Professional Documents
Culture Documents
jk ω0t
e
kω0t
3 Re
3.1 Introduction
x(t) P XX (f)
t
(a)
P XX (f)
x(t)
(b)
Figure 3.1 The concentration or spread of power in frequency indicates the correlated or
random character of a signal: (a) a predictable signal, (b) a random signal.
Sec. 3.2 Fourier Series 3
Notation: In this chapter the symbols t and m denote continuous and discrete
time variables, and f and k denote continuous and discrete frequency variables
respectively. The variable ω=2πf denotes the angular frequency in units of rad/s
and is used interchangably (within a scaling of factor of 2π) with the frequency
variable f in units of Hz.
The following three sinusoidal functions form the basis functions for the Fourier
analysis
x1 (t ) = cosω 0 t (3.1)
x 2 (t ) = sinω 0 t (3.2)
(3.3)
x3 (t ) = cosω 0 t + j sinω 0 t = e jω0t
Fig. 3.2.a shows the cosine and the sine components of the complex exponential
(cisoidal) signal of Eq. (3.3), and Fig. 3.2.b shows a vector representation of the
complex exponential in a complex plane with real (Re) and imaginary (Im)
dimensions. The Fourier basis functions are periodic with an angular frequency of
ω0 rad/s and a period of T0=2π/ω0=1/F0 seconds, where F0 is the frequency in
Hz. The following properties make the sinusoids the ideal choice as the
elementary building block basis functions for signal analysis and synthesis:
Im
sin(kωot) cos(kωot)
jk ω0t
e
kω0t
Re
T0
(a) (b)
Figure 3.2 - Fourier basis functions: (a) real and imaginary parts of a complex sinusoid,
(b) vector representation of a complex exponential.
4 Chap.3 Fourier Analysis and Synthesis
∞ ∞ ∞
1 1
∫−∞sin(ω1t ) sin(ω 2 t ) dt = − 2 −∫∞cos(ω1 + ω 2 )t dt + 2 −∫∞cos(ω1 − ω 2 )t dt = 0
(3.4)
For harmonically related sinusoids the integration can be taken over one period.
Similar equations can be derived for the product of cosines, or sine and cosine,
of different frequencies. Orthogonality implies that the sinusoidal basis
functions are independent and can be processed independently. For example in
a graphic equaliser we can change the relative amplitudes of one set of
frequencies, such as the bass, without affecting other frequencies, and in
subband coding different frequency bands are coded independently and
allocated different number of bits.
(ii) Sinusoidal functions are infinitely differentiable. This is important, as most
signal analysis, synthesis and manipulation methods require the signals to be
differentiable.
(iii) Sine and cosine signals of the same frequency have only a phase difference of
π/2 or equivalently a relative time delay of a quarter of one period i.e. T0/4.
[1, e ± jω 0 t
,e ± j2ω 0t ,e ± j3ω 0t , K ] (3.5)
The set of exponential signals in Eq. (3.5) are periodic with a fundamental
frequency ω0=2π/T0=2πF0 where T0 is the period and F0 is the fundamental
frequency. These signals form the set of basis functions for the Fourier analysis.
Any linear combination of these signals of the form
∑c e
k = −∞
k
jkω 0t
(3.6)
is also periodic with a period of T0. Conversely any periodic signal x(t) can be
synthesised from a linear combination of harmonically related exponentials.
The Fourier series representation of a periodic signal are given by the following
synthesis and analysis equations:
∞
x(t ) = ∑c e
k = −∞
k
jkω 0t
k = L − 1,0,1,L Synthesis equation (3.7)
Sec. 3.2 Fourier Series 5
T0 / 2
1
∫ x(t )e
− jkω 0t
ck = dt k = L − 1,0,1,L Analysis equation (3.8)
T0 −T0 / 2
Example 3.1 Given the Fourier synthesis Eq. (3.7), obtain the frequency analysis
Eq. (3.8).
Solution: Multiply both sides of Eq. (3.7) by e − jmω0t and integrate over one period
to obtain
T0 / 2 ∞ T0 / 2 ∞ T0 / 2
∫ x(t )e
− jmω 0t
dt = ∑ c ∫e
k = −∞
k
jkω 0t
e − jmω 0t
dt = ∑ c ∫e
k = −∞
k
j( k − m )ω 0t
dt (3.9)
−T0 / 2 −T0 / 2 −T0 / 2
From the orthogonality principle the integral in the r.h.s of Eq. (3.9) is zero unless
k=m in which case the integral is equal to T0. Hence
T0 / 2
1
∫ x(t )e
− jmω 0t
cm = dt (3.10)
T0 −T0 / 2
6 Chap.3 Fourier Analysis and Synthesis
x(t) | ck |
t -1 1 k
-103 103 f H
T0=1 ms
(a) (b)
Example 3.2 Find the frequency spectrum of a 1 kHz sinewave shown in Fig
3.3.a
x(t ) = sin( 2000πt ) − ∞<t <∞ (3.11)
1 1
c1 = , c −1 = − and c k ≠ ±1 = 0 ( 3.14)
2j 2j
Fig 3.3.b shows the magnitude spectrum of the sinewave, where the spectral lines
c1 and c−1 correspond to the 1 kHz and −1 kHz frequencies respectively.
1 j2000πt 1 − j2000πt
Solution B: Substituting sin( 2000πt ) = e − e in the Fourier
2j 2j
analysis Eq. (3.8) yields
T0 / 2
1 1 j 2000πt 1 − j2000πt − jk 2000πt
ck =
T0 ∫ e
−T0 / 2
2j
− e
2j
e
dt
T0 / 2 T0 / 2
1 1
∫ e j (1− k ) 2000πt dt − ∫e
− j(1+ k ) 2000πt (3.15)
= dt
2 jT0 −T0 / 2
2 jT0 −T0 / 2
Sec. 3.2 Fourier Series 7
Since sine and cosine functions are positive-valued over one half a period and
negative-valued over the other half, it follows that Eq. (3.15) is zero unless k=1 or
k=−1.
1 1
c1 = and c −1 = − and c k ≠ ±1 = 0 (3.16)
2j 2j
Example 3.3 Find the frequency spectrum of a periodic train of pulses with
amplitude of 1.0, a period of 1.o kHz and a pulse 'on' duration of 0.3 milliseconds.
x(t)
t
c(k)
0 .3
0 .2 5
0 .2
0 .1 5
0 .1
0 .0 5
0
1 2
-0 .0 5
k
-0 .1
Figure 3.4 - A rectangular pulse train and its discrete frequency ‘line’ spectrum.
8 Chap.3 Fourier Analysis and Synthesis
Example 3.4 For the example 3.3 write the formula for synthesising the signal up
to the Nth harmonic, and plot a few examples for the increasing number of
harmonics.
Solution: The equation for the synthesis of a signal upto the Nth harmonic content
is given by
N N N
x(t ) = ∑c
k =− N
ke
jkω 0 t
= c0 + ∑c e
k =1
k
jkω 0 t
+ ∑c
k =1
−k e
− jkω 0 t
N
= c0 + ∑ [Re(c
k =1
k ) + j Im(c k )][cos(kω 0 t ) + j sin( kω 0 t )]
N
+ ∑ [Re(c
k =1
k ) − j Im(c k )][cos(kω 0 t ) − j sin(kω 0 t )]
N
= c0 + ∑ [2 Re(c
k =1
k ) cos(kω 0 t ) − 2 Im(c k ) sin( kω 0 t )] (3.19)
x(t) x(t)
(a) (b)
t t
x(t) x(t)
(c) (d)
t t
Figure 3.5 Illustration of the Fourier synthesis of a periodic pulse train, and the Gibbs
phenomenon, with the increasing number of harmonics in the Fourier synthesis:
(a) N=1, (b) N=3, (c) N=6, and (d) N=100.
Sec. 3.3 Fourier Transform 9
The sinusoidal basis functions of the Fourier transform are smooth and infinitely
differentiable. In the vicinity of a discontinuity the Fourier synthesis of a signal
exhibits ripples as shown in the Fig 3.5. The peak amplitude of the ripples does
not decrease as the number of harmonics used in the signal synthesis increases.
This behaviour is known as the Gibbs phenomenon. For a discontinuity of unity
height, the partial sum of the harmonics exhibits a maximum value of 1.09 (that is
an overshoot of 9%) irrespective of the number of harmonics used in the Fourier
series. As the number of harmonics used in the signal synthesis increases, the
ripples become compressed toward the discontinuity but the peak amplitude of the
ripples remains constant.
c(k)
x(t)
Ton
Toff t 1
(a) T0 k
T0=Ton+Toff
X(f)
x(t)
Toff = ∞
(b) t
f
Figure 3.6 – (a) A periodic pulse train and its line spectrum, (b) a single pulse from the
periodic train in (a) with an imagined ‘off’ duration of infinity; its spectrum is the envelope
of the spectrum of the periodic signal in (a).
10 Chap.3 Fourier Analysis and Synthesis
∫ X ( f )e
j2πft
x(t ) = df (3.20)
−∞
∞
−∞
∫
X ( f ) = x(t )e − j2πft dt (3.21)
∫
2
x(t ) dt < ∞ (3.22)
−∞
Solution: Consider the Fourier analysis Eq. (3.8) for a periodic signal and Eq.
(3.21) for its non-periodic version (consisting of one period only). Comparing
these equations reproduced below
T0 / 2 ∞
1
ck =
T0 ∫ x(t )e − j2πkF0 t dt
− T0 / 2
∫
X ( f ) = x(t )e − j2πft dt
−∞
we have
Sec. 3.3 Fourier Transform 11
1 (3.23)
ck = X (kF0 ) as T0 → ∞
T0
where F0=1/T0. Using Eq. (3.23) the Fourier synthesis Eq. (3.7) for a periodic
signal can be rewritten as
∞
x(t ) = ∑ X ( k / T )e
k = −∞
0
j2πkF0 t
∆F (3.24)
∫ X ( f )e
j2πft
x(t ) ⇒ df (3.25)
T0 →∞
−∞
Consider the unit-area pulse p(t) shown in Fig 3.7.a. As the pulse width ∆ tends to
zero the pulse tends to an impulse. The impulse function shown in Fig 3.7.b is
defined as a pulse with an infinitesimal time width as
1 / ∆ t ≤ ∆ / 2
δ (t ) = limit p (t ) = (3.26)
∆→0 0 t > ∆ / 2
∆ t t f
(a) (b) (c)
Figure 3.7 (a) A unit-area pulse, (b) the pulse becomes an impulse as ∆ → 0 , (c) the
spectrum of the impulse function.
12 Chap.3 Fourier Analysis and Synthesis
∞
1
∫ δ (t )dt = ∆ × ∆ =1
−∞
(3.27)
∫ x(t )δ (t − T ) dt = x(T )
−∞
(3.28)
∫
∆ ( f ) = δ (t )e − j 2πft dt = e 0 = 1
−∞
(3.29)
The impulse function is used as a test function to obtain the impulse response of a
system. This is because as shown in Fig 3.7.c an impulse is a spectrally rich signal
containing all frequencies in equal amounts.
Example 3.7 Find the spectrum of the 10 kHz periodic pulse train shown in Fig.
3.8.a with an amplitude of 1 mV and expressed as
∞ ∞
x(t ) = ∑
m = −∞
A× limit p(t − mT0 ) × ∆ =
∆ →0
∑ A δ (t − mT )
m = −∞
0 (3.30)
where p(t) is a unit-area pulse with width ∆ as shown in Fig. 3.7.a, and the
function δ(t-mT0) is now assumed to have a unit amplitude representing the area
under the impulse.
Solution: T0=1/F0=0.1 ms and A=1 mV. For the time interval −T0/2<t<T0/2
x(t)=Aδ(t), hence we have
ck
x(t )
... ...
T0 t 1/T0 k
(a) (b)
Figure 3.8 A periodic impulse train x(t) and its Fourier series spectrum ck.
Sec. 3.3 Fourier Transform 13
T0 / 2 T0 / 2
1 1 A
∫ x(t )e ∫ Aδ (t )e
− jkω 0t − jkω 0t
ck = dt = dt = = 10 V (3.31)
T0 −T0 / 2
T0 −T0 / 2
T0
The rectangular pulse is particularly important in digital signal analysis and digital
communication. For example, the spectrum of a rectangular pulse can be used to
calculate the bandwidth required by pulse radar systems or communication
systems that transmit pulses. The Fourier transform of a rectangular pulse Fig.
3.9.a of duration T seconds is obtained as
∞ T /2
R ( f ) = ∫ r (t ) e − j2πft dt = ∫ 1.e
− j2πft
dt
−∞ −T / 2
(3.32)
e j2πfT / 2 − e − j2πfT / 2 sin(πfT )
= =T = T sinc(πfT )
j2πf πfT
Fig. 3.9.b shows the spectrum of the rectangular pulse. Note that most of the pulse
energy is concentrated in the main lobe within a bandwidth of 2/T. However there
are pulse energy in the side lobes that may interfere with other electronic devices
operating at the side lobe frequencies.
Matlab code for drawing the spectrum of a rectangular pulse.
R(f)
r(t)
T
-2 -1 1 2 3
t T T T T T
f
T
(a) (b)
Figure 3.9 A rectangular pulse and its spectrum.
14 Chap.3 Fourier Analysis and Synthesis
m m m
σ>0 σ =0 σ <0
Figure 3.10 – The Laplace basis functions.
3.3.1 The Relation Between the Laplace and the Fourier Transforms
∞
X ( s ) = ∫ x(t )e − st dt (3.33)
0−
where the complex variable s=σ+jω, and the lower limit of t=0− allows the
possibility that the signal x(t) may include an impulse. The inverse Laplace
transform is defined by
σ 1 + j∞
∫ X (s)e
st
x(t ) = ds (3.34)
σ1 − j∞
where σ1 is selected so that X(s) is analytic (no singularities) for s>σ1. The basis
functions for the Laplace transform are damped or growing sinusoids of the form
e − st = e −σt e − jωt as shown in Fig. 3.10. These are particularly suitable for transient
signal analysis. The Fourier basis functions are steady complex exponential, e − jωt ,
of time-invariant amplitudes and phase, suitable for steady state or time-invariant
signal analysis.
The Laplace transform is a one-sided transform with the lower limit of integration
at t = 0 − , whereas the Fourier transform Eq. (3.21) is a two-sided transform with
the lower limit of integration at t = −∞ . However for a one-sided signal, which is
zero-valued for t < 0 − , the limits of integration for the Laplace and the Fourier
transforms are identical. In that case if the variable s in the Laplace transform is
replaced with the frequency variable jω then the Laplace integral becomes the
Fourier integral. Hence for a one-sided signal, the Fourier transform is a special
case of the Laplace transform corresponding to s=jω and σ=0. The relation
between the Fourier and the Laplace transforms are discussed further in Chapter 4.
3.3.2 Properties of the Fourier Transform
Sec. 3.3 Fourier Transform 15
There are a number of Fourier Transform properties that provide further insight
into the transform and are useful in reducing the complexity of the solutions of
Fourier transforms and inverse transforms. These are:
Linearity
The Fourier transform is a linear operation, this mean the principle of
superposition applies. Hence
if:
z (t ) = ax(t )+by (t ) (3.35)
then
Z ( f ) = aX ( f )+bY ( f ) (3.36)
Symmetry
This property states that if the time domain signal x(t) is real (as is often the case
in practice) then
X ( f ) = X * (− f ) (3.37)
Where the superscript asterisk * denotes the complex conjugate operation. From
Eq. (3.37) it follows that Re{ X ( f ) } is an even function of f and Im{ X ( f ) } is an
odd function of f. Similarly the magnitude of X ( f ) is an even function and the
phase angle is an odd function.
Note that the modulation Eq. (3.39) states that multiplying a signal x(t) by e j 2πF0t
translates the spectrum of x(t) onto the frequency F0, this is the frequency
modulation (FM) principle.
∫ X ( f )e
j2πft
x(t ) = df (3.40)
−∞
16 Chap.3 Fourier Analysis and Synthesis
Then by differentiating both sides of the Fourier transform Eq. (3.40) we obtain
∞
d x(t )
dt ∫
= j2πf X ( f ) e j2πft df
−∞
14243
(3.41)
FourierTransform
of d x ( t ) / dt
t
F 1
∫ x(τ )dτ ←→ j2πf X ( f )+πX (0)δ ( f )
−∞
(3.42)
Where the impulse term on the right-hand side reflects the dc or average value that
can result from the integration.
F 1 f
x(αt ) ←→ X (3.43)
α α
For example try to say something very slowly, then α >1, your voice spectrum
will be compressed and you may sound like a slowed down tape or disc, you can
do the reverse and the spectrum would be expanded and your voice shifts to higher
frequencies, This property is further illustrated in section 3.9.4.
Convolution
The convolution integral of two signals x(t) and h(t) is defined as
∫
y (t ) = x(τ ) h(t − τ )dτ
−∞
(3.44)
∞
∞ ∞ ∞
x(τ )h(t − τ )dτ e − j2πft dt = x(τ ) e − j2πfτ dτ h(t − τ ) e − j2πf (t −τ ) dt
∫ ∫
−∞ −∞
−∞
∫ −∞
∫
(3.46)
= X ( f )H ( f )
Duality
Comparing the Fourier transform and the inverse Fourier transform relations we
observe a symmetric relation between them. In fact the main difference between
Eqs. (3.20) and (3.21) is a negative sign in the exponent of e − j2πft in Eq (3.21).
This symmetry leads to a property of Fourier transform known as the duality
principle and stated as
F
x(t ) ←→ X ( f )
F (3.47)
X (t ) ←→ x( f )
X (f)
x(t)
T
-2 -1 1 2 3
T T T T T
f
T t
X (t) x(f)
-1 1
t T1 T1 f
T1
∞ ∞
Energy = ∫
−∞ −∞
∫
| x(t ) | 2 dt = | X ( f ) | 2 df (3.48)
Find and sketch the frequency spectrum of the following finite duration signal
Solution: Substitute for x(t) and its non-zero valued limits in the Fourier transform
Eq. (3.21)
NT0 / 2
∫ sin(2πF t )e
− j 2πft
X(f ) = 0 dt
− NT0 / 2
(3.50)
NT0 / 2
e j2πF0t − e − j2πF0t − j2πft
X(f ) = ∫
− NT0 /2
2 j
e dt
NT0 / 2 NT0 / 2
e − j2π ( f − F0 )t e − j2π ( f + F0 )t
= ∫
− NT0 /2
2j
dt −
− NT
∫ /2
2j
dt
0 (3.51)
X(f)
Nc=4
Nc=2
Nc=1
F0
f
Nc=3
Figure 3.12 The spectrum of a finite duration sine wave with the increasing length of
observation. Nc is the number of cycles in the observation window. Note that the
width of the main lobe depends inversely on the duration of the signal.
The signal spectrum can be expressed as the sum of two shifted sinc functions.
Note that energy of a finite duration sinewave is spread in frequency between the
main lobe and the side lobes of the sinc function. Fig. 3.12 demonstrates that as
the window length increases the energy becomes more concentrated in the main
lobe and in the limit for an infinite duration window the spectrum tends to an
impulse positioned at the frequency F0.
20 Chap.3 Fourier Analysis and Synthesis
N −1
x(t ) = ∑ A(m) p (t − mTb ) ()
m =0
1 t ≤ Tb / 2
p (t ) = ()
0 t >0
N −1 N −1
X ( f ) = ∑ A(m) P( f )e − j 2πmTb f = P( f ) ∑ A(m)e − j 2πmTb f ()
m=0 m=0
∞ ∞
E [X ( f ) X * ( f )] = E ∑ A(m) P( f )e − j 2πmTb f ∑ A(n) P * ( f )e j 2πmTb f
m = −∞ n = −∞ ()
N −1 N −1
∑ ∑ E[ A(m) A(n)]e
2 − j 2π ( m − n )Tb f
= P( f )
m =0 n =0
1 m=n
E[ A(m) A(n)] = ()
0 m≠n
E [ X ( f ) X * ( f )] = N P ( f )
2
()
∞
x ( m) = ∑ x(t )δ (t − mT ) s
m = −∞ (3.54)
where m is the discrete time variable and Ts is the sampling period. The Fourier
transform of x(m), a sampled version of a continuous signal x(t), can be obtained
from Eq. (3.21) as
∞ ∞ ∞ ∞
X ( f )= ∫ ∑ x(t )δ (t − m)e − j 2πft dt = ∑ ∫ x(t )δ (t − mT )e s
− j 2πft
dt
−∞ m = −∞ m = −∞ −∞
∞
= ∑ x(mT )e
m = −∞
s
− j 2πmf / Fs
(3.55)
For convenience of notation and without loss of generality it is often assumed that
sampling frequency Fs=1/Ts=1, hence
∞
Xs( f ) = ∑ x(m)e − j 2πmf
m = −∞ (3.56)
1/ 2
x ( m) = ∫X
−1 / 2
s ( f ) e j2πfm df (3.57)
Note that x(m) and X s ( f ) are equivalent in that they contain the same
information in different domains. In particular, as expressed by the Parseval's
theorem, the energy of the signal may be computed either in the time or in the
frequency domain a
∞ 1/ 2
Signal Energy = ∑ x 2 (m) = ∫ | X s ( f ) |2 df
m = −∞
(3.58)
−1 / 2
Example 3.10 Show that the spectrum of a sampled signal is periodic with a
period equal to the sampling frequency Fs.
22 Chap.3 Fourier Analysis and Synthesis
∞ ( f + kFs ) ∞ f kFs
− j2πm − j2πm − j2πm
X ( f + kFs ) = ∑ x(mT )e
m = −∞
s
Fs
= ∑ x(mT )e
m = −∞
s
Fs
e1 3= X(f )
424
Fs
=1 (3.59)
X(f) Xs(f)
... ...
0 -2Fs -Fs 0 Fs 2Fs
f f
(a) (b)
Figure 3.13 The spectrum of : (a) a continuous signal, and (b) its sampled version.
x(0) X(0)
x(1) Discrete Fourier X(1)
x(2) Transform X(2)
. N–1 2 π kn .
–j
.
. X(k) = ∑ x(m) e N
.
.
x(N–2) m=0 X(N – 2)
x(N – 1) X(N– 1)
N −1 2π
−j mk
X ( k ) = ∑ x ( m) e N k = 0, . . ., N−1 (3.60)
m=0
Comparing Eqs. (3.60) and (3.56) we see that the DFT consists of N equi-spaced
samples taken from one period (2π) of the continuous spectrum of the discrete
time signal x(m). The inverse discrete Fourier transform (IDFT) is given by
2π
1 N −1 j mk
x ( m) = ∑
N k =0
X (k ) e N m= 0, . . ., N−1 (3.61)
N −1 2π
−j km
X ( k ) = ∑ x ( m) e N k = 0, . . ., N−1
m =0
Multiply both sides of the DFT equation by e-j2πkn/N and take the summation
as
24 Chap.3 Fourier Analysis and Synthesis
N −1 2π N −1 N −1 2π 2π
+j kn −j km +j kn
∑ X (k ) e
k =0
N
= ∑∑ x(m) e
k =0 m=0
N
e N
N −1 N −1 2π
−j k ( m−n)
= ∑ x ( m) ∑ e N
k =0 m=0
14 4244
3
N if m = n
0 otherwise
1 1 F
∆f = = = s (3.62)
∆T NTs N
Note that the frequency resolution ∆f and the time resolution ∆T are inversely
proportional in that they can not both be simultanously increased, in fact ∆T∆f=1.
This is known as the uncertainty principle.
Example 3.11 A DFT is used in a DSP system for the analysis of an analog
signal with a frequency content of up to 10 kHz. Calculate: (i) the minimum
sampling rate Fs required, and (ii) the number of samples required for the DFT to
achieve a frequency resolution of 10 Hz at the minimum sampling rate.
Sec. 3.5 Short-Time Fourier Transform 25
Solution:
(i) Sampling rate > 2×10 kHz, say 22 kHz, and
Fs 22000
(ii) ∆f = 10 = N ≥ 2200 .
N N
Solution:
In the following program the two sinwaves have frequencies of 100 Hz and 110
Hz, and the sampling rate is 1 kHz. We experiment with two time windows of
length N1=1024 with a theoretical frequency resolution of ∆f=1000/1024=0.98 Hz,
and N2=64 with a theoretical frequency resolution ∆f=1000/64=15.7 Hz.
2
Amplitude
- 1
- 2
0 2 0 0 4 0 0 6 0 0 8 0 0 1 0 0 0
T im e
(a)
80 450
Amplitude
Amplitude
60
300
40
150
20
(b) (c)
Figure 3.15 Illustration of time and frequency resolutions: (a) sum of two sinewaves with 10
Hz difference in their frequencies, (b) the spectrum of a segment of 64 samples from
demonstrating insufficient frequency resolution to separate the sinewaves, (c) the spectrum of
a segment of 1024 samples has sufficient resolution to show the two sinewaves.
Y1=abs(fft(y(1:N1)));Y2=abs(fft(y(1:N2)));
figure(1); plot(y); figure(2); plot(Y1(1:N1/2)); figure(3); plot(Y2(1:N2/2));
In a practical situation we have either with a short length signal, or with a long
signal of which the DFT can only handle one segment at a time. Having a short
segment of N samples of a signal or taking a slice of N samples from a signal is
equivalent to multiplying the signal by a unit-amplitude rectangular pulse window
of N samples. Therefore an N-sample segment of a signal x(m) is equivalent to
1 0≤m ≤ N − 1
w(m)= (3.64)
0 otherwise
1 0 ≤ m ≤ N − 1
w(m)= (3.66)
0 otherwise
Solution: Taking the DFT of w(m), and using the convergence formula for the
partial sum of a geomertic series, described in the appendix A, we have
2π ( N −1)
N −1 −j mk 1 − e − j2πk −j πk sin(πk )
W (k ) = ∑
m =0
w(m)e N =
−j
2π
k
=e N
sin(πk / N )
(3.67)
1− e N
Note that for the integer values of k, w(k) is zero except for k=0.
N −1 2π N −1 − j2π ( F − k ) m
−j mk
∑e − j2πF0 m
∑
0
X (k ) = e N = e N
m =0 m =0
(3.68)
( N −1)
1 − e − j2π ( NF0 − k ) −j π ( NF0 − k ) sin(π ( NF0 − k ))
= − j2π ( NF0 − k ) / N
=e N
1− e sin(π ( NF0 − k ) / N )
Note that for integer values of k, X(k) is zero at all samples but one k=0.
In DFT the input signal is assumed to be periodic, with a period equal to the length
of the observation window of N samples. Now, for a sinusoidal input signal if
there is an integer number of cycles within the observation window, as in Fig.
3.16.a, then the assumed periodic waveform is the same as an infinite length pure
sinusoid. But if the observation window contains a non-integer number of cycles
of a sinusoid then the assumed periodic waveform will not be a pure sine wave and
will have end-point discontinuities. The spectrum of the signal then differs from
the spectrum of sinewave as illustrated in Fig. 3.16.b. The overall effects of finite
length window and end-point discontinuities are:
time time
F0 Frequency F0 Frequency
(a) (b)
Figure 3.16 The DFT spectrum of exp(j2πfm) : (a) an integer number of cycles with in the
N-sample analysis window, (b) a non-integer number of cycles in the window.
1. The spectral energy which could have been concentrated at a single point, or
in a narrow band of frequencies is spread over a larger band of frequencies.
28 Chap.3 Fourier Analysis and Synthesis
The end-point problems may be alleviated using a window that gently drops to
zero. One such window is a raised cosine window of the form
2πm
α − (1 − α ) cos 0≤m≤ N − 1
w(m) = N (3.69)
0 otherwise
For α=0.5 we have the Hanning window also known as the raised cosine window
2πm
wHan (m) = 0.5−0.5cos 0≤m≤ N − 1 (3.70)
N
For α=0.54 we have the Hamming window
2πm
wHam (m) = 0.54 − 0.46 cos 0 ≤ m ≤ N −1 (3.71)
N
-5
-1 0
-1 5
-2 0
-2 5
-3 0
0 T im e 6 4 0 F re q u en c y F s /2
0
-1 0
-2 0
-3 0
-4 0
-5 0
-6 0
0 6 4 0 F s /2
0
-1 0
-2 0
-3 0
-4 0
-5 0
-6 0
-7 0
-8 0
0 6 4 0 F s /2
Figure 3.17 (a) Rectangular window frequency response, (b) Hamming window frequency
response, (c) Hanning window frequency response.
2 N −1 2π
−j mk
X (k ) = ∑ x ( m) e
m=0
2N
k = 0, . . ., 2N−1 (3.72)
N −1 π
−j mk
= ∑ x ( m) e
m =0
N
r(m) |R(f)|
Time Frequency
|R(f)| |R(f)|
Frequency Frequency
Figure 3.18 Illustration of the interpolating effect, in the frequency domain, of zero
padding a signal in the time domain.
30 Chap.3 Fourier Analysis and Synthesis
time
In Fourier transform it is assumed that the signal is stationary, meaning the signal
statistics, such as the mean, the power, and the power spectrum, are time-invariant.
Most real life signals such as speech, music and image signals are nonstationary in
that their amplitude, power, spectral composition and other features changes
continuously with time. To apply Fourier transform to nonstationary signals the
signal is divided into appropriately short-time windows, such that within each
window the signal can be assumed to be time-invariant. The Fourier transform
applied to the short signal segment within each window is known as the short-time
Fourier Transform. Fig. 3.19 illustrates the segmentation of a speech signal into a
sequence of overlapping, hamming windowed, short segments. The choice of
window length is a compromise between the time resolution and the frequency
resolution. For Audio signals a time window of about 25 ms, corresponding to a
frequency resolution of 40 Hz is normally adopted.
walker.qxp 3/24/98 2:16 PM Page 658
Fourier Analysis
and Wavelet
Analysis
James S. Walker
I
n this article we will compare the classical Frequency Information, Denoising
methods of Fourier analysis with the newer As an example of the importance of frequency in-
methods of wavelet analysis. Given a sig- formation, we will examine how Fourier analysis can
nal, say a sound or an image, Fourier analy- be used for removing noise from signals. Consider
sis easily calculates the frequencies and a signal f (x) defined over the unit interval (where here
the amplitudes of those frequencies which make x stands for time). The periodP1 Fourier series ex-
up the signal. This provides a broad overview of i2π nx , with
the characteristics of the signal, which is im- R 1 of f is defined by
pansion n∈Z cn e
cn = 0 f (x)e−i2π nx dx. Each Fourier coefficient, cn , is
portant for theoretical considerations. However,
an amplitude associated with the frequency n of the
although Fourier inversion is possible under
exponential ei2π nx. Although each of these expo-
certain circumstances, Fourier methods are not
nentials has a precise frequency, they all suffer from
always a good tool to recapture the signal, par-
a complete absence of time localization in that their
ticularly if it is highly nonsmooth: too much
magnitudes, |ei2π nx | , equal 1 for all time x .
Fourier information is needed to reconstruct
the signal locally. In these cases, wavelet analy- To see the importance of frequency information,
sis is often very effective because it provides a let us examine a problem in noise removal. In
simple approach for dealing with local aspects Figure 1(a)[top] we show the graph of the signal
of a signal. Wavelet analysis also provides us with
(1) f (x) =
new methods for removing noise from signals
2
that complement the classical methods of (5 cos 2π νx) [ e−640π (x−1/8)
Fourier analysis. These two methodologies are 2 2
+ e−640π (x−3/8) + e−640π (x−4/8)
major elements in a powerful set of tools for the-
2 2
oretical and applied analysis. + e−640π (x−6/8) + e−640π (x−7/8) ]
This article contains many graphs of discrete
signals. These graphs were created by the com- where the frequency, ν , of the cosine factor is 280 .
puter program FAWAV, A Fourier–Wavelet An- Such a signal might be used by a modem for trans-
alyzer, being developed by the author. mitting the bit sequence 1 0 1 1 0 1 1 . The Fourier co-
efficients for this signal are shown in Figure 1(b)[top].
The highest magnitude coefficients are concentrated
around the frequencies ±280. Suppose that when this
James S. Walker is professor of mathematics at the
University of Wisconsin-Eau Claire. His e-mail address signal is received, it is severely distorted by added
is walkerjs@uwec.edu. noise; see Figure 1(a)[middle]. Using Fourier analy-
The author would like to thank Hugo Rossi, Steven
sis, we can remove most of this noise. Computing the
Krantz, his colleague Marc Goulet, and two anony- noisy signal’s Fourier coefficients, we obtain the
mous reviewers for their helpful comments during the graph shown in Figure 1(b)[middle]. The original sig-
writing of this article. nal’s largest magnitude Fourier coefficients are clus-
Figure 1. (a)[top] Signal. (b)[top] Fourier coefficients of signal. (a)[middle] Signal after adding noise.
(b)[middle] Fourier coefficients of noisy signal and filter function. (b)[bottom] Fourier coefficients
after multiplication by filter function. (a)[bottom] Recovered signal.
tered around the frequency positions ±280. The for n = − 12 M + 1, . . . , 0, . . . , 12 M . The discrete
Fourier coefficients of the added noise are lo- Fourier coefficients f̂n can be calculated by a fast
calized around the origin, and they decrease in Fourier transform (FFT) algorithm and are the
magnitude until they are essentially zero near discrete analog of the Fourier coefficients cn
the frequencies ±280 . Thus, the original sig- for F , when fj = F(j/M) . Moreover, f̂n is just a
nal’s coefficients and the noise’s coefficients Riemann sum approximation of the integral that
are well separated. To remove the noise from the defines cn .1 The magnitudes of the discrete
signal, we multiply the noisy signal’s coefficients Fourier coefficients for this transient damp down
by a filter function, which is 1 where the signal’s to zero very slowly (their graph is a very wide
coefficients are concentrated and 0 where the bell-shaped curve with maximum at the origin).
noise’s coefficients are concentrated. We then re- Consequently, to represent the transient well, one
cover essentially all of the signal’s coefficients; must retain most if not all of these Fourier co-
see Figure 1(b)[bottom]. Performing a Fourier efficients. In Figure 2(a)[bottom] we show the re-
series partial sum with these recovered coeffi- sults obtained from trying to compress the tran-
cients, we obtain the denoised signal, which is sient
P104 by computing a discrete partial sum
shown in Figure 1(a)[bottom]. Clearly, the bit se- i2π nj using only one-fifth of the
n=−104 f̂n e
quence 1 0 1 1 0 1 1 can now be determined from Fourier coefficients.2 Clearly, even a moderate
the denoised signal, and the denoised signal is compression ratio of 5:1 is not effective.
a close match of the original signal. In the sec- Wavelets, however, are often very effective at
tion “Signal Denoising” we shall look at another representing transients. This is because they are
example of this method and also discuss how designed to capture information over a large
wavelets can be used for noise removal. range of scales. A wavelet series expansion of a
function f is defined by
Signal Compression X
As the example above shows, Fourier analysis is βnk2
n/2
ψ (2n x − k)
very effective in problems dealing with frequency n,k∈Z
location. However, it is often very ineffective at Z∞
with
representing functions. In particular, there are βn f (x) 2n/2 ψ (2n x − k) dx .
k =
severe problems with trying to analyze tran- −∞
sient signals using classical Fourier methods.
For example, in Figure 2(a)[top] we show a dis- The function ψ (x) is called the wavelet, and the
crete signal obtained from M = 1024 values coefficients βn
k are called the wavelet coeffi-
{fj = F(j/M)}M−1 of the function F(x) = cients. The function 2n/2 ψ (2n x − k) is the
j=0
5 2
e−10 π (x−.6) . For this example, we compute the 1For further discussion of discrete Fourier coefficients,
discrete Fourier series coefficients f̂n defined by see [2] or [11].
P512
P
−1 M−1
2The sum
n=−511 f̂n e
i2π nj , which uses all of the dis-
f̂n = M j=0 fj e−i2π nj/M crete Fourier coefficients, equals fj.
Figure 2. (a)[top] Signal. (a)[bottom] Fourier series using 205 coefficients, 5:1 compression. (b)[top]
Wavelet coefficients for signal. (b)[bottom] Wavelet series using only largest 4% in magnitude of
wavelet coefficients, 25:1 compression.
n
Vn = step functions in L2 (R), constant Note that these wavelets have period 1 . Fur-
o thermore, ψ̃n,k ≡ 0 for n < 0 , and
k k+1
on the intervals [ , ), k ∈ Z . ψ̃n,k+2n = ψ̃n,k for all k ∈ Z and n ≥ 0. On the in-
2n 2n
terval [0, 1), the periodic Haar wavelets ψ̃n,k sat-
This set of subspaces {Vn }n∈Z satisfies the fol-
lowing five axioms [6]: isfy ψ̃n,k (x) = 2n/2 ψ (2n x − k) for n ≥ 0 and
Axioms for a Multi-Resolution Analysis k = 0, 1, . . . , 2n − 1. So we have the following the-
orem as a consequence of Theorem 1.
(MRA)
Scaling: f (x) ∈ Vn if and only if f (2x) ∈ Vn+1. Theorem 2. The functions 1 and ψ̃n,k for n ≥ 0
and k = 0, 1, . . . , 2n − 1 are an orthonormal basis
Inclusion: Vn ⊂ Vn+1, for each n.
for L2 [0, 1) .
S
Density: closure Vn = L2 (R) .
T n∈Z
Remark. In the section “Daubechies Wavelets”
Maximality: Vn = {0}.
n∈Z we will make use of periodized scaling func-
Basis: ∃φ(x) such that {φ(x − k)}k∈Z is an tions, φ̃n,k, defined by
orthonormal basis for V0 .
To satisfy the basis axiom, we shall use the X
(4) φ̃n,k (x) = 2n/2 φ 2n (x + j) − k
Haar scaling function φ defined above. Then, by j∈Z
combining the scaling axiom with the basis
axiom, we find that {2n/2 φ(2n x − k)}k∈Z is an
for n ≥ 0 and k = 0, 1, . . . , 2n − 1.
orthonormal basis for Vn . But the totality of all
these orthonormal bases, consisting of the set
{2n/2 φ(2n x − k)}k,n∈Z, is not an orthonormal
Fast Haar Transform
basis for L2 (R) because the spaces Vn are not The relation between the Haar scaling function
mutually orthogonal. To remedy this difficulty, φ and wavelet ψ leads to a beautiful set of re-
we need what are called wavelet subspaces. De- lations between their coefficients as bases. Let
fine the wavelet subspace Wn to be the orthog- {αkn } and {βnk } be defined by
onal complement of Vn in Vn+1 . That is, Wn sat- Z∞
isfies the equation Vn+1 = Vn ⊕ Wn where ⊕
αkn = f (x) 2n/2 φ(2n x − k) dx,
denotes the sum of mutually orthogonal sub- −∞
(5) Z∞
spaces. From the density axiom and repeated ap-
plication ofLthe last equation, we obtain βn
k = f (x) 2n/2 ψ (2n x − k) dx.
−∞
L2 (R) = V0 ⊕ ∞ n=0 Wn . Decomposing
L V0 in a
similar way, we obtain L2 (R) = n∈Z Wn . Thus, Substituting 2n x in place of x in the identities
L2 (R) is an orthogonal sum of the wavelet sub- in (2), we obtain
spaces Wn.
Using (2) and the MRA axioms, it is easy to 1 n+1
2n/2 φ(2n x) = √ [2 2 φ(2n+1 x)]
prove the following lemma. 2
1 n+1
Lemma 1. The functions {ψ (x − k)}k∈Z are an + √ [2 2 φ(2n+1 x − 1)]
2
orthonormal basis for the subspace W0 .
It follows from the scaling axiom that 1 n+1
H = (H2 ⊕ IM−2 )
Successively applying this orthogonal matrix O ,
we obtain {αkn } and {βn · · · (PM/2 ⊕ IM/2 ) (HM/2 ⊕ IM/2 ) PM HM
k } starting from some
highest level coefficients {αkN } for some large where IN is the N × N identity matrix.
N . Because of the density axiom, N can be cho- These matrix multiplications can be per-
sen
P large enough to approximate f by formed rapidly on a computer. Multiplying by
N N/2
k∈Z αk 2 φ(2N x − k) in L2-norm as closely HM requires only O(M) operations, since HM
as desired. consists mostly of zeroes. Similarly, the per-
Let us now discretize these results. Suppose mutation PM requires O(M) operations. There-
that we are working with data {fj }M−1
j=0 associ- fore, the whole transformation requires
M−1 O(M) + O(M/2) + · · · + O(2) = O(M) operations.
ated with the time values {xj = j/M}j=0 on the
unit interval. If we shrink the Haar scaling func- The transformation H is called a fast Haar
tion φ(x) = 1[0,1) (x) enough, it covers only the transform. It should be noted that FFTs, which
first point, x0 = 0 . Consequently, by choosing a have revolutionized scientific practice during
large enough N , we may assume that our scal- the last thirty years, are O(M log M) algorithms.
ing coefficients {αkN } satisfy αkN = fk , for Since each Hk is an orthogonal matrix, and
k = 0, 1, . . . , M − 1 . Assuming that M is a power so is every permutation Pk , it follows that H is
of 2 , say M = 2R , it follows that N = R . invertible. Its inverse is
The next step involves expressing the coeffi-
H −1 = HM
T T T
PM (HM/2 T
⊕ IM/2 ) (PM/2 ⊕ IM/2 )
cient relations in (6) in a matrix form. Let A ⊕ B
stand for the orthogonal of the matrices A
sum · · · (H2T ⊕ IM−2 ).
A0
and B, that is, A ⊕ B = 0 B . Now let HM denote
Therefore, the inverse operation is also an O( M )
the M × M orthogonal matrix defined by
operation.
HM = O ⊕ O ⊕ · · · ⊕ O , where the orthogonal
matrix sums are applied M/2 times and O is the Discrete Haar Series
matrix defined in (7). Then, by applying the co-
The fast Haar transform can be used for com-
efficient relations in (6) and using the fact that
puting partial sums of the discretized version
fk = αkR, we obtain
of the following Haar wavelet series in L2 [0, 1) :
HM [f0 , f1 , . . . , fM−1 ]T
h ∞ 2X
X
n
−1
= α0R−1 , βR−10 , α1R−1 , βR−1
1 , (8) α00 + βn
k2
n/2
ψ (2n x − k).
T n=0 k=0
. . . , αR−1
1 , βR−11 .
2
M−1 2
M−1 Let us assume, as in the previous section, that
we have a discrete signal {fj }M−1
j=0 associated
To sort the coefficients properly into two groups, with the time values {xj = j/M}M−1
j=0 on the unit
we apply an M × M permutation matrix PM as interval. Substituting these time values into (8)
follows: and restricting the upper limit of n, we obtain
T
PM α0R−1 , βR−1
0 , . . . , α R−1
1 , βR−1
1 n
−1
2
M−1 2
M−1 X 2X
R−1
T fj = α00 CM + βn n/2
ψ (2n xj − k).
k CM 2
= α0R−1 , . . . , αR−1
1 , βR−1
0 , . . . , βR−1
1 . n=0 k=0
2
M−1 2
M−1
Figure 3. (a)[top] Haar series partial sum, 229 terms. (b)[top] Fourier series partial sum, 229 terms.
(a)[bottom] Haar series partial sum, 92 terms. (b)[bottom] Daub4 series partial sum, 22 terms.
sist of multiplying the data by H , then setting created by the energy method, which contains
some of the resulting coefficients equal to 0 , and 99.5% of the energy of this signal. This partial
then multiplying by H −1 . A widely used method sum, which used 229 coefficients out of a pos-
involves specifying a threshold. All coefficients sible 8192, provides an acceptable visual repre-
whose magnitudes lie below this threshold are sentation of the signal. In fact, the sum of the
set equal to 0 . This method is frequently used squares of the errors is 2.0 × 10−3 . By compar-
for noise removal, where coefficients whose ison a 229 coefficient Fourier series partial sum
magnitudes are significant only because of the suffers from serious drawbacks (see Figure
added noise will often lie below a well-chosen 3(b)[top]). The sum of the squares of the errors
threshold. We shall give an example of this in the is 4.5 × 10−1 , and there is severe oscillation and
section “Signal Denoising”. A second method a Gibbs’ effect near x = 0.5 . Although these lat-
keeps only the largest magnitude coefficients, ter two defects could be ameliorated using other
while setting the rest equal to 0 . This method summation methods ([11], Ch. 4), there would
is convenient for making comparisons when it still be a significant deviation from the original
is known in advance how many terms are needed. signal (especially near x = 0.5 ).
We used it in the compression example in the sec- This example illustrates how wavelet analy-
tion “Signal Compression”. A third method, sis homes right in on regions of high variability
which we shall call the energy method, involves of signals and that Fourier methods try to
specifying a fraction of the signal’s energy, where smooth them out. The size of a function’s Fourier
the energy is the square root of the sum of the coefficients is related to the frequency content
squares of the coefficients.3 We then retain the of the function, which is measured by integra-
least number of the largest magnitude coeffi- tion of the function against completely unlo-
cients whose energy exceeds this fraction of the calized basis functions. For a function having a
signal’s energy and set all other coefficients discontinuity, or some type of transient behav-
equal to 0 . The energy method is useful for the- ior, this produces Fourier coefficients that de-
oretical purposes: it is clearly helpful to be able crease in magnitude at a very slow rate. Conse-
to specify in advance what fraction of the sig- quently, a large number of Fourier coefficients
nal’s energy is contained in a partial sum. We are needed to accurately represent such sig-
shall use the energy method frequently below. nals.4 Wavelet series, however, use compactly
Let us look at an example. Suppose our sig- supported basis functions which, at increasing
nal is levels of resolution, have rapidly decreasing sup-
{fj = F(j/8192)}8191
j=0
ports and can zoom in on transient behavior. The
transient behaviors contribute to the magnitude
where F(x) = x1[0,.5) (x) + (x − 1)1[.5,1) (x) . In Fig- of only a small portion of the wavelet coefficients.
ure 3(a)[top] we show a Haar series partial sum,
4In recent years, though, significant improvements
3The Haar transform is orthogonal, so it makes sense
have been achieved using local cosine bases [3, 7,
to specify energy in this way. 9, 5, 1].
Figure 4. (a) Magnitudes of highest level coefficients for a function F: [top] Haar coefficients, [middle]
Daub4 coefficients; [bottom] Graph of F. (b) Sums of squares of all coefficients: [top] Haar coefficients,
[middle] Daub4 coefficients; [bottom] Graph of F. Vertical scales for highest level coefficients and sums
of squares are logarithmic.
Consequently, a small number of wavelet coeffi- Daub4R ∞wavelet is supported R ∞ on [0, 3) and satis-
cients are needed to accurately represent such sig- fies −∞ ψ (x) dx = 0 and −∞ xψ (x) dx = 0. Con-
nals. sequently, if the signal is constant or linear over
The Haar system performs well when the signal an interval [a, b) which contains
is constant over long stretches. This is because the [k2−n , 3(k + 1)2−n ) , then the wavelet coefficient
Haar wavelet is supported on R ∞[0, 1) and satisfies a βnk will equal 0. In Figure 4(a) we show graphs
0th order moment condition, −∞ ψ (x) dx = 0 . There- of the magnitudes of the highest level Haar co-
fore, if the signal {fj } is constant over an interval efficients and Daub4 coefficients. Each magni-
a ≤ xj < b such that [k2−n , (k + 1)2−n ) ⊂ [a, b) , then tude |β12k | is plotted at the x -coordinate k2
−12
n 12
for k = 0, 1, . . . , 2 − 1 . These graphs show that
the wavelet coefficient βk equals 0 . For example, sup-
the highest level Haar coefficients are near 0 over
pose {fj = F(j/8192)}8191
j=0 where the constant parts of F , while the highest level
F(x) = (8x − 1)1[.125,.25) (x) Daub4 coefficients are near 0 over the constant
and linear parts of F . In Figure 4(b) we show
+ 1[.25,.75) (x) + (7 − 8x)1[.75,.875) (x). graphs of the sums of the squares of all the co-
efficients, which show that almost all the Daub4
In Figure 3(a)[bottom] we show a Haar series partial coefficients are near 0 over the constant and lin-
sum which contains 99.5% of the energy of this sig- ear parts of F , while the Haar coefficients are
nal and uses only 92 coefficients out of a possible near 0 only over the constant parts of F . Fur-
8192. The fact that the signal is constant over three thermore, the largest magnitude Daub4 coeffi-
large subintervals of [0, 1) accounts for the excellent cients are concentrated around the locations of
compression in this example. In order to obtain the points of nondifferentiability of F . This kind
wavelet bases that provide considerably more com- of local analysis illustrates one of the powerful
pression, we need a compactly supported wavelet features of wavelet analysis.
ψ (x) which has more moments equal to zero. That Looking again at Figure 3(b)[bottom], we see
is, we want that the most serious defects of the Daub4 com-
Z∞ pressed signal are near the points where F is non-
differentiable. If, however, we consider the in-
(9) xj ψ (x) dx = 0, for j = 0, 1, . . . , L − 1
−∞ terval [0.4, 0.6] where F is constant, the Daub4
compressed signal values differ from the values
for an integer L ≥ 2 . We say that such a wavelet has of F by no more than 1.2 × 10−15 at all of the
its first L moments equal to zero. For example, a 1641 discrete values of x in this interval. In
Daub4 wavelet has its first 2 moments equal to zero. contrast, a Fourier series partial sum using 23
Using a Daub4 wavelet series for the signal above, it coefficients differs by more than 10−3 at 1441
is possible to capture 99.5% of the energy using only of these 1641 values of x . The Fourier series par-
22 coefficients! See Figure 3(b)[bottom]. This im- tial sum exhibits oscillations of amplitude
provement in compression is due to the fact that the 6.5 × 10−3 around the value 1 over this subin-
Figure 5. (a)[top] Signal. (b)[top] 37-term Daub4 approximation. (a)[bottom] 257-term Fourier cosine
series approximation. (b)[bottom] Highest level Daub4 wavelet coefficients of signal.
terval. Because the Fourier coefficients for F are The MRA axioms tell us that φ(x) must gener-
only O (n−2 ) , using just the first 23 coefficients ate a subspace V0 and that V0 ⊂ V1 . Therefore,
produces an oscillatory approximation to F over
all of [0, 1), including the subinterval [0.4, 0.6]. X p
(11) φ(x) = ck 2 φ(2x − k)
The highest magnitude wavelet coefficients,
k∈Z
however, are concentrated at the corner points
for F , and their terms affect only a small por- for some constants {ck } . A wavelet ψ (x) , for
tion of the partial sum (since their basis func- which {ψ (x − k)} spans the wavelet subspace
tions are compactly supported). Consequently, W0 , can be defined by5
the wavelet series provides an extremely close X p
approximation of F over the subinterval (12) ψ (x) = (−1)k c1−k 2 φ(2x − k).
[0.4, 0.6]. k∈Z
A major defect of the Haar wavelet is its dis-
continuity. For one thing, it is unsatisfying to use Equations (11) and (12) generalize the equations
discontinuous functions to approximate con- in (2) for the Haar case. The orthogonality of φ
tinuous ones. Even with discrete signals there can and ψ leads to the following equation
be undesirable jumps in Haar series partial sum
values (see Figure 3(a)[bottom]). Therefore, we X
want to have a wavelet that is continuous. In the (13) (−1)k c1−k ck = 0.
k∈Z
next section we will describe the Daubechies
wavelets, which have their first L ≥ 2 moments This equation, and the second equation in (14)
equal to zero and are continuous. below, imply the orthogonality of the matrices,
WN, used in the fast wavelet transform which we
Daubechies Wavelets shall discuss later in this section.
It is possible to generalize the construction of Combining (11) with the first two integrals in
the Haar wavelet so as to obtain a continuous (10), it follows that
scaling function φ(x) and a continuous wavelet
ψ (x) . Moreover, Daubechies has shown how to X p X
make them compactly supported. We will briefly (14) ck = 2, |ck |2 = 1.
k∈Z k∈Z
sketch the main ideas; more details can be found
in [4, 5, 8, 10]. Similarly, assuming that L = 2 , the equations in
Generalizing from the case of the Haar (9) combined with (12) imply
wavelets, we require that φ(x) and ψ (x) satisfy
X X
Z∞ Z∞ (15) (−1)k ck = 0, k(−1)k ck = 0.
k∈Z k∈Z
φ(x) dx = 1, |φ(x)|2 dx = 1,
−∞ −∞
(10) Z∞ 5 A simple proof, based on the MRA axioms, that
R1
And, for L > 2 , equations (9) and (12) yield ad- α̃kn by α̃kn = 0 f (x)φ̃n,k (x) dx . And, we periodi-
ditional equations similar to the ones in (15).
cally extend f with period 1 , also denoting this
There is a finite set of coefficients that solves
periodic extension by f . Then, for n ≥ 0 and
the equations in (14) and (15), namely,
k = 0, 1, . . . , 2n − 1 , we have
√ √
1+ 3 3+ 3 Z1 X
c0 = √ , c1 = √ , n
4 2 4 2 α̃k = f (x)2n/2 φ 2n (x + j) − k dx
(16) √ √ 0 j∈Z
3− 3 1− 3
c2 = √
4 2
, c3 = √
4 2 (20) X Z j+1
= f (x − j)2n/2 φ(2n x − k) dx
j∈Z j
with all other ck = 0 . Using these values of ck ,
the following iterative solution of (11) = αkn .
Z∞
If we use the coefficients c0 , c1 , c2 , c3 defined
in (16), then for N > 2 , WN has the following F(x) 2R φ(2R (x − xk )) dx ≈ F (xk ) .
−∞
structure :
This approximation will hold for all period 1 con-
tinuous functions F and will be more accurate the
larger the value of 2R = M .
The higher the order of a Daubechies wavelet,
the more of its moments are zero. A Daubechies
wavelet of order 2L is defined by 2L nonzero coef-
ficients {ck } , has its first L moments equal to zero,
and is supported on the interval [0, 2L − 1] . Gener-
ally speaking, the more moments that are zero, the
more wavelet coefficients that are nearly vanishing
for smooth functions F . This follows from consid-
ering Taylor expansions. Suppose F(x) has an L -
c +c c +c term Taylor expansion about the point xk = k2−n .
2 0 3 1
For N = 2, W2 = c1 +c3 −c0 −c2 . For other That is,
Daubechies wavelets, there are other finite co-
efficient sequences {ck } , and the matrices WN X
L−1
1 (j)
F(x) = F (xk ) (x − xk )j
are defined similarly. The permutation matrix PN j!
j=0
is the same one that we defined for the Haar
transform and is used to sort the (N − 1)st level 1 (L)
+ F (tx ) (x − xk )L
coefficients so that WN−1 can be applied to the L!
scaling coefficients {α̃kN−1 }. where tx lies between x and xk . Suppose also that
As initial data for the wavelet transform we ψ is supported on [−a, a] and that ψ has its first L
can, as we did for the Haar transform, use dis- moments equal to 0 and that |F (L) (x)| is bounded
crete data of the form {fj }M−1 j=0 . The equations by a constant B on [(k − a)2−n , (k + a)2−n ] . It then
in (15) then provide a discrete analog of the zero follows that
moment conditions in (9) [for L = 2]; hence the
wavelet coefficients will be 0 where the data is L+1/2
B a
linear. In the last section, we saw how this can (22) |β̃n
k| ≤
p .
L + 1/2 L! 2n
produce effective compression of signals when
just the 0th and 1st moments of ψ are 0 .
It is often the case that the initial data are val- This inequality shows why ψ (x) having zero mo-
ues of a measured signal, i.e. {fk } = {F(xk )} , for ments produces a large number of small wavelet co-
xk = k2−n , where F is a signal obtained from a efficients. If F has some smoothness on an interval
measurement process. As shown in the previous (c, d), then wavelet coefficients β̃n k corresponding to
section, we can interpret the behavior of the dis- the basis functions ψ (2n (x − xk )) whose supports
crete case based on properties of the function are contained in (c, d) will approach 0 rapidly as n
F . A measured signal increases to ∞ .
R ∞F is often described by a In addition to the Daubechies wavelets, there is
convolution: F(x) = −∞ g(t)µ(x − t) dt, where g
is the signal being measured and µ is called the another class of compactly supported wavelets called
instrument function. Such convolutions generally coiflets. These wavelets are also constructed using
have greater regularity than a typical function the method outlined above. A coiflet of order 3L is
in L2 (R). For instance, if g ∈ L1 (R) and is sup- defined by 3L nonzero coefficients {ck } and has its
ported on a finite interval and µ = 1[−r ,r ] for first L moments equal to zero and is supported on
some positive r , then F is continuous and sup- the interval [−L, 2L − 1] . A coiflet of order 3L is dis-
ported on a finite interval. By a linear change of tinguished from a Daubechies wavelet of order 2L
variables, we may then assume that F is sup- in that, in addition to ψ having its first L moments
ported on [0, 1]. The data {F(xk )} then provide equal to zero, the scaling function φ for the coiflet
approximations for the highest level scale coef- also
R ∞ has L − 1 moments vanishing. In particular,
x j φ(x) dx = 0, for j = 1, . . . , L − 1. For a coiflet of
ficients {α̃kR } . If we assume that F has period 1 , −∞
then order 3L , supported on [−a, a] , an argument simi-
lar to the one that proves (22) shows that
Z∞ L+1/2
B a
α̃kR = 2−R/2 F(x) 2R φ 2R (x − xk ) dx |α̃kR − CM F(xk )| ≤ p .
−∞ L + 1/2L! 2R
≈ CM F (xk ) .
√
Here we have replaced 1/ 2R by the scale fac- This inequality provides a stronger theoretical jus-
tor CM and used tification for using the data {CM F(xk )} in place of
Figure 6. (a)[top] Signal. (b)[top] Signal’s coiflet30 transform, 7th level coefficients lie above the dotted
line. (b)[middle] 7th level coefficients. (a)[bottom] Signal with added noise. (b)[bottom] Noisy signal’s
coiflet30 transform. The horizontal lines are thresholds equal to ±0.15.
the highest level scaling coefficients, beyond menting the interval and performing a cosine ex-
the argument we gave above for Daubechies pansion on each segment, or by using a smoother
wavelets. The construction of coiflets was first version of the same idea involving local cosine
carried out by Daubechies and named after Coif- bases [3, 9, 12, 1, 5].
man (who first suggested them). One way to quantify the accuracy of these
approximations is to use relative R.M.S. differ-
Compression of Signals ences. Given two sets of data {fj }M−1j=0 and
One of the most important applications of {gj }M−1
j=0 , their relative R.M.S. difference,
wavelet analysis is to the compression of signals.
relative to {fj } , is defined by
As an example, let us use a Daub4 series to com-
press the signal {fj = F(j/1024)}1023 j=0 where v v
uM−1 uM−1
uX .u
F(x) = − log |x − 0.2|. See Figure 5(a)[top]. For u uX
this signal, a partial sum containing 99% of the D(f , g) = t |fj − gj | t
2 |fj |2 .
j=0 j=0
energy required only 37 coefficients (see Figure
5(b)[top]). It certainly provides a visually ac-
For the example above, if we denote the wavelet
ceptable approximation of {fj } . In particular, the
approximation by f w , then D(f , f w ) =
sharp maximum in the signal near x = 0.2 seems
9.8 × 10−3 . For the Fourier cosine series ap-
to be reproduced quite well. The compression
proximation, call it f c , we have
ratio is 1024: 37 ≈ 27: 1 , which is an excellent re- c −2
D(f , f ) = 2.7 × 10 . A rule of thumb for a vi-
sult considering that we also have 99% accu-
sually acceptable approximation is to have a rel-
racy. In addition, wavelet analysis has identi-
ative R.M.S. difference of less than 10−2. The ap-
fied the singularities of F . Notice in Figure
5(b)[bottom] the peak in the wavelet coefficients proximations in this example are consistent with
is near x = 0.2 , where F has a singularity, and this rule of thumb.
the largest wavelet coefficient is near x = 1 , We can also do more localized analysis with
where the periodic extension of F has a jump dis- R.M.S. differences. For example, over the subin-
continuity. terval [.075, .325] centered on the singularity
Turning to Fourier series, since the even pe- of F , we find that D(f , f w ) = 9.7 × 10−3 and
riodic extension is continuous, we used a discrete D(f , f c ) = 3.2 × 10−2. These numbers confirm
Fourier cosine series to compress this signal. In our visual impression that the wavelet series
Figure 5(a)[bottom] we show a 257-term discrete does a better job reproducing the sharp peak in
Fourier cosine series partial sum for {fj } . Even the signal. Or, using the subinterval [.25, .75], we
using seven times as many coefficients as the get D(f , f w ) = 1.0 × 10−2 and D(f , f c ) =
wavelet series, the cosine series cannot repro- 3.3 × 10−3 , confirming our impression that both
duce the sharp peak in the signal. Better results series do an adequate job approximating {fj }
could be obtained in this case by either seg- over this subinterval.
Figure 7. (a)[top] Denoised signal using wavelet analysis. (a)[bottom] Denoised signal using Fourier
analysis. (b)[top] Fourier coefficients of noisy signal and filter function. (b)[middle] Moduli-squared of
Fourier coefficients of original signal. (b)[bottom] Moduli-squared of Fourier coefficients of wavelet
denoised signal.
Although in the examples we have discussed Signal Denoising
so far Fourier analysis did not compress the sig- Wavelet analysis can also be used for removing
nals very well, we do not wish to create the im- noise from signals. As an example, we show in
pression that this will always be true. In fact, if Figure 6(a)[top] a discrete signal {f (j/1024)}1023
j=0
a signal is composed of relatively few sinusoids, where f (x) is defined by formula (1) with ν = 80.
then Fourier analysis will provide very good Each term of the form
compression. For example, consider the signal
{fj = f (j/1024)}1023
2
j=0 where f (x) is defined in (1) (5 cos 2π νx)e−640π (x−k/8)
with ν = 280 . The Fourier coefficients for f are
graphed in Figure 1(b)[top]. They tend rapidly to we shall refer to as a blip. Notice that each blip
0 away from the frequencies ±280 ; hence the is concentrated around x = k/8 , since
2
signal is composed of relatively few sinusoids. e−640π (x−k/8) rapidly decreases to 0 away from
By computing a Fourier series partial sum that x = k/8 . This signal can be interpreted as rep-
uses only the 122 Fourier coefficients whose resenting the bit sequence 1 0 1 1 0 1 1 . In Figure
frequencies are within ±30 of ±280 , we ob- 6(a)[bottom] we show this signal after it has
tained a signal g that was visually indistin- been corrupted by adding noise. In Figure
guishable from the original signal. In fact, 6(b)[top] we show the coiflet30 wavelet coeffi-
D(f , g) = 5.1 × 10−3 . However, by compressing cients for the original signal. The rationale for
{fj } with the largest 122 Daub4 wavelet coef- using wavelets to remove the noise is that the
ficients, we obtained D(f , f w ) = 2.7 × 10−1 and original signal’s wavelet coefficients are closely
the compressed signal f w was only a crude ap- correlated with the points near x = k/8 where the
proximation of the original signal. The reason blips are concentrated. To demonstrate this, we
that compactly supported wavelets perform show in Figure 6(b)[middle] a graph of the 7th
poorly in this case is that the large number of level wavelet coefficients {β̃7k }
−7 27 −1
rapid oscillations in the signal produce a corre- corresponding to the points {k2 }k=0 on the
spondingly large number of high magnitude unit interval. Comparing this to Figure 6(a)[top],
wavelet coefficients at the highest levels. Con- we can see that the positions of this level’s
sequently, a significant fraction of all the wavelet largest magnitude wavelet coefficients are closely
coefficients are of high magnitude, so it is not correlated with the positions of the blips. Simi-
possible to significantly compress the signal lar graphs could also be drawn for other levels,
using compactly supported wavelets. This ex- but the 7th level coefficients have the largest
ample illustrates that wavelet analysis is not a magnitude. In Figure 6(b)[bottom] we show the
panacea for the problem of signal compression. coiflet30 transform of the noisy signal. In spite
In fact, much work has been done in creating of the noise, the 7th level coefficients clearly
large collections of wavelet bases and Fourier stand out, although in a distorted form. By in-
bases and choosing for each signal a basis which troducing a threshold, in this case 0.15 , we can
best compresses it [12, 9, 3, 5]. retain these 7th level coefficients and remove
Conclusion
In this paper we have tried to show how the two
methodologies of Fourier analysis and wavelet
analysis are used for various kinds of work. Of
course, we have only scratched the surface of
both fields. Much more information can be found
in the references and their bibliographies.
This spectral density is shown in Figure 10. From the shape of this density
we see that in order to obtain segments in which the spectrum is flat, we
need to partition the spectrum finely at low frequencies, but only coarsely
at high frequencies. The subbands we obtain by this procedure will be ap-
proximately vectors of white noise with variances proportional to the power
spectrum over their frequency range. We can use an procedure similar to
that described for the KLT for coding the output. As we will see below,
this particular partition of the spectrum is closely related to the wavelet
transform.
where fn is the height of the step in [n, n + 1). A natural value for the
heights fn is simply the average value of f(x) in the interval [n, n + 1).
n+1
This gives fn = n f(x)dx.
We can generalize this approximation procedure to building blocks other
than the box function. Our more generalized approximation will have the
24 Geoffrey M. Davis, Aria Nosratinia
form
Af(x) = φ̃(x − n), f(x)φ(x − n). (1.23)
n
Let Vj be the space spanned by the functions {φj (x−2−j k)}. Our resolution
j approximation Aj f is simply a projection (not necessarily an orthogonal
one) of f(x) onto the span of the functions φj (x − 2−j k).
For our box function example, the approximation Aj f(x) corresponds to
an orthogonal projection of f(x) onto the space of step functions with step
width 2−j . Figure 11 shows the difference between the coarse approxima-
tion A0 f(x) on the left and the higher resolution approximation A1 f(x) on
the right. Dilating scaling functions give us a way to construct approxima-
tions to a given function at various resolutions. An important observation
is that if a given function is sufficiently smooth, the differences between
approximations at successive resolutions will be small.
Constructing our function φ(x) so that approximations at scale j are
special cases of approximations at scale j + 1 will make the analysis of
differences of functions at successive resolutions much easier. The function
φ(x) from our box function example has this property, since step functions
with width 2−j are special cases of step functions with step width 2−j−1 .
For such φ(x)’s the spaces of approximations at successive scales will be
nested, i.e. we have Vj ⊂ Vj+1 .
The observation that the differences Aj+1 f −Aj f will be small for smooth
functions is the motivation for the Laplacian pyramid [26], a way of trans-
forming an image into a set of small coefficients. The 1-D analog of the
procedure is as follows: we start with an initial discrete representation of
a function, the N coefficients of Aj f. We first split this function into the
sum
Aj f(x) = Aj−1 f(x) + [Aj f(x) − Aj−1 f(x)]. (1.25)
Because of the nesting property of the spaces Vj , the difference Aj f(x) −
Aj−1 f(x) can be represented exactly as a sum of N translates of the func-
1. Wavelet-based Image Coding: An Overview 25
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
FIGURE 11. A continuous function f (x) (plotted as a dotted line) and box func-
tion approximations (solid lines) at two resolutions. On the left is the coarse
approximation A0 f (x) and on the right is the higher resolution approximation
A1 f (x).
tion φj (x). The key point is that the coefficients of these φj translates will
be small provided that f(x) is sufficiently smooth, and hence easy to code.
Moreover, the dimension of the space Vj−1 is only half that of the space
Vj , so we need only N2 coefficients to represent Aj−1 f. (In our box-function
example, the function Aj−1 f is a step function with steps twice as wide as
Aj f, so we need only half as many coefficients to specify Aj−1 f.) We have
partitioned Aj f into N difference coefficients that are easy to code and
N
2 coarse-scale coefficients. We can repeat this process on the coarse-scale
coefficients, obtaining N2 easy-to-code difference coefficients and N4 coarser
scale coefficients, and so on. The end result is 2N − 1 difference coefficients
and a single coarse-scale coefficient.
Burt and Adelson [26] have employed a two-dimensional version of the
above procedure with some success for an image coding scheme. The main
problem with this procedure is that the Laplacian pyramid representation
has more coefficients to code than the original image. In 1-D we have twice
as many coefficients to code, and in 2-D we have 43 as many.
4.2 Wavelets
We can improve on the Laplacian pyramid idea by finding a more efficient
representation of the difference Dj−1 f = Aj f − Aj−1 f. The idea is that
to decompose a space of fine-scale approximations Vj into a direct sum
of two subspaces, a space Vj−1 of coarser-scale approximations and its
complement, Wj−1 . This space Wj−1 is a space of differences between coarse
and fine-scale approximations. In particular, Aj f − Aj−1 f ∈ W j−1 for any
f. Elements of the space can be thought of as the additional details that
must be supplied to generate a finer-scale approximation from a coarse one.
Consider our box-function example. If we limit our attention to functions
on the unit interval, then the space Vj is a space of dimension 2j . We can
26 Geoffrey M. Davis, Aria Nosratinia
0.5
0.4
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
−0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
FIGURE 12. D f (x), the difference between the coarse approximation A0 f (x)
0
φ (x) ψ (x)
1 1
-1 -1
decompose Vj into the space Vj−1 , the space of resolution 2−j+1 approxi-
mations, and Wj−1 , the space of details. Because Vj−1 is of dimension 2j−1 ,
Wj−1 must also have dimension 2j−1 for the combined space Vj to have
dimension 2j . This observation about the dimension of Wj provides us with
a means to circumvent the Laplacian pyramid’s problems with expansion.
Recall that in the Laplacian pyramid we represent the difference Dj−1 f
as a sum of N fine-scale basis functions φj (x). This is more information than
we need, however, because the space of functions Dj−1 f is spanned by just
N j j −j
2 basis functions. Let ck be the expansion coefficient φ̃ (x − 2 k), f(x)
in the resolution j approximation to f(x). For our step functions, each
coefficient cj−1
k is the average of the coefficients cj2k and cj2k+1 from the
resolution j approximation. In order to reconstruct Aj f from Aj−1 f, we
only need the N2 differences cj2k+1 − cj2k . Unlike the Laplacian pyramid,
there is no expansion in the number of coefficients needed if we store these
differences together with the coefficients for Aj−1 f.
The differences cj2k+1 − cj2k in our box function example correspond
(up to a normalizing constant) to coefficients of a basis expansion of the
space of details Wj−1 . Mallat has shown that in general the basis for
Wj consists of translates and dilates of a single prototype function ψ(x),
called a wavelet [27]. The basis for Wj is of the form ψj (x − 2−j k) where
j
ψj (x) = 2 2 ψ(x).
1. Wavelet-based Image Coding: An Overview 27
Figure 13 shows the scaling function (a box function) for our box func-
tion example together with the corresponding wavelet, the Haar wavelet.
Figure 12 shows the function D0 f(x), the difference between the approxi-
mations A1 f(x) and A1 f(x) from Figure 11. Note that each of the intervals
separated by the dotted lines contains a translated multiple of ψ(x).
The dynamic range of the differences D0 f(x) in Figure 12 is much smaller
than that of A1 f(x). As a result, it is easier to code the expansion coeffi-
cients of D0 f(x) than to code those of the higher resolution approximation
A1 f(x). The splitting A1 f(x) into the sum A0 f(x) + D0 f(x) performs a
packing much like that done by the Karhunen-Loève transform. For smooth
functions f(x) the result of the splitting of A1 f(x) into a sum of a coarser
approximation and details is that most of the variation is contained in A0 f,
and D0 f is near zero. By repeating this splitting procedure, partitioning
A0 f(x) into A−1 f(x) + D−1 f(x), we obtain the wavelet transform. The
result is that an initial function approximation Aj f(x) is decomposed into
the telescoping sum
Aj f(x) = Dj−1 f(x) + Dj−2 f(x) + . . . + Dj−n f(x) + Aj−n f(x). (1.26)
The coefficients of the differences Dj−k f(x) are easier to code than the
expansion coefficients of the original approximation Aj f(x), and there is
no expansion of coefficients as in the Laplacian pyramid.
Similarly, we have
φ̃(x) = h̃k φ̃(2x − k)
k
ψ(x) = gk φ(2x − k)
k
ψ̃(x) = g̃k φ̃(2x − k). (1.28)
k
28 Geoffrey M. Davis, Aria Nosratinia
time
FIGURE 14. Partition of the time-frequency plane created by the wavelet trans-
form.
1. Wavelet-based Image Coding: An Overview 29
There is a tradeoff between wavelet support and the regularity and accu-
racy of approximation. Wavelets with short support have strong constraints
on their regularity and accuracy of approximation, but as the support is
increased they can be made to have arbitrary degrees of smoothness and
numbers of vanishing moments. This limitation on support is equivalent to
keeping the analysis filters short. Limiting filter length is also an important
consideration in the subband coding literature, because long filters lead to
ringing artifacts around edges.
7 C++ source code for a coder that implements these components is available from
Dead zone
x
x=0
techniques with ideas from wavelet constructions and yields filters that
perform significantly better than the popular 7/9 filter set from [37].
5.2 Boundaries
Careful handling of image boundaries when performing the wavelet trans-
form is essential for effective compression algorithms. Naive techniques for
artificially extending images beyond given boundaries such as periodization
or zero-padding lead to significant coding inefficiencies. For symmetrical
wavelets an effective strategy for handling boundaries is to extend the im-
age via reflection. Such an extension preserves continuity at the boundaries
and usually leads to much smaller wavelet coefficients than if discontinuities
were present at the boundaries. Brislawn [43] describes in detail procedures
for non-expansive symmetric extensions of boundaries. An alternative ap-
proach is to modify the filter near the boundary. Boundary filters [44, 45]
can be constructed that preserve filter orthogonality at boundaries. The
lifting scheme [46] provides a related method for handling filtering near the
boundaries.
5.3 Quantization
Most current wavelet coders employ scalar quantization for coding. There
are two basic strategies for performing the scalar quantization stage. If
we knew the distribution of coefficients for each subband in advance, the
optimal strategy would be to use entropy-constrained Lloyd-Max quantizers
for each subband. In general we do not have such knowledge, but we can
provide a parametric description of coefficient distributions by sending side
information. Coefficients in the high pass subbands of a wavelet transform
are known a priori to be distributed as generalized Gaussians [27] centered
around zero.
A much simpler quantizer that is commonly employed in practice is a
uniform quantizer with a dead zone. The quantization bins, as shown in
Figure 15, are of the form [n∆, (n + 1)∆) for n ∈ Z except for the central
bin [−∆, ∆). Each bin is decoded to the value at its center in the simplest
case, or to the centroid of the bin. In the case of asymptotically high rates,
uniform quantization is optimal [47]. Although in practical regimes these
dead-zone quantizers are suboptimal, they work almost as well as Lloyd-
34 Geoffrey M. Davis, Aria Nosratinia
Max coders when we decode to the bin centroids [48]. Moreover, dead-zone
quantizers have the advantage that of being very low complexity and robust
to changes in the distribution of coefficients in source. An additional ad-
vantage of these dead-zone quantizers is that they can be nested to produce
an embedded bitstream following a procedure in [49].
The covariance matrix is circulant, so the KLT for this process is the simply
the Fourier transform. The Fourier transform of Y is a very inefficient repre-
k
sentation for coding Y . The energy at frequency k will be |1+e2πi N |2 which
means that the energy of Y is spread out over the entire low-frequency half
of the Fourier basis with some spill-over into the high-frequency half. The
KLT has “packed” the energy of the two non-zero coefficients of Y into
roughly N2 coefficients. It is obvious that Y was much more compact in its
original form, and could be coded better without transformation: Only two
coefficients in Y are non-zero, and we need only specify the values of these
coefficients and their positions.
As suggested by the example above, the essence of the extensions to
traditional transform coding is the idea of selection operators. Instead of
quantizing the transform coefficients in a pre-determined order of prior-
ity, the wavelet framework lends itself to improvements, through judicious
choice of which elements to code. This is made possible primarily because
wavelet basis elements are spatially as well as spectrally compact. In parts
of the image where the energy is spatially but not spectrally compact (like
the example above) one can use selection operators to choose subsets of
the wavelet coefficients that represent that signal efficiently. A most no-
table example is the Zerotree coder and its variants (Section 7).
1. Wavelet-based Image Coding: An Overview 37
TABLE 1.1. Peak signal to noise ratios in decibels for coders discussed
in the paper. Higher values indicate better performance.
Lena (bits/pixel) Barbara (bits/pixel)
Type of Coder 1.0 0.5 0.25 1.0 0.5 0.25
JPEG [59] 37.9 34.9 31.6 33.1 28.3 25.2
Optimized JPEG [60] 39.6 35.9 32.3 35.9 30.6 26.7
Baseline Wavelet [61] 39.4 36.2 33.2 34.6 29.5 26.6
Zerotree (Shapiro) [62] 39.6 36.3 33.2 35.1 30.5 26.8
Zerotree (Said & Pearlman) [63] 40.5 37.2 34.1 36.9 31.7 27.8
Zerotree (R/D optimized) [64] 40.5 37.4 34.3 37.0 31.3 27.2
Frequency-adaptive [65] 39.3 36.4 33.4 36.4 31.8 28.2
Space-frequency adaptive [66] 40.1 36.9 33.8 37.0 32.3 28.7
Frequency-adaptive + zerotrees [67] 40.6 37.4 34.4 37.7 33.1 29.3
TCQ subband [68] 41.1 37.7 34.3 – – –
Bkwd. mixture estimation (EQ) [69] 40.9 37.7 34.6 – – –
noise ratios for the coders we discuss in the paper.8 The test images are
the 512×512 Lena image and the 512×512 Barbara image. Figure 16 shows
the Barbara image as compressed by JPEG, a baseline wavelet transform
coder, and the zerotree coder of Said and Pearlman [63]. The Barbara image
is particularly difficult to code, and we have compressed the image at a low
rate to emphasize coder errors. The blocking artifacts produced by the dis-
crete cosine transform are highly visible in the image on the top right. The
difference between the two wavelet coded images is more subtle but quite
visible at close range. Because of the more efficient coefficient encoding (to
be discussed below), the zerotree-coded image has much sharper edges and
better preserves the striped texture than does the baseline transform coder.
7 Zerotree Coding
The rate-distortion analysis of the previous sections showed that optimal
bitrate allocation is achieved when the signal is divided into subbands such
that each subband contains a “white” signal. It was also shown that for
typical signals of interest, this leads to narrower bands in the low frequen-
cies and wider bands in the high frequencies. Hence, wavelet transforms
have very good energy compaction properties.
This energy compaction leads to efficient utilization of scalar quantizers.
However, a cursory examination of the transform in Figure 17 shows that
a significant amount of structure is present, particularly in the fine scale
coefficients. Wherever there is structure, there is room for compression, and
FIGURE 16. Results of different compression schemes for the 512 × 512 Barbara
test image at 0.25 bits per pixel. Top left: original image. Top right: baseline
JPEG, PSNR = 24.4 dB. Bottom left: baseline wavelet transform coder [61],
PSNR = 26.6 dB. Bottom right: Said and Pearlman zerotree coder, PSNR =
27.6 dB.
40 Geoffrey M. Davis, Aria Nosratinia
LL3 LH3
LH2
HL3 HH3
LH1
HH2
HL2
HH1
HL1
Typical
Wavelet tree
Low frequency
H 2
0 Component
H 2
0
Input H 2 H 2
0 1
Signal Detail
H 2
1 Components
H 2
1
MSB 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 ....
x x x 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 ....
x x x x x x 1 1 1 1 0 0 1 0 0 0 0 0 ....
x x x x x x x x x x 1 1 1 1 1 1 0 0 ....
......
......
......
......
......
......
......
......
......
LSB
Wavelet Coefficients
in scan order
FIGURE 20. Bit plane profile for raster scan ordered wavelet coefficients.
mization. In phase one, the uniform scalar quantizers are fixed, and optimal
zerotrees are chosen. In phase two, zerotrees are fixed and the quantization
level of uniform scalar quantizers is optimized. This algorithm is guaranteed
to converge to a local optimum [74].
There are further details of this algorithm involving prediction and de-
scription of zerotrees, which we leave out of the current discussion. The
advantage of this method is mainly in performance, compared to both
EZW and SPHIT (the latter only slightly). The main disadvantages of this
method are its complexity, and perhaps more importantly, that it does not
generate an embedded bitstream.
where the hk and gk are the same as those in the recurrence equations
( 4.3) and ( 4.3).
The idea of wavelet packets is most easily seen in the frequency do-
main. Recall from Figure 14 that each step of the wavelet transform splits
the current low frequency subband into two subbands of equal width, one
high-pass and one low-pass. With wavelet packets there is a new degree of
freedom in the transform. Again there are N stages to the transform for a
signal of length 2N , but at each stage we have the option of splitting the
low-pass subband, the high-pass subband, both, or neither. The high and
low pass filters used in each case are the same filters used in the wavelet
transform. In fact, the wavelet transform is the special case of a wavelet
packet transform in which we always split the low-pass subband. With this
increased flexibility we can generate 2N possible different transforms in 1-
D. The possible transforms give rise to all possible dyadic partitions of the
frequency axis. The increased flexibility does not lead to a large increase
in complexity; the worst-case complexity for a wavelet packet transform is
O(N log N ).
D0 D2 D0 D2
S
D0 D1 D2 D3 D0 D1 D2 D3 0
D1 D3 D1 D3
S
1
D0 D2
D1 D3
D0 D2
D1 D3
D2 D0
D3 D1
D2 D0
D3 D1
FIGURE 22. An 8-state TCQ trellis with subset labeling. The bits that specify the
sets within the superset also dictate the path through the trellis.
states, is given in Figure 22. The subsets {D0 , D1 , D2 , D3 } are also used to
label the branches of the trellis, so the same bit that specifies the subset
(at a given state) also determines the next state of the trellis.
Encoding is achieved by spending one bit per sample on specifying the
path through the trellis, while the remaining R − 1 bits specify a codeword
out of the active subset. It may seem that we are back to a non-optimal
rate-R quantizer (either S0 or S1 ). So why all this effort? The answer
is that we have more codewords than a rate-R quantizer, because there
is some freedom of choosing from symbols of either S0 or S1 . Of course
this choice is not completely free: the decision made at each sample is
linked to decisions made at past and future sample points, through the
permissible paths of the trellis. But it is this additional flexibility that
leads to the improved performance. Availability of both S0 and S1 means
that the reproduction levels of the quantizer are, in effect, allowed to “slide
around” and fit themselves to the data, subject to the permissible paths
on the trellis.
Before we continue with further developments of TCQ and subband cod-
ing, we should note that in terms of both efficiency and computational
requirements, TCQ is much more similar to VQ than to scalar quantiza-
tion. Since our entire discussion of transform coding has been motivated by
an attempt to avoid VQ, what is the motivation for using TCQ in subband
coding, instead of standard VQ? The answer lies in the recursive struc-
ture of trellis coding and the existence of a simple dynamic programming
method, known as the Viterbi algorithm [83], for finding the TCQ code-
words. Although it is true that block quantizers, such as VQ, are asymptot-
ically as efficient as TCQ, the process of approaching the limit is far from
trivial for VQ. For a given realization of a random process, the code vectors
generated by the VQ of size N − 1 have no clear relationship to those with
vector dimension N . In contrast, the trellis encoding algorithm increases
52 Geoffrey M. Davis, Aria Nosratinia
the coefficients bears a strong resemblance to the problem of best linear transform.
Both operations, namely the linear transform and partitioning, conserve energy. The
goal in both is to minimize overall distortion through optimal allocation of a finite rate.
Not surprisingly, the solution techniques are similar (Lagrange multipliers), and they
both generate sets with maximum separation between low and high energies (maximum
arithmetic to geometric mean ratio).
1. Wavelet-based Image Coding: An Overview 53
a good choice. Suppose that we want to classify the samples into J groups,
and that all samples assigned to a given class i ∈ {1, ..., J} are grouped into
a source Xi . Let the total number of samples assigned to Xi be Ni , and the
total number of samples in all groups be N . Define pi = Ni /N to be the
probability of a sample belonging to the source Xi . Encoding the source
Xi at rate Ri results in a mean squared error distortion of the form [90]
where 5i is a constant depending on the shape of the pdf. The rate allocation
problem can now be solved using a Lagrange multiplier approach, much in
the same way as was shown for optimal linear transforms, resulting in the
following optimal rates:
R 1 52 σ 2
Ri = + log2 J i i (1.34)
J 2 2 2 pj
j=1 (5j σj )
where R is the total rate and Ri are the rates assigned to each group.
Classification gain is defined as the ratio of the quantization error of the
original signal X, divided by that of the optimally bit-allocated classified
version.
52 σ 2
Gc = J (1.35)
2 2 pj
j=1 (5j σj )
One aims to maximize this gain over {pi }. It is not unexpected that
the optimization process can often yield non-uniform {pi }, resulting in
unequal population of the classification groups. It is noteworthy that non-
uniform populations not only have better classification gain in general, but
also lower overhead: Compared to a uniform {pi }, any other distribution
has smaller entropy, which implies smaller side information to specify the
classes.
The classification gain is defined for Xi taken from one subband. A gen-
eralization of this result in [68] combines it with the conventional coding
gain of the subbands. Another refinement takes into account the side in-
formation required for classification. The coding algorithm then optimizes
the resulting expression to determine the classifications. ECTCQ is then
used for final coding.
Practical implementation of this algorithm requires attention to a great
many details, for which the interested reader is referred to [68]. For example,
the classification maps determine energy levels of the signal, which are
related to the location of the edges in the image, and are thus related in
different subbands. A variety of methods can be used to reduce the overhead
information (in fact, the coder to be discussed in the next section makes the
management of side information the focus of its efforts) Other issues include
alternative measures for classification, and the usage of arithmetic coded
TCQ. The coding results of the ECTCQ based subband coding are some of
54 Geoffrey M. Davis, Aria Nosratinia
10 Future Trends
Current research in image coding is progressing along a number of fronts.
At the most basic level, a new interpretation of the wavelet transform
has appeared in the literature. This new theoretical framework, called the
lifting scheme [41], provides a simpler and more flexible method for design-
ing wavelets than standard Fourier-based methods. New families of non-
separable wavelets constructed using lifting have the potential to improve
56 Geoffrey M. Davis, Aria Nosratinia
coders. One very intriguing avenue for future research is the exploration of
the nonlinear analogs of the wavelet transform that lifting makes possible.
The area of classification and backward estimation based coders is an
active one. Several research groups are reporting promising results [92, 93].
One very promising research direction is the development of coded images
that are robust to channel noise via joint source and channel coding. See
for example [94], [95] and [96].
The adoption of wavelet based coding to video signals presents special
challenges. One can apply 2-D wavelet coding in combination to temporal
prediction (motion estimated prediction), which will be a direct counterpart
of current DCT-based video coding methods. It is also possible to consider
the video signal as a three-dimensional array of data and attempt to com-
press it with 3-D wavelet analysis. This approach presents difficulties that
arise from the fundamental properties of the discrete wavelet transform.
The discrete wavelet transform (as well as any subband decomposition) is
a space-varying operator, due to the presence of decimation and interpo-
lation. This space variance is not conducive to compact representation of
video signals, as described below.
Video signals are best modeled by 2-D projections whose position in
consecutive frames of the video signal varies by unknown amounts. Because
vast amounts of information are repeated in this way, one can achieve
considerable gain by representing the repeated information only once. This
is the basis of motion compensated coding. However, since the wavelet
representation of the same 2-D signal will vary once it is shifted10 , this
redundancy is difficult to reproduce in the wavelet domain. A frequency
domain study of the difficulties of 3-D wavelet coding of video is presented
in [97], and leads to the same insights. Some attempts have also been made
on applying 3-D wavelet coding on the residual 3-D data after motion
compensation, but have met with indifferent success.
downsampling rate
1. Wavelet-based Image Coding: An Overview 57
12 References
[1] A. Gersho and R. M. Gray, Vector Quantization and Signal Compres-
sion. Kluwer Academic Publishers, 1992.
[38] M. Vetterli and C. Herley, “Wavelets and filter banks: Theory and
design,” IEEE Trans. Acoust. Speech Signal Proc., vol. 40, no. 9,
pp. 2207–2232, 1992.
[52] Y. Shoham and A. Gersho, “Efficient bit allocation for an arbitrary set
of quantizers,” IEEE Trans. Acoustics, Speech, and Signal Processing,
vol. 36, pp. 1445–1453, Sept. 1988.
[63] A. Said and W. A. Pearlman, “A new, fast, and efficient image codec
based on set partitioning in hierarchichal trees,” IEEE Trans. Circuits
and Systems for Video Technology, vol. 6, pp. 243–250, June 1996.
[70] A. S. Lewis and G. Knowles, “Image compression using the 2-d wavelet
transform,” IEEE Transactions on Image Processing, vol. 1, pp. 244–
250, April 1992.
[72] A. Said and W. A. Pearlman, “A new, fast, and efficient image codec
based on set partitioning in hierarchical trees,” IEEE Transactions on
Circuits and Systems for Video Technology, vol. 6, pp. 243–250, June
1996.
Copyright
c 1995, 2002 Javier R. Movellan.
This is an open source document. Permission is granted to copy, distribute and/or modify
this document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant Sections,
no Front-Cover Texts, and no Back-Cover Texts.
Endorsements
This document is endorsed by its copyright holder: Javier R. Movellan. Modified versions
of this document should delete this endorsement.
1 Notation for discrete HMM
Let λ represent the current model and λ̄ represent a candidate model. Our objective is to
make Pλ̄ (o) ≥ Pλ (o), or equivalently log Pλ̄ (o) ≥ log Pλ (o).
P
Due to the presence of stochastic constraints (e.g., aij ≥ and j aij = 1) it turns out to
be easier to maximize an auxiliary function Q(·) rather than to directly maximize log P λ̄ .
The E-M auxiliary function is defined as follows:
X
Q(λ, λ̄) = Pλ (q|o)log Pλ̄ (qo) (1)
q
Here we show that Q(λ, λ̄) ≥ Q(λ, λ) → log Pλ̄ (o) ≥ log Pλ (o).
For any model λ or λ̄ it must be true that
Pλ̄ (oq)
Pλ̄ (o) = (2)
Pλ̄ (q|o)
or logPλ̄ (o) = logPλ̄ (oq) − logPλ̄ (q|o).
Also, X
logPλ̄ (o) = Pλ (q|o)logPλ̄ (o) (3)
q
Applying the Q(·, ·) function to (λ, λ̄) and to (λ, λ), it follows that,
Q(λ, λ̄) − Q(λ, λ) = log Pλ̄ (o) − log Pλ (o) − KL(λ, λ̄) (6)
where KL(·, ·) is the Kullback-Leibler criterion (relative entropy) of the probability distri-
bution Pλ (q|o) with respect to the probability distribution Pλ̄ (q|o)
X Pλ (q|o)
KL(λ, λ̄) = Pλ (q|o)log (7)
q
Pλ̄ (q|o)
Rearranging terms,
log Pλ̄ (o) − log Pλ (o) = Q(λ, λ̄) − Q(λ, λ) + KL(λ, λ̄) (8)
with o including the entire set of sequences in the training sample. Assuming that the
sequences are independent, it follows that
X X XX
Q(λ, λ̄) = Pλ (q|o) log Pλ̄ (q l ol ) = (log Pλ̄ (q l ol ))Pλ (q|o) (12)
q l l q
And since each state sequence q l depends only on the corresponding observation sequence
ol it follows that
XX X K
Y
Q(λ, λ̄) = (log Pλ̄ (q l ol )) Pλ (q m |o) (13)
l q l q−q l m=1
XX X K
Y
Q(λ, λ̄) = (log Pλ̄ (q l ol ))Pλ (q l |ol ) Pλ (q m |o) = (14)
l q l q−q l m6=l
XX X Y
= (log Pλ̄ (q l ol ))Pλ (q l |ol ) Pλ (q m |o) (15)
l ql q−q l m6=l
and since
K
X Y
Pλ (q m |o) = 1 (16)
q−q l m6=l
it follows that the E-M function can be decomposed into additive E-M functions, one per
observation sequence in the training sample:
XX X
Q(λ, λ̄) = Pλ (q l |ol ) log Pλ̄ (q l ol ) = Ql (λ, λ̄) (17)
l ql l
2.3 Maximizing the E-M function
For simplicity let us start with the case in which there is a single sequence. The results
easily generalize to multiple sequences. Since we work with a single sequence we may
drop the l superscript.
X
Q(λ, λ̄) = Pλ (q|o)log Pλ̄ (qo) (18)
q
And since,
Pλ̄ (qo) = π̄q1 b̄q1 (o1 )āq1 q2 b̄q2 (o2 )āq2 q3 ... (19)
it follows that, X
Q(λ, λ̄) = Pλ (q|o)log π̄q1 + (20)
q
Tl X
X
+ Pλ (q|o)log b̄qt (ot )+ (21)
t=1 q
TX
l −1 X
where δ(i, qt )δ(j, ot ) tells us to include only those cases for which qt = i and ot = j.
Therefore,
X
Pλ (q|o)δ(i, qt )δ(j, ot ) = Pλ (qt = i|o)δ(ot , j) = γt (i)δ(ot , j) = (26)
q
where δ(i, qt )δ(j, qt+1 ) tells us to include only those cases for which qt = i and qt+1 = j.
Therefore,
X
Pλ (q|o)δ(i, qt )δ(j, qt+1 ) = Pλ (qt = i qt+1 = j|o) = ξt (i, j) (28)
q
Putting it together
X
Q(λ, λ̄) = γ1 (j) log π̄j + (29)
j
XX Tl
X
+ log b̄i (j) ( γt (i)δ(ot , j))+ (30)
i j t=1
XX TX
l −1
When there is more than a sequence, we just need to add up over sequences to obtain the
overall Q(·, ·) function
X X
Q(λ, λ̄) = log π̄j ( γ1l (j))+ (32)
j l
XX Tl
XX
+ log b̄i (j) ( γtl (i)δ(olt , j))+ (33)
i j l t=1
XX X TX
l −1
Note that the part of the overall Q(λ, λ̄) function dependent on π̄j is of the form wj log xj
with xj = π̄j and
XK
wj = γ1l (j) (35)
l=1
P
with constraints j xj = 1, and xj ≥ 0.
Equivalently, the part of the overall Q(λ, λ̄) function dependent of b̄i (j) is of the form
wj log xj with xj = b̄i (j) and
Tl
K X
X
wj = γtl (j)δ(olt , j) (36)
l=1 t=1
P
with constraints j xj = 1, and xj ≥ 0.
Finally, the part of the overall Q(λ, λ̄) function dependent of āij is also of the form
wj log xj with xj = āij and
XK TXl −1
The parameters of the new model λ̄ that maximize the overall Q(λ, λ̄) function easily
follow:
PK
γ1 (i)
π̄i = l=1
K
PK PTl
γt (i)δ(olt ,j)
b̄i (j) = l=1
PK
t=1
P Tl l (39)
l=1 t=1 γt (i)
PK PTl −1
t=1 ξt (i,j)
āij = Pl=1
K
PTl −1 l
l=1 t=1 γt (i)
2.4 Obtaining the E-M parameters from the scaled forward and backward
algorithms
2.5 MM Training
PK
δ(i,q̂1l )
π̄i = l=1
K
PK P Tl
δ(i,q̂tl )δ(j,olt )
b̄i (j) = l=1
PK
t=1
P Tl l (40)
l=1 t=1 δ(i,q̂t )
PK PTl −1
δ(i,q̂tl )δ(j,q̂t+1
l
)
āij = l=1
PK
t=1
PTl −1 l)
l=1 t=1 δ(i,q̂ t
Note that Viterbi decoding maximizes log Pλ (q, o) with respect to q for a fixed model (the
first M step) then we maximizes log Pλ (q, o) with respect to λ, for a fixed q (the second
M step). Since we are always maximizing with respect to some variables, log P λ (q o) can
only increase and convergence to a local maximum is guaranteed.
3 Notation for Continuous Density Models
The notation for the continuous case is the same as the discrete case with the following
additional terms.
P Number of dimensions per observation (e.g. cepstral coefficients): o t = (ot1 ...otP ).
M. Number of clusters within a state.
V = {v11 , ..., v1M ..., vN 1 ..., vN M }. Set of possible clusters of external observations (M
clusters per state). Each cluster vij is identified by a state index i and a cluster index j.
mt variable identifying the cluster index of the cluster that occurred at time t. For example
if cluster vik occurred at time t then qt = i and mt = k.
m = (m1 ...mt ). A sequence of cluster indexes, one per time step.
gik = Pλ (mt = k|qt = j). Emission probability (gain) of cluster kvik by state Sj .
φ(·). A kernel function (e.g. Gaussian) to model the probability density of a cluster of
observations produced by a state.
bj (ot ) = Pλ (ot |qt = j).
µik = (µik1 , ..., µikP ). The centroid or prototype of cluster vik .
2
σik1 σik12 ... σik1p
2
σ σik2 ... σik2p
Σik = ik21... The covariance matrix of cluster vik .
2
σikp1 σikp2 ... σikp
2
σikn . The variance of the l th dimension (e.g. cepstral) within cluster vik , a diagonal ele-
ment of Σik .
σiknm , The covariance between dimension l and dimension m within the cluster v ik an
off-diagonal elements of Σik . Usually assumed zero.
In the discrete case the observations are discrete, represented by an integer. In the con-
tinuous case the observations at each time step are P-dimensional real-valued vectors (e.g.
cepstrals coefficients). In our notation ot = (ot1 , ..., otP ).
The continuous observation model produces sequences of observations in the following
way: At each time step the system generates a hidden state qt according to a a state to state
transition probability distribution aqt−1 qt . Once qt has been generated, the system gener-
ates a hidden cluster mt according to a state to cluster emission probability distribution
gqt mt . Once the hidden cluster has been determined, an observation vector is produced
probabilistically according to some kernel probability distribution (e.g. Multivariate Gaus-
sian). We can think of the clusters as low level hidden states embedded within high level
hidden states qt . For example, the high level hidden states may represent phonemes and
the low-level hidden clusters may represent acoustic categories within the same phoneme.
For simplicity each state is assumed to have the same number of clusters (M) but the set of
clusters is different from state to state. Thus, there is a total of NxM clusters, M per state.
We represent cluster k of state Sj as vjk and we use the variable mt to identify the cluster
number within a state at time t. Thus if qt = i and mt = k it means that at time t cluster
vjk occurred.
If we know the cluster at time t, the probability density of a vector of continuous obser-
vations (e.g. cepstral coefficients) is modeled by a kernel function, usually a multivariate
Gaussian
Pλ (ot |vjk ) = Pλ (ot |qt = j mt = k) = φ(ot , µjk , Σjk ) (41)
Where φ(·) is the kernel function (e.g. multivariate Gaussian) µjk = (µjk1 , ..., µjkP ) is
a centroid or prototype that determines the position of the cluster in P-dimensional space,
and 2
σjk1 σjk12 ... σjk1P
σjk21 σ 2 ... σjk2P
Σjk = jk2 (42)
...
2
σjkP 1 σjkP 2 ... σjkP
is a covariance matrix that determines the width and tilt of the cluster in P-dimensional
2 2
space. The diagonal terms {σjk1 ...σjkP } are the cluster variances for each dimension.
They determine the spread of the cluster on each dimension. The off-diagonal elements are
known as the cluster covariances and they determine the tilt of the cluster. For the Gaussian
case, the kernel function is
1
φ(ot , µjk , Σjk ) = e−d(ot ,µjk ) (43)
(2π)P/2 |Σjk |1/2
where |Σjk | is the determinant of the variance matrix, and d(ot , µjk ) is known as the Ma-
halanobis distance between the observation and the kernel’s centroid.
1
d(ot , µjk ) = (ot − µjk )Σ−1
jk (ot − µjk )
0
(44)
2
Since the covariance matrices Σjk are symmetric, each kernel is defined by P (P2+3) param-
eters (P for the centroids and P (P + 1)/2 for the variances). In practice the off-diagonal
variances are assumed zero, reducing the number of parameters per kernel to 2P. In such
QP 2
case the determinant |Σjk | is just the product of P scalar variances |Σjk | = l=1 σjkl , and
the Mahalanobis distance becomes a scaled Euclidean distance:
P
1 X (otl − µjkl )2
d(ot , µjk ) = 2 (45)
2 σjkl
l=1
Given a state Sj , the system randomly chooses one of its M possible clusters with state to
cluster emission probability P (mt = k|qt = j). This probability is assumed independent
of t and thus it can be represented by a parameter with no time index. In our notation
P (mt = k|qt = j) = gjk , the gain of the k th cluster embedded in state Sj .
Thus, the overall probability density of the observations generated by a state S j is given by
a weighted mixture of kernel functions.
M
X
Pλ (ot |qt = j) = P (mt = l|qt = j)P (ot |qt = j, mt = k) = (46)
l=1
or
M
X
bj (ot ) = gjk φ(ot , µjk , Σjk ) (47)
k=1
4.2 Forward and backward variables
The un-scaled and the scaled algorithms work the same as in the discrete case. Only now
the emission probability terms bj (ot ) are modeled by a mixture of densities.
M
X
bj (ot ) = gjk φ(ot , µjk , Σjk ) (48)
k=1
4.3 EM Training
In the continuous case the clusters are low level hidden states mt embedded within high
level hidden states qt . Thus, the E-M function is defined over all possible q m sequences
of high-level and low-level hidden states
XX
Q(λ, λ̄) = Pλ (qm|o)log Pλ̄ (qmo) (49)
q m
and since
Pλ̄ (qmo) = π̄q1 ḡq1 m1 φ(o1 , µ̄q1 m1 , Σ̄q1 m1 ), ..., āqT −1 qT ḡqT mT φ(oT , µ̄qT mT , Σ̄qT mT )
(50)
it follows that
XX
Q(λ, λ̄) = Pλ (qm|o)log π̄q1 + (51)
q m
T
X −1 X X
Pλ (qm|o)log āqt qt+1 + (52)
t=1 q m
T XX
X
+ Pλ (qm|o)log ḡqt mt (53)
t=1 q m
T XX
X
+ Pλ (qm|o)log φ(ot , µ̄qt mt , Σ̄qt mt ) (54)
t=1 q m
Since the factors in the first two terms are independent of m they simplify into
XX X
Pλ (qm|o)log π̄q1 = Pλ (q|o)log π̄q1 (55)
q m q
and
T
X −1 X X T
X −1 X
Pλ (qm|o)log āqt qt+1 = Pλ (q|o)log āqt qt+1 (56)
t=1 q m t=1 q
These terms are identical as in the discrete case and thus the same training rules for initial
state probabilities and for state transition probabilities apply here.
To find the training formulas for the cluster gains, we focus on the part of Q(·) dependent
on the gain terms. This part can be transformed as follows
T XX
X
Pλ (qm|o)log ḡqt (mt ) = (57)
t=1 q m
T X
X N X
M XX
= Pλ (qm|o)log ḡik δ(i, qt )δ(j, mt ) (58)
t=1 i=1 k=1 q m
T X
X N X
M
= Pλ (qt = i mt = k|o)log ḡik (59)
t=1 i=1 k=1
Thus the part of Q(·) that depends on ḡik is of the form wj log xj with xj = ḡik and
T
X
wj = Pλ (qt = i mt = k|o) (60)
t=1
P
with constraints j xj = 1 and xj ≤ 0, with maximum achieved for
wj
xj = P (61)
j wj
Thus
PT PT
Pλ (qt = i mt = k|o) t=1 Pλ (qt = i mt = k|o)
ḡik = PT t=1
PM = PT (62)
t=1 k=1 Pλ (qt = i mt = k|o) t=1 Pλ (qt = i|o)
To find the learning rules for the centroids and variances we focus on the part of Q(·) that
depends on the cluster centroids and variances, which is given by the following expression:
T XX
X
Pλ (qm|o)log φ(ot , µ̄qt mt , Σ̄qt mt ) (63)
t=1 q m
This expression can be transformed as follows
T XX
X
Pλ (qm|o)log φ(ot , µ̄qt mt , Σ̄qt mt ) = (64)
t=1 q m
T X
X N X
M XX
= Pλ (qm|o)log φ(ot , µ̄ik , Σ̄ik )δ(i, qt )δ(k, mt ) (65)
t=1 i=1 k=1 q m
T X
X N X
M
= Pλ (qt = i mt = k|o)log φ(ot , µ̄ik , Σ̄ik ) (66)
t=1 i=1 k=1
Thus, at a maximum,
T N M
∂ XXX
Pλ (qt = i mt = k|o)log φ(ot , µ̄ik , Σ̄ik ) = 0 (67)
∂ µ̄ikn t=1 i=1
k=1
and since
PT
t=1 Pλ (qt = i mt = k|o) otn
µ̄ikn = PT (70)
t=1 Pλ (qt = i|o)
2
A similar argument can be made for the diagonal variance σikl . In this case
Thus, at a maximum
T
X (otl − µ̄ikn )2
Pλ (qt = i mt = k|o)(1 − 2 )=0 (72)
t=1
σ̄ikn
PT
Pλ (qt = i mt = k|o)(otl − µ̄ikn )2
σ¯2 ikn = t=1
PT (73)
t=1 Pλ (qt = i|o)
Training for the mixture gains, mixture centroids, and mixture variances requires the
P (qt = jmt = k|o) terms, for t = 1...T , j = 1..N , k = 1...M . To obtain these
terms note the following:
P (ot mt = k|qt = j)
= P (qt = j|o) (76)
P (ot |qt = j)
Thus
gjk φ(ot , µik , Σik )
P (qt = jmt = k|o) = P (qt = j|o) (78)
bj (ot )
As in the discrete case, the P (qt = j|o) can be obtained through the scaled feed-forward
algorithm.
For the case with multiple training sequences the overall Q(·) decomposes into additive
Ql (·), one per training sequence. As a consequence we have to add in the numerator and
denominator of the training formulas the effects of each training sequence.
Summarizing, the E-M learning rules for the mixture of Gaussian densities case with diag-
onal covariance matrices are as follows:
PK
Pλ (q1 =j|ol )
π̄i = l=1
K
PK PTl −1
Pλ (qt =i qt+1 =j|ol )
āij = l=1
PK
t=1
PTl −1 l
l=1 t=1 Pλ (qt =i|o )
P K P Tl
P (q =i mt =k|ol )
ḡik = PK t=1
l=1 PT λ t l (79)
l=1 t=1 Pλ (qt =i|o )
PK P Tl l l
t=1 Pλ (qt =i mt =k|o ) otn
µ̄ikn = l=1 P T
P (q =i|o l)
t=1 λ t
PK P Tl
2 Pλ (qt =i mt =k|ol ) (otn −µ̄ikn )2
σ̄ikn = l=1 t=1
P K P Tl l
l=1 t=1 Pλ (qt =i|o )
where l = 1, ..., K indexes the training sequence, t = 1, ..., Tl indexes the time within
a training sequence, i = 1, ..., N and j = 1, ..., N index the hidden state, k = 1, ..., M
indexes the cluster embedded within a state, and n = 1, ..., P indexes the dimension of the
continuous vector of observations (e.g. the cepstral coefficient). The P (q t = jmt = k|o),
Pλ (qt = i|ol ) and Pλ (qt = i qt+1 = j|ol ) terms are obtained from the scaled forward-
backward algorithms according to the following formulas:
PM
bi (ot ) = k=1 gik φ(ot , µik , Σik )
1
φ(ot , µik , Σik ) = (2π)P/2
Q P
σikn
e−d(ot ,µik )
n=1
(80)
1
PP otn −µikn 2
d(ot , µik ) = 2 n=1 ( σikn )
We can use Viterbi decoding to find the best possible sequence of high level hidden states
q and low level clusters m. There are two approaches to this problem. One approach
attempts to find simultaneously the best joint sequence of high-level and low-level states.
Thus the goal is to find q̂ m̂ = arg maxqm Pλ (qm|o). The second approach first finds the
best possible sequence of high level states q̂ = arg maxq Pλ (q|o) and once q̂ has been
found, m̂ is defined as m̂ = arg maxm Pλ (m̂|q̂ o). The two approaches do not necessarily
yield the same results. The second approach is the standard in the literature.
For the second, most used version, of Viterbi decoding, the same Viterbi recurrence as in
the discrete case applies but using the continuous version of b i (ot ). Once we have qˆt , the
desired m̂t is simply the cluster within Sq̂t which is closest (in Mahalanobis distance) to
ot .
The objective in Viterbi training (also known as segmental k-means) is to find an opti-
mal point (local maximum) of log Pλ (oq̂ m̂) with q and λ being the optimizing vari-
ables. It does not matter how q̂ m̂ are found as long as a consistent procedure is used
throughout training. As in the discrete case we define a dummy model λ̂ such that
Pλ̂ (q l ml |o) = δ(q̂ l , q l )δ(m̂l , ml ). For such model, Pλ̂ (q1l = j|ol ) = δ(i, q̂1l ), Pλ̂ (qtl =
i|ol ) = δ(i, q̂tl ), Pλ̂ (qtl = i qt+1
l
= j|ol ) = δ(i, q̂tl )δ(j, q̂t+1
l
), and Pλ̂ (qtl = i mlt = k|ol ) =
l l
δ(i, q̂t )δ(k, m̂t ).
As in the discrete case note that Q(λ̂, λ̄) = log Pλ̄ (o q̂ m̂), the function we want to op-
timize. Thus, substituting the λ̂ parameters in the standard E-M formulas guarantees that
Q(λ̂, λ̄) ≥ Q(λ̂, λ), and thus log Pλ̄ (o q̂ m̂) ≥ log Pλ (o q̂ m̂)
Thus, the Viterbi training rules are as follows:
PK
δ(i,q̂1l )
π̄i = l=1
K
PK PTl −1
δ(i,q̂tl )δ(j,q̂t+1
l
)
āij = l=1
PK
t=1
PTl −1 l
l=1 t=1 δ(i,q̂t )
PK PTl
δ(i,q̂tl )δ(k,m̂lt )
ḡik = PK t=1
l=1 PT l (81)
l=1 t=1 δ(i,q̂t )
PK PTl
δ(i,q̂tl )δ(k,m̂lt ) oltn
µ̄ikn = l=1 t=1
PT l
t=1 δ(i,q̂t )
PK PTl
δ(i,q̂tl )δ(k,m̂lt )(otn −µ̄ikn )2
σ¯2 ikn = l=1 t=1
P K P Tl l
l=1 t=1 δ(i,q̂t )
Since Viterbi decoding maximizes Pλ (q, m, o) with respect to q and m and Viterbi train-
ing maximizes Pλ (q, m, o) with respect to λ, repeatedly applying Viterbi decoding and
Viterbi training, can only make Pλ (q, m, o) increase and convergence to a local maximum
is guaranteed.
5 Factored Sampling Methods for Continuous State Models
Many recognition problems can be framed in terms of infering something about q t the
internal state of a system, based on a sequence of observations o = o 1 · · · ot . These infer-
ences are in many cases based on estimates of p(qt |o1 · · · ot ). When the states are discrete
and countable, these conditonal state probabilities can be obtained using the forwards algo-
rithm. However, the algorithm cannot be used when the states are continuous. In such case,
direct sampling methods are appropriate. Here is an example of how these methods work.
We start with a sensor model: p(ot |qt ) and a Markovian state dynamics model p(qt+1 |qt ).
Our goal is to obtain estimates of p(qt |ot ) for all t.
1. Recursion
Assume we have an estimate p̂(qt |ot ). Our goal is to update that estimate for the
next time step p(qt+1 |ot+1 ).
(a) First we draw a random sample X from p̂(qt |ot ). This sample will implicitely
define a re-estimation of p(qt |ot ) in terms of a mixture of delta functions:
p̂(qt |ot ) = N1 N
P
i=1 δ(qt , xi )
(b) For each observation xi we obtain another random observation yi using the
state dynamics p(qt+1 |qt = xi ). The new sample Y = {y1 · · · yn } im-
plicitely defines our estimates of p(qt+1 |ot ).
PN
p(qt |ot ) = N1 i=1 δ(qt+1 , yi )
(c) We know that
p(o1 · · · ot )
p(qt+1 |o1 · · · ot+1 ) = p(qt+1 ot+1 |o1 · · · ot ) = (82)
p(o1 · · · ot+1 )
p(o1 · · · ot )
p(qt+1 |o1 · · · ot )p(ot+1 |qt+1 )
p(o1 · · · ot+1 )
The fraction is a constant K(ot+1 ) independent of qt+1 , we already have an
estimate of p(qt+1 |ot ) so we just need to weight it by p(ot+1 |qt+1 ).
p̂(qt+1 |ot+1 ) = K N1 N
P
i=1 δ(qt+1 , yi )p(ot+1 |qt+1 ) =
N
1 X
= P δ(qt+1 , yi )p(ot+1 |yi )
(N ) i p(ot+1 |yi ) i=1
We can now use p̂(qt+1 |ot+1 ) to estimate parameters like the mean or the
variance
R of the distribution. More generally,
ω̂ = dqt+1 Q(p(qt+1 |ot+1 ), qt+1 )
2. Initialization
The initialization step is basically the same as the recursion step only that instead
of using the state transition probabilities we use the initial state probabilities
(a) Obtain a sample of N random states : X= {x1 · · · xN } from the initial state
probability function π(·). These N samples will implicitely work as our
estimate of the initial state probability.
N
1 X
p̂(q1 ) = δ(q1 , yi ) (83)
N i=1
(b) Weight each observation by the sensor probability p(o1 |q1 = yi ). This de-
fines our initial distribution estimates
1
p̂(q0 |o1 ) = PN δ(q0 , yi )p(o0 |yi ) (84)
N i=1 p(o1 |xi )
!#"$&%')( "$*,+-.* *21 9:;12-2
" +-?>@;12+-,$ *
/0 0436587 3 3=< 3 5
A BDCFEGIHKJLNMOGIPRQFCTSUSVGWHEYX[Z]\#EGIH_^ L`\&aUJ)PTZ]bYcdGWHfegL
EEZ]HKP
h Z]\iGIjIPUJEJGIP
kIlUmon)prqtsvu`lUw
myx?z{k|z?}I~m|w=p ( -.*@ - 1 !
3 5 0
@
.,;:
,8
-?
.:
$ ¡¢£¤ *2!¥¥¦-§§
¨]©ª «¢«¬4 ® < +-¯12)1 0 ,"$4*,1.,1.*¥r" 1,+o 3o0° ° "$; 0 91,+-± 043 ;² 12* 0 9³´µ"$!¥ 3 *2" 043 3 |*,1.,-;12-24,"$*2¶
!#" "$!#"$·41," ¹1,+- )*22" (d¥ " ¹º)² ² 21?³';1 .+ " * º³ * 9 *.;²-.4()$i *2;²-.4(
4123 2¶04" 36¼ 5i¸ 1,+- 4+T 3 122" "$3 4K;½)!¹² $0 `" )120 4" 7 `3 *.2" (`
7 t!¥0 .+- "$4K 4 3 4§ 3-04 3 » F)"$*.;-*2*
r+- º³ 0 3 * 1," 0 *¹.¹ 3 04"$3 ¿»-]° Wr+- 12+ §W 3 2¥4 5¾(-4¸ ¥ *2," (d¥+ V*2 3 ² ² ,1 3 ;0 1 K12.3 4" " `
(d&² 3 .;12" 7 4 §` 0 "$!#² 04 3 !# 12 3 `)3" *.;-*.*@" 3 ;1.4" '1,+-K¶0 ;5] ¸ 4² ² " ¯12.+ 0" ¿-¹+)"$0 .+`"$°* *.0 =1 3 3 *,1.,-;3 1
À!¥
º³ * 1," *rr+ "$.+¯3. "3 r" 1,+-3 )41. *2+ o3 + Uº ² ² 3 21Á³Á;31 r!¥2+ " *r + 0 04,3 §¥$.
7 0 043 3-043 3 3 5À¸ 0 0 0 0 3 3 ° °
° 3 " 3 Â 3 " 1. ³´Ã)" !¥ 3 *2" 043 (§` 0 !#² 1," 3
1,+-#³´U)"$!¥ 3 *," 043 9 0 + 0 ! 0 3 0 -*&² 0 § 3-0 !#" 48 3 IÄ4-*2*2"$ 3
."$4r(-*2"$*¥9 ;1," ¶ ;$* + " $ 2§¼+)"$4+|³´Å)"$!# *2" $ 2!¥4 §v( ±" r9 ¥ 24 "$·412"
²d29 .!¥ 3 ¥r043 + " $43 1Á²-.5t *. ¸ 181,+-.° ;½"$*212* 12+ ,§ir3 + "$.04+¯ 3 *2+ 0 *Á1,+-3-40 1r ] 0 .4 "$·41,0 " ²d3 29 .!¥ 043
"$* )0 2-3 {
.
Æ 9 3 &º³ * 1,+-.i3 .¹*. .4Á.4-3-!¥0 1.*?0 r+ "$2+¼*2)² ² 0 21?1,+- (-*200 I3 + "$4+t 04.3 ;§ 0 9º³ 3 *
r+ "$.+`8&2 "$;0 7 *, 12* 9'* !¥K;½²d° 2" !¥ 12*r+ 3 "$.+`2" *2²)"$.0 `(§¯1,+-0 *.&.° 4-!¥ 1.*@.K4$* ² 0 .*. 1.7
@4" -!¥° -5*ÁÇ ½!#² 0 * 0 ¥²- 9È* 96! 3 *,1 912+ ¶;§#1,+-3 .!¥* + 2" * ;!#43 1.2" 4 i0 ÉÀ+ ²3 12+ 41 5
¸1,+-K.° 3 " 0 ¯12+-41@ 3 $=!#040 1.20" 46"$*0 *,0 1@" #9:.*2+¯ "$0 4+1 5< 3 3 0
 3 ° 3=0 3 5
Ê
ËÌÍ&¡d¢4 ª £ º ² ² 0 ,1³';1 0 7 .+ " 3 * º12412"$*,12" 4'Î6 3 " 3 < -+ 0 2§ ³´FÏ" !¥ 3 2* " 40 3 %41,12 3`Ç 0 3 " 12" 40 3
Ð zÑnÒ8ÓÔ ÕÖ@×ØÓ)Ù,ÕÁÒ
z
}IÕÁ×@ÒÖTÕ'ÒÓW!6Ò"Ô$#%2Ù'&
#ÓÙ2Õ'ÒTuÔ$(ÕÔ$)*#Ò@Ø
`Õ+(,#Iu#ÓÓÔ-ÒÃm-6ØÕ+.ÁÒ@Ù,ÓÙ2Õ'Ò0/12#Ô-Ò,3
Ù,Ò".546#ÀØ7Ù,Ò"
ò<PD
ó
vò< DKL Àó òø ó
àãÜWãÛåã÷ÿiÛÀÜèÃå|çÜè'áäÈãê@ò<PDKLÀóiÜëä{áã«á÷
vò<PD LÀó]ý|å êoíÁÜWÿiß«ä:ããÜè@ò<PDKLÀó < LÁô¯Ú¥ÛÀä{á¯äá¯å
èÀä{éÜ
ÿ¥å êtàdâ@ÿiß«äÈãäè`ã«ÛÀÜ
ãß«ÞÀÜ=ývÜåèFÜß«ßàß÷6íÀÞãiÞÀèìÈÜ)áá¹ÿ#Ü
Û'å ædܱådèFÜ)áã«äý|å-ãܱàâÿiÛ'å-ãvò<PDKLó¹ä{á÷
äÈãiäá¥èàãiædÜßê|Þ'áÜâöÞÀì.ô
Ü !ÞådèãäÈãê A¼ò ó=ä{á`éådìÈìÜ)çTãÛÀÜoÜëÝ'Ü)éã«Üçïßä{áõÁ÷àdß Þáã`ãÛÜ|ßä{áõÁô¯Üß«Üvÿ¹Ü¼ÿiäììKéådìÈìäÈã`ãÛÜ
Ú¥ÛÀ/
ådéãÞådì@ß«ä{áõÁ÷ÁãàÜývÝÀÛ'ådáäÜ=ãÛåã
äÈã
äá¯ãÛÜ !Þådèã«äÈãêoã«Ûå-ã`ÿ#ÜIåß«ÜWÞì:ã«äý|å-ãÜìê|äèã«Üß«ÜáãÜ)çTäÈèrô`Ú¥ÛÜ
õ Aò óKä{á¹çÜ èÜç¼ã«àWíÁÜ Þáã&ã«ÛÀܱývÜ)ådáÞÀß«ÜçIývÜåèvÜßß«àdß&ß«åãÜ]àèvãÛÀܯãß4åäèÀäè
4ÜývÝÀäßä{éåì6ßä{á
áÜãWò{âöàß]å ÀëÜç8 ÷ 'èÀäÈãÜ=è6ÞÀýIíÁÜߥàdâàíáÜßæ-åãäàdèá4!ó
"#$ò%ó1H ø (
A
: '=*) & L
=+
ò<,= D
ó
I ò
ó
#
àãܾãÛåãFèÀàOÝÀßàíåíäÈìäÈãêµçä{áãß«äÈíÞãäàdèDåÝÀÝÁÜådß«áoÛÀÜß«Üô A òóvä{áoå ÀëÜç è6ÞÀýtí'Üßvâöàdßå
Ýådßã«äéÞÀì{åß#é4ÛÀàdä{éÜ
àâ,µådèç|âöàß]åtÝåßãä{éÞìådß&ã«ß«ådäèÀäèIáÜã"-7<,= /.ô
D L
=
Ú¥ÛÀÜ !Þådèã«äÈãê # L =
ò< = D
ó ä{á=éådìÈìÜ)çãÛÀܼìàá«áô àdß
ãÛܼéådáÜtçÀÜá«éß«äÈíÁÜç¾ÛÀÜß«Ü÷räÈã`éådè¾àdèìÈê
ã«ådõdÜWãÛÀܼæ-åìÞÀÜ)áS åèçOøô à-ÿ é4ÛààáÜIáàývÜ0ÃáÞé4ÛÃã«Ûå-ã S12031 øôNÚ¥ÛÀÜèÃâöàdß=ìàá«áÜáiã4åõ6äè
ãÛÜáÜ`æ ådìÞÀÜá÷ÿiä:ã«ÛFÝßàíåíÀäìäÈãêFø450Á÷ÀãÛܱâöàììÈà-ÿiäèWí'àÞÀèçÛÀàdì{çÀá=ò,ðKåÝÀèÀäõÁ÷røùùúó
Ú¥ÛÀÜ¥ð ÅçÀäÈývÜèáäÈàè`ä{áå±ÝÀß«àdÝÁÜßãê`àâ8å
áÜã?àdââöÞÀèéãäàdè', á -
ò ó .Iò,å ådäÈèr÷ ÿ#Ü¥Þá6 Ü Uåáå ÜèÀÜßä{é¹áÜã
àâÝådß«ådývÜãÜß«áå¼é4ÛÀàä{éÜ
àâ,µáÝÁÜéäÜá¥åtÝåßãä{éÞÀì{åß¹âöÞÀèéã«äàdè'ó÷åèçTéådèFíÁÜNçÜèÀÜçFâöàßiæ ådßäàÞá
éì{åááÜá¹àâ@âöÞÀèéã«äàdè
ô ]Üß«Ü=ÿ¹Ü=ÿiäìì'àèÀìê|éàèáä{çÜß&âöÞèéã«äÈàèá#ãÛ'å-ã¯éàßß«ÜáÝÁàdèç¼ã«à¼ãÛÀÜ
ãÿ#à ;éì{ådá«á
ÝåããÜßèß«Üéà dèÀäÈã«äÈàè|éåáÜd÷áàIãÛåã
ò <PD
ó ? - `
ø Dø .E<"D
&ô à-ÿ äÈâå däædÜèFáÜãiàP â :ÝÁàdäèã«á¥éådè
íÁÜvìådí'ÜìÜçÃäèVåìì?ÝÁàá«áäíÀìM Ü ÿ#å êá÷åèç âöàdß`Ü)ådé4ÛVì{åíÁÜìä è À÷råývÜýtíÁÜß=àâ#ãÛÀÜ|áÜ ã -
ò ó .Féådè
íÁÜ=âöàdÞÀèçÿiÛÀä{é4ÛTéàßß«Üéãìêvådá«& áä dè'á&ã«ÛÀàáÜ=ì{åíÁÜì{á÷ÿ#Ü`áå êvã«Ûå-ãiã«Ûå-ã¯áÜã]àdâÝÁàdäèã«á#ä{ á G M
M!
í6ê
ãÛ'å-ãNáÜã`àâ&âöÞèéã«äÈàèáôtÚ¥ÛÀܼð çÀäÈývÜèáäÈàèTâöàdß
ãÛÜ|áÜã=àdâKâöÞÀèéãäàdè' á -
ò ó .|ä{á=çÜ èÜçïådá
ãÛÜ
ý|å-ëäýtÞÀý è6ÞÀýtí'Üß&àdârãß4åäèÀä è `ÝÁàdäèã«áKãÛåã¥éåèoíÁÜ=áÛåãã«Üß«Üçví6ê -
%ò
ó .ô àdãܱãÛåã÷ÀäÈâãÛÀÜ
ð
çäývÜè'áäàdètä{á ÷6ãÛÀÜèvãÛÜܱ߫Üë6ä{áã«á¹å-ã¹ìÜ)ådáã&àdèÀÜ
áÜã#àdâ ÝÁàdäèã«á&ãÛ'å-ãiéåèoí'Ü
áÛåãã«Üß«Üç8÷íÀÞã¥äÈã¹äè
dÜèÀÜß4åì8äÈãiÿiäìÈ< ì8èÀàdã]í'Ü=ã«ßÞܱã«Ûå-ã M"CM#`áÜãiàâ Ý'àäèã4ái < éådèFíÁÜNáÛ'å-ãã«Üß«Üç8ô
<
$%'&() M *,+.-4I(* G/ $10 M 2M 43#!5 M )576 M G8* m:9
î6ÞÀÝÝ'àáܯãÛåã¹ã«ÛÀÜ
áÝåéÜ]äè|ÿiÛÀä{é4Û¼ã«ÛÀÜ
çå-ã«åWìäÈæÜiä{á m ÷Àåèç¼ã«ÛÀÜ=áÜ4 ã -
ò%ó
.`éàdèáä{áã4áKàâràßäÜèã«Üç
á ãß4åädÛãKìäèÀÜá÷áàNã«Ûå-ã#âöàdߥå äÈæÜè¼ìäèÀÜ÷6åìì'ÝÁàdäèã«á&àdè|àèÀܱáä{çܱådßܱådá«áädèÜçIãÛÀÜ=éì{åáá¯ø÷åèç|ådìì
ÝÁàdäèã«á±àdèÃã«ÛÀÜtàãÛÜß`áä{çÜd÷'ã«ÛÀÜvéì{ådá«á`ødôNÚ¥ÛÀÜIàßäÜèã4å-ã«äÈàèTäá
áÛÀà-ÿièÃäè ädÞÀß«Ü|øtí6êTådè ådßß«à-ÿ=÷
áÝ'Ü)éäÈâöê6äÈ
è
àè¼ÿiÛÀä{é4Û¼áä{çÜ¥àâ8ãÛÜ]ìäÈèÜ#ÝÁàdäèã«á?ådßÜ¥ã«à`í'ܱådá«áä dèÜçWãÛÀÜ]ì{ådí'ÜìøôP ÛÀäìÜ#äÈãKä{áÝÁàá«áäíÀìÜ
ãà 'èçãÛÀß«ÜÜ`ÝÁàdäèã«á#ã«Ûå-ã
éåèíÁÜWáÛ'å-ãã«Üß«Üçí6ê|ã«ÛÀä{áiáÜã¯àââöÞÀèéãäàdèá÷äÈãiä{á¥èÀàã¯ÝÁàá«áäÈíìÈܱãà 'èç
âöàdÞßôKÚ¥Û6Þá¥ã«ÛÀÜ`ð çäývÜèáäàè|àâ@ãÛÀÜNáÜã]àdâàßäÜèã«ÜçvìäèÀÜá#äè m ä{á#ãÛÀß«ÜÜdô
$IAG M GJICKM"GJM I 5 I(* G * m B I ICGJM #I M I $ M 5 I(* G G I( +(*
M * M 5 I(* G M G () M M2# I(M M2 #!5AM)56 M G 7 I 6 #1 $ M 5 CI G2 I
"CM I GI * M MJK * *,+@5AI* G
M86 * M 6 # * M 5 M7 M %
*ò ),D
ó213''ò,áäÈèrò%4)Áóó D5),D
? m I ò@B6ó
) = H øS87
=
DRF"H ø :9;9;9DK: I
D ò2úó
Ú¥ÛÀÜè
ò% ó+äædÜá&ã& Ûäá¥ì{ådí'ÜìÈäèIäÈâ
¹é4ÛààáÜ Vã«à¼í'Ü
;Á$), > ¦ 0 -² 0 " 3 1.*@12+-41 3 3-0 1@(d?*2+-41,1.2±(§ *2" 3 d2 *,² " 12&" 3)Â 3 " 12&³´)" !¥ 3 2* " 40 35
1.4
1.2
1
VC Confidence
0.8
0.6
0.4
0.2
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
h / l = VC Dimension / Sample Size
;Á$),
> ³´F 043 Â ) 3 K"$*! 043-0 1 043 "$K" 3
/I K 56*M G
¹àdè'áä{çÜßKãÛÜ ã«ÛFèÀÜ)åß«Üáã¹èÀÜädÛ6í'àÞÀß¹éì{ådá«áäÜß÷ÿiä:ã«ÛH ødôKÚ¥ÛÀä{á¹áÜãiàââöÞÀèéã«äàdèá#Ûåá&äèèä:ã«Ü
ð çäývÜèáäàdèåèç Üß«àvÜývÝÀäß«äéåì8ß«äáõÁ÷'áäÈè'éÜNåè6êFèÞýIíÁÜß]àdâ?Ý'àäÈèã4á÷Áìådí'ÜìÜçåß«íÀäÈãß4åß«äÈìêd÷ÿiäìì
íÁÜWáÞ'ééÜ)ááâöÞÀììê|ìÜådßèÀÜ)çoí6êoãÛÀÜWådìdàdß«äÈãÛÀýfò,ÝÀß«à-æä{çÜ)çoèà¼ãÿ¹àvÝÁàdäèã«áiàdâ?àdÝÀÝÁàáäÈãÜ`éì{ådá«á¥ìäÜ
ß«ädÛã
àdèÃãàdÝ àâ&Ü)ådé4ÛÃàdãÛÀÜß4óô=Ú¥Û6Þá±ã«ÛÀÜIíÁàdÞèçÃÝÀß«à-æ6ä{çÜáièàoäèâöàdß«ý|å-ã«äàdèô
;è¾â,ådéã÷8âöàdß`åè6êTéì{ådá«áä 'Üß
ÿiäÈãÛ|ä è 'èÀäÈãܯð çäývÜèáäàdè÷ã«ÛÀÜ
í'àÞÀèçFäá&èÀàã¥ÜæÜèFæ ådìä! ç )ô,¯à-ÿ¹ÜædÜß÷ÜædÜè|ã«ÛÀàdÞ ÛvãÛÜ
íÁàdÞèç
ä{á&èàãiæ-åìä{ç8÷èÀÜ)åß«Üáã#èÜädÛ6íÁàdÞÀß#éì{ådá«áäÜß«á&éåèáã«äìÈì'Ý'Üßâöàßýþÿ#Üìì.ô?Ú¥ÛÞ'á&ã«ÛÀä{á 'ß«áã#ÜëÀåývÝÀìÜ]ä{á#å
éådÞãäàèåß«êtã«åìÜ ä ê Içà6Üá¥èÀàdã dÞådß«ådèã«ÜÜ
Ý'à6àßiÝ'Üßâöàßý|ådèéÜdô
è 'èÀäÈãÜ éådÝådéäÈã2
ã $á&âöàdììà-ÿÅãÛÀܯãäývÜiÛÀàèÀàdÞÀß«ÜçIã«ß«åçäÈãäàdètàâÞÀèçÀÜß4áã4åèçäè=ã«ÛÀäèáKíêIãß«ê6äè=ãàNíßÜ)åõNã«ÛÀÜýT÷
Ü
å è'ç áÜÜväÈâ¹ÿ#ÜvéådèïéàdývÜtÞÀÝVÿiä:ã«ÛïåUéì{ådá«áäÜß]âöàdß`ÿiÛäé4Û¾ãÛÀÜvíÁàdÞèçïäá`áÞÀÝÀÝÁàáÜç¾ãàUÛàdì{ç8÷ríÀÞã
ÿiÛÀä{é4ÛæäàìåãÜ)á?ãÛÜ`í'àÞÀèç8ô VÜ`ÿ¥åèã¥ãÛÀÜ`ìÜâ{ã]Û'åèçáä{çÜ=àdâ 0 !ô]ò%Eó¹ãàvíÁÜNådá¥ì{ådß Ü±ådá¥ÝÁàá«áäÈíìÈÜd÷
åè'çoã«ÛÀÜ`ß«ä Ûã#ÛådèçTáä{çÜ
ã«àví'ÜNåáiáý|ådìÈìåá¥Ý'àááäíÀìÜô&î6àvÿ#Ü
ÿ¥åèã]åtâ,åýväìê¼àâ?éì{ådá«áä 'Üß4áKÿiÛÀä{é4Û
däæÜáãÛܯÿ#àdß4áãKÝÁàá«áäíÀìÜiådéãÞådìß«ä{áõIà, â SI ú6' ÷ Üß«à`ÜývÝäÈß«ä{éådìß«ä{áõIÞÀݼãàtáàývÜ¥èÞýIíÁÜßKàdâãß4åäèÀä è
àdí'áÜßæ-å-ã«äàdèá÷råèç ÿiÛÀàáܼð çäývÜèáäàèUäá=Ü)ådáêTãàUéàývÝÀÞã«Ü¼ådèç äá
ìÜá«á±ãÛådè :=ò2áàãÛ'å-ã`ãÛÜ
íÁàdÞÀèç ä{á¯èàdèÃã«ßäæ6ä{åìöó4ô èïÜëÀådývÝÀìÜNä{á¯ã«ÛÀÜIâöàdììà-ÿiäÈ è À÷'ÿiÛäé4Û
±éådììã«ÛÀÜ èÀàã«ÜíÁà6àdõÃéì{åááä Üßô
Ú¥ÛÀä{á¼éì{ådá«áä 'Üß¼éàdè'áä{áã«áIàâ`å èàãÜí'à6àõµÿiäÈãÛDÜèÀà Þ dÛ ßà6àdý ãàVÿißäÈã«ÜTçÀà-ÿièOã«ÛÀÜÃéì{ådá«áÜáIàD â
ãß4åäèÀä è vàíáÜß«æ åãäàdè'á÷ÿiÛÜß«Ü 1 :;ô Ààdß=åììáÞÀí'áÜ !ÞÜèã±Ýåãã«Üß«èá÷8ã«ÛÀÜtéì{ådá«áä Üß]áäývÝÀìêoá«å êá
ãÛ'å-ãIådìÈì?Ý'å-ãã«Üß«èáNÛ'å ædÜtãÛÀÜFáådývÜ|éì{ådá«áôTî6ÞÀÝÀÝÁàáÜoåì{áàoã«Ûå-ãWãÛÀÜoçå-ã«åÃÛå ædÜvådáNý|åè6êÃÝÁàáä:ã«äædÜ
ò L H 7Iø)óiåá]èÀ Ü åãäædÜ|'ò L!H `ø)óiÜëÀåývÝÀìÜ)á÷'ådèçTã«Ûå-ã±ãÛÀÜtáådývÝÀìÜáiåß«ÜWé4ÛÀàáÜèUß«ådèçàývìÈêdô¥Ú¥ÛÜ
èÀàdãÜíÁà6àdõ|éìåááä Üß¹ÿiäìÈìÁÛå æÜ Üß«àIÜývÝÀäßä{éådìÀß«äáõ¼âöàdߥÞÀÝFã& à yàíáÜßæ-åãäàdèáû SI ú`ã«ß«ådäÈèäÈ
è NÜß«ßàß
âöàdßNådììáÞÀíá Ü !ÞÀÜèãNàdíáÜß«æ-å-ãäàèá> û SI$úåéãÞ'åì?Üß«ß«àdß÷åèç ð çÀäÈývÜèáäÈàè
< H
¾ôFîÞÀíáãäÈãÞã«ä è
ãÛÜáÜ`æ ådìÞÀÜá#ä, è 0 !'ôi%ò Eó÷ÀãÛÀÜ`íÁàdÞèçí'Ü)éàdývÜ)á
D
B : 1 ìÈèrò : = Uó 7 4ø
òø =ETóìÈèrò0A=
Bó ò,üó
0 äÈâ
ÿiÛÀä{é4Ûäá¥éÜßã«åäèÀìê¼ývÜã¥âöàß]åìì o
ò6ó H
ÜëÝ
7 #
1 ø
D 1 ò =$:2ó D
S 1N1 ø ò,ùó
ÿiÛÀä{é4Ûäá¹ã«ßÞÀÜd÷'áäÈè'éÜ
ò6ó&ä{á¥ývàèÀàã«àdèÀä{éiäèéßÜ)ådáäÈèÀ÷åèç
òH ø)ó H SI LEÀô
('& $ $ 6 G * K. ( I
;Á$),
> >@*,1.±*2 (-*2;1.* 0 99: 3 1," 40 3 * 0 . 2¯(§]³´F)" !¥ 3 2* " 40 35
VÜvÛå ædܼèÀà-ÿ[ì{åä{çTãÛÜ dß«àdÞÀèçÀÿ¹àßõTèÀÜ)éÜá«á«åß«êTãàTí'Ü däèïàdÞÀß=ÜëÝÀìàß«åãäàdèÃàdâ#áÞÀÝÀÝÁàdßã`ædÜéãàß
ý|ådé4ÛäÈèÜáô
<>=-9
7 I7 ø âöàdß L
=H 7Iø òøSó
<>=-9
7 1 `
ø âöàdß L
=H ø
` òødø ó
ò
L = < = 9
27ó Dø;S EF òøó
ãÛ'å-+
ç ådß ÜIÝåß4åììÜìKòöãÛÀÜêU
Ûå ædÜW 7 ãÛÀܼáådývÜNèÀàßý|ådì{ó¥ådèçTã«Ûå-ã=èÀàFãß4åäèÀäè ¼Ý'àäèã4á]â,ådìì
H H
ã
# åè',
íÁÜãÿ#ÜÜèFãÛÜýTôKÚ¥Û6Þá¥ÿ¹Ü=éådè 'èç|ã«ÛÀÜ=ÝåäߥàâÛ6êÝÁÜß«ÝÀì{åèÜáKÿiÛäé4Û däædÜ)áKã«ÛÀÜ=ý|å-ëäýIÞý[ý|ådß äÈè
í6ê|ýväèÀäývä ä è -
,J-÷'áÞ í Üéãiãàvéàèáã«ß«ådäèã4á=òø ó4ô
Ú¥Û6Þá=ÿ#ܼÜëÝ'Ü)éã=ãÛÜ|áàìÈÞÀãäàdèTâöàßNå|ãê6ÝÀä{éådìãÿ¹àTçäývÜèáäàdèådì@éådáÜWã«àÛ'å ædÜIãÛÀÜtâöàdß«ý áÛÀà-ÿiè
äè ädÞÀß«ÜúôDÚ¥ÛÀàáÜFãß4åäèÀäè Ý'àäèã4áNâöàßvÿiÛÀä{é4ÛÅãÛÜÜ !ÞåìäÈãêVäÈè)01!'ôgòøóNÛÀàì{çÀáTòöä.ô ÜdôµãÛàáÜ
ÿ ÛÀä{é4ÛTÿiäèçÞÀÝUìêäè¼àèàèÀÜ`àâã«ÛÀÜNÛ6ê6Ý'ÜßÝÀì{ådèÀÜá. # ÷/ ó÷8ådèçÿiÛÀàáÜ`ßÜývà-æ-åìÿ#àdÞÀì{çé4ÛåèÜ
i
ãÛÜtáàdìÞã«äÈàèoâöàÞÀèç8÷ådßÜIéåììÜçTáÞÀÝÀÝÁàdßã±æÜéã«àdß4áûÁãÛÜêÃåß«ÜNäèçä{éå-ãÜ)çäÈè @äÞÀßÜWú|í6êFã«ÛÀÜtÜë6ãß4å
éäß4éìÜáô
VܼÿiäÈììèÀà-ÿ áÿiä:ã4é4ÛUãàTå!å ß«ådèdä{ådèFâöàßýtÞÀì{å-ã«äàdèàâKã«ÛÀܼÝÀßàíÀìÜýTô`Ú¥ÛÀÜܼ߫åß«ÜNãÿ#àFß«Üåáàèá
âöàdß
çàdäè¼ã«ÛÀä{áô
Ú¥ÛÀÜ ß4áã±ä{á¥ãÛ'å-ã¯ã«ÛÀÜIéàdèáãß4åäèã«áWòøó¥ÿiäììrí'ÜIßÜÝÀì{ådéÜçí6êTéàdèáãß4åäèã«á¥àèUãÛÜ
å dß4å
è Ü¥ýIÞÀìÈã«äÈÝìÈäÜß«árãÛÀÜý|áÜìÈæÜá÷ÿiÛÀä{é4Û¼ÿiäÈììíÁܯýtÞé4Û¼ÜåáäÜßã«àNÛåè'çìÜô?Ú¥ÛÀܯáÜéàèçtä{áãÛåã¹äè
ãÛäá¹ß«Üâöàdß«ýtÞÀì{å-ãäàèWàâã«ÛÀÜ
ÝÀßàíÀìÜýT÷ã«ÛÀܱãß4åäèÀä è NçÀåã«åNÿiäììÀàèÀìê¼ådÝÀÝÁÜåß
òöäèvã«ÛÀÜ
åéã«Þåì'ãß4åäèÀä
è
åè'çtãÜáã#ådì dàdß«äÈãÛÀý|á4óäè¼ã«ÛÀÜ]âöàßý àârçàdã&ÝÀß«à6çÀÞéã4áKíÁÜãÿ#ÜÜèoædÜéãàß«áôÚ¥ÛÀä{á?ä{á¹åNéß«Þéä{ådìÝÀßàÝ'Üßãê
ÿiÛÀä{é4ÛÿiäÈìì8åììà-ÿ Þá#ãà ÜèÀÜß«ådìä ÜiãÛÜ`ÝÀßàéÜçÞÀܱ߫ãà¼ãÛÜ`èÀàdèÀìäèÀÜ)åߥéådáܼò2î6Ü)éãäàè B6ó4ô
Ú¥Û6Þá÷iÿ¹ÜTäèãß«à6çÀÞéÜTÝÁàáä:ã«äædÜ0@å ß«ådèdÜýIÞì:ã«äÝÀìäÈÜß«á= D F H ø
D;9:9;9JD :;÷¥àdèÀÜâöàdßoÜådé4ÛDàâ
ãÛÜ
äèÀÜ !Þådìä:ãêFéàdè'áã«ß«ådäÈèã4áNòø7dó4ô iÜ)éådìÈìãÛåã±ã«ÛÀÜtßÞÀìÜNä{á¯ãÛåã±âöàß
éàdèáãß4åäèã«á]àâKã«ÛÀÜWâöàßý10J=RS÷
w
H2
-b
|w| H1
Origin
Margin
;Á$), > Î " 3 @*2² .41," 3 i+§²d2² $ 3 *9 0 1,+-K*.;²-24( $K*. 5Á< +-?*, ² ² 0 2 1 ° ;1 0 2*2;"$.;$ 5
ãÛÜïéàdèáãß4åäèã|Ü !ÞåãäàdèáFåܾ߫ýtÞÀìÈãäÝÀìäÜçOíê 5AI G *" M @å dß4åèdÜÃýtÞÀìÈãäÝÀìäÜß4áIådèç áÞÀíã«ß«åéãÜ)ç
âöß«àdý[ãÛÀܱàd í Ü)éãäæܹâöÞèéã«äÈàè÷ãà`âöàßý[ãÛÀ
Ü å dß4åèäådèôÀàß&Ü !Þ'åìäÈãêNéàdè'áã«ß«ådäÈèã4á÷ã«ÛÀÜ@å ß«ådèdÜ
ýtÞÀìÈãäÝÀìäÜß4á?åß«Ü
Þèéàèáã«ß«ådäèÀÜç8ô?Ú¥ÛÀä{á däæÜá å ß«åd è dä{ådè
1 ø
, ( & ='L
=ò'<,=
9
7 ó 7 ( & "= Eó
òø
= ) =*)
# #
VÜTýIÞáãIèà-ÿ ýväèÀäýväÜ ÿiäÈãÛ ßÜ)áÝÁÜéãWã«à
!D-)÷¹ådèçOáäýIÞÀìÈã4åèÀÜàdÞáìêUßÜ !ÞäÈß«Üvã«Ûå-ã¼ãÛÜ
çÜßäæ-å-ã«äÈæÜá
àâ ÿiä:ã«ÛÅßÜ)áÝÁÜéãNãà åììKã«ÛÀÜ "=¯æ-åèäáÛ÷?åìì&áÞÀíÜ)éãNã«àÃãÛÜoéàèáã«ß«ådäèã4á =&GS
òöìÜã $áKéådìÈì6ã«ÛÀä{áÝåßãä{éÞÀì{åßáÜãKàâéàdè'áã«ß«ådäÈèã4á
# ó4ô à-ÿ ãÛÀä{áä{á?å`éàè6ædÜë !Þ'ådçß4å-ã«äé¹ÝÀß«àß«ådývýväÈè
ÝÀß«àdíÀìÜýT÷áäèéÜoãÛÀÜàd í Ü)éã«äÈæÜvâöÞÀèéãäàdèOäáIäÈã«áÜìÈâ]éàèæÜë8÷KåèçµãÛÀàáÜFÝ'àäèã4áWÿiÛäé4Ûµá«å-ã«ä{áâöêïãÛÜ
éàèáãß4åäèã«á]ådìáàtâöàdß«ýyåoéàè6ædÜëTáÜãtò2åè6êTìÈäèÀÜ)åß]éàdèáãß4åäèã¯çÜ 'èÀÜá
å|éàè6ædÜëUáÜã÷rådèçÃå|áÜã
àâ
áäýtÞÀìÈã«ådèÀÜàÞá]ìäèÀÜådß
éàdèáãß4åäèã«á
çÜ èÀÜá`ãÛÀÜväèãÜß«áÜéãäàdèÃàd â éàè6ædÜëïáÜã«á÷@ÿiÛÀä{é4Û¾ä{á`ådìáà
å¾éàdè6ædÜëµáÜã4ó4ôµÚ¥ÛÀä{áIývÜådèá`ã«Ûå-ã¼ÿ¹ÜTéådèµ Ü !ÞÀäæ-åìÜèã«ìê¾áàìædÜvã«ÛÀÜoâöàììÈà-ÿiä è çÞåì UÝßàíÀìÜý
K K4 M ÷@áÞí Üéã±ã«àFã«ÛÀܼéàdè'áã«ß«ådäÈèã4á]ã«Ûå-ã
ãÛÀÜ ß«åçäÜèã¯àâ ÿiäÈã«Û¾ß«ÜáÝ'Ü)éã
ã à
fåèç
æ-åèÀä{áÛ÷åèçUáÞÀ í Ü)éã`åì{áàtã«à|ãÛÜtéàdèáãß4åäèã«á#ã«Ûå-ã
ãÛÀ Ü = SÃòöìÜã $á±éåìì *( Ýådßã«äéÞÀì{åß]áÜã
àâ
éàèáãß4åäèã«
á ó4ôTÚ¥ÛÀä{á`Ýådßã«äéÞÀì{åß=çÞådì?âöàdß«ýIÞìåãäàdè¾àâ#ãÛÀÜ|ÝßàíÀìÜýyä{áNéådììÈÜ)çUã«ÛÀ Ü ïàìÈâöÜ|çÞåì
*ò ìÜã4é4ÛÀÜß÷?øù'ü óô
ã
Ûåá]ã«ÛÀܼÝÀß«àdÝÁÜßãêFã«Ûå-ã
ãÛÜIý|åë6äýtÞÀýYàd â ÷áÞÀ í Ü)éã
ãàTéàèáã«ß«ådäèã4 á ÷
Ü
F÷ ¯ådèç &÷Áåá¹ãÛÀÜ`ýväèÀäýtÞÀý[àd â ¹÷áÞ
àééÞÀß«áiåã¥ãÛÀÜNá«åývÜ
æ-åìÞÀÜá#àdâãÛ+ í Üéãiãàvéàèáã«ß«ådäèã4á
# ô
iÜ !ÞÀäßäèWãÛåãiãÛÀÜ ß«åçäÜèã#àdâ
ådèç]æ-ådèÀä{áÛ däædܱãÛÀÜNéàdèçÀä:ã«äàdèá
ÿiäÈãÛß«ÜáÝ'Ü)éã¥ãà Y
H ( = L = < = òøJB6ó
=
( = L = H SI òø)úó
=
M
$%
$ G $ $ M $ I IAG
Ü `ådßÞ'á
Ú¥ÛÀ
Û =ÞÀÛÀ è Ú@Þé4õdÜßUò
`Ú]óvéàdèçäÈã«äÈàèá¼ÝÀì{å ê åOéÜèã«ß«ådì]ßàìÜUäÈè íÁàã«Û ãÛÀܾãÛÀÜàdß«ê ådèç
ÝÀß4ådéãä{éÜtàâ¥éàdè'áã«ß«ådäÈèÜçÃàÝãäýväåãäàdèrô Ààdß=ã«ÛÀÜvÝÀß«äý|åìÝÀßàíÀìÜýfådí'à-æÜ÷rã«ÛÀÜ
`Ú éàdèçäÈã«äÈàèá
ý|å êvíÁÜNáã4å-ãÜ)ç ò ìÈÜã«é4ÛÀÜß÷rø)ùd ü ó
5H ( "= L
=*)+= H S H ø :9;9:9JD
D òø dó
=
2
H ( = L = H S òøüó
=
-9
7ó? ø+
ò
L
= '<>= S FPH ø :9;9:9JDK:
D òøùó
=
S EF ò Só
=«òL
=«ò
9 <,=F
7 ó D
ø)ó H S EF ò ø ó
)N= 2
H Nòø4 ?= Àó 9 KòO
O
A
7 ø)ó
7 ?= O
AO
7 ø
ò
ó
`ø `ø
<
# Ñ
H ò
E D
D
ó
Ñò
`ø
`ø ó
<
H
D
E D
Ñò
`ø `ø
ó ò LEó
<
8
H
D
D
E
$±èÀÜTéàèá Ü !ÞÀÜè'éÜ|àdâiãÛÀÜTáê6ývývÜãß«ê¾ä{á`ã«Ûå-ãtãÛÀÜTåèdìÜví'Üãÿ¹ÜÜèDåè6ê Ý'åäßWàdâ¯ædÜ)éãàß«áNäáWãÛÜ
á«åývܼò,åè'çFä{á#Ü !Þådì8ãà|ådß«ééàáò`ø =$Oóó
<,= H A ò CB6ó
â?ãÛÜNßÜ)áÞì:ã«äèv áàìÈÞÀãäàdèçà6Üá]äèTâ,åéã±á«å-ã«äáâöê = S.EF«÷ã«ÛÀÜèUÿ#ÜNÿiäììÛå ædÜ=âöàÞÀèçTã«ÛÀÜ ÜèÀÜß«ådì
áàdìÞãäàè÷áäèéÜ=ãÛÀÜIådéã«Þåì@ý|å-ëäýIÞý àdâ ÿiäìÈìrãÛÀÜèUìÈäÜ=äèãÛÀÜWâöÜådáäíÀìÜ=ßÜ d äàdè÷ÀÝßà-æ6ä{çÜçoãÛÜ
Ü ! Þådìä:ãêÃéàdèáãß4åäèã÷ 01' ! ôÅòø)úó4÷ä{áNåì{áàývÜãô ;
èÅàdß4çÜß=ãàÃäývÝÁàáÜIã«ÛÀÜ|Ü ! Þ'åìäÈãê¾éàèáã«ß«ådäèã=ÿ#Ü
äèãß«à6çÀÞéÜ=åèUådççäÈãäàdèådìE@å ß«ådèd ܱýIÞÀìÈã«äÈÝìÈäÜßô¹Ú¥Û6Þá¥ÿ#Ü`áÜÜõvã«àvý|å-ëäývä Ü
B #
1 ( = ø ( #
B
= =
B #
( = L = D ò
dó
= ) #
= )
# *)
=
#
ÿiÛÀÜßÜ=ÿ#Ü`Ûå ædÜ
äèãß«à6çÀÞéÜ)çvãÛÜ]Üá«áäådè
òl ?ó = 7NL
=H ø EF ò ùó
="H ; ø 7* L
= D
7 ; 4ø L
=
D ò%E
Só
ÿiÛÀÜßÜ åèç
åܱ߫ÞÀèõèà-ÿièáô1.&ìÞäÈè`ãÛÀä{á#âöàßý äè 01!'ô]ò d ùó däædÜ)á
7 ø 7 ;
7 ø4L
; D ;
D ò%EÀø ó
O L = =
O O D H
A
ÿiÛÀÜßÜ6Uä{á¥çÜ 'èÀÜçí6ê
#
I1 (
B
L
= I ò%E
ó
= )
#
Ú¥Û6Þá
7 H
A
$O
ò'O7 ø)ó
ò%E
Eó
ådèçTáÞÀí'áã«ä:ã«ÞãäèIãÛäá¥äèãàIãÛÜ`Ü !ÞåìäÈãêvéàèáã«ß«ådäèã 01!'ô]òø údó&ãà 'èç ÷ äÈæÜá
O
;ø4
O
;ø7 ò%ELB6ó
H
A O ò 7 ø)ó O 7 Fø D D H
A ò
'O 7 ø ó O 7 ø D
ÿiÛÀä{é4Û,däædÜ)áKâöàߥãÛÀÜNáàdìÞã«äàdè
O
; 4ø L
= ò%Eúó
= H
A 'O ò 7 ø)ó O 7 Aø D
ìáàÀ÷
ó
òl ø4
L
= ò%Eó
=PH
O 7 ø
I
L
¯ÜèéÜ
, H
B
( # = L = L < = 9 < H l
= )
#
( =$; ø4 L
=
B # B
( # O
ø4 ; O7 øAD
ò%E'dó
O 7 øAD
H H = H
= ) =*)
A
# #
àãÜ=ã«Ûå-ãiã«ÛÀä{á¥ä{á¹àèÀÜ`àâ@ãÛÀàáÜ`éådáÜá¥ÿiÛÀÜßÜ
ãÛÜ å ß«ådèdܱýIÞì:ã«äÝÀìäÈÜß UéåèßÜý|åäè|ÞÀèçÀÜãÜß
ýväèÀÜçÅò2åìÈãÛàdÞÛ¾çÜã«Üß«ýväèÀäèväÈã±ä{á]ã«ßäæ6ä{åìöó4ôïÜtÛå ædÜIèÀà-ÿ áàdìædÜ)çãÛÀÜtÝÀß«àdíÀìÜýT÷8áäèéÜIådìÈì@ãÛÜ
"=¯åß«Ü|éìÜ)åß«ìÈêTÝÁàáäÈãäædܼàdß Üß«àÅòöäèVâ,ådéãNãÛÜ =¯ÿiäìÈì?àèÀìêÃíÁÜ ÜßàTäÈâ]åììãß4åäèÀäèÝÁàdäèã«á`Ûå æÜ
ãÛÜNáådývܱéì{åáá4ó4ô àdãÜ
ãÛ'å-ã
çÜÝÁÜè'çÀáiàdèìÈêvàèoã«ÛÀÜ $ KCM¹àâ@ÝÁàáäÈãäædÜtòöèÀÜ åãäædÜ óKÝÁàdì{åß«äÈãê
ÝÁàdäèã«á÷åèçtèÀàdã?àdè¼ÛÀà-ÿµã«ÛÀܯéìåááì{åíÁÜì{áåß«Üiådá«áä dèÜç`ã«à=ãÛÀÜiÝÁàdäèã«áäè 01!'ôiò
ó4ô?Ú¥ÛÀä{áäá?éìÜåß«ìê
èÀàdãiãß«ÞÀÜ=à$ â
äÈã«áÜìÈâ;÷ÿiÛÀä{é4ÛFä{+
á däædÜèoí6ê
( # ; L
=>
O B
ò%Edüó
A òO 7 O 7 Fø D <,=
H
ø)ó = )
#
Ú¥ÛÀÜ`ý|ådß äè÷ H * =
÷'äá¹ã«ÛÞ'á däædÜèoí6ê
ò%Edùó
A
¼òø4Dò
= ò'O 7
H I
O ø óó ó
MG - GM
$±èéܱÿ#Ü=Ûå ædܯãß4åäèÀÜçFåtîÞÀÝÀÝÁàdßã¥ð?Ü)éã«àdß]ñÃåé4ÛÀäèÀÜ÷6ÛÀà-ÿ éåèÿ¹Ü
ÞáÜ
äÈã 0ïÜ`áäÈývÝìÈê¼çÜãÜß«ýväèÀÜ
àdèvÿiÛäé4Û|áä{çÜ]àdâÁã«ÛÀܱçÀÜéä{áäàdè¼í'àÞÀèçÀådßêTò{ã«Ûå-ã#Û6êÝÁÜß«ÝÀì{åèÜiìÈê6äè=Ûådì:âÿ¥å êWíÁÜãÿ#ÜÜè # åè'ç-
åè'çÃÝådß«ådìÈìÜìrãàFãÛÜýoó±å äÈæÜèUã«Üáã`Ýå-ããÜß«è < ìäÈÜ)á¯ådèç åááä dèTã«ÛÀÜvéàdß«ß«ÜáÝ'àèçäè|éì{åáá±ì{åíÁÜì.÷
â < ãàvíÁ
ä.ô Üdô?ÿ¹Ü
ã«ådõdÜ
ãÛÜNéì{ådá«á#à Ü OKò
9< 7ó4ô
M"I( & M 5 6 M GM
Ú¥ÛÀÜvåíÁà-ædÜtåì dàßäÈã«ÛÀý âöàß=áÜÝ'åß4åíÀìÜWçÀåã«åÀ÷rÿiÛÀÜèVåÝÀÝìÈäÜ)çTãàèÀàdèáÜÝådß«ådíÀìÜNçÀå-ã4å÷ÿiäìì 'èç¾èà
âöÜåáäíÀìÜ]áàdìÞã«äÈàè rã«ÛÀä{áKÿiäììÀí'Ü
Üæ6ä{çÜèéÜ)ç¼íêIãÛܱàíÜéãäædÜ¥âöÞÀè'éãäàè¾òöä.ô ÜôãÛÀÜ=çÞådì+ådß4åèä{åè'ó
dß«à-ÿiäè Iådßíä:ã«ß«ådßäìêWì{ådß ÜôKî6àvÛÀà-ÿ éåèÿ¹Ü`Üë6ãÜè'çoã«ÛÀÜáÜ`ä{çÜåá¹ãàvÛådèçìÜ=èÀàdè;áÜÝådß«ådíÀìܯçÀå-ã4å
VÜWÿ#àdÞÀì{çTìäõdÜ
ãàFßÜìåëoã«ÛÀÜIéàdèáãß4åäèã«á`ò ø Só¯åèçOòødø ó4÷8íÀÞã
àdèÀìêÿiÛÀÜèÃèÜéÜ)áá«åß«ê÷'ãÛåã±ä{á÷'ÿ#Ü
ÿ#àdÞÀì{ç|ìäõdܯãà¼äèãß«à6çÀÞéÜ=åIâöÞÀßã«ÛÀÜß±éàáãNòöä.ô ÜôKåèäèéß«ÜåáܱäÈèFãÛÜ`ÝÀßäý|ådì'à í ÜéãäædܯâöÞÀèéã«äàdè'ó¹âöàß
çàäÈè FáàÀôtÚ¥Ûäá=éåè¾íÁÜ|çàèÀÜIí6êÃäèãß«àçÞéäÈ á = D FH
ø D;9:9;9D : äèÃãÛÜ
è |Ý'àáäÈã«äÈæÜWáìåé4õTæ ådßä{ådíÀìÜ
éàèáãß4åäèã«á
ò ¹àdßãÜ)áiåèçFðKåÝèÀäõ'÷rø)ùdùúdó4÷ÿiÛÀä{é4ÛFãÛÀÜèTí'Ü)éàdývÜ
<>= -9
7 I7 ø4 = âöàß L
=H 7Iø òB
Só
<>= -9
7 1 ` ø7 = âöàß L
=H `ø òBø ó
= 4
S FKI òBó
G $JM I
SN1 = 1
D òB 6
Bó
( = L
=>H SI òBúó
=
H ( = L = < = I òBó
= ) #
ÿiÛÀÜßÜ
ïä{á
ãÛÀÜvè6ÞÀýtí'Üß`àâ¥áÞÀÝÝ'àßãNæÜéã«àdß4áô|Ú¥Û6Þá
ã«ÛÀÜ|àdèìÈêÃçÀä 8Üß«Üè'éÜtâößàýyãÛÜvàdÝã«äÈý|ådì
Û6ê6Ý'ÜßÝÀì{ådèÀÜ]éåáÜ]ä{áã«Ûå-ã¹ãÛÀÜ = èÀà-ÿ Ûå æÜ]åèvÞÀÝÀÝÁÜß#íÁàdÞÀèç|àâ
IôKÚ¥ÛÀܱáäÈãÞåãäàdètä{á&áÞÀývý|åß«äÜ)ç
á«é4ÛÀÜý|åãä{éådìÈìêNäè ä ÞÀß«
Ü Àô
VÜTÿiäÈìì¹èÀÜÜ)çµã«ÛÀÜ `ådßÞ'áÛ =ÞÀÛÀèÚ@Þé4õdÜßWéàèçäÈãäàèá`âöàßWã«ÛÀÜTÝÀßäý|ådìKÝÀß«àdíÀìÜýTôÅÚ¥ÛÀÜTÝÀß«äý|åì
ådß4åèä{åèvä{á
H
ø
, 7
( = ( = $- L ò
= '< = 9
7 ó ø 7 = ." (
= = òBdó
= = =
ÿiÛÀÜßÜ=ã«ÛÀÜ
,=¹åܱ߫ã«ÛÀÜ@å dß4åèdÜ
ýIÞÀìÈã«äÈÝìÈäÜß«á?äèãß«àçÞéÜçFãàvÜèâöàdß4éÜ`ÝÁàáäÈãäæ6ä:ãê¼àdâã«ÛÀÜ=ô¥Ú¥ÛÜ
`Ú éàdèçÀä:ã«äàdèá¹âöàdߥã«ÛÀÜ`ÝÀß«äý|åì8ÝÀß«àdíÀìÜý ådßÜ
ã«ÛÀÜß«ÜâöàßÜvòöèàãÜ F¹ßÞèá#âößàý ø
ãà¼ã«ÛÀÜ`è6ÞÀýtí'Üßiàâ
ãß4åäèÀä
è IÝ'àäèã4á÷Àådè ç |âöß«àdý ø±ã«à¼ãÛÀÜNçÀäÈývÜèáäÈàèvàâãÛÀÜNçå-ã«åó
H ( "= L
=*)+= H S òBüó
=
2
H ( = L = H S òBùó
=
H : =>
,=>H S
ò.ú
S ó
=
ò
L = '< = 9
7 ó? ø7 = ;S ò.úø ó
= ; S ò.
ú ó
=$;S ò.Lú Eó
,=$;
S ò.Cú B6ó
=!-7L
=ò<=-9
7 ó?Dø7 = .CH S ò.údúó
,= =H S ò.ú ó
-b
|w|
−ξ
|w|
;Á$), > Î " 3 @*2² .41," 3 i+§²d2² $ 3 *9 0 1,+- 3-043 » *2;²-.4( &*2 5
( àdß4éÜ)á H ( ='L
=
G
H S ò.údó
=
%ò ]Üß«Ü
ãÛÀÜ = å߫ܯã«ÛÀÜ`áÞÀÝÝ'àßãiædÜ)éã«àdß4á÷åè
ç µçÀÜèÀàdãÜá¹ã«ÛÀÜ=ædÜéãàß#ÝßàçÞéãô ó ÀàdßiçÀåã«åIäè m B ÷
é ìÜ)åß«ìÈêFãÛÜ|éàdè'çäÈãäàdèUã«Ûå-ã`ã«ÛÀÜ|áÞÀýfàâ¹âöàdß4éÜ)á=æ ådèÀä{áÛ ä{á
áãäììývÜãô $±èܼéåèïÜåáäìêáÛÀà-ÿ ã Û åã
ãÛÜ=ãàdß !ÞÀÜ`åì{áàWæ-åèÀä{áÛÀÜáô
Ú¥ÛÀä{á@ývÜé4ÛådèÀä{éådìdåèådìà dê±çÜÝÁÜèçá@àèÀìê±àèNã«ÛÀÜ&âöàßý àdâÀãÛÀÜ¥áàdìÞãäàè¼òBó4÷åèçNãÛÀÜßÜâöàdß«ÜKÛÀàìçá
âöàdßWí'àdãÛïãÛÀÜFáÜÝåß4åíÀìÜtåèç ãÛÜ|èÀàdèáÜÝådß«ådíÀìÜIéådáÜáô/
;è â,åéãNã«ÛÀä{á`åèådìàdêTÛÀàì{çÀá=äè ÜèÀÜß«ådì
òöä.ô ÜôÈ÷8ådì{áà¼âöà߯ã«ÛÀÜWèÀàèÀìäèÀÜådß]éådáÜIçÀÜá«éß«äÈíÁÜçíÁÜìà-ÿ¯ó4ô=Ú¥ÛÀÜIådèåì
à dê|ÜývÝÀÛådáä Ü)á#ãÛÀÜtäèã«Üß«Üáãä
è
ÝÁàdäèã¹ã«Ûå-ã#ãÛÀÜ-ývàáã¹äývÝ'àßã4åèãK`çå-ã«åIÝÁàdäèã«á#åß«Ü]ãÛÜ=áÞÀÝÀÝÁàdßã¥ædÜéãàß«á¹ÿiäÈãÛoÛäÛÀÜáã¹æ-ådìÈÞÜá¹àâ
&÷áäèéÜ]ãÛÜê|Üë6Üßã#ãÛܱÛÀä ÛÀÜáã¹âöàß«éÜá¹àdèoãÛÀÜ`çÜ)éä{áäàè|áÛÀÜÜãô àdß#ãÛÜ
èàdè ;áÜÝåß4åíìÈÜ]éådáÜ÷6ãÛÜ
ÞÀÝÀÝÁÜß#íÁàdÞÀèç = 1
éàßß«ÜáÝ'àèçÀáã«àWåèoÞÀÝÀÝÁÜß&íÁàdÞèçvàdè|ãÛÀܯâöàdß4éܯådèê däædÜè¼Ý'àäÈèã&ä{áKådììÈà-ÿ#Üç
ãàTÜë6ÜßãNàè¾ã«ÛÀÜ|áÛÀÜÜãôoÚ¥ÛÀä{á=åèådìà dêUådìáàÝÀßà-æ6ä{çÜ)á±åß«ÜådáàdèDò,åá dà6àç åá`åè6êÃàã«ÛÀÜßó¯ãàUéådìì
ãÛÜáÜ`Ýåßãä{éÞìådß&æÜéã«àdß4á áÞÀÝÀÝÁàdßã¥æÜéãàdß4Ká '# ô
; $),
> +-K " *. *.;²-24( $ $;91 3 3-0 1 2" 4+1 5< +-?(-.¶ 0 3 N 0 0 -?*,+ 0 *12+-&*2+ 4² 0 912+-
Á
;"$*2" *2-< ,9È 3
043 5
z UÕÁÒ%2Ù,Ò"#ÀÔFpr×
¹ÕÔ Ó!C6ØdÓÕ'Ô046#ÀØ7Ù,Ò"
]à-ÿ éåèFãÛÜ=ådí'à-æܯývÜãÛÀàçÀá#íÁ Ü ÜèÀÜß«ådìä Üçtã«àWã«ÛÀÜ`éådáÜ
ÿiÛÀÜ߫ܯã«ÛÀÜNçÜ)éä{áäàè¼âöÞÀèéã«äàd
è ##iä{á#èÀàdã
å ìäèÀÜådߥâöÞÀèéãäàdèUàdâã«ÛÀÜIçå-ã«å ò*#àáÜß÷ % ÞÀêdàèÃåèçTðKådÝÀèÀäõÁ÷@øùù
ó4÷ÁáÛÀà-ÿ#ÜçFãÛ'å-ã
å|ß«åãÛÀÜß]àdì{ç
v
ãß«ä{é4õïò äÜßý|ådè÷@ø)ù BóiéåèÃíÁÜIÞ'áÜ)çTãàådééàdývÝÀìä{áÛoã«ÛÀä{á¯äè¾ådè¾ådáãàèÀä{áÛÀäèìê|áã«ß«ådäÛãâöàdß«ÿ¥åß4ç
ÿ¥å êô @äß4áãKèàãä{éÜiãÛåã&ã«ÛÀܯàèÀìêWÿ¥å êIäèvÿiÛÀä{é4ÛtãÛÀÜ
çÀå-ã4åWådÝÀÝ'Ü)åß4á?äè¼ãÛÀܯãß4åäèÀä è =ÝÀß«àdíÀìÜýT ÷ 01!6áô
@ò B E+ó `@ò B6údó÷Áä{áiäèTãÛÜNâöàdß«ý àâKçàdã±ÝÀß«àçÞéã4á÷ <,= 9< ô à-ÿ áÞÀÝÝ'àáÜIÿ¹Ü 'ß«áã¯ý|åÝÝ'Ü)çTãÛÀÜtçÀå-ã4å
ãàtáàývÜiàã«ÛÀÜß
òöÝÁàá«áäíÀìêWä è 'èÀäÈãÜ]çäývÜè'áäàdè'åìö ó 0&Þéìä{çÜ)åè¼áÝåé
Ü Ã÷Þáä
è Wå`ý|ådÝÀÝÀä è `ÿiÛÀä{é4Û¼ÿ#Ü
ÿiäìì8éåìì
m I ò.úùó
Ú¥ÛÀÜèWàdâéàdÞß«áÜ?ãÛÜKãß4åäèÀä
è iådì àdß«äÈãÛÀý ÿ#àdÞìç=àèÀìê±çÜÝ'ÜèçNàè`ãÛܹçÀåã«åiã«ÛÀß«àdÞÛNçàãÝÀßàçÞ'éã«á
äè
U÷Áä.ô Üô¥àdèTâöÞÀè'éãäàèáiàâã«ÛÀÜNâöàßý
ò<>=ó 9
ò< )ó4ô à-ÿ ä:âãÛÜß«ÜNÿ¹ÜßÜNå 4õdÜßèÜì8âöÞÀèéãäàdè+
áÞé4Û¼ãÛåã O'ò <,= DK< )ó H
'ò <>=28 ó 9
ò < )ó4÷6ÿ¹Ü]ÿ#àdÞìçtàèÀìêNèÀÜÜ)ç¼ãàWÞáÜ
fäÈètã«ÛÀÜ]ã«ß«ådäÈèäÈè
ådìàdß«äÈãÛÀýT÷
åè'çoÿ#àdÞÀì{çFèÀÜæÜߥèÀÜÜçFãà¼ÜëÝÀìä{éäÈã«ìÈêtÜæÜèõ6èÀà-ÿ ÿiÛå ã ä{á1ô $±èÜ
ÜëÀåývÝÀìܯä{á
Oò'<,= DK< )ó1H 7 <
7
<
$I ò
Só
;èoãÛÀä{áKÝ'åßãä{éÞÀì{ådß?ÜëÀåývÝÀìÜd÷ ä{áKäè'èÀäÈãܯçäývÜèáäàèåì.÷áàIäÈã&ÿ¹àÞÀì{ç¼èÀàã#í'Ü
ædÜßêtÜåáêIãàtÿ#àdß«õ
ÿiäÈãÛ ÜëÝÀìäéäÈãìêô]à-ÿ#ÜæÜß÷'äÈâKàdèÜWß«ÜÝÀì{ådéÜá <=
9$< Wí6ê
ò<>= D < )ó]ÜæÜß«êÿiÛÜß«ÜNäèTãÛÀÜIãß4åäèÀäè
åì àdß«ä:ã«ÛÀýT÷Àã«ÛÀÜvåìdàßäÈãÛýYÿiäììÛådÝÀÝÀäìêFÝÀß«àçÞéÜtåáÞÝÀÝ'àßã=ædÜ)éã«àdß
ý|ådé4ÛÀäèÀÜNÿiÛäé4Û¾ìÈäæÜá]äè¾ådè
ä è èÀäÈã«Ü¼çÀäÈývÜèáäÈàèåì?áÝådéÜ÷åèçïâöÞÀßãÛÜß«ývàdß«Ütçà áàUäèVßà Þ dÛìÈêTã«ÛÀÜoá«åývÜ|ådývàdÞÀèã`àâ#ãäývÜväÈã
ÿ#àdÞÀì{ç¼ã4åõdܱãàtã«ß«ådäÈèoàdè|ã«ÛÀÜ=ÞÀ è ý|åÝÀÝÁÜçFçÀå-ã4åô ìì8ã«ÛÀÜ`éàdè'áä{çÜß«åãäàdèáàdâã«ÛÀÜ=ÝÀßÜæ6äÈàÞá#áÜ)éãäàèá
ÛÀàìç÷áäèéܱÿ#Ü`åß«Ü=áãäììçàdä è ¼åtìäÈèÜåߥáÜÝådß«åãäàdè÷6íÀÞã]äèTå¼çä ÁÜßÜèã]áÝådéÜô
#Þã]ÛÀà-ÿgéådèÿ#Ü=Þáܱã«ÛÀä{á¥ý|ådé4ÛÀäèÀÜ {â ãÜ߯åìì.÷6ÿ¹Ü=èÀÜÜç
÷'åèçoãÛ'å-ã]ÿiäììÁìäædܱäè yåì{áàFò2áÜÜ
01!'ô¾òBóó4ô #ÞãNäèïãÜ)áãNÝÛådáÜvåèµî6ð±ñ {ä á=ÞáÜçïí6ê éàdývÝÀÞã«äèçÀàãNÝÀß«àçÞéã4á=àâiå däædÜèÃã«Üáã
ÝÁàdäèã <ïÿiäÈãÛ
÷àdߥývàdß«Ü=áÝÁÜéä ÁéåììêtíêéàývÝÀÞãäèW ãÛÀÜNáäd èFàâ
ò<@ó H ( = L
=
ò =ó 9
ò'<@ó 7 H ( ='L
= Oò = DK<ó 7 òÀø ó
= ) = )
# #
ÿiÛÀÜßÜtã«ÛÀÜ = ådßÜtã«ÛÀÜ|áÞÝÀÝ'àßã`æÜéã«àdß4áô|î6àUå åäè¾ÿ#Üvéådèïå æàdä{çÃéàývÝÀÞã«äÈè
ò<ó
ÜëÝÀìä{éäÈãìê
åè'çoÞáÜ=ãÛÜ Oò = DK<ó1H
ò = ó49
ò'<@ó#äèáãÜåç8ô
ÜãiÞáiéådìì'ã«ÛÀÜ`áÝ'ådéÜ
äèoÿiÛÀä{é4ÛoãÛÀÜ`çÀåã«åIìÈäæÜ÷ ]ô¯ò¯Üß«Ü`åèçFí'Üìà-ÿ ÿ¹Ü=ÞáÜ
åáiåWývèÀÜývàdèÀä{é
;èïãÛÜvèÀÜë6ãIî6Ü)éãäàè¾ÿ#Ütÿiäìì?çä{áéÞá«á±ÿiÛäé4Û âöÞÀèéãäàdè'
á ådßÜvåììà-ÿ¥åíÀìÜNåè'ç¾ÿiÛÀä{é4Ûïåß«Ütèàãô
rÜã¯Þá±Üèçã«ÛÀä{á±î6Ü)éã«äÈàèTÿiä:ã«ÛUåvædÜßêáäývÝÀìÜ=ÜëÀåývÝìÈÜ=àdâ?ådèÃåììà-ÿ¹Ü)çvõdÜßèÜì.÷Àâöàdß±ÿiÛÀä{é4ÛTÿ¹ÜI
éàèáãß«Þéã¥ãÛÜ`ý|åÝÀÝÀäè ¯ô
î6ÞÝÀÝ'àáÜIãÛ'å-ã`êdàÞÀß
çå-ã«åFåß«ÜWæÜéã«àdß4á¯äè m ÷ådèçÃêàdÞ¾é4ÛÀà6àáÜ Oò<,= D < )óH ò'<>=9$< )ó ôWÚ¥ÛÀÜè
äÈã$á#Üådáê¼ãà èçå¼áÝ'ådéÜ U÷ådèç|ý|ådÝÀÝÀäè Dâößàý m ãà U÷áÞé4Û|ã«Ûå-ãNò'<J9 ó H
ò'<ó9
ò ó
ÿ#Ü=é4ÛààáÜ H m 8 åèç
)
#
ò<@ó H ) #) ò
ó
)N
á !Þ'åß«Ü *`ø D)ø `ø
Dø @? m ò,åãê6ÝÀä{éådì&áäÈã«Þå-ã«äÈàè÷râöàß dß«Üê ìÈÜædÜì?äÈý|ådܼçÀå-ã4åó÷@ã«ÛÀܾòöÜèãäßÜ ó
äý|å Ü]àâ ä{á¥áÛà-ÿièoäè,ädÞßÜ
üô&Ú¥ÛÀä{á äÞÀß«Ü
ådìáàWäììÞáãß4å-ãÜ)á?ÛÀà-ÿ ãàtãÛÀäèÀõoàâã«ÛÀä{á#ý|åÝÀÝÀäè
ãÛÜtäý|å dÜNàdâ ý|å êTìäædÜNäèïåFáÝådéÜIàdâ¹æÜß«êÛäÛ çäývÜèáäàè÷ÁíÞã`ä:ã=ä{á ÞáãNåVòöÝÁàá«áäÈíìÈêædÜßê
éàèãàdßãÜ)çó#áÞßâ,åéÜ=ÿiÛÀàáÜ
äÈèã«ßäèáä{é¯çäývÜè'áäàdè # äá ÞáãiãÛ'å-ã]àâ ]ô
ø
ò,) )
#
ó
ò<@ó H ) #)
Eó
ò
,ò ) # 7 ) ó
1
0.8
0.6
0.4
0.2
0
1
0.5
0.2 0
0.4 -0.5
0.6 -1
0.8
1
)
Ú¥ÛÀÜ`ìäÈãÜß«åãÞÀß«Ü]àèTî6ð
ñÃáiÞáÞådìÈìêtßÜâöÜß4á¹ãà¼ãÛÜ=áÝåéÜ _åá¥å¯äìÈíÁÜßãiáÝåéÜd÷áà¼ìÜã$á#ÜèçFã«ÛÀä{á
î Ü)éãäàè|ÿiäÈãÛoåNâöÜÿ èÀàdãÜ)á¹àèvãÛäá¹ÝÁàdäèãô 6KàdÞéåè|ãÛÀäèÀõ¼àâå]äìíÁÜßã#áÝådéܱådá#å ÜèÀÜß«ådìä)å-ã«äÈàè
6
àâ 0KÞ'éìä{çÜådèIáÝådéܹã«Ûå-ãKíÁÜÛ'å ædÜáäèvådÜèã«ìÜý|åèìÈê
â,åáÛÀäàèôî6ÝÁÜéä'éådìÈìêd÷ äÈã?ä{á?åè6êNìäèÀÜådß?áÝåéÜd÷
ÿiäÈãÛOådèµäèÀèÀÜßWÝÀß«àçÞéã¼çÜ 'èÀÜç8÷KÿiÛäé4ÛÅä{áIåì{áà éàdývÝìÈÜãÜvÿiäÈãÛÅß«ÜáÝÁÜéãIãà¾ãÛÀÜTéàßß«ÜáÝ'àèçäè
èÀàßý ò{ã«Ûå-ãvä{á÷&åè6ê #åÞé4Û6êVáÜ !ÞÜèéÜFàdâ¯Ý'àäèã4áWéàdè6ædÜß Üá`ã«à å Ý'àäèãtäèÅãÛÀÜTáÝådéÜ)ó4ôOî6àývÜ
åÞÀãÛÀàß«á=òöÜdô ô¯ò =àìÈývà dàßà-æÁ÷Àø ù $Sóó#åì{áàWß« Ü !ÞÀäßܯãÛåã]äÈã¥íÁÜNáÜÝådß«ådíÀìÜIò{ã«Ûå-ã]ä{á÷ÀäÈã¥ýtÞáã#Û'å ædÜ=å
éàÞÀèã«ådíÀìÜKáÞÀí'áÜãÿiÛÀàáÜKéìàáÞÀß«Ü?ä{á8ãÛÜ&áÝ'ådéÜ&ä:ã4áÜìÈâ«ó4÷-åèç`áàdývÜ]ò,Üô Àô ¯åìývàá÷dø)ù ó8çàè ãô
ã á@å
dÜèÀÜß4åìä åãäàdè
ý|åäèÀìê±í'Ü)éåÞ'áÜ#äÈã«áäèÀèÀÜßÝÀß«àçÞéã?éådèWíÁÜ #¹äèÀèÀÜßÝÀß«àçÞéã÷èÀàd ã Þ'áããÛÜ¥á«éådì{åß
ò çàã ó=ÝßàçÞéãWÞáÜçVÛÀÜßÜÃò2åèçVäè 0&Þéìä{çÜådè áÝådéÜá`äè ÜèÀÜß«ådì{óô
ã áWäÈèã«Üß«Üáãä
K è FãÛåãWãÛÜ
àdì{çÜߥý|å-ãÛÜý|å-ã«ä{éåì'ìä:ã«Üß4å-ã«ÞÀßÜtòöÜdô ô =àìývà dàßà-æÁ÷Àø)' ù $Só&ådì{áàtßÜ !ÞÀäßÜ)ç¼ãÛåã ]äìí'ÜßãiáÝådéÜá¥íÁÜ
ä è èÀäÈã«Ü¯çäývÜèáäàdèådì.÷-ådèç¼ã«Ûå-ã#ý|å-ãÛÜý|å-ã«ä{éä{åèáådßÜ !ÞÀäÈã«Ü]ÛåÝÝêvçÜ èÀä è Näè 'èÀäÈãܯçäývÜèáäàèåì
0KÞ'éìä{çÜådè áÝådéÜáô ]ÜáÜåß4é4ÛVàdè ]äìíÁÜßãWáÝåéÜ)áNéÜèã«Üß4á`àdèÅàÝ'Üß«åãàß«á=äèVãÛàáÜ|áÝåéÜ)á÷?áäèéÜ
ãÛÜ|íådáä{éWÝßàÝ'Üßã«äÈÜ)á
Ûå ædܼìàd è UáäèéܼíÁÜÜèVÿ#àdß«õdÜ)çÃàdÞÀãôTîäÈè'éÜváàývܼÝ'ÜàdÝÀìܼÞÀèçÜß«áã«ådèçÀådíÀìê
íÀì{åè'é4Ûïåã`ãÛÀÜvývÜèã«äàdè¾àd6 â ¯äìÈíÁÜßãNáÝ'ådéÜ)á÷
Oò'<PD ó H (
ò'<@ó =
ò ó = òúó
=
ò'<ó¥áÞ'é4ÛFã«Ûå-ã
äÈâ?åèçFàdèÀìêväÈâ;÷âöàdß]ådè6ê 8
ã«ÛÀÜè
ò( ) = L = ò<ó ò ó
ó
< SI òdüó
= )
#
Ú ÛÀÜNãêÝäéåì8ã«Üß«ý
¥ äèFãÛÀÜNýtÞÀìÈãäèÀàývä{åìÜëÝåèáäàdèFàâ]ò =*)
#
)N= L
=ó éàdèãß«äíÀÞã«ÜáiåtãÜßý àâãÛÜ
âöàdß«ý
#
:9;9:9ò # 9;9:9 ó )# ) 9:9;9 L# L 9;9;9 8ò'<ó ò ó < òdùó
ã«à¼ãÛÀÜ`ìÜâ{ãiÛådèçTáä{çÜ
àâ 01!'ô]ò*ó÷ÀÿiÛÀä{é4Ûoâ,åéãàßäÜá
H
#
;9:9;9)ò # 9:9;9 ó ò/ ) # ) 9:9;9 8ò'<ó @ó
< ;SI ò $Só
áÜãFàdâWåìì¯Û6ê6ÝÁÜß«ÝÀì{åèÀÜ)á -
0D J.Ååß«ÜÃÝåß4åývÜãÜßäÜ)çOí6ê çäýÃò ó 7IøÃè6ÞÀýtí'Üß«áô[ñUàáãoÝ'å-ãã«Üß«è
ß«ÜéàdèÀäÈã«äÈàèWáê6áãÜý|áÿiäÈãÛ¼íÀäììÈäàèá÷ àdß&ÜædÜèvåè¼äèèÀäÈãÜd÷dè6ÞÀýtí'Üß?àâÝåß4åývÜãÜß«á@ÿ¹àÞÀì{çWèàã&ý|ådõdÜ
äÈã¯Ýåáãiã«ÛÀÜIáã«ådßã å-ãÜd ô ]à-ÿ éàývÜ`îð±ñÃá
çàoáàvÿ¹Üìì $±èÀÜNýväÛã]ådß ÞÀÜ=ãÛåã÷ däædÜèoã«ÛÀÜ`âöàßý
àâáàdìÞã«äàdè÷ã«ÛÀÜܱ߫ådßܱå-ã&ývàáã : 7µø¯åç Þáã«åíìÈܯÝåß4åývÜãÜß4áiòöÿiÛÜß« Ü :ä{á?ã«ÛÀܯè6ÞÀýtíÁÜß&àdâãß4åäèÀä
è
á«åývÝÀìÜá4ó4÷íÀÞã&ã«ÛÀä{á&áÜÜý|áãàIí'ܱí' Ü dä
è =ã«ÛÀÜ !ÞÀÜ)áã«äàd
è # 8 ô
ã¹ýtÞáã&í'Ü
áàývÜã«ÛÀä
è
ãàWçàIÿiäÈãÛvàdÞß
ß« Ü !ÞÀäßÜývÜèãKà$â K K $ K K +*IÛ6ê6ÝÁÜß«ÝÀì{åèÀÜ)á?ãÛ'å-ã¥äá¹áå æ6ä è =ã«ÛÀÜ
çå êô á#ÿ#ܱáÛåìì'áÜܱí'Üìà-ÿ=÷
å¼áã«ßà
è véåáÜ=éådèFíÁÜ`ý|ådçÜ
âöàߥãÛÀä{áiéì{ådäÈýTô
î6äèéÜtãÛÀÜ|ý|ådÝÀÝÁÜçïáÞÀßâ,ådéܼäá
àdâ#äèãß«äèáä{éNçÀäÈývÜèáäÈàè¾çäýÃò ¥ó4÷ÞÀèÀìÜá«á=çäýÃò #ó HYçäýÃò Fó÷äÈã
ä{á`àdí6æ6äàdÞá
ã«Ûå-ãIãÛÀÜFý|åÝÀÝäÈèTéåèÀèàãWíÁÜoàèã«àµò,áÞÀßÜéã«äædÜ)óô
ãtåì{áàTèÀÜÜ)çÅèÀàãIí'ÜFàdèÀÜvã«àÃàdèÜ
òöíä Ü)éã«äÈæÜ)ó &éàdè'áä{çÜß ) # ) # D ) ) äè 01!'ô±*ò dóô#Ú¥ÛÀÜ`äý|å ܱàdâ èÀÜÜçTèÀàdã]ä:ã4áÜìÈâ@íÁÜ
è
ãÛÀÜ=åíÁà-ædܯáäývÝÀìÜ !Þådçß4å-ã«ä{éiÜëÀåývÝìÈÜd÷dã«ÛÀܱæÜéãàd"
åNæÜéãàdߥáÝådéÜ ?å ådäè÷éàdèáä{çÜß«ä ß
ò <ó&ä{á
èÀàdã#äèvã«ÛÀܱäý|å dÜiàâ ÞÀèÀìÜá«1 á <QH Sô ÀÞÀßãÛÜß÷åWìä:ããìÜ]ÝÀì{å ê6ä
è =ÿiäÈãÛ|ã«ÛÀÜ
äÈèÛÀàdýv à dÜèÜàdÞ'áõdÜß«èÀÜì
*ò )Áó1H
7 ( ò
# ) ó87
éàáò Á
) óó
áäèrò ' ò
ó
) #
då èç
ãÛÀä{áéåè=í'Ü&æäÜÿ¹Ü)ç±åáråiçàãrÝßàçÞéãrí'Üãÿ¹ÜÜè=ãÿ#à¥ædÜéãàß«á8äè m # # H ò
ó÷
# #
#
D DIIIJD DIII
åè'ç|ã«ÛÀÜ=ý|åÝÀÝÁÜç
ò*)'ó H ò # D4éàáò*)Áó D«éàá)ò F)'ó DIIID«áäèrò*)Áó D«áäèò F)'ó DIII$óô&Ú¥ÛÀÜèFãÛÜ=éàdß«ßÜ)áÝÁàdè'ç
äè T*ò -±äÈß«ä{é4ÛÀìÜãóKõÜß«èÀÜìéådèTí'ÜNéàdývÝÀÞã«ÜçoäÈèTéìàáÜ)çvâöàdß«ý
ò *) = ó49
ò*) ó H
ø
7 ( éàáò ) = óéàáò ) >ó 7OáäÈèrò ) = óáäèrò ) ó
)
#
H ø
7 ( éàáò ? ó 2
H
ø
7 A L- (
=
.
) )
ø
H 7*A L-òø4 = #
ó =Àòø4
=
ó .
HÑò,áäèròò 7 øC= dó?óó
= ?áäèÁò,? = dó I
äèåììê÷räÈã`ä{á`éìÜådß
ãÛåã`ãÛÀÜ|ådí'à-æÜIäývÝÀìä{éäÈã¯ý|ådÝÀÝÀäè Fãß«äé4õUÿiäììÿ¹àßõTâöàß #tådì dàdß«äÈãÛÀý äè
ÿiÛÀä{é4Û|ã«ÛÀÜNçÀåã«åIàdèÀìê|åÝÝ'Ü)åßiådáiçàdãiÝÀßàçÞ'éã«á`òöâöàdßiÜëådývÝÀìÜ÷ãÛÀÜ=èÀÜ)åß«ÜáãièÀÜädÛ6íÁàdߥåìdàßäÈã«ÛÀýoó4ô
Ú¥ÛÀä{á=â,åéãWÛåá`íÁÜÜèÅÞ'áÜ)çïãàÃçÜßäædÜvåUèàdèÀìäèÀÜådß±æÜß4áäàè¾àâiÝÀß«äèéäÝådì?éàdývÝÁàdèÜèãWåèådìê6áä{á¯í6ê
ò2îÀé4
Û àd ìõdàÝâ;÷@î6ývàìåFåèçV) ñ Þ ìÈìÜß÷?øùùdüdí'ó4ûä:ãNáÜÜý|á±ìäõdÜìêoã«Ûå-ã`ã«ÛÀä{á
ãß«äé4õTÿiäìì?éàdèã«äÈè6ÞÀÜIãà 'èç
ÞáÜá¥ÜìáÜÿiÛÜß«Üô
!'& ICKM K 56*M GID"I( 6 * M &
G
Ú¥ÛÀÜ ß4áã¥õÜß«èÀÜì{á¹äèæÜáãäåãÜ)ç¼âöàdߥã«ÛÀÜ`Ýå-ããÜßèßÜ)éàdèÀäÈãäàèvÝÀßàíÀìÜý ÿ#Üܱ߫ãÛÀÜ=âöàìÈìà-ÿiäè
Oò'<PD ó H ò<L9 7 ø)ó
Bó
ò 6
àèÀÜ ædÜß«áÞá .ß«Üáã¥éì{ådá«áä Üß«á=ò,á«å ê1÷ 4àèÀÜ ¼ÝÁàáäÈãäædÜdP÷ 4ß«ÜáKã ¼èÜ åãäædÜ ó¹âöàd߯ãÛÀÜ éì{ådá«áiéådáÜWådèç
ã«ådõdÜ)áNã«ÛÀÜUéì{ådá«áNâöàßtå¾ãÜ)áãvÝÁàdäèãWã«àïíÁÜFãÛåãvéàdß«ßÜ)áÝÁàdè'çäèTãàïãÛÜFì{åßdÜ)áãIÝ'àáäÈãäæÜoçä{áã4åèéÜ
*ò #àáÜß÷ % ÞÀêàdèTåèçFðKåÝèÀäõ'÷rø)ùdù dó4ô
*I
6 6 & I 6 $ IAG 7 ;$AM
M G
G
ÛÀÜèFä{á¹ãÛÀÜ`áàdìÞã«äÈàè¼ãàIãÛÀÜ`áÞÀÝÀÝÁàdßãiædÜéãàß&ã«ß«ådäÈèäÈèN ÝÀß«àdíÀìÜý ìÈàíåì.÷6åèçoÿiÛÀÜèä{á¹äÈã¥ÞÀèÀä
!ÞÀÜ
#ê ìÈàíåì À÷'ÿ#ÜtývÜ)åèUã«Ûå-ã=ã«ÛÀÜܼ߫Üëä{áã4á¯èÀàTàdãÛÀÜß
ÝÁàdäèã=äÈè¾ãÛÀÜtâöÜ)ådáäÈíìÈÜIßÜ däàdèÃåã`ÿiÛÀä{é4ÛÃãÛÜ
àdíÜéã«äædÜâöÞÀèéãäàdè=ã4åõÜáråiìà-ÿ#Üßæ-åìÞÀÜdô,VÜ&ÿiäììdådçÀçÀßÜ)áá8ãÿ#àiõ6äèçÀáràdâÿ#å êáäè`ÿiÛÀä{é4Û=ÞÀèÀä
!ÞÀÜèÜá«á
ý|å ê=èÀàã?ÛÀàì{ç áàdìÞã«äÈàèáâöàßÿiÛÀä{é4ÛN-
!DJ.]ådßܹãÛÜý|áÜìædÜ)áÞÀèä! ÞÀÜd÷íÀÞã?âöàß?ÿiÛÀä{é4ÛNãÛÜ#ÜëÝåèáäàdè
àâ$
_äÈè 0 ! ô
ò@B ó#ä{áièÀàãûÁådèçTáàìÈÞÀãäàdèá¹ÿiÛÀàáÜ -'
0DJ.IçäÁÜßô # àãÛUådßÜ=àâ?äèãÜßÜ)áã K
ÜædÜèTä:âãÛÜ
ÝådäÈß -
0D .Uä{áIÞÀèÀä
! ÞÀÜ÷&äÈâ]ã«ÛÀÜ=`ådßÜFèÀàã÷&ãÛÜß«ÜFý|å êïí'ÜTÜ ! ÞäÈæ-ådìÈÜèãWÜëÝådèáäàèáNàdâ
ÿiÛÀä{é4Û
ß«Ü ! ÞÀäßÜ]âöÜÿ#ÜߥáÞÀÝÝ'àßã#ædÜéãàß«á±ò,åWãß«äÈæ6ä{åìÜëÀåývÝÀìÜiàdârãÛÀä{á¹ä{á d äæÜèvíÁÜìà-ÿ¯ó4÷åèç|ÿiÛÀä{é4ÛvãÛÀÜßÜâöàdß«Ü
ß«Ü ! ÞÀäßܱâöÜÿ#Üßiäèáãß«Þéãäàdèá#çÞÀß«äèI ãÜáã]ÝÀÛåáÜdô
ãvã«ÞÀßè'áIàÞã¼ãÛåãvÜædÜßêVìàéådì¥áàdìÞã«äàdèÅä{á¼åì{áà dìàdíådì.ôÅÚ¥Ûäátä{á¼å¾ÝßàÝ'ÜßãêÅàdâ
åè6êµéàdè6ædÜë
ÝÀß«àß«ådývýväÈè ÝÀß«àdíìÈÜý òìÈÜã«é4ÛÀÜß÷tøùü'óô ÞÀßãÛÀÜßývàßÜd÷]ãÛÀÜVáàdìÞãäàè äá dÞådß«ådèãÜÜ)ç ãàDíÁÜ
ÞÀèÀ
ä !ÞÀÜUäÈâ=ãÛÀܾàdíÜ)éãäæÜFâöÞèéã«äÈàè ò*0 !ô òB Eóó¼ä{á|áãß«äéãìêµéàè6ædÜë8÷iÿiÛÀä{é4Û äèDàÞÀßoéådáÜÃývÜådèá
ãÛ'å-ã=ãÛÀÜ ¯Üá«áä{ådè¾ýtÞáã=íÁÜtÝÁàáäÈãäædÜtçÜ 'èÀäÈãÜUòöèÀàdãÜtãÛ'å-ã`âöàdß !Þåçß«åãä{éNà í ÜéãäædÜIâöÞÀèéãäàdè'á t÷
ãÛ Ü ¯Üá«áä{åèUä{á±Ý'àáäÈã«äÈæÜWçÜ èÀäÈãÜIä:â&ådèçÃàèÀìêFäÈâ äá±áã«ßä{éãìêéàdè6ædÜë8ûÁã«ÛÀä{á¯ä{á±èÀàã
ãß«ÞÀÜWâöàdß=èÀàd è
!ÞådçÀß«åãä{é ¹ãÛÀÜßÜd÷Áå¼ÝÁàáäÈãäædÜ=çÜ 'èÀäÈã Ü ]Ü)ááä{åèäývÝÀìäÈÜ)á¹åváãß«ä{éã«ìÈêFéàdè6æÜëoà í ÜéãäædܱâöÞÀèéãäàdè÷
íÀÞã¼èÀàdãtæ6ä{éÜ|æÜß4áåÅò2éàèáä{çÜß H ) -ó|*ò ìÜã4é4ÛÀÜß÷]øù' ü dóó4ô ]à-ÿ#ÜæÜß÷ÜædÜèµäÈâ#ã«ÛÀ Ü ]Üá«áäådèÅä{á
ÝÁàáä:ã«äædÜváÜývä{çÜ èä:ã«Ü÷@ãÛÀÜáàìÈÞÀãäàdèVéådèµáã«äÈììKíÁÜ|ÞÀèä !ÞÀÜ ¼éàdèáä{çÜßNãÿ¹àUÝÁàdäèã«áIåìàd è ã«ÛÀÜoß«Üådì
ìäèÀܱÿiäÈã«ÛTéà6àdß4çäèå-ã«Üá ) # H øNåè'I ç ) HR6÷'ådèçÿiä:ã«ÛFÝÁàdì{ådßäÈãäÜ) á 7 åè5 ç = ô ¯Üܱ߫ã«ÛÀ Ü ¯Üá«áä{åèFä{á
ÝÁàáä:ã«äædÜ=áÜýväçÀÜ èÀäÈã«Ü÷íÀÞã]ã«ÛÀÜNáàdìÞãäàè ò
H 2D @HED
="HRS¼ä, è 0 !6áô=@ò B
Só4÷@ò Bø ó4÷@ò Bdóó#ä{á
ÞÀèÀ
ä !ÞÀÜdô
ã¥ä{á¹ådì{áàNÜ)ådáêtãà èçoáàdìÞã«äàdèá?ÿiÛäé4ÛFå߫ܯèÀàdã#ÞÀèä !ÞÀÜ
äèvãÛÜ
áÜèáܱãÛåã¹ã«ÛÀÜ =@äè|ãÛÜ
ÜëÝådèáäÈàè|àâ
Yåß«Ü
èÀàãiÞÀèÀ
ä !ÞÀÜ âöàdߥÜëådývÝÀìÜ÷éàèáä{çÜß&ã«ÛÀÜ`ÝÀß«àdíÀìÜý àdââöàÞÀß]áÜÝåß4åíìÈܯÝ'àäÈèã4á
àdèUåvá !Þåß«Ü=äÈè m ) # H
ø Dø 2
÷ ) H `
ø Dø 2
÷ ) 8 H `
ø D `ø åè/
ç ) H
ø D `ø .÷ÀÿiäÈãÛTÝÁàdì{åß«äÈãäÜá
7D CD CD 7 ¯ß«ÜáÝ'Ü)éãäæÜìêô $±èÀÜ á àdìÞã«äàdèDä
á
H
ø DKS 2÷ H S÷ H SI d2
ú D SI dú DKSI d
ú DKSI ú .û
åèàãÛÜ߯Û'ådáiã«ÛÀܼáådýv+
Ü
ådèç )÷8íÀÞã H SI 2 ú D SI 2
ú D SDKS &òöèÀàdãÜWãÛåã±íÁàãÛïáàìÞãäàdè'áiáåãä{áâöêFãÛÜ
éàèáãß4åäèã« á = . SIådèç = = L = H SóP
8M * M4" 6*M G > I( * M 5 *I" 6*M G -
!DJ. 8M * M3M GG2 I(
* M 57 I 6 M KHM5 ICG2 *"CMGJM K. M M ?GJI $ ( $ M IFJM *" M $ - I GDI "CM M 7
M /I 5AI * G / * * MIFJM *" M $ - I ()) * G G K.* K 6 " 6 $AM M $ M MM G G #
5 (*
H ¾ò ó H òø ó
7 # D ?
SDø $G $ $ ( ¾ò ó8 G GJI(6 $ I(+ I( 6*6
u=Ô-ÕrÕE( rÜã
ãÛÜIýväèÀäýtÞÀý
vò ó H vò # ó H 8 = B ô #êFéàdè6ædÜë6äÈãê|àdâ t÷ vò ïò óó òø ó vB ò ó vò # ó H
ÀÞßã«ÛÀÜß«ývàdß«Ü÷í6êoìäèÀÜådßäÈãêd÷dãÛÜ
?
"1
87
æ-åìÞÀÜIàâKã«ÛÀÜIàíÜéãäædÜNâöÞÀèéã«äàdèÃíÁÜ
B B
= ô`Ú¥ÛÜè¾í6êÃåááÞÀývÝã«äÈàè÷
? = ô
B
ò óiáåãä{áâöê¼ãÛÀÜWéàdèáãß4åäèã«á+01!'ô±ò
Só4÷@ò ø)ó KÜëÝÀìä{éäÈãìêTò,ååäè
éàýIíäÈèäÈ
è Ní'àdãÛUéàèáãß4åäèã«á¹äÈèã«àtàèÀÜ)ó
#9< = 7dó
òòø4 óò
9< =
7 ó 7 rò
# 9< = 7 # óó
L = ò H L =
ì:ã«ÛÀàd
Þ dÛ|áäÈývÝìÈÜd÷-ãÛÀä{áKãÛÀÜàdß«Üý äá1!Þä:ã«Ü]äèáãß«Þéã«äÈæÜô àdß¹ÜëÀåývÝÀìÜd÷dàdèÀܯýväÛãã«ÛÀäèÀõIãÛåã&ãÛÜ
ÝÀß«àdíÀìÜý|áçÜÝÀä{éãÜ)çWäè @äÞÀßÜ
øS±Ûå æÜ#áÜæÜß4åì6çäÁÜßÜèã?àÝãäý|åì6áàdìÞãäàèá&òöâöàdßãÛÜ¥éådáÜ¥àâ'ìäèÀÜ)åß
áÞÀÝÀÝÁàdßã¯ædÜ)éãàß]ý|ådé4ÛäÈèÜá4ó4ô4]à-ÿ#ÜædÜß÷áäèéÜ`àèÀÜWéåèÀèÀàdã±áývààdãÛÀìê|ývà-æÜ
ãÛÀÜIÛê6ÝÁÜß«ÝÀì{åèÀÜ
âöß«àdý
àdèÜWÝÀß«àdÝÁàáÜçÃáàdìÞã«äàdèãàåèÀàdãÛÀÜß±ÿiäÈãÛàdÞ ã dÜèÜß4å-ãä
è vÛ6ê6ÝÁÜß«ÝÀì{åèÀÜ)á]ÿiÛÀä{é4Û¾ådßÜNèàã`áàìÈÞÀãäàdèá÷
ÿ#Üoõ6èÀà-ÿ ãÛ'å-ãtãÛÀÜ)áÜÝÀß«àdÝÁàáÜçµáàdìÞã«äÈàèáNådßÜFäÈèµâ,ådéãtèÀàãváàdìÞãäàèáNå-ãvåìì.ô
;èOâ,ådéã÷KâöàßIÜ)ådé4Û
àâ&ãÛÀÜ)áÜtéådáÜá÷ÁãÛÀÜtàdÝã«äý|åìÞèÀ
ä !ÞÜWáàìÞãäàdèTä{á
å-ã
H6SÀ÷ÿiä:ã«Û¾åoáÞÀäÈã«ådíÀìÜNé4ÛÀàdä{éÜNàâ |òöÿiÛÀä{é4Û
Ûåá
ãÛÀÜ|Ü ÁÜ)éãNàdâ¥åááä dèÀä è oãÛÀÜ|á«åývܼì{åíÁÜìã«àTådììãÛÀÜvÝÁàdäèã«á4ó4ô àã«Ü¼ãÛåã`ãÛÀä{á=ä{á`åTÝ'ÜßâöÜ)éã«ìÈê
ådééÜÝã«ådíÀìܼáàdìÞã«äàdèÃã«àTãÛÀÜFéì{ådá«áä 'éå-ã«äÈàèUÝÀß«àdíÀìÜý Nådèê¾ÝÀßàÝ'àáÜ)ç¾Û6ê6ÝÁÜß«ÝÀì{åèÀÜUòöÿiäÈã«Û
HG Só
ÿiäìì8éåÞáÜ
ãÛÀÜ`Ýßäý|åìÁàd í Üéã«äædܱâöÞÀèéã«äàdèFãàtã«ådõdÜ`åtÛÀä dÛÀÜß#æ-ådìÈÞÜô
ÕE(±pÕ+%2×Ó)Ù,ÕÁÒ
z46ÓÕrÖ I
;èµýIê àdÝÀäèÀäàè¾ã«ÛÀÜoÛådß«çÀÜáãNãÛÀäèTã«à ÜãWß«ädÛãNä{á`Ûåè'çìäÈèÝßÜ)éä{áäàè ÝÀß«àdíìÈÜý|á`éàßß«Üéãìê
ÜædÜßê6ÿiÛÀÜßÜdô
âã«ÛÀä{á?ä{áKèÀàdã¹çàdèÜ÷ãÛÀÜ
åìdàßäÈãÛý ý|å êWèÀàdã#éàè6ædÜßdÜd÷àdß¹ý|å êWíÁܯýtÞé4Ûváìà-ÿ#Üß
ãÛådèäÈãièÀÜÜçÀá#ã«àví'Üdô
dà6àç|ÿ¥å êtãà¼é4ÛÀÜé4õtã«Ûå-ãiêdàÞÀߥåìdàßäÈãÛý ä{á&ÿ#àdß«õ6äè`ä{áKã«à¼é4ÛÜé4õtãÛåã¹ã«ÛÀÜ`áàìÞãäàdè|á«å-ã«ä{á Ü)á
åìì?ã«ÛÀÜ `ådßÞ'áÛ =ÞÀÛÀèÚ@Þé4õdÜß`éàdè'çäÈãäàdèá
âöàßNã«ÛÀÜ|ÝÀß«äý|åì?ÝÀß«àdíÀìÜýT÷áäÈè'éܼãÛÜáÜoåß«Ü|èÀÜ)éÜá«á«åß«ê
åè'çÃá Þ véäÜèã±éàèçäÈãäàèá¥ãÛ'å-ã¯ã«ÛÀܼáàìÈÞÀãäàdèí'ÜtàdÝÀãäý|åì.ô]Ú¥ÛÀÜ
`Ú éàdèçäÈã«äÈàèá]ådßÜ 0 !6áôIòBüó
ãÛßà Þ dÛ ò.ú ó÷?ÿiä:ã«ÛµçàãIÝÀßàçÞ'éã«áWí'Üãÿ¹ÜÜèOçå-ã«åÃædÜéãàß«á`ß«ÜÝìåéÜ)ç¾í6êVõdÜßèÀÜì{á=ÿiÛÀÜß«ÜæÜß`ã«ÛÀÜê
åÝÝ'Ü)åß]òöèàã. Ü
[ýIÞ'áãKíÁÜiÜëÝådèçÜç¼åáäÈè 0 !ô¥@ò Büó ß4áã÷áäèéÜ
ä{áèÀàdãKäè ÜèÀÜß«ådì6ãÛÀÜiý|ådÝÀÝÀä è
ÞÀÝÀÝÁÜ߯í'àÞÀèç8÷'ådèç
=$: ød÷Àÿ¹Ü=Ûå æÜ ¾àâ¼ò : ó4ô
:2óô äÈè'åììê÷8ä:âKývàáãNî6ð
á
ådßÜtå-ã=ãÛÜ
àdßNì{åßdÜß±ÝßàíÀìÜý|á÷Áãÿ#àÃçÜ)éàdývÝÁàáäÈãäàdèÃådìdàdß«äÈãÛÀý|á¯Ûå ædÜtíÁÜÜèÅÝßàÝ'àáÜ)çÃã«àÃçÀå-ã«Üô
;èïãÛÜ
é4Û6ÞÀèõäè oývÜãÛÀàç ò¹ àáÜß÷ % ÞÀêdàèµåèçïðKådÝÀèÀäõÁ÷#ø)ùdù
ó4÷àdèÀÜoáã«ådßã4á`ÿiäÈãÛÅåÃáý|åìì.÷åß«íÀäÈãß4åß«ê
($ M G *,+
;è ãÜáãTÝÀÛådáÜ÷±àdèÀÜ ýIÞáãáäývÝÀìêOÜæ ådìÞå-ã«Ü 01!'ôyò*Àø ó4÷
ÿiÛÀä{é4Û ÿiäÈìì±ßÜ !ÞäÈß«Ü
ò
ó#àÝ'Üß«åãäàdè'á÷ÿiÛÀÜß«Ü ä{á&ã«ÛÀÜ=è6ÞÀýIíÁÜß#àâàdÝÁÜß4å-ã«äÈàèá¹ßÜ !ÞäÈß«ÜçvãàtÜæ-åìÞåãܱãÛÀÜ=õdÜßèÜì.ô
¼
Ààß±çàdã±ÝÀß«àçÞéã
åè'çQ DõdÜß«èÀÜìá÷ ä{á ¼ò ó÷8ãÛÀÜtçäývÜèáäàdèàâ?ãÛÜIçÀåã«åvæÜéã«àdß4áô ådäÈèr÷
íÁàãÛã«ÛÀÜ`Üæ-åìÞåãäàdèoàâãÛÀÜ`õÜß«èÀÜìåè'çFàdâã«ÛÀÜNáÞý åß«Ü
ÛÀädÛÀìêvÝådß«ådìÈìÜìÈä)åíÀìÈÜ¥ÝÀßàéÜçÞÀß«Üáô
;èFãÛÀÜ=åí'áÜèéÜ
àâÝ'åß4åììÜìÀÛådß«çÀÿ#ådßÜd÷àdèÜ
éåèFáãäìì'áÝÁÜÜçoÞÀÝFãÜáã¥ÝÀÛådáÜ
íê|åWì{åßdÜiâ,åéãàß÷Àåá
çÜ)áéßäíÁÜç|äèUî6ÜéãäàdèTùô
z q "
ok Ù')*6Ò Ù2Õ'Ò Õ+(±p×&Õ'Ô ÓC6ØdÓÕ'Ô-46#ÀØ7Ù,Ò"
VÜvèÀà-ÿ áÛÀà-ÿ ã«Ûå-ã`ã«ÛÀܼð çäývÜèáäàèUàâ¥îð±ñÃáNéåèïí'ÜvæÜß«êUì{åßdÜFò,ÜæÜèïäÈèèÀäÈãÜ ó4ô ïÜvÿiäìì
ãÛÜè ÜëÝÀìàdܾ߫áÜædÜß«ådì±åßdÞývÜèã«áFådáoãàDÿiÛêd÷±äè áÝÀäÈãÜ àâ`ãÛäá÷=îð±ñÃáÞáÞåììêOÜëÛÀäíÀäÈã dà6àç
dÜèÀÜß4åìä åãäàdèVÝÁÜßâöàdß«ý|åèéÜô ]à-ÿ#ÜæÜßtäÈãoáÛÀàÞÀì{çµíÁÜÃÜývÝÛådáäÜçÅãÛåãvãÛÀÜ)áÜÃådßÜTÜá«áÜèãä{åììê
ÝÀì{åÞ'áäíÀäìäÈãêïådß ÞÀývÜèã«áô ¹ÞÀß«ßÜèã«ìêÅãÛÀÜßÜ Üëä{áã4ávèÀàÅã«ÛÀÜàßêDÿiÛÀä{é4Û +F$ M
M GtãÛåãå äÈæÜè
â,åýväìê¼àâ?î6ð±ñÃá]ÿiäÈìì8Ûå æܱÛÀä ÛåééÞß«åéêvàdèTå däædÜèoÝÀß«àdíìÈÜýTô
VÜIÿiäìì@éåììådèêTõÜß«èÀÜìãÛåã=á«å-ã«äá Ü)á]ñUÜß4éÜß$á¯éàdèçÀä:ã«äàdèUå|ÝÁàáäÈãäædÜ`õÜß«èÀÜì.÷ÁådèçTã«ÛÀÜtéàdß«ßÜ
áÝ'àèçäè¯áÝådéÜ ãÛÀÜ#Üýtí'Ü)çÀçäè¯áÝåéÜôïÜ#ÿiäììåì{áà¯éådììåè6ê=ÜýtíÁÜçÀçäè±áÝåéܹÿiä:ã«ÛNýväèÀäý|åì
çäývÜè'áäàdè|âöàdß]å däædÜèoõÜß«èÀÜìåQ4ýväèÀäý|åì'Üýtí'Ü)çÀçäè¼áÝ'ådéÜô"ïÜ`Û'å ædÜ
ãÛÀÜ=âöàììÈà-ÿiäè
8M M 5AI G *" M M M6 / * I
M G)5AI7CG I . K * K 66M KCM2*,+3G5 M
1 M $ M
( KNM GI I $ MJI M G)5 I(7(*%+G $ 55 I( " M I( K * M / M M $ M M
I
5AM 6 #
*
G 6*6*I/M2 I ) M 6*6 " 6 $AM G G F Vò ó>7 ø
u=Ô-ÕrÕE(
DÞÀèçÜß&ã«ÛÀܱý|ådÝÀÝÀäè
âãÛܯýväèÀäý|åì6ÜýtíÁÜçÀçäèNáÝåéÜiÛåá&çäývÜè'áäàdè
÷ãÛÀÜè ÝÁàdäèã«á&äÈètã«ÛÀܯäý|å Ü¥àâ
é åè|íÁܱâöàdÞÀèçoÿiÛÀàáܯÝÁàáäÈãäàdè¼æÜéã«àdß4áKäè YådßܱìäÈèÜåß«ìê`äèçÜÝÁÜè'çÜèãô
Àß«àdý Ú¥ÛÀÜàßÜýød÷#ã«ÛÀÜáÜUædÜ)éã«àdß4ávéåè íÁܾáÛå-ããÜß«ÜçDí6êDÛê6ÝÁÜß«ÝÀì{åèÀÜ)átäÈè Uô Ú¥Û6Þáví6êOÜäÈãÛÀÜß
ß«Üáãß«äéãäèàdÞÀß4áÜìædÜ)á
ãà î6ð
ñÃáWâöàdßNã«ÛÀÜoáÜÝådß«ådíÀìÜvéådáܾò.î6Üéãäàdè Eôø)ó÷?àdß`âöàßWÿiÛÀä{é4ÛïãÛÀÜFÜß«ßàß
ÝÁÜèådì:ãê
ä{á#åììà-ÿ¹Ü)çNã«àWã4åõܱåììÁæ-åìÞÀÜá±ò,áàWã«Ûå-ã÷ÀäÈâãÛÀÜ
Ý'àäèã4á¹ådßܱìäÈèÜåß«ìêWáÜÝådß«ådíÀìÜ÷å
éådè
íÁÜIâöàÞÀèçVáÞé4Û¾ã«Ûå-ã=ã«ÛÀÜváàìÈÞÀãäàdè¾çÀàÜ)á=äèçÜÜ)ç¾áÜÝådß«åãÜIãÛÀÜýoó4÷rã«ÛÀÜtâ,åýväìêFàdâ#áÞÀÝÀÝÁàdßã`ædÜéãàß
ý|ådé4ÛäÈèÜáÿiäÈãÛ¼õdÜßèÀÜì yéådèvåì{áà`áÛå-ããÜß?ãÛÀÜ)áܯÝ'àäÈèã4á÷ådèç¼ÛÀÜèéÜiÛåáKð OçäývÜèáäàdè
7ïødô
Üã$áiìà6àdõ|åã#ãÿ#à¼ÜëÀåývÝìÈÜ)áô
Oò'<
# DK<
ó1H ò< # 9< ó D <
# D@<
? m- ò -üó
áNäèVãÛÀÜFéådáÜvÿiÛÀÜè
H Tåèçïã«ÛÀÜ|õdÜßèÜìKä{á !Þåçß«åãä{éTò2î6Ü)éã«äÈàè Bó4÷?àèÀÜ|éådèÅÜëÝÀìä{éäÈãìê
é àèáãß«ÞéãKã«ÛÀܯý|ådÝ ¯ô rÜããäè 7=HB) # =,) =÷áà`ã«Ûå-ã Oò'< # D < ó H ò # 7 9:9;9 7 ó ÷ÿ#ܱáÜÜ]ãÛåã
Üåé4ÛçÀäÈývÜèáäÈàèvàâ éàdß«ßÜ)áÝÁàdè'çÀá¹ãàvåIãÜß«ý ÿiäÈãÛ,däæÜèoÝÁà-ÿ#Üß4á¹àdâã«ÛÀÜ 7=äÈèFãÛÜ`ÜëÝåè'áäàdèoàâ
Å ô
;èUâ,ådéã±äÈâKÿ#ÜWé4ÛÀà6àáÜ`ãàFì{åíÁÜìrã«ÛÀÜIéàdývÝÁàdèÀÜèã«á]àdâ
ò<@ó]äèTãÛÀä{á]ý|ådèÀèÀÜß÷8ÿ¹ÜIéåè¾ÜëÝÀìä{éäÈãìê
ÿiß«ä:ã«Ü¯ã«ÛÀÜ`ý|åÝÝÀä è Wâöàß]åè6ê Uåè'ç
( =
# ;9;9 /D #
9 ; ) ) 9:9;9) ,D = ò -ùó
ò<@ó H
:9 D
= )
#
H S
Ú¥ÛÀä{á¥ìÜåçÀá¹ãà
* * MG)5
M * / * $ M ) 6 *" M G KMAG2I( M
H m * M
KM G I I $ M K.* K 6 MJK&M2(*%+NG)5 M I ILKNI!+ M MI $ G 5AI6 # ICK. 6 M M 6 G I M +(
M
M,
Oò'< # D < ó H ò< # 9 < ó D < # D< ? m G 7 #
ò2Ú¥ÛÀÜvÝÀß«ààdâ&ä{á=äèÃãÛÜ ÝÀÝÁÜèçäÈëó4ô|Ú¥Û6Þá
ã«ÛÀÜvð çäývÜèáäàèUàâ¥îð±ñÃá=ÿiäÈãÛ ãÛÜáÜtõÜß«èÀÜì{á±ä{á
7 # 7 ødô á¥èÀàdãÜçTådí'à-æÜ÷6ãÛÀä{á Üã«á¥æÜß«ê|ì{åßdܱædÜß«ê/!ÞÀä{é4õ6ìÈêdô
M
KM G I
$%
I 6 G G
$ I(
M M6 G
$IAG M * M%6 GG I M M M M6 G I / * Oò< # D < ó S GJ< # -<
7 (I @/ * Oò'<PD <ó@ G 0 7 GG $ KM * * M8 ( M ICG M ; *6 # 2 ICK
m : 7 M * M K.*6 #ID6 G
G MGDIAG G *%+NI4G $ 55AI "CM IK * M G>$ G*,+ * M GJM M M6 G
(I 8/ * * MM I 5AM 6 # G 6*6 I/M2 I M 6*6 " 6 $AM G G* M
KMAGI
u=Ô-ÕrÕE( Ú¥ÛÀÜ`õdÜßèÀÜìrý|å-ã«ßäÈë8÷ = 1 Oò'< = DK< ó4÷Áäáiå % ß4åý ý|å-ã«ßäÈëÃò,å¼ý|åãß«ä:ë|àdâçÀàã]ÝÀß«àçÞéã4á
áÜܵò ¯àdß«è÷NøùüúóóIäè Ãô ¹ìÜådßìêÅÿ#ÜÃéådè é4ÛÀà6àáÜFã«ß«ådäèÀäèïçå-ã«åÅáÞé4Û ã«Ûå-ãFådììià çä{ådàdè'åì
ÜìÜývÜèã«á =) éåè¼íÁÜ]ý|ådçÀÜ¥ådßíä:ã«ß«ådßäìê±áý|åìì.÷åèçtí6êIåááÞÀývÝã«äàdèWådìÈìÀçÀäådàèåì6ÜìÜývÜèã«á = )
åß«Ü]àd â $tòø)óôÚ¥Ûܯý|å-ã«ßäÈ ë ä{áKãÛÀÜè|àââöÞÀììÁß«ådèÀõÁûÛÜèéÜ]ãÛܱáÜã¹àdâædÜéãàß«á÷ÿiÛÀàáÜ
çàã¹ÝÀßàçÞ'éã«á
äè âöàdß« ý ï÷ådßÜ#ìäèÀÜådßìê¯äèçÀÜÝÁÜèçÜèãi%ò ]àßè÷'øùdüúdóûdÛÀÜè'éÜ÷íêIÚ¥ÛÀÜàßÜýþø÷ã«ÛÀÜiÝ'àäèã4áéåètíÁÜ
áÛå-ããÜß«Üç|íêvÛ6ê6ÝÁÜß«ÝÀì{åèÀÜ)á?äè U÷åè'çvÛÀÜè'éÜ=åì{áà`í6ê|áÞÀÝÀÝÁàdßã¹æÜéã«àdß#ý|ådé4ÛÀäèÀÜ)á?ÿiäÈãÛoáÞ |éäÜèã«ìê
ì{åß dÜ
Üß«ßàßiÝ'ÜèåìÈãêdô#î6äèéÜ=ã«ÛÀä{áiä{á#ãß«ÞÀÜ`âöà߯åè6/ ê 'èÀäÈãÜ`è6ÞÀýtí'ÜßiàâÝ'àäÈèã4á÷Àã«ÛÀÜNð çäývÜè'áäàdèoàâ
ãÛÜáÜNéì{ådá«áä Üß«á&ä{á¹ä è 'èÀäÈãÜô
àãÜ=ã«Ûå-ãiã«ÛÀÜWåááÞÀývÝã«äÈàèá¹äÈèFãÛÜ=ãÛÀÜàdß«ÜýYådßÜ`áãß«àd è Üß#ãÛådèTèÀÜéÜá«áådßêÃòöãÛÀÜêoÿ#Üß«Ü=é4ÛàáÜè
ãàý|åõÜNã«ÛÀÜvéàdèèÀÜéãäàdèUã«àTß«åçä{åìí'ådáäá¯âöÞÀèéãäàdè'á
éìÜåßó4ô
;è â,ådéã
äÈã=ä{á±àèÀìêTèÀÜéÜá«áådßêãÛ'å-ã :
ãß4åäèÀäè tÝÁàdäèã«áiéådèTí'ÜWé4ÛàáÜèUáÞ'é4ÛFã«Ûå-ãiã«ÛÀÜNß4åèÀõFàâãÛÀÜWý|å-ãß«äÈë = äèéßÜ)ådáÜá¹ÿiäÈãÛÀàÞã]ìäÈýväÈã
ådá :@äèéß«ÜåáÜ)áô ÀàdߥÜëådývÝÀìÜ÷âöàdß % åÞá«áäåd-è ÅõÜß«èÀÜì{á÷dãÛäá¥éåèFådì{áàIí'Ü`ådééàývÝÀìäáÛÀÜ)çFò,ÜædÜè
âöàdßNãß4åäèÀäè TçÀå-ã4åTßÜ)áã«ßä{éã«Üç¾ãàÃìäܼäèVåTíÁàdÞèçÜçÅáÞÀíáÜãIàâ m ó=í6ê é4ÛÀà6àáä è Táý|ådìÈì?ÜèÀàd Þ dÛ
Ãÿiä{ç6ã«Ûáô ¯à-ÿ¹ÜædÜßäè dÜèÀÜß4åì6ã«ÛÀÜið µçÀäÈývÜèáäÈàèWàdâî6ð
ñ éìåááä Üß4áéåèvéÜßã4åäèÀìê`íÁÜ
èÀäÈã«Ü÷ådèçWäèçÀÜÜç÷-âöàdßKçÀåã«å±ßÜ)áã«ßä{éã«ÜçNãà=ìäÈܹäèIå
í'àÞÀèçÜ)çIáÞÀíáÜã?àdâ m ÷é4ÛÀà6àáä è ]ß«Üáãß«ä{éã«äÈàèá
àdèFãÛÜ µÿiä{ç6ãÛ'á¥äáiå ààçFÿ#å ê¼ã«àvéàdèã«ßàì8ãÛÀÜ`ð çäývÜèáäàdèô
Ú¥ÛÀä{á`éådáÜ äædÜá
ÞáWåáÜéàdèçïàÝÀÝ'àßã«ÞÀèÀäÈãêTãàUÝßÜ)áÜèãNåUáäÈãÞåãäàdèÃÿiÛÜܼ߫ãÛÀÜoî6ð
ñ áàìÞãäàdè
éådèÅí'ÜéàývÝÀÞãÜ)çÅåèådìêã«ä{éåììê÷ÿiÛÀä{é4ÛµådìáàUåývàÞÀèã«á=ã«à åÃáÜéàdèç8÷Kéàdèáãß«Þéã«äædܼÝÀß«ààdâiàâiãÛÜ
Ú¥ÛÀÜàdß«ÜýTô Ààßvéàdè'éß«ÜãÜèÀÜá«átÿ¹Üÿiäìì&ã4åõdÜFãÛÜTéådáÜoâöàß % åÞá«áä{ådè* õdÜßèÀÜì{áWàdâ]ãÛÀÜâöàßý
O'ò < DK< ó H
#
7 < 7 < ô rÜã±Þá¯é4ÛÀà6àáÜ
ãß4åäèÀäè tÝÁàdäèã«á]áÞé4ÛãÛåã]ãÛÜWáý|ådììÈÜ)áã¥çä{áã4åèéÜ
íÁÜãÿ#ÜÜèåè6êvÝådäÈß#àâ@Ý'àäÈèã4á¹ä{á¹ýIÞ'é4Û|ì{åß dÜß&ãÛådè|ã«ÛÀÜ=ÿiäçãÛ ô+¹àèáäçÀÜß¹ãÛÀÜ`çÜ)éä{áäàè¼âöÞÀèéã«äàdè
Üæ-ådìÈÞ'å-ãÜ)ç|àèFã«ÛÀÜNáÞÀÝÀÝÁàdßã¥æÜéãàdß
ò )ó H ( "=L
= 7
7 7 I ò2ü
Só
=
7 Ñø
H
7 7 H `ø
7 7 H S ò2üÀø ó
ÿiÛÀä{é4ÛTåß«Ü`áåãä{á 'Üç|í6ê
7
H
7 7
7 H
7 7
7 ò2ü
ó
H
7 7
è = ådßÜ=åì{áàIÝÁàáäÈãäædÜd÷6ådìì'ã«ÛÀÜ
`Ú éàdèçÀä:ã«äàdèá#åèçTéàdèáãß4åäèã«á#åß«Ü
Ú¥Û6Þá÷'áäÈè'éܱãÛÀÜ`ß«ÜáÞÀìÈãä
«á å-ãä{á Üç÷ådèçFÿ¹Ü`ýtÞáã#Ûå æÜ
âöàdÞÀèçoãÛÜ dìàdíådìáàìÈÞÀãäàdè¾ò,ÿiäÈãÛ,Üß«àIãß4åäèÀäèWÜßß«àdß4á4ó4ôKî6äèéÜ
ãÛÜ
è6ÞÀýtí'Üß]àâ?ã«ß«ådäÈèäÈètÝÁàdäèã«á÷ÁådèçãÛÜäß]ì{åíÁÜìäèÀ÷Àä{á±åß«íÀäÈãß4åß«ê÷åè'çãÛÀÜêTådßÜNáÜÝ'åß4å-ãÜ)çTÿiäÈãÛÀàÞã
Üß«ß«àdß÷ã«ÛÀÜ`ð çÀäÈývÜèáäÈàè|ä{á#äÈ è èÀäÈãÜdô
Ú¥ÛÀÜNáäÈãÞåãäàdèoäáiáÞÀývý|ådßäÜ)çváé4ÛÜý|å-ã«ä{éåììêNäè @äÞÀßÜtødødô
;Á$),
4<<> Ä 4-*.*,"$ ¦ º³ * 9?*,i;"$ 1, §*.!¥4 r"$)1,+T ;$*2*2" 9§| 2( " 12.," §o$. -!K(d 9
Å
12.4" " i² " 21 * .23T 1, ǧ / ¯ 12+7 -*r+-0 ?" " 3 12&³´F)" !¥ *2" 3 3 3 0
3 3 0 3 0 3 ° 3 Â)3 3 043
z q " !6ÒÔ$#%2Ù'&
#ÀÓÙ,ÕÁÒ uÔ$(ÕÔ$)*#ÀÒØ
¾Õ+(±p 04
B ÷?ÿiäÈãÛ Þ'áãtãÛÀß«ÜÜFèÀÜáãÜ)ç
éì{åááäÜß4á`éàßß«ÜáÝ'àèçÀá=ã«à¾àèÀÜoàdâiãÛÀÜTáÜã«áIàâ¯éìåááäÜß4á
äè äÞÀß«Ü À
áÞÀíáÜã«á¥àdâ@âöÞèéã«äÈàèá÷åèçFÿiäÈãÛ # H ø÷
< H ÷åèç
< 8 <
H E
À ô
Φ=0
Φ=1
D=2
M = 3/2
Φ=0
Φ=−1
Φ=0
8I: ) *
m $ M
KMAG2I(
I+ 5 I6*M 6 GG MG I K4* K+$ K
K +* = B 7K K+$ K KM M N
: G>
I $ <7 !M I"CM # # K.* -
= = B D . 7
ø
àdß
ãÛÀÜtÝÀß«à6àâ&ÿ#ÜtåááÞÀývÜ`ã«ÛÀÜtâöàdììà-ÿiäè¼ìÈÜývý|å÷ÁÿiÛÀä{é4ÛÃäèOòöðKådÝÀèÀäõÁ÷Køù'ùó¥ä{á±ÛÀÜì{çÃãàFâöàdììà-ÿ
âöß«àdýYáê6ývývÜãß«êvåßdÞÀývÜèã«á
/ ) )*# ¹àèáäçÀÜ1 ß O51
7Oø¯ÝÁàdäèã«á¹ìêäè`äèoåWíåìì ? m ô1Üã¹ãÛÀÜ
Ý'àäèã4á&íÁÜ=áÛåãã«Üß4åíÀìÜ
í6ê åÝtãàìÈÜß«ådèãéìåááäÜß4á@ÿiä:ã«ÛIý|ådß äè ô?Ú¥ÛÀÜè¼äè¼àdß4çÜßâöàß ã«à`í'Üiý|åëäÈýväÜç8÷)ãÛÀܯÝ'àäÈèã4á
ý ÞáãiìäÜ=àdèãÛÀÜWædÜßãä{éÜá#àâ?åèÅòODø)ó çäývÜèáäàèåì8áê6ývývÜãß«äé±áäývÝÀìÜëÁ÷åè'çFýtÞáã¯åì{áàIìäܱàè
t
ãÛÜNáÞÀßâ,ådéÜ=àâã«ÛÀÜ`íådìÈì.ô
u=Ô-ÕrÕE(
òO(.
ïÜWèÀÜÜ)çFàèÀìêFéàdèáä{çÜß#ãÛÜWéåáÜ`ÿiÛÀÜßÜ=ã«ÛÀÜNè6ÞÀýtí'Üßiàâ?ÝÁàdäèã«á OÅáåãä{á Üá O 1 7 øô
7 VøiÝÁàdäèã«á?ÿiäììèÀàã¹í'Ü
áÛåãã«Üß4åíÀìÜ÷áäèéÜ¥ãÛÀܱð OçÀäÈývÜèáäÈàèWàdâ8àßäÜèã«ÜçtÛ6ê6Ý'ÜßÝÀì{ådèÀÜáäè
m äá 7Ãø÷6åèç¼ådèêNçÀäáãß«äíÀÞã«äÈàèNàâ8ÝÁàdäèã«áÿiÛÀä{é4Ûtéåètí'Ü]áÛåããÜßÜ)çWí6êWååÝIãàdìÜß«ådèãéì{ådá«áä 'Üß
éådèFådì áàWíÁÜ=áÛåãã«Üß«Üç|í6êoådèoàßäÜèã«ÜçvÛ6ê6Ý'ÜßÝìådèÀÜûãÛäá¥ådìáàWáÛÀà-ÿ]á&ãÛåã
< 1 7Dø ó4ô å6äô"è|Vÿ#ÜÜ
éàèáäçÀÜßKÝÁàdäèã«á?àèoåNáÝÀÛÀÜßÜ]àdâçä{åývÜãÜß U÷6ÿiÛÜß«ÜiãÛÀÜ
áÝÛÀÜß«Ü]äÈã«áÜìÈâä{á?àâçäývÜèáäàè
ÿiäììèÜÜçIãÿ#à±ß«ÜáÞÀìÈã«á@âöß«àdý î6Üéãäàd è Eô E÷dèådývÜìêvòø óä:E
â OFä{áÜæÜè÷ÿ¹Ü]éåè èç¼å=çä{áã«ßäíÀÞã«äàdèNàd+â O
ÝÁàdäèã«á#ò{ãÛÜ#æÜßãä{éÜ)áàdâÀãÛÜ='ò O|ø ó ;çäývÜèáäàdèådìáê6ývývÜã«ßä{é?áäývÝÀìÜëóÁÿiÛÀä{é4ÛIéåèWíÁÜiáÛ'å-ãã«Üß«Üç`í6ê
ådÝtãàdìÜß«ådèãKéìåááä Üß4á@ä:â
= = B H OÃø÷6åèçTò óäÈâ Oä{áàçÀç8÷6ÿ#ܯéådè 'èçvåNçä{áãß«äÈíÞãäàdè
àP â O¾ÝÁàdäèã«á¥ÿiÛäé4ÛéådèFíÁÜNáàváÛå-ããÜßÜ)çoäÈâ -
= = H ò O ø ó ò O 7 ø ó =$O, ô
B
âPO ä{á#ÜæÜè÷Áå-ãiývàáã ï
O ÝÁàdäèã«áiéådèFíÁÜNáÛå-ããÜß«ÜçFÿiÛÀÜèÀÜæÜß
O ø 1
= =
B
A OPI Eó
ò2ü
Ú¥Û6ÞáNâöàdß O ÜæÜèïãÛÀÜvý|åëäÈýtÞÀý è6ÞÀýtí'ÜßNàâ¥Ý'àäèã4á
ãÛåãIéådèVí'ÜoáÛåããÜßÜ)ç¾ý|å ê¾í'Ü|ÿiß«äÈãã«Üè
-
= =
B
7 ødô
â Å
O äá±àçÀç8÷åã
ývàáãOOÝ'àäÈèã4á±éåèÃíÁÜtáÛåããÜßÜ)çÃÿiÛÀÜè
= = YòO
H 7 ø)ó
=7O
ø ó 'ò O g ô
]à-ÿ#ÜæÜß÷6ãÛÀÜ ! Þådèã«äÈãêvàdèFãÛÀÜ`ß«äd Ûã#Û'åèçáä{çÜ`á«å-ã«äá Ü)á B
5+A
O òO ø ó 'ò O 7 ø ó =7O A O Dø ò2üLB6ó
ýtÞáã]á«å-ã«äáâöê
< àâ åÝïãàìÜß4åèã=éì{ådá«áäÜß«á
<
1
7
= B
ø
I ò2üúó
Ú¥ÛÀä{á¥ß«ÜáÞÀìÈã÷6ãàdÜãÛÀÜßiÿiä:ã«Û
< 1 7 ø÷Áéàdè'éìÞçÜ)á&ã«ÛÀÜ`ÝÀß«à6àâ;ô
$% 5 (I 6*M 6
G
G MG & $ $ 6 G * K. ( I &
G
rÜ ã $áiáÜܱÛà-ÿ ÿ¹Ü=éådèFçàváãß«Þéã«ÞÀß4åì8ßä{áõIýväèÀäýväåãäàdèIÿiäÈãÛ åÝoãàdìÜß«ådèã¥éì{ådá«áä 'Üß4áô,VÜ=èÀÜÜ)ç
àdèìÈê=éàèáäçÀÜßã«Ûå-ã?áÞíáÜãàdâãÛÀÜ
¯÷déådììdäÈã
r÷-âöàdß@ÿiÛÀä{é4Û`ãß4åäèÀäè áÞééÜÜçÀáK÷ ÿiÛÀÜß«Ü&í6ê`áÞééÜá«á
ÿ#ÜWývÜådèTãÛåã=ådììrãß4åäèÀä è vçÀåã«åFåß«ÜIådá«áä èÀÜçTåoìådí'Üì1? - IøC.òöèÀàdãÜIãÛåã±ã«ÛÀÜáÜWì{åíÁÜì{á]çÀàoèÀàdã
Ûå æÜWã«àéàdäèéä{çÜIÿiäÈãÛÃã«ÛÀÜ|ådéãÞådììådí'Üì{á÷ä.ô ÜdôNãß4åäèÀä è FÜß«ßàß«á±åß«ÜtådììÈà-ÿ#Üç'ó4ô äÈãÛÀäè
÷ 'èç
ãÛÜNáÞÀí'áÜã]ÿiÛÀä{é4Û äÈæÜá¹ãÛÜ=âöÜÿ#Üáãiãß4åäèÀä è WÜßß«àdß4á ¹éådìì'ã«ÛÀä{á¥è6ÞÀýIíÁÜßiàdâÜßß«àdß4á = B ô äÈãÛÀäè
ãÛ'å-ãiáÞÀí'áÜã ÷ èçoãÛÀܱâöÞÀèéãäàdL
è =TÿiÛÀä{é4/ Û däæÜá?ý|åë6äýtÞÀý ý|ådß äèTò2åèçoÛÀÜè'éܱãÛÀÜ
ìà-ÿ¹Ü)áã&íÁàdÞèç
àdè¾ãÛÀܼð çÀäÈývÜèáäÈàè'ó4ô àdãÜIãÛÀܼæ-åìÞÀÜIàâ¹ãÛÀÜtßÜ)áÞì:ã«ä è |ß«äáõí'àÞÀèçDò{ãÛÜtß«ä dÛã±Û'åèç áä{çÜIàâ
01!'ôiò Eó4÷ÞáäÈ è =ãÛÀÜ
í'àÞÀèçvàè¼ãÛÀÜ
ð çäývÜèáäàdètäèvÝÀì{ådéܯàâ8ãÛÀÜ
ð çäývÜèáäàdè'óô Üë6ã÷6ÿiäÈãÛÀäè
÷ è'çTãÛåã
áÞÀíáÜã¯ÿiÛäé4Û äædÜ á = B 7 ø`ã«ß«ådäèÀä è ¼Üß«ßàß«áô ådäÈèr÷'ÿiäÈãÛÀäèãÛ'å-ã
áÞÀíáÜã÷ 'èç
ãÛC Ü =µÿiÛÀä{é4Û äÈæÜá¯ãÛÀܼý|åë6äýtÞÀý_ý|åß däè÷åèç èÀàã«ÜtãÛÀÜ|éàdß«ßÜ)áÝÁàdèçÀäÈ è |ßä{áõTí'àÞÀèç8ô
ãÜß«åãÜd÷
åè'ç|ã4åõdÜ
ã«Ûå-ã¯éìåááä Üß¹ÿiÛÀä{é4Û äædÜá&ãÛÜ`à-ædÜß4åììÁýväèÀäýIÞÀý ß«äáõvíÁàdÞÀèçô
è¼ådìÈãÜß«èåãäædÜ&ådÝÀÝÀß«àåé4ÛWä{á@ãàNçäæ6ä{çÜKã«ÛÀÜ¥âöÞÀèéãäàdèá Åäèã«à±èÜáãÜç¼áÞÀíáÜã4á = D F ? D F ø÷
då á¯âöàdììà-ÿ]á¥ådìÈì =*? =¹Ûå ædÜ- D .¼áåãä{áâöêäè = 1 F«ô=Ú¥Û6Þá±ãÛÜNâ,åýväìêoàdâKâöÞÀèéãäàdè'á
äè =KÛåá]ð çäývÜèáäàdèFí'àÞÀèçÜ)çÃåíÁà-ædÜ`í6êýväÈèò'F D ó?7 øô àdãÜWådì{áàtãÛ'å-ã = = # ô±î]ñ
ãÛÜèÅÝÀß«àéÜÜ)çÀá`í6ê ã«åõ6äèãÛ'å-ã&=µâöàdßWÿiÛÀä{é4Û ã«ß«ådäèÀäèáÞééÜÜ)çÀáNäèVÜåé4ÛÅáÞÀí'áÜãIåèçïâöàßNÿiÛÀä{é4Û
ãÛÜIÜývÝÀäßä{éådì8ßä{áõFä{á]ýväèÀäývä Ü)ç|äèUã«Ûå-ã=áÞÀí'áÜã÷råèç¾å ådäè÷Áé4ÛÀà6àáä è ¼ã«Ûå-ã = ÿiÛÀä{é4"Û däædÜ)á¥ãÛÜ
ìà-ÿ¹Ü)áã¥à-ædÜß«ådììß«äáõvíÁàdÞÀèçô
àãÜ&ãÛåã@äÈãä{áÜ)ááÜèãä{ådì-ãàiã«ÛÀÜáܹådß ÞÀývÜèã4áÁãÛåããÛÜ&íÁàdÞÀèço
ò EórÛÀàdì{çÀáâöàdß #Ké4ÛÀàáÜèNçÜéä{áäàdè
âöÞÀèéãäàdèr÷èÀàdãÞáãKãÛܯàdèÜiãÛåã&ýväèÀäýväÜáãÛÜ]ÜývÝÀäß«ä{éåì6ß«ä{áõoò,àã«ÛÀÜß«ÿiä{áÜiÜìÈäýväèåãäè¯áàdìÞã«äÈàèá
âöàdßiÿiÛäé4ÛTáàdývܱãß4åäèÀä è NÝÁàdäèã <ïá«å-ã«äá Ü)á ='ò <@1 ó H Stÿ#àdÞÀì{çFäèæ-ådìÈä{çÀåãÜiã«ÛÀÜNåß dÞývÜèã4óô
Ú¥ÛÀÜFßÜ)áÞì:ã«ä
è åÝVã«àdìÜß4åèãWéìåááä Üß=ä{áNäèVâ,åéãtå¾áÝÁÜéäådì&õ6äèçVàâ]áÞÀÝÀÝÁàdßãIæÜéã«àdßIý|ådé4ÛÀäèÀÜ
ÿiÛÀä{é4ÛváäývÝÀìê`çà6Üá&èÀàã&éàdÞÀèã¹çå-ã«å
â,ådììÈäè±àdÞã4áä{çÜ¥ã«ÛÀܯáÝÀÛÀÜßÜ]éàdèã«ådäÈèäÈè±åììã«ÛÀÜiã«ß«ådäÈèäÈè
çÀåã«åÀ÷
àdß¼äèáä{çÜ|ã«ÛÀÜUáÜÝåß4å-ã«äÈ è Ãý|åß däè÷KådávådèOÜßß«àdßô
ã|áÜÜý|áIædÜßêÅß«ÜåáàèåíÀìÜvã«àïéàdèéìÞçÜFãÛåã
áÞÀÝÀÝÁàdßã#æÜéã«àdߥý|ådé4ÛäÈèÜá÷ÿiÛÀä{é4Ûoådßܱãß4åäèÀÜ)çvÿiäÈãÛoæÜß«êváäýväì{åß?àd í Üéã«äædÜá÷åì{áà åäèoåtáäÈýväì{åß
õ6äèçVàâ±éåÝåéäÈãê éàdèãß«àdì&âößàý ã«ÛÀÜäßIãß4åäèÀä è 3
ô ]à-ÿ#ÜædÜß÷?" å åÝVã«àdìÜß4åèãWéì{åááä Üß`ä{áNèàãtådè
î6ð
ñV÷ÁådèçUáà¼ã«ÛÀÜWådß ÞÀývÜèã]çà6Üá¯èÀàã¯éàdèáãäÈãÞÀãÜWå¼ß«ä dàdß«àdÞ'á¹çÜývàdèáãß4å-ã«äàdèoàdâKáã«ßÞéãÞÀß4åìßä{áõ
ýväèÀäývä å-ã«äàdè±âöàdß¹î6ð
ñÃáô?Ú¥ÛÜiàdß«ä äèåìåß dÞÀývÜèãâöàßKáã«ßÞéãÞÀß4åìß«ä{áõWýväÈèäÈývä )å-ãäàè¯âöàß¹î6ð±ñ¾á?ä{á
õ6èÀà-ÿiètãà`íÁÜ 'å ÿ#Üç8÷áäèéÜ¥ãÛÀÜ]áãß«Þéã«ÞÀß«Ü¥ãÛÀÜßÜ]ä{á?çÜãÜßýväèÀÜ)çWí6êNã«ÛÀܯçÀåã«åvò2áÜÜNòöðKåÝèÀäõ'÷Áøùùúdó÷
î6Ü)éãäàè¾ú6ôødø)óô
iíÁÜìäÜæÜ`ãÛ'å-ã
ãÛÀÜßÜIäá±åvâöÞÀßãÛÜß`áÞíãìÜNÝÀß«àdíìÈÜý ÿiäÈã«ÛUãÛÀÜtàdß«ä däèådìråß dÞÀývÜèãô
Ú¥ÛÀܯáãß«ÞéãÞÀß«Üiä{á?çÜ èÜçváà=ãÛåã&èÀà=ãß4åäèÀä è
ÝÁàdäèã«áKådßÜ¥ývÜýtíÁÜß4á@àdâ8ãÛÀÜ]ý|ådß äèWáÜã$ ô ]à-ÿ#ÜæÜß÷
àdèÜKýIÞ'áã@áãäììdáÝ'Ü)éäÈâöê=ÛÀà-ÿïãÜ)áãÝ'àäÈèã4áãÛ'å-ãâ,ådììdäèãà¥ãÛÜ&ý|åß däè=åß«Ü?ãà±í'ܹì{åíÁÜìÜç8ô
â'àdèÀÜ&áäývÝÀìê
ådá«áä èá@ãÛÀÜ
áådývÜ÷ ë6Ü)çvéì{ådá«áãà=ã«ÛÀÜýyò2áå ê7Iø)ó4÷ãÛÀÜè¼ãÛÀܯð OçÀäÈývÜèáäÈàèWÿiäììí'ܯÛÀädÛÀÜß #&ãÛådè
ãÛÜ`í'àÞÀèççÜß«äædÜçoäèFÚ¥ÛÀÜàdß«Üý ô¯à-ÿ¹ÜædÜß÷ãÛÀÜNá«åývܱä{á&ã«ßÞÜ`ä:â@àdèÀÜ=ì{åíÁÜì{á&ãÛÀÜý åììådá#Üßß«àdß4á
ò,áÜܹãÛÜ ÝÀÝÁÜèçäÈëó4ô
â'àèÀÜ#ìådí'Üì{áãÛÜý ådìÈì6ådá@éàdß«ß«Üéã ÷ àdèÀÜ¥åß«ß«äÈæÜáråã åÝNãàdìÜß«ådèãéì{ådá«áä 'Üß4áô
$±èFãÛÀÜ`àdãÛÀÜߥÛåèç÷Àä:ãiä{á#õ6èÀà-ÿièoÛÀà-ÿ ãàvçàváã«ßÞ'éãÞß«ådìÁß«ä{áõ¼ýväèÀäýväåãäàdèIâöàdßiáê6áãÜý|á#ÿiÛÀÜß«Ü
ãÛÜáãß«Þéã«ÞÀß«Üçà6Ü)átçÀÜÝÁÜèçOàdèµã«ÛÀÜTçÀå-ã4å ò.î6Ûå ÿ#Ü ;Ú@å ê6ìàdß`Üã¼ådì2ôÈ÷]øùùåû#î6Û'å ÿ¹ÜÚå ê6ìÈàß`Üãvåì.ôÈ÷
øùù dí'ó4ô ¯èâöàdßãÞèå-ã«ÜìêÃã«ÛÀÜvßÜ)áÞì:ã«ä
è FíÁàdÞèçÀáWådßܼýtÞé4ÛVìààáÜß
ãÛådè ã«ÛÀÜ|ð íÁàdÞÀèçáNåíÁà-ædÜd÷
ÿiÛÀä{é4ÛTåß«ÜNåìßÜ)ådçê¼æÜß«êoìà6àáÜtò,ÿ¹Ü`ÿiäììÜëÀåýväèÀÜ=åIãê6ÝÀä{éåìréådáÜ`í'Üìà-ÿ ÿiÛÀÜßÜ
ãÛÜNð íÁàdÞèçTä{á
åoâ,ådéã«àdß=àâ±ø S
SoÛÀä dÛÀÜ߯ã«Ûåè¾ãÛÀܼývÜ)ådáÞÀßÜ)çãÜ)áã`Üß«ß«àdßó4ôWÚ¥Û6Þá=å-ã
ãÛÀܼývàývÜèã±áãß«Þéã«ÞÀß«ådìßä{áõ
ýväèÀäývä å-ã«äàdè¯ådìÈàèÀܹçÀàÜ)áèÀàãÝÀß«à-æä{çܹå + I I $ GÜëÝÀì{ådèå-ã«äÈàè`ådá@ãà
ÿiÛ6êWî6ð±ñÃáàdâ{ãÜètÛå æ1 Ü dà6àç
dÜèÀÜß4åìä åãäàdè±Ý'Üßâöàßý|ådèéÜdô ]à-ÿ#ÜæÜß÷)ã«ÛÀÜ¥åíÁà-ædܹådß ÞÀývÜèã4áráã«ßàè dìê=á
Þ dÜ)áããÛåãådì àdß«äÈãÛÀý|á
ãÛ'å-ã±ýväèÀäývä Ü = éådèUí'ÜIÜëÝÁÜéã«Üçã«à däæÜ
íÁÜããÜß ÜèÀÜß«ådìÈä )å-ãäàèvÝ'Üßâöàßý|ådèéÜdô ÞÀßã«ÛÀÜß
Üæ6ä{çÜèéÜ?âöàßã«ÛÀä{áä{árâöàdÞÀèçWäè=ãÛÀܹâöàdììà-ÿiäè &ã«ÛÀÜàßÜý àâ@ò,ðKåÝÀèÀäõÁ÷øùùdüó÷)ÿiÛäé4ÛWÿ¹+ Ü !ÞàãܹÿiäÈãÛÀàÞã
ÝÀß«à6àâ K
8 II25 K 6 #!5AM)576 MG 5 G
G*,+ *
IE$+ *
MI +* / M " M
vò ó 1 =
ò2üó
/ M M6vò ó G * M57 I *6 #I M I("I $ M M G GJM * MM 5AM ( II * M 6*M* G"I(" M
:
6*6 * *%+ GJM GIG M ïø 7 * MM 5 M ) I(I( $ M + GI"CM 6*6 * *%+NG M GIG2 M
:
:
ò,) = ó
k 1 A F ò2ü'dó
ò,ÿiÛÀÜß«Ü`ÿ¹Ü=Ûå æÜ
åådäÈèFßÜÝÀì{ådéÜç
ò'<,=.ó 9
ò< )ó#í6ê ò<,= D < )óó¥áÞÀíÜéãiã«à
( = H ø ò2ù
Só
=
N= S ò2ùÀø ó
ÿiäÈãÛUáàdìÞã«äÈàè/däæÜè|í6ê
k H ( N=
ò<,=ó I ò2ù
ó
=
; $),
< > º ² ² 21 ;1 . * ;"$.;$* 3 (d 0 !#. 0 . * 0 2* * 4912Á.! 0° 4d 3 ]. » 12.4" 3 " 3 1,+- 0 1,12] " 3
Á
12*@1,+- ; Ã0 ° " *2" 0 *,-2:9
3-0 3 043 5
& I$ 7 G 7 * M D $ 6 G
rÜãIÞáIÝÀÞãtãÛÜáÜoàdí'áÜßæ-å-ã«äàdèá
ã«à áàdývÜ|ÞáÜô áIývÜèãäàèÀÜçVåíÁà-ædÜd÷@ã«ß«ådäèÀäèÃåèOî6ð
ñ
éì{åááäÜßvÿiäìì±ådÞãàý|å-ãä{éåììê äædÜUæ-åìÞÀÜ)ávâöàdßFãÛÜ ÿ#ÜädÛã«á÷]è6ÞÀýtí'ÜßoàâWéÜèãÜß«á÷±éÜèãÜß
ÝÁàáä:ã«äàdèá÷åèç¾ãÛßÜ)áÛÀàì{ç8/ ô Ààdß % ådÞá«áä{åè á÷ã«ÛÀÜß«Üvä{á
àdèÀìêÃàèÀܼÝåß4åývÜãÜß
ìÜâ{ã `ãÛÀÜ
ÿiä{ç6ãÛgò äÈè 01!'ô ò,
ü SóóFòöÿ#ÜTådá«áÞÀývÜoÛÀÜß«ÜoàèÀìêïàèÀÜ ÿiäçãÛµâöàßWã«ÛÀÜTÝÀß«àdíÀìÜýoóô #ådèµÿ#Ü
èç|ãÛÀܱàdÝã«äý|åìæ-ådìÈÞÜiâöàdß¹ãÛåã&ã«àà÷íêvé4ÛÀà6àáä è =ã«Ûå-ã UÿiÛÀä{é4ÛvýväèÀäývä Üá = ä dÞßÜNø B
áÛÀà-ÿ]áå]áÜß«äÜáàdâÜëÝ'ÜßäývÜèã4árçàèÀÜKàd è dü-Në ü
îÚÅçä däÈãçÀåã«åÀ÷ ÿiäÈã«Ûvø S÷ S
S
S#ãß4åäèÀä
è ¥ÝÁàdäèã«árådèç
SÀ÷ S
S
S]ãÜáã?Ý'àäèã4áôÚ¥ÛÀÜ#ãàÝtéÞÀßæܹäèWãÛÀÜiìÜâ{ãÛådèçWÝ'åèÀÜìÀáÛÀà-ÿ]á@ã«ÛÀÜ¥ð VíÁàdÞÀèçTò,ä2ô Üô@ãÛÀÜiíÁàdÞèç
ß«ÜáÞÀìÈãä è ±âößàý[åÝÝÀßà ëäý|å-ã«äè iãÛÜið OçÀäÈývÜèáäÈàèNäè 01!'ô¥%ò Eó 8 í6ê 01!'ô¥ò2üúóó4÷ãÛÀܯýväççìÜ#éÞÀßæÜ
áÛÀà-ÿ]á¹ãÛÀÜ`íÁàdÞÀè'ç|âöß«àdý ìÜ)å ædÜ àdèÜ àdÞã±ò 01!'ô]ò,ù Eóó÷'åè'ç|ã«ÛÀÜ=í'àdããàý éÞÀßæÜ
áÛÀà-ÿ]á¹ãÛÜ`ývÜådáÞÀß«Üç
ãÜ)áã¥Üß«ß«àdßô ¹ìÜådßìê÷äèvãÛÀä{á#éåáÜd÷ã«ÛÀܱíÁàdÞèçÀá¥å߫ܯæÜß«êIìà6àáÜô?Ú¥ÛÀÜ
ßä dÛã&Û'åèçvÝ'åèÀÜì8áÛÀà-ÿ]á Þáã
ãÛÜtð í'àÞÀèçDò{ãÛÜIã«àdÝVéÞÀß«ædÜd÷8âöàdß .
S
Só÷8ãà Üã«ÛÀÜß
ÿiäÈãÛÃã«ÛÀÜtãÜ)áã`Üß«ß«àdß÷8ÿiäÈã«ÛÃãÛÜIì{å-ããÜß
á«éåìÜ)ç¾ÞÀÝÅí6êïåâ,ådéã«àdßNàdâ
ø S
SVòöèàãÜ|ãÛåãNãÛÀÜ|ãÿ¹àUéÞßæÜáNéßàáá4ó4ô
ãIäá`áãß«äõ6äÈ è oã«Ûå-ãNã«ÛÀÜvãÿ#à
éÞßæÜá`Ûå æÜvýväèÀäý|å|äèïãÛÀÜoá«åývÜvÝìåéÜ NãÛ6ÞáWäÈèïã«ÛÀä{áNéåáÜd÷@ã«ÛÀÜoð íÁàdÞÀèç÷?ådìÈãÛÀà Þ dÛïìààáÜd÷
áÜÜý|á?ãàWíÁÜ
èÀÜæÜßãÛÀÜìÈÜ)ááÝßÜ)çä{éãäæÜ
ô 0KëÝ'ÜßäývÜèã4áàdèoçä ä:ã4á =ã«ÛÀßà Þ dÛ|ùIáÛà-ÿ¹Ü)çtãÛåã&ãÛܯð
íÁàdÞÀèç å ædÜiå=ýväèÀäýIÞý âöàßKÿiÛÀä{é4Û ÿ¥ådá?ÿiäÈãÛäÈè¼å
â,ådéãàßKàâ8ãÿ#à
àdâ8ãÛåã&ÿiÛÀä{é4ÛIýväèÀäývä Ü)ç=ãÛÜ
ãÜ)áã#Üß«ßàß]ò,çä ä:ã]ø¯ÿ#åá?äèéàèéìÞáäædÜ)óô
;èãÜß«Üáãä è ìê÷äè¼ãÛÀàáܱéådáÜá?ãÛÀܱð µíÁàdÞèç|éàèáäáãÜèãìê
å æÜiå=ìÈà-ÿ#ÜßÝÀß«ÜçÀäéãäàdèWâöàdß ãÛådè¼ãÛåã&ÿiÛÀä{é4ÛIýväèÀäývä Ü)ç=ãÛÜ¥ãÜáã&Üßß«àdßô $±ètãÛÜ]àã«ÛÀÜßKÛ'åèç8÷
ãÛÜIìÜå æÜ àdèÀÜ .àÞã¥í'àÞÀèç8÷ådì:ã«ÛÀàd Þ dÛãä dÛã«Üß÷8çà6Ü)á¯èÀàdã
áÜÜý ã«àoíÁÜIÝßÜ)çä{éãäæÜ÷'áäèéÜIä:ã±ÛådçUèà
ýväèÀäýIÞý âöàߥãÛÀÜ`æ-åìÞÀÜ)á¥àâ
ãÜáãÜ)ç8ô
0.7 0.7
Actual Risk : SV Bound : VC Bound
0.6 0.65
VC Bound : Actual Risk * 100
0.5 0.6
0.4 0.55
0.3 0.5
0.2 0.45
0.1 0.4
0 0.35
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Sigma Squared Sigma Squared
z w < Ó 6Ò)Ù,ÕÁÒ
VܱæÜß«êtíßäÜ 'ê¼çÀÜá«éß«äÈíÁÜiãÿ#àIàdârãÛÜ
áäývÝÀìÜáã÷ådèç|ývàáã#ÜJ8Üéã«äædÜ÷6ývÜã«ÛÀàçÀá¹âöàdß#äývÝÀßà-æ6äè
ãÛÜ
ÝÁÜßâöàdß«ý|åèéܱàdâî6ð
ñÃáô
Ú¥ÛÀܾæäßãÞådì]áÞÝÀÝ'àßãFædÜéãàßvývÜãÛà6ç ò.îé4 Û à ìÈõàdÝâ;÷ #ÞÀß dÜávådèç ðKåÝèÀäõ'÷=øùùÀû #ÞÀß Üávådèç
î é4Ûà ìõdàdÝÀâ;÷@øùù'dó4÷å-ãã«ÜývÝã4á]ã«àoäèéàdß«Ý'àß«åãÜ`õ6èÀà-ÿièUäè6æ ådßä{ådèéÜ)á¥àâKã«ÛÀÜWÝÀß«àdíìÈÜý òöâöàdß
ÜëÀåývÝìÈÜd÷
ãß4åè'áì{å-ã«äàdèïäÈè6æ-åß«ä{åèéÜtâöàdßNã«ÛÀÜoäý|ådÜvß«ÜéàdèÀäÈã«äÈàè¾ÝÀß«àdíìÈÜýoó
í6ê 'ß«áãWã«ß«ådäÈèäÈèTå áêáãÜýT÷ådèç
ãÛÜèoéßÜ)å-ãä è `èÜÿ çÀåã«å`í6êvçä{áã«àdßãä è
ãÛܯ߫ÜáÞÀìÈãäè `áÞÝÀÝ'àßã#ædÜ)éãàß«á¯ò{ãß4åè'áì{å-ã«äè
ãÛÀÜýT÷äè¼ãÛÜ
éåáÜWývÜèãäàdèÜçó÷Áåè', ç èåììê|ã«ß«ådäèÀäè våvèÀÜÿ áêáã«Üý àdèTã«ÛÀÜIçÀäáãàßã«ÜçÅò2åèçTã«ÛÀÜWÞèçä{áã«àdßãÜ)çó
çÀåã«åô Ú¥ÛÀÜTä{çÜå ä{átÜådáêVãàVäývÝÀìÜývÜèãIåèçDáÜÜý|áIã«àïÿ#àdß«õÅíÁÜãã«Üß¼ã«Ûåè àã«ÛÀÜß¼ývÜãÛÀàçÀátâöàß
äèéàßÝÁàdß4å-ã«ä è Näè6æ ådßä{ådèéÜ)áKÝÀß«àdÝÁàáÜçFáàtâ,åßô
Ú¥ÛÀÜvß«ÜçÀÞéÜ)ç áÜãNývÜãÛÀàçD*ò #ÞÀß dÜá÷&ø)ùdù Àû #ÞÀß dÜá=ådèçÅîé4
Û àd ìõàdÝâ;÷?øù'ù dó]ÿ¥ådá=äèãß«àçÞéÜ)çUã«à
ådççßÜ)áá]ãÛÀÜtáÝÁÜÜ)çUàâKáÞÀÝÀÝÁàdßã±æÜéã«àdß]ý|åé4ÛÀäèÀÜá]äÈèTã«Üáã±ÝÀÛ'ådáÜ÷åèçUådìáà|áã«åßã«á¯ÿiäÈãÛÃå¼ã«ß«ådäÈèÜç
î6ð
ñVô6Ú¥ÛÜ]ä{çÜå=ä{á?ã«àNß«ÜÝÀì{ådéÜ#ãÛÀÜ
áÞÀý ä/ è 01!'ôi@ò Bóí6êtåWáäýväì{åßáÞÀýT÷ÿiÛÀÜß«Ü]äèáãÜ)ådçtàâráÞÀÝÀÝÁàdßã
ædÜ)éã«àdß4á÷éàdývÝÀÞÀãÜçtæÜéã«àdß4á]òöÿiÛäé4ÛvådßÜièÀàdã&ÜìÜývÜèã«áàâ8ã«ÛÀÜ]ã«ß«ådäÈèäÈ è
áÜãó?åß«ÜiÞáÜç8÷åèçtäèáãÜ)ådç
àâ¯ãÛÀ Ü =÷¹å çä ÁÜßÜèã¼áÜãtàdâ¯ÿ#Üä dÛã«áWåß«ÜFéàývÝÀÞã«Üç8ôµÚ¥ÛÀÜè6ÞÀýIíÁÜßIàâ¯Ý'åß4åývÜã«Üß4á`ä{áIé4ÛàáÜè
íÁÜâöàdß«ÜÛ'åèçVã«à äædÜvã«ÛÀÜáÝÁÜÜçÀÞÀÝDçÀÜáäÈß«ÜçôVÚ¥ÛÀÜß«ÜáÞÀìÈãä è UædÜ)éã«àdßWä{átáã«äìÈì#å¾æÜéãàdßIäè U÷¹ådèç
ãÛÜWÝådß«ådývÜãÜß«á¥ådßÜ`âöàÞÀèçTí6êýväÈèäÈývä äÈ è `ãÛÀÜ 0KÞéìäçÀÜåèFèÀàdß«ý_àâãÛÀÜtçä ÁÜßÜèéÜ`íÁÜãÿ#ÜÜèUãÛÜ
àdß«ä däèådìÀædÜ)éãàß
YåèçoãÛÜ=ådÝÀÝÀß«à ëäÈý|åãäàdèIãà¼äÈãôKÚ¥ÛÀÜ`á«åývܱãÜ)é4ÛÀèÀ
ä !ÞÀÜ
éàdÞÀì{çoí'Ü=ÞáÜçoâöàd߯î6ð
ñ
ß«Ü dß«Üá«áäÈàè¾ã«"
à è'çµýtÞé4ÛÅývàdß«Ü|Ü |éäÈÜèãNâöÞÀèéãäàdèµß«ÜÝßÜ)áÜèã4å-ã«äÈàèá|ò,ÿiÛÀä{é4ÛµéàdÞìçVíÁÜÞáÜ)ç8÷Kâöàß
ÜëÀådývÝÀìÜ÷6äÈèTçÀåã«å¼éàývÝÀßÜ)ááäàdè'óô
¹àýIíÀäèÀäèNãÛÀÜ)áܱãÿ¹àtývÜã«ÛÀàçÀá+å æܱåIâ,ådéãàdß#àâ?ú StáÝÁÜÜ)çÞÀÝVòöÿiÛÀäìÜ]ã«ÛÀÜ=Üß«ßàß¹ß4å-ã«Ü±äèéßÜ)ådáÜç
âöß«àdýfødô S [ãàUødôÈø Ió¹àèoã«ÛÀÜ
îÚgçädäÈã4á
ò#ÞÀß Üá¥åè'çTîÀé4Ûàd ìõdàÝâ;÷8øùdùóô
oØrÒÕ
%'6Ö.E)*6Ò8Ó
ý æÜß«ê ß«åãÜâöÞìÀãà .ô =èÀäß4áé4Ûr÷ ¯ô àdÛÀì.÷ 0iô $
áÞÀèåÀ ÷ 0] ô iäÜãý|åèr÷ ±ôîé4Û àd ìõdàÝâ;÷ 6WôÀî6ä
è Üß÷ ô
î ývàìåÀ÷¯ô8îã«Üèådß«ç÷8ådèçTðNôÁðKåÝÀèÀäõÁ÷âöàß]ã«ÛÀÜä߯éàdývývÜèã4á¥àdèTãÛÜNý|åè6Þá«éß«äÝãôiÚ¥Ûåèõ6á¯åì{áàtã«à
6
ãÛܯ߫Üæ6äÜÿ#Üß4á÷ådèçtãàWãÛÀÜ0&çäÈã«àdß÷ `ôå ê6êåç8÷dâöàß&Üëã«ÜèáäædÜd÷ÞáÜâöÞÀì8éàývývÜèã«áôîÝ'Ü)éä{åìã«Ûåèõ6á
åß«ÜWçÞÜNãàFðNôðKåÝèÀäõ'÷8ÞÀèçÀÜß
ÿiÛÀàáÜNÝåãäÜèã ÞÀä{çÀåèéÜ
iìÜådßèÀÜ)çFãÛÀÜtßàÝ'Ü)áûÁãà ôî6ývàìå¼ådèç
±ô8îé4Û àd ìõdàÝâ;÷Àâöàdß±ý|åè6êoäèãÜßÜ)áã«äè våèçâöß«ÞÀäÈãâöÞÀìçä{á«éÞá«áäàèáûådèçãà &ôî6Ûå ÿ#Ü ;Ú@å ê6ìàdߥåè'" ç -Nô
îé4Û6ÞÀÞÀß«ý|åè'á÷6âöàdßiæ-ådìÈÞ'åíÀìܱçÀäá«éÞ'ááäàdèá&àdèUáã«ßÞ'éãÞß«ådìÁß«ä{áõvýväèÀäývä åãäàdèrô
- 16ÒÖ@Ù <
Uz Ð z
ÕE(]q
u=Ô ÕÕ+( v "Õ'Ô$2)%
q "6Õ$Ô 2) "
â
ãÛÀÜUçÀåã«åïä{á ;çäývÜè'áäàdè'åì±ò,ä.ô Üdô
H m ó4÷#ãÛÀÜÃçäývÜèáäàèµàâ±ãÛÀÜUýväèÀä
<
ÝÁÜß«ýIÞã4å-ã«äÈàèá¹àâãÛÀÜ`à í Üéã«á÷ÀÿiÛÀä{é4Ûoä{á 7 # ô
Üë6ã?ÿ¹Ü¥ýtÞáãáÛà-ÿÅãÛåããÛÀÜiáÜã?àâæÜéãàdß4á@ÿiä:ã«ÛtéàdývÝÁàdèÀÜèã«á
ò<@óáÝåèWãÛÀÜiáÝådéÜ Uô
Ú¥ÛÀä{á¹âöàdììà-ÿ]áKâöß«àdý ãÛÀÜ=â,åéã¥ãÛåãiãÛÀÜNéàdývÝÁàdèÀÜèã«á#àâ
ò<@ó#åß«Ü=ìäÈèÜåß«ìêWäèçÜÝ'ÜèçÜèãiâöÞèéã«äÈàèáô
Ààß`áÞÀÝÝ'àáÜIäÈè'áã«ÜådçUãÛåã
ã«ÛÀÜtäÈý|ådÜNàdâ åéãäèFàdè < ? ä{á±åáÞÀíáÝådéÜWàâ UôIÚ¥ÛÀÜèÃãÛÜß«Ü
? @I ò ô$úó
*)
=
#
±áäètãÛÜ`ìådí'ÜìäÈèNäèã«ßàçÞ'éÜçFåíÁà-ædÜd÷éàdè'áä{çÜß#åtÝådßã«äéÞÀì{åߥéàývÝ'àèÀÜèã¥àâ
ò ô üó
H SI
D:9;9;9D Hr÷ÜæÜß«êéàdývÝÁàdèÀÜèã±àâ
áÞé4ÛãÛ'å-ã =*) #
î6äèéÜNã«ÛÀä{á]ä{áiã«ßÞÀÜIâöàdß
åìì@é4Ûàdä{éÜ)áiàâ #
=1H
ýtÞáãiæ-åèäáÛô,]ÜèéÜ=ã«ÛÀÜ`äý|å dܱàâ åéãäètàdè!<5? áÝådèá Uô
Uz z
!# q#Õ+% Ô$#ÀÒ8Óok %'# )Ù PÔ #ÀÒÖ(ok }tÕ'×Ò@Ö
Ú¥ÛÀÜ¥âöàìÈìà-ÿiäè]ÝÁàdäèã?ä{áéÜèãß4åì6ã«à=ãÛÀÜ]ådß ÞÀývÜèãô $±èÀÜ¥èÀàdß«ý|åììê±ãÛÀäèÀõáàâå=éàìÈìÜ)éãäàè`àâ8Ý'àäÈèã4á
ådáíÁÜä
è MáÛå-ããÜß«ÜN
ç ]í6êWå
áÜãàdââöÞÀèéãäàdè'á÷äÈâÀâöàßådè6ê=é4Ûàdä{éܹàâ'ì{ådí'ÜìárâöàdßãÛÜ#ÝÁàdäèã«á÷å±âöÞÀèéã«äàdè
âöß«àdýgãÛÀÜ¥áÜãéåèNíÁܹâöàdÞÀèçWÿiÛÀä{é4ÛNådá«áäèá8ãÛàáÜ&ì{åíÁÜì{áã«à]ã«ÛÀܹÝÁàdäèã«áô@Ú¥ÛÀÜ#ð ïçäývÜè'áäàdè
àâãÛåã
áÜã¥àârâöÞÀè'éãäàèá&ä{á&ãÛÀÜèFçÜ èÀÜ)çFåáKã«ÛÀܱý|åëäÈýtÞÀý[èÞýIíÁÜß#àârÝÁàdäèã«á&ãÛåã¥éåèFí'Ü=áà¼áÛå-ããÜß«Üçô
]à-ÿ#ÜæÜß÷réàdè'áä{çÜß
åáìä dÛã«ìÈêTçä 8Üß«Üèã=çÀÜ èÀäÈã«äÈàèô ÜãNåFáÜã`àdâ¹ÝÁàdäèã«á±í'Ü|áÛå-ããÜßÜ)çÃí6ê¾åáÜã
àâ?âöÞèéã«äÈàèáiäÈâ?âöàdß
åè6êFé4Ûàdä{éÜ`àdâKìådí'Üì{á#âöàdß]ã«ÛÀÜWÝÁàdäèã«á÷Áå|âöÞÀèéãäàdèTâöß«àdý ãÛÀÜtáÜã
éåèUíÁÜNâöàdÞèç
ÿiÛÀä{é4ÛTådá«áä dè'áKã«ÛÀ: Ü * I
M ì{åíÁÜì{á¹ãàvådìÈì8ã«ÛÀÜ`ÝÁàdäèã«áô ådäÈèìÜã¥ãÛÜ`ð çäývÜèáäàèvàâãÛåã¯áÜã
àââöÞÀèéãäàdè'á¹íÁÜNçÜ èÀÜçTåá#ãÛÀÜ`ý|åë6äýtÞÀý è6ÞÀýIíÁÜߥàdâÝÁàdäèã«á¹ãÛ'å-ã¯éådèFíÁÜNáà¼áÛåããÜßÜ)ç8ô
ãoä{átäèOâ,åéãvã«ÛÀä{á|áÜ)éàèç çÀÜ èÀäÈã«äÈàè ò,ÿiÛÀä{é4Û ÿ#ÜÃådçàÝã¼âößàý ÛÀÜßÜUàè'óIãÛåã|ÜèãÜß«átã«ÛÀÜUð
íÁàdÞÀèçµÝÀßà6àdâ,á|òöðKådÝÀèÀäõÁ÷#ø)ù'-ùÀû -¯Üæß«à-êdÜd÷ % ê àd ß µåèç Þdàáä.÷#øùdùóô"$¯â]éàÞÀß4áܼâöàßNâöÞÀèéãäàdè'á
ÿiÛÀàáÜ]ß4å è dÜiä{á - Iø .tòöä.ô Üô?åìì'çå-ã«åNÿiäÈììÀíÁܱåááädèÀÜ)çtÜäÈãÛÀÜßKÝÁàáä:ã«äædÜiàdß&èÀÜå-ãäæÜ]éì{ådá«á«ó÷dã«ÛÀÜ]ãÿ#à
çÜ èÀäÈãäàdè'á¹ådßÜ
ãÛÜ=á«åývÜdô ¯à-ÿ¹ÜædÜß÷äÈâ@åìì8ÝÁàdäèã«á¹â,åììä è NäèáàdývÜ±ß Ü däàdèFåß«Ü=áäývÝÀìêtçÀÜÜývÜ)çvã«à
íÁ0 Ü 4Üß«ß«àdß4á À÷à ß éàdß«ßÜ)éKã À÷ã«ÛÀÜ`ãÿ#àvçÜ èä:ã«äàdèá¥åß«Ü`çä ÁÜßÜèãô á]å¼éàdè'éß«ÜãÜ`ÜëÀåývÝÀìÜ÷áÞÀÝÀÝÁàáÜ
ÿ#Ü
çÜ èÀÜ åÝ|äèãàìÈÜß«ådèã&éì{ådá«áä Üß«Ká ÷ÿiÛÀä{é4Ûoå߫ܯìäõdÜ åÝvã«àdìÜß4åèã¹éì{ådá«áä Üß«á÷díÀÞãiÿiÛÀä{é4Û|ì{ådí'Üì
åìì6ÝÁàdäèã«áìêä è ¯äÈèIãÛÜ#ý|ådß äè`àdßàdÞã4áä{çܹãÛÜ¥áÝÀÛÀÜßÜiå6 á M I(G«ô ¹àèáä{çÜßå ådäÈèNãÛÀÜiáä:ã«Þå-ã«äàdèNäè
ä dÞÀß«ÜNø ÷íÀÞã]ådá«áä ètÝÁàáäÈãäædܱéì{ådá«á?ã«àtådìÈì'ãÛßÜܱÝÁàdäèã«áôÚ¥ÛÜèFå ådÝ|äèãàdìÜß«ådèã&éì{ådá«áä ÜßKÿiäÈãÛ
ý|åß däèÃÿiä{ç6ã«Û dß«Üå-ã«Üß
ãÛ'åèïãÛÀÜoíåììKçä{åývÜãÜß
éåèÀèÀàdãWáÛ'å-ãã«Üß=ã«ÛÀÜ|ÝÁàdäèã«á=äÈâ¹ÿ#ÜvÞáܼãÛÀÜ ß4áã
çÜ èÀäÈãäàdèFà â áÛå-ããÜKß À÷6íÞã¯éådèáÛå-ããÜߥãÛÀÜ`ÝÁàdäèã«á¥äÈâ@ÿ#Ü=Þáܱã«ÛÀÜNáÜéàèçïò,éàßß«Üéã4ó#çÜ èÀäÈãäàdèrô
äÈãÛ¼ã«ÛÀä{á&éå ædÜåã?äÈè¼ýväèç8÷ÿ¹Üièà-ÿDàÞãìäèÀÜ¥ÛÀà-ÿ ãÛÀܱð Oí'àÞÀèçÀá¹éåè|ådÝÀÝÀìê`ã«à`âöÞÀèéãäàdèá&ÿiäÈãÛ
ß4åèÜ- Iø DKS .÷rÿiÛÀÜßܼã«ÛÀÜvì{åíÁÜì SývÜådèá
ãÛ'å-ã`ãÛÜvÝ'àäÈèã`ä{á=ì{åíÁÜìÜç;éàßß«Üéãô Yò2Ú¥ÛÀÜví'àÞÀèçÀá
ÿiäììåì{áà¼ådÝÀÝÀìê|ã«à|âöÞÀè'éãäàèá]ÿiÛÀÜßÜ Soä{á]çÜ 'èÀÜçTã«à|ývÜådè 4Üß«ß«àdß ÷íÀÞÀã¯ãÛÜIéàßß«ÜáÝÁàdèçäètð
çäývÜè'áäàdèIÿiäìÈìí'Ü]Ûä ÛÀÜß÷dÿ#ÜådõdÜèäÈè±ã«ÛÀܯíÁàdÞèç8÷åèç¼äè¼àdÞÀß#éåáÜd÷ý|åõ6äè
äÈã&ÞáÜìÜá«á«óô>Vܯÿiäìì
âöàdììà-ÿ ãÛÜ`èÀàã4å-ãäàèoàâ¥*ò -±Üæ6ß«à-êdÜ÷ % ê à ß Uåè! ç rÞ dàáä.÷øùd
ù óô
¹àèáä{çÜßiÝ'àäèã4á <5?QA ÷ÁådèçìÜã$@ò'<óiçÜèÀàã«ÜNåvçÜèáäÈãê|àdè A ô rÜãD=ÃíÁÜWåtâöÞÀè'éãäàèàè A
ÿiäÈãÛÅß4åèÜ - Iø D S .6÷åè'çVìÈÜã í'ÜåÃáÜãWàâ]áÞé4ÛVâöÞÀèéãäàdè'áô rÜãIÜåé4Û5< Ûå ædÜ|ådèVåááàéä{å-ã«Üç
ì{åíÁÜì L ?3- IøC.ô9rÜã-7< # D:9;9:9JD < B .WíÁÜNåè6ê èä:ã«Ü=èÞýIíÁÜßiàdâ@ÝÁàdäèã«á¥äè!A ?ãÛÀÜèTÿ¹Ü=ß«Ü !ÞÀäßÜ
ãàvÛ'å ædÜ
ãÛÀÜ`ÝßàÝ'Üßãêvã«Ûå-ãiã«ÛÀÜß«Ü`Üëä{áã«á¥å-ã]ìÜåáãiàèÀÜ = ? áÞ'é4Ûoã«Ûå-ãD=ò<"=2ó ?3- Iø .@)N=ô Ààß
däæÜL
è =÷'çÀÜ èÀÜ=ã«ÛÀÜNáÜãiàâÝ'àäèã4á íê
:- ),
L H
H ø D =@ò,)'ó1H `ø .-E),
LH2`ø D = ò*)'ó H øC. ò ô ùó
= íÁÜNáÞé4Ûoã«Ûå-ã]ådìÈìáÜã4á
VÜ`ßÜ !ÞÀäßܯãÛåãiãÛÀÜ U
åܱ߫ývÜ)ådáÞÀß«ådíÀìÜôÜã çÜèàãÜ
ã«ÛÀÜNáÜãiàâådìì
ô
! ÒÙöÓ)Ù,ÕÁÒ 0Üã<>= DKF H ø
D;9:9;9JDKO í'Ü!O ÝÁàdäèã«áô5ïÜUçÜèÀÜFãÛÀÜÜývÝÀäß«ä{éåì&ßä{áõ¾âöàßIã«ÛÀÜTáÜã
-7<,= D =+.NãàvíÁÜ
! = ã«àví'Ü
ÒÙöÓ)Ù,ÕÁÒ PVÜNçÜ èÜ=ãÛÀÜNåéãÞ'åìß«äáõ¼âöàdߥã«ÛÀÜ=âöÞÀèéãäàdè T
ò 8ó ,= H
H vò<5? ó I ò ôødø ó
! ÒÙöÓ)Ù,ÕÁÒ Ààß ÀëÜçVò<
# ;9:9;9D <
D
B
ó ? A ÷ìÜã í'Ü=ã«ÛÀÜ`è6ÞÀýtí'Üß#àdâçÀä 8Üß«ÜèãiáÜã«áiäè
ÿiÛÀÜßÜ
ã«ÛÀÜNáÜã«á ä{áiçÜèÀÜ)ç
ådßÜ=çÜèÀÜçTådí'à-æÜôKÚ¥ÛÀÜ`è.ãÛTáÛ'å-ãã«Üß]éàÜ|éäÜèãiàâ [
5 M2 2 ICK M" I(# M #
I 7 $+ I G
M
I( MJK *"CM B ò -7<,= D =+.ó
= ó 7ò D Oó M M2 I"CM 7+*" M O 5 I(* G
ò< DIII DK< ó ?!A 6*M M I M * G $
GJM I G $ * 6*6 = ? G ( G 2#+=ò<=.ó ?3- IøC.8 <,# = B G8 M G * I5K #
M " M/M2 G
ò Á
5 I * M * *,+ 6 + I( $ K M I #NG $ =
vò B ò 7- <,= D =+.ó ,= .4ó1 ü ò D Oó6ÜëÝ
7
ò 8ó B 8 ò ôøJB6ó
UÕ'Ó
1.0 *,1K*.;1
*2" ·3 12 043I *° 1 " 3 " 1{§ $0 *.* 312+-," 0 *.K*21241. 3 043 0
3 0 3 Â 3 3 0 5
5 @ Ä " 12+- !¥ ,21.*218*.;1 .&²d,+-4²-*Á8*2+ $]4$* -*. ,212.4" *.;1 & .&( )1812+ + ( (§" *212* 1Á12+-2
2*2° 1 3 3 0 0 3 0 0
 @-5 *2@1,+-@12.! , ,"$ 12±+§²d2² /.¹1 !#² + *2"$·1,+-4181,+-!¥41,+-!¥41,"$4 ( ;1r *2"$ 2
*
5i"$¸ *À12+ ²-4"
0 r+-3 .
O" *1,+-*.;13 9² 0 " 1.*'+)"$.+¥ " " 12+-r+§²d2² @ 0 "$*'²-043 21,"$; $
0 0 3
.+ " ?9 12+-? " 1 .!¥4 +-*
.?" d2 1 2"$ 1.=+§²d2²)$ * 3 3 3
0 0 3 3 0 5< 3 3 0 3 3 5
º-2+=¥*21 9 ² " 1.* r+ "$2+`*2²- )"$!¥ *," 4À*, (-*2² 9À& " *,²- 2?*.4"$
%
5 1 ( ,2" 0 24-² 0 3*2" 1," . ! 3 3 3+-r 043 ;½¥+ 9d?0 *21 93 O² " 12*Á" .4
"(
² 0 *2" 1," 3 3 * 0 043 )"$!¥ 0 *," 0 0 40*2°"$!#²)$;½ 512+ < 2041,"$3° * 9r+ "$0 .+¯.r1,0 +-² " 0 1.3 *81,+-3 !¥*.;3 *
0 +-)043 ," 41, " 3 93 1,+-( ¯*.3 *,-!¥ 043 *Á1,+-4112+ !#² "$° ,"$462" *2¶¥ 0 .* " 9 .!# 0 3 §i1 12+ ;12° -4 5
(
5]<2" *2¶K*À12+-° -04!K 3
(d0 9-12.40 " 3 " (-*. 41," *Á" 2*.* ³Á4² " ¶ 043° 3 0 0
3 0 3 3 0 ° 4
0 3 3 3 "( " 5 3 *2*.2§& 3 ¹*2 ¥" 3 1
)" 1," 9 1,+ "$*?"$*?1,+-41? "$ ! È r+-2 "$*?1,+- -!K( 9812.4" " =*2!#² $*K
043 : " *041,3 +-0 ³´F 3 12 0 ²§ 0 91,+-?*21 0 9À ;"$*," 043 9: 3 ;12" 043 * ³Á4² 3 " 3 ¶ "( "'& 0 ³'4² 3 " 3 ¶ 3 "$"$* 5 ¦ 0 r 3 3 §
*.;1 9-9 ;12" *'r" 1,+¥" " 12³´I)" !¥ *2" 1,+-8³´t 12 ²§K"$* ! ++- r9 12+ *2r *.*2" .*
1,+-?0 .¿ 3 "$.
043 " 9 .!O3 Â)3 . # 3 * 0431r+ $ ± 3 * 0 ;" 12+ 0 *r12+ 3 ( 0 Â
3 0 043° 3 0 3-0 0 3 0&3 0 0 3 5
# +-.&"$*? " # !#122" ¥" 122²-2;1.41," 9 12+ ¥)-4²- ()$! +" 1?"$*( *2"$4 § "
1,+-#1{
5]<; *.*,1@² " 3 1.* 9À 0 ;½
+ 3 * 91,+-?1{ 043 *.;0 12* º 0 Â 3 3 0
0 ? 0 )3 ? 0 12+-0431 ° .¿-?1 (d0 0 5 / 3 3 1,1 3 / . 3 *,1.;" 3 "$"$# 5
"
!
5 @3 3 Â)3 0 0
"$# &% % % # ( ' *) #
% % % # # +' , #
r+-.].;²d41.¼" )"$*i.]*2-!#!¥
1,+-¯2"$4+1#+ v*2"$) tr+-2 ) "$*¹1,+-]1 124 §
5 %
& 5 É 3 1,+- 0 ,"$4" 3 49 0 .!K $41," 043 ³Á4² 3 " ¶ "( " 1,+-;§±24 $ ,.;½12.!¥ ° ;1 0 2* 5 .
5ö/ § ,. ;"$*2" 043 9: 3 ;12" 043 .=¥!¥ 3 ¯9: 3 ;12" 043;: =< r+ 0 *2i*2"$ 3 .;²-2*2 3 12*?12+-i;$*2*¹*2*2"$ 3 W1 0
412#² " 1 <
0 3 5
! 5ö/ § ,2" 3 122" 3 *2"$)"$!¥ 3 *," 043 .K@!# 3 1,+- 3 !&(d 0 9d²-.!¥;12.*Á2¿ " .i1 0 *,²;" 9§i² 0 " 3 1 043 12+
!¥ " 9 $
3 0 5
« 5 @ 1. 3 412" ° ; § 043 ¹ 3 24-&1,+-41 4" ° 3 12+-K9 0 .! 0 91,+-K* 0 1," 0436 12+ K² 0 *2*2" ( Í !K-*21 "$K" 3
*2)(-*2²- 9)"$!# *2"
0 3 043 5
5{¸W0 2¶]" 3 ²-.;²-2412" 0435
* 5ö< +- 3 ¶*1 0 5 º! 0 ¥9 0 ² 0 " 3 12" 3 ]1,+ "$* 0 1 5
% 5,7 3 §¯1,+- 3 ¶*r1 0i043 0 91,+-K. ° "$;.*r9 0 ² 0 " 3 12" 3 i12+ " * 0 1 5
( 5ö< +-] 0 .]¿ .41,"$ 0 ² 1,"$!#" ·¥"$*¹4( 0 1 (> " 3 * 0 9´
?@? 5N< +-]+ "$4+-#$ ° ; 0 1 0 +- 3 $
.+ " 9À 1² )-;1.* .+ ¶" É ;12 "$*¿)" 1.K !#² ½± = *,"$ .4() §
$2
3 0 0 0 3 3 ¹ 0 3 043 5
# 5ö< +- 3 ¶*1 0 Î 5 ?4 9:!¥ 3 9 0 ²- 0° "$" 3 ]!¥r" 1,+=1,+-*.K.*2) 1.* 5
" 5öÇ 4 d1,+-4181,+- ,.;" " 3 $ .¥*2"$ B ADC 4!# * ,.*.!¥4 $*21" 1.r.41281,+- 8¿-41 .N@ * 12+-2
"$*¹1{§² " 1,+-K;12-49 .!K3 $¹ " 3 " ³'4² " ¶ 3 "$"$* r+ " .+`ÉÁ+- 3¯ ?0 ..;12±0 +-5 . 0
0 3 0 ° 3 3 3 ° 0 5
!> > 12 9 @½!#² 1,+-411,+-¹"$*212 #(d;1{ 2§=² 4"$ 9 ,12" * 912+ #*2§!¥!¥;122"$&*2" !#² $;½
5 "$*Á0 1,+-*20!¥$ +À*.F E'¿ !$% 3 3 ° 0 ° 0
5 H5 G@0 ° ¹,"$ 0 0 *8²- 00 9"$* 3 3 ]*89È'*ɶ 3-0 "$*8$2¶" 3 5
! +- ¶*1 º+ § 9 ² " 12" i12+ " * 1
5ö< 3 0 I-5
6 »{< 0 0 0 3 3 0 5
!$! ³ ³Á4² " ¶ %Á," 412¹´ !#!& "$412"
5 5 3 ° 0 3 0435
! +-.?"$*@ 4 1. 412" #( &!#"$4+1-*2 !¥; §=1,+-41 2.*2² )" ¯1 12+ ¹*21 91 124 §
5ö(< 3 412" ¹9: 1," * EÁ¿-41," È -3 !$# " ³'4² " ¶ 0 "$"$*04,3 3 80 9 @ 0 *.*0 9
3 ° 0 3 4
0 3
1," 0 3 *124¶" 3-04±3 1,»È+-3 4 -° ¹· 3 04 3 `" 049'3 1,+-¹5 !¹² "$23 " 4'2" *23 ¶¯"$*· 12 5 +)"$G*0 ( ° ="$*0 0 *21,+-3 »
043 3 ° 0¯0 043 R 3 - .
R Y 0 0 3 00 3
1,+-41@" EÁ¿ r+- K JMLDN OPQL A JTSVU SW$N OPQLDX S Z % r+ "$.+=" *1,+-?*.?+-.
3 5 3 ° 5
!4 ³ $ · %'2" 41.¹´ !¥!K "$41,"
5 5)/ 3 ° 0 3 043
m-2( Ô$ÒØ
t
E
7& !¥5 ;1,5 + ±" ²-4361,12 5 7t. 5À / ° " 1," $36
"$·.!¥ . 2!¥ FÎ É
3" 5
·
5 Ç0 043 5Y< 0
+- .;12" 49 412" * 912+
² 1. 12"$49: 1,"
0 3 043 + 0 0 3 3 40 3
0 3 3 0 3 40 3 3 3 \5 [ ) {>]#ÈÈIÆ^r
.]¥{
!_-, !$* #$! )#( "$%4 5
1,+ §v v> "$* %À¯$ " W ;-24 ;1{ 2¶* É
?Æ2ti,V)
IÆ
7t5 3 043
K
), K
'3 4 ²-56/ * %$5 "4«)%$"( "$"$3 * 3 5 3 3 3 0 5 3
3 412+-3 !¥41,5 "$40 *2* " 412" 3 9À@3 !¥3 2"$5 3
>]¥
ÈF=
% ;121= E 2 *212;" Ä !¥;1.,§T" $ " É ²-W1 4² ²d
5 5*2/ + " 3 43 1 Ï 3 ´ 5"$/ "$# 0
´ ¸ 3 4
0
3 5
" *2+ ² ?
2?
';r{
;K^
.-È ´Á$2 $
5 Á
5 7 0 4
0 N
3 0 5 %Á2*.* ½9 . "$"$*
³ 5 7t5)$ / · 0 5 º.+ ¶ ² 9 ) 12+ ´ -2* ³ ³'45 ² " ¶ 3 043 ³'1,12+ 0 ´ !#²-,"$5* 9 "$; (-*.
5( / ;3 1¹28/ 5 " 12" 0 0 4
G2¹ " 1,+-5À!¥ / *&
-*20" `.45 / "$*,12"$ t! 5 ;$* 3 É ´ 3 <?5 54$*2( 0 . 043F0 º° ;$ ´
³0 2(- - 0 3 043 º 0 )+ )" 31 .* 0 5 3 5°043 7 ¸¾56°043 363I-5 5
0 6
3 3
/ 5 3 0 0 5[ ;È ÁÈ ?
), K
' 4 >_ [ ²-* !$*K !$*$% -/ , " 36
"$"$% º²-2" @Î6;1,-.K> 12*" ´ !#² 12º;"$ ³ !
/5 56/0
E 5 *2 3 É 567t5 Ä@ § 043 0 3 W³ 3 5 ³'0 4² 3 " ¶ 5 µ12.3 4" 3 " 3 `0 45 $ 0 ," 1,+-5 !g9 0 0 ² 12" !¥4!¥24" 3 ;$*2*2" Â 2* 5 É 3 ;Á :
-- = 2) v; _Q ]){:6
;-I
. %'" 1,1.*,( -24+ "$"$! 5 ´ 7t5
[ !¥* .+W `Î" ]?4 9:!¥ º !¥&*2124( $¹!¥;12+ *9 @4$; 412" =" 21,"$¯ `* " ¯*2§!¥!¥;122"$
I " @ÇK*25-§/*212!¥ 3 * 3
. 3
] Q
] 365 0
0 0 3 3 3 0 ° 3
!¥3 * .+N 5 =Î" )±?4 9:!¥ { :!¹ ² 124 ö12«" ( 4+ 8 %!¥ ;1, + ( " = 9 "($r( 125 +-K" ) " 12¥¿ .41,"$¹² .!#!#"
¥
:
K
& .
)
I ²- ( $ÇK! 5/ 3
3
«3 ,±Æ¯: 36!5 :0 È- 4 + 4043 ( "$0 #> 0 3 Â)3 0 3
´ 0 ´ 5 .*@ [ [
º2+ ¶ ² 9 É{!#² " ¥1,+-K;-.;§± =*2²d 9À*2)² ² 21 ;1 @$ " ]!¥.+)" * 5
É M5 I-5 5 / · 3 /. 5 0 0 5 %À12*.2+-0° 3 )" 1 .* Æ#"4
K3 K
),0 $;%;]¥ 0 :° &0 ,
3 3 ('${
.]#3 5
3²-` t
7 )
5
* ( *$ # ´!K(-,"$ =
7 0 -7 t 5 I
0
3 3 K
<
"$"( 5 É %'.*.* 0 [
-2* º"$!#² " =*2 d7²)² ,1 ;1 57 <;"$*2" 2 5 * É Î . ·¥º4" 1,1. )" 1
/ ;-
;5 :Â _%
,
0
¥° 0
&043
;-: 5 ²-3 0 * (3 '($( 2" É1.4 0 § r"$,"$%
Æ.K ?
¯: )
´ ´
5{I-
5 -:5* ?4 9È!#
´ ´ -.* " $)" ¯ 4 §`" 2"$ 1¶ ;$* É r,
Æ# K-/:)=
< (+',=5Á7= 02) v3 - 'd !)36;5
5 .dI-
.5 {5 / 5¥
/
0 /ö±3 !)0
.2 1 "$3"$° # 5 3
3 5 3
´ ´ -.* % " .*.2+ 241.*2.+ º)² ² 21 ;1 ;(=²-$+8+121,² + 3#3«* ! 2*...+ (; $4(-* !
5 I-5 .+ 5 / " 4À2;² ,51 Î3 - 1 3 .+ ÇK+5 G 4" * 5 0 ° 0 ° 5 5 » 5 0 5
"$"$%
´ < ´ 23 12*@ ±³ 0 ³Á4² " ¶ 3 º< ² ² 3-0 21 0 ;1 ;1{5 2¶* !> + !(4)
5 ´ 0 -. 13 =Ï 5 " 3 (,5 1 0:)Æ° & 0 3
.]¥ 0 :
;
5 84: É 12.*.;"$
:
&
;
-
!$"( "$"$*
5
ÇK
Î-5 ¹Ï 0 3 § 3 6 Î 5 *.·;5 7 G5 Ä@§ 0 5
W Ä 45 ( ?Î- *2"
-
r ,
.
5 :I3 )
=3 8 "$*
;5 ^
.-: º²-,"
³',$ ° 0 ² ² "$41," 0 * 9 0 Â641,+-3 !¥41,"$*0 ³ 0 ; 5 "$[ "$% 5 3
4
0 3 0 7 0 5 5
G¹5 Ï#Æ "4,-.¶
? :& ´ ´
5 I-5 5d/
-.* Î ?4)9È!¥ º! $ W³ ³'4² " ¶ º ² ² ,1 ;1 ¹22*.*2" !¥.+)" *
5 ,3
5 (
'$0 {.
]# 3 " + *$5 * % 3 "$5 "( 5
0 ° 0 043 3 5
[ ¦$;1.2+- r?,
),:$ %
: )]#ÆÈKÈ 9 (:; ]# <: 5 I0 3=¸ + " $
;
§ `
º * É ! ±
" 2
1 " "$#(
ÇK º1,5 -21Ä!¥5 !E' " " *21 .¶ >@;-.4 1{ 2¶* ¯12+-?3 ( " *8043 3 23 "$ 5$ K3 )" !#!¥04 36 ?
25 _>])!)
: + ) *$# "$"$! 3 3 / 3 3 0 5 3 0 3 ° 3 5
¦ Ä@"$ *2" ¿ " 45 (d1{8 *,²-.*284² ²- ½"$!¥41," #*, ² ² 21 ;1 '!¥.+ " * K
),
_'> ]d{:
5 /ö{¯0 !)5
2173 =@_ _6° 3
1]# < 3 [ "$"$#
0
5
043 3 0 ° 0 3 5
É Ä@)§ ³ ³Á4² " ¶ *. Î 1,1 Tº º $ º122-;1,-.42"$*,¶|!#" "$!¹"$·412" 9 i.+-212
5 . 0436" 1," 5 #Æ "43 /
? :5'& /?0
),$ 58% / 0 ]#0 ÈÈ& 3 r,.5
5 0 ( 5 {.
]# + 3 043 0
0 34$! 0436* 5 [
$
'
( ( " "$"$!
5
%
5 Ç&5 G 0 5 [ =: ´Á.
+-;>2'?$*
@r,+ *
1] @ È5 ;B Ï A 5 ³Á 3 > :0 *,1.2´8 3 !&`(-´ ,"$0 !# ²- ¤ 3 § " É 3 . *2"5$1{§= "$%Á%2( 5 *.* "$#$*
Ç0 .+ 5 " !¥ G* 0 3 ;3 ½1&412 2"$Ç& ·45 1,"I0 3 "041236+t50*2 ² ² ,1 [ ;1 ¹!¥5 2+ " * 2+ 3 "$° 4.;² 21 ÎÀº=³É,É2É? 5 >-!K(
!
<?¤ $ 5 I0 " .*2" 1{§ 5Å9< Ï ,1.!K 0 "$"( 043 9:1,² + 3#3912² 4" " 0 9 .° !#4120 " ¶ " 3 212!K5V < %3 3² )($3 ;²0 212*3«2;² ,1 ! ² *
Î ?4 9:!¥ 3 ° 0 0 3
º " 1,+-8¿²¹² ( $!Ã9 *2 ² ² ,1 1 1.24" " É ,
.
Æ:@ 8:)
< +'D= 42) 5 » 5 3 0 5 3 » 0 3 5 Ç 0 0 5 5 C
À 5
5 & 'd !)36;0 5 .d0
.° {3
0 0 /ö± !)0
.2 1 "$0 "$# ° 0 3 3 5 3
> ! `º ³ ¦ !#" -,Æ {5 ^
%Á2 1,"$ 4 É "(
¹5
5 Î 5 0 0.0 *,"$0 ° 3 K-5 :5
0
5 7 3 6
3 5 r36,.5 ,> ]\]# 5¹7 Ä. [ G " >;5 F E 3 2¶ » "$G %$"
0 5
3 5$ 5
Ä21,+N% ´ .!¹"$.¶ KW
6 ,2> ] ]#4 G=)
« [ ;: ]¹#Æ [ ? È:- 5 I0 + 3`¸ " $;§N 3
º * É 57 "$0 # 5
Ï ´ 043 3 125$ !¥2§¯ 5 E %À.¶ ;-È2Æ ÈW{#:
.F
^
ö,
: [ + " $;§
=º * É
5 5)7=043 0 3 5 5 5 5 I0 3±¸ 3 043 3 5$
! ±" 12" "$"$!
3 4 5 4
0 6
3 2"$4+1 :;dÈ=]¹B<: ):Æ
ºÉ{ 5
5 3 ±5 Ä.. 0¹<0 24$
0 5 @3 5 1,+-7t
7=0 3 ¸ 4 @ "$"
* 1," 5 96 .¿- 2412"$² .!#!#" K²- ()$!¥*" 12+¯(
I0 . *212I.5)47= " 12* 'H 0
>:; ]# <: ö4 + " « 043= 0 0 0 3 0 0 3
º 043 ¶+-3 5 E [JI
"$"
*2 `¦ Ä@"$ *2" > " ²-.)" 5 1," 9Á2+- 12" ¹1,"$!¥&*.2"$* *2" ]i*, ² ² 21 ;1
5d!¥7 .+)" É 5 r,.
3
Æ:3 N `5 :)
0K>K+5 KL04= 3 3 , M K04
3I ,0 K
N 0
Á 8 'd O 3 ,
0 ° ²-0 *
3 5 3
*) * " !# "$iÉ{*2$ 3 ¦ÀÎ "$"( 5
$ º! $ Ä 41.*2.+ º.+ ¶ ² 9 + $! . N³ ³'4² " ¶ %'.)"$;1," =1,"$!¥¥*2,"$*
r5 »{ÇK I-{5
;0 :0 _36% ;3
,
5
] 3 [ ;È5 ÁÈ;
57 *, ² ² 21 5 ;1 & 0 !#.+ 5" Ç * É / 25
.Æ0 0 ,
" 1,+t
0 ° 0 3 5 3 K
3 ,; K
' 4
²- "$"$" º²-," Î6;1,-.K> 12*" ´ !#² 12º" "$"(
EÁ) *2 5 3 ÁÇ3 0 (d,1i¦-.; 3 0 3 o¦-3 0 ,"$ 0 Ä@"$ 0 *2" 5 3 3 "$!#²- 5 0° v1.24" 3 " 3 t4$ 0 ," 12+ ! 9 0 #*, ² ² 0 21 ° ;1 0
!¥.+)" * É r,.
Æ:] ¥ B
< (PK+K>KQ= 2) ?
2R K
N
Á S ;@ 'd$T r,
.:« KrÆ >
>I Tr;3 5
3 7 >
,> > È)+ K >
; ² * !( %& !$#$* @!¥; "$iÉ{*,$ 3 ¦ÀÎ "$"( 5
EÁ) *, 3 Ç0 (d21K¦-.; 3 3 W¦-),"$ 0 Ä@"$ 0 *," 5¾< .4" 3 " 3 =*, ² ² 0 21 ° ;1 0 &!¥.+ " 3 *+& 3 4²)² "$41," 043 1 0
04365 3 PK+K>K _%
,
K _> ]){
( .::`6 Æ
; ^
.-: ²-* «> ]«$ % "$"( 5
9: ;12;1," É
EÁ) *2 W o¦- 2"$ Ä@" *2" )-;" I12+-
2 12"$!#
!#² $;½" 1{§ 9?*2 ²)² ,1 ;1 ]!¥.+)" * É
-{
;:3 3 _%;
,
#0 r0 È{5
; Ç ^
.3 -È /ö 3 .» ]#{
.Æ71 0 "$"$# 5 0 0 ° 0 3 5 3
" "$! %Á2*.* .4" % ¦$ 2§ º4 ' ; ¶ $*2¶§ " "$! ³';12122" ]¥
;È.,
#
¸
_>ÀG
#G¹ 5
;8? ;:/
-È Á 3¹>5 ] )33 5 ´!K(-,"$ 5d< ¤ 3 0" ° 2*2" 1{§
3 %Á2¸ *2* ! 3 =<K)" 5 1," 0436 "$3"$! 5 5
º.+-!¹"$)1 É{ 1," 9§" i*2²d4¶8" 12+±*2 ²)² ,1 ;1 ;1{ ,¶)* É ;-
;.
9 8,
.
Æ: º§ ;§ "$"$%
7t5 º.+ ¶ ² 59 3 3 0 $)° ( 0 -3 .i³'0 2$ 5 3 "$.+ 3 5
/5 º.+ 0 ¶ 0 ² 9 5 ´ d' ! -.;0* .
{=³
³';4-² " ¶ 5?ÇKE'5(½12.;3 12" 0
*, ² ² 21 741.]3 9 i4"$" " ( 5 1.*,¶ É ¤ ¦-§§N
/5 ¤ 0 12 +-0 ,-*.!K5§ / )" 1 .* 3
ÇK5 0 ,.
5
Æ:3 2 5 ;Á;8 -{
3 ;ÈÈ0 _%;
,
0
W° &3 #
r
5 Æ
r K;#"«
&3 {
3 5d7t5
-: @É8%'.*.* 7 3 0 %À,¶ ´ "$"$* 5
º.+ ¶ 5 ² 9 ´ -2* ]³ ³'4² " ¶ É 2² .41," ¥" ,"$ *" *, ² ² 21 ;1 $ " ¥!#.+ " * É
/´5 0 0 4$5*,/ ( -. 3 º5 3 5 ´ 3 ³ 0 2( 0 -3 3° 3 º )+3 )0 " 1 2° * 0 ; Á:3 #K 3 3 5 3
5
° 4
0 3 7 ¸
5
° 4
0 3
3 I
- 5 5 0 6
3 3
/ 5 3 0 0 [
),4K
'4
L_ [ ²
* (K *$!
,
" "$"$%
º -
² ,
"
6
Î
,
1 -
2
&
> .
1
r
* " ´ #
! ² )
.
1
º ;
$
"
³ ! 5
º.+ ¶ ²)9 % º"$!#. º! )/$ 36&³ ³Á45 ² " ¶ 3 %'2" ¶ r$ 0 " *23 ²)² 0 ,1 1 '¶3 ;$* 0 É 5 2
/5 ?0 0 * 5 ]º º $ 5 )" 1 0 .* Æ#3 "«.5
r: 3 ?
5 ),40 %3-0 ]¥:(3 2
0 6'$° {0
1]¹ < 3 ´85 !K3¥(-7t 2" 5&I 0 36
7t5 3 É %Á3 2*2* 5 É 0 ²-.*2* 0 M[ )7
"$"$#
º.+ 5Á 7 ¶ ² < 9 º!5 3 $ `5 $ > " K !#² 1 4 §*2"$*K*i¶ ;;"$ 4 -¥²- ( $!
/K 5
),0 _ 0 > ])5 {:0 3 5 »{Ç&5-7 5 043 3 0 043 3 3 3 3° 0 5
"$"$# É ²-2*.*
º2+ ¶ ² 9 º! 5 3 $ ´ ´ 5 -2* =³ ³Á4² " ¶ º ² ² 21 1 !¥;1,+ *" "
/ 5 ¯9È0 41,0 -.?;½51.21,0 " É 5 »{Ç&5-7 5 I-5 5-/ _,
K3 &K5
), 3 K
5
Á400 /ö{¯° !0
21 "$"$0 # 3 3 3
3 º2+ ¶ ² 9 º 04 365 ´ 3 -:.- * [ ¦ Ä@," :*29 " % >@" § 4" % 4" t³ ³Á4² " ¶ ´ !¹²-52" `*2 ² ² ,1
/5 ;1 0 0
!#.+ " 5 *r" 3 1,+N4-5*./ *,"$ ¶ ;5 $*1 0 .)" 45 (-*20"$*@9:Á< ?5 1," 0 ;$0 *.*," 3 .* 5 PK+K>K 3 -,5 - >R0 'd >;3,
0
° + 0 3 3 3 0 3 043 Â 5
* !( *$#¹
!( %$* "$"( 5
+ º+ § %À;12±Î ,12 1,1 (,1`´ " "$!¥* ,12" 12+ § 9:.!# 2¶|9
I0 *21223 1,-24,»{<"$*2¶¥!#0 " "$!¹"$·412" 58/ É ,
.Ç
Æ :
0 2 : 5Á¸ --) _043%;
,
3
¹7 _'3 >]d3 {043 :5
;-N0 )
0
²-* %$#' ( % "$"$% 3 0435 3 [
5
I0 !#+ "$3 ·4º1,+-" 8 »< §) 40 1. %);²d12 & Î 5/ 1'+),"$12 .1,.1 2Ç+ @"$0 * (d21¹ ´ " "$!#*
.+ 5"$¸ 4-.;² 2041 36>@ 3 ´ 7 Î 3 .3 + "$043 4 5 ² ,1À>?´
21," 1,+ § º122-;1,-.4'2" *2¶`!#" "
"$% $*
3 »
"$"$%
4
0 i
3
0 ° » 3 3 r
5 < 3 0 0 ¼
< < 3 Ç 0 {
» r
<
Ç » »
º! 5 $¯ º.+ ¶ ² 9 i¶ ; (-*.N!¥;12+ `9 @²-41,1. 2 " 12" .2*2*2" 4²)²- ½"$!#412"
5 0 ²243 1 r/" 5 2*2" 0 0 5 @3;: ]¹È.9 3 /ö{» ¯!)
21 0 0
"$"$#
3 0 3 0436 043 0 043
º! $ 3 0 0
3 °
º.+ ¶ ² 9 ` 4
0
3 5 [ $ Ä .4À *21@9: ;12" *9 *2 ²)² ,1 ;1 ..*.*," 5 É ?-:
5 )È02 Èd/ 5 _'0 2
0 K&
B 3 K
,5 » Ç&?5
7
' 0 5 /ö¯3 !
2 1 0 "$"$3# 5 043 0 0 ° 0 04365 3
@[ ½ º! $ +-2Tº.+ ¶ ² 9 F 4-* (d21 $ +-
1," (d;1{ .4 $,"$·412"
²d.I-45 1 .*0
/*, ² ² 3 21 ;1 0 ¶0 ;$* 3 ?»{Ç
0
' 7 /ö± !)5
.2 < 1 "$"$04# 3)5 3 043 3 043
0º12" 1.* 3 Ä0 !¥!¥° 2!¥0 ³ 3 ³Á45 ² ?"
¶ 2³
0 ³ ¶ ´ 41,¶" * *21 º ² ² 21 1 ¥
7t 5 ¹5 !#² *2" 041,36" 5 .+ " 42² 36 ,1 5 §4 3 5 0§N ° ´ $5¸ ;² 3 21 3 -!&I-(d56@ ¸ ´rºÏ 04365 "( !$0! "$° "( 0 3-0°
Ä@" (d0 21?º120 . 04 365¹-< ,Æ3 ):t{ 0 ?)Ç@ È0
Æ G0 :)1
0 ]#ÈÈ 0 ; Ç$*2$;§ 0 ´8!& 3 ( 2"$ #%Á.*2»{* <rÇr» »
"$#$%
5
³' 23 (d;5 " É 1.2" @² " 1!¥[ ;1,+ * +@ 2" 1,+-!¥*5 ¸
9 .!K » $41," * :H^'
> _ >
] 5
ÇKM5 I-5 3
"$"4
5 3 0 0 3 0 0 3 0 043 5 [ I ' d %) {4 + $ !4
³Á5 ),(d" Î
ÇKº5 I125 412" *21,3 "$*
+-
5 ²d.41,@" * 3 *.3 .2+ 0 %Á,0" 3 ;1 0 ¤ 0 " 2*2" 1{§ "$"4 0
" 122" ² " 1 K9 ¿ .41,"$¹² .!#!#"
3 5]<
.+ "$4Á2² ,1 %' .!O"
3 0 0 3
³ ³Á4² " ¶ 4
0 3 Ç
K]¥:F &
P)
Æ
i@;
.Ƽ K]:;: &{ { ^): 5 3 4
0 3 3 °
5 @
> 4
4
¶ .
* "( "
5 E 3 4 " 3 *2+`5 12. 3 *,$412" 043 +rº² 2" 3 ³Á2 >; E 0 2¶ "$#$! 5 Á7=0 0 5
³ ³Á4² " ¶ 6)@
?
)
,
¥
'
d
È
:
È
È
;
-
I
)
º²-," ³'2$ >@; E ,¶ "$"$*
³ 5 ³Á4² 3 " ¶ 5 'd{:.È.d6
.;-I)
+ " $;§= 5 `º 3 * É » >@;FE ,¶ 0 " ²-2;²-.5 41,"
5
³ ³'4² " ¶ º Ä r"$.+ v º! $ 3 5 5 @ I
0 º)² ² 21 ;3 1 i04!#3 1,+ 3 o5$ 9 ¹9 ;0 1," 43 ² ²- ½" !¥41204" 365 ..*.*,"
=
3 ¸
5 *21,"$!¥43 1," 5 ±0 *20 "$ 4²- 3 *2*2" 5 0 !Æ "«5
?& 0 K
° , 0 ;% ; ]¥0 :& 0 r,3
.:043 ( '$.{1
0 ]¹ " + !$04#36) !$#( "$"$04% 3 5
4
0 6
3 3 3
Ä. 4+(- º ²)² ,1 ;1 r!¥.+ " * .;²- -;" ¥¶ ;+ " (d21*2² * ]12+-? É r,.
Æ:& 0 3 5 [
:)
< ¸ 5+',=0 2)° ¼0& 'd !3 ; .
0 ;3 :
0 3 /ö{¯ !
2 1 "$"$# 5 3
°d5 3
*21 Ä!#!¥.!¥ º1," 12* ³ ³'4² " ¶ ³ ³ ¶ N´ 412¶" * Ï *," 1{§=*,12"$!#412" -*,"
I-5 *2)¸ ² ² 0421 36 ;1 5 !¥.+ " * 36-7t 2+ (5 ¹5"$4.;² 043621 5 §4 3 5 §¯ 0° ´ 3 5 ;¸ ² 21 3 -!K 5 (d3 ´rºÏ 043 3
0 ° 0 3 5@< 3 0 Ç0 G@0 0 0 )Ç 0 3 »{<rÇr» "( » ! "$"( 5