symbolic dynamics of digital signal processing systemseprints.lincoln.ac.uk › 4008 › 2 ›...
TRANSCRIPT
1
Symbolic Dynamics of Digital Signal Processing Systems
Dr. Bingo Wing-Kuen Ling
School of Engineering, University of Lincoln.Brayford Pool, Lincoln, Lincolnshire, LN6
7TS, United Kingdom.Email: [email protected]
2
My research interestsSymbolic dynamics
Limit cycle, fractal and chaotic behaviors of digital filters with two’s complement arithmetic, sigma delta modulators and perceptron training algorithms
Optimization theorySemi-infinite programming, nonconvex optimization and nonsmooth optimization with applications to filter, filter bank, wavelet kernel and pulse designs
Time frequency analysisNonuniform filter banks, filter banks with block samplers to multimedia and biomedical signal processing
Control theoryImpulsive control, optimal control, fuzzy control and chaos control for HIV model and avian influenza model
3
Acknowledgements
Prof. Xinghuo Yu (Royal Melbourne of Institute of Technology)Dr. Joshua D. Reiss (Queen Mary University of London)Dr. Herbert Ho-Ching Iu (University of Western Australia)Dr. Hak-Keung Lam (King’s College London)
4
OutlineIntroductionDigital filters with two’s complement arithmeticSigma delta modulatorsPerceptron training algorithmsConclusionsQ&A Session
5
IntroductionDefinition
Symbolic dynamics is a kind of system dynamics which involves multi-level signals.
MotivationsMany practical signal processing systems, such as digital filters with two’s complement arithmetic, sigma delta modulators and perceptron training algorithms, are symbolic dynamical systems. They are found in almost everywhere.
6
IntroductionChallenges
Symbolic dynamical systems could lead to chaotic, fractals and divergent behaviors.Stability conditions are generally unknown.Admissibility conditions are generally unknown.
7
Digital filters with two’s complement arithmetic
What are digital filters?Digital filters are systems that are characterized in the frequency domain.
( ) ( )( )ωωω
UYH =
h(k)( )ku ( )ky
( ) ( )∑∀
−=k
kjekuU ωω ( ) ( )∑∀
−=k
kjekyY ωω( ) ( )∑∀
−=k
kjekhH ωω
8
Types of digital filtersLowpass filters
Allow a low frequency band of a signal to pass through and attenuate a high frequency band.
Digital filters with two’s complement arithmetic
9
Types of digital filtersBandpass filters
Allow intermediate frequency bands of a signal to pass through and attenuate both low and high frequency bands.
Digital filters with two’s complement arithmetic
10
Types of digital filtersHighpass filters
Allow a high frequency band of a signal to pass through and attenuate a low frequency band.
Digital filters with two’s complement arithmetic
11
Types of digital filtersBand reject filters
Allow both low and high frequency bands of a signal to pass through and attenuate intermediate frequency bands.
Digital filters with two’s complement arithmetic
12
Types of digital filtersNotch filters
Allow almost all frequency components to pass through but attenuate particular frequencies.
Digital filters with two’s complement arithmetic
13
Types of digital filtersOscillators
Allow particular frequency components to pass through and attenuate almost all frequency components.
Digital filters with two’s complement arithmetic
14
Types of digital filtersAllpass filters
Allow all frequency components to pass through.
0 1 2 3 40
0.5
1
1.5
2
Frequency ω
Mag
nitu
de re
spon
se |H
( ω)|
(g)
Digital filters with two’s complement arithmetic
15
Digital filters with two’s complement arithmetic
Hardware schematic
z-1
z-1
a
b
Accumulator f(•)( )ku ( )ky
16
When no input is present, the filter can be described by the following nonlinear state space difference equation:
( ) ( )( ) ( )( )
( )( ) ( )( ) ⎟
⎟⎠
⎞⎜⎜⎝
⎛⎥⎦
⎤⎢⎣
⎡⋅+⋅
=
=⎥⎦
⎤⎢⎣
⎡++
≡+
kxakxbfkx
kkxkx
k
21
2
2
1 F11
1 xxformDirect
Digital filters with two’s complement arithmetic
17
where f(•) is the nonlinear function associated with the two’s complement arithmetic
1 3-1-3
1
-1
v
f(v) linesStraight
overflow No OverflowOverflow
Digital filters with two’s complement arithmetic
18
where
and m is the minimum integer satisfying
( )( ) ( ){ }
( ) { }mmks
xxxxIkxkx
ab
,,1,0,1,,
11,11:,
10
21212
2
1
LL −−∈
<≤−<≤−≡∈⎥⎦
⎤⎢⎣
⎡
⎥⎦
⎤⎢⎣
⎡=A
( ) ( )( ) ( )kskxkx
k ⋅⎥⎦
⎤⎢⎣
⎡+⎥
⎦
⎤⎢⎣
⎡⋅=+
20
12
1Ax
1212 21 +⋅≤⋅+⋅≤−⋅− mxaxbm
Digital filters with two’s complement arithmetic
19
is called symbolic sequences.( )ks
( ) ( ) ( )( ) ( ) ( )( ) ( ) ( ) 311
110131
21
21
21
<⋅+⋅≤⇒−=<⋅+⋅≤−⇒=−<⋅+⋅≤−⇒=
kxakxbkskxakxbkskxakxbks
Digital filters with two’s complement arithmetic
20
For example: ( ) [ ] 5.0 and 1,6135.06135.00 =−=−= abTx( ) ( ) ( )( ) ( ) ( )
( ) ( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) ( ) ( )( ) ( ) ( ) ( )
( )LL
M
M
11001000
9161.019 and 1180839.118189311.018 and 1170689.11717
0166184.016160157597.01515
9982.015 and 1140018.11414
011534.011009203.000
221
221
21
21
221
21
21
−−=
==⇒−=⋅+⋅−=−=⇒=⋅+⋅
=⇒=⋅+⋅=⇒−=⋅+⋅
−=−=⇒=⋅+⋅
=⇒−=⋅+⋅=⇒=⋅+⋅
s
xsxaxbxsxaxb
sxaxbsxaxb
xsxaxb
sxaxbsxaxb
valuecomplement sTwo'
Digital filters with two’s complement arithmetic
21
Stability analysis of the corresponding linear system
Eigenvalues of A
242 baa ⋅+±
=λ
Digital filters with two’s complement arithmetic
22
R1 R3
R5
R2
R4
-1
2-2
1
a
b
R3R1
Digital filters with two’s complement arithmetic
23
R5 :Magnitudes of both eigenvalues <1.⇒The corresponding linear system is stable.
R1 and R3 : Either one of the magnitudes of the eigenvalues < 1.⇒The corresponding linear system is unstable.R2 and R4 :Magnitudes of both eigenvalues are greater than 1.⇒The corresponding linear system is unstable.
Digital filters with two’s complement arithmetic
24
Effects of different initial conditions when the eigenvalues of system matrix are on the unit circle
Type I trajectoryThere is a single rotated and translated ellipse in the phase portrait.
( ) ⎥⎦
⎤⎢⎣
⎡−
=−==612.0
612.00,1,5.0 xba
Digital filters with two’s complement arithmetic
25
Digital filters with two’s complement arithmetic
26
Type II trajectoryThere are some rotated and translated ellipses in the phase portrait.
( ) ⎥⎦
⎤⎢⎣
⎡−
=−==616.0
616.00,1,5.0 xba
Digital filters with two’s complement arithmetic
27This corresponds to the limit cycle behavior.
Digital filters with two’s complement arithmetic
28
Type III trajectoryThere is a fractal pattern on the phase portrait.
( ) ⎥⎦
⎤⎢⎣
⎡−
=−==6135.0
6135.00,1,5.0 xba
Digital filters with two’s complement arithmetic
29Conclusion: Very sensitive to initial conditions
Digital filters with two’s complement arithmetic
30
Step response of second order digital filters with two’s complement arithmetic
The system can be represented as:
where and⇒DC offset on symbolic sequences.
( ) ( )( ) ( ) ( )( )
( )( ) ( ) ( )kskukxkx
kukxakxbfkx
k
⋅⎥⎦
⎤⎢⎣
⎡+⋅+⎥
⎦
⎤⎢⎣
⎡⋅=
⎟⎟⎠
⎞⎜⎜⎝
⎛⎥⎦
⎤⎢⎣
⎡+⋅+⋅
=+
20
1
2
1
21
2
BA
x
⎥⎦
⎤⎢⎣
⎡=
10
B⎥⎦
⎤⎢⎣
⎡=
ab10
A
Digital filters with two’s complement arithmetic
31
For ( ) ⎥⎦
⎤⎢⎣
⎡==−=−=
3.03.0
0,1,1,5.1 xcba
Digital filters with two’s complement arithmetic
32
For 1 and 1,5.1 =−=−= cba
s(k)=0
s(k)=1
s(k)=-1
s(k)=-2
Digital filters with two’s complement arithmetic
33
For ( ) ( )L,0,1,1,0,7.07.0
0,1,1,5.1 −−=⎥⎦
⎤⎢⎣
⎡−−
==−=−= scba x
Digital filters with two’s complement arithmetic
34
For 1 and 1,5.1 =−=−= cba
Digital filters with two’s complement arithmetic
35
For ( ) ⎥⎦
⎤⎢⎣
⎡−−
==−=−=99.099.0
0,1,1,5.1 xcba
Digital filters with two’s complement arithmetic
36
For 1 and 1,5.1 =−=−= cba
Digital filters with two’s complement arithmetic
37
Sinusoidal response of second order digital filters with two’s complement arithmetic
Digital filters with two’s complement arithmetic
38
Digital filters with two’s complement arithmetic
39
Digital filters with two’s complement arithmetic
Autonomous case when the eigenvalues of A are inside the unit circle
40
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1–1
–0.8
–0.6
–0.4
–0.2
0
0.2
0.4
0.6
0.8
1
x1
x2
Digital filters with two’s complement arithmetic
41
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
x1
x2
Digital filters with two’s complement arithmetic
42
Autonomous case for third order digital filters with two’s complement arithmetic realized in cascade form
System block diagram:
( )kx3
Delay
Delay
Delay
c
( )•f
( )kx2
( )kx1b
a
( )•f( )ky1
( )ky
Digital filters with two’s complement arithmetic
43
Digital filters with two’s complement arithmetic
44
Digital filters with two’s complement arithmetic
45
Autonomous case for third-order digital filters with two’s complement arithmetic realized in parallel form
System block diagram:
( )kx3
Delay
Delay
Delay
c
( )•f
( )kx2
( )kx1b
a
( )•f
( )ky1
( )ky2
( )ky
Digital filters with two’s complement arithmetic
46
Digital filters with two’s complement arithmetic
47
Digital filters with two’s complement arithmetic
48
Admissibility of second order digital filters with two’s complement arithmetic
Admissible set of periodic symbolic sequences is defined as a set of periodic symbolic sequences such that there exists an initial condition that produces the symbolic sequences.
sx(0)
Admissible set of periodic symbolic sequences
Set of initial conditions
Digital filters with two’s complement arithmetic
49
For example, whens Admissible
,0,1,5.0 =−== cbaM1 ( )L,0=s2 ( )L,1,1−=s ( )L,1,1 −=s
3 ( )L,1,0,0=s ( )L,0,1,0=s( )L,0,0,1=s( )L,1,0,0 −=s ( )L,0,1,0 −=s
( )L,0,0,1−=s
s admissibleNot ( )L,1=s ( )L,1−=s( )L,0,1=s ( )L,1,0=s( )L,0,1−=s ( )L,1,0 −=s( )L,1,1,0=s ( )L,1,1,0 −=s( )L,1,1,0 −=s( )L,1,0,1=s ( )L,1,0,1 −=s
( )L,1,1,0 −−=s
( )L,0,1,1 −=s ( )L,0,1,1 −=sL15 foundNot L
Digital filters with two’s complement arithmetic
50
A periodic sequence with period is admissible if and only if for
s M
( )( )1
sin2
sin
12
cos,mod1
1
0 <⋅⎟⎠⎞
⎜⎝⎛ ⋅
⎟⎠
⎞⎜⎝
⎛⋅⎟⎠⎞
⎜⎝⎛ −−⋅+
≤−∑−
=
θθ
θ
M
jMMjisM
j
1,,1,0 −= Mi L
Digital filters with two’s complement arithmetic
51
What is sigma delta modulators?Sigma delta modulators are devices implementing sigma delta modulations and are widely used in analogue-to-digital conversions.The input signals are first oversampled to obtain the inputs of the sigma delta modulators. The loop filters are to separate the quantization noises and the input signals so that very high signal-to-noise ratios could be achieved at very coarse quantization schemes.
Sigma delta modulators
52
Sigma delta modulatorsBlock diagram
( )( )
( )( )zFzF
zUzS
+=
1
F(z) Qu(k) y(k) s(k)
( )( ) ( )zFzNzS
+=
11
53
( ) ( ) ( )[ ]Tkykyk 12 −−≡x
( ) ( ) ( )( ) ( )( )( ) ( ) ( )( )( )2211cos2
21cos2−−−−−−−
+−−−=⇒kyQkykyQku
kykykyθ
θ
( ) 21
21
cos21cos2F −−
−−
+−−
=zz
zzzθ
θ
( ) ( ) ( )( ) ( )( )( ) ( ) ( )( )( )2211cos2
21cos2−−−−−−−=
−+−−⇒kyQkykyQku
kykykyθ
θ
( ) ( ) ( )[ ]Tkukuk 12 −−≡u
( ) ( )( ) ( )( )[ ]TkyQkyQk 12 −−≡s
Sigma delta modulatorsNonlinear state space dynamical model:
54
There are only finite numbers of possibilities of s(k). Hence, s(k) can be viewed as symbol and s(k) is called a symbolic sequence.
⎥⎦
⎤⎢⎣
⎡−
≡θcos21
10A ⎥
⎦
⎤⎢⎣
⎡−
≡θcos21
00B
( ) ( ) ( ) ( )( )kkkk suBAxx −+=+1
( )⎩⎨⎧−
≥≡
otherwise101 y
yQ
Sigma delta modulators
55-1 -0.5 0 0.5 1 1.5 2
-1
-0.5
0
0.5
1
1.5
2
x1
x 2
Second order marginally stable bandpasssigma delta modulators
Sigma delta modulators
56
Sigma delta modulators
57
Second order strictly stable bandpass sigma delta modulators
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x 104
-1
-0.5
0
0.5
1
1.5
2
Clock cycle
x 1(k)
Sigma delta modulators
58
Sigma delta modulators
59
Sigma delta modulators
60
Sigma delta modulatorsGeneral second order bandpass sigma delta modulators
61
Admissibility of periodic sequence
⎥⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢⎢
⎣
⎡
≡
−−
−−
−
−−
102
21
110
0121
MM
MM
M
MM
DDDDD
DDDDDDD
D
LL
OM
MOOOM
O
L
( ) ( ) ( ) ( )[ ]TMsMsss 1,1,0,0 2121 −−≡ Lζ
( ) ( ) ⎥⎦
⎤⎢⎣
⎡++
≡θθ
θθ1sin1sin
sinsinjcjd
jcjdjD
Sigma delta modulators
62
Σ be the admissible set of periodic output sequences with period M
( ) ( )( )11 ,, −−−−≡ MMdiag AIAIK L
[ ]Tυυτ ,, L≡
( ) ( ) ( )( )1,,1,0 −≡ Mssss L
( ) ( )⎥⎦
⎤⎢⎣
⎡ ++≡ ∑∑
=
−
=
M
j
M
jjucdjucd
1
1
0sin
sin,sin
sinθ
θθ
θυ
( )( ){ }ζτDζKQs =+=Σ :
Sigma delta modulators
63
What is a perceptron training algorithm?A perceptron is a single neuron that employs a single bit quantization function as its activation function.
where
Perceptron training algorithms
( )kx1
( )kxd
( )kw1
( )kwd
M
( )kw0
Q ( )ky
( ) ( ) ( )( )kkQky T xw=
( )⎩⎨⎧
<−≥
≡0101
zz
zQ
( ) ( ) ( )[ ]Td kxkxk ,,,1 1 L≡x( ) ( ) ( ) ( )[ ]Td kwkwkwk ,,, 10 L≡w
64
What is a perceptron training algorithm?The weights of a perceptron is usually found by the perceptron training algorithm.
where t(k) is the desired output corresponding to x(k).If w(k) converges, then the steady state value of w(k) could be employed as the weights of the perceptron.
( ) ( ) ( ) ( ) ( )kkyktkk xww2
1 −+=+
Perceptron training algorithms
65
Example 1Iteration index k wold(k) x(k) t(k) y(k) wnew(k)
0 [0 0 0]T [1 5 1]T -1 1 [-1 -5 -1]T
1 [-1 -5 -1]T [1 2 1]T -1 -1 [-1 -5 -1]T
2 [-1 -5 -1]T [1 1 1]T 1 -1 [0 -4 0]T
3 [0 -4 0]T [1 3 3]T 1 -1 [1 -1 3]T
4 [1 -1 3]T [1 4 2]T -1 1 [0 -5 1]T
5 [0 -5 1]T [1 2 3]T 1 -1 [1 -3 4]T
6 [1 -3 4]T [1 5 1]T -1 -1 [1 -3 4]T
7 [1 -3 4]T [1 2 1]T -1 -1 [1 -3 4]T
8 [1 -3 4]T [1 1 1]T 1 1 [1 -3 4]T
9 [1 -3 4]T [1 3 3]T 1 1 [1 -3 4]T
10 [1 -3 4]T [1 4 2]T -1 -1 [1 -3 4]T
11 [1 -3 4]T [1 2 3]T 1 1 [1 -3 4]T
Converge
( ) ( ) ( ) ( ) ( )kkyktkk xww2
1 −+=+
Perceptron training algorithms
66
Example 1
0 2 4 60
1
2
3
4
x1
x 2
LinearlySeparable
Perceptron training algorithms
67
Example 2Iteration index k wold(k) x(k) t(k) y(k) wnew(k)
0 [0 0 0]T [1 0 0]T -1 1 [-1 0 0]T
1 [-1 0 0]T [1 0 1]T 1 -1 [0 0 1]T
2 [0 0 1]T [1 1 0]T 1 1 [0 0 1]T
3 [0 0 1]T [1 1 1]T -1 1 [-1 -1 0]T
4 [-1 -1 0]T [1 0 0]T -1 -1 [-1 -1 0]T
5 [-1 -1 0]T [1 0 1]T 1 -1 [0 -1 1]T
6 [0 -1 1]T [1 1 0]T 1 -1 [1 0 1]T
7 [1 0 1]T [1 1 1]T -1 1 [0 -1 0]T
8 [0 -1 0]T [1 0 0]T -1 1 [-1 -1 0]T
9 [-1 -1 0]T [1 0 1]T 1 -1 [0 -1 1]T
10 [0 -1 1]T [1 1 0]T 1 -1 [1 0 1]T
11 [1 0 1]T [1 1 1]T -1 1 [0 -1 0]T
Limit Cycle
( ) ( ) ( ) ( ) ( )kkyktkk xww2
1 −+=+
Perceptron training algorithms
68
Example 2
- 1 0 1 2- 1
0
1
2
x 1
x 2
NonlinearlySeparable
Perceptron training algorithms
69
Perceptron training algorithms
-1 -0.5 0 0.5 1-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
x1
Class 1Class -1
0 10 20 30 40-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time index k
w0(k
)
0 10 20 30 401
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
2.8
3
Time index k
w1(k
)
0 10 20 30 40-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time index k
w2(k
)
Example 3
70
Perceptron training algorithmsExample 4
The set of training feature vectorsDesirable outputInitial weightResult:Set of weights of the perceptron:
Invariant set of weights of the perceptron:Invariant map:Note: is not bijective because
⎪⎭
⎪⎬
⎫
⎪⎩
⎪⎨
⎧
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡−
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
11
1,
11
1,
111
,111
{ }1,1,1,1 −−( ) [ ]T1,1,10 −−−=w
( ) [ ]Tk 0,2,0=w Zk∈∀
[ ]{ }T1,1,1 −−−=℘[ ]{ } [ ]{ }TTF 1,1,11,1,1: −−−→−−−ℑ
33:~ ℜ→ℜℑF ( ) ( ) ( )kkk T xwx >2
Zk∈∀
⎪⎭
⎪⎬
⎫
⎪⎩
⎪⎨
⎧
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡−
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡−−
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−−−
02
0,
111
,000
,111
71
Perceptron training algorithms
-1 -0.5 0 0.5 1-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
x1
x 2
Class 1Class -1
0 10 20 30 40-1
-0.9
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
Time index k
w0(k
)
0 10 20 30 40-2
-1.8
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
Time index k
w1(k
)
0 10 20 30 40-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Time index k
w2(k
)
72
Perceptron training algorithmsExample 5
The set of training feature vectorsDesirable outputInitial weightInvariant set of the weights of the perceptron consists of three hyperplanes.Note: is not bijective because
{ }1,1,1,1 −−
33:~ ℜ→ℜℑF
( ) ( ) ( )kkk T xwx >2Zk∈∃
⎪⎭
⎪⎬
⎫
⎪⎩
⎪⎨
⎧
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
−⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡
− 0668.11139.01
,1364.0
1832.21
,5883.0
7258.01
,1867.0
1746.01
( ) [ ]T2133.0,7923.0,10 −−=w
73
Perceptron training algorithms
0 1 2 3-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
x1
x 2
Class 1Class -1
-2-1.5
-1-0.5
0
-1
0
1
2
3
4-1.5
-1
-0.5
0
0.5
1
1.5
w0
w1
w2
74
Many digital signal processing systems are symbolic dynamical systems.These symbolic dynamical systems could exhibit fractal, chaotic and divergent behaviors.By investigating the properties of these symbolic dynamical systems, unwanted behaviors could be avoided.
Conclusions
75
Q&A Session
Thank you!
Let me think…
Bingo