TFW HW4

電信所 甲組 電波組 r10942024 陳嘉佑

1. Suppose that there are 6 signals. Their time length are all 10 seconds and their
frequencies are all from -2000Hz ~ 2000Hz. Also suppose that the channel is
almost full except for f  [200000 - 200t, 210000 - 200t] and f  [-210000 +
200t, -200000 + 200t], 0 < t < 12. How do we transmit the 6 signals by the
channel?
ANS:
There are six different signals with frequency spectrum ranging from -2000 Hz to
2000 Hz and duration 10 seconds. Here, I have assumed that all these signals range on
time domain from 0 to 10 seconds. However, here, we could first use Hilbert transform
plus with the original signal, where it could make Fourier transform of real-world signal
(real time function) without components distributing on negative frequency parts, i.e.
making their frequency spectrum only ranging from 0 to 2000 Hz. Then, now, we could
make the first five signals applied with Hilbert transform on the upper channel, which
is ranging for 𝑓 ∈ [200000 − 200𝑡, 210000 − 200𝑡], ∀𝑡 ∈ [0,12]. How to do that?
2
We could first multiply these five signals with chirp function 𝑒 𝑗𝜋𝑎𝑡 , where 𝑎 = −200.
And, do the following vertical shifting as the following:
×𝑒 400000𝑗π𝑡
1𝑠𝑡 → 𝑆𝑥 (𝑡, 𝑓 − 200000)
×𝑒 404000𝑗𝜋𝑡
2𝑛𝑑 → 𝑆𝑥 (𝑡, 𝑓 − 202000)
×𝑒 408000𝑗𝜋𝑡
3𝑟𝑑 → 𝑆𝑥 (𝑡, 𝑓 − 204000)
×𝑒 412000𝑗𝜋𝑡
4𝑡ℎ → 𝑆𝑥 (𝑡, 𝑓 − 206000)
×𝑒 416000𝑗𝜋𝑡
{ 5𝑡ℎ → 𝑆𝑥 (𝑡, 𝑓 − 208000)
Therefore, the overall transform for the first five signals would just like the below
shows:
2
×𝑒 −200𝑗𝜋𝑡
𝑥(𝑡) → 𝑥𝑎 (𝑡) = 𝑥(𝑡) + 𝑗𝑥𝐻 (𝑡) → 𝑠ℎ𝑒𝑎𝑟𝑖𝑛𝑔
×𝑒 400000𝑗𝜋𝑡
1𝑠𝑡 → 𝑆𝑥 (𝑡, 𝑓 − 200000)
×𝑒 404000𝑗𝜋𝑡
2𝑛𝑑 → 𝑆𝑥 (𝑡, 𝑓 − 202000)
×𝑒 408000𝑗𝜋𝑡
→ 3𝑟𝑑 → 𝑆𝑥 (𝑡, 𝑓 − 204000) (𝑣𝑒𝑟𝑡𝑖𝑐𝑎𝑙 𝑠ℎ𝑖𝑓𝑡𝑖𝑛𝑔)
×𝑒 412000𝑗𝜋𝑡
4𝑡ℎ → 𝑆𝑥 (𝑡, 𝑓 − 206000)
×𝑒 416000𝑗𝜋𝑡
{ 5𝑡ℎ → 𝑆𝑥 (𝑡, 𝑓 − 208000)
We have shown how to transmit first five signals to the channel ranging for 𝑓 ∈
 [200000 − 200𝑡, 210000 − 200𝑡], ∀𝑡 ∈ [0,12].
The last thing we should do is to transmit the last signal to the channel ranging for
𝑓 ∈ [−200000 + 200𝑡, −210000 + 200𝑡], ∀𝑡 ∈ [0,12] . Here, just like the above
procedure, first we use Hilbert transform to help us let the signal only ranging from 0
to 2000 Hz, and then multiply the signal after using Hilbert transform with chirp
2
function 𝑒 200𝑗𝜋𝑡 to perform shearing and 𝑒 −404000𝑗𝜋𝑡 to achieve vertical shifting.
Hence, overall, it would look like the following:
2
×𝑒 200𝑗𝜋𝑡 ×𝑒 −404000𝑗𝜋𝑡
𝑥(𝑡) → 𝑥𝑎 (𝑡) = 𝑥(𝑡) + 𝑗𝑥𝐻 (𝑡) → 𝑠ℎ𝑒𝑎𝑟𝑖𝑛𝑔 → 𝑣𝑒𝑟𝑡𝑖𝑐𝑎𝑙 𝑠ℎ𝑖𝑓𝑡𝑖𝑛𝑔

2. Suppose that x(t) is a stationary random process. Which of the following

functions are also stationary random processes? Why?
(i) exp(j10t)x(t); (ii) exp(j10t2)x(t); (iii) FrFT of x(t); (iv) the Fresnel
transform of x(t).
ANS:
(i.) 𝑒 10𝑗𝜋𝑡 𝑥(𝑡) is still a stationary random process, since multiplying 𝑥(𝑡) with
𝑒 10𝑗𝜋𝑡 merely means vertical shifting on frequency axis, which implies
𝐸[𝑊𝑥 (𝑡, 𝑓 − 𝑓0 )] = 𝑆𝑥 (𝑓 − 𝑓0 ) is still invariant with 𝑡.
2
(ii.) 𝑒 10𝑗𝜋𝑡 𝑥(𝑡) is not a stationary random process, since multiplying chirp function
means performing shearing on time axis, just like the below figure:

If 𝐸[𝑊𝑥 (𝑡, 𝑓)] = 𝑆𝑥 (𝑓) is not invariant with 𝑓, it would make the expectation
2
value of WDF of 𝑒 10𝑗𝜋𝑡 𝑥(𝑡) no longer invariant with 𝑡.
(iii.) FrFT of 𝑥(𝑡) is not a stationary random process. Just, take Fourier transform,
which is the special case of fractional Fourier transform, of 𝑥(𝑡) for example. We
all know that applying Fourier transform to a signal would make it turn 90°
around origin, which would make signal no longer invariant with 𝑡 unless the
expectation value of WDF of original random process 𝑥(𝑡) is invariant with 𝑓.
(iv.) The Fresnel transform of 𝑥(𝑡) is still a stationary random process, since the
Fresnel transform of 𝑥(𝑡) means taking convolution with chirp function, which
 would perform shearing on 𝑓 axis, just like the above figure. Observing the above
figure, since originally, the expectation of WDF of 𝑥(𝑡) is invariant with 𝑡 ,
performing shearing on frequency axis would not make it variant with 𝑡.
In conlusion, (i) and (iv) would still be stantionary random processess.

3. (a) What are the two main differences between the IMF and a sinusoid
function? (b) Which function is an IMF? Why? (i) (2+cos(10t))cos(2t); (ii)
cos(t3).
ANS:
(a) The two main differences between the IMF and a sinusoid function are 1. time-
variant amplitude and 2. time-variant instantaneous frequency.
(b) (2 + 𝑐𝑜𝑠(10𝜋𝑡)) 𝑐𝑜𝑠(2𝜋𝑡) is not an IMF, since the local minima of it are equal to
-1 but the local maxima of it are equal to 3, which obeys the second demand of IMF
stating at any point, the mean value of the envelope defined by the local maxima
and the envelope defined by the local minima is near to zero.
Yet, cos 𝜋𝑡 3 is like sinusoid function just with varying instantaneous frequency
but would not make the unequal number of local maxima and minima, and due to
its sinusoid-like function, the average of envelope defined by local maxima and that
defined by local minima is equal to zero.
Hence, only cos 𝜋𝑡 3 is IMF.

4. Among the Gabor transform, the WDF, the HHT, and the wavelet transform,
which one is better for the applications of (a) signal modulation, (b) random
process analysis, (c) climate data analysis, and (d) signal sampling? Also
illustrate the reasons.
ANS:
(a) Gabor transform is the most suitable technique for signal modulation, since in signal
modulation, we hope to add a signal into empty band by horizontal and vertical
shifting, dilation, shearing, generalized shearing and rotation. Therefore, we would
hope that cross terms would not occur, or otherwise, it is hard to analyze where
exists empty band for us to transmit the signal wanted to send into it. Using Gabor
transform would not cause problems of cross terms and also give us a suitable clarity.
(b) For random process analysis, WDF is the most suitable technique, since in its
definition, it naturally contains the information concerning to auto-covariance
function of a certain random process by the following explanation:
∞
Wx (t, f) = ∫ x(t − τ/2)x ∗ (t − τ/2) ∙ e−j2πfτ dτ
−∞
 And now, if we take expectation value of the above expression, we would naturally
include auto-covariance function into our expression just as below:
∞ ∞
𝐸[Wx (t, f)] = ∫ E[x(t − τ/2)x ∗ (t − τ/2)] ∙ e−j2πfτ dτ = ∫ 𝑅𝑥 (𝑡, τ)𝑒 −𝑗2π𝑓τ 𝑑τ
−∞ −∞

where R x (t, τ) is the auto-covariance function of a random process.

(c) For climate data analysis, since we hope to observe out the trend of climate, Hilbert-
Huang transform would be the most suitable technique to achieve such goal. The
basic spirit of Hilbert-Huang transform is to decompose a signal into many different
sinusoid-like components plus trend. With the help of it, we could decompose signal
out from lots of sinusoid-like components and gain the trend of climate data which
we really care about.
(d) For single sampling, we should base on what stages of sampling we are going to do.
Gabor transform is suitable for separation of components of signals, since we do
not hope cross terms problem to happen. After decomposing our signals into several
different components, we would rather use WDF to get better clarity instead of
Gabor transform, i.e. in the stage of using shearing or rotation to minimize the “area”
to each component, which is proportional to the number of points of sampling. In
conclusion, we could say that Gabor transform is used in stage two about
decomposition of signals into several independent components, whereas WDF is
recommended for the stage three for the goal of minimization of the area of each
independent component separated from the signal.

5. (a) What is the most important advantage of the Haar transform nowadays?
(b) Write the 7th row of the 16-point Haar transform.
ANS:
(a) Although due to more and more advanced technologies nowadays, the
computational advantage of the Haar transform become faded compared to before,
it is still the most convenient and fastest way to analyze out the higher frequency
components in different locations with different scales just by using addition and
subtraction. Especially, in image process, high frequency components are useful for
edge detections, which is helpful for image recognition.
(b) The 7th row of the 16-point Harr transform is the 7th row of 8-point Harr transform
⨂[1, 1] . Thus, it would be [0 0 0 0 1 −1 0 0] ⨂[1, 1] =
[0 0 0 0 0 0 0 0 1 1 −1 −1 0 0 0 0]
6. (a) What is the role of the vanish moment in the wavelet transform? (b)
Suppose that x(t) = 1+ at + bt2 for -2 < t < 2, x(t) = 0 otherwise. If x(t) has the
vanish moment of 2, determine a and b.
ANS:
(a) The role of the vanish moment in the wavelet transform is to represent the order of
high frequency of the mother wavelet function. If mother wavelet function has
larger p vanishing moments, it represents higher frequency function, which would
be more suitable to separate out fast changing parts of a signal.
∞ ∞
(b) If 𝑥(𝑡) has vanish moment of 2, it means that ∫−∞ 𝑡𝑥(𝑡)𝑑𝑡 = 0 & ∫−∞ 𝑥(𝑡)𝑑𝑡 =

0. For the second condition,

∞
𝑎 𝑏 2 16 3
∫ 𝑥(𝑡)𝑑𝑡 = 𝑡 + 𝑡 2 + 𝑡 3 | = 4 + 𝑏 = 0 ⇒ 𝑏 = −
−∞ 2 3 −2 3 4

For the first condition,

∞ 2
𝑡2 𝑎 𝑏 16
∫ 𝑡𝑥(𝑡)𝑑𝑡 = + 𝑡 3 + 𝑡 4 | = 𝑎=0⇒𝑎=0
−∞ 2 3 4 −2 3

7. Write a Matlab or Python program of the Hilbert-Huang transform.

y = hht(x, t, thr)
x: input, y: output (each row of y is one of the IMFs of x), t: samples on the t-
axis, thr: the threshold used in Step 7.
In Step 8, the number of non-boundary extremes can be no more than 3.
Step 9 is not required. The code should be handed out by NTUCool.
ANS:
In this homework, I do not consider boundary point as extreme point.

(extra): problem (學號尾數 4 和 9): 我們以前學過的哪個 transform 跟小波轉換

一樣，width of function 是跟頻率有關？
ANS:
S transform.