity-tracking solution on the p = 1
(middle left), p = 2 (middle), p = 3 (middle right), p = 4 (right) meshes with mesh regularization α = 0.1.

34
Figure 23: Fluid density (top), entropy (middle), and enthalpy (bottom) corresponding to viscosity solution on the non-aligned
p = 1 mesh (left) and the inviscid, discontinuity-tracking solution on the p = 1 (middle left), p = 2 (middle), p = 3 (middle
right), p = 4 (right) meshes with mesh regularization α = 0.1.

35
since the mesh is extremely coarse and the output quantities used to quantify the error are complicated
nonlinear functionals. Furthermore, the entropy plot in Figure 23 shows a minor imperfection along the
shock near the right boundary for the p = 3 and p = 4 solution, which suggests the shock is not perfectly
aligned with its true, physical location. This again may be attributed to non-monotonicity of the solution
error on coarse meshes or due to the approximate Riemann solver, Roe’s method with entropy fix, not being
consistent with the governing equations.
Table 1 also reports the number of optimization iterations required for SNOPT to satsify the first-order
optimality conditions to within a tolerance of 10−6 . Based on the discussion in Section 3.4, an optimization
iteration is comparable to a standard nonlinear iteration since the main additional cost comes from the
∂r
evaluation of the objective function and its derivatives and , which is quite small compared to a linear
∂x
solve with the Jacboian matrix. Since a p-homotopy strategy is used, the cost of, e.g., the p = 2 solution is
396 iterations on the p = 1 mesh and 283 iterations on the p = 2 mesh. These are reasonable iteration counts
given that pseudo-transient continutation is the standard nonlinear solver for supersonic flow problems, which
usually require tens to hundreds of iterations for convergence. In fact, the iteration count in Table 1 are
competitive given that our method only requires an extremely coarse mesh with 48 elements to deliver an
accurate solution.

5. Conclusion

This paper introduces a high-order accurate, nonlinearly stable discontinuity-tracking framework for
solving steady conservation laws with discontinuous solution features such as shock waves. The method
leverages the discontinuities between computational elements present in the finite-dimensional solution basis
in the context of a discontinuous Galerkin or finite volume discretization to track discontinuities in the
underlying solution. The approximate Riemann solver ensures the numerical flux is consistent with the
governing equations and provides appropriate stabilization through upwinding. Central to the proposed
tracking framework is a PDE-constrained optimization formulation of the discrete conservation law whose
objective drives the computational mesh to align with discontinuities and constraints ensure the discrete
conservation law is satisfied. The proposed objective function not only attains its minimum when the mesh
is aligned with discontinuities, but also approaches the minimum monotonically in a neighborhood of radius
h/2, which is critical for a gradient-based optimizer to locate it. A mesh regularization term is included
to avoid poor-quality meshes and is shown to not destroy the minima or monotonicity of the indicator, for
reasonable values of the mesh quality parameter α. The optimization problem is solved using a full space
optimization solver whereby the PDE solution and mesh simultaneously converge to their optimal values,
ensuring the discrete PDE solution is never required on non-aligned meshes and avoiding non-robustness
issues that arise due to Gibbs’ phenomena in these situations. This simultaneous convergence behavior of
the full space approach is demonstrated in Figure 10 where the discrete conservation law residual is quite
large until the objective is sufficiently reduced, i.e., the discontinuity is tracked.
A host of numerical experiments in one and two spatial dimensions are provided to demonstrate the
merit of proposed framework. Simple one-dimensional tests problems are provided in Section 3.2 to show
the objective function in (18) possesses the appropriate minima and monotonicity property, whereas other
popular shock indicators have the appropriate minima, but are highly oscillatory even when the mesh is nearly
aligned with the discontinuity. The entire tracking framework is studied on a slew of one- and two-dimensional
test problems, including the L2 projection of discontinuous functions onto a piecewise polynomial basis, the
modified steady inviscid Burgers’ equation with a discontinuous source term, the quasi-one-dimensional Euler
equations (transonic, inviscid flow through a nozzle), and the supersonic two-dimensional Euler equations.
We showed the framework successfully tracks discontinuities in the solution even when the nodal positions
of the entire mesh are taken as optimization variables, provided the mesh regularization parameter is set
appropriately. In the two-dimensional test problems with p > 1, the high-order mesh curves to provide a
high-order alignment with the shock surface. We also showed optimal O(hp+1 ) convergence rates in the L1
norm up to p = 6 for the L2 projection problem and inviscid Burgers’ equation and showed the discontinuity
tracking framework can significantly outperform popular alternatives such as uniform or adaptive mesh

36
refinement. Finally, the discontinuity-tracking framework proved extremely effective in resolving supersonic
(Mach 2) flow around a cylinder.
While the proposed discontinuity-tracking framework shows considerable promise as an efficient, high-
order, nonlinearly stable method for resolving discontinuous solutions of conservation laws, a number of
important research issues must be investigated for this to be a truly viable approach. First, as discussed in
Section 3.3, there is considerable structure in the optimization problem (17), e.g., only equality constraints,
partition of optimization variables into u and φ, invertible constraint Jacobian with respect to u, efficient
sparse parallel solvers available for operations involving r and its derivatives, that black-box optimizers
cannot leverage. Therefore, there is significant opportunity for efficiency and robustness by developing a
specialized solver that can leverage this structure and incorporate well-known homotopy strategies into the
constraint since r(u, x) = 0 is difficult to solve for high-speed flows that contain shocks. Additionally, it
is often the case that, during the full space iterations, the discontinuity indicator and mesh quality can be
significantly improved by performing local topology changes between elements, for example using the so-called
edge flip and face swap operations. Future work will focus on incorporating these discrete mesh updates
into the optimization solution procedure since it has the potential to drastically improve the quality of the
converged mesh and reduce the number of optimization iterations. We also intend to extend the proposed
framework to the time-dependent setting, where discrete mesh operations will be crucial. Further research
into the choice of numerical flux is also warranted since the inter-element jumps do not tend to zero under
refinement along the shock surface, an unfamiliar situation for high-order discontinuous Galerkin methods.
While the method was only introduced and tested for inviscid conservation laws, where perfect discontinuities
arise, we expect the framework to also be useful for viscous problems with smooth, high-gradient solution
features, but further investigation is required. Finally, this paper only considers relatively simple model
problems in one and two dimensions. Further testing is required for more complex flows in two and three
dimensions and will be the subject of future work.

Acknowledgments

This work was supported in part by the Luis W. Alvarez Postdoctoral Fellowship (MZ) by the Director,
Office of Science, Office of Advanced Scientific Computing Research, U.S. Department of Energy under
Contract No. DE-AC02-05CH11231 (MZ, PP) and by the AFOSR Computational Mathematics program
under grant number FA9550-15-1-0010 (MZ, PP). The content of this publication does not necessarily reflect
the position or policy of any of these supporters, and no official endorsement should be inferred.

References

[1] Mark Ainsworth and J. Tinsley Oden. A posteriori error estimation in finite element analysis. Computer
Methods in Applied Mechanics and Engineering, 142(1-2):1–88, 1997.

[2] Frédéric Alauzet and Adrien Loseille. A decade of progress on anisotropic mesh adaptation for compu-
tational fluid dynamics. Computer-Aided Design, 72:13–39, 2016.
[3] Douglas Arnold, Franco Brezzi, Bernardo Cockburn, and L Donatella Marini. Unified analysis of discon-
tinuous Galerkin methods for elliptic problems. SIAM Journal on Numerical Analysis, 39(5):1749–1779,
2002.
[4] Michael Baines, Stephen Leary, and Matthew Hubbard. Multidimensional least squares fluctuation
distribution schemes with adaptive mesh movement for steady hyperbolic equations. SIAM Journal on
Scientific Computing, 23(5):1485–1502, 2002.
[5] Garrett Barter. Shock capturing with PDE-based artificial viscosity for an adaptive, higher-order dis-
continuous Galerkin finite element method. PhD thesis, M.I.T., June 2008.

37
 [6] Francesco Bassi and S. Rebay. Accurate 2D Euler computations by means of a high order discontin-
uous finite element method. In Fourteenth International Conference on Numerical Methods in Fluid
Dynamics, pages 234–240. Springer, 1995.
[7] Carlos Baumann and J. Tinsley Oden. A discontinuous hp finite element method for the Euler and
Navier-Stokes equations. International Journal for Numerical Methods in Fluids, 31(1):79–95, 1999.
Tenth International Conference on Finite Elements in Fluids (Tucson, AZ, 1998).
[8] John Bell, Gregory Shubin, and Jay Solomon. Fully implicit shock tracking. Journal of Computational
Physics, 48(2):223–245, 1982.
[9] Ankit Bhagatwala and Sanjiva Lele. A modified artificial viscosity approach for compressible turbulence
simulations. Journal of Computational Physics, 228(14):4965–4969, 2009.
[10] George Biros and Omar Ghattas. Parallel lagrange–Newton–Krylov–Schur methods for PDE-
constrained optimization. part I: The Krylov–Schur solver. SIAM Journal on Scientific Computing,
27(2):687–713, 2005.
[11] Paul Boggs and Jon Tolle. Sequential quadratic programming for large-scale nonlinear optimization.
Journal of Computational and Applied Mathematics, 124(1):123–137, 2000.
[12] A. Burbeau, P. Sagaut, and Ch.-H. Bruneau. A problem-independent limiter for high-order Runge-Kutta
discontinuous Galerkin methods. Journal of Computational Physics, 169(1):111–150, 2001.
[13] Bernardo Cockburn and Chi-Wang Shu. The Runge-Kutta discontinuous Galerkin method for conser-
vation laws. V. Multidimensional systems. Journal of Computational Physics, 141(2):199–224, 1998.
[14] Bernardo Cockburn and Chi-Wang Shu. Runge–Kutta discontinuous Galerkin methods for convection-
dominated problems. Journal of Scientific Computing, 16(3):173–261, 2001.
[15] Bernardo Cockburn and Chi-Wang Shu. Runge-Kutta discontinuous Galerkin methods for convection-
dominated problems. J. Sci. Comput., 16(3):173–261, 2001.
[16] Alain Dervieux, David Leservoisier, Paul-Louis George, and Yves Coudière. About theoretical and
practical impact of mesh adaptation on approximation of functions and PDE solutions. International
Journal for Numerical Methods in Fluids, 43(5):507–516, 2003. ECCOMAS Computational Fluid Dy-
namics Conference, Part I (Swansea, 2001).
[17] Krzysztof Fidkowski and David L. Darmofal. Review of output-based error estimation and mesh adap-
tation in computational fluid dynamics. AIAA Journal, 49(4):673–694, 2011.
[18] A. Gargallo-Peiró, X. Roca, Jaime Peraire, and J. Sarrate. A distortion measure to validate and gen-
erate curved high-order meshes on CAD surfaces with independence of parameterization. International
Journal for Numerical Methods in Engineering, 106(13):1100–1130, 2016.
[19] Philip Gill, Walter Murray, and Michael Saunders. SNOPT: An SQP algorithm for large-scale con-
strained optimization. SIAM Journal on Optimization, 12(4):979–1006, 2002.
[20] James Glimm, Xiaolin Li, Yingjie Liu, Zhiliang Xu, and Ning Zhao. Conservative front tracking with
improved accuracy. SIAM Journal on Numerical Analysis, 41(5):1926–1947, 2003.
[21] Max Gunzburger. Perspectives in Flow Control and Optimization. SIAM, 2002.
[22] Ami Harten, Björn Engquist, Stanley Osher, and Sukumar R. Chakravarthy. Uniformly high-order
accurate essentially nonoscillatory schemes. III. Journal of Computational Physics, 71(2):231–303, 1987.
[23] Ami Harten and James Hyman. Self adjusting grid methods for one-dimensional hyperbolic conservation
laws. Journal of Computational Physics, 50(2):235–269, 1983.

38
[24] Ralf Hartmann and Paul Houston. Adaptive discontinuous Galerkin finite element methods for the
compressible Euler equations. Journal of Computational Physics, 183(2):508–532, 2002.
[25] Jan S. Hesthaven and Tim Warburton. Nodal Discontinuous Galerkin Methods: Algorithms, Analysis,
and Applications, volume 54 of Texts in Applied Mathematics. Springer, New York, 2008.

[26] Michael Hinze, René Pinnau, Michael Ulbrich, and Stefan Ulbrich. Optimization with PDE Constraints,
volume 23. Springer Science & Business Media, 2008.
[27] Antonio Huerta, E. Casoni, and Jaime Peraire. A simple shock-capturing technique for high-order
discontinuous Galerkin methods. International Journal for Numerical Methods in Fluids, 69(10):1614–
1632, 2012.

[28] Antony Jameson, W. Schmidt, and Eli Turkel. Numerical solution of the Euler equations by finite volume
methods using Runge Kutta time stepping schemes. In 14th Fluid and Plasma Dynamics Conference,
1981.
[29] Guang-Shan Jiang and Chi-Wang Shu. Efficient implementation of weighted ENO schemes. Journal of
Computational Physics, 126(1):202–228, 1996.

[30] Andreas Klöckner, Tim Warburton, and Jan S. Hesthaven. Viscous shock capturing in a time-explicit
discontinuous Galerkin method. Mathematical Modelling of Natural Phenomena, 6(3):57–83, 2011.
arXiv:1102.3190.
[31] Patrick M. Knupp. Algebraic mesh quality metrics. SIAM Journal on Scientific Computing, 23(1):193–
218, 2001.
[32] L. Krivodonova, J. Xin, J.-F. Remacle, N. Chevaugeon, and J. E. Flaherty. Shock detection and limiting
with discontinuous Galerkin methods for hyperbolic conservation laws. Applied Numerical Mathematics,
48(3):323–338, 2004.
[33] Randall LeVeque. Finite Volume Methods for Hyperbolic Problems. Cambridge Texts in Applied Math-
ematics. Cambridge University Press, Cambridge, 2002.
[34] Xu-Dong Liu, Stanley Osher, and Tony Chan. Weighted essentially non-oscillatory schemes. Journal of
Computational Physics, 115(1):200–212, 1994.
[35] David Moro-Ludena. An adaptive high order Reynolds-averaged Navier-Stokes solver with transition
prediction. PhD thesis, M.I.T., December 2014.
[36] Jorge Nocedal and Stephen Wright. Numerical Optimization. Springer, 2006.
[37] Jayandran Palaniappan, Scott T. Miller, and Robert B. Haber. Sub-cell shock capturing and spacetime
discontinuity tracking for nonlinear conservation laws. International Journal for Numerical Methods in
Fluids, 57(9):1115–1135, 2008.

[38] Ruben E. Perez, Peter W. Jansen, and Joaquim R. R. A. Martins. pyOpt: A Python-based object-
oriented framework for nonlinear constrained optimization. Structures and Multidisciplinary Optimiza-
tion, 45(1):101–118, 2012.
[39] Per-Olof Persson. Shock capturing for high-order discontinuous Galerkin simulation of transient flow
problems. In 21st AIAA Computational Fluid Dynamics Conference, San Diego, CA, Jun 2013. AIAA-
2013-3061.
[40] Per-Olof Persson, Javier Bonet, and Jaime Peraire. Discontinuous Galerkin solution of the Navier–
Stokes equations on deformable domains. Computer Methods in Applied Mechanics and Engineering,
198(17):1585–1595, 2009.

39
[41] Per-Olof Persson and Jaime Peraire. Sub-cell shock capturing for discontinuous Galerkin methods. In
44th AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada, 2006. AIAA-2006-0112.
[42] Jianxian Qiu and Chi-Wang Shu. Runge-Kutta discontinuous Galerkin method using WENO limiters.
SIAM Journal on Scientific Computing, 26(3):907–929, 2005.

[43] Philip Roe. Approximate Riemann solvers, parameter vectors, and difference schemes. Journal of
Computational Physics, 43(2):357–372, 1981.
[44] Philip Roe and Hiroaki Nishikawa. Adaptive grid generation by minimizing residuals. International
Journal for Numerical Methods in Fluids, 40(1-2):121–136, 2002.
[45] Gregory Shubin, AB Stephens, and Harland Glaz. Steady shock tracking and Newton’s method applied
to one-dimensional duct flow. Journal of Computational Physics, 39(2):364–374, 1981.
[46] Gregory Shubin, AB Stephens, Harland Glaz, Andrew Wardlaw, and LB Hackerman. Steady shock
tracking, Newton’s method, and the supersonic blunt body problem. SIAM Journal on Scientific and
Statistical Computing, 3(2):127–144, 1982.

[47] F Taghaddosi, WG Habashi, G Guevremont, and D Ait-Ali-Yahia. An adaptive least-squares method for
the compressible Euler equations. International Journal for Numerical Methods in Fluids, 31(7):1121–
1139, 1999.
[48] J.-Y. Trepanier, M. Paraschivoiu, M. Reggio, and R. Camarero. A conservative shock fitting method
on unstructured grids. Journal of Computational Physics, 126(2):421 – 433, 1996.

[49] John Van Rosendale. Floating shock fitting via lagrangian adaptive meshes. Technical Report ICASE
Report No. 94-89, Institute for Computer Applications in Science and Engineering, 1994.
[50] John Von Neumann and Robert Richtmyer. A method for the numerical calculation of hydrodynamic
shocks. Journal of Applied Physics, 21:232–237, 1950.

[51] Z.J. Wang, Krzysztof Fidkowski, Rémi Abgrall, Francesco Bassi, Doru Caraeni, Andrew Cary, Herman
Deconinck, Ralf Hartmann, Koen Hillewaert, H.T. Huynh, Norbert Kroll, Georg May, Per-Olof Pers-
son, Bram van Leer, and Miguel Visbal. High-order CFD methods: current status and perspective.
International Journal for Numerical Methods in Fluids, 72(8):811–845, 2013.
[52] Vincent Wheatley, Harish Kumar, and Patrick Huguenot. On the role of Riemann solvers in discon-
tinuous Galerkin methods for magnetohydrodynamics. Journal of Computational Physics, 229(3):660 –
680, 2010.
[53] Masayuki Yano and David L. Darmofal. An optimization-based framework for anisotropic simplex mesh
adaptation. Journal of Computational Physics, 231(22):7626–7649, 2012.
[54] Matthew J. Zahr and Per-Olof Persson. An adjoint method for a high-order discretization of deform-
ing domain conservation laws for optimization of flow problems. Journal of Computational Physics,
326(Supplement C):516 – 543, 2016.
[55] Xiangxiong Zhang and Chi-Wang Shu. On maximum-principle-satisfying high order schemes for scalar
conservation laws. Journal of Computational Physics, 229(9):3091–3120, 2010.

[56] Xiaolin Zhong. High-order finite-difference schemes for numerical simulation of hypersonic boundary-
layer transition. Journal of Computational Physics, 144(2):662–709, 1998.
[57] Jun Zhu, Jianxian Qiu, Chi-Wang Shu, and Michael Dumbser. Runge-Kutta discontinuous Galerkin
method using WENO limiters. II. Unstructured meshes. Journal of Computational Physics, 227(9):4330–
4353, 2008.

40