Download - Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Transcript
Page 1: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Numerical methods for stiff Ordinarydifferential equations. Application to the

Finite Element Method (FEM)

Elisabete Alberdi Celaya

EUIT de Minas y Obras Publicas UPV/EHU, Paseo Rafael Moreno Pitxitxi 2,48013 Bilbao (Vizcaya)

April 7, 2013

Page 2: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Index

1 Introduction

2 Numerical methods for first order ODEs

3 Changing the predictor in EBDF and MEBDF methods

4 Linear multistep methods for second order ODEs

5 BDF-α method

6 Object Oriented Programming methodology

7 Results

Page 3: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Page 4: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

Page 5: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

Page 6: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

0

1

Ni(x

j)=δ

ij

Page 7: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

0

1

Ni(x

j)=δ

ij

FEM approximation: u(x, t) ≈ uh(x, t) =∑ n−1

j=2 dj (t)Nj (x)

Page 8: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

0

1

Ni(x

j)=δ

ij

FEM approximation: u(x, t) ≈ uh(x, t) =∑ n−1

j=2 dj (t)Nj (x)

Orthogonality of the residual: ∫ L0 Ni (ρcput − kuxx )dx = 0, i = 2, ..., n − 1

Page 9: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

0

1

Ni(x

j)=δ

ij

FEM approximation: u(x, t) ≈ uh(x, t) =∑ n−1

j=2 dj (t)Nj (x)

Orthogonality of the residual: ∫ L0 Ni (ρcput − kuxx )dx = 0, i = 2, ..., n − 1

Weak formulation: ∫ L0 Ni ρcput dx =

∫ L0 N′

i kuhx dx, i = 2, ..., n − 1

Page 10: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

0

1

Ni(x

j)=δ

ij

FEM approximation: u(x, t) ≈ uh(x, t) =∑ n−1

j=2 dj (t)Nj (x)

Orthogonality of the residual: ∫ L0 Ni (ρcput − kuxx )dx = 0, i = 2, ..., n − 1

Weak formulation: ∫ L0 Ni ρcput dx =

∫ L0 N′

i kuhx dx, i = 2, ..., n − 1

Ordinary Differential Equations System:

∑ n−1j=2

∫ L

0ρcpNi (x)Nj (x)dx

︸ ︷︷ ︸

mij

d ′

j (t) = −∑ n−1

j=2

∫ L

0kN′

i (x)N′

j (x)dx

︸ ︷︷ ︸

kij

dj (t), i, j = 2, ..., n − 1

Page 11: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

0

1

Ni(x

j)=δ

ij

FEM approximation: u(x, t) ≈ uh(x, t) =∑ n−1

j=2 dj (t)Nj (x)

Orthogonality of the residual: ∫ L0 Ni (ρcput − kuxx )dx = 0, i = 2, ..., n − 1

Weak formulation: ∫ L0 Ni ρcput dx =

∫ L0 N′

i kuhx dx, i = 2, ..., n − 1

Ordinary Differential Equations System:

∑ n−1j=2

∫ L

0ρcpNi (x)Nj (x)dx

︸ ︷︷ ︸

mij

d ′

j (t) = −∑ n−1

j=2

∫ L

0kN′

i (x)N′

j (x)dx

︸ ︷︷ ︸

kij

dj (t), i, j = 2, ..., n − 1

DIFFUSION EQUATION:

Md′(t) = α2K d(t),

IC : d0i = g(x i), ∀i ∈ ηd

Page 12: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A FEM application to the 1D linear diffusionequation

PDEs→ FEM approximation

Difussion:

ρcput = kuxx , ∀x ∈ [0, L], t ∈ [0,∞)

BC : u(0, t) = 0 = u(L, t)

IC : u(x , 0) = g(x), ∀x ∈ [0, L]

0

1

Ni(x

j)=δ

ij

FEM approximation: u(x, t) ≈ uh(x, t) =∑ n−1

j=2 dj (t)Nj (x)

Orthogonality of the residual: ∫ L0 Ni (ρcput − kuxx )dx = 0, i = 2, ..., n − 1

Weak formulation: ∫ L0 Ni ρcput dx =

∫ L0 N′

i kuhx dx, i = 2, ..., n − 1

Ordinary Differential Equations System:

∑ n−1j=2

∫ L

0ρcpNi (x)Nj (x)dx

︸ ︷︷ ︸

mij

d ′

j (t) = −∑ n−1

j=2

∫ L

0kN′

i (x)N′

j (x)dx

︸ ︷︷ ︸

kij

dj (t), i, j = 2, ..., n − 1

DIFFUSION EQUATION:

Md′(t) = α2K d(t),

IC : d0i = g(x i), ∀i ∈ ηd

WAVE EQUATION:

Md′′(t) = α2K d(t),

IC : d0i = g1(x i), ∀i ∈ ηd ,

(d ′

i )0

= g2(x i), ∀i ∈ ηd

Page 13: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

The 1D linear wave equation: L = 8cm, T = 16s, α2 = 1 and three Initial Conditions (g(x)):

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

1.2

Page 14: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

The 1D linear wave equation: L = 8cm, T = 16s, α2 = 1 and three Initial Conditions (g(x)):

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

1.2

Continuous solution: Separation of variables:

ρutt = Tuxx ⇒ u(x, t) =∑

k=1 Ak sin(

kπx8

)cos(ωk t), where:

ωk = kπ

8 , φk = sin(

kπx8

)

Ak = 2L

∫ L0 g(x) sin

(kπx

8

)dx

Page 15: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

The 1D linear wave equation: L = 8cm, T = 16s, α2 = 1 and three Initial Conditions (g(x)):

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

1.2

Continuous solution: Separation of variables:

ρutt = Tuxx ⇒ u(x, t) =∑

k=1 Ak sin(

kπx8

)cos(ωk t), where:

ωk = kπ

8 , φk = sin(

kπx8

)

Ak = 2L

∫ L0 g(x) sin

(kπx

8

)dx

Solution of the discrete model: Modal superposition.

Md′′(t) = −K d(t) ⇒ u(x, t) ≈ uh(x, t) =∑ n−2

k=1 Yk (0)φk (x) cos(ωk t), where:

ωk , φk

Yk (0) =φT

kMgh(x)

φTk

Mφk

Page 16: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

The 1D linear wave equation: L = 8cm, T = 16s, α2 = 1 and three Initial Conditions (g(x)):

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

1.2

Continuous solution: Separation of variables:

ρutt = Tuxx ⇒ u(x, t) =∑

k=1 Ak sin(

kπx8

)cos(ωk t), where:

ωk = kπ

8 , φk = sin(

kπx8

)

Ak = 2L

∫ L0 g(x) sin

(kπx

8

)dx

Solution of the discrete model: Modal superposition.

Md′′(t) = −K d(t) ⇒ u(x, t) ≈ uh(x, t) =∑ n−2

k=1 Yk (0)φk (x) cos(ωk t), where:

ωk , φk

Yk (0) =φT

kMgh(x)

φTk

Mφk

100 element discretization:

0 20 40 60 80 1000

5

10

15

20

25

30

35

40

45

number of the frequence

valu

e of

the

freq

uenc

e

discretcontinuous

Figure: Frequencies of thediscrete and continuous models.

Page 17: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

The 1D linear wave equation: L = 8cm, T = 16s, α2 = 1 and three Initial Conditions (g(x)):

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7 80

0.2

0.4

0.6

0.8

1

1.2

Continuous solution: Separation of variables:

ρutt = Tuxx ⇒ u(x, t) =∑

k=1 Ak sin(

kπx8

)cos(ωk t), where:

ωk = kπ

8 , φk = sin(

kπx8

)

Ak = 2L

∫ L0 g(x) sin

(kπx

8

)dx

Solution of the discrete model: Modal superposition.

Md′′(t) = −K d(t) ⇒ u(x, t) ≈ uh(x, t) =∑ n−2

k=1 Yk (0)φk (x) cos(ωk t), where:

ωk , φk

Yk (0) =φT

kMgh(x)

φTk

Mφk

100 element discretization:

0 20 40 60 80 1000

5

10

15

20

25

30

35

40

45

number of the frequence

valu

e of

the

freq

uenc

e

discretcontinuous

Figure: Frequencies of thediscrete and continuous models.

0 1 2 3 4 5 6 7 8−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

Figure: Modes 1, 2 and 10 (continuous and discrete).

Page 18: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

0 1 2 3 4 5 6 7 8−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

Figure: Mode 99 of the continuous.

0 1 2 3 4 5 6 7 8−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

Figure: Mode 99 of the discrete model.

Page 19: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

0 1 2 3 4 5 6 7 8−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

Figure: Mode 99 of the continuous.

0 1 2 3 4 5 6 7 8−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1

Figure: Mode 99 of the discrete model.

0 20 40 60 80 1000

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

discretecontinuous

Figure: Modal participation factors|Ak |, |Yi (0)| for pulse IC.

52 54 56 58 60 620

0.01

0.02

0.03

0.04

0.05

0.06

discretecontinuous

Figure: Modal participation factors|Ak |, |Yi (0)| for pulse IC (detail).

Page 20: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

Page 21: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

Page 22: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

Page 23: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

Page 24: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5

25 modos continuos

desplamiento nodos − tiempo

Page 25: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5

25 modos continuos

desplamiento nodos − tiempo

DISCRETES

Page 26: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5

25 modos continuos

desplamiento nodos − tiempo

DISCRETES

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=1599

desplamiento nodos − tiempo

t= 0t= 2

Page 27: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5

25 modos continuos

desplamiento nodos − tiempo

DISCRETES

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=1599

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=399

desplamiento nodos − tiempo

t= 0t= 2

Page 28: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5

25 modos continuos

desplamiento nodos − tiempo

DISCRETES

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=1599

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=399

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=99

desplamiento nodos − tiempo

t= 0t= 2

Page 29: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5

25 modos continuos

desplamiento nodos − tiempo

DISCRETES

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=1599

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=399

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=99

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=25

desplamiento nodos − tiempo

t= 0t= 2

Page 30: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

CONTINUOUS

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.51599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5399 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.599 modos continuos

desplamiento nodos − tiempo

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5

25 modos continuos

desplamiento nodos − tiempo

DISCRETES

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=1599

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=399

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=99

desplamiento nodos − tiempo

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=1600, nº modos=25

desplamiento nodos − tiempo

t= 0t= 2

The discrete model presents noise because of the high modes. By eliminating high modes,the noise disappears but the solution loses precision.

Page 31: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Integration of the ODE system which comes from FEM.

- Stiffness, makes the solution expensive (mores steps).- Stiffness → existence of eigenvalues of different magnitude in the solution.- Increase of the number of elements, increases stiffness.- Matlab odesuite: ode45, ode15s. Adaptative step size.

Difussion equation:

Md′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

Eigenvalues:100 elements: λmax = −1875, λmin = −0.15421000 elements: λmax = −187500, λmin = −0.1542

Page 32: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Integration of the ODE system which comes from FEM.

- Stiffness, makes the solution expensive (mores steps).- Stiffness → existence of eigenvalues of different magnitude in the solution.- Increase of the number of elements, increases stiffness.- Matlab odesuite: ode45, ode15s. Adaptative step size.

Difussion equation:

Md′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

Eigenvalues:100 elements: λmax = −1875, λmin = −0.15421000 elements: λmax = −187500, λmin = −0.1542

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 tiempo=16, nele=1000

desplazamiento nodos − tiempo

t= 0t= 2t= 4t= 8t= 16

Senoidal: The ode15s is 83times quicker (lesscomputation time).

Page 33: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Integration of the ODE system which comes from FEM.

- Stiffness, makes the solution expensive (mores steps).- Stiffness → existence of eigenvalues of different magnitude in the solution.- Increase of the number of elements, increases stiffness.- Matlab odesuite: ode45, ode15s. Adaptative step size.

Difussion equation:

Md′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

Eigenvalues:100 elements: λmax = −1875, λmin = −0.15421000 elements: λmax = −187500, λmin = −0.1542

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 tiempo=16, nele=1000

desplazamiento nodos − tiempo

t= 0t= 2t= 4t= 8t= 16

Senoidal: The ode15s is 83times quicker (lesscomputation time).

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1tiempo=16, nele=100

desplamiento nodos − tiempo

t= 0t= 2t= 4t= 8t= 16

Triangular: The ode15s is 114times quicker.

Page 34: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Integration of the ODE system which comes from FEM.

- Stiffness, makes the solution expensive (mores steps).- Stiffness → existence of eigenvalues of different magnitude in the solution.- Increase of the number of elements, increases stiffness.- Matlab odesuite: ode45, ode15s. Adaptative step size.

Difussion equation:

Md′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

Eigenvalues:100 elements: λmax = −1875, λmin = −0.15421000 elements: λmax = −187500, λmin = −0.1542

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 tiempo=16, nele=1000

desplazamiento nodos − tiempo

t= 0t= 2t= 4t= 8t= 16

Senoidal: The ode15s is 83times quicker (lesscomputation time).

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1tiempo=16, nele=100

desplamiento nodos − tiempo

t= 0t= 2t= 4t= 8t= 16

Triangular: The ode15s is 114times quicker.

0 1 2 3 4 5 6 7 80

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1tiempo=16, nele=100

desplamiento nodos − tiempo

t= 0t= 2t= 4t= 8t= 16

Rectangular pulse: Theode15s is 38 times quicker.

Page 35: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Wave equation:

Md′′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

d′(0) = d′0 = (0, ..., 0))T

Eigenvalues:100 elements: λmax = ±43.29i, λmin = ±0.3927i1000 elements: λmax = ±433.01i, λmin = ±0.3927i

Page 36: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Wave equation:

Md′′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

d′(0) = d′0 = (0, ..., 0))T

Eigenvalues:100 elements: λmax = ±43.29i, λmin = ±0.3927i1000 elements: λmax = ±433.01i, λmin = ±0.3927i

Senoidal:

0 1 2 3 4 5 6 7 8−1.5

−1

−0.5

0

0.5

1

1.5tiempo=16, nele=100

desplazamiento nodos− tiempo

t= 0t= 2t= 4t= 8t= 16

The ode15s is 11 times quicker (it was 83in diffusion).

Page 37: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Wave equation:

Md′′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

d′(0) = d′0 = (0, ..., 0))T

Eigenvalues:100 elements: λmax = ±43.29i, λmin = ±0.3927i1000 elements: λmax = ±433.01i, λmin = ±0.3927i

Senoidal:

0 1 2 3 4 5 6 7 8−1.5

−1

−0.5

0

0.5

1

1.5tiempo=16, nele=100

desplazamiento nodos− tiempo

t= 0t= 2t= 4t= 8t= 16

The ode15s is 11 times quicker (it was 83in diffusion).

Triangular:

0 1 2 3 4 5 6 7 8−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1tiempo=16, nele=100

desplazamiento nodos− tiempo

t= 0t= 2t= 4t= 8t= 10t= 16

The advantage of the ode15s disappears.

Page 38: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

Wave equation:

Md′′(t) = −K d(t),IC : d(0) = d0 = (g(x2), ..., g(xn−1))

T

d′(0) = d′0 = (0, ..., 0))T

Eigenvalues:100 elements: λmax = ±43.29i, λmin = ±0.3927i1000 elements: λmax = ±433.01i, λmin = ±0.3927i

Senoidal:

0 1 2 3 4 5 6 7 8−1.5

−1

−0.5

0

0.5

1

1.5tiempo=16, nele=100

desplazamiento nodos− tiempo

t= 0t= 2t= 4t= 8t= 16

The ode15s is 11 times quicker (it was 83in diffusion).

Triangular:

0 1 2 3 4 5 6 7 8−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

0.8

1tiempo=16, nele=100

desplazamiento nodos− tiempo

t= 0t= 2t= 4t= 8t= 10t= 16

The advantage of the ode15s disappears.

Pulse: The advantage of theode15s disappears.

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método ode15s, nele=400, pasos=12837, masa=cons

desplazamiento nodos − tiempo

t= 0t= 2

Page 39: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

400 elements:

Page 40: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

400 elements:

Ode15s, 12837 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método ode15s, nele=400, pasos=12837, masa=cons

desplazamiento nodos − tiempo

t= 0t= 2

Page 41: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

400 elements:

Ode15s, 12837 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método ode15s, nele=400, pasos=12837, masa=cons

desplazamiento nodos − tiempo

t= 0t= 2

Modal superposition:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=400, nº modos= 399

desplamiento nodos − tiempo

t= 0t= 2

Page 42: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

400 elements:

Ode15s, 12837 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método ode15s, nele=400, pasos=12837, masa=cons

desplazamiento nodos − tiempo

t= 0t= 2

Modal superposition:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=400, nº modos= 399

desplamiento nodos − tiempo

t= 0t= 2

HHT-α method (“α” method),1400 steps:

Page 43: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

400 elements:

Ode15s, 12837 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método ode15s, nele=400, pasos=12837, masa=cons

desplazamiento nodos − tiempo

t= 0t= 2

Modal superposition:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=400, nº modos= 399

desplamiento nodos − tiempo

t= 0t= 2

HHT-α method (“α” method),1400 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método HHT−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=0.3 γ=0.8 β=0.4225

t= 0t= 2

Page 44: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

400 elements:

Ode15s, 12837 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método ode15s, nele=400, pasos=12837, masa=cons

desplazamiento nodos − tiempo

t= 0t= 2

Modal superposition:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=400, nº modos= 399

desplamiento nodos − tiempo

t= 0t= 2

HHT-α method (“α” method),1400 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método HHT−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=0.3 γ=0.8 β=0.4225

t= 0t= 2

Newmark’s method β = 1/6,γ = 0.5, 800 steps →Superconvergence:

Page 45: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear diffusion and wave equation examples inMATLAB

400 elements:

Ode15s, 12837 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método ode15s, nele=400, pasos=12837, masa=cons

desplazamiento nodos − tiempo

t= 0t= 2

Modal superposition:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método supmod., nele=400, nº modos= 399

desplamiento nodos − tiempo

t= 0t= 2

HHT-α method (“α” method),1400 steps:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método HHT−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=0.3 γ=0.8 β=0.4225

t= 0t= 2

Newmark’s method β = 1/6,γ = 0.5, 800 steps →Superconvergence:

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método Newmark, tiempo=16, nele=400, pasos=800, masa=cons

γ=0.5, β=1/6

Page 46: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A non-linear version of the wave equation

Non linear PDE of a guitar string:

ρutt (x, t) =

T + E · S(√

1 + u2x (x, t) − 1

)

︸ ︷︷ ︸

T

uxx (x, t)

Real data: L = 0.648 m, diam = 0.41 · 10−3 m,S = 0.25 · π · diam2, frec = 329,T = 1.8002 · 102.IC: First mode of vibration.Time interval: 1 and 5 linear periods (5 linearperiods=0.015198 s)

20 elements are considered:

Page 47: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A non-linear version of the wave equation

Non linear PDE of a guitar string:

ρutt (x, t) =

T + E · S(√

1 + u2x (x, t) − 1

)

︸ ︷︷ ︸

T

uxx (x, t)

Real data: L = 0.648 m, diam = 0.41 · 10−3 m,S = 0.25 · π · diam2, frec = 329,T = 1.8002 · 102.IC: First mode of vibration.Time interval: 1 and 5 linear periods (5 linearperiods=0.015198 s)

20 elements are considered:

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método trap., tiempo=0.0030395, nele=20, pasos=200

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método ode15s, tiempo=0.0030395, nele=20, pasos=1410

Page 48: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A non-linear version of the wave equation

Non linear PDE of a guitar string:

ρutt (x, t) =

T + E · S(√

1 + u2x (x, t) − 1

)

︸ ︷︷ ︸

T

uxx (x, t)

Real data: L = 0.648 m, diam = 0.41 · 10−3 m,S = 0.25 · π · diam2, frec = 329,T = 1.8002 · 102.IC: First mode of vibration.Time interval: 1 and 5 linear periods (5 linearperiods=0.015198 s)

20 elements are considered:

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método trap., tiempo=0.0030395, nele=20, pasos=200

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método ode15s, tiempo=0.0030395, nele=20, pasos=1410

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.0030395, nele=20, pasos=200, masa=cons

α=0.3 , γ=0.8, β=0.4225

Page 49: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A non-linear version of the wave equation

Non linear PDE of a guitar string:

ρutt (x, t) =

T + E · S(√

1 + u2x (x, t) − 1

)

︸ ︷︷ ︸

T

uxx (x, t)

Real data: L = 0.648 m, diam = 0.41 · 10−3 m,S = 0.25 · π · diam2, frec = 329,T = 1.8002 · 102.IC: First mode of vibration.Time interval: 1 and 5 linear periods (5 linearperiods=0.015198 s)

20 elements are considered:

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método trap., tiempo=0.0030395, nele=20, pasos=200

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método ode15s, tiempo=0.0030395, nele=20, pasos=1410

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.0030395, nele=20, pasos=200, masa=cons

α=0.3 , γ=0.8, β=0.4225

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método trap., tiempo=0.015198, nele=20, pasos=1000

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método ode15s, tiempo=0.015198, nele=20, pasos=9425

Page 50: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

A non-linear version of the wave equation

Non linear PDE of a guitar string:

ρutt (x, t) =

T + E · S(√

1 + u2x (x, t) − 1

)

︸ ︷︷ ︸

T

uxx (x, t)

Real data: L = 0.648 m, diam = 0.41 · 10−3 m,S = 0.25 · π · diam2, frec = 329,T = 1.8002 · 102.IC: First mode of vibration.Time interval: 1 and 5 linear periods (5 linearperiods=0.015198 s)

20 elements are considered:

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método trap., tiempo=0.0030395, nele=20, pasos=200

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método ode15s, tiempo=0.0030395, nele=20, pasos=1410

0 0.5 1 1.5 2 2.5 3 3.5

x 10−3

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.0030395, nele=20, pasos=200, masa=cons

α=0.3 , γ=0.8, β=0.4225

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método trap., tiempo=0.015198, nele=20, pasos=1000

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método ode15s, tiempo=0.015198, nele=20, pasos=9425

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.015198, nele=20, pasos=1000, masa=cons

α=0.3 , γ=0.8, β=0.4225

Page 51: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Numerical methods for first order ODEs

A first order ODE is given by: y ′(t) = f (t, y(t)), y(a) = y0

Runge-Kutta methods ⇒ ode45

Linear multistep methods ⇒ BDFs ⇒ ode15s

Page 52: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Numerical methods for first order ODEs

A first order ODE is given by: y ′(t) = f (t, y(t)), y(a) = y0

Runge-Kutta methods ⇒ ode45

Linear multistep methods ⇒ BDFs ⇒ ode15s

Search of better linear multistep methods

The search of linear multistep methods with better stability and precision characteristicsfollowing 3 directions:

using high order derivatives

using superfuture-points

combining two existing methods or techniques to generate them

Page 53: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Error, stability and stiffness

Amplification factor

A method is stable if the perturbations are not amplified. Apply the method to the testequation: y ′ = λy .- Linear multistep method:

∑ kj=0 αj yn+j = h

∑ kj=0 βj yn+j , where h = λh ⇒

yn+1yn+2...

yn+k

=

a11 a12 . . . a1ka21 a22 . . . a2k

...... · · ·

...ak1 ak2 . . . akk

·

ynyn+1...

yn+k−1

⇒ Yn+k = A

(

h)

Yn+k−1

where: Yn+k = (yn+1, yn+2, ..., yn+k )T , Yn+k−1 = (yn, yn+1, ..., yn+k−1)T and A the

amplification factor.

- One-step method ⇒ Matrix A is an escalar function: yn+1 = R(

h)

yn

Numerical stability: The module of the eigenvalues of A is less than or equal to 1.

The spectral radius is the maximum module of the eigenvalues:ρ = max |ρi | : ρi eigenvalue of A

Stability region:

S =

h ∈ C :∣∣∣rj

(

h)∣∣∣ ≤ 1 ∀ h, rj root of the characteristic polynomial of A

The frontier of the stability region: h = hλ : r(h) = 1. To draw it we do: r = eiθ andθ ∈ [0, 2pi).

Page 54: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Error, stability and stiffness

0

0

Figure: A-stability (or unconditional stability).

0

−α

Figure: A(α) stability.

Precision of a method

Global truncation error: GTEn+k = y(tn+k ) − yn+kLocal truncation error: LTEn+k = y(tn+k ) − y∗

n+kLocalizing assumption to calculate y∗

n+k : yn+j = y(tn+j ), for j = 0, 1, ..., k − 1Method of order p: LTE = O(hp+1) ⇒ GTE = O(hp)

Page 55: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Runge-Kutta methods

yn+1 = yn + hs∑

i=1

biki , where: ki = f (tn + cih, yn + hs∑

j=1aijkj), i = 1, 2, ...., s

Page 56: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Runge-Kutta methods

yn+1 = yn + hs∑

i=1

biki , where: ki = f (tn + cih, yn + hs∑

j=1aijkj), i = 1, 2, ...., s

b = [b1, b2, ..., bs]T

, c = [c1, c2, ..., cs]T

, A =[aij]

c A

bT

Table: Butcher Table.

Page 57: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Runge-Kutta methods

yn+1 = yn + hs∑

i=1

biki , where: ki = f (tn + cih, yn + hs∑

j=1aijkj), i = 1, 2, ...., s

b = [b1, b2, ..., bs]T

, c = [c1, c2, ..., cs]T

, A =[aij]

c A

bT

Table: Butcher Table.

Stability:

yn+1 = R(h)yn, h = hλ

R(h) = 1 + hbT(

I − hA)−1

e, e = [1, 1, ..., 1]T ∈ Rs

Page 58: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Runge-Kutta methods

yn+1 = yn + hs∑

i=1

biki , where: ki = f (tn + cih, yn + hs∑

j=1aijkj), i = 1, 2, ...., s

b = [b1, b2, ..., bs]T

, c = [c1, c2, ..., cs]T

, A =[aij]

c A

bT

Table: Butcher Table.

Stability:

yn+1 = R(h)yn, h = hλ

R(h) = 1 + hbT(

I − hA)−1

e, e = [1, 1, ..., 1]T ∈ Rs

−4 −3 −2 −1 0 1−3i

−2i

−i

0

i

2i

3i

p=1

p=2

p=3

p=4

Page 59: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Runge-Kutta methods

yn+1 = yn + hs∑

i=1

biki , where: ki = f (tn + cih, yn + hs∑

j=1aijkj), i = 1, 2, ...., s

b = [b1, b2, ..., bs]T

, c = [c1, c2, ..., cs]T

, A =[aij]

c A

bT

Table: Butcher Table.

Stability:

yn+1 = R(h)yn, h = hλ

R(h) = 1 + hbT(

I − hA)−1

e, e = [1, 1, ..., 1]T ∈ Rs

−4 −3 −2 −1 0 1−3i

−2i

−i

0

i

2i

3i

p=1

p=2

p=3

p=4

Embedded Runge-Kutta methods: Methods oforder p and p + 1 share the coefficients ci , aij .DOPRI(5,4)→ ode45.

c A

bT

bT

ET

Table: Embedded Runge-Kutta Butcher Table.

Page 60: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Runge-Kutta methods

yn+1 = yn + hs∑

i=1

biki , where: ki = f (tn + cih, yn + hs∑

j=1aijkj), i = 1, 2, ...., s

b = [b1, b2, ..., bs]T

, c = [c1, c2, ..., cs]T

, A =[aij]

c A

bT

Table: Butcher Table.

Stability:

yn+1 = R(h)yn, h = hλ

R(h) = 1 + hbT(

I − hA)−1

e, e = [1, 1, ..., 1]T ∈ Rs

−4 −3 −2 −1 0 1−3i

−2i

−i

0

i

2i

3i

p=1

p=2

p=3

p=4

Embedded Runge-Kutta methods: Methods oforder p and p + 1 share the coefficients ci , aij .DOPRI(5,4)→ ode45.

c A

bT

bT

ET

Table: Embedded Runge-Kutta Butcher Table.

−6 −4 −2 0 2−4i

−3i

−2i

−i

0

i

2i

3i

4i

p=4

p=5

Page 61: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods

Linear multistep methods:k∑

j=0

αj yn+j = hk∑

j=0

βj fn+j

Page 62: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods

Linear multistep methods:k∑

j=0

αj yn+j = hk∑

j=0

βj fn+j

-Backward Differentiation Formulae (BDF):∑ k

j=11j ∇

j yn+k = hfn+k

-Numerical Differentiation Formulae (NDF):∑ k

j=11j ∇

j yn+k = hfn+k + κ∇k+1yn+k

−10 −5 0 5 10 15 20−15i

−10i

−5i

0

5i

10i

15i

BDF2BDF3

BDF4

BDF5

BDF1

Figure: BDF stability regions(exterior to the curves).

k κ NDF %step size BDF’s A(α) NDF’s A(α)1 -0.1850 26% 90 902 -1/9 26% 90 903 -0.0823 26% 86 804 -0.0415 12% 73 66

Table: NDFs: efficiency and stability respect to BDFs.

Page 63: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods

Linear multistep methods:k∑

j=0

αj yn+j = hk∑

j=0

βj fn+j

-Backward Differentiation Formulae (BDF):∑ k

j=11j ∇

j yn+k = hfn+k

-Numerical Differentiation Formulae (NDF):∑ k

j=11j ∇

j yn+k = hfn+k + κ∇k+1yn+k

−10 −5 0 5 10 15 20−15i

−10i

−5i

0

5i

10i

15i

BDF2BDF3

BDF4

BDF5

BDF1

Figure: BDF stability regions(exterior to the curves).

k κ NDF %step size BDF’s A(α) NDF’s A(α)1 -0.1850 26% 90 902 -1/9 26% 90 903 -0.0823 26% 86 804 -0.0415 12% 73 66

Table: NDFs: efficiency and stability respect to BDFs.

Some modifications to the linear multistep methods:- Extended BDF (EBDF):

∑ kj=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1

- Modified Extended BDF (MEBDF):∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1 + h(

βk − βk

)

fn+k

Both A-stable up to order 4.

Page 64: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Changing the predictor in EBDF and MEBDFmethods

The motivation of the change is double

- NDF-s imply few computational additional cost with respect to BDFs.- Good stability characteristics of EBDF and MEBDF.

Page 65: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Changing the predictor in EBDF and MEBDFmethods

The motivation of the change is double

- NDF-s imply few computational additional cost with respect to BDFs.- Good stability characteristics of EBDF and MEBDF.

Predictor-corrector scheme of EBDF and MEBDF:

Predict yn+k using the k step BDF.

Predict yn+k+1 of the instant tn+k+1 using the k step BDF.

Evaluate fn+k+1 = f (tn+k+1, yn+k+1) and also fn+k = f (tn+k , yn+k ) for MEBDFs.

Substitute these values in the correctors:EBDF:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1

MEBDF:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1 + h(

βk − βk

)

fn+k

Page 66: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Changing the predictor in EBDF and MEBDFmethods

The motivation of the change is double

- NDF-s imply few computational additional cost with respect to BDFs.- Good stability characteristics of EBDF and MEBDF.

Predictor-corrector scheme of EBDF and MEBDF:

Predict yn+k using the k step BDF.

Predict yn+k+1 of the instant tn+k+1 using the k step BDF.

Evaluate fn+k+1 = f (tn+k+1, yn+k+1) and also fn+k = f (tn+k , yn+k ) for MEBDFs.

Substitute these values in the correctors:EBDF:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1

MEBDF:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1 + h(

βk − βk

)

fn+k

Lemma

If the predictors used are of order k and the correctors of order k + 1, the whole algorithm isof order (k + 1).

Page 67: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Changing the predictor in EBDF and MEBDFmethods

The motivation of the change is double

- NDF-s imply few computational additional cost with respect to BDFs.- Good stability characteristics of EBDF and MEBDF.

Predictor-corrector scheme of EBDF and MEBDF:

Predict yn+k using the k step BDF.

Predict yn+k+1 of the instant tn+k+1 using the k step BDF.

Evaluate fn+k+1 = f (tn+k+1, yn+k+1) and also fn+k = f (tn+k , yn+k ) for MEBDFs.

Substitute these values in the correctors:EBDF:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1

MEBDF:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1 + h(

βk − βk

)

fn+k

Lemma

If the predictors used are of order k and the correctors of order k + 1, the whole algorithm isof order (k + 1).

Local truncation errors: EBDF: LTEk = hk+2

βk+1C1

(

1 −αk−1

αk

)∂f

∂yy (k+1)

+ C3y (k+2)

(tn) + O(

hk+3)

MEBDF: LTEk = hk+2

C1

(

βk+1

(

1 −αk−1

αk

)

+ (βk − βk )

)∂f

∂yy (k+1)

+ C4y (k+2)

(tn) + O(

hk+3)

Page 68: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Order and stability of EBDF and MEBDF

Stability:The characteristic polynomial in both cases: Ah3 + Bh2 + Ch + D = 0

where:

A = −βk r k

B = 2αk βk r k + T − βk+1S−(βk − βk )RC = −βk α2

k r k − 2αk T + αk βk+1S − βk+1αk−1R+(βk − βk )Rαk

D = α2k T

R =∑ k−1

j=0 αj rj , S =

∑ k−2j=0 αj r

j+1, T =∑ k

j=0 αj rj ,

The red coefficients are substituted by βk in EBDFs.The blue coefficients are characteristic of MEBDFs.

Page 69: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Order and stability of EBDF and MEBDF

Stability:The characteristic polynomial in both cases: Ah3 + Bh2 + Ch + D = 0

where:

A = −βk r k

B = 2αk βk r k + T − βk+1S−(βk − βk )RC = −βk α2

k r k − 2αk T + αk βk+1S − βk+1αk−1R+(βk − βk )Rαk

D = α2k T

R =∑ k−1

j=0 αj rj , S =

∑ k−2j=0 αj r

j+1, T =∑ k

j=0 αj rj ,

The red coefficients are substituted by βk in EBDFs.The blue coefficients are characteristic of MEBDFs.

Scheme of the new methods:

Mantaining the corrector:- Two NDF predictors ⇒ ENDF and MENDF.- First corrector BDF and second corrector NDF ⇒ EBNDF and MEBNDF.- First corrector NDF and second corrector BDF ⇒ ENBDF and MENBDF.

Page 70: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Order and stability of EBDF and MEBDF

Stability:The characteristic polynomial in both cases: Ah3 + Bh2 + Ch + D = 0

where:

A = −βk r k

B = 2αk βk r k + T − βk+1S−(βk − βk )RC = −βk α2

k r k − 2αk T + αk βk+1S − βk+1αk−1R+(βk − βk )Rαk

D = α2k T

R =∑ k−1

j=0 αj rj , S =

∑ k−2j=0 αj r

j+1, T =∑ k

j=0 αj rj ,

The red coefficients are substituted by βk in EBDFs.The blue coefficients are characteristic of MEBDFs.

Scheme of the new methods:

Mantaining the corrector:- Two NDF predictors ⇒ ENDF and MENDF.- First corrector BDF and second corrector NDF ⇒ EBNDF and MEBNDF.- First corrector NDF and second corrector BDF ⇒ ENBDF and MENBDF.

All the methods are of order p = k + 1:

LTEk = hk+2

(

βk+1Ak + Ci (βk − βk )) ∂f

∂yy (k+1)

+ Di y(k+2)

(tn) + O(

hk+3)

where:- the boxed part depends on the predictors.- The coefficients Ak depend on the error coefficients of the predictors.- The blue coefficients are characteristic of MEBDFs.- The coefficient Di depends on the correctors.

Page 71: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Order and stability of EBDF and MEBDF

k p (order) A(α) EBDF A(α) EBNDF A(α) ENBDF A(α) ENDF1 2 90 90 90 902 3 90 90 90 903 4 90 90 90 904 5 87.61 87.68 87.49 87.54

Table: A(α)-estabilidad de los mtodos EBDF, EBNDF,ENBDF, ENDF.

Page 72: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Order and stability of EBDF and MEBDF

k p (order) A(α) EBDF A(α) EBNDF A(α) ENBDF A(α) ENDF1 2 90 90 90 902 3 90 90 90 903 4 90 90 90 904 5 87.61 87.68 87.49 87.54

Table: A(α)-estabilidad de los mtodos EBDF, EBNDF,ENBDF, ENDF.

−1 0 1 2 3 4 5 6 7

−4i

−3i

−2i

−i

0

i

2i

3i

4ik=4

k=3

k=2

k=1

EBDFENDF

Page 73: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Order and stability of EBDF and MEBDF

k p (order) A(α) EBDF A(α) EBNDF A(α) ENBDF A(α) ENDF1 2 90 90 90 902 3 90 90 90 903 4 90 90 90 904 5 87.61 87.68 87.49 87.54

Table: A(α)-estabilidad de los mtodos EBDF, EBNDF,ENBDF, ENDF.

−1 0 1 2 3 4 5 6 7

−4i

−3i

−2i

−i

0

i

2i

3i

4ik=4

k=3

k=2

k=1

EBDFENDF

k p (order) A(α) MEBDF A(α) MEBNDF A(α) MENBDF A(α) MENDF1 2 90 90 90 902 3 90 90 90 903 4 90 90 90 904 5 88.36 88.41 88.88 88.93

Table: A(α)-estabilidad de los mtodos MEBDF, MEBNDF,MENBDF, MENDF.

Page 74: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Order and stability of EBDF and MEBDF

k p (order) A(α) EBDF A(α) EBNDF A(α) ENBDF A(α) ENDF1 2 90 90 90 902 3 90 90 90 903 4 90 90 90 904 5 87.61 87.68 87.49 87.54

Table: A(α)-estabilidad de los mtodos EBDF, EBNDF,ENBDF, ENDF.

−1 0 1 2 3 4 5 6 7

−4i

−3i

−2i

−i

0

i

2i

3i

4ik=4

k=3

k=2

k=1

EBDFENDF

k p (order) A(α) MEBDF A(α) MEBNDF A(α) MENBDF A(α) MENDF1 2 90 90 90 902 3 90 90 90 903 4 90 90 90 904 5 88.36 88.41 88.88 88.93

Table: A(α)-estabilidad de los mtodos MEBDF, MEBNDF,MENBDF, MENDF.

−1 0 1 2 3 4 5 6 7

−4i

−3i

−2i

−i

0

i

2i

3i

4i

k=1

k=3

k=2

MEBDF

MENDF

k=4

Page 75: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods for second orderODEs

Stiffness

The second order ODE systems obtained after discretizing the wave-type PDE by the FEMare stiff. The high modes are result of the discretization and they are not representative. Theirresolution requires:- the use of good stability numerical methods.- controlled numerical dissipation in the range of the high frequencies.

Page 76: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods for second orderODEs

Stiffness

The second order ODE systems obtained after discretizing the wave-type PDE by the FEMare stiff. The high modes are result of the discretization and they are not representative. Theirresolution requires:- the use of good stability numerical methods.- controlled numerical dissipation in the range of the high frequencies.

Alfa-generalized method: Different weighting of the inertia forces and the rest of the addends:

Man+1−αm + Cvn+1−αf+ Kdn+1−αf

= F(

tn+1−αf

)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

where:

dn+1−αf= (1 − αf ) dn+1 + αf dn

vn+1−αf= (1 − αf ) vn+1 + αf vn

an+1−αm = (1 − αm) an+1 + αman

tn+1−αf= (1 − αf ) tn+1 + αf tn

Page 77: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods for second orderODEs

Stiffness

The second order ODE systems obtained after discretizing the wave-type PDE by the FEMare stiff. The high modes are result of the discretization and they are not representative. Theirresolution requires:- the use of good stability numerical methods.- controlled numerical dissipation in the range of the high frequencies.

Alfa-generalized method: Different weighting of the inertia forces and the rest of the addends:

Man+1−αm + Cvn+1−αf+ Kdn+1−αf

= F(

tn+1−αf

)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

where:

dn+1−αf= (1 − αf ) dn+1 + αf dn

vn+1−αf= (1 − αf ) vn+1 + αf vn

an+1−αm = (1 − αm) an+1 + αman

tn+1−αf= (1 − αf ) tn+1 + αf tn

- If αm = 0 ⇒ HHT-α method:

Man+1 + Cvn+1−αf+ Kdn+1−αf

= F(

tn+1−αf

)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

Page 78: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods for second orderODEs

Stiffness

The second order ODE systems obtained after discretizing the wave-type PDE by the FEMare stiff. The high modes are result of the discretization and they are not representative. Theirresolution requires:- the use of good stability numerical methods.- controlled numerical dissipation in the range of the high frequencies.

Alfa-generalized method: Different weighting of the inertia forces and the rest of the addends:

Man+1−αm + Cvn+1−αf+ Kdn+1−αf

= F(

tn+1−αf

)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

where:

dn+1−αf= (1 − αf ) dn+1 + αf dn

vn+1−αf= (1 − αf ) vn+1 + αf vn

an+1−αm = (1 − αm) an+1 + αman

tn+1−αf= (1 − αf ) tn+1 + αf tn

- If αm = 0 ⇒ HHT-α method:

Man+1 + Cvn+1−αf+ Kdn+1−αf

= F(

tn+1−αf

)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

- If αf = αm = 0 ⇒ Newmark method:

Man+1 + Cvn+1 + Kdn+1 = F (tn+1)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

Page 79: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Linear multistep methods for second orderODEs

Stiffness

The second order ODE systems obtained after discretizing the wave-type PDE by the FEMare stiff. The high modes are result of the discretization and they are not representative. Theirresolution requires:- the use of good stability numerical methods.- controlled numerical dissipation in the range of the high frequencies.

Alfa-generalized method: Different weighting of the inertia forces and the rest of the addends:

Man+1−αm + Cvn+1−αf+ Kdn+1−αf

= F(

tn+1−αf

)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

where:

dn+1−αf= (1 − αf ) dn+1 + αf dn

vn+1−αf= (1 − αf ) vn+1 + αf vn

an+1−αm = (1 − αm) an+1 + αman

tn+1−αf= (1 − αf ) tn+1 + αf tn

- If αm = 0 ⇒ HHT-α method:

Man+1 + Cvn+1−αf+ Kdn+1−αf

= F(

tn+1−αf

)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

- If αf = αm = 0 ⇒ Newmark method:

Man+1 + Cvn+1 + Kdn+1 = F (tn+1)

dn+1 = dn + ∆tvn + ∆t22 [(1 − 2β) an + 2βan+1]

vn+1 = vn + ∆t [(1 − γ) an + γan+1]

Order of precision: Second order

- Generalized-alfa: γ = −αm + αf + 12

- HHT-α: γ = αf + 12

- Newmark: γ = 12

Page 80: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability and spectral radius

Stability study → Applying the method to the second order test equation: u′′ + ω2u = 0

Page 81: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability and spectral radius

Stability study → Applying the method to the second order test equation: u′′ + ω2u = 0Newmark method: stability and dissipation of high frequencies

- Unconditionally stable: 12 ≤ γ < 2β

- Conditionally stable: γ ≥ 12 and γ > 2β

- Dissipation of high frequencies ρ∞ < 1: β =

(γ+ 1

2

)2

4 . There is not high frequencydissipation in second order Newmark method.

Page 82: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability and spectral radius

Stability study → Applying the method to the second order test equation: u′′ + ω2u = 0Newmark method: stability and dissipation of high frequencies

- Unconditionally stable: 12 ≤ γ < 2β

- Conditionally stable: γ ≥ 12 and γ > 2β

- Dissipation of high frequencies ρ∞ < 1: β =

(γ+ 1

2

)2

4 . There is not high frequencydissipation in second order Newmark method.

HHT-α method: stability and dissipation of high frequencies

Unconditionally stable and dissipation of high frequencies: α ∈[0, 1

3

]and β =

(1+α)2

4

10−2

10−1

100

101

102

103

104

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Ω/(2π)

ρ

Collocation(γ=0.5,β=0.16,θ=1.514951)

Houbolt

(γ=0.5,β=0.18,θ=1.287301)

(γ=0.5,β=1/6,θ=1.4)Wilson

Collocation

Newmark

TrapezoidalHHT−

(β=0.3025,γ=0.6)

α (α= 0.05)

α (α= 0.3)HHT−

EDMC−1 χ1=χ

2=0.2998

Figure: Spectral radius of some methods as function of Ω/(2π).

Page 83: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

BDF-α method: linear multistep method withcontrolled numerical dissipation

Spectral radius of the BDFs:Second order ODEs ⇒ test equation u′′ + ω2u = 0This second order test equation is transformed in an equivalent first order ODE system:

(uu′

)′

=

(0 1

−ω2 0

) (uu′

)

⇒ y ′= ±iωy

Page 84: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

BDF-α method: linear multistep method withcontrolled numerical dissipation

Spectral radius of the BDFs:Second order ODEs ⇒ test equation u′′ + ω2u = 0This second order test equation is transformed in an equivalent first order ODE system:

(uu′

)′

=

(0 1

−ω2 0

) (uu′

)

⇒ y ′= ±iωy

Apply the method to the test equation y ′ = λy , where λ = ±iω:

Yn+k = A(

h)

· Yn+k−1

where h = hλ, Yn+k = (yn+1, yn+2, ..., yn+k )T , Yn+k−1 = (yn, yn+1, ..., yn+k−1)T and A the

amplification matrix.

Page 85: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

BDF-α method: linear multistep method withcontrolled numerical dissipation

Spectral radius of the BDFs:Second order ODEs ⇒ test equation u′′ + ω2u = 0This second order test equation is transformed in an equivalent first order ODE system:

(uu′

)′

=

(0 1

−ω2 0

) (uu′

)

⇒ y ′= ±iωy

Apply the method to the test equation y ′ = λy , where λ = ±iω:

Yn+k = A(

h)

· Yn+k−1

where h = hλ, Yn+k = (yn+1, yn+2, ..., yn+k )T , Yn+k−1 = (yn, yn+1, ..., yn+k−1)T and A the

amplification matrix.

The eigenvalues of A and the spectral radiusare calculated → BDF-s have high dissipationof the high frequency modes.

10−2

10−1

100

101

102

103

104

0

0.2

0.4

0.6

0.8

1

1.2

1.4

Ω/(2π)

ρ

(β=0.3025,γ=0.6)

BDF3

BDF5

BDF1

Houbolt

BDF4

HHT−

BDF2(γ=0.5,β=0.16,θ=1.514951)

Park

HHT−α (α= 0.3)

Newmark

α (α= 0.05)

Collocation

Page 86: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

New method based on the BDF2

BDF2: 32 yn+2 − 2yn+1 + 1

2 yn = hfn+2

- Second order and A-stable.- With a bigger range of spectral radius ρ∞ than the BDF2.

Page 87: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

New method based on the BDF2

BDF2: 32 yn+2 − 2yn+1 + 1

2 yn = hfn+2

- Second order and A-stable.- With a bigger range of spectral radius ρ∞ than the BDF2.

Expression of the method: Weighting with 3 free parameters:32 ((1 + β)yn+2 − βyn+1) − 2 ((1 + γ)yn+1 − γyn) + 1

2 yn = h ((1 + α)fn+2 − αfn+1)

Page 88: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

New method based on the BDF2

BDF2: 32 yn+2 − 2yn+1 + 1

2 yn = hfn+2

- Second order and A-stable.- With a bigger range of spectral radius ρ∞ than the BDF2.

Expression of the method: Weighting with 3 free parameters:32 ((1 + β)yn+2 − βyn+1) − 2 ((1 + γ)yn+1 − γyn) + 1

2 yn = h ((1 + α)fn+2 − αfn+1)

Reagrouping terms it results a linear multistep method:∑ 2

j=0 αj yn+j = h∑ 2

j=0 βj fn+j

where :

α2 = 32 (1 + β), α1 = − 3

2 β − 2(1 + γ), α0 = 2γ + 12

β2 = 1 + α, β1 = −α, β0 = 0

Page 89: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

Order of precision: LTE = C0y(tn) + C1hy ′(tn) + C2h2y ′′(tn) + ... + Cqhqy (q)(tn + ...

where:

C0 =∑ k

i=0 αi

C1 =∑ k

i=0 iαi −∑ k

i=0 βi

Cq = 1q!

(∑ k

i=0 iqαi

)

− 1(q−1)!

(∑ k

i=0 iq−1βi

)

, q ≥ 2

Page 90: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

Order of precision: LTE = C0y(tn) + C1hy ′(tn) + C2h2y ′′(tn) + ... + Cqhqy (q)(tn + ...

where:

C0 =∑ k

i=0 αi

C1 =∑ k

i=0 iαi −∑ k

i=0 βi

Cq = 1q!

(∑ k

i=0 iqαi

)

− 1(q−1)!

(∑ k

i=0 iq−1βi

)

, q ≥ 2

C0 =∑ 2

i=0 αi = 0C1 =

∑ 2i=0 iαi −

∑ 2i=0 βi = −2γ + 3

2 β

C2 = 12!

(∑ 2

i=0 i2αi

)

−(∑ 2

i=0 iβi

)

= −γ + 94 β − α

C3 = 13!

(∑ 2

i=0 i3αi

)

− 12!

(∑ 2

i=0 i2βi

)

= 74 β − 1

3 − γ3 − 3

2 α

Page 91: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

Order of precision: LTE = C0y(tn) + C1hy ′(tn) + C2h2y ′′(tn) + ... + Cqhqy (q)(tn + ...

where:

C0 =∑ k

i=0 αi

C1 =∑ k

i=0 iαi −∑ k

i=0 βi

Cq = 1q!

(∑ k

i=0 iqαi

)

− 1(q−1)!

(∑ k

i=0 iq−1βi

)

, q ≥ 2

C0 =∑ 2

i=0 αi = 0C1 =

∑ 2i=0 iαi −

∑ 2i=0 βi = −2γ + 3

2 β

C2 = 12!

(∑ 2

i=0 i2αi

)

−(∑ 2

i=0 iβi

)

= −γ + 94 β − α

C3 = 13!

(∑ 2

i=0 i3αi

)

− 12!

(∑ 2

i=0 i2βi

)

= 74 β − 1

3 − γ3 − 3

2 α

The method is of order 2: α = 32 β = 2γ

Error constant: C = −2−3α6

Page 92: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

Order of precision: LTE = C0y(tn) + C1hy ′(tn) + C2h2y ′′(tn) + ... + Cqhqy (q)(tn + ...

where:

C0 =∑ k

i=0 αi

C1 =∑ k

i=0 iαi −∑ k

i=0 βi

Cq = 1q!

(∑ k

i=0 iqαi

)

− 1(q−1)!

(∑ k

i=0 iq−1βi

)

, q ≥ 2

C0 =∑ 2

i=0 αi = 0C1 =

∑ 2i=0 iαi −

∑ 2i=0 βi = −2γ + 3

2 β

C2 = 12!

(∑ 2

i=0 i2αi

)

−(∑ 2

i=0 iβi

)

= −γ + 94 β − α

C3 = 13!

(∑ 2

i=0 i3αi

)

− 12!

(∑ 2

i=0 i2βi

)

= 74 β − 1

3 − γ3 − 3

2 α

The method is of order 2: α = 32 β = 2γ

Error constant: C = −2−3α6

Second order BDF-α:(

3

2+ α

)

yn+2 + (−2 − 2α) yn+1 +

(1

2+ α

)

yn = h(1 + α)fn+2 − hαfn+1

Page 93: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

Order of precision: LTE = C0y(tn) + C1hy ′(tn) + C2h2y ′′(tn) + ... + Cqhqy (q)(tn + ...

where:

C0 =∑ k

i=0 αi

C1 =∑ k

i=0 iαi −∑ k

i=0 βi

Cq = 1q!

(∑ k

i=0 iqαi

)

− 1(q−1)!

(∑ k

i=0 iq−1βi

)

, q ≥ 2

C0 =∑ 2

i=0 αi = 0C1 =

∑ 2i=0 iαi −

∑ 2i=0 βi = −2γ + 3

2 β

C2 = 12!

(∑ 2

i=0 i2αi

)

−(∑ 2

i=0 iβi

)

= −γ + 94 β − α

C3 = 13!

(∑ 2

i=0 i3αi

)

− 12!

(∑ 2

i=0 i2βi

)

= 74 β − 1

3 − γ3 − 3

2 α

The method is of order 2: α = 32 β = 2γ

Error constant: C = −2−3α6

Second order BDF-α:(

3

2+ α

)

yn+2 + (−2 − 2α) yn+1 +

(1

2+ α

)

yn = h(1 + α)fn+2 − hαfn+1

Cases:

α = −0.5 ⇒ Trapezoidal method

α = 0 ⇒ BDF2 method

Page 94: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability regions

After applying the method to the test equation:(

32 + α

)yn+2 + (−2 − 2α) yn+1 +

(12 + α

)yn = h(1 + α)yn+2 − hαyn+1

Page 95: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability regions

After applying the method to the test equation:(

32 + α

)yn+2 + (−2 − 2α) yn+1 +

(12 + α

)yn = h(1 + α)yn+2 − hαyn+1

Frontier: h =

(32 +α

)r2+(−2−2α)r+

(12 +α

)

(1+α)r2−αr

Page 96: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability regions

After applying the method to the test equation:(

32 + α

)yn+2 + (−2 − 2α) yn+1 +

(12 + α

)yn = h(1 + α)yn+2 − hαyn+1

Frontier: h =

(32 +α

)r2+(−2−2α)r+

(12 +α

)

(1+α)r2−αr

After substituting r = eiθ : h(θ) =(1+2α)(cosθ−1)2+isinθ

[(1+2α)(1−cosθ)+ 1

1+α

]

(1+α)

[(cosθ−

α1+α

)2+sin2θ

]

Page 97: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability regions

After applying the method to the test equation:(

32 + α

)yn+2 + (−2 − 2α) yn+1 +

(12 + α

)yn = h(1 + α)yn+2 − hαyn+1

Frontier: h =

(32 +α

)r2+(−2−2α)r+

(12 +α

)

(1+α)r2−αr

After substituting r = eiθ : h(θ) =(1+2α)(cosθ−1)2+isinθ

[(1+2α)(1−cosθ)+ 1

1+α

]

(1+α)

[(cosθ−

α1+α

)2+sin2θ

]

For α ≥ −0.5 the denominator of h(θ) is lower bounded. Fixing α ≥ −0.5, for a sufficientlybig real number which depends on α and independent of θ, R (α) ∈ R, the real part h(θ)verifies: 0 ≤ Re(h(θ)) ≤ R(α)

The frontier of the stability region h(θ) lies inthe right semiplane C

+.For h ∈ C

−, A-stability is achieved and usingcontinuity, C

− belongs to the stability region.A-stable when α ∈ [−0.5, +∞)

Page 98: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Stability regions

After applying the method to the test equation:(

32 + α

)yn+2 + (−2 − 2α) yn+1 +

(12 + α

)yn = h(1 + α)yn+2 − hαyn+1

Frontier: h =

(32 +α

)r2+(−2−2α)r+

(12 +α

)

(1+α)r2−αr

After substituting r = eiθ : h(θ) =(1+2α)(cosθ−1)2+isinθ

[(1+2α)(1−cosθ)+ 1

1+α

]

(1+α)

[(cosθ−

α1+α

)2+sin2θ

]

For α ≥ −0.5 the denominator of h(θ) is lower bounded. Fixing α ≥ −0.5, for a sufficientlybig real number which depends on α and independent of θ, R (α) ∈ R, the real part h(θ)verifies: 0 ≤ Re(h(θ)) ≤ R(α)

The frontier of the stability region h(θ) lies inthe right semiplane C

+.For h ∈ C

−, A-stability is achieved and usingcontinuity, C

− belongs to the stability region.A-stable when α ∈ [−0.5, +∞)

−2 0 2 4 6 8 10 12 14−8i

−6i

−4i

−2i

0

2i

4i

6i

8i

α=−0.4

α=−0.3

α=−0.2

α=0

α=4

α=100

α=1

++++++

α=−0.1

Page 99: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

The expression obtained after applying the method to the test equation in matrix form:Yn+2 = AYn+1

where:

Yn+2 = (yn+1, yn+2)T , Yn+1 = (yn, yn+1)

T , A = A−11 A2

A1 =

(1 00 3

2 + α − h(1 + α)

)

, A2 =

(0 1

− 12 − α 2 + 2α − hα

)

Page 100: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

The expression obtained after applying the method to the test equation in matrix form:Yn+2 = AYn+1

where:

Yn+2 = (yn+1, yn+2)T , Yn+1 = (yn, yn+1)

T , A = A−11 A2

A1 =

(1 00 3

2 + α − h(1 + α)

)

, A2 =

(0 1

− 12 − α 2 + 2α − hα

)

Eigenvalues of the amplification matrix:

λ1,2 =−2 − 2α + hα ±

h2α2 + 2h(α + 1) + 1

−3 − 2α + 2h(1 + α)(1)

To characterize the numerical dissipation, the espectral radius when h → ∞ is calculated. Forthe A-stable BDF-α, that is to say, α ∈ [−0.5, +∞), we obtain:

ρ∞ =

1, α = −0.5−2α2+2α

< 1, α ∈ [−0.5, 0)2α

2+2α< 1, α ∈ [0, +∞)

Which means that fixing α ∈ [−0.5, +∞) ρ∞ takes all the values of (0, 1].

Page 101: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Considerations about the new method

10−2

10−1

100

101

102

103

104

0

0.2

0.4

0.6

0.8

1

Ω/(2π)

ρ

Trapezoidal

BDF−α=9.50

Collocation

(γ=0.5,β=0.16,θ=1.514951)

BDF−α=0Houbolt

BDF−α=1.17

HHT−α (α= 0.05)

BDF−α=−0.35

HHT−α (α= 0.3)

BDF−α=−0.475065

10−0.8

10−0.6

10−0.4

10−0.2

100

0.88

0.9

0.92

0.94

0.96

0.98

1

1.02

Ω/(2π)

ρ

Trapezoidal

HHT−α (α= 0.05)

BDF−α=−0.475065

Page 102: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Page 103: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Formulation of the FEM approximation of the Laplace-Poisson non-linear PDE:

Page 104: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Formulation of the FEM approximation of the Laplace-Poisson non-linear PDE:

Strong formulation:φ = −k(u)∇u. Find u(x) ∈ C2(Ω) which verifies:

∇ · φ(u(x)) = f (x), ∀x ∈ Ωu(x) = g(x), ∀x ∈ Γg

φ(u(x)) · n = −h(x), ∀x ∈ Γh

Page 105: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Formulation of the FEM approximation of the Laplace-Poisson non-linear PDE:

Strong formulation:φ = −k(u)∇u. Find u(x) ∈ C2(Ω) which verifies:

∇ · φ(u(x)) = f (x), ∀x ∈ Ωu(x) = g(x), ∀x ∈ Γg

φ(u(x)) · n = −h(x), ∀x ∈ Γh

Weak formulation: ∫

Γhw(x) (φ(u(x)) · n) dΓh −

Ω∇w(x) · φ(u(x))dΩ =

Ωw(x)f (x)dΩ, ∀w(x) ∈ V

Page 106: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Formulation of the FEM approximation of the Laplace-Poisson non-linear PDE:

Strong formulation:φ = −k(u)∇u. Find u(x) ∈ C2(Ω) which verifies:

∇ · φ(u(x)) = f (x), ∀x ∈ Ωu(x) = g(x), ∀x ∈ Γg

φ(u(x)) · n = −h(x), ∀x ∈ Γh

Weak formulation: ∫

Γhw(x) (φ(u(x)) · n) dΓh −

Ω∇w(x) · φ(u(x))dΩ =

Ωw(x)f (x)dΩ, ∀w(x) ∈ V

FEM approximation: uh(x) =∑

j∈η

uj Nj (x)

Page 107: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Formulation of the FEM approximation of the Laplace-Poisson non-linear PDE:

Strong formulation:φ = −k(u)∇u. Find u(x) ∈ C2(Ω) which verifies:

∇ · φ(u(x)) = f (x), ∀x ∈ Ωu(x) = g(x), ∀x ∈ Γg

φ(u(x)) · n = −h(x), ∀x ∈ Γh

Weak formulation: ∫

Γhw(x) (φ(u(x)) · n) dΓh −

Ω∇w(x) · φ(u(x))dΩ =

Ωw(x)f (x)dΩ, ∀w(x) ∈ V

FEM approximation: uh(x) =∑

j∈η

uj Nj (x)

Matricial format: uh(x) =

n∑

j=1

uj Nj (x) ⇒ uh(x) = N · U = (N1(x), ..., Nn(x))

︸ ︷︷ ︸

N

·

u1

...un

︸ ︷︷ ︸

U

= N · U

∇uh(x) =∑ n

j=1 uj∇Nj (x) ⇒ ∇uh(x) =

(∂N1∂x · · · ∂Nn

∂x∂N1∂y · · · ∂Nn

∂y

)

︸ ︷︷ ︸

B

·

u1

...un

︸ ︷︷ ︸

U

= B · U

φ = −k(u)BU

Page 108: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Formulation of the FEM approximation of the Laplace-Poisson non-linear PDE:

Strong formulation:φ = −k(u)∇u. Find u(x) ∈ C2(Ω) which verifies:

∇ · φ(u(x)) = f (x), ∀x ∈ Ωu(x) = g(x), ∀x ∈ Γg

φ(u(x)) · n = −h(x), ∀x ∈ Γh

Weak formulation: ∫

Γhw(x) (φ(u(x)) · n) dΓh −

Ω∇w(x) · φ(u(x))dΩ =

Ωw(x)f (x)dΩ, ∀w(x) ∈ V

FEM approximation: uh(x) =∑

j∈η

uj Nj (x)

Matricial format: uh(x) =

n∑

j=1

uj Nj (x) ⇒ uh(x) = N · U = (N1(x), ..., Nn(x))

︸ ︷︷ ︸

N

·

u1

...un

︸ ︷︷ ︸

U

= N · U

∇uh(x) =∑ n

j=1 uj∇Nj (x) ⇒ ∇uh(x) =

(∂N1∂x · · · ∂Nn

∂x∂N1∂y · · · ∂Nn

∂y

)

︸ ︷︷ ︸

B

·

u1

...un

︸ ︷︷ ︸

U

= B · U

φ = −k(u)BU

Non-linear equations system: Fint (U) = Fext

where:

Fint (U) = −

Ωh BTa φh(U)dΩh, donde φh = −k(u)∇uh(x) = −k(u)BU

Fext =∫

Ωh NTa f (x)dΩh +

Γhh

NTa h(x)dΓh

h

Page 109: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

Formulation of the FEM approximation of the Laplace-Poisson non-linear PDE:

Strong formulation:φ = −k(u)∇u. Find u(x) ∈ C2(Ω) which verifies:

∇ · φ(u(x)) = f (x), ∀x ∈ Ωu(x) = g(x), ∀x ∈ Γg

φ(u(x)) · n = −h(x), ∀x ∈ Γh

Weak formulation: ∫

Γhw(x) (φ(u(x)) · n) dΓh −

Ω∇w(x) · φ(u(x))dΩ =

Ωw(x)f (x)dΩ, ∀w(x) ∈ V

FEM approximation: uh(x) =∑

j∈η

uj Nj (x)

Matricial format: uh(x) =

n∑

j=1

uj Nj (x) ⇒ uh(x) = N · U = (N1(x), ..., Nn(x))

︸ ︷︷ ︸

N

·

u1

...un

︸ ︷︷ ︸

U

= N · U

∇uh(x) =∑ n

j=1 uj∇Nj (x) ⇒ ∇uh(x) =

(∂N1∂x · · · ∂Nn

∂x∂N1∂y · · · ∂Nn

∂y

)

︸ ︷︷ ︸

B

·

u1

...un

︸ ︷︷ ︸

U

= B · U

φ = −k(u)BU

Non-linear equations system: Fint (U) = Fext

where:

Fint (U) = −

Ωh BTa φh(U)dΩh, donde φh = −k(u)∇uh(x) = −k(u)BU

Fext =∫

Ωh NTa f (x)dΩh +

Γhh

NTa h(x)dΓh

h

Newton Raphson resolution:

R(Uk ) = Fext − Fint (Uk )

Uk+1 = Uk − J−1R(Uk )k = 0,1,2,... until ‖R (Uk+1)‖ < tol ≈ 0 J(Uk ) = −K t (Uk ) =

∂R(Uk )

∂Uk= −

∂Fkint

(Uk )

∂Uk

Page 110: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

The dynamic problem:

Page 111: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

The dynamic problem:

FEM approximation: uh(x) = N(x)U(t)

Page 112: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

The dynamic problem:

FEM approximation: uh(x) = N(x)U(t)

Function f (x) of the anterior section becomes:

f(x, t) − ρutt (x, t)︸ ︷︷ ︸

inertia forces

− cut (x, t)︸ ︷︷ ︸

damping forces

≈ f(x, t) − ρN(x)U′′(t) − cN(x)U′

(t)

Page 113: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

The dynamic problem:

FEM approximation: uh(x) = N(x)U(t)

Function f (x) of the anterior section becomes:

f(x, t) − ρutt (x, t)︸ ︷︷ ︸

inertia forces

− cut (x, t)︸ ︷︷ ︸

damping forces

≈ f(x, t) − ρN(x)U′′(t) − cN(x)U′

(t)

Second order ODE system: MU′′ + CU′ + Fint (U) = Fext

where:

M =∫

ΩNT ρN dΩ , C =

ΩNT cN dΩ

Fint =∫

ΩBT σh(U) dΩ , Fext =

ΓtNT t dSt +

ΩNT f dΩ

Page 114: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

The dynamic problem:

FEM approximation: uh(x) = N(x)U(t)

Function f (x) of the anterior section becomes:

f(x, t) − ρutt (x, t)︸ ︷︷ ︸

inertia forces

− cut (x, t)︸ ︷︷ ︸

damping forces

≈ f(x, t) − ρN(x)U′′(t) − cN(x)U′

(t)

Second order ODE system: MU′′ + CU′ + Fint (U) = Fext

where:

M =∫

ΩNT ρN dΩ , C =

ΩNT cN dΩ

Fint =∫

ΩBT σh(U) dΩ , Fext =

ΓtNT t dSt +

ΩNT f dΩ

If the problem is linear

Fint = KU, where K =∫

ΩBT kBU dΩ

Page 115: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology

The dynamic problem:

FEM approximation: uh(x) = N(x)U(t)

Function f (x) of the anterior section becomes:

f(x, t) − ρutt (x, t)︸ ︷︷ ︸

inertia forces

− cut (x, t)︸ ︷︷ ︸

damping forces

≈ f(x, t) − ρN(x)U′′(t) − cN(x)U′

(t)

Second order ODE system: MU′′ + CU′ + Fint (U) = Fext

where:

M =∫

ΩNT ρN dΩ , C =

ΩNT cN dΩ

Fint =∫

ΩBT σh(U) dΩ , Fext =

ΓtNT t dSt +

ΩNT f dΩ

If the problem is linear

Fint = KU, where K =∫

ΩBT kBU dΩ

First order equivalent ODE system:

d(t) = U(t)v(t) = U′(t) ⇒

d ′(t) = U′(t) = v(t)

Mv ′(t) = MU′′(t) = −Fint (d) − Cv + Fext ⇒

(I 00 M

)

︸ ︷︷ ︸

M∗

(d ′(t)v ′(t)

)

︸ ︷︷ ︸

y′

=

(v(t)

R(d) − Cv

)

︸ ︷︷ ︸

f (t,y)

⇒ M∗y ′= f (t, y)

where:

f1(t, y) ≡ y2 = v , f2(t, y) ≡ R(y1) − Cy2 = R(d) − Cv

j =∂f (t,y)

∂y =

∂f1∂y1

∂f1∂y2

∂f2∂y1

∂f2∂y2

=

( ∂v∂d

∂v∂v

∂(R(d)−Cv)∂d

∂(R(d)−Cv)∂v

)

=

(0 I

∂R(d)∂d −C

)

=

(0 I

J(U) −C

)

Page 116: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object Oriented Programming methodology(OOP)

Resulting OOP architecture: 3 blocks

Mesh object (“objmalla”): Related to the calculus of the domain integrals.Problem object (“objprob”): Related to the management of the specific aspects of the problem.Method object (“objmetodo”): Related to the resolution of the problem (linear, non-linear,dynamic).

objmef

objelef

objelep

objfty

objpint

objdef

objprobobjmetodo

objnr

objJ

objmalla

objresYcc objmat objU

objsislin

Page 117: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode)

The step by step advancing process of the different methods has a common estructure:initialization, loop in steps, actualization.

Page 118: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode)

Trapezoidal rule → trapezoidal object:

Page 119: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode)

Trapezoidal rule → trapezoidal object:

Non-linear problem:

M·y ′ = f (t, y) ⇒ M · yn+1 = M · yn + h2 (fn + fn+1) Each iteration is solved by Mewton-Raphson:

ym+1n+1 = ym

n+1 −(J(ym

n+1))−1 R

(ym

n+1

), m = 0, 1, 2, ... paso(step)

where:

R(ym

n+1

)= 2

h M(ym

n+1 − yn)−(fn + f m

n+1

)calcR

J(ym

n+1

)=

∂R(ymn+1)

∂yn+1= 2

h M − jmn+1 calcJ

Page 120: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode)

Trapezoidal rule → trapezoidal object:

Non-linear problem:

M·y ′ = f (t, y) ⇒ M · yn+1 = M · yn + h2 (fn + fn+1) Each iteration is solved by Mewton-Raphson:

ym+1n+1 = ym

n+1 −(J(ym

n+1))−1 R

(ym

n+1

), m = 0, 1, 2, ... paso(step)

where:

R(ym

n+1

)= 2

h M(ym

n+1 − yn)−(fn + f m

n+1

)calcR

J(ym

n+1

)=

∂R(ymn+1)

∂yn+1= 2

h M − jmn+1 calcJ

Linear problem:

M·y ′ = A · y + g(t) ⇒ where:

First order Laplace: A → −K , y → d(t), g(t) → Fext (t)

Wave: M →

(I 00 M

)

, A →

(0 I

−K −C

)

, y →

(d(t)v(t)

)

, g(t) →

(0

Fext (t)

)

Generic case: M, A, g(t)

Advancing formula:

Myn+1 = Myn + h2 (Ayn + g(tn) + Ayn+1 + g(tn+1)) ⇒

(M − h

2 A)

yn+1 =(M + h

2 A)

yn + h2 (g(tn) + g(tn+1))

yn+1 = C1

[

C2yn +h

2(g(tn) + g(tn+1))

]

where: C1 =(M − h

2 A)−1, C2 =

(M + h

2 A). C1, C2, M and A are calculated once in the initialization.

Page 121: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode): BDF method

Multistep methods → BDFs:

Page 122: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode): BDF method

Multistep methods → BDFs:

Non-linear problem:

∑ kj=0 αj Myn+j = hfn+k ⇒

R(ym

n+k

)= 1

h

(

αk Mymn+k +

∑ k−1j=0 αj Myn+j

)

− f mn+k

J(ym

n+k

)=

∂R(ymn+k

)

∂yn+k= 1

h αk M − jmn+k

Programming:

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− f m

n+kJ(ym

n+k

)= 1

h αk M − jmn+k

where:

C1 = (αk M − hA)−1 , only for the linear caseC2 =

(αk−1M αk−2M αk−3M . . . α1M α0M

)

Yn+k−1 =(yn+k−1 yn+k−2 yn+k−3 . . . yn+1 yn

)T

Page 123: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode): BDF method

Multistep methods → BDFs:

Non-linear problem:

∑ kj=0 αj Myn+j = hfn+k ⇒

R(ym

n+k

)= 1

h

(

αk Mymn+k +

∑ k−1j=0 αj Myn+j

)

− f mn+k

J(ym

n+k

)=

∂R(ymn+k

)

∂yn+k= 1

h αk M − jmn+k

Programming:

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− f m

n+kJ(ym

n+k

)= 1

h αk M − jmn+k

where:

C1 = (αk M − hA)−1 , only for the linear caseC2 =

(αk−1M αk−2M αk−3M . . . α1M α0M

)

Yn+k−1 =(yn+k−1 yn+k−2 yn+k−3 . . . yn+1 yn

)T

Linear problem:∑ k

j=0 αj Myn+j = h (Ayn+k + g(tn+k )) ⇒ (αk M − hA) yn+k = −∑ k−1

j=0 αj Myn+j + hg(tn+k )

(αk M − hA) yn+k = −C2Yn+k−1 + hg(tn+k )

Page 124: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode): superfuturepoints methods

Superfuture points methods → EBDFs:

Non-linear problem:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− βk f m

n+k − βk+1 fn+k+1

J(ym

n+k

)= 1

h αk M − βk jmn+k

where:

C2 =(αk−1M αk−2M αk−3M . . . α1M α0M

)

Yn+k−1 =(yn+k−1 yn+k−2 yn+k−3 . . . yn+1 yn

)T

Page 125: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode): superfuturepoints methods

Superfuture points methods → EBDFs:

Non-linear problem:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− βk f m

n+k − βk+1 fn+k+1

J(ym

n+k

)= 1

h αk M − βk jmn+k

where:

C2 =(αk−1M αk−2M αk−3M . . . α1M α0M

)

Yn+k−1 =(yn+k−1 yn+k−2 yn+k−3 . . . yn+1 yn

)T

Superfuture points methods → MEBDFs:

Non-linear problem:

∑ kj=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1 + h

(

βk − βk

)

fn+k

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− βk f m

n+k − βk+1 fn+k+1 −(

βk − βk

)

fn+k

J(ym

n+k

)=

∂R(ymn+k

)

∂yn+k= 1

h αk M − βk jmn+k

Page 126: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Object method ODE (objode): superfuturepoints methods

Superfuture points methods → EBDFs:

Non-linear problem:∑ k

j=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− βk f m

n+k − βk+1 fn+k+1

J(ym

n+k

)= 1

h αk M − βk jmn+k

where:

C2 =(αk−1M αk−2M αk−3M . . . α1M α0M

)

Yn+k−1 =(yn+k−1 yn+k−2 yn+k−3 . . . yn+1 yn

)T

Superfuture points methods → MEBDFs:

Non-linear problem:

∑ kj=0 αj yn+j = hβk fn+k + hβk+1 fn+k+1 + h

(

βk − βk

)

fn+k

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− βk f m

n+k − βk+1 fn+k+1 −(

βk − βk

)

fn+k

J(ym

n+k

)=

∂R(ymn+k

)

∂yn+k= 1

h αk M − βk jmn+k

Programming both:

R(ym

n+k

)= 1

h

(αk Mym

n+k + C2Yn+k−1)− γk f m

n+k − βk+1 fn+k+1 − γk fn+k

J(ym

n+k

)=

∂R(ymn+k

)

∂yn+k= 1

h αk M − γk jmn+k

where:For EBDFs: γk = βk and γk = 0.

For MEBDFs: γk = βk and γk =(

βk − βk

)

.

Page 127: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 1: Linear wave equation with rectangular pulse IC (1D).400 elements and 1400 steps.

Page 128: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 1: Linear wave equation with rectangular pulse IC (1D).400 elements and 1400 steps.

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método HHT−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=0.3 γ=0.8 β=0.4225

t= 0t= 2

Page 129: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 1: Linear wave equation with rectangular pulse IC (1D).400 elements and 1400 steps.

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método HHT−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=0.3 γ=0.8 β=0.4225

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método BDF−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=−0.35

Page 130: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 1: Linear wave equation with rectangular pulse IC (1D).400 elements and 1400 steps.

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método HHT−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=0.3 γ=0.8 β=0.4225

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método BDF−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=−0.35

Computational times → HHT-α: 1.88 seconds, BDF-α: 6.17 secondsBoth very quick but HHT-α 3 times quicker.

Page 131: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 1: Linear wave equation with rectangular pulse IC (1D).400 elements and 1400 steps.

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método HHT−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=0.3 γ=0.8 β=0.4225

t= 0t= 2

0 1 2 3 4 5 6 7 8−0.5

0

0.5

1

1.5Método BDF−alfa, tiempo=16, nele=400, pasos=1400, masa=cons

α=−0.35

Computational times → HHT-α: 1.88 seconds, BDF-α: 6.17 secondsBoth very quick but HHT-α 3 times quicker.The most expensive operations:- HHT-α: an+1 = MCK−1 · Fn+1

where:

MCK = (M + (1 − α)Chγ + (1 − α)Kh2β)

Fn+1 = Fext (tn+1−α) − C ((1 − α)vn+1 + αvn) − K(

(1 − α)dn+1 + αdn

)

- BDF-α: yn+2 = C−11 · [−C2Yn+1 − hαAyn+1 + h(1 + α)g(tn+2) − hαg(tn+1)]

︸ ︷︷ ︸

TI

⇒ yn+2 = C1 · TI

where:

C1 = (β2M − h(1 + α)A) ,

C2 = (β1M β0M), Yn+1 = (yn+1 yn)T

The dimension of the matrix C1 = (β2M − h(1 + α)A) is the double of MCK .

Page 132: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 2: Non-linear PDE of a guitar string.20 elements and 1000 steps.

Page 133: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 2: Non-linear PDE of a guitar string.20 elements and 1000 steps.

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.015198, nele=20, pasos=1000, masa=cons

α=0.3 , γ=0.8, β=0.4225

Figure: HHT-α = 0.3.

Page 134: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 2: Non-linear PDE of a guitar string.20 elements and 1000 steps.

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.015198, nele=20, pasos=1000, masa=cons

α=0.3 , γ=0.8, β=0.4225

Figure: HHT-α = 0.3.

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

α=−0.35

Método BDF−α, tiempo=0.015198, nele=20, pasos=1000, masa=cons

Figure: BDF-α = −0.35.

Page 135: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 2: Non-linear PDE of a guitar string.20 elements and 1000 steps.

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.015198, nele=20, pasos=1000, masa=cons

α=0.3 , γ=0.8, β=0.4225

Figure: HHT-α = 0.3.

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

α=−0.35

Método BDF−α, tiempo=0.015198, nele=20, pasos=1000, masa=cons

Figure: BDF-α = −0.35.

Computational times → HHT-α = 0.3: 244.76 seconds, 9479 iterations. BDF-α = −0.35:259.89 seconds, 9516 iterations.

Page 136: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 2: Non-linear PDE of a guitar string.20 elements and 1000 steps.

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08Método HHT−alfa, tiempo=0.015198, nele=20, pasos=1000, masa=cons

α=0.3 , γ=0.8, β=0.4225

Figure: HHT-α = 0.3.

0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

α=−0.35

Método BDF−α, tiempo=0.015198, nele=20, pasos=1000, masa=cons

Figure: BDF-α = −0.35.

Computational times → HHT-α = 0.3: 244.76 seconds, 9479 iterations. BDF-α = −0.35:259.89 seconds, 9516 iterations.Again, the dimension of the matrices of the BDF-α method is the double.Time for solving the equation system of the total iterations→ HHT-α: 1.54 seconds andBDF-α 0.1 seconds.This difference is not important in the final balance as it is the calculation of R and J of themethods which more time consumes.

Page 137: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D.

A plane deformation square block of 2 units.Data: E = 1000, ν = 0.25, ρ = 1, IC that corresponds to a vertical lengthening of 1 unit, BCDirichlet nule for the verticals of the horizontal simmetry axis and for the horizontaldisplacement of the vertical simmetry axis.Only the forth part of the block has been discretized due to the symmetry.

Page 138: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D.

A plane deformation square block of 2 units.Data: E = 1000, ν = 0.25, ρ = 1, IC that corresponds to a vertical lengthening of 1 unit, BCDirichlet nule for the verticals of the horizontal simmetry axis and for the horizontaldisplacement of the vertical simmetry axis.Only the forth part of the block has been discretized due to the symmetry.

0 0.2 0.4 0.6 0.8 1 1.2 1.40

0.2

0.4

0.6

0.8

1

tensiones: yy

−300

−200

−100

0

100

200

300

400

500

Figure: Final stresses anddeformation (t = 0.5).

Page 139: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D.

A plane deformation square block of 2 units.Data: E = 1000, ν = 0.25, ρ = 1, IC that corresponds to a vertical lengthening of 1 unit, BCDirichlet nule for the verticals of the horizontal simmetry axis and for the horizontaldisplacement of the vertical simmetry axis.Only the forth part of the block has been discretized due to the symmetry.

0 0.2 0.4 0.6 0.8 1 1.2 1.40

0.2

0.4

0.6

0.8

1

tensiones: yy

−300

−200

−100

0

100

200

300

400

500

Figure: Final stresses anddeformation (t = 0.5).

0 0.1 0.2 0.3 0.4 0.5−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

0.6

Figure: Nodal displacements(Trap. method 200 steps).

Page 140: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D → Harmonic response

Page 141: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D → Harmonic response

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

1.2

1.4Modo nº 1 de 60 ; Frec: w=44.7771 ; Partic: F=0.33037

0 0.1 0.2 0.3 0.4 0.5−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1

Figure: First mode deformation as IC (left) and nodal displacements (right) (Trap. 200steps).

Page 142: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D → Harmonic response

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

1.2

1.4Modo nº 1 de 60 ; Frec: w=44.7771 ; Partic: F=0.33037

0 0.1 0.2 0.3 0.4 0.5−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1

Figure: First mode deformation as IC (left) and nodal displacements (right) (Trap. 200steps).

0 0.2 0.4 0.6 0.8 1 1.2 1.40

0.2

0.4

0.6

0.8

1

1.2

1.4Modo nº 60 de 60, Frec: w=680.1678 ; Partic: F=3.3282e−005

0 0.01 0.02 0.03 0.04 0.05−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1

Figure: Last mode deformation as IC (left) and nodal displacements (right) (Trap. 1000

Page 143: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D.

Page 144: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D.

0 0.5 1 1.5 2−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1BDF−α=−0.35, 317 pasos

0 0.5 1 1.5 2−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1BDF−α=−0.35, 1000 pasos

0 0.5 1 1.5 2−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1BDF−α=−0.35, 2000 pasos

Figure: Algorithmic damping for high frequency, BDF-α = −0.35 (317, 1000 and 2000steps). Last mode deformation.

Page 145: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Results

Example 3: Dynamic linear elasticity 2D.

0 0.5 1 1.5 2−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1BDF−α=−0.35, 317 pasos

0 0.5 1 1.5 2−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1BDF−α=−0.35, 1000 pasos

0 0.5 1 1.5 2−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1BDF−α=−0.35, 2000 pasos

Figure: Algorithmic damping for high frequency, BDF-α = −0.35 (317, 1000 and 2000steps). Last mode deformation.

0 0.5 1 1.5 2−0.1

−0.08

−0.06

−0.04

−0.02

0

0.02

0.04

0.06

0.08

0.1BDF−α=−0.35, 317 pasos, evolución del primer modo

Figure: First mode evolution, BDF-α = −0.35 (317steps).

The solution does not lose too muchprecision.

Page 146: Numerical stiff ODEs Numerical methods for stiff Ordinary ...Alberdi Celaya Introduction First order ODEs Changing EBDFs and MEBDFs LMS for second order ODEs BDF-α method OOP methodology

Numericalmethods forstiff ODEs

ElisabeteAlberdi Celaya

Introduction

First orderODEs

ChangingEBDFs andMEBDFs

LMS forsecond orderODEs

BDF-αmethod

OOPmethodology

Results

Thank You foryour attention