Root Mean Square Error Fourier Series

r Pn For a 2L periodic function f (x) centered about the origin.

i=1 [yi − f (xi )]2 ∞ ∞
RM SE = a0 X  nπx  X  nπx 
n f (x) = + an cos + bn sin
2 n=1
L n=1
L
Euler Numerical Method Z L
1
x i+1 = x i + hf (x i , ti ) a0 = f (x) dx
L −L
1 L
Z  nπx 
RK2 Numerical Method an = f (x) cos dx
L −L L
k 1 = f (x i , ti )
1 L
Z  nπx 
k 2 = f (x i + hk 1 /2, ti + h/2) bn = f (x) sin dx
L −L L
x i+1 = x i + hk 2
Fourier Sine Series
RK4 Numerical Method The odd continuation of f (x) defined for 0 ≤ x ≤ L
∞
k 1 = f (x i , ti ) X  nπx 
f (x) = bn sin
k 2 = f (x i + hk 1 /2, ti + h/2) n=1
L
k 3 = f (x i + hk 2 /2, ti + h/2) 2
Z L  nπx 
k 4 = f (x i + hk 3 , ti + h) bn = f (x) sin dx
L 0 L
x i+1 = x i + h(k 1 + 2k 2 + 2k 3 + k 4 )/6
Fourier Cosine Series
Converting Higher Order DEs The even continuation of f (x) defined for 0 ≤ x ≤ L
By letting x = x1 convert a nth order DE into a system ∞
of n first order DEs shown below. a0 X  nπx 
f (x) = + an cos
2 n=1
L
dxi /dt = xi+1 i = 1...n − 1 Z L
2
dxn /dt = f (x1 . . . xn ) a0 = f (x) dx
L 0
2 L
Z
Finite Difference Equations  nπx 
an = f (x) cos dx
Forward difference of the first derivative is L 0 L
yi+1 − yi
yi′ = (1st Order) Classifying PDEs
h
Backward difference of the first derivative is Auxx + 2Buxt + Cutt + f (u, ux , ut ) = 0
yi − yi−1 AC − B 2 = 0 (Parabolic)
yi′ = (1st Order) 2
h AC − B < 0 (Hyperbolic)
2
Centred difference of the first derivative is AC − B > 0 (Elliptic)
2
yi+1 − yi−1 AC − B = f (x, t) (Mixed)
yi′ = (2nd Order)
2h
Zero Dirichlet Heat PDE
Centred difference of the second derivative is
ut = c2 uxx u(x, 0) = f (x)
yi+1 − 2yi + yi−1
yi′′ = (2nd Order) u(0, t) = 0 u(L, t) = 0
h2
∞  2 2 2 
Shooting Method X c n π t  nπx 
To solve a BVP, first determine numerical solutions to u(x, t) = b n exp − sin
n=1
L2 L
IVP1 = S(x) and IVP2 = U (x) by any numerical method.
The shooting method below solves the BVP. bn are the coefficients of the Fourier sine series of f (x).

y(x) = αS(x) + βU (x) Zero Neumann Heat PDE

y ′ (x) = αS ′ (x) + βU ′ (x)
ut = c2 uxx u(x, 0) = f (x)
Symmetric Integrals ux (0, t) = 0 ux (L, t) = 0
Z L Z L ∞  2 2 2 

a0 X c n π t  nπx 
f (x) dx = 2 f (x) dx Even f (x) u(x, t) = + an exp − cos
−L 0 2 n=1
L2 L
Z L
a0 and an are the coefficients of the Fourier cosine series


f (x) dx = 0 Odd f (x)
−L of f (x).

1
Explicit Forward Finite Difference Method Zero Dirichlet d’Alembert’s Solution

ui,j+1 − ui,j ui−1,j − 2ui,j + ui+1,j utt = c2 uxx u(x, 0) = f (x) ut (x, 0) = g(x)
ut = uxx =
k h2 u(0, t) = 0 u(L, t) = 0
Z x+ct
Stable if c2 k/h2 ≤ 0.5 F (x + ct) + F (x − ct) 1
u(x, t) = + G(s) ds
2 2c x−ct
Implicit Backward Finite Difference Method F and G are Fourier sine series of f and g respectively.
ui,j+1 − ui,j
ut = Zero Neumann d’Alembert’s Solution
k
ui−1,j+1 − 2ui,j+1 + ui+1,j+1 utt = c2 uxx u(x, 0) = f (x) ut (x, 0) = g(x)
uxx =
h2
ux (0, t) = 0 ux (L, t) = 0
Always stable. F (x + ct) + F (x − ct) 1
Z x+ct
u(x, t) = + G(s) ds
2 2c x−ct
Crank-Nicholson Method
F and G are Fourier cosine series of f and g respectively.
ui,j+1 − ui,j
ut =
k CTCS Explicit Wave Method
ui−1,j − 2ui,j + ui+1,j
uxx = ui,j−1 − 2ui,j + ui,j+1
2h2 utt =
ui−1,j+1 − 2ui,j+1 + ui+1,j+1 k2
+ ui−1,j − 2ui,j + ui+1,j
2h2 uxx =
h2
Always stable. Initially (j = 0) use the second order centred difference
for the derivative condition to eliminate the ui,−1 term.
Method of Characteristics Stable if |ck/h| ≤ 1
aux + but = c
Implicit Wave Method
Taking 2 of the 3 differential combinations solves the PDE. ui,j−1 − 2ui,j + ui,j+1
utt =
dx dt du k2
= = u i−1,j−1 − 2ui,j−1 + ui+1,j−1
a b c uxx =
2h2
Solution is unique on all characteristic curves passing ui−1,j+1 − 2ui,j+1 + ui+1,j+1
through the initial condition. +
2h2
Initially the finite difference equation is the same as the
FTFS Method
CTCS method. Always stable.
ui,j+1 − ui,j ui+1,j − ui,j
ut = ux =
k h Elliptic PDEs
Laplace PDE is uxx + uyy = 0
Stable if a/b ≤ 0 and |ak/(bh)| ≤ 1
Poisson PDE is uxx + uyy = f (x, y)
Both PDEs are solved analytically by using separation of
FTBS Method
variables u = F (x)G(y) or u = F (x) + G(y)
ui,j+1 − ui,j ui,j − ui−1,j
ut = ux = Implicit Method
k h
Stable if a/b ≥ 0 and |ak/(bh)| ≤ 1 ui−1,j − 2ui,j + ui+1,j
uxx =
h2
Lax-Wendroff Method ui,j−1 − 2ui,j + ui,j+1
uyy =
Convert the PDE using second order centred differences. k2
Then create another scheme by differentiating the PDE Always stable.
with respect to each variable to create another 2 PDEs.
Eliminate the mixed derivative, then convert the PDE Gauss-Newton Method
using second order centred differences. Eliminate the The (model - actual) differences f vector, the Jacobian J
ui,j−1 term. Stable if |ak/(bh)| ≤ 1 matrix, and the b vectors are
   
f0 ∂f0 /∂A ∂f0 /∂B
d’Alembert’s Solution  
A
f =  ...  J = .. .. b =
   
The 1D wave PDE defined on the interval 0 ≤ x ≤ L . .  B
fn ∂fn /∂A ∂fn /∂B
utt = c2 uxx u(x, 0) = f (x) ut (x, 0) = g(x)
Better estimates for the parameters are found by using
Z x+ct
f (x + ct) + f (x − ct) 1
u(x, t) = + g(s) ds b N EW = b OLD − (J T J )−1 (J T f )
2 2c x−ct