Introduction

Motivation and content

Neoclassical growth model
Function approximation
Numerical integration
Summary

Monash Macroeconomics
Summer School
Intro and basic tools

Petr Sedláček

University of Bonn

February 2016

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model
Function approximation
Numerical integration
Summary

Goal

1 motivation and overview of entire course

what tools and how are they used in practice?
2 describe simple DSGE model
3 functional approximation
basics, polynomials, splines
4 numerical integration
monte-carlo, quadrature methods

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

Motivation and overview

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

Resources

much of this course is based on Wouter den Haan’s work

excellent lecture notes, slides on his webpage

other great resources include e.g.

DeJong, Dave (2011). “Structural Macroeconometrics”
Heer, Maussner (2009). “Dynamic General Equilibrium Modelling”
Judd (1998). “Numerical Methods in Economics”
lecture notes and slides by

Kopecky, Mukoyama, Nakajima etc.

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

What tools are needed in modern macro?

1 theoretical tools
use models to look at data

2 computational tools
characterize model outcomes
solve models for particular parameter values
assign the “right” parameter values

3 empirical tools
analyze statistical properties of data and model

4 common sense
Sedláček Monash Macro
 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

Why DSGE models?

prior to DSGE’s
long tradition of large macroeconometric models
these reduced-form systems have certain drawbacks
no “GE”
no forward-looking behavior

changes after Kydland and Prescott (1982)

other critical contributions by Hansen, Lucas, Sargent and Sims

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

Why DSGE models?

attempt to put discipline on reduced-form methods

discipline comes from “cross-equation” restrictions
stochastics of exogenous variables
together with forward-looking behavior of agents
result in implication for evolution of endogenous variables

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

What will we cover?

1) Computational tools for “Rep-Agent models”

what do we need to solve for?

policy rules (functions)

why is this a tough problem?

forward looking behavior
dynamics today depend on expectations of future dynamics
focus on recursive problems
even then
analytical solutions are rare
“S” in DSGE necessitates computation of expectations

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

What will we cover?

1) Computational tools for “Rep-Agent models”

characterize unknown functions and their integrals

function approximation
numerical integration

we’ll then be ready to solve models in several ways

projection
perturbation
other methods (which we won’t cover):
value (policy) function iteration

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

What will we cover?

2) Tools for parametrizing models

discussion of alternative methods

calibration
estimation
matching moments

we’ll cover estimation in detail

Maximum Likelihood
Bayesian methods

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

What will we cover?

3) Computational tools for “Hetero-Agent models”

solving for equilibrium prices and market clearing is tougher

Krusell-Smith algorithm
simulation methods

often compute “only” approximate equilibria:

need to check accuracy of solutions

if time permits, we’ll check out a few other topics:

practical issues regarding the Krusell-Smith algorithm
additional solution algorithm (Sedláček and Sterk; 2016)

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

What will we cover?

1 theoretical tools
use models to look at data

2 computational tools
characterize model outcomes
solve models for particular parameter values
assign the “right” parameter values

3 empirical tools
analyze statistical properties of data and model

4 common sense
Sedláček Monash Macro
 Introduction
Motivation and content
Neoclassical growth model Why DSGE’s?
Function approximation What will we cover?
Numerical integration
Summary

Main thing to remember

“The goal of computing is insight, not numbers”

Richard Hamming

Mathematician and
computer scientist

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

A simple DSGE model

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Neoclassical growth model

representative household maximizing expected lifetime utility

household owns production technology
capital is the only factor of production
resources spent on consumption and investment into capital
each period existing capital depreciates at certain rate
production subject to exogenous fluctuations in productivity

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Production

yt = Zt ktα
Zt = 1 − ρ + ρZt−1 + t
Et = 0
E2t = σz2

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Household decision

∞
X
max ∞ E0 β t u(ct )
{ct ,kt+1 }t=0
t=0

s.t. ct + kt+1 = yt + (1 − δ)kt

k0 given
Z0 given

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Solution

What is the solution?

a sequence {ct , kt+1 }∞

t=0

maximizing the expected discounted sum of per-period utilities

Sounds like a tough problem!

different k0 → optimal sequences different!

different realizations of Zt → optimal sequences different!

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Solution
Trick is to

replace sequential problem with 2-period decisions

recursiveness

must make sure that each 2-period decision is globally optimal

Principle of Optimality (Bellman)

decisions depend on state variables

agents facing the same state variables
make the same decisions
independent of in which period they are

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Policy rules

what are the state variables?

beginning-of-period capital and productivity

what are the policy rules?

ct = c(kt , Zt )

kt+1 = k(kt , Zt )
how are they determined?
α−1


uc (ct ) = βEt uc (ct+1 ) αZt+1 kt+1 +1−δ

ct + kt+1 = yt + (1 − δ)kt
Sedláček Monash Macro
 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Where do we use our computational tools?

Function approximation?

analytical solutions rarely exist

→ need to approximate the policy functions

ct ≈ ce(kt , Zt ; ψc )

kt+1 ≈ k(k
e t , Zt ; ψk )

what are we solving for?

the coefficients of the approximations: ψc and ψk

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Where do we use our computational tools?

Numerical integration?

analytical solutions rarely exist

→ need to calculate expectations of (unknown) functions
α−1
 
uc (ct ) = βEt uc (ct+1 ) αZt+1 kt+1 +1−δ

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Where do we use our computational tools?

Numerical integration?

analytical solutions rarely exist

we need to calculate expectations of (unknown) functions

uc (e
c (kt , Zt ; ψc )) =
βEt uc (e e t , Zt ; ψk ), (1 − ρ + ρZt + t+1 ); ψc ))∗
c (k(k
h i
α(1 − ρ + ρZt + t+1 )k(k e t , Zt ; ψk )α−1 + 1 − δ

Sedláček Monash Macro

 Introduction
Motivation and content
Production and household
Neoclassical growth model
Solution
Function approximation
Use of computational tools
Numerical integration
Summary

Special case of analytical solution

assume log utility (γ = 1)

and full depreciation (δ = 1)
this is the Brock-Mirman model

Turns out that this version has an analytical solution:

kt+1 = αβZt ktα

ct = (1 − αβ)Zt ktα

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Function approximation

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Why function approximation?

we are after policy rules

these are functions of state variables
moreover, closed form solutions rarely exist
→ work with approximations of true functions

let’s think about this problem more generally first

later we’ll talk about how to implement it with DSGE models

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Main idea of function approximation

Consider we want to approximate a function

y = f (x)

1. choose a family of functions to use as interpolants

popular choice is the family of polynomials
but others also exist
trigonometric functions (fourier approximation)
rational functions (pade approximation)

2. find coefficients of interpolant

such that interpolant and true function agree at certain points
Sedláček Monash Macro
 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Main idea of function approximation

We’ve already made enormous progress at this point:
we have reduced the problem to a finite dimension!
y ≈ f (x) = a0 T0 (x) + a1 T1 (x) + ... + an Tn (x)

where aj are coefficients of the polynomial for j = 0, .., n

and Tj (x) are basis functions

aj ’s solve the above at each chosen node, i = 1, ..., n + 1

n
X
yi = f (xi ) = a0 + aj Tj (xi )
j=1

y = T (x)a
Sedláček Monash Macro
 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Main idea of function approximation

the above (interpolation) method is a special case

regression is a method of interpolation!
when the number of grid points i
is generally larger than the number of basis functions
what if you run a regression of n points on n regressors?

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Why polynomial approximation?

Weierstrass’ theorem (a sloppy version)
if f (x) is a continuous function in [a, b]
then there exists a polynomial p for which
supa≤x≤b |f (x) − p(x)| ≤ 
with  > 0

In other words
there exists a polynomial
that approximates any continuous function
arbitrarily well
Sedláček Monash Macro
 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Problem with Weierstrass...

it is useless from a practical point of view

because it gives no guidance on how to find p

There are (at least) 2 important choices to be made

what type of polynomial to use

and where to evaluate it (grid points)

An example is a Taylor approximation at equidistant nodes

turns out to be a bad idea

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Why not monomials?

the choice of monomial basis functions implies

1 x1 · · · x1n−1
    
y1 a0
 ..   .. .. . . ..   .. 
 . = . . . .  . 
yn 1 xn · · · xn n−1 an

the first matrix on the RHS is a Vandermonde matrix

even though it is non-singular, it is often ill-behaved
intuition from regression?

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Orthogonal polynomials

monomials often suffer from high correlation

orthogonal polynomials are constructed
to have orthogonal basis functions (w.r.t. some measure)
Z b
Ti (x)Tj (x)w (x) = 0 ∀i, j i 6= j
a

w (x) is some weighting function

Popular orthogonal polynomials are Chebyshev polynomials

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Chebyshev polynomials

defined on interval [−1, 1]

weighting function
1
w (x) =
(1 − x 2 )1/2

basis functions
T0 (x) = 1
T1 (x) = x
Tj+1 (x) = 2xTj (x) − Tj−1 (x) j > 1

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Why not an equidistant grid?

suppose we have an equidistant grid

it turns out that the higher the order of polynomial
the larger the “swings” between grid points
these oscillations become more dramatic at the end points!

Weierstraas’ theorem

there exists a uniformly converging polynomial approximation

to find it, however, we have to be smart about the nodes

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Chebyshev nodes

Chebyshev nodes ensure uniform convergence

the roots (zj ) at which the basis functions are equal to 0

e.g. T2 (x) = 2x 2 − 1 → nodes of z1 = −1/2 and z2 = 1/2
i.e. get n Chebyshev nodes by solving the nth basis function
this is the reason for the popularity of Chebyshev polynomials
Chebyshev nodes can be computed according to
 
(2j − 1)π
zj = −cos
2n

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Interval conversion

to approximate on an interval [a, b]

we must rescale the Chebyshev nodes
find α and β for which

x = αz + β

β−α=a
β+α=b
then if z ∈ [−1, 1] and x ∈ [a, b] then
b−a a+b
x= z+
2 2
Sedláček Monash Macro
 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Splines - main idea

polynomials approximate over entire domain

spectral method

splines split support into sections

finite element method

splines can be expressed as linear combinations of basis fces

but they are not polynomials
basis functions are zero for most of the domain

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Piece-wise linear splines

the easiest type is piece-wise linear

   
x − xi x − xi
f (x) ≈ 1 − fi + fi+1 x ∈ [xi , xi+1 ]
xi+1 − xi xi+1 − xi

in general not differentiable at nodes

could be problematic → use higher-order polynomials

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Cubic splines

fit a 3-rd order polynomial in each segment

f (x) ≈ ai + bi x + ci x 2 + di x 3 x ∈ [xi , xi+1 ]

( S (x) x ≤ x ≤ x
1 0 1
S(x) = Si (x) xi−1 ≤ x ≤ xi
Sn (x) xn−1 ≤ x ≤ xn
i.e. we have n separate cubic splines for n + 1 nodes

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Cubic splines

n splines give 4n coefficients to determine

what conditions pin down the 4n coefficients?
2 + 2(n − 1) function values at the nodes
2(n − 1) smoothness conditions (for 0 < i < n)

Si0 (xi ) = Si+1

0
(xi )

Si00 (xi ) = Si+1

00
(xi )
2 boundary conditions
“natural (simple)” S100 (x0 ) = 0 and Sn00 (xn ) = 0
or “clamped” S10 (x0 ) = 0 and Sn0 (xn ) = 0

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Why polynomials
Neoclassical growth model Choice of basis functions
Function approximation Choice of nodes
Numerical integration Interval conversion
Summary Splines

Before we move on...

sovling DSGE models means getting policy functions

so how does function approximation work in DSGE models?

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Numerical integration

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Why numerical integration?

in economics, there are plenty of integrals

expectations

evaluating integrals can be a tough problem

the functional form may be nasty
we may not even have the functional form
we may be able to evaluate it, but not draw from it
as in e.g. Bayesian estimation

therefore, we need a way around this...

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Intuitive integration method

Rb
Consider you want to compute the integral a h(x)dF (x)
where x is a random variable with CDF F (x)
Monte-Carlo integration uses the following approximation
Z b PT
h(xt )
h(x)dF (x) ≈ t=1
a T

{xt }T
t=1 is a series drawn from a random number generator
This procedure is very simple and intuitive but
it is not very accurate (fast)
more powerful procedures available → numerical integration
Sedláček Monash Macro
 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Main idea of numerical integration

we want to calculate
Z b
I = f (x)dx
a

the basic idea is to approximate it with

n
X
I ≈ ωi f (xi )
i=1

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Main idea of numerical integration

n
X
I ≈ ωi f (xi )
i=1

therefore, we face (at least) three choices

choice of quadrature weights, ωi ’s
choice of quadrature nodes, xi ’s
choice of number of evaluations, n

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Types of quadrature methods

Newton-Cotes quadrature methods

break interval into equidistant intervals

approximate f (x) with a low-order polynomial
use integrals of polynomials as the approximations

Gaussain quadrature methods

same idea as with Newton-Cotes

more clever in choosing quadrature nodes

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Newton-Cotes

Types of Newton-Cotes quadrature methods

mid-point rule
interpolant is a constant

trapezoid rule
interpolant is linear

Simpson’s rule
interpolant is quadratic

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Simpson’s quadrature

X
I ≈ ωi P2 (xi )
i

choose 3 equidistant nodes in (each) interval [a, b] :

x0 = a, x1 = a + h and x2 = a + 2h = b

choose polynomial type to approximate f (x)

Simpson’s quadrature uses Lagrange polynomials
the beauty is that Simpson’s rule can be standardized!

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Lagrange basis functions

Pn (x) = a0 L0 (x) + ... + an Ln (x)

(x − x0 )...(x − xj−1 )(x − xj+1 )...(x − xn )
Lj (x) =
(xj − x0 )...(xj − xj−1 )(xj − xj+1 )...(xj − xn )

Lj ’s are polynomials so the approximation is a polynomial

the approximation gives an exact fit at the n + 1 nodes
(
1 if x = xj
Lj (x) =
0 if x ∈ {x0 , ..., xn } \ xj

what are the coefficients of the polynomial?

Sedláček Monash Macro
 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Simpson’s quadrature

Z b
I ≈ (f0 L0 (x) + f1 L1 (x) + f2 L2 (x)) dx
a
Z b Z b Z b
=f0 L0 (x)dx + f1 L1 (x)dx + f2 L2 (x)dx
a a a

doing the integration results in

Z b Z b Z b
L0 (x)dx = 1/3h L1 (x)dx = 4/3h L2 (x)dx = 1/3h
a a a

i.e. you can find quadrature weights

independent of the functional form of f!
Sedláček Monash Macro
 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Simpson’s quadrature

The above can be easily extended to n + 1 equidistant nodes

total number of nodes must be odd

this gives us n/2 segments of lenght h
apply the above idea for each of the segments

Z b 
1 4 2 4 2
f (x)dx ≈ f0 + f1 + f2 + f3 + f4 + · · ·
a 3 3 3 3 3

2 4 1
+ fn−2 + fn−1 + fn h
3 3 3

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Gaussian quadrature

Newton-Cotes formulas are simple due to equidistant nodes

moreover, one can show that with Newton-Cotes
we get the exact answer when the true function
is a polynomial of order n − 1

but we can get more accuracy by choosing nodes cleverly

we get the exact answer if the true function
is a polynomial of order 2n − 1!
i.e. 5 nodes give exact (accurate) answers for true functions
which are (approximated by) a 9th order polynomial!

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Procedure of Gaussian quadrature

R1
Using n nodes, we can approximate −1 f (x)dx as
Z 1 n
X
f (x)dx ≈ ωi f (ζi )
−1 i=1

i.e. we have 2n parameters

want correct answer for any polynomial of order 2n − 1
→ make sure we get correct answer for basis functions

1, x, x 2 , · · · , x 2n−1

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Procedure of Gaussian quadrature

How to choose weights and nodes?

to get the basis functions

we must solve for ωi and ζi ∀i s.t.
Z 1 n
ωi ζij j = 0, 1, · · · , 2n − 1
X
x j dx =
−1 i=1

i.e. solve a system of 2n equations in 2n unknowns

note that the solution is independent of f !

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Gauss-Legendre quadrature

the above method (for interval between −1 and 1)

is called Gauss-Legendre quadrature
nodes (ζiGL ) and weights (ωiGL ) satisfy above 2n conditions
the approximation is then given by
Z 1 n
X
f (x)dx ≈ ωiGL f (ζiGL )
−1 i=1

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Gauss-Hermite quadrature

when the true function is given by f (x) = g (x)W (x) where

g (x) can be approximated well by a polynomial
but f (x) cannot

then adjust the quadrature procedure depending on W (x)

Gauss-Hermite quadrature is used when W (x) = exp(−x 2 )
why is this an interesting case?

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Gauss-Hermite quadrature

nodes and weights chosen s.t.

Z ∞ n
ωi ζij j = 0, 1, · · · , 2n − 1
X
x j exp(−x 2 )dx =
−∞ i=1

the approximation is then given by

Z ∞ n
X
2
g (x) exp(−x )dx ≈ ωiGH f (ζiGH )
−∞ i=1

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Chebyshev quadrature

1
case when W (x) = (1−x 2 )1/2

nodes and weights again choses such that

Z 1 n
1
ωi ζij
X
j
x 2 1/2
dx =
−1 (1 − x ) i=1

resulting nodes ζiGC = nodes in Chebyshev polynomials

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Change of variable
suppose we want to calculate
Z ∞
(y − µ)2
 
g (y )
√ exp − dy
−∞ σ 2π 2σ 2
how to use Gauss-Hermite quadrature?
y −µ
√
define x = √
σ 2
or y = xσ 2 + µ
this gives
∞
√
Z
g (xσ 2 + µ) √
√ exp(−x 2 )σ 2dx
−∞ σ 2π
then use Gauss-Hermite quadrature
Z ∞ n √
(y − µ)2 GH
 
g (y ) X
GH h(ζi σ 2 + µ)
√ exp − dy ≈ ωi √
−∞ σ 2π 2σ 2 π
i=1

Sedláček Monash Macro

 Introduction Main idea
Motivation and content Newton-Cotes
Neoclassical growth model Gaussian quadrature
Function approximation Gauss-Hermite quadrature
Numerical integration Chebyshev quadrature
Summary Change of variable

Change of variable

could you instead simply do the following?

 
(y −µ)2
exp − 2
g (y ) 2σ
g (y ) = √
σ 2π exp(−y 2 )
R∞
and approximate −∞ g (y ) exp(−y 2 )?

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model
Function approximation
Numerical integration
Summary

What did we do?

1 motivation and overview of entire course

what tools and how are they used in practice?
2 a simple DSGE model
3 functional approximation
basics, polynomials, splines
4 numerical integration
quadrature methods

Sedláček Monash Macro

 Introduction
Motivation and content
Neoclassical growth model
Function approximation
Numerical integration
Summary

What’s next?

we now know the basic tools

apply them in two solution methods for DSGE models
projection
perturbation

Sedláček Monash Macro