Z Transforms

. First difference system: y[n] = x[n] - x[n-1].

. Delays can be represented in the z domain. For example, a delay of 1 would become z^-1 in the z
domain.
. When you evaluate z transforms at z = e^jw, you can use the value of w embedded in your x[n].
. Nulling frequencies in the z-domain?
- We know how to do that in the n or discrete time domain.
- We use zeros to null frequencies. A zero on the unit circle at a certain frequency means all sinusoids at
that frequency will be set to zero.
- To null a frequency at 2000 Hz at an Fs of 8000 Hz, you will need a zero at 2pi*(2000/8000) or pi/2. Use
you the system function to null the frequency in question.
. Poles closer to the unit circle push magnitude response up at that frequency. Infinite if on the unit
circle. For this, we need a magnitude of 1.
. Zeros closer to the unit circle pull magnitude response down at that frequency. Zero if on the unit
circle. For this, we need a magnitude of 1.
. The angle in the p-z plot for the z transform is the frequency.
. Poles at the origin with a bunch of zeros on the unit circle = sine shaped frequency response.
. FIR filters have no poles or only poles at zero.
. IIR filters have both poles and zeros.
. The closer your pole/zero is to 1 in terms of magnitude, the stronger their impact in terms of pulling
down or pushing up the frequency response.
. For an infinite geometric sum to converge for example, 1/(1-A), the magnitude of the A in the
denominator needs to be smaller than 1.
. ROC defined the magnitude of z.
. 1 or a delta[n] converges for all values of z.
. Finite geometric sums = (1-a^L)/(1-a).
. Absolutely summable = finite magnitude of the geometric sum.
. Signal that are zero for all negative time, that type of signals are called causal signals, while the signals
that are zero for all positive value of time are called anti-causal signal.
. Right-sided sequences' ROCs extend outside a circle while for left-sided sequences, the ROC extends
inwards.

. To determine the unit step response of a system, multiply its system function H(z) with the z transform
of u[n]. This will be a convolution in the discrete time domain.

. If you have access to the system function, H(z), and you want to figure out the output in response to a
cosine or sine based signal, you must evaluate H(z) at z = e^jw.
- x(n) = Acos(wt+phi)
- y(n) = A|H(e^jw)|cos(wt+phi+<H(e^jw)).
- If it's a DC signal you are dealing with, use an omega or w of 0.

. High pass filters have the following z transforms: 1/(1+az^-1) where the value of a signifies how
strong/good the filter is.
. Low pass filters have the following z transforms: 1/(1-az^-1) where the value of a signifies how
strong/good the filter is.
. z^i - 1 = 0 yields i number of roots or angular frequencies of i/2*pi.
. What does the p/z plot and z transform look like for the different kinds of filters?

Stability and the ROC:

. The stability of a system can be determined by knowing the ROC alone. If the ROC contains the unit
circle (i.e., |z| = 1) then the system is stable.
. If you need stability then the ROC must contain the unit circle.
. If you need a causal system then the ROC must contain infinity and the system function will be a rightsided sequence. It must be outside the outermost pole.
. If you need an anticausal system then the ROC must contain the origin and the system function will be
a left-sided sequence. Must be inside the innermost pole.
. If you need both, stability and causality, all the poles of the system function must be inside the unit
circle.
. An absolutely summable sequence is stable and finite therefore it must include the unit circle and will
look like a washer.
. For the fourier transorm to exist for a signal, the roc must include the unit circle.

HOURLY 2
To Do List:
. What is leakage? Leakage occurs for peaks at non-integer values.

. Convolution
- Length of linear convolution = L + M - 1 where L and M are the discrete bandwidths of the signal.
- Length of circular convolution = 0 to N-1 where N is the number of samples in a linear convolution.
. DTFT and DFT of discrete time plots.
- DFT: L is the number of samples and N is basically the index number associated with the period. Refere
to Proakis Page 253.
. DFT
- Simple cosine/exponential based signals.
-- Simple cosine/exponential based signals in the continuous time domain in the case of the fourier
series would always produce peaks at 1 because their frequencies corresponded to the fundamental
frequency. In the case of the discrete time domain, the discrete fourier series is a sampled version of the
DTFT at a frequency of N. We calculate peak location on the basis of N.
- For simple cosine based signals, make sure you make adjustments to your signal according to the value
of N i.e. the frequency term needs to be in the form of 2pi/N and the amplitude at k_0 needs to be
equal to N.
- Complicated signals that rely on the use of the table of DFT pairs.
. DTFT
- Simple cosine/exponential based signals.
- Complicated signals that rely on the use of the table of DTFT pairs.
. Sampling
- A-to-D conversion etc.
- Aliasing; how to figure out the different aliases for a given signal.
- Downsampling.
- Upsampling.

. DFS or DFT = no negative values of k.

. Circular Convolution = Linear Convoultion + Aliases.
. Sampling in one domain introduces periodicity in the other.
. The fundamental frequency must be rational.
. The inverse DFT is just like the DFT. Look at the discrete impulses and reduce everything in terms of
sines and cosines.

. Circular Convlution:
- Do linear convolution on your signals.
- Linear convolution -> just shift the origin of your signal.
- N - 1 is the length of a single period of the convolved signal.
- Samples at n = N + all subsequent samples start over again at n = 0 i.e. they add to the samples at n = 0
and onwards.

. e^jw = DTFT = continuous and periodic.

. jw or w = DFS or DFT.
. N = discrete time period or the period as a result of sampling at Fs.
. N = Fs/F.
. Going from the box function in the discrete time domain to the continuous frequency domain is easy.
We are dealing with a bunch of impulses only.
. Aliases repeat after every 2pi so a given impulse will make an appearance after every 2pi in the discrete
domain and after every fs in the continuous frequency domain.
. Aliases at exactly 2pi are DC components.
. Realize that the frequency spectrum as a result of sampling will always be periodic so the spectrum of
x[n] will be periodic.
. Folding is when you use negative aliases. Even symmetry holds but make sure you change the sign of
the phase.
. The geometric sum as a result of the DFT = 1-a^L / 1 - a for k = 1 to N-1.
. The highest frequency in the k domain or the discrete frequency domain as a result of the DFT is one
that corresponds to the highest value of k possible which would be N-1.
. IDFT = 1/N * Inverse DFT.
. A single exponential in the frequency domain = an impulse or delta in the time domain.

Questions?
. Why is the phase delay exponential always decaying?
. How do you do the DFT or DFS for infinite sequences?
- The DFT and IDFT yield one period of the infinite sequences.

- It does not matter if the sequence is finite or infinite, we only use one period or N.
- For a finite length sequence, the value of N-1 in the sigma must correspond to the number of samples
in your signal.
. For a finite length sequence, does the number of k values also correspond to the number of samples in
your signal?
. Example 5.15, Signals and Systems, Oppenheim.
. pset05_3, McClellan.
. DFTs of real sequences and symmetry?
. The spectrum repeats itself for a DFT. Does the property of symmetry involve conjugates? Yes. For real
valued sequences. X_k = *X_N-k.
. N = Fs/F where F is the frequency resolution.
. Is L the number of samples for a DFT? No. It is the upper limit to your discrete signal.

DSP Preexam
. Revise Z-transforms + filters.
. Revise filters and their pz-plots and z-transforms.
. Revise all the transforms.
. Low priority: revise continuous convolution + look s09, Q2.a.
. Medium priority: revise the hourlies etc.
. Revise circular convolution.
. Look up the filter command in MATLAB.
. The multiplication property of the discrete fourier series.
. L = N for discrete time signals with no jumps.
. DFT needs some work.
. Revise freq. domain characteristics of lti systems.

. The length of continuous linear convolution = the two bandwidths starting from 0.
. The trick for cosine based functions in the case of the z transform which involves evaluating the system
function at z = e^jw also works for the continuous time fourier transform. Works for deltas. The output
is the basically the input with a phase change and a different amplitude.
. Remember there is always symmetry for filters.

. The cut-off frequency for a discretized and filtered signal should be handled with care. You need to
take into account the sampling rate.
. The range for a digital filter is between -fs/2 and fs/2.
. The combined system function is the product of the cascaded system functions and gives you the total
impulse response.
. The frequency response of a system or the magnitude of the system function is zero at the zeros. So
rearrange your z-transform or e^jw based expression to figure out the poles and zeros.

DSP Exam
. Nulling frequencies in the time domain.
- You sample them at frequencies where the undesired components overlap.
. DC component in the discrete time domain: signals between 0 and up or down.
. The z transform of a constant, A, is A/(1-z^-1).
. The DFT of a period impulse train produces a plot that has a DC component and a couple of
components like a cosine plot.
. Figuring out the DC gain from a PZ Plot?
. Filters and impulse response?
. For a signal to be periodic, it must contain frequency components that are multiples of the
fundamental frequency only. All other components MUST GO AWAY.
. The maximum frequency out of an ideal D to C converter is equal to Fs/2.
. The time delay property of the fourier transform: x(t-a) <=> X(F)e^-jwa.