of 32
7/29/2019 2 Linear Transformations
1/32
AGC
DSP
Professor A G Constantinides 1
Hilbert Spaces
Linear Transformations and Least Squares:
Hilbert Spaces
7/29/2019 2 Linear Transformations
2/32
AGC
DSP
Professor A G Constantinides 2
Linear Transformations A transformation from a vector space to a vector
space with the same scalar field denoted by
is linear when
Where
We can think of the transformation as an operator
YXL :
X
Y
2121
)( xxxxL
)()( xaLaxL
Xxxx 21,,
7/29/2019 2 Linear Transformations
3/32
AGC
DSP
Professor A G Constantinides 3
Linear Transformations Example: Mapping a vector space from to
can be expressed as a mxn matrix.
Thus the transformation
can be written as
nR mR
)43,2().,( 3221321 xxxxxxxL
3
2
1
2
1
430
021
x
x
x
y
y
7/29/2019 2 Linear Transformations
4/32
AGC
DSP
Professor A G Constantinides 4
Range space & Null space The range space of a transformation is
the set of all vectors that can be reached by the
transformation
The null space of the transformation is the set of all
vectors inX that are transformed to the null vector inY.
XLLR xxy :)()(
YXL :
XLLN x0x :)()(
7/29/2019 2 Linear Transformations
5/32
AGC
DSP
Professor A G Constantinides 5
Range space & Null space If is a projection operator then so is
Hence we have
Thus the vector is decomposed into two disjoint
parts. These parts are not necessarily orthogonal
If the range and null space are orthogonal then the
projections is said to be orthogonal
P PI
xP)x(IPx
x
7/29/2019 2 Linear Transformations
6/32
AGC
DSP
Professor A G Constantinides 6
Linear Transformations Example: Let
and let the transformation a nxm matrix
Then
Thus, the range of the linear transformation (orcolumn space of the matrix ) is the span of thebasis vectors.
The null space is the set which yields
Tmxxxx ...321x
mm332211 ppppAx xxxx ...
A
0Ax
7/29/2019 2 Linear Transformations
7/32
AGC
DSP
Professor A G Constantinides 7
A Problem Given a signal vector in the vector space S, we
want to find a point in the subset Vof the space ,
nearest to
x
v
x
x
0v1v
2v
0w
V
W00 vxw
7/29/2019 2 Linear Transformations
8/32
AGC
DSP
Professor A G Constantinides 8
A Problem Let us agree that nearest to in the figure is taken in
the Euclidean distance sense.
The projection orthogonal to the set Vgives thedesired solution.
Moreover the error of representation is
This vector is clearly orthogonal to the set V (Moreon this later)
00 vxw
0v
7/29/2019 2 Linear Transformations
9/32
AGC
DSP
Professor A G Constantinides 9
Another perspective We can look at the above problem as seeking to find
a solution to the set of linear equations
where the given vector is not in the range of
as is the case with an overspecified set of equations.
There is no exact solution. If we project orthogonallythe given vector into the range of then we have
the shortest norm solution in terms of the Euclidean
distance of the error.
xAv
v
x A
A
7/29/2019 2 Linear Transformations
10/32
AGC
DSP
Professor A G Constantinides 10
Another perspective The least error is then orthogonal to the data into
which we are projecting
Set
Then as in the above figure we can write
Where is the error, which is orthogonal to eachof the members of above.
m321 ppppA ...m321 ppppv mvvvv ...321
wAvx w
A
7/29/2019 2 Linear Transformations
11/32
AGC
DSP
Professor A G Constantinides 11
Another perspective Thus we can write
Or
mjj ,...,3,2,1,0, pAvx
0Avx
p
p
p
)(.2
1
H
m
H
H
0AvxA )(H
xAAAvHH 1)(
7/29/2019 2 Linear Transformations
12/32
AGC
DSP
Professor A G Constantinides 12
Another perspective Thus
and hence the projection matrix is
ie this is the matrix that projects orthogonally into the
column space of
HHAAAP
1)(
PxxAAAv HH 1)(
A
7/29/2019 2 Linear Transformations
13/32
AGC
DSP
Professor A G Constantinides 13
Another perspective
If we adopt the weighted form
The induced norm is
Then the projection matrix is
Where is positive definite
WAWAAAPHH 1)(
Wyxyx,W
H
WxxxW
H2
W
7/29/2019 2 Linear Transformations
14/32
AGC
DSP
Professor A G Constantinides 14
Least Squares Projection
PROJECTION THEOREM
In a Hilbert space the orthogonal projection of asignal into a smaller dimensional space minimises
the norm of the error, and the error vector is
orthogonal to the data (ie the smaller dimensional
space).
7/29/2019 2 Linear Transformations
15/32
AGC
DSP
Professor A G Constantinides 15
Orthogonality Principle Let be a set of
independent vectors in a vector spaceS.
We wish to express any vector inSas
If is in the span of the independent vectors then
the representation will be exact.
If on the other hand it is not then there will be an
error
}...{ m321 pppp
x
mm332211 ppppx xxxx ...
x
7/29/2019 2 Linear Transformations
16/32
AGC
DSP
Professor A G Constantinides 16
Orthogonality Principle In the latter case we can write
Where
is an approximation to given vector with error
We wish to find that approximation which minimises
the Euclidean error norm (squared)
i
m
iix px 1
exx
e
i
m
iii
m
iim xxxxJ pxpx
111 ,),...,(
7/29/2019 2 Linear Transformations
17/32
AGC
DSP
Professor A G Constantinides 17
Orthogonality Principle Expand to
Where T
mxxx ...21
m
iiji
m
jji
m
iim xxxxxJ
1
*
11
*1 ,,Re2,),...,( pppxxx
Rpx THHm
xxJ Re2),...,(2
1
Tmpxpxpxp ,...,, 21
7/29/2019 2 Linear Transformations
18/32
AGC
DSP
Professor A G Constantinides 18
Reminders
0axx
*
aax
x
H*
2/)Re(*
aaxx
H RxRxxx
H*
7/29/2019 2 Linear Transformations
19/32
AGC
DSP
Professor A G Constantinides 19
Orthogonality Principle
On setting this to zero we obtain the solution
This is a minimum because on differentiating we
have a positive definite matrix
RpRpx
THHRe22*
pR 1
7/29/2019 2 Linear Transformations
20/32
AGC
DSP
Professor A G Constantinides 20
Alternatively The norm squared of the error is
where
We note that
and
eeT
J
ee
T
iix
Jx
2
i
ix
pe
7/29/2019 2 Linear Transformations
21/32
AGC
DSP
Professor A G Constantinides 21
Orthogonality Principle At the minimum
Thus we haveand hence
Thus,
1) At the optimum the error is orthogonal to the data (Principle oforthogonality)
2)
0 xx,pe,p ii
022
epee
T
i
T
ii x
J
x
mixm
jjjiii ,...,1,,
1
ppxpx,p
7/29/2019 2 Linear Transformations
22/32
AGC
DSP
Professor A G Constantinides 22
Orthogonality Principle Thus for
Hence or
Tmpxpxpxp ,...,, 21
Tm
xxx ...21
mmmm
m
m
pppppp
pppppp
pppppp
R
,..,,
........
,..,,
,..,,
21
22212
12111
pR1
Rp
7/29/2019 2 Linear Transformations
23/32
AGC
DSP
Professor A G Constantinides 23
Orthogonalisation A signal may be projected into any linear space.
The computation of its coefficients in the various
vectors of the selected space is easier when thevectors in the space are orthogonal in that they are
then non-interacting, ie the evaluation of one such
coefficient will not influence the others
The error norm is easier to compute
Thus it makes sense to use an orthogonal set of
vectors in the space into which we are to project a
signal
7/29/2019 2 Linear Transformations
24/32
AGC
DSP
Professor A G Constantinides 24
Orthogonalisation Given any set of linearly independent vectors that
span a certain space, there is another set ofindependent vectors of the same cardinality, pair-
wise orthogonal, that spans the same space
We can think of the given set as a linear combinationof orthogonal vectors
Hence because of independence, the orthogonal
vectors is a linear combination of the given vectors
This is the basic idea behind the Gram-Schmidtprocedure
7/29/2019 2 Linear Transformations
25/32
AGC
DSP
Professor A G Constantinides 25
Gram-Schmidt Orthogonalisation The problem: (we consider finite dimensional spaces
only)
Given a set of linearly independent vectorsto determine a set of vectors that are pair-
wise orthogonal
Write the ith vector as
}{x
}{p
mixxxx mi
m
iii
i ,...,1...)(
3)(
32)(
21)(
1 ppppx
7/29/2019 2 Linear Transformations
26/32
AGC
DSP
Professor A G Constantinides 26
Gram-Schmidt Orthogonalisation If we knew the orthogonal set then the
coefficients of the expression can be determined as
the inner product
Step(1) The unknown orthogonal vector can be
oriented such that one of its members coincides withone of the members of the given set
Choose to be coincident with
}{x
}{p
2, jjji x ppx
1p 1x
7/29/2019 2 Linear Transformations
27/32
AGC
DSP
Professor A G Constantinides 27
Gram-Schmidt Orthogonalisation Step (2) Each member of has a projection
onto given by
Step(3) We construct
Step(4) Repeat the above on
}{x
21
)(11, ppx ii x
1p
}{u
mixi
ii ,...,21)(
1 pxu
7/29/2019 2 Linear Transformations
28/32
AGC
DSP
Professor A G Constantinides 28
Gram-Schmidt Orthogonalisation Example: Let
Then
And the projection of onto is
1
11x
2
12x
1
11p
11
121
2x 1p
7/29/2019 2 Linear Transformations
29/32
AGC
DSP
Professor A G Constantinides 29
Gram-Schmidt Orthogonalisation Form
Then
5.1
5.1
2/1
1
12
1
/1
2
112 ppx
5.15.1
2p
7/29/2019 2 Linear Transformations
30/32
AGC
DSP
Professor A G Constantinides 30
Gram-Schmidt
-2 -1.5 -1 -0.5 0 0.5 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1 1x
212 xxp a
11 xp
2x
Projection ofin
2x
1x
2
p
1
xa
7/29/2019 2 Linear Transformations
31/32
AGC
DSP
Professor A G Constantinides 31
3-D G-S Orthogonalisation
-1
-0.5
0
0.5
1
-1
-0.5
0
0.5
1-1
-0.5
0
0.5
1
7/29/2019 2 Linear Transformations
32/32
AGC
DSP
P f A G C t ti id 32
Gram-Schmidt Orthogonalisation Note that in the previous 4 steps we have
considerable freedom at Step 1 to choose any
vector not necessarily coincident with onefrom the given set of data vectors.
This enables us to avoid certain numerical ill-conditioning problems that may arise in the
Gram-Schmidt case.
Can you suggest when we are likely to haveill-conditioning in the G-S procedure?