Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Simplices and Finite Elements From Scratch: February 19, 2020

Download as pdf or txt
Download as pdf or txt
You are on page 1of 3

Simplices and finite elements from scratch

February 19, 2020

We denote by kxk the Euclidien norm of x ∈ Rd . The Euclidean scalar product of x, y ∈ Rd is either
written as xT y = hx, yi = x · y. For a measurable set A ⊂ Rd we denote by |A| its Lebesgue measure and for a
hyperplane H ⊂ Rd and B ⊂ H we debote by |B| its (d − 1)-dimensional measure.

Lemma 1. For any non-degenerate hyperplane H = x ∈ Rd nT (x − xH ) = 0 (n 6= 0) and x ∈ Rd , there


exists a unique xb ∈ H such that h := d(x, H) = inf y∈H ky − xk = kb x − xk. So we can define the projector
PH : Rd → H by PH (x) = x b. Moreover, we have PH (x) = x + hn.

Proof. The strictly convexe function 21 kx − xp k2 has a unique minimiser over the convexe set H. It can be
computed as the stationary point of the Lagrangian L(y, q) = 21 ky − xk2 + qnT (y − xH ) defined on Rd × R.
Then we have x b − x + qn = 0. This implies q = kb x − xk = h.

Lemma 2. For any non-degenerate hyperplane H = x ∈ Rd nT (x − xH ) = 0 (n 6= 0) and xp 6= H, there


exists a unique λ ∈ P 1 (Rd ) such that H = λ−1 (0) and λ(xp ) = 1; it is given by for any xH ∈ H

nT (x − xH ) d(x, H)
λ(H, xP )(x) = λ(x) = , λ(x) = . (1)
nT (xp − xH ) d(xp , H)

Proof. Since (1) is independent of the choice of xh ∈ H, we can take xH = PH (xp ), such that

nT (PH (xp ) − x) nT (PH (x) − x) d(x, H)


λ(H, xP )(x) = T
= T
=
n (PH (xp ) − xp ) n (PH (xp ) − xp ) d(xp , H)

1 Barycentric coordinates
Let K = conv x0 , x1 , · · · , xd be a non-degenerate simplex in Rd . We denote by Si = conv x0 , x1 , · · · , 
i , · · · xd
x
the face opposite to the vertex i and by ni the outward unit vector of Si , by hi = d(xi , Si ) the altitude and
xbi = PSi (xi ) ∈ Si the foot of the altitude, such that hi = kxbi − xi k.
We define the barycentric coordinates as

nT (xbi − x) d(x, Si )
λi = λ(Si , xi ) = = . (2)
hi hi

Lemma 3. (λi )0≤i≤d is the canonical basis of the P 1 finite element, and the gradients are aligned with the
corresponding normals.
d d
X X 1
λi (xj ) = δij , λk (x) = 1, λk (x)xk = x, ∇λi = − ni . (3)
hi
k=0 k=0

1
Proof. For j 6= i we have xj ∈ Si , such that λi (xj ) = 0. The following formulas follow from the unisolvance of
the P 1 finite element (they clearly hold for x = xi , and since the two functions on each side of the equal sign
are affine, they have to be identical). The gradient formula follows from differentiation of (2).

Corollary 1.
d d
X 1 X ni T xi
ni = 0, I=− (4)
hi hi
i=0 i=0

Proof. Taking derivatives in (3).

Lemma 4. We have
|K| |Sj |
Z Z
λi = , λi = (1 − δij ) (5)
K d +1 Sj d

Proof. Follows from the trapezoidal rule.

As a consequence we have for


Z
1
ψi := 1 − dλi ⇒ ψi = δij , (6)
|Sj | Sj

so (ψi )0≤i≤d is the basis of the nonconforming finite element of lowest order.

Lemma 5.
d |K| = |Si | hi (7)

Proof.
d Z
d |K|
Z Z X
ni = ∇ψi = ψi n∂K = ψi nj = |Si | ni .
hi K ∂K Sj
j=0

2 Centers
The three classical centers of a simplex K, the enter of gravity x∗K , the circumcenter x
eK and the in x̌K can be
caracterized as solutions to minimization problems inf x∈Rd f (x) with the following functions
d
X

f (x) = kx − xi k2 , fe(x) = max {kx − xi k | 0 ≤ i ≤ d} , fˇ(x) = max {d(x, Si ) | 0 ≤ i ≤ d} . (8)
i=0

Expressing these functions with the help of barycentric coordinates gives a solution procedure.
All functions are strictly convex, which yields a unique minimizer. The function for the center of gravity is
quadratic and differentiation leads to the necessary and sufficient first-order condition
d d
∗ ∗
X
∗ ∗ 1 X 1
∇f (x ) = 0 ⇔ (x − xi ) = 0 ⇔ x = xi ⇔ λi (x∗ ) = .
d+1 d+1
i=0 i=0

From (5) we get the following result.

2
Lemma 6. The center of gravity is equally given by
Z Z
∗ 1
xK = x, p = |K| p(x∗ ) ∀p ∈ P 1 (K). (9)
|K| K K

Proof. The second formula immediately holds for all λi and then by linearity for all p ∈ P 1 (K).

For the incenter and the circumcenter we note that for smooth f1 , f2 the function f (x) := max {f1 (x), f2 (x)}
is differentiable at all points with f1 (x) 6= f2 (x). If f1 (x) > f2 (x) we have ∇f (x) = ∇f1 (x) and similarly in the
other case.
We have the following result.

Lemma 7. The incenter x̌K is given in barycentric coordinates as

h−1 |Si |
λi (x̌K ) = Pd i −1 = Pd . (10)
k=0 hk k=0 |Sk |

Proof. We note that thanks to (2) and (3) ∇d(x, Si ) = −ni 6= 0. Therefore we must have d(exK , Si ) = hi λi (e
xK )
xK ) = hαi and α = 1/ dk=0 h−1
P
all equal. Let α be the common value, then λi (e k . Finally we apply (7).

For the circumcenter suppose now that for j 6= i we have ke x − xi k > ke x − xj k. Then it follows that xe = xi ,
but this means 0 > kex − xj k, which is impossible. So we have proved the classical result: all kb
x − xi k are equal,
meaning that xe is the center of the unique sphere passing through all points of the simplex. We have
 
2 2 xi + xj
kex − xi k = kex − xj k ⇒ e−
x , xj − xi = 0
2

The orthocenter [1, 2]

References
[1] N. Hadjisavvas, J.-B. Hiriart-Urruty, and P.-J. Laurent. “A characterization by optimization of the Monge
point of a tetrahedron”. In: J. Optim. Theory Appl. 171.3 (2016), pp. 856–864.
[2] J.-B. Hiriart-Urruty and P.-J. Laurent. “A characterization by optimization of the orthocenter of a triangle”.
In: Elem. Math. 70.2 (2015), pp. 45–48.

You might also like