Chapter 3 projection

Projection Methods

Jesús Fernández-Villaverde

University of Pennsylvania

July 10, 2011

Jesús Fernández-Villaverde (PENN) Projection Methods July 10, 2011 1 / 52

Introduction

Introduction

We come back to our functional equation:

H (d ) = 0

Projection methods solve the problem by specifying:

n
d n (x, θ ) = ∑ θ i Ψi (x )
i =0

We pick a basis fΨi (x )gi∞ 0 and “project” H ( ) against that basis to
=
…nd the θ i ’s.

How?

Introduction

Points to Emphasize

1 We may want to approximate di¤erent objects d: for instance a
decision rule, a value function, or an expectation.

2 In general we will have with the same number of parameters than
basis functions.

3 We will work with linear combinations of basis functions. Why? The
theory of nonlinear approximations is not yet as developed as the
linear case.


Introduction

Basic Algorithm
1 De…ne n known linearly independent functions ψi : Ω ! <m where
n < ∞. We call the ψ1 ( ) , ψ2 ( ) , ..., ψn ( ) the basis functions.
2 De…ne a vector of parameters θ = [θ 1 , θ 2 , ..., θ n ].
3 De…ne a combination of the basis functions and the θ’ s:
n
d n ( j θ) = ∑ θ i ψn ( )
i =1

4 Plug dn ( j θ ) into H ( ) to …nd the residual equation:
R ( j θ ) = H (d n ( j θ ))
5 Find the value of b that make the residual equation as close to 0 as
θ
possible given some objective function ρ : J 1 J 1 ! J 2 :
b = arg min ρ (R ( j θ ) , 0)
θ nθ 2<


Introduction

Relation with Econometrics

Looks a lot like OLS. Explore this similarity later in more detail.

Also with semi-nonparametric methods as Sieves.

Compare with:

1 Policy iteration.

2 Parameterized Expectations.


Introduction

Two Issues

We need to decide:

1 Which basis we use?

1 Pick a global basis)spectral methods.

2 Pick a local basis)…nite elements methods.

2 How do we “project”?

Di¤erent choices in 1 and 2 will result in slightly di¤erent projection
methods.

Introduction

Spectral Methods

Main reference: Judd (1992).

Spectral techniques use basis functions that are nonzero and smooth
almost everywhere in Ω.

Advantages: simplicity.

Disadvantages: di¢ cult to capture local behavior. Gibbs phenomenon.


Introduction

Spectral Basis I
Monomials: c, x, x 2 , x 3 , ...
Simple and intuitive.
Even if this basis is not composed by orthogonal functions, if J1 is the
space of bounded measurable functions on a compact set, the
Stone-Weierstrass theorem assures completeness in the L1 norm.
Problems:
1 (Nearly) multicollinearity. Compare the graph of x 10 with x 11 .
The solution of a projection involves matrices inversion. When the
basis functions are similar, the condition number of these matrices (the
ratio of the largest and smallest absolute eigenvalues) are too high.
Just the six …rst monomials can generate conditions numbers of 1010 .
The matrix of the LS problem of …tting a polynomial of degree 6 to a
function (the Hilbert Matrix), is a popular test of numerical accuracy
since it maximizes rounding errors!
2 Monomials vary considerably in size, leading to scaling problems and
accumulation of numerical errors.
We want an orthogonal basis.

Introduction

Spectral Basis II
Trigonometric series

1/ (2π )0.5 , cos x / (2π )0.5 , sin x / (2π )0.5 , ...,
cos kx / (2π )0.5 , sin kx / (2π )0.5 , ...

Periodic functions.

However economic problems are generally not periodic.

Periodic approximations to nonperiodic functions su¤er from the
Gibbs phenomenon, requiring many terms to achieve good numerical
performance (the rate of convergence to the true solution as n ! ∞
is only O (n)).


Introduction

Spectral Basis III
Flexible class: orthogonal polynomials of Jacobi (or hypergeometric)
type. Why orthogonal?

α,β
The Jacobi polynomial of degree n, Pn (x ) for α, β > 1, is de…ned
by the orthogonality condition:
Z 1
α,β α,β
(1 x )α (1 + x ) β Pn (x ) Pm (x ) dx = 0 for m 6= n
1

The two most important cases of Jacobi polynomials:

Legendre: α = β = 1
2.
1

2 Chebyshev: α = β = 0.

Introduction

Alternative Expressions
The orthogonality condition implies, with the customary
normalizations:
α,β n+α
Pn (1) =
n

that the general n term is given by:
n
n+α n+β
2 n
∑ k n k
(x 1)n k
(x + 1)k
k =0

Recursively:
2 (n + 1) (n + α + β + 1) (2n + α + β) Pn +1 =
(2n + α + β + 1) α2 β2
Pn
+ (2n + α + β) (2n + α + β + 1) (2n + α + β + 2) x
2 (n + α) (n + β) (2n + α + β + 2) Pn 1


Introduction

Chebyshev Polynomials
One of the most common tools of Applied Mathematics.
References:
Chebyshev and Fourier Spectral Methods, John P. Boyd (2001).
A Practical Guide to Pseudospectral Methods, Bengt Fornberg (1998).

Advantages of Chebyshev Polynomials:
1 Numerous simple close-form expressions are available.
2 The change between the coe¢ cients of a Chebyshev expansion of a
function and the values of the function at the Chebyshev nodes are
quickly performed by the cosine transform.
3 They are more robust than their alternatives for interpolation.
4 They are bounded between [ 1, 1] while Legendre polynomials are not,
o¤ering a better performance close to the boundaries of the problems.
5 They are smooth functions.
6 Several theorems bound the errors for Chebyshev polynomials
interpolations.

Introduction

De…nition of Chebyshev Polynomials I
Recursive de…nition:

T0 ( x ) = 1
T1 ( x ) = x
Tn +1 (x ) = 2xTn (x ) Tn 1 (x ) for a general n
The …rst few polynomials are then 1, x, 2x 2 1, 4x 3 3x,
8x 4 8x 2 + 1, etc...

The n zeros of the polynomial Tn (xk ) = 0 are given by:
2k 1
xk = cos π, k = 1, ..., n
2n

Note that zeros are clustered quadratically towards 1.

Introduction


Introduction

De…nition of Chebyshev Polynomials II

Explicit de…nition:

Tn (x ) = cos (n arccos x )
1 1 1 1
= z n + n where z+ =x
2 z 2 z
1 0.5 n 0.5 n
= x + x2 1 + x x2 1
2
[n/2 ]
1 (n k 1) !
=
2 ∑ ( 1)k
k! (n 2k )!
(2x )n 2k

k =0
( 1)n π 0.5 0.5 dn n 1
= n 1 x2 1 x2 2

2 Γ n+ 1 2
dx n


Introduction

Remarks

The domain of the Chebyshev polynomials is [ 1, 1]. Since our state
space is, in general, di¤erent, we use a linear translation from [a, b ]
into [ 1, 1] :
x a
2 1
b a
Chebyshev polynomials are orthogonal with respect to the weight
function:
1
(1 x 2 )0.5

Chebyshev Interpolation Theorem
th
if an approximating function is exact at the roots of the n1 order
Chebyshev polynomial then, as n1 ! ∞, the approximation error becomes
arbitrarily small.


Introduction

Multidimensional Problems

Chebyshev polynomials are de…ned on [ 1, 1].

However, most problems in economics are multidimensional.

How do we generalize the basis?

Curse of dimensionality.


Introduction

Tensors
Assume we want to approximate F : [ 1, 1]d ! R.

Let Tj denote the Chebyshev polynomial of degree j = 0, 1, .., κ.

We can approximate F with tensor product of Chebyshev polynomials
of degree κ:

κ κ
ˆ
F (x ) = ∑ ... ∑ ξ n1 ,...,nd Tn1 (x1 ) Tnd (xd )
n 1 =0 n d =0

Beyond simplicity, an advantage of the tensor basis is that if the
one-dimensional basis is orthogonal in a norm, the tensor basis is
orthogonal in the product norm.

Disadvantage: number of elements increases exponentially. We end
up having terms x1 x2
κ κ xd , total number of (κ + 1)d .
κ


Introduction

Complete Polynomials

Solution: eliminate some elements of the tensor in such a way that
there is not much numerical degradation.

Judd and Gaspar (1997): Use complete polynomials instead
( )
d
d
Pκ i
x11 i
xdd with ∑ il κ, 0 i1 , ..., id
l =1

Advantage: much smaller number of terms, no terms of order dκ to
evaluate.

Disadvantage: still too many elements.


Introduction

Smolyak’ Algorithm I
s

De…ne m1 = 1 and mi = 2i 1 + 1, i = 2, ....

De…ne G i = fx1 , ..., xm i g
i i [ 1, 1] as the set of the extrema of the
Chebyshev polynomials
π (j 1)
xji = cos j = 1, ..., mi
mi 1

with G 1 = f0g. It is crucial that G i G i +1 , 8i = 1, 2, . . .

Example:

i = 1, mi = 1, G i = f0g
i = 2, mi = 3, G i = f 1, 0, 1g
π 3π
i = 3, mi = 5, G i = f 1, cos , 0, cos , 1g
4 4


Introduction

Smolyak’ Algorithm II
s
For q > d, de…ne a sparse grid
[
H(q, d ) = (G i1 ... G id ),
q d +1 ji j q

where ji j = i1 + . . . + id . The number q de…nes the size of the grid
and thus the precision of the approximation.
For example, let q = d + 2 = 5:
[
H(5, 3) = (G i1 ... G id ).
3 ji j 5

G3 G1 G1, G1 G3 G1, G1 G1 G3
G2 G2 G1, G2 G1 G2, G1 G2 G2
G2 G1 G1, G1 G2 G1, G1 G1 G2
G1 G1 G1

Introduction

Smolyak’ Algorithm III
s

Number of points for q = d + 2
d (d 1)
1 + 4d + 4
2
Largest number of points along one dimension

i = q d +1
mi = 2q d + 1

Rectangular grid
h id
2q d
+1
Key: with rectangular grid, the number of grid points increases
exponentially in the number of dimensions. With the Smolyak
algorithm number of points increases polynomially in the number of
dimensions.

Introduction

Smolyak’ Algorithm IV
s

Size of the Grid for q = d + 2
d
d 2q d +1 #H(q, d ) 2q d + 1
2 5 13 25
3 5 25 125
4 5 41 625
5 5 61 3, 125
12 5 313 244, 140, 625


Introduction

Smolyak’ Algorithm V
s

For one dimension denote the interpolating Chebyshev polynomials as
mi
U i (x i ) = ∑ ξ ij Tj (x i )
j =1

and the d-dimensional tensor product by U i1 ... U id (x ).
For q > d, approximating function (Smolyak’ algorithm) given by
s

d 1
A(q, d )(x ) = ∑ ( 1)q ji j
q ji j
(U i1 ... U id )(x )
q d +1 ji j q

Method is (almost) optimal within the set of polynomial
approximations (Barthelmann, Novak, and Ritter, 1999).
Method is universal, that is, almost optimal for many di¤erent
function spaces.

Introduction

Boyd’ Moral Principal
s

1 When in doubt, use Chebyshev polynomials unless the solution is
spatially periodic, in which case an ordinary Fourier series is better.

2 Unless you are sure another set of basis functions is better, use
Chebyshev polynomials.

3 Unless you are really, really sure another set of basis functions is
better, use Chebyshev polynomials.


Introduction

Finite Elements

Standard Reference: McGrattan (1999).

Bound the domain Ω in small of the state variables.

Partition Ω in small in nonintersecting elements.

These small sections are called elements.

The boundaries of the elements are called nodes.


Introduction

Partition into Elements

Elements may be of unequal size.

We can have small elements in the areas of Ω where the economy will
spend most of the time while just a few, big size elements will cover
wide areas of the state space infrequently visited.

Also, through elements, we can easily handle issues like kinks or
constraints.

There is a whole area of research concentrated on the optimal
generation of an element grid. See Thomson, Warsi, and Mastin
(1985).


Introduction

Structure
Choose a basis for the policy functions in each element.
Since the elements are small, a linear basis is often good enough:
8 x x
> x xi 1 if x 2 [xi 1 , xi ]
< i i 1
x i +1 x
ψi (k ) = if k 2 [xi , xi +1 ]
> x i +1 x i
: 0 elsewhere

Plug the policy function in the Equilibrium Conditions and …nd the
unknown coe¢ cients.
Paste it together to ensure continuity.
Why is this an smart strategy?
Advantages: we will need to invert an sparse matrix.
When should be choose this strategy? speed of computation versus
accuracy.

Introduction

Three Di¤erent Re…nements

1 h-re…nement: subdivide each element into smaller elements to
improve resolution uniformly over the domain.

2 r-re…nement: subdivide each element only in those regions where
there are high nonlinearities.

3 p-re…nement: increase the order of the approximation in each
element. If the order of the expansion is high enough, we will
generate in that way an hybrid of …nite and spectral methods knows
as spectral elements.


Introduction

Choosing the Objective Function

The most common answer to the second question is given by a
weighted residual.
That is why often projection methods are also called weighted
residual methods
This set of techniques propose to get the residual close to 0 in the
weighted integral sense.
Given some weight functions φi : Ω ! <m :
R
0 if Ω φi (x ) R ( j θ ) dx = 0, i = 1, .., n
ρ (R ( j θ ) , 0) =
1 otherwise

Then the problem is to choose the θ that solve the system of
equations: Z
φi (x ) R ( j θ ) dx = 0, i = 1, .., n
Ω


Introduction

Remarks

With the approximation of d by some functions ψi and the de…nition
of some weight functions φi ( ), we have transform a rather
intractable functional equation problem into the standard nonlinear
equations system!

The solution of this system can be found using standard methods, as
a Newton for relatively small problems or a conjugate gradient for
bigger ones.

Issue: we have di¤erent choices for an weight function:


Introduction

Weight Function I: Least Squares
∂R ( x jθ )
φi (x ) = ∂θ i .

This choice is motivated by the solution of the variational problem:
Z
min R 2 ( j θ ) dx
θ Ω

with …rst order condition:
Z
∂R ( x j θ )
R ( j θ ) dx = 0, i = 1, .., n
Ω ∂θ i

Variational problem is mathematically equivalent to a standard
regression problem in econometrics.

OLS or NLLS are regression against a manifold spanned by the
observations.

Introduction

Weight Function I: Least Squares

Least Squares always generates symmetric matrices even if the
operator H is not self-adjoint.

Symmetric matrices are convenient theoretically (they simplify the
proofs) and computationally (there are algorithms that exploit their
structure to increase speed and decrease memory requirements).

However, least squares may lead to ill-conditioning and systems of
equations complicated to solve numerically.


Introduction

Weight Function II: Subdomain

We divide the domain Ω in n subdomains Ωi and de…ne the n step
functions:
1 if x 2 Ωi
φi (x ) =
0 otherwise

This choice is then equivalent to solve the system:
Z
R ( j θ ) dx = 0, i = 1, .., n
Ωi


Introduction

Weight Function III: Moments
Take 0, x, x 2 , ..., x n 1 and compute the …rst n periods of the
residual function:
Z
x i R ( j θ ) dx = 0, i = 0, .., n
Ωi

This approach, widely used in engineering works well for a low n (2 or
3).

However, for higher orders, its numerical performance is very low:
high orders of x are highly collinear and arise serious rounding error
problems.

Hence, moments are to be avoided as weight functions.

Introduction

Weight Function III: Collocation or Pseudospectral or
Method of Selected Points
φi (x ) = δ (x xi ) where δ is the dirac delta function and xi are the
collocation points.
This method implies that the residual function is zero at the n
collocation points.
Simple to compute since the integral only needs to be evaluated in
one point. Specially attractive when dealing with strong nonlinearities.
A systematic way to pick collocation points is to use a density
function:
Γ 3 γ 2
µ γ (x ) = 1 γ<1
(1 x 2 ) γ π 2 Γ (1 γ)
and …nd the collocation points as the xj , j = 0, ..., n 1 solutions to:
Z xj
j
µγ (x ) dx =
1 n
For γ = 0, the density function implies equispaced points.

Introduction

Weight Function IV: Orthogonal Collocation

Variation of the collocation method:

1 Basis functions are a set of orthogonal polynomials.

2 Collocation points given by the roots of the n th polynomial.

When we use Chebyshev polynomials, their roots are the collocation
points implied by µ 1 (x ) and their clustering can be shown to be
2
optimal as n ! ∞.

Surprisingly good performance of orthogonal collocation methods.


Introduction

Weight Function V: Galerkin or Rayleigh-Ritz
φi (x ) = ψi (x ) with a linear approximating function ∑n=1 θ i ψi (x ).
i
Then: !
Z n

Ω
ψi (x ) H ∑ θ i ψi (x ) dx = 0, i = 1, .., n
i =1
that is, the residual has to be orthogonal to each of the basis
functions.
Galerkin is a highly accurate and robust but di¢ cult to code.
If the basis functions are complete over J1 (they are indeed a basis of
the space), then the Galerkin solution will converge pointwise to the
true solution as n goes to in…nity:
n

n !∞
∑ θ i ψi ( ) = d ( )
lim
i =1

Experience suggests that a Galerkin approximation of order n is as
accurate as a Pseudospectral n + 1 or n + 2 expansion.

Introduction

A Simple Example

Imagine that the law of motion for the price x of a good is given by:

d 0 (x ) + d (x ) = 0

Let us apply a simple projection to solve this di¤erential equation.

Code: test.m, test2.m, test3.m


Introduction

Analysis of Error

As with projection, it is important to study the Euler equation errors.

We can improve errors:

1 Adding additional functions in the basis.

2 Re…ne the elements.

Multigrid schemes.


Introduction

A More Serious Example
Representative agent with utility function

1 τ
∞ ctθ (1 lt )1 θ

U = E0 ∑ βt 1 τ
t =0

One good produced according to yt = e zt Aktα lt1 α
with α 2 (0, 1) .

Productivity evolves zt = ρzt 1 + t, jρj < 1 and t N (0, σ ).

Law of motion for capital kt +1 = it + (1 δ)kt .

Resource constrain ct + it = yt .

Introduction

Solve for c ( , ) and l ( , ) given initial conditions.

Characterized by:

Uc (t ) = βEt Uc (t + 1) 1 + αAe zt +1 ktα+1 l (kt +1 , zt +1 )α
1
δ

1 θ c (kt , zt )
= (1 α) e zt Aktα l (kt , zt ) α
θ 1 l (kt , zt )

A system of functional equations with no known analytical solution.

Fortran code using Chebyshev and Finite Elements.


Introduction

Chebyshev I
We approximate the decision rules for labor as lt = ∑n=1 θ i ψi (kt , zt )
i
where fψi (k, z )gn=1 are basis functions and θ = [fθ i gn=1 ] unknown
i i
coe¢ cients.

We use that policy function to solve for consumption using the static
…rst order condition.

We build a residual function R (k, z, θ ) using the Euler equation and
the static …rst order condition.

Then we choose θ by solving:
Z Z
φi (k, z ) R (k, z, θ ) = 0 for i = 1, ..., n
[kmin ,kmax ] [zmin ,zmax ]

where fφi (k, z )gn=1 are some weight functions.
i

Introduction

Chebyshev I

We use a collocation method that sets φi (k, z ) = δ (k kj , z zv )
where δ (.) is the dirac delta function, j = 1, ..., n1 , v = 1, ..., n2 and
n = n1 n2 and collocation points fkj gjn=1 and fzv gn2 1 .
1
v=

For the technology shocks and transition probabilities we use Tauchen
(1986)’ …nite approximation to an AR(1) process and obtain n2
s
points.

We solve the system of n equations R (ki , zi , θ ) = 0 in n unknowns θ
using a Quasi-Newton method.

We use an iteration based on the increment of the number of basis
functions and a nonlinear transform of the objective function (apply
(u 0 ) 1 ).


Introduction

Finite Elements

Rewrite Euler equation as
Z ∞ 2
β t +1
Uc (kt , zt ) = [Uc (kt +1 , zt +1 )(r (kt +1 , zt +1 )] exp( )d t +1
(2πσ)0.5 ∞ 2σ2

where

Uc (t ) = Uc (kt , zt )
z t +1 α 1 α
kt +1 = e kt lt + (1 δ)kt c (kt , zt )
r (kt +1 , zt +1 ) = 1 + αe zt +1 ktα+1 l (kt +1 , zt +1 )1 α
1
δ

and
zt +1 = ρzt + t +1


Introduction

Goal

The problem is to …nd two policy functions
c (k, z ) : R + [0, ∞] ! R + and l (k, z ) : R + [0, ∞] ! [0, 1] that
satisfy the model equilibrium conditions.

Since the static …rst order condition gives a relation between the two
policy functions, we only need to solve for one of them.

For the rest of the exposition we will assume that we actually solve
for l (k, z ) and then we …nd c (l (k, z )).


Introduction

Bounding the State Space I

We bound the domain of the state variables to partition it in
nonintersecting elements.

To bound the productivity level of the economy de…ne λt = tanh(zt ).

Since λt 2 [ 1, 1] we can write the stochastic process as:

1 0.5
λt = tanh(ρ tanh (zt 1 ) + 2 σvt )

where vt = t
2 0.5 σ
.


Introduction

Bounding the State Space II
1 (1 +λt +1 )0.5 b
Now, since exp(tanh (zt 1 )) = = λt +1 , we have:
(1 λt +1 )0.5

Z 1
β
Uc (t ) = 0.5 [Uc (kt +1 , zt +1 )r (kt +1 , zt +1 )] exp( vt2+1 )dvt +1
π 1

where

b
kt +1 = λt +1 ktα l (kt , zt )1 α + (1 δ)kt c (l (kt , zt ))
b
r (kt +1 , zt +1 ) = 1 + αλt +1 k α 1 l (kt +1 , zt +1 )1 α δ
t +1
1
and zt +1 = tanh(ρ tanh (zt ) + 20.5 σvt +1 ).

To bound the capital we …x an ex-ante upper bound kmax , picked
su¢ ciently high that it will only bind with an extremely low
probability.

Introduction

Partition into Elements

De…ne Ω = [0, kmax ] [ 1, 1] as the domain of lfe (k, z; θ ).

Divide Ω into nonoverlapping rectangles [ki , ki +1 ] [zj , zj +1 ], where
ki is the ith grid point for capital and zj is jth grid point for the
technology shock.

Clearly Ω = [i ,j [ki , ki +1 ] [zj , zj +1 ].


Introduction

Our Functional Basis
b e
Set lfe (k, z; θ ) = ∑i ,j θ ij Ψij (k, z ) = ∑i ,j θ ij Ψi (k ) Ψj (z ) where

8 k ki 1
>
< ki ki 1 if k 2 [ki 1 , ki ]
b
Ψi (k ) = k i +1 k
if k 2 [ki , ki +1 ]
>
: k i +1 k i
0 elsewhere
8 z zj 1
>
< zj zj 1 if z 2 [zj 1 , zj ]
e
Ψj (z ) = z j +1 z
if z 2 [zj , zj +1 ]
>
: z j +1 z j
0 elsewhere

Note that:
1 Ψ (k, z ) = 0 if (k, z ) 2 [k
/ i 1 , ki ] zj 1 , zj [ [ki , ki +1 ] zj , zj +1
ij
8i, j, i.e. the function is 0 everywhere except inside two elements.
2 l (k , z ; θ ) = θ
fe i j ij 8i, j, i.e. the values of θ specify the values of cfe at
the corners of each subinterval [ki , ki +1 ] zj , zj +1 .

Introduction

Residual Function I

De…ne Uc (kt +1 , zt +1 )fe as the marginal utility of consumption
evaluated at the …nite element approximation values of consumption
and leisure.
From the Euler equation we have a residual equation:

R (kt , zt ; θ ) =
Z 1
β Uc (kt +1 , zt +1 )fe
r (kt +1 , zt +1 ) exp( vt2+1 )dvt +1 1
π 0.5 1 Uc (kt +1 , zt +1 )fe

A Galerkin scheme implies that we weight the residual function by the
basis functions and solve the system of θ equations
Z
Ψi ,j (k, z ) R (k, z; θ )dzdk = 0 8i, j
[0,kmax ] [ 1,1 ]

on the θ unknowns.

Introduction

Residual Function II

Since Ψij (k, z ) = 0 if
(k, z ) 2 [ki 1 , ki ] [zj
/ 1 , zj ] [ [ki , ki +1 ] [zj , zj +1 ] 8i, j we have
Z
Ψi ,j (k, z ) R (k, z; θ )dzdk = 0 8i, j
[k i 1 ,k i ] [z j 1 ,z j ][[k i ,k i +1 ] [zj ,zj +1 ]

We use Gauss-Hermite for the integral in the residual equation and
Gauss-Legendre for the integrals in Euler equation.

We use 71 unequal elements in the capital dimension and 31 on the λ
axis. To solve the associated system of 2201 nonlinear equations we
use a Quasi-Newton algorithm.


Chapter 3 projection

More Related Content

Chapter 3 projection