What is a Good Linear Finite Element?

Interpolation, Conditioning, Anisotropy, and Quality Measures

Jonathan Richard Shewchuk

Computer Science Division
University of California at Berkeley
Berkeley, California
[email protected]
 Two Communities:
Mesh Generation and Error Analysis
Mesh generation people:
Most don’t really understand the goals of
their own field!
Know from experience small & large angles
are bad (and are faintly aware why).

Numerical analysts:
Tend to derive asymptotic error bounds &
estimators (functional analysis, embedding theorems)−
not very useful to meshing people!
Meshing people can’t read functional
analysis anyway.
 Error Bounds & Quality Measures
The connections are still fuzzy.
interpolation error
element shape
element size
? discretization error
matrix conditioning
(Especially in anisotropic cases.)

My goals:
(Nearly) tight bounds on worst−case errors,
element stiffness matrix eigenvalues.
Quality measures that can choose the better
of two elements of intermediate quality.
(Suitable for numerical optimization.)
Guide mesh generators to make good elements.
 Three Criteria for Linear Elements
Let f be a function. f
Let g be a piecewise linear interpolant of
f over some triangulation. g
Criterion
Interpolation error Size very important.
|| f − g|| Shape only marginally
8

important.
Gradient interpolation error Size important.
|| f − g || Large angles bad;
8

small okay.
Element stiffness matrix Small angles bad;
maximum eigenvalue large okay.
max
 Main Assumption
Curvature of f is bounded:
f
| f d’’( p) | < c
(second directional derivative along any direction d)

Then over an element t, t

r mc
|| f − g|| 8 < c2 r 2mc
where r mc is the radius of the
t
min−containment circle/sphere
of t. [Waldron 1998.] r mc
Sharp for triangles, tetrahedra, higher dimensions...
The Importance of Approximating Gradients Accurately

|| f − g || affects discretization error in FEM.

8

In mechanics, f is the strains.

Classical Error Bounds on Gradients

Approximation theory: error bound proportional

to l max / rin . inradius of element (See Johnson.)
maximum edge length of element
Not asymptotically tight − overestimates
error for elements with small angles.
Functional analysis: asymptotically tight error
^
bound for triangles [Babuska and Aziz 1976].
^^
And tetrahedra [Jamet 1976, Krízek 1992]?
But nobody knows the constant!
 Error Bound: Gradients on Triangles
(Same assumption: bounded curvature.) f
Over a triangle t,
edge lengths inradius of t
l max l med(l min + 4 rin )
|| f − g || < c
8

4A
y
1.4 Area of t
1.2

1
Bound tight within factor of 2.
0.8

0.6

0.4

0.2

x
0 0.2 0.4 0.6 0.8 1 1.2
 Error Bound: Gradients on Triangles
(Same assumption: bounded curvature.) f
Over a triangle t,
edge lengths inradius of t
l max l med(l min + 4 rin )
|| f − g || < c < 3 c rcirc
8

4A
1.4
y
Area of t Circumradius of t
1.2
Bound tight within factor of 2.
1 o
0.8
Angle near 180
0.6 large circumradius large
0.4
t error.
0.2

x
0 0.2 0.4 0.6 0.8 1 1.2
 Error Bound: Gradients on Tetrahedra
Over a tetrahedron t,
Edge lengths of t
1
6V 1 <i < j < 4 Ai Aj l ij + max i j=i Aj l ij
|| f − g || < c 4
8

m =1 Am

Volume of t Face areas of t

z z
1.4

1.2 y
1

0.8

0.6

0.4

0.2
x
x
0 0.2 0.4 0.6 0.8 1 1.2
Good and Bad Tetrahedra for Interpolation

Good

Bad
Deriving the New Gradient Error Bounds
Start with standard approximation theory:
Choose a point p .
Take Taylor expansion of f−g about p .
Set it to zero at element vertices (d+1 equations).
Eliminate f(p)−g(p) term from equations.
Curvature bounds yield naive bound on
|| f(p) − g(p) || .
8

Repeat for each point p

in the triangle...
Deriving the New Gradient Error Bounds
Naive bound on But the naive error bound
|| f(p) − g(p) ||is is minimized at the incenter.

8
parabolic. Worst point Curvature of f is
p’ gives standard bounded, so gradient of
l max / rin bound. || f − g || is bounded.

8
p’ p’
Error Bound: Triangle Normals on Surfaces

c How much can triangle normal m

deviate from surface normal n ?

n
m Assumption: spheres tangent to
surface with radius c do not
enclose any portion of surface.

Angle between m and n (at any vertex) is at most

α + arcsin( 2 sin 2α ) + arcsin lmed , α = arcsin rcirc .
3 2c c
[Amenta, Choi, Dey, Leekha 2002.]
 Delaunay Optimality
A set of vertices has many triangulations.

In any dimensionality, the Delaunay triangulation

minimizes the largest r mc .

In two dimensions, the Delaunay triangulation

minimizes the largest r circ.

A domain has many triangulations that

respect its boundaries. Among these,
the constrained Delaunay triangulation
is optimal.
Conditioning of Global Stiffness Matrix

max : Dominated by the single worst element.

Depends on shape of element(s).
2D: Independent of element size.
3D: Largest element usually dominates.

min : Relatively independent of shape.

Directly proportional to areas/volumes
of elements.
Somewhere between smallest and
largest elements (times a constant).
 Conditioning: Maximum Eigenvalue
of Element Stiffness Matrix
(for Poisson’s Equation)
edge lengths
2 2 2 2 2 2 2 2
l + l + l + 2 (l + l + l ) − 48 A
1 2 3 1 2 3
max =
8A
Area of t
y
1.4

1.2
Maximum eigenvalue
1
is a quality measure that
0.8
prefers equilateral triangles.
0.6
Small angles are deleterious.
0.4

0.2

x
0 0.2 0.4 0.6 0.8 1 1.2
 Maximum Eigenvalue in 3D
(for Poisson’s Equation)

Eigenvalue for tetrahedron requires solving

a cubic equation.
Eigenvalue smallest for equilateral tetrahedra.
z
1.4

1.2
Dihedral angles, not planar
1
angles, are related to quality.
0.8 It’s a ‘‘well−known fact’’ that
0.6 both small and large dihedral
0.4 angles hurt conditioning...
0.2

x
0 0.2 0.4 0.6 0.8 1 1.2
 WRONG!!!
Surprise #1

A tetrahedron can have a dihedral angle arbitrarily

close to 180° with no dihedral smaller than 60°.
Such a tetrahedron does not hurt conditioning at all!
Good and Bad Tetrahedra for Conditioning

Good

Bad
 Quality Measures
Used to evaluate & choose elements.
Reciprocal of interpolation error or max eigenvalue.
Behave well as objective functions for mesh smoothing
by numerical optimization.
y z
1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

x x
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

triangles || f − g || tetrahedra
8
 Scale−Invariant Quality Measures
...measure how effective an element’s shape is for
a fixed amount of area.
|| f − g || max

8
y y
1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

x x
0 0.2 0.4 0.6 0.8 1 1.2 0 0.2 0.4 0.6 0.8 1 1.2

ideal okay ideal

 Two Types of Quality Measures

Scale−invariant quality measures:

Separate the effects of shape from the
effects of size.
Popular in mesh generation because they’re
easy to understand.

Size−and−shape measures & error bounds:

Incorporate the effects of shape and size
in one number.
Require more understanding of application.
What I advocate for most purposes.
Why Scale−Invariant Measures are Misleading

Interpolation/discretization error and

(in 3D) conditioning depend on element size too!
It’s okay (usually) for smaller elements to be worse
shaped than big ones. Smoothing, cleanup,
Delaunay refinement should take this into account.
Example: 2D scale−invariant measure for
interpolated gradients suggests ‘‘minimize large
angles.’’ But size−and−shape measure says
‘‘minimize circumradii’’ Delaunay.
Uses of Error Bounds & Quality Measures
Mesh refinement: Refine element if either
|| f − g|| or || f − g || is too large. Use
8

8
Delaunay refinement if conditioning/shape bad.
Mesh smoothing: Quality measures are designed
for numerical optimization of vertex positions.
Smoother, slightly weaker measures available.
Topological mesh improvement: Use quality
measures plus refinement bound to judge elements.
Vertex placement in advancing front meshing.
 Anisotropy and Interpolation Error

2
H = Hessian of f. Let E = H with E symmetric pos−def.
You can judge the error || f − g|| of an element t
8
by judging Et by isotropic error bounds/measures.

E
Anisotropy and Gradient Interpolation

Large angles are fine if aligned correctly and not

too large.
A good element t for controlling || f − g ||

8
is one for which Et has no large angle.

But...
Surprise #2: Superaccurate Gradients

Longer, thinner elements than expected sometimes

give the best accuracy! (For a fixed # of elements.)

But no large angles allowed for these extra−long

ones. And they must be very precisely aligned.
 Superaccurate Gradients
f−g
f−g
vmax
vmin

f−g
 Anisotropy and Conditioning
Ideal element for stiffness matrix depends on anisotropy
of the PDE, not the solution.

− ·B f =0

2 −1 F
F =B

Large and small angles are fine if Ft looks good by

isotropic standards.
Equilateral elements can be quite bad for conditioning.
 Surprise #3: Anisotropy Blues

Interpolation and conditioning don’t always

agree on the ideal aspect ratio or orientation!
Superaccuracy implies that discretization
error sometimes disagrees with both
interpolation and conditioning!
Advice:
Interpolation/discretization error can always
be improved by refining. Conditioning cannot.
With finite elements, always choose
small discretization error over interpolation.
 Concluding Questions

What about quadratic triangles and tetrahedra?

What about bilinear quads and trilinear hexes?
What if the curvature bound is in the L2 norm?