
Signals and Systems
Slide a flipped signal to convolve, add harmonics until a square wave appears, and drag poles to shape a filter
22 chapters, 22 interactives, few words. Every chapter is mapped to the Anna University, VTU, SPPU, Mumbai University and KTU units it covers: elementary signals and operations, classification of signals and systems, convolution sum and integral, impulse and step response, difference equations, Fourier series and the Fourier transform, the Laplace transform and its region of convergence, s-domain analysis, correlation and spectral density, sampling and aliasing, the DTFT and DFT, the z-transform and FIR and IIR systems, and probability and random variables.
- AUEC3354Signals and Systems
- VTUBEC402Basic Signal Processing (modules 3-5)
- SPPU204191Signals & Systems
- MUECC404Signals & Systems
- KTUPCECT402Signals and Systems
By Skill-Lync · Updated 19 Nov 2026 · 22 chapters · 22 interactives · free, no sign-up
Elementary signals: step, ramp, impulse, exponentials, sinusoids, sinc and signum
Which syllabus covers thisAU EC3354 Unit IVTU BEC402 Module 3SPPU 204191 Unit IMU ECC404 Module 1KTU PCECT402 Module 1
How are the unit step, the ramp and the unit impulse related?
The unit step u(t) is 0 for negative t and 1 from t = 0 on (this book takes u(0) = 1). Integrate it and you get the ramp r(t) = t·u(t); differentiate it and you get the impulse δ(t). Picture δ(t) as a rectangular pulse of width ε and height 1/ε: its area is always 1, but as ε shrinks it gets taller and thinner. At ε = 0.5 the height is 2; at ε = 0.05 it is 20. In the limit the width is zero, the height has no finite value and only the area, 1, survives.
What does the impulse do inside an integral?
It picks out one value. Multiply any smooth x(t) by δ(t − t0) and integrate: the result is x(t0), because the product is zero everywhere except at t0, where the area is 1. So ∫(t² + 3t)δ(t − 2)dt = 4 + 6 = 10. Two cautions: if t0 lies outside the limits of integration the answer is 0, and a scaled impulse obeys δ(at) = δ(t)/|a|, so ∫(t + 4)δ(3t − 6)dt = (2 + 4)/3 = 2.
Why is the complex exponential e^(st) the most important signal of all?
With s = σ + jω it holds every other smooth signal of this chapter. σ sets growth (positive σ) or decay (negative σ) of the envelope e^(σt); ω sets how fast it turns. ω = 0 gives a real exponential, σ = 0 gives e^(jωt) = cos ωt + j sin ωt, and adding e^(jωt) and e^(−jωt) gives 2 cos ωt. India's 230 V, 50 Hz mains is the sinusoid 325.3 sin(314.2t): peak 230√2 V, period 20 ms. Linear time-invariant systems only scale and delay e^(st), which is why Fourier and Laplace are built on it.
What are the sinc and signum functions, and where do they appear?
sinc(t) = sin(πt)/(πt), with sinc(0) = 1. It crosses zero at every non-zero integer and its total area is 1; stretched to sinc(t/T) its zeros sit at ±T, ±2T and its area and energy are both T. It is the spectrum of a rectangular pulse and the impulse response of an ideal low-pass filter. The signum sgn(t) is −1 for negative t, +1 for positive t and 0 at t = 0; it equals 2u(t) − 1 and is odd, so it shows up in even/odd splits and Fourier tables.
How do the discrete-time versions differ?
They live only at integer n. The unit sample δ[n] is simply 1 at n = 0 and 0 elsewhere: no limit, no infinite height, and δ[n] = u[n] − u[n − 1]. A DT sinusoid cos(Ωn) has a twist: adding 2π to Ω gives the same samples, so cos(0.25πn) and cos(2.25πn) are identical sequences, and the fastest possible oscillation is Ω = π, the alternating sequence (−1)ⁿ. The complex exponential becomes zⁿ = rⁿe^(jΩn): r below 1 decays, r above 1 grows.
The unit-area pulse that models δ(t) is narrowed from ε = 0.5 s to ε = 0.05 s. Its area and its energy ∫p²dt
The height is 1/ε, so the area (1/ε)·ε is always 1, while the energy (1/ε)²·ε = 1/ε grows from 2 to 20. That is why the ideal impulse has unit area but is neither an energy nor a power signal.
Worked examples
Mains voltage as a sinusoid · India's supply: 230 V RMS at 50 Hz, modelled as v(t) = Vp sin(ωt)
- Peak Vp = 230 × √2 = 325.27 V
- ω = 2π × 50 = 314.16 rad/s, period T = 1/50 s = 20 ms
- At t = 3 ms: v = 325.27 × sin(314.16 × 0.003) = 325.27 × sin(0.3π) = 263.1 V
- Mean of v² over one period = Vp²/2 = 52,900 V², whose square root is the 230 V RMS value
Answer: v(t) = 325.3 sin(314.2t) V; v(3 ms) = 263.1 V
Sifting with the impulse · Evaluate (a) ∫(t² + 3t)δ(t − 2)dt over all t, (b) ∫ from 0 to 5 of e^(−t)δ(t + 1)dt, (c) ∫(t + 4)δ(3t − 6)dt over all t
- (a) The impulse sits at t = 2, inside the limits: 2² + 3 × 2 = 10
- (b) The impulse sits at t = −1, outside 0 to 5: the integrand is zero everywhere in range, so 0
- (c) δ(3t − 6) = δ(3(t − 2)) = δ(t − 2)/3
- So (c) = (2 + 4)/3 = 2
Answer: (a) 10, (b) 0, (c) 2
A decaying complex exponential · x(t) = e^((−0.5 + j4)t); find x(1), its envelope and how fast it turns
- |x(1)| = e^(−0.5) = 0.6065
- Re x(1) = 0.6065 × cos 4 = −0.3965, Im x(1) = 0.6065 × sin 4 = −0.459 (4 rad, not degrees)
- It turns once every 2π/4 = 1.5708 s; the envelope falls by e^(−0.5 × 1.5708) = 0.456 per turn
- Time constant of the envelope = 1/0.5 = 2 s
Answer: x(1) = −0.396 − j0.459; it shrinks to 45.6% every 1.571 s turn
Operations on signals: shifting, reversal, time scaling and building waveforms from steps and ramps
Which syllabus covers thisAU not in syllabusVTU BEC402 Module 3SPPU 204191 Unit IMU ECC404 Module 1KTU PCECT402 Module 1
What do x(t − t0) and x(−t) do to a signal?
x(t − t0) is a time shift. If t0 is positive the whole waveform moves right, later by t0 seconds: whatever x did at t = 0, the new signal does at t = t0. x(t + t0) moves it left, earlier. x(−t) is time reversal, a mirror image about t = 0, like playing a recording backwards. A combination such as x(−t + 2) = x(−(t − 2)) is the mirror image moved right by 2: a feature of x at t = −1 now appears at t = 3.
What does time scaling x(at) do, in continuous and in discrete time?
In CT, x(2t) runs twice as fast: the waveform is squeezed towards t = 0 to half its width, and its area halves. x(t/2) stretches it to double width. A negative a also reverses it. In DT the index must stay an integer, so x[2n] keeps only the even-numbered samples and throws the rest away (decimation; information is lost), while x[n/2] is defined only for even n and the gaps are filled with zeros (interpolation needs a rule). The book's seven-sample x[n] keeps 4 samples under x[2n − 1].
In y(t) = x(at + b), where does each point of x land, and why does the order of operations matter?
Set the argument equal to an old time: a·t′ + b = t, so a feature of x at time t lands at t′ = (t − b)/a. That one line beats any recipe. If you prefer steps: shift first, x(t + b), then replace t by a·t. Or scale first, x(at), and then shift by b/a, not by b, because x(a(t + b/a)) = x(at + b). Scaling first and then shifting by b gives x(at + ab), the most common exam slip.
How do you write a waveform as a sum of steps and ramps?
Walk along the waveform from left to right. At every jump of size J at time t1 add J·u(t − t1); at every change of slope by S at time t2 add S·r(t − t2). The trapezoid that rises from 0 at t = −1 to 2 at t = 0, stays at 2 until t = 1 and falls to 0 at t = 3 becomes 2r(t + 1) − 2r(t) − r(t − 1) + r(t − 3). Check the far right: the slopes 2 − 2 − 1 + 1 add to 0, so the signal stays flat, as it should. Its area is 5.
What are amplitude scaling, addition and multiplication of signals?
They act value by value at each instant. A·x(t) stretches the waveform vertically; a negative A also flips it upside down. x1(t) + x2(t) adds the two heights at every t, which is how a waveform is built from steps and ramps. x1(t)·x2(t) multiplies them: multiplying by u(t) keeps only the part for t ≥ 0, multiplying by a rectangular pulse cuts out a window, and multiplying by cos(ω0t) is the amplitude modulation a radio transmitter performs.
To draw y(t) = x(2t − 4) from x(t), which recipe is right?
Shift first: x(t − 4), then replace t by 2t to get x(2t − 4). Compressing first gives x(2t), and that must then move right by b/a = 2, not 4: shifting by 4 gives x(2(t − 4)) = x(2t − 8). Option 3 mixes the two and gives x(2t − 2).
Worked examples
Reverse, scale and shift in one go · x(t) is the trapezoid with corners (−1, 0), (0, 2), (1, 2), (3, 0). Sketch y(t) = x(−2t + 3)
- Each corner at time t lands at t′ = (t − 3)/(−2) = (3 − t)/2
- t = −1 → 2, t = 0 → 1.5, t = 1 → 1, t = 3 → 0
- So y rises from 0 at t = 0 to 2 at t = 1, stays at 2 until t = 1.5 and falls to 0 at t = 2: reversed and half as wide
- Check one point: y(0.5) = x(−1 + 3) = x(2) = 1, on the falling edge of x
- Area: x has area 5; scaling time by |a| = 2 halves it to 2.5
Answer: y(t) is non-zero on 0 < t < 2, peak 2 for 1 ≤ t ≤ 1.5, area 2.5
A waveform from steps and ramps · g(t) = 0 for t < 0, rises with slope 2 to 4 at t = 2, stays at 4 until t = 3, drops at once to 1, stays at 1 until t = 5, then drops to 0
- Slope changes: +2 at t = 0, −2 at t = 2 → 2r(t) − 2r(t − 2)
- Jumps: −3 at t = 3 and −1 at t = 5 → −3u(t − 3) − u(t − 5)
- Check: g(1) = 2 × 1 = 2, g(2.5) = 5 − 1 = 4, g(4) = 8 − 4 − 3 = 1, g(6) = 12 − 8 − 3 − 1 = 0
- Area = triangle ½ × 2 × 4 + 4 × 1 + 1 × 2 = 10
Answer: g(t) = 2r(t) − 2r(t − 2) − 3u(t − 3) − u(t − 5); area 10
Decimation and interpolation of a sequence · x[n] = 1, 2, 2, 2, 1.5, 1, 0.5 for n = −1 to 5 (zero elsewhere). Find y[n] = x[2n − 1] and describe z[n] = x[n/2]
- y[n] needs 2n − 1 between −1 and 5, so n = 0 to 3
- y[0] = x[−1] = 1, y[1] = x[1] = 2, y[2] = x[3] = 1.5, y[3] = x[5] = 0.5
- The even-indexed samples x[0], x[2], x[4] are lost; the sum of samples drops from 10 to 5
- z[n] = x[n/2] exists only for even n: the 7 samples spread over n = −2 to 10 with a zero in each of the 6 gaps
Answer: y[n] = 1, 2, 1.5, 0.5 for n = 0 to 3; z[n] has 7 non-zero samples and 6 inserted zeros
Classifying signals: even and odd, periodic and aperiodic, energy and power, deterministic and random
Which syllabus covers thisAU EC3354 Unit IVTU BEC402 Module 3SPPU 204191 Unit IMU ECC404 Module 1KTU PCECT402 Module 1
How is any signal split into an even part and an odd part?
An even signal is its own mirror image, x(−t) = x(t), like cos t. An odd one flips sign, x(−t) = −x(t), like sin t or sgn t. Every signal is the sum of one of each: average x(t) with its mirror for the even part, take half the difference for the odd part. For u(t) that gives 1/2 plus ½·sgn(t). The two parts are orthogonal, so their energies (or powers) simply add up to that of x. Shift a signal and its split changes completely.
When is a sum of sinusoids periodic, and when is a discrete sinusoid periodic?
x(t) is periodic if x(t + T) = x(t) for all t; the smallest such T is the fundamental period. A sum of two periodic signals is periodic only if the ratio of their periods is rational; the period is then their least common multiple. cos(2πt) + cos(3πt), at 1 Hz and 1.5 Hz, repeats every 2 s; with 1 Hz and √2 Hz it never repeats. In DT, cos(Ω0n) is periodic only if Ω0/2π = k/N is rational: cos(0.3πn) repeats every 20 samples, but cos(n) never does.
What separates an energy signal from a power signal?
Energy is the total area under |x(t)|²; power is that area per second, averaged over a window that grows without limit. A pulse or a decaying exponential has finite energy (e^(−0.5t)u(t) has E = 1), so its average power is 0: an energy signal. A step, a sinusoid or a periodic wave has infinite energy but finite power (2u(t) has P = 2): a power signal. A ramp or a growing exponential has neither. No signal can be both.
What are deterministic and random signals, and causal and anticausal ones?
A deterministic signal is known exactly at every instant from a formula or a table: a test tone, a clock, a step command. A random signal can only be described by statistics such as its mean, variance and spectrum, because its next value cannot be predicted exactly: thermal noise in a resistor, speech, a sensor reading jittering about its true value. Separately, a signal that is zero for all negative t is causal, one that is zero for all positive t is anticausal, and anything else is non-causal.
Is x(t) = cos(6t) + cos(10t) periodic?
The periods are 2π/6 = π/3 and 2π/10 = π/5. Their ratio is 5/3, rational, so the sum is periodic with period equal to the LCM: 2π/gcd(6, 10) = 2π/2 = π s. The π cancels in the ratio; only an irrational ratio of the two periods breaks periodicity.
Worked examples
Even and odd parts of a decaying exponential · x(t) = e^(−2t)u(t)
- x(−t) = e^(2t)u(−t), so xe(t) = ½e^(−2|t|) and xo(t) = ½sgn(t)e^(−2|t|)
- At t = 0.5: x = e^(−1) = 0.3679, xe = 0.1839, xo = 0.1839; they add to x
- At t = −0.5: x = 0, xe = 0.1839, xo = −0.1839; they cancel
- Energies: E(x) = ∫e^(−4t)dt from 0 = 0.25, E(xe) = E(xo) = 0.125, and 0.125 + 0.125 = 0.25
Answer: xe = ½e^(−2|t|), xo = ½sgn(t)e^(−2|t|), each with energy 0.125
Period and power of a two-tone signal · x(t) = 3cos(80πt) + 2sin(120πt + π/4)
- Frequencies 40 Hz and 60 Hz; periods 25 ms and 16.67 ms
- Ratio 60/40 = 3/2 is rational; fundamental frequency gcd(40, 60) = 20 Hz, so T0 = 50 ms
- In 50 ms the first tone makes 2 cycles and the second 3
- Power: each sinusoid gives (amplitude)²/2 and the cross term averages to zero, P = 9/2 + 4/2 = 6.5
Answer: Periodic, T0 = 50 ms; P = 6.5 (power signal)
Energy, power or neither? · (a) 4e^(−3|t|), (b) 3 + 2cos(10t), (c) x[n] = (0.8)ⁿu[n]
- (a) E = 2∫16e^(−6t)dt from 0 = 32/6 = 5.333: finite, so an energy signal with P = 0
- (b) Mean of (3 + 2cos 10t)² = 9 + 0 + 4 × ½ = 11: P finite, E infinite, a power signal
- (c) E = Σ(0.64)ⁿ for n ≥ 0 = 1/(1 − 0.64) = 2.778: an energy sequence
- A signal with finite non-zero power always has infinite energy, and vice versa
Answer: (a) energy, E = 5.333; (b) power, P = 11; (c) energy, E = 2.778
System properties: linearity, time invariance, causality, memory, stability and invertibility
Which syllabus covers thisAU EC3354 Unit IVTU BEC402 Module 3SPPU 204191 Unit IMU ECC404 Module 1KTU PCECT402 Module 1
What makes a system linear, and why is y = 2x + 3 not linear?
A linear system obeys superposition: scale an input and the output scales by the same factor; add two inputs and the outputs add. One test covers both: feed a·x1 + b·x2 and compare with a·y1 + b·y2. A quick screen: a linear system must give zero output for zero input. y = 2x + 3 gives 3 for x = 0, so it fails; it is called incrementally linear, because changes in the output are linear in changes of the input. y = x² fails because doubling x quadruples y.
How do you test for time invariance?
A time-invariant system behaves the same today as tomorrow: delay the input by t0 and the output is the old output delayed by t0, nothing else. Test it in two orders. First delay the input and pass it through the system; then pass the original and delay the output. If the two agree for every t0 the system is time-invariant. A coefficient that depends on t breaks it (y = t·x(t)), and so does any change to the time axis itself (y = x(2t), y = x(−t)).
What do causal and memoryless mean?
A causal system's output at time t depends only on the input at t and before, never after: y(t) = x(t − 2) is causal, y(t) = x(t + 1) is not, and y(t) = x(−t) is not either, because y(−3) needs x(3). A memoryless (static) system uses only the input at the same instant: y = 2x + 3, y = x² and y = t·x(t). Anything that stores, delays or integrates has memory (dynamic). Real-time systems must be causal; processing recorded data offline need not be.
What is BIBO stability, and why is an integrator unstable?
Bounded input, bounded output: if every input with |x| ≤ M gives an output with |y| ≤ some finite B, the system is BIBO stable. y = 2x + 3 is stable, with B = 2M + 3. The running integral is not: feed it the bounded step u(t) and the output is the ramp t, which reaches 10 at t = 10 s and 100 at t = 100 s, growing for ever. y = t·x(t) fails the same way. One unbounded output from one bounded input is enough to prove instability.
When is a system invertible?
When different inputs always give different outputs, so a second system can recover the input. y = 2x + 3 is undone by x = (y − 3)/2, a delay by an advance, the running integral by a derivative. y = x² is not invertible because x and −x give the same output. Neither is y[n] = x[n] + x[n − 1]: the inputs 0 and (−1)ⁿ both give y = 0. y = t·x(t) loses x(0), since the output at t = 0 is always 0.
y[n] = x[n] + x[n − 1] is linear, time-invariant, causal and BIBO stable. Which property does it lack?
The inputs x[n] = 0 and x[n] = (−1)ⁿ both give y[n] = 0 for every n, so the output cannot tell them apart and no inverse system exists. It also has memory, since it uses x[n − 1], but that was not an option.
Worked examples
Testing y(t) = x(t)·cos(5t) · A modulator multiplies its input by cos(5t). Check linearity, time invariance, causality, memory, stability and invertibility
- Linear: (a x1 + b x2)cos 5t = a·x1 cos 5t + b·x2 cos 5t ✓
- Time invariance, with x(t) = t, delay 0.2 s, at t = 0.5 s: system on the delayed input gives 0.3 × cos 2.5 = −0.2403; the delayed output gives 0.3 × cos 1.5 = 0.0212. They differ, so time-varying
- Causal and memoryless: y(t) uses only x(t) at the same instant
- Stable: |y| ≤ |x| ≤ M. Not invertible: at t = π/10 = 0.314 s, cos(5t) = 0 and x there is lost
Answer: Linear, time-varying, causal, memoryless, BIBO stable, not invertible
A bound for a three-tap system · y[n] = x[n] − 2x[n − 1] + 3x[n − 3], any input with |x[n]| ≤ 1
- Linear and time-invariant (constant coefficients, no n outside the brackets); causal; has memory
- |y[n]| ≤ 1·|x[n]| + 2·|x[n − 1]| + 3·|x[n − 3]| ≤ 1 + 2 + 3 = 6
- The bound is reached: x[n] = 1, x[n − 1] = −1, x[n − 3] = 1 gives y[n] = 1 + 2 + 3 = 6
- So the system is BIBO stable with B = 6M
Answer: BIBO stable; the largest possible |y| for |x| ≤ 1 is 6
Slowing a signal down: y(t) = x(t/3) · Is y(t) = x(t/3) causal and time-invariant?
- y(−6) = x(−2): the output at t = −6 s needs the input 4 s in the future, so it is not causal (for positive t the system only looks back, but one failure is enough)
- Delay the input by 1 s, then apply the system: x(t/3 − 1). Apply the system, then delay: x((t − 1)/3) = x(t/3 − 1/3)
- With x(t) = e^(−t²) at t = 0: e^(−1) = 0.3679 against e^(−1/9) = 0.8948, so time-varying
- It is linear, has memory, is stable, and is invertible: x(t) = y(3t)
Answer: Not causal and not time-invariant; linear, stable and invertible
Discrete-time LTI systems and the convolution sum
Which syllabus covers thisAU EC3354 Unit VVTU BEC402 Module 4SPPU 204191 Unit IIMU ECC404 Module 2KTU PCECT402 Module 1
Why is the impulse response enough to describe a discrete-time LTI system?
Any sequence is a stack of scaled, shifted impulses: x[n] is x[0]δ[n] plus x[1]δ[n − 1] plus x[2]δ[n − 2], and so on. Feed in one impulse and the system returns h[n], its impulse response. Time invariance says a shifted impulse returns a shifted h, and linearity says the scaled pieces add. So the output is a sum of shifted copies of h, each weighted by one input sample. Know h[n] and you can predict the output for every input.
How do you work out the convolution sum by hand, the graphical way?
Plot x[k] against k. Flip h to get h[−k], then slide it right by n to get h[n − k]. Multiply the two sample by sample and add the products: that sum is y[n]. Move n by one and repeat. Two 4-sample pulses of height 1 give 1, 2, 3, 4, 3, 2, 1: the overlap grows by one sample per step, peaks at 4 when the pulses line up at n = 3, then shrinks. Two equal rectangles always give a triangle.
How long is the output, and where does it start?
If x has N samples starting at n1 and h has M samples starting at n2, the output has N + M − 1 samples and starts at n1 + n2. A quick way to get the numbers is to multiply the two sequences like the coefficients of two polynomials: the convolution sum and polynomial multiplication are the same operation. A 5-reading record through a 3-point averager gives 7 output samples; the first and last two are only partly filled.
Which closed forms are worth remembering?
u[n] * u[n] = (n + 1)u[n], a ramp. aⁿu[n] * u[n] = (1 − aⁿ⁺¹)/(1 − a), which settles at 1/(1 − a) when |a| < 1: for a = 0.8 that is 5. aⁿu[n] * bⁿu[n] = (aⁿ⁺¹ − bⁿ⁺¹)/(a − b) for a ≠ b, and (n + 1)aⁿ when a = b. With a = 0.9 and b = 0.5 the output rises to a peak of 1.51 at n = 2 before decaying. Every one of these is a geometric series summed over the overlap.
x[n] = 0.5ⁿu[n] is fed to an accumulator, h[n] = u[n]. What does y[n] approach as n grows?
y[n] = Σ 0.5ᵏ for k = 0 to n = (1 − 0.5ⁿ⁺¹)/0.5, which tends to 2. The accumulator is not BIBO stable (a step makes it climb forever), but this input has a finite sum, so the output settles.
Worked examples
A 3-point moving average · Readings 21, 24, 18, 27, 30 (illustrative, one per second from n = 0) pass through h[n] = 1/3 for n = 0, 1, 2
- N = 5, M = 3, so the output has 5 + 3 − 1 = 7 samples, n = 0 to 6
- y[0] = 21/3 = 7, y[1] = (21 + 24)/3 = 15: the window is only partly full
- y[2] = (21 + 24 + 18)/3 = 21, y[3] = (24 + 18 + 27)/3 = 23, y[4] = (18 + 27 + 30)/3 = 25
- y[5] = (27 + 30)/3 = 19, y[6] = 30/3 = 10
Answer: y = 7, 15, 21, 23, 25, 19, 10 for n = 0 to 6; only n = 2, 3, 4 are true 3-point averages
Sequences that do not start at zero · x[n] = 2, −1, 3 for n = −1, 0, 1 and h[n] = 1, 2, −2 for n = 1, 2, 3
- Start index = −1 + 1 = 0, length = 3 + 3 − 1 = 5, so y runs from n = 0 to 4
- Multiply like polynomials: (2 − D + 3D²)(1 + 2D − 2D²), the power of D counting samples from each start
- Collect: 2, then 4 − 1 = 3, then −4 − 2 + 3 = −3, then 2 + 6 = 8, then −6
- Check y[2] directly: Σ x[k]h[2 − k] = x[−1]h[3] + x[0]h[2] + x[1]h[1] = −4 − 2 + 3 = −3
Answer: y[n] = 2, 3, −3, 8, −6 for n = 0, 1, 2, 3, 4
Step response of a decaying system · h[n] = 0.6ⁿu[n], input x[n] = u[n]
- y[n] = Σ 0.6ᵏ for k = 0 to n, a geometric series = (1 − 0.6ⁿ⁺¹)/(1 − 0.6)
- y[0] = 1, y[1] = 1.6, y[3] = (1 − 0.6⁴)/0.4 = 2.176
- Final value = 1/(1 − 0.6) = 2.5
- 95 % of it needs 0.6ⁿ⁺¹ ≤ 0.05, so n + 1 ≥ ln 0.05/ln 0.6 = 5.86: n = 5
Answer: y[n] = 2.5(1 − 0.6ⁿ⁺¹), reaching 95 % of its final 2.5 at n = 5
Continuous-time convolution integral by the graphical method
Which syllabus covers thisAU EC3354 Unit IIIVTU not in syllabusSPPU 204191 Unit IIMU ECC404 Module 2KTU PCECT402 Module 1
What does the convolution integral calculate?
It is the continuous-time version of the convolution sum: the impulse response h(t) is the system's reply to a very short, unit-area kick, and any input can be cut into such kicks, x(τ)dτ each. The output at time t adds up every past kick, weighted by how strongly the system still remembers it, h(t − τ). A system with a short h forgets fast and follows the input closely; a long h smooths and delays it.
How does the graphical method work, step by step?
Draw x(τ) on a τ axis. Flip h to get h(−τ) and shift it right by t to get h(t − τ). For each range of t, find where the two overlap, multiply them, and integrate over the overlap. The answer usually comes in pieces: no overlap, partial overlap, full overlap, leaving. For a unit step through a 2 s pulse, y(t) = t for 0 < t < 2 and stays at 2 afterwards: the area under the pulse.
How do you get the limits of integration right?
If x(τ) is non-zero for a1 < τ < b1 and h(t) for a2 < t < b2, then h(t − τ) is non-zero for t − b2 < τ < t − a2. The overlap runs from the larger lower end to the smaller upper end, and is empty if they cross. The output lives from a1 + a2 to b1 + b2: durations add, with no minus one as in the discrete case. A 1 s pulse with a 2 s pulse gives a 3 s trapezoid.
What do two exponentials give, and what happens when the rates match?
e^(−at)u(t) * e^(−bt)u(t) = (e^(−at) − e^(−bt))/(b − a) for t > 0: it starts at zero, rises, peaks and decays. The formula looks like 0/0 when a = b, but the limit is t·e^(−at). For a = 2 per second the output peaks at t = 1/a = 0.5 s with height 0.184. Two identical RC stages with a buffer between them behave this way: their impulse response starts at zero instead of jumping.
Two rectangular pulses of height 1, one 2 s wide and one 3 s wide, both starting at t = 0, are convolved. The result is
The overlap grows for 2 s, stays at the full width of the shorter pulse (2) while the shorter one slides inside the longer one, and shrinks for the last 2 s. Durations add: 2 + 3 = 5 s. Only equal widths give a triangle.
Worked examples
RC low-pass charged by a step · R = 10 kΩ, C = 10 µF (illustrative), h(t) = (1/RC)e^(−t/RC)u(t), input a 5 V step at t = 0
- RC = 10⁴ × 10⁻⁵ = 0.1 s, so h(t) = 10e^(−10t) per second
- Overlap of u(τ) and h(t − τ) is 0 < τ < t: y(t) = ∫₀ᵗ 5 × 10e^(−10(t − τ)) dτ
- y(t) = 5(1 − e^(−10t)) V; at t = RC = 0.1 s, y = 3.16 V (63 % of 5 V)
- Time to 4.5 V: e^(−10t) = 0.1, t = 0.1 ln 10 = 0.23 s
Answer: y(t) = 5(1 − e^(−10t)) V for t > 0; 4.5 V after 0.23 s
A pulse into a first-order system · x(t) = 1 for 0 < t < 2 s, h(t) = e^(−3t)u(t)
- t < 0: no overlap, y = 0
- 0 < t < 2: overlap 0 < τ < t, y = ∫₀ᵗ e^(−3(t − τ)) dτ = (1 − e^(−3t))/3
- t > 2: overlap 0 < τ < 2, y = (e^(−3(t − 2)) − e^(−3t))/3
- y(1) = 0.3167, y(2) = 0.3325, y(3) = 0.0166
Answer: Rises towards 1/3 while the pulse is on (y(2) = 0.3325), then decays as e^(−3(t − 2))
Two exponentials and the peak · x(t) = e^(−t)u(t), h(t) = e^(−4t)u(t)
- Overlap 0 < τ < t: y = ∫₀ᵗ e^(−τ) e^(−4(t − τ)) dτ = e^(−4t) ∫₀ᵗ e^(3τ) dτ
- y(t) = (e^(−t) − e^(−4t))/3 for t > 0
- Peak where dy/dt = 0: e^(−t) = 4e^(−4t), so t = ln 4/3 = 0.4621 s
- y at the peak = (e^(−0.4621) − e^(−1.8484))/3 = 0.1575
Answer: y(t) = (e^(−t) − e^(−4t))/3, peaking at 0.1575 when t = 0.4621 s
LTI properties from the impulse response: causality, stability, step response and interconnections
Which syllabus covers thisAU EC3354 Unit IIIVTU BEC402 Module 4SPPU 204191 Unit IIMU ECC404 Module 2KTU PCECT402 Module 1
How do you read causality, memory and stability straight from h?
Causal: h is zero for every t < 0 (or n < 0), so the output never uses future input. Memoryless: h is a single scaled impulse at zero, K·δ(t), so y = K·x. BIBO stable: the area under |h| is finite, ∫|h(t)|dt for continuous time or Σ|h[n]| for discrete time. The sequence 0.9ⁿu[n] sums to 10, so it is stable; u[n] sums to infinity, so the accumulator is not.
How is the step response related to the impulse response?
The step is the running integral of the impulse, so for an LTI system the step response is the running integral of h: s(t) = ∫ h(τ)dτ from −∞ to t, and h = ds/dt. In discrete time s[n] is the running sum of h[n], and h[n] = s[n] − s[n − 1]. The final value of s is the total area of h, the DC gain. A lab step test therefore gives you h too, by differentiating the measured curve.
What is the overall impulse response of systems in series and in parallel?
In series (cascade) the impulse responses convolve: h = h1 * h2. Convolution is commutative, so swapping the two LTI blocks changes nothing at the output. In parallel the outputs add, so h = h1 + h2. With h1 = e^(−t)u(t) and h2 = e^(−2t)u(t), the cascade gives e^(−t) − e^(−2t) with area 0.5, and the parallel pair has area 1 + 0.5 = 1.5. The cascade is stable and causal if both parts are.
When is a system invertible, and what is deconvolution?
A system is invertible if a second system can undo it: h * h_inv = δ. An echo h[n] = δ[n] + 0.5δ[n − 1] is undone by h_inv[n] = (−0.5)ⁿu[n], which is stable because its terms shrink. Deconvolution recovers x from y and h by running the convolution backwards: x[n] = (y[n] − Σ h[k]x[n − k] for k ≥ 1)/h[0], one sample at a time. Noise in y grows if h_inv is unstable.
An accumulator, h1[n] = u[n], is followed by a first difference, h2[n] = δ[n] − δ[n − 1]. The overall system is
h = u[n] * (δ[n] − δ[n − 1]) = u[n] − u[n − 1] = δ[n]. The difference undoes the running sum, so the cascade passes the input unchanged. Stability belongs to the overall h, not to each block.
Worked examples
Two first-order stages in series · h1(t) = e^(−2t)u(t), h2(t) = e^(−5t)u(t) connected in series
- h = h1 * h2 = ∫₀ᵗ e^(−2τ) e^(−5(t − τ)) dτ = (e^(−2t) − e^(−5t))/3 for t > 0
- Peak: 2e^(−2t) = 5e^(−5t), t = ln(5/2)/3 = 0.3054 s
- DC gain = ∫h dt = (1/2 − 1/5)/3 = 0.1, finite, so BIBO stable; h = 0 for t < 0, so causal
- Step response s(t) = (1 − e^(−2t))/6 − (1 − e^(−5t))/15; s(1) = 0.0779
Answer: h(t) = (e^(−2t) − e^(−5t))/3; causal and stable; the step response reaches 77.9 % of its final 0.1 after 1 s
Three systems, three verdicts · h_A[n] = 0.7ⁿu[n], h_B[n] = 0.7^|n| (all n), h_C[n] = 1.1ⁿu[n]
- A: zero for n < 0, so causal; Σ|h| = 1/(1 − 0.7) = 3.333, so stable
- B: non-zero for n < 0, so non-causal; Σ|h| = 1 + 2 × 0.7/0.3 = 5.667, so stable
- C: causal; the terms grow, the first 50 alone add up to 1164, so unstable
- None is memoryless: each h has more than one non-zero sample
Answer: A causal and stable, B stable but non-causal, C causal but unstable
Undoing an echo · h[n] = 1, 0.5 for n = 0, 1 (a half-strength echo one sample later); measured y[n] = 4, 0, 2, 1.5
- x[0] = y[0]/h[0] = 4
- x[1] = y[1] − 0.5x[0] = 0 − 2 = −2
- x[2] = y[2] − 0.5x[1] = 2 + 1 = 3, and x[3] = 1.5 − 0.5 × 3 = 0, so the input has ended
- The same thing as a system: h_inv[n] = (−0.5)ⁿu[n], Σ|h_inv| = 2, stable
Answer: x[n] = 4, −2, 3 for n = 0, 1, 2; check: 4, −2, 3 convolved with 1, 0.5 gives 4, 0, 2, 1.5
Differential and difference equations: natural, forced and total response, recursive and non-recursive systems
Which syllabus covers thisAU EC3354 Unit VVTU not in syllabusSPPU not in syllabusMU ECC404 Module 2KTU not in syllabus
What is a linear constant-coefficient difference equation?
It gives each output sample from a few past outputs and a few inputs, with fixed weights: y[n] = a1·y[n − 1] + a2·y[n − 2] + b0·x[n] + b1·x[n − 1]. The highest output delay is the order. It is the discrete twin of a differential equation such as RC·dy/dt + y = x, and like it, it needs initial conditions, here y[−1] and y[−2], before it has a unique answer. A computer solves it by simply running the recursion. In continuous time, RC = 0.1 s and a capacitor left at 3 V with no input give y = 3e^(−10t): 1.1 V after 0.1 s.
What are the natural, forced and total responses?
The natural (zero-input) response is what the stored initial conditions do with the input switched off. The forced (zero-state) response is what the input does to a system starting at rest. Because the system is linear, total = natural + forced. For y[n] = 0.5y[n − 1] + x[n] with y[−1] = 4 and a step input, the natural part is 2(0.5)ⁿ, the forced part 2 − 0.5ⁿ and the total 2 + 0.5ⁿ. Some books split the total instead into a homogeneous and a particular solution.
How do the characteristic roots shape the natural response?
Put y[n] = zⁿ into the equation with no input: z² − a1·z − a2 = 0. Two distinct real roots give c1·r1ⁿ + c2·r2ⁿ; a negative root makes its term flip sign every sample. A repeated root r gives (c1 + c2·n)rⁿ. A complex pair r·e^(±jθ) gives rⁿ·cos(θn + φ), a ringing that repeats every 2π/θ samples. The constants come from the initial conditions. The response dies out, and the system is stable, only if every root has |z| < 1.
What separates recursive (IIR) and non-recursive (FIR) systems?
A non-recursive system uses no past outputs: all a's are zero, the impulse response is just the list of b's, and it ends after M + 1 samples. That finite impulse response (FIR) is always stable. A recursive system feeds outputs back, so one impulse rings forever: y[n] = 0.9y[n − 1] + 0.1x[n] has h[n] = 0.1(0.9)ⁿ, an infinite impulse response (IIR) with DC gain 1. IIR does more with fewer coefficients, but it can go unstable.
y[n] = 1.2y[n − 1] − 0.36y[n − 2] + x[n]. With no input and non-zero initial conditions, the output
z² − 1.2z + 0.36 = (z − 0.6)², a repeated root (discriminant 0). A repeated root adds the factor n. |0.6| < 1, so it decays: stability depends on the roots, not on any single coefficient.
Worked examples
Natural, forced and total for a first-order system · y[n] = 0.5y[n − 1] + x[n], y[−1] = 4, x[n] = u[n]
- Natural (x = 0): y[n] = 0.5ⁿ⁺¹ × 4 = 2(0.5)ⁿ, so 2, 1, 0.5, ...
- Forced (y[−1] = 0): y[n] = Σ 0.5ᵏ for k = 0 to n = 2 − 0.5ⁿ, so 1, 1.5, 1.75, ...
- Total = 2 + 0.5ⁿ: 3, 2.5, 2.25, 2.125, ...
- Check by recursion: y[0] = 0.5 × 4 + 1 = 3, y[1] = 0.5 × 3 + 1 = 2.5
Answer: y[n] = 2 + 0.5ⁿ for n ≥ 0, settling at 2 (the DC gain 1/(1 − 0.5) times the step)
Impulse response from distinct roots · y[n] = 0.1y[n − 1] + 0.2y[n − 2] + x[n], system at rest, x[n] = δ[n]
- z² − 0.1z − 0.2 = 0 gives z = (0.1 ± 0.9)/2 = 0.5 and −0.4
- h[n] = c1(0.5)ⁿ + c2(−0.4)ⁿ; by recursion h[0] = 1, h[1] = 0.1
- c1 + c2 = 1 and 0.5c1 − 0.4c2 = 0.1 give c1 = 5/9 = 0.5556, c2 = 4/9 = 0.4444
- Check: h[2] = 0.1 × 0.1 + 0.2 × 1 = 0.21; formula (5/9)(0.25) + (4/9)(0.16) = 0.21
Answer: h[n] = (5/9)(0.5)ⁿ + (4/9)(−0.4)ⁿ, stable (both |z| < 1), Σh = DC gain = 1.4286
Complex roots and a sampled mains tone · y[n] = 1.2y[n − 1] − 0.72y[n − 2] + x[n], x[n] = cos(πn/6): a 50 Hz mains tone sampled at an illustrative 600 samples/s
- ω0 = 2π × 50/600 = π/6 rad/sample, 12 samples per cycle
- z² − 1.2z + 0.72 = 0 gives z = 0.6 ± 0.6j: |z| = √0.72 = 0.8485, angle π/4, so the natural part rings with an 8-sample period inside an envelope (0.849)ⁿ that drops below 1 % at n = 29
- H(e^jω0) = 1/(1 − 1.2e^(−jπ/6) + 0.72e^(−jπ/3)): |H| = 3.109, phase 4.2°
- Once the ringing has died out, y[n] = |H| cos(πn/6 + phase)
Answer: Steady output amplitude 3.11, phase 4°; stable, since both roots lie inside the unit circle
Fourier series: harmonics, trigonometric and exponential forms, and the Gibbs phenomenon
Which syllabus covers thisAU EC3354 Unit IIVTU not in syllabusSPPU 204191 Unit IIIMU ECC404 Module 3KTU PCECT402 Module 2
What does a Fourier series say about a periodic signal?
A signal that repeats every T seconds can be rebuilt from a constant plus sinusoids at whole-number multiples of its fundamental frequency f0 = 1/T, called harmonics. A waveform with a 2 ms period has a 500 Hz fundamental and harmonics at 1,000, 1,500, 2,000 Hz and so on, with nothing in between. The Dirichlet conditions guarantee this works: over one period the signal must be absolutely integrable and have a finite number of maxima, minima and jumps. Every practical waveform passes.
How do you find the trigonometric coefficients, and what do symmetries save you?
Multiply x(t) by the harmonic you want and average over one period. Every other harmonic averages to zero, because sines and cosines of different whole-number frequencies are orthogonal over a period. Symmetry cuts the work: an even signal has only cosine terms, an odd signal only sine terms, and half-wave symmetry, x(t + T/2) = −x(t), leaves only odd harmonics. A ±1 V square wave is odd and half-wave symmetric, so bn = 4/(nπ) for odd n: 1.273, 0.424, 0.255 V.
What is the exponential form, and how does it relate to an and bn?
The same series can be written as a sum of rotating phasors e^(jnω0t) for n from minus to plus infinity. Each pair cn e^(jnω0t) + c−n e^(−jnω0t) is one real sinusoid. The link is c0 = a0 and cn = (an − j bn)/2 for n > 0, and for a real signal c−n is the complex conjugate of cn. So |cn| is half the amplitude of harmonic n and the angle of cn is its phase. One integral gives every coefficient, which is why most derivations use this form.
Why does a truncated series ring at a jump, and does it ever stop?
Near a jump the partial sum overshoots, and adding harmonics only squeezes the ripple closer to the jump. The overshoot never shrinks: it settles at 8.95% of the jump, so a ±1 V square wave's sum peaks near 1.179 V. This is the Gibbs phenomenon. At the jump itself the series gives the midpoint. Smooth waveforms behave better: a triangle has no jump, its coefficients fall as 1/n² rather than 1/n, so its fifth harmonic is 4% of the fundamental against 20% for the square wave.
You keep adding harmonics to the Fourier series of a square wave. What happens to the overshoot next to each jump?
The peak of the partial sum settles at 8.95% of the jump above the top value (the Gibbs phenomenon). More terms make the ripple narrower, so the energy of the error falls to zero even though the peak does not.
Worked examples
A 0 to 5 V square wave · A clock-like square wave: 5 V for the first half of each period, 0 V for the second half, period 2 ms (illustrative)
- f0 = 1/T = 1/0.002 = 500 Hz
- a0 = average value = (5 × 1 ms)/2 ms = 2.5 V
- The wave minus 2.5 V is odd about t = 0 and half-wave symmetric, so an = 0 and only odd sine terms remain
- bn = (2/T) ∫ from 0 to T/2 of 5 sin(nω0t) dt = 10/(nπ) for odd n: b1 = 3.183 V, b3 = 1.061 V, b5 = 0.637 V
- Even harmonics: bn = (5/(nπ))(1 − cos nπ) = 0 for even n
Answer: x(t) = 2.5 + 3.183 sin(2π·500t) + 1.061 sin(2π·1500t) + 0.637 sin(2π·2500t) + … V
Full-wave rectified mains · 230 V rms, 50 Hz mains through a bridge rectifier into a resistor, no smoothing capacitor; diode drops ignored
- Peak Vp = 230√2 = 325.27 V; |sin| repeats every half cycle, so T = 10 ms and f0 = 100 Hz
- a0 = (1/T) ∫ Vp |sin(2π·50t)| dt = 2Vp/π = 207.07 V (the DC a moving-coil meter reads)
- The waveform is even about t = 0, so bn = 0; integrating gives an = −4Vp/(π(4n² − 1))
- n = 1 (100 Hz): 138.05 V; n = 2 (200 Hz): 27.61 V; n = 3 (300 Hz): 11.83 V
- Check: √(207.07² + 138.05²/2 + 27.61²/2 + 11.83²/2) = 229.9 V, against the true rms of 230 V
Answer: 207.1 V DC plus ripple at 100, 200, 300 Hz of 138.0, 27.6 and 11.8 V peak: no 50 Hz term at all
Exponential coefficients of a ramp · A sawtooth rises linearly from 0 to 4 V over T = 1 ms and drops back to 0 (illustrative)
- c0 = average = 2 V
- cn = (1/T) ∫ from 0 to T of (4t/T) e^(−j2πnt/T) dt. By parts, the leftover integral of e^(−j2πnt/T) over a whole period is zero, and the end term at t = T gives cn = 4/(−j2πn) = j·2/(nπ) V
- |c1| = 0.6366 V, |c2| = 0.3183 V, |c3| = 0.2122 V, each at phase +90°
- Trigonometric form: an = 2 Re cn = 0, bn = −2 Im cn = −4/(nπ), so b1 = −1.273 V
- At the jump (t = 0) the series gives the midpoint of 4 V and 0 V
Answer: x(t) = 2 − (4/π)[sin ω0t + ½ sin 2ω0t + ⅓ sin 3ω0t + …] V, ω0 = 2π × 1,000 rad/s; it converges to 2 V at each jump
Line spectra, symmetry and Parseval's theorem for periodic signals
Which syllabus covers thisAU EC3354 Unit IIVTU not in syllabusSPPU 204191 Unit IIIMU ECC404 Module 3KTU PCECT402 Module 2
What is a line spectrum, and why is it drawn on both sides of zero?
Plot |cn| and the angle of cn against frequency nf0 and you get the amplitude and phase spectra. They are lines, not curves, because a periodic signal contains only the harmonics. The exponential form needs negative frequencies too: each real sinusoid is two counter-rotating phasors at ±nf0, each carrying half its amplitude. For a real signal the amplitude spectrum is even and the phase spectrum is odd, so the left half adds no new information; the one-sided amplitude of harmonic n is 2|cn|.
What does the spectrum of a rectangular pulse train look like?
For pulses of height A, width τ and period T, cn = A·d·sinc(nd) with duty cycle d = τ/T and sinc(u) = sin(πu)/(πu). The period sets the line spacing 1/T. The pulse width sets the envelope, whose nulls fall at 1/τ, 2/τ and so on. Narrow pulses therefore spread their power over many lines. With d = 0.25 every fourth harmonic is missing, because it lands exactly on a null. Delaying the train by τ/2 leaves every |cn| alone and turns the phase of c1 by −45°.
What does Parseval's theorem say for a periodic signal?
The average power can be found in either domain: the mean of x² over a period equals the sum of |cn|² over all lines. Each line carries its own share, and the shares never interfere, because the harmonics are orthogonal. For a ±1 V square wave the fundamental alone carries 8/π² = 81.1% of the power. For a narrow pulse train, the lines inside the first null of the sinc envelope carry about 90% of the power, a handy rule for choosing a bandwidth.
How do symmetry and a time shift show up in the coefficients?
An even signal has real coefficients, so its phase is 0 or 180°. An odd signal has purely imaginary coefficients, at ±90°. Half-wave symmetry kills every even harmonic, which is why square and triangle waves have only odd lines. Delaying a signal by t0 multiplies cn by e^(−jnω0t0). The amplitude spectrum is unchanged and the phase falls in a straight line with n. Total harmonic distortion compares the rms of all the other harmonics with the fundamental's, so it uses the amplitudes alone.
A pulse train keeps its pulse width τ, but its period T is doubled. What happens to the line spectrum?
The line spacing is 1/T, so it halves. The envelope A·d·sinc(fτ) keeps its nulls at multiples of 1/τ, but d = τ/T halves (for example from 0.25 to 0.125), so every line is half as tall.
Worked examples
Coefficients of a pulse train · 3 V pulses, 0.25 ms wide, every 1 ms, centred on t = 0 (illustrative)
- d = τ/T = 0.25, f0 = 1 kHz, envelope nulls at multiples of 1/τ = 4 kHz
- c0 = A·d = 0.75 V (the average)
- c1 = 0.75 sinc(0.25) = 0.6752 V, c2 = 0.75 sinc(0.5) = 0.4775 V, c3 = 0.75 sinc(0.75) = 0.2251 V
- c4 = 0.75 sinc(1) = 0 (the 4 kHz line is missing); c5 = −0.1350 V, negative, so its phase is 180°
- The one-sided amplitude of the 1 kHz component is 2|c1| = 1.350 V
Answer: Lines every 1 kHz: 0.75, 0.675, 0.477, 0.225, 0 V for n = 0 to 4; real coefficients because the train is even
Parseval: how much power is in the first three harmonics? · The same 3 V, 0.25 ms, 1 ms pulse train, delivering power into 1 Ω
- Time domain: P = (1/T) ∫ x² dt = 3² × 0.25 ms / 1 ms = 2.25 W
- Lines |n| ≤ 3: c0² + 2(c1² + c2² + c3²) = 0.5625 + 2(0.4559 + 0.2280 + 0.0507) = 2.0317 W
- Share = 2.0317/2.25 = 90.3%
- These are all the lines inside the first envelope null at 4 kHz, so a 3 kHz bandwidth passes them
Answer: 90.3% of the 2.25 W lies in DC and the first three harmonics
Power and THD of a square-wave inverter · An illustrative inverter that switches between +12 V and −12 V at 50 Hz, into a 1 Ω resistive load
- Total power = 12² = 144 W (x² is 144 at every instant)
- Fundamental: b1 = 4 × 12/π = 15.279 V peak, power b1²/2 = 116.72 W, which is 81.1% of the total
- Third harmonic (150 Hz): b3 = b1/3 = 5.093 V peak
- THD = √(P − P1)/√P1 = √(144 − 116.72)/√116.72 = 48.3%
Answer: THD ≈ 48.3%: almost a fifth of the power (18.9%) is in harmonics, which is why sine-wave inverters filter or modulate their output
The Fourier transform: from series to spectrum, standard pairs and the time-bandwidth trade-off
Which syllabus covers thisAU EC3354 Unit IIVTU not in syllabusSPPU 204191 Unit IVMU ECC404 Module 3KTU PCECT402 Module 2
How does the Fourier series turn into the Fourier transform?
Take a pulse train and stretch its period while the pulse stays the same. The lines move closer together, 1/T apart, and each shrinks in proportion to 1/T. Multiply the coefficients by T and they settle onto one fixed curve: T·cn = X(n/T). In the limit of a single pulse, T → ∞, the lines merge into a continuous spectrum X(ω), a density in volt-seconds rather than volts. The interactives plot X against ordinary frequency f = ω/2π in hertz.
When does a signal have a Fourier transform?
A sufficient condition is that x(t) is absolutely integrable, with finitely many maxima, minima and jumps in any finite interval: these are the Dirichlet conditions again. Pulses and decaying exponentials qualify; e^(−at)u(t) needs a > 0. Signals that never die out, such as the unit step, a constant or a sinusoid, have transforms only if impulses are allowed: 1 ↔ 2πδ(ω), and cos ω0t ↔ π[δ(ω − ω0) + δ(ω + ω0)]. A periodic signal's transform is therefore a row of impulses of area 2πcn at nω0.
Which transform pairs should you know by heart?
A rectangular pulse of width τ gives τ·sinc(fτ), with nulls at multiples of 1/τ. A triangle of half-width τ gives τ·sinc², because it is two rectangles convolved, so its sidelobes fall faster. The one-sided exponential e^(−at)u(t) gives 1/(a + jω): magnitude 1/√(a² + ω²), phase −tan⁻¹(ω/a), 3 dB down and −45° at ω = a. A Gaussian transforms into a Gaussian. And δ(t) ↔ 1: an impulse contains every frequency equally.
Why can't a signal be both short and narrow-band?
Squeeze a pulse in time by a factor k and its spectrum stretches by k. So for a given shape, duration × bandwidth is a fixed number. With the half-height width and the one-sided 3 dB bandwidth used in the interactive, that product is 0.443 for a rectangle, 0.319 for a triangle, 0.312 for a Gaussian and 0.110 for e^(−t/τ)u(t). With root-mean-square widths there is a hard floor of 1/(4π) ≈ 0.080, and only the Gaussian reaches it. Short radar pulses and fast data bits need wide bandwidth for this reason.
A rectangular pulse is made 4 times shorter, at the same height. What happens to its spectrum?
X(f) = Aτ·sinc(fτ): the first null is at 1/τ, so it moves up 4 times, and X(0) = Aτ is the pulse area, which falls to a quarter. The spectrum gets wider and lower.
Worked examples
Spectrum of a rectangular pulse · A single 5 V pulse lasting 2 ms, centred on t = 0 (illustrative)
- X(f) = Aτ sinc(fτ) = 5 × 0.002 × sinc(0.002f) = 0.01 sinc(f/500) V·s
- X(0) = 10 mV·s, the area of the pulse
- First null at f = 1/τ = 500 Hz (ω = 2π/τ = 3142 rad/s)
- X(250 Hz) = 0.01 sinc(0.5) = 6.366 mV·s; X(750 Hz) = 0.01 sinc(1.5) = −2.122 mV·s, negative, phase 180°
Answer: A sinc with peak 10 mV·s and nulls every 500 Hz; the first sidelobe is negative
A decaying exponential: magnitude, phase and energy · x(t) = 10 e^(−200t) u(t) V, the voltage on a capacitor discharging with a time constant of 5 ms (illustrative)
- X(ω) = 10/(200 + jω) V·s
- ω = 0: |X| = 0.050 V·s, phase 0
- ω = 200 rad/s (31.83 Hz): |X| = 0.05/√2 = 0.0354 V·s, phase −45°
- ω = 1,000 rad/s: |X| = 10/√(200² + 1,000²) = 0.00981 V·s, phase −tan⁻¹(5) = −78.69°
- Energy = ∫ 100 e^(−400t) dt = 0.25 V²·s; the band |ω| < 200 rad/s holds (2/π) tan⁻¹(1) = 50% of it
Answer: |X| falls from 0.05 to 0.0098 V·s by 1,000 rad/s, the phase heads towards −90°, and half the 0.25 V²·s of energy lies below 200 rad/s
A triangle pulse as two rectangles convolved · A triangular pulse peaking at 2 V at t = 0 and reaching 0 at ±1 ms (illustrative)
- With τ = 1 ms: a unit-height rectangle of width τ convolved with itself is a triangle of half-width τ and peak τ, so x(t) = (2/τ) × (rect * rect)
- Each rectangle transforms to 1 ms × sinc(f × 1 ms); convolution in time multiplies the transforms
- X(f) = (2/0.001) × (0.001 sinc(0.001f))² = 0.002 sinc²(0.001f) V·s
- X(0) = 2 mV·s (area = ½ × 2 ms × 2 V); X(500 Hz) = 0.002 × (2/π)² = 0.8106 mV·s; first null at 1 kHz
Answer: X(f) = 2 sinc²(f/1000) mV·s: never negative, nulls every 1 kHz, sidelobes falling as 1/f² against 1/f for a rectangle
Fourier transform properties: shift, modulation, scaling, convolution and frequency response
Which syllabus covers thisAU EC3354 Unit IIIVTU not in syllabusSPPU 204191 Unit IVMU ECC404 Module 3KTU PCECT402 Module 4
What do a time shift and a time scaling do to the spectrum?
Delaying a signal by t0 multiplies its transform by e^(−jωt0). Every magnitude stays the same, and the phase gains a straight line of slope −t0, so a pure delay is invisible on a magnitude plot. Scaling time by a, x(at), compresses the signal when a > 1 and stretches its spectrum by the same factor, while the 1/|a| keeps the area right. Watch the order in x(2t − 3): write it as x(2(t − 1.5)), so the delay is 1.5, not 3.
What does multiplying by a cosine do?
Multiplying x(t) by cos ω0t splits its spectrum into two half-height copies centred at +ω0 and −ω0: x(t) cos ω0t ↔ ½X(ω − ω0) + ½X(ω + ω0). This is how a radio transmitter lifts an audio signal up to its carrier frequency, and how the receiver brings it back down by multiplying by the same cosine again and low-pass filtering. If the carrier frequency is below the signal's bandwidth, the two copies overlap and the original cannot be recovered cleanly. The dual property: a time shift multiplies by a phase factor, a frequency shift multiplies by e^(jω0t).
Why does convolution become multiplication?
An LTI system's output is the input convolved with the impulse response, y = x * h. In the frequency domain this becomes the much simpler Y(ω) = X(ω)H(ω), where H is the transform of h, the frequency response. An RC low-pass has h(t) = (1/RC)e^(−t/RC)u(t) and H = 1/(1 + jωRC). With RC = 0.15 s on a 1 s pulse (the interactive's default), the cut-off is 1.06 Hz. The edges round off with a 10-90% rise of 0.33 s, and the output keeps 85% of the pulse's energy.
What is the frequency response, and how is it measured?
Feed an LTI system a sinusoid and, once transients die away, the output is a sinusoid at the same frequency, scaled by |H(ω)| and shifted by the angle of H(ω). Sweeping the frequency and recording gain and phase therefore measures H directly. For a 1 kΩ, 1 µF RC low-pass, a 50 Hz input comes out at 0.954 of its amplitude, lagging by 17.4°. At the cut-off, 159.2 Hz, the gain is 1/√2 (−3 dB) and the lag is 45°.
Which other properties save work in exams?
Linearity: the transform of a sum is the sum of the transforms. Differentiation multiplies by jω, which boosts high frequencies; integration divides by jω and adds an impulse for the DC part. Duality: if x(t) ↔ X(ω), then X(t) ↔ 2πx(−ω), so a sinc pulse in time has a rectangular spectrum, the ideal low-pass filter's impulse response. Parseval: the energy ∫|x|² dt equals (1/2π)∫|X(ω)|² dω, so energy can be counted band by band.
A signal band-limited to 3 kHz is multiplied by cos(2π × 20,000 t). Where does its spectrum sit now (positive frequencies)?
Modulation moves half-height copies of the whole spectrum, −3 to +3 kHz, to ±20 kHz. The copy at +20 kHz spans 17 to 23 kHz: twice the original one-sided bandwidth (double sideband).
Worked examples
RC low-pass at 50 Hz and at 1 kHz · R = 1 kΩ, C = 1 µF, H(f) = 1/(1 + j2πfRC); 1 V sinusoidal inputs (illustrative)
- RC = 1 ms, cut-off fc = 1/(2πRC) = 159.15 Hz
- 50 Hz: 2πfRC = 0.3142, |H| = 1/√(1 + 0.3142²) = 0.9540, angle = −tan⁻¹(0.3142) = −17.44°
- The lag in time is 17.44°/(360° × 50 Hz) = 0.969 ms
- 1 kHz: 2πfRC = 6.2832, |H| = 0.1572, angle = −80.96°
- Check: a step-by-step (RK4) solution of RC·y′ + y = sin(2π·50t) settles to a peak of 0.9540 V
Answer: 50 Hz passes at 0.954 V lagging 17.4°; 1 kHz is cut to 0.157 V lagging 81.0°
A modulated pulse · A 1 V pulse lasting 2 ms (|t| < 1 ms) multiplied by cos(2π × 5,000 t) (illustrative)
- The pulse alone: X(f) = 0.002 sinc(0.002f) V·s, peak 2 mV·s, nulls every 500 Hz
- Modulation: Y(f) = ½X(f − 5,000) + ½X(f + 5,000)
- At 5 kHz: ½ × 2 mV·s + ½ × 0.002 sinc(20) = 1.000 mV·s (the far copy is at a null)
- Main lobe of the upper copy: 4500 to 5500 Hz, 1 kHz wide, the same width as the pulse's own main lobe from −500 to +500 Hz
Answer: Two half-height sinc copies, peaks of 1.0 mV·s at ±5 kHz, each with a 1 kHz-wide main lobe
Shift and scale together: x(2t − 3) · x(t) is a 1 V pulse on |t| < 1 s, so X(f) = 2 sinc(2f). Find the transform of y(t) = x(2t − 3)
- Rewrite: y(t) = x(2(t − 1.5)), a compression by 2 followed by a delay of 1.5 s
- Scaling: x(2t) ↔ ½X(f/2) = sinc(f)
- Delay by 1.5 s: multiply by e^(−j2πf × 1.5) = e^(−j3πf)
- So Y(f) = sinc(f) e^(−j3πf). Check in time: y is 1 V for 1 < t < 2, a 1 s pulse centred at 1.5 s
- At f = 0.25 Hz: |Y| = sinc(0.25) = 0.9003 V·s, angle = −3π × 0.25 = −135°
Answer: Y(f) = sinc(f) e^(−j3πf); delaying by 3 instead of 1.5 is the classic slip
The Laplace transform and its region of convergence
Which syllabus covers thisAU EC3354 Unit IIVTU not in syllabusSPPU 204191 Unit VMU ECC404 Module 4KTU PCECT402 Module 2
Why do we need the Laplace transform when the Fourier transform exists?
The Fourier transform needs a signal that dies away: e^(2t)u(t) grows without bound and has none. The Laplace transform first multiplies the signal by a decaying exponential e^(−σt) and then takes the Fourier transform. Pick σ large enough and almost anything converges. Writing s = σ + jω folds both steps into one kernel e^(−st). For e^(2t)u(t) the result is 1/(s − 2), valid for σ > 2: at s = 3 the integral of e^(2t)e^(−3t) from 0 to infinity is exactly 1.
What is the region of convergence, and why must it be stated with X(s)?
The ROC is the set of s for which the integral converges. It depends only on σ, so it is always a vertical strip, a half-plane or the whole plane. The ROC is not optional: e^(−3t)u(t) and −e^(−3t)u(−t) both give 1/(s + 3). The first converges for σ > −3, the second for σ < −3. Same formula, different signals. In the interactive, switch between right- and left-sided and watch only the shading move.
What rules does the ROC follow?
No pole is ever inside the ROC, because X(s) is infinite there. A right-sided signal has an ROC to the right of its rightmost pole; a left-sided signal, to the left of its leftmost pole; a two-sided signal, a strip between two poles, or nothing at all if the strip closes up. A finite-duration signal converges everywhere. For e^(−2t)u(t) + e^(3t)u(−t) the strip is −2 < σ < 3. Push the left part's exponent below −2 and the strip vanishes: there is no Laplace transform.
How do the Laplace and Fourier transforms connect, and what does the ROC say about stability?
On the line σ = 0, s = jω and the Laplace transform becomes the Fourier transform, but only if that jω axis lies inside the ROC. An LTI system is BIBO stable exactly when its impulse response is absolutely integrable, which is the same test: the ROC of H(s) contains the jω axis. A causal system has an ROC to the right of every pole, so a causal system is stable only when every pole sits in the left half-plane. The damped sine e^(−2t) sin 5t u(t) passes: poles at −2 ± j5.
x(t) = e^(−3t)u(t) + e^(2t)u(t). What is the region of convergence of X(s) = 1/(s + 3) + 1/(s − 2)?
Both parts are right-sided, so each needs σ to the right of its own pole: σ > −3 and σ > 2. Both must hold at once, which leaves σ > 2. The jω axis is outside, so this signal has no Fourier transform: the e^(2t) part grows.
Worked examples
A two-sided signal and its strip · x(t) = e^(−2t)u(t) + e^(3t)u(−t) (illustrative)
- Right part: ∫ from 0 to ∞ of e^(−2t)e^(−st) dt = 1/(s + 2), needs σ > −2
- Left part: ∫ from −∞ to 0 of e^(3t)e^(−st) dt = 1/(3 − s) = −1/(s − 3), needs σ < 3
- X(s) = 1/(s + 2) − 1/(s − 3) = 5/((s + 2)(3 − s)), ROC −2 < σ < 3
- Check inside the strip, s = 1: 1/3 + 1/2 = 0.8333; numerical integration of the definition gives the same
- The jω axis (σ = 0) is inside, so the Fourier transform exists: X(j2) = 5/((2 + j2)(3 − j2)), |X(j2)| = 5/(√8·√13) = 0.4903, phase −45° + 33.7° = −11.3°
Answer: X(s) = 5/((s + 2)(3 − s)), ROC −2 < σ < 3, Fourier transform exists, |X(j2)| = 0.490
One X(s), three signals · X(s) = 1/((s + 1)(s + 4)); list every signal it could be
- Partial fractions: X(s) = (1/3)/(s + 1) − (1/3)/(s + 4)
- Poles at −1 and −4 split the s-plane into three possible ROCs
- σ > −1: x(t) = (1/3)(e^(−t) − e^(−4t))u(t), causal and stable; check X(0) = 0.25 = ∫x dt
- −4 < σ < −1: x(t) = −(1/3)e^(−t)u(−t) − (1/3)e^(−4t)u(t), two-sided and unstable (the ROC misses σ = 0); X(−2) = −0.50
- σ < −4: x(t) = −(1/3)(e^(−t) − e^(−4t))u(−t), anticausal and unstable; X(−5) = 0.25
Answer: Three different signals; only the ROC σ > −1 gives a causal, stable one
A damped sine · x(t) = e^(−2t) sin(5t) u(t) (illustrative)
- sin(5t)u(t) ↔ 5/(s² + 25) for σ > 0
- Multiplying by e^(−2t) shifts s to s + 2: X(s) = 5/((s + 2)² + 25), ROC σ > −2
- Poles at −2 ± j5, both in the left half-plane, so the jω axis is in the ROC
- Area under x(t) = X(0) = 5/29 = 0.1724 (Simpson on the integral agrees)
- |X(j5)| = 5/|(2 + j5)² + 25| = 5/|4 + j20| = 0.2451
Answer: X(s) = 5/((s + 2)² + 25), σ > −2; area 0.1724; Fourier magnitude at ω = 5 rad/s is 0.245
Laplace transform properties and the inverse by partial fractions
Which syllabus covers thisAU EC3354 Unit IIVTU not in syllabusSPPU 204191 Unit VMU ECC404 Module 4KTU not in syllabus
Which Laplace properties do the most work?
Linearity lets you transform term by term. A delay of t0 multiplies by e^(−st0): u(t − 2) becomes e^(−2s)/s. Multiplying by e^(−at) shifts s to s + a, which is how damped sines get their (s + a)² + ω² denominators. Differentiation multiplies by s and, in the one-sided transform, subtracts the starting value x(0⁻), which is how initial conditions enter. Integration divides by s. Convolution in time becomes multiplication, so a system's output is simply Y(s) = H(s)X(s).
How can you read x(0+) and x(∞) without inverting?
The initial value theorem gives x(0+) as the limit of sX(s) as s goes to infinity; the final value theorem gives x(∞) as the limit of sX(s) as s goes to 0. For X(s) = 4(s + 2)/(s(s + 4)), the first gives 4 and the second 2, and indeed x(t) = 2 + 2e^(−4t). The final value theorem has a condition: every pole of sX(s) must lie in the left half-plane. For 1/(s − 1) it would give 0, while e^t runs off to infinity.
How does partial-fraction expansion invert a rational X(s)?
First make it proper: if the numerator's degree is not lower than the denominator's, divide, and the quotient gives impulses. Then split it into one simple fraction per pole. Each r/(s − p) inverts to r·e^(pt)u(t) for a causal signal, so a pole at −2 becomes an e^(−2t) term and its residue r sets the size. For a distinct pole, cover up its factor and substitute s = p: that is the residue. In the interactive each pole has its own colour and its term is drawn in the same colour.
What changes for repeated and complex poles?
A double pole at −a needs two fractions, B/(s + a)² and A/(s + a). B comes from the cover-up rule; A is the derivative of the covered-up expression at s = −a. The pair inverts to (A + Bt)e^(−at): the t factor is the signature of a repeated pole. Complex poles come in conjugate pairs with conjugate residues r and r*, and their two terms add to the real 2|r|e^(σt)cos(ωt + ∠r). Completing the square gives the same answer as cosine and sine terms.
In the interactive, H(s) = 1/((s + 1)(s + 2)(s + 4)). At t = 0 its three exponential terms add up to
The residues are 1/3, −1/2 and 1/6, which sum to zero. That is no accident: the denominator is three degrees above the numerator, so h(0+) = lim sH(s) = 0 (initial value theorem) and h′(0+) = 0 too. The response starts flat and rises.
Worked examples
Three distinct poles · X(s) = (s + 6)/((s + 1)(s + 2)(s + 4)), causal (illustrative)
- X(s) = A/(s + 1) + B/(s + 2) + C/(s + 4)
- A = (s + 6)/((s + 2)(s + 4)) at s = −1 = 5/(1 × 3) = 5/3
- B = (s + 6)/((s + 1)(s + 4)) at s = −2 = 4/(−1 × 2) = −2
- C = (s + 6)/((s + 1)(s + 2)) at s = −4 = 2/(−3 × −2) = 1/3
- x(t) = [(5/3)e^(−t) − 2e^(−2t) + (1/3)e^(−4t)]u(t); x(0) = 5/3 − 2 + 1/3 = 0, x(1) = 0.3486
Answer: x(t) = (5/3)e^(−t) − 2e^(−2t) + (1/3)e^(−4t) for t ≥ 0; x(1 s) = 0.3486
A repeated pole · X(s) = (s + 3)/((s + 1)²(s + 4)), causal (illustrative)
- X(s) = A/(s + 1) + B/(s + 1)² + C/(s + 4)
- B = (s + 3)/(s + 4) at s = −1 = 2/3
- A = d/ds[(s + 3)/(s + 4)] at s = −1 = 1/(s + 4)² = 1/9
- C = (s + 3)/(s + 1)² at s = −4 = −1/9
- x(t) = [(1/9)e^(−t) + (2/3)t e^(−t) − (1/9)e^(−4t)]u(t); x(1) = 0.2841
Answer: x(t) = (1/9 + 2t/3)e^(−t) − (1/9)e^(−4t) for t ≥ 0; x(1 s) = 0.2841
Total response with initial conditions · y″ + 2y′ + 10y = 10u(t), y(0⁻) = 0, y′(0⁻) = 2 (illustrative)
- One-sided transform: s²Y − 0·s − 2 + 2(sY − 0) + 10Y = 10/s
- Y(s) = (2s + 10)/(s(s² + 2s + 10)); the quadratic has roots −1 ± j3
- Residue at s = 0: 10/10 = 1; the remainder is −s/((s + 1)² + 9) = −(s + 1)/((s + 1)² + 9) + (1/3)·3/((s + 1)² + 9)
- y(t) = 1 − e^(−t)cos 3t + (1/3)e^(−t)sin 3t for t ≥ 0
- Checks: y(0) = 0, y′(0) = 1 + 1 = 2, y(∞) = 1 (final value: sY at s = 0 is 10/10); y(1) = 1.3815, first peak 1.435 at t = 0.83 s
Answer: y(t) = 1 − e^(−t)(cos 3t − (1/3)sin 3t); it rings at 3 rad/s and settles at 1, with y(1 s) = 1.382
Analysing CT LTI systems in the s-domain: transfer function, poles, zeros and frequency response
Which syllabus covers thisAU EC3354 Unit IIIVTU not in syllabusSPPU 204191 Unit VMU ECC404 Module 4KTU PCECT402 Module 4
What is a transfer function, and how do you get one from a differential equation?
The transfer function H(s) is the Laplace transform of the impulse response, and equally the ratio Y(s)/X(s) with every initial condition set to zero. From a differential equation it is one line: replace each derivative by a power of s. For y″ + 5y′ + 6y = 2x′ + 3x, H(s) = (2s + 3)/(s² + 5s + 6) = (2s + 3)/((s + 2)(s + 3)). The roots of the numerator are zeros (here −1.5), those of the denominator are poles (−2 and −3). Systems in series multiply their H(s); in parallel they add.
How do the pole positions decide causality and stability?
Each pole p contributes a term like e^(pt) to h(t). For a causal system, poles in the left half-plane give decaying terms, so the system is stable; one pole in the right half-plane gives a growing term. Poles on the jω axis are the borderline: 1/(s² + 16) rings at 4 rad/s forever, and feeding it sin 4t, a bounded input, gives an output that grows like t. If you drop causality, a right-half-plane pole can still be stable, with an ROC strip that holds the jω axis and an h(t) that runs backwards in time.
How can you read |H(jω)| and its phase straight off the s-plane?
Factor H(s) = K(s − z1)…/((s − p1)(s − p2)…). At s = jω each factor jω − p is an arrow from the pole p to the point jω on the axis. So |H(jω)| is |K| times the product of zero-arrow lengths divided by the product of pole-arrow lengths, and the phase is the sum of zero-arrow angles minus the sum of pole-arrow angles. Slide ω past a pole close to the axis and its arrow becomes short: a peak. With ζ = 0.3 and ωn = 10 rad/s the peak is 1.747 at 9.06 rad/s.
What do zeros do, and what is a non-minimum-phase system?
A zero at the origin blocks DC and turns a low-pass into a band-pass; a zero on the jω axis makes a perfect notch. Mirror a left-half-plane zero at −10 to +10 and every zero-arrow keeps its length, so the magnitude response is identical. The angles are not: the right-half-plane version lags up to 180° more, and its step response starts off in the wrong direction. Systems with every zero in the left half-plane have the least phase lag for their magnitude: minimum phase.
In the interactive, move the zero from s = −10 to its mirror image s = +10 (poles unchanged, DC gain kept at 1). The magnitude response
The arrow from +10 to jω has the same length as the arrow from −10 to jω, √(100 + ω²), at every frequency, so |H(jω)| cannot change. Only the phase does: the right-half-plane zero adds lag instead of lead, up to 180° more at high frequency.
Worked examples
Reading a response off the arrows · H(s) = 25/(s² + 6s + 25); find |H| and the phase at ω = 4 rad/s, then the resonant peak (illustrative)
- Poles: s = −3 ± j4, so ωn = 5 rad/s and ζ = 3/5 = 0.6
- Arrow from −3 + j4 to j4: (3, 0), length 3, angle 0°
- Arrow from −3 − j4 to j4: (3, 8), length √73 = 8.544, angle 69.4°
- |H(j4)| = 25/(3 × 8.544) = 0.9753; phase = −(0° + 69.4°) = −69.4°
- ζ < 0.707, so there is a peak: ωr = 5√(1 − 2 × 0.36) = 2.646 rad/s, Mr = 1/(2 × 0.6 × 0.8) = 1.0417
Answer: |H(j4)| = 0.975 at −69.4°; peak 1.042 at 2.65 rad/s
Causal or stable, not both · H(s) = (s + 3)/((s + 1)(s − 2)) (illustrative)
- Residues: at s = −1, (s + 3)/(s − 2) = 2/(−3) = −2/3; at s = 2, (s + 3)/(s + 1) = 5/3
- Causal choice, ROC σ > 2: h(t) = [−(2/3)e^(−t) + (5/3)e^(2t)]u(t), which grows: unstable
- Stable choice, ROC −1 < σ < 2 (holds the jω axis): h(t) = −(2/3)e^(−t)u(t) − (5/3)e^(2t)u(−t)
- The stable h is absolutely integrable: ∫|h| dt = 2/3 + 5/6 = 1.5, and ∫h dt = H(0) = 3/(−2) = −1.5
- No ROC is both to the right of s = 2 and around the jω axis, so no causal stable system has this H(s)
Answer: The right-half-plane pole forces a choice: causal and unstable, or stable and non-causal
Mains hum through a sensor filter · A second-order low-pass on a sensor signal: fn = 5 Hz, ζ = 0.707 (illustrative); India's mains frequency is 50 Hz
- ωn = 2π × 5 = 31.42 rad/s; poles at −22.21 ± j22.21
- At 50 Hz, jω = j314.2: arrow lengths √(22.21² + (314.2 − 22.21)²) = 292.8 and √(22.21² + (314.2 + 22.21)²) = 337.1
- |H| = ωn²/(product) = 987.0/(292.8 × 337.1) = 0.01000, that is −40.0 dB
- Check with r = f/fn = 10: |H| = 1/√((1 − r²)² + (2ζr)²) = 1/√(1 + r⁴) = 0.01000
- At 5 Hz itself |H| = 0.7071 (−3 dB); phase at 50 Hz −171.9°
Answer: 50 Hz hum is cut to 1.00% of its size (−40 dB): a decade above fn, two poles give 40 dB
Correlation, energy spectral density and power spectral density
Which syllabus covers thisAU not in syllabusVTU not in syllabusSPPU not in syllabusMU ECC404 Module 2KTU not in syllabus
What does cross-correlation measure, and how is it different from convolution?
Cross-correlation slides one signal past another and, at each shift τ, multiplies and integrates: R_xy(τ) = ∫x(t + τ)y(t)dt. A large value means the two line up well at that shift. Convolution does the same sliding but flips one signal first, so correlating with y is convolving with y(−t). The discrete version swaps the integral for a sum. Correlating [1, 2] with itself gives 2, 5, 2 for shifts −1, 0, 1: the middle value is the energy.
How does correlation find an echo's delay in noise, and why use a Barker code?
Correlate the received record with the transmitted code. Where the buried echo lines up, all 13 products add in step and the sum peaks; noise products add at random and mostly cancel. The lag of the peak is the delay. A plain block of thirteen 1s has a broad triangular correlation, so noise easily moves the peak sideways. The Barker-13 code has autocorrelation sidelobes of at most 1 against a peak of 13 (−22.3 dB), so its peak is a single sharp spike.
What properties does autocorrelation always have?
Autocorrelation R_x(τ) = R_xx(τ) is even, R(−τ) = R(τ), and never larger than its value at zero shift. For an energy signal R(0) is the total energy; for a power signal, with a time average, it is the average power. A periodic signal has a periodic autocorrelation with the same period: 3cos(2π·50t) gives R(τ) = 4.5cos(2π·50τ), so R(0) = 4.5 is its power. Random noise decorrelates quickly, so R falls off fast away from zero.
What are energy and power spectral density, and how do they link to autocorrelation?
Energy spectral density |X(f)|² says where an energy signal's energy sits in frequency; its integral over f is the total energy (Parseval). For power signals the power spectral density S_x(f) plays the same role for average power. The Wiener-Khinchin theorem ties both to correlation: the PSD is the Fourier transform of the autocorrelation. A 50 Hz cosine of power 4.5 has two PSD lines of 2.25 at ±50 Hz. White noise has a flat PSD and an autocorrelation that is a single impulse at zero.
In the interactive, swap the Barker-13 code for a block of 13 ones (same energy, 13). The correlation peak
The peak height is the code energy, 13 for both. But the ones code still matches well when shifted by one or two samples (12, 11, …), so the peak is a wide hill; Barker-13 drops to 0 or 1 one sample away, a needle that noise cannot easily move.
Worked examples
Autocorrelation of a short sequence · x[n] = [2, −1, 3] for n = 0, 1, 2 (illustrative)
- r[k] = Σ x[n + k] x[n]
- k = 0: 2² + (−1)² + 3² = 14, the energy
- k = 1: x[1]x[0] + x[2]x[1] = (−1)(2) + (3)(−1) = −5
- k = 2: x[2]x[0] = 3 × 2 = 6
- Even symmetry gives r[−1] = −5, r[−2] = 6; the length is 2 × 3 − 1 = 5
Answer: r = [6, −5, 14, −5, 6] for k = −2 … 2; the largest value 14 sits at k = 0 and equals the energy
Range from a correlation peak · A radar sends Barker-13 with 1 µs chips; the echo is sampled at 1 MHz and the correlation peak falls at lag 40 (illustrative); c = 299,792,458 m/s
- Round-trip delay τ = 40/1,000,000 s = 40 µs
- The pulse goes out and back, so range = cτ/2 = 299,792,458 × 40 × 10⁻⁶/2 = 5,996 m
- One lag step is 1 µs, worth c × 1 µs/2 = 149.9 m of range
- Peak sidelobe of the code: 1 against 13, 20 log10(1/13) = −22.3 dB, so a second echo up to 22 dB weaker than the first still stands above the first echo's sidelobes
Answer: Target at about 6.00 km, measured in steps of 150 m
Energy spectral density of a decaying exponential · x(t) = e^(−2t)u(t) (illustrative)
- Energy E = ∫ e^(−4t) dt from 0 to ∞ = 1/4 = 0.25
- X(f) = 1/(2 + j2πf), so the ESD is |X(f)|² = 1/(4 + 4π²f²)
- Autocorrelation R(τ) = ∫x(t + τ)x(t)dt = (1/4)e^(−2|τ|); its Fourier transform is 1/(4 + 4π²f²), the ESD again (Wiener-Khinchin)
- Energy in |f| < 1 Hz = ∫ from −1 to 1 of df/(4 + 4π²f²) = atan(π)/(2π) = 0.2010
- Share = 0.2010/0.25 = 0.8038
Answer: 80.4% of the energy lies below 1 Hz; the ESD falls to half its peak at f = 1/π = 0.318 Hz
Sampling: the sampling theorem, aliasing and reconstruction
Which syllabus covers thisAU EC3354 Unit IVVTU not in syllabusSPPU not in syllabusMU not in syllabusKTU PCECT402 Module 3
What does the sampling theorem promise, and what is the Nyquist rate?
If a signal has nothing above a highest frequency fmax, its samples fix it completely, provided the sampling rate fs is more than 2 fmax. The rate 2 fmax is the Nyquist rate; half the sampling rate, fs/2, is the folding frequency. A telephone call keeps roughly 300 to 3,400 Hz of speech and is sampled at 8 kHz, so the folding frequency of 4 kHz sits 600 Hz above the top of the voice band. The theorem says nothing about how to rebuild the signal cheaply: that needs filters, and real filters need room.
Why does a fast sinusoid pretend to be a slow one?
Samples of cos(2πfn/fs) do not change if f moves by any whole multiple of fs, and f and fs − f give the same samples with the phase reversed. So every input frequency lands on one alias between 0 and fs/2. At fs = 1,000 Hz, a 700 Hz tone and a 1,300 Hz tone both give exactly the samples of a 300 Hz tone: the 700 Hz one folds back from above 500 Hz and flips its phase, the 1,300 Hz one does not. After sampling, nothing can tell them apart.
What does sampling do to the spectrum?
Sampling is multiplication by a train of impulses every Ts = 1/fs, and that copies the whole spectrum to every multiple of fs (scaled by fs). If the copies do not overlap, a low-pass filter picks out the original. For the telephone band the first copy around 8 kHz runs from 4,600 to 11,400 Hz, so a gap of 1,200 Hz separates it from the 3,400 Hz top of the voice band. If fs is below 2 fmax the copies overlap, the tails fold into the band and no filter can separate them again.
How are samples turned back into a signal, and why filter before the ADC?
The ideal rebuild puts a sinc pulse on every sample, which is an ideal low-pass filter at fs/2: perfect but infinitely long. A real DAC holds each sample for Ts (a zero-order hold) and a smoothing filter follows; the hold itself droops, to 2/π = 0.637 (−3.92 dB) at fs/2. Before the ADC an analogue anti-aliasing filter must remove everything above fs/2, because aliasing cannot be undone afterwards. CD audio samples at 44.1 kHz for hearing up to 20 kHz, leaving that filter a 2.05 kHz transition band.
A 900 Hz tone is sampled at 1,000 Hz and played back through an ideal reconstruction filter. What comes out?
900 Hz is above fs/2 = 500 Hz, so it folds to 1,000 − 900 = 100 Hz (with its phase reversed). The reconstruction filter passes only 0 to 500 Hz, and the only tone there that fits the samples is 100 Hz.
Worked examples
Nyquist rate of a product of tones · x(t) = 2cos(2π·400t)·cos(2π·1,100t) + sin(2π·250t); then sampled at 2,000 Hz
- 2cos A cos B = cos(A − B) + cos(A + B), so the product is cos(2π·700t) + cos(2π·1,500t)
- Frequencies present: 250, 700 and 1,500 Hz, so fmax = 1,500 Hz
- Nyquist rate = 2 × 1,500 = 3,000 Hz: sample faster than this to keep every tone
- At 2,000 Hz the folding frequency is 1,000 Hz: 250 and 700 Hz survive, 1,500 Hz folds to 2,000 − 1,500 = 500 Hz
Answer: Nyquist rate 3,000 Hz; at 2 kHz the 1,500 Hz term masquerades as 500 Hz
Mains harmonics in an ECG sampled at 360 Hz · ECG digitised at 360 samples/s (the MIT-BIH Arrhythmia Database rate); interference at 50 Hz (Indian mains) and its harmonics leaks past a weak front-end filter
- Folding frequency = 360/2 = 180 Hz; harmonics at 50, 100 and 150 Hz are below it and stay put
- 200 Hz → 360 − 200 = 160 Hz; 250 Hz → 110 Hz; 300 Hz → 60 Hz (all folded)
- 350 Hz → 360 − 350 = 10 Hz; 400 Hz → 400 − 360 = 40 Hz
- A 50 Hz notch filter applied after sampling removes only the 50 Hz line, not these aliases
Answer: The 7th harmonic (350 Hz) appears as a 10 Hz ripple; the fix is an analogue low-pass before the ADC, not a digital filter after it
A whistle on a telephone line · Telephone channel sampled at 8 kHz, voice band about 300-3,400 Hz; a 5.2 kHz test whistle reaches the sampler with no anti-aliasing filter (illustrative)
- Folding frequency = 4 kHz; 5.2 kHz is above it
- Alias = 8,000 − 5,200 = 2,800 Hz, with the phase reversed
- 2,800 Hz lies inside the 300-3,400 Hz band, so the receiving filter passes it
- A 3.9 kHz tone would not alias (it is below 4 kHz); it would simply sit above the voice band
Answer: The listener hears a 2,800 Hz tone that was never sent
The discrete-time Fourier transform and the frequency response of DT systems
Which syllabus covers thisAU EC3354 Unit IVVTU not in syllabusSPPU not in syllabusMU ECC404 Module 3KTU PCECT402 Module 3
What is the DTFT, and what does ω mean for a sequence?
The DTFT weighs a sequence against a complex sinusoid e^(jωn) at every frequency ω and adds up: X(e^jω) = Σ x[n] e^(−jωn). The sequence is discrete, but X is a continuous function of ω. ω is in radians per sample, linked to hertz by ω = 2πf/fs. For an ECG sampled at 360 Hz, 50 Hz mains sits at ω = 2π × 50/360 = 0.873 rad/sample, or 0.278π. The sum converges when x is absolutely summable.
Why does the DTFT repeat every 2π, and why is ω = π the top frequency?
For whole numbers n, e^(−j(ω + 2π)n) equals e^(−jωn), so X repeats every 2π. These are the sampling replicas again: ω = 2π is fs. A sequence at ω = 1.8π has the same samples as one at −0.2π. So only −π to π matters, and for real x the magnitude is even and the phase odd, so 0 to π says everything. ω = π is (−1)^n, the fastest a sequence can alternate: it is fs/2.
What is the frequency response of a DT system, and how is it used?
For an LTI system, H(e^jω) is the DTFT of the impulse response h[n]. An input A cos(ω0 n) comes out as the same sinusoid, scaled by |H| and shifted by the phase of H, once the start-up transient has died. A 4-point moving average fed cos(πn/4) settles to 0.653 cos(πn/4 − 67.5°). Every LTI system works one frequency at a time like this, which is why a filter is specified by its frequency response.
How do a moving average and a first-order IIR smoother differ?
An M-point moving average has |H| = |sin(Mω/2)/(M sin(ω/2))|, which is exactly zero at ω = 2πk/M, and a straight-line phase −(M − 1)ω/2: every frequency is delayed by the same (M − 1)/2 samples. The IIR smoother y[n] = a·y[n − 1] + (1 − a)·x[n] needs one multiply-add per sample whatever its smoothing; its gain is 1 at DC and (1 − a)/(1 + a) at π (0.0526 for a = 0.9), it has no exact zeros and its phase is not linear.
Which DTFT properties are used most?
A delay of n0 samples multiplies X by e^(−jωn0): magnitude unchanged, phase tilted. Convolution becomes multiplication, so cascaded systems multiply their responses. Parseval: the energy Σ|x[n]|² equals the area under |X|² over one period divided by 2π. For x = [1, 2, 1], X = e^(−jω)(2 + 2cos ω); the energy is 6, and integrating |X|² from −π to π and dividing by 2π also gives 6.0000.
A 5-point moving average gets the input cos(0.4πn). Once the first four samples have passed, the output is
0.4π = 2π/5 is a zero of the 5-point average: any 5 consecutive samples of this cosine cover exactly one period and add to zero. The interactive shows the dip to zero at ω = 0.4π with M = 5.
Worked examples
DTFT of a decaying sequence · x[n] = 0.75ⁿ u[n]
- X(e^jω) = Σ (0.75 e^(−jω))ⁿ for n ≥ 0 = 1/(1 − 0.75 e^(−jω)), a geometric series that converges since 0.75 < 1
- ω = 0: X = 1/0.25 = 4; ω = π: X = 1/1.75 = 0.5714
- ω = π/2: 1 − 0.75(−j) = 1 + 0.75j, so |X| = 1/1.25 = 0.80 and the phase is −atan(0.75) = −36.87°
- Energy Σ 0.75^(2n) = 1/(1 − 0.5625) = 2.2857; the Parseval integral of |X|² over −π..π divided by 2π gives 2.2857
Answer: A low-pass shape: |X| falls from 4 at DC to 0.571 at ω = π
Removing 50 Hz hum with a moving average · A slow 5 Hz sensor signal logged at 200 samples/s (illustrative) with 50 Hz mains hum; 4-point moving average
- 50 Hz → ω = 2π × 50/200 = π/2; 100 Hz → ω = π; 5 Hz → ω = 2π × 5/200 = π/20
- The 4-point average has zeros at ω = 2πk/4 = π/2, π, 3π/2: the 50 Hz hum and its 100 Hz harmonic are removed exactly
- At 5 Hz: |H| = sin(4 × π/40)/(4 sin(π/40)) = 0.3090/0.3138 = 0.9846
- Delay = (4 − 1)/2 = 1.5 samples = 7.5 ms at 200 samples/s, the same for every frequency
Answer: Hum gone, the 5 Hz signal kept at 98.5% amplitude and delayed 7.5 ms
First-order IIR smoother · y[n] = 0.9 y[n − 1] + 0.1 x[n], input x[n] = 2cos(0.1πn)
- H(e^jω) = 0.1/(1 − 0.9 e^(−jω)); at ω = 0.1π: |1 − 0.9 e^(−j0.1π)| = 0.3132
- |H| = 0.1/0.3132 = 0.3193, phase = −62.6°
- Steady-state output = 2 × 0.3193 cos(0.1πn − 62.6°) = 0.639 cos(0.1πn − 62.6°); running the recursion for 600 samples agrees
- Half-power frequency: cos ωc = 1 − (0.1)²/(2 × 0.9), ωc = 0.1055 rad/sample = 0.0336π, so 0.1π is about 3.0 times the cut-off
Answer: Output amplitude 0.639, lagging by 62.6°
The discrete Fourier series and the DFT: bins, zero padding, leakage and windows
Which syllabus covers thisAU EC3354 Unit VVTU not in syllabusSPPU not in syllabusMU not in syllabusKTU PCECT402 Module 3
What is the discrete-time Fourier series?
A sequence that repeats every N samples needs only N harmonics e^(j2πkn/N), k = 0 to N − 1, because harmonic k + N is the same sequence as harmonic k. Each coefficient is a finite sum, ck = (1/N) Σ x[n] e^(−j2πkn/N), so there is no convergence question and no Gibbs overshoot. The period-4 sequence 3, 1, −1, 1 has c0 = 1, c1 = c3 = 1 and c2 = 0, so it is exactly 1 + 2cos(πn/2).
How is the DFT related to the DTFT and the DTFS?
Take a block of N samples. Its DFT X[k] is the block's DTFT sampled at N equally spaced frequencies ωk = 2πk/N, and it is also N times the DTFS coefficients of the block repeated forever. Bin k stands for the frequency k fs/N in hertz. Because the DFT treats the block as one period of a periodic signal, whatever happens at the seam between the end and the start of the block shows up in the spectrum.
What sets frequency resolution, and what does zero padding do?
Bin spacing is fs/N = 1/T, the reciprocal of the record length. Two seconds of an ECG at 360 Hz gives N = 720 and 0.5 Hz bins, and 50 Hz lands exactly on bin 100. Zero padding appends zeros before the DFT: the bins get closer, but they are samples of the same DTFT, so the curve looks smoother while its peaks stay just as wide. Two tones 0.6 bin apart in a 64-sample record still merge into one lobe after padding to 512. Only a longer record separates them.
Why does a pure tone spread across many bins, and how does a window help?
If the block holds a whole number of cycles, the tone sits on one bin. Otherwise the repeated block has a jump at the seam and energy leaks into every bin. For a tone half-way between bins of a 32-point DFT, the nearest bins read 3.87 dB low and 9.4% of the energy lands more than 2 bins away; the highest sidelobe is −13.2 dB. A Hann window tapers the ends to zero: sidelobes drop to −31.5 dB and the far leakage to 0.09%, at the price of a main lobe twice as wide.
Why does everyone compute the DFT with an FFT?
The direct DFT needs about N² complex multiplications. A radix-2 FFT splits the block into even and odd samples again and again, reusing the same twiddle factors, and needs about (N/2) log2 N. For N = 1,024 that is 5,120 instead of 10,48,576, about 205 times fewer. The answer is the same DFT, not an approximation: the speed-up is pure bookkeeping, which is why N is usually chosen as a power of two.
A 64-sample record has two tones 0.6 bin apart. You zero-pad it to 512 points before the DFT. The two tones
Zero padding samples the same DTFT more finely. The main lobe width is set by the 64-sample record (2 bins wide null to null for a rectangular window), so tones 0.6 bin apart merge. The same record with tones 2.4 bins apart shows two peaks.
Worked examples
A 4-point DFT by hand · x[n] = 2, −1, 0, 3 for n = 0 to 3
- With N = 4, e^(−j2πkn/4) = (−j)^(kn)
- X[0] = 2 − 1 + 0 + 3 = 4
- X[1] = 2 + (−1)(−j) + 0 + 3(j) = 2 + 4j; X[2] = 2 + 1 + 0 − 3 = 0
- X[3] = 2 + (−1)(j) + 0 + 3(−j) = 2 − 4j, the conjugate of X[1] as it must be for real x
- Parseval: Σ|x|² = 4 + 1 + 0 + 9 = 14; (1/4)(16 + 20 + 0 + 20) = 14
Answer: X = [4, 2 + 4j, 0, 2 − 4j]
Choosing N so mains hum lands on a bin · ECG sampled at 360 Hz (the MIT-BIH rate), 50 Hz mains interference; compare N = 1,024 and N = 720
- N = 1,024: bin spacing = 360/1,024 = 0.3516 Hz, and 50 Hz falls at k = 50/0.3516 = 142.22
- The nearest bins are 142 (49.92 Hz) and 143 (50.27 Hz); bin 142 reads 0.71 dB low and 4.2% of the hum's energy leaks more than 2 bins away
- N = 720 (2 s): spacing = 0.5 Hz, 50 Hz falls exactly on bin 100, every other bin is zero
- A non-power-of-two N still has fast algorithms (720 = 2⁴ × 3² × 5), or keep 1,024 and apply a Hann window
Answer: With N = 720 the hum is one clean bin; with N = 1,024 it smears around bin 142.2
Leakage, rectangular against Hann · N = 32, a tone at 8.5 bins (8.5 cycles in the block), no zero padding
- Rectangular: the peak bin reads 3.87 dB below an on-bin tone; 9.4% of the energy sits more than 2 bins away; highest sidelobe −13.2 dB
- Hann, w[n] = 0.5 − 0.5cos(2πn/32): peak loss 1.42 dB, far leakage 0.088%, highest sidelobe −31.5 dB
- An on-bin Hann tone fills 3 bins in the ratio 1/2 : 1 : 1/2, so the main lobe is 4 bins wide null to null against 2 for rectangular
- For very long blocks the half-bin losses tend to 20 log10(2/π) = −3.92 dB and 20 log10(8/(3π)) = −1.42 dB
Answer: Hann cuts far leakage by about 107 times and sidelobes by 18 dB, at the cost of a wider peak
The z-transform: region of convergence, standard pairs, properties and the inverse
Which syllabus covers thisAU EC3354 Unit IVVTU BEC402 Module 5SPPU not in syllabusMU ECC404 Module 5KTU PCECT402 Module 3
What is the z-transform, and how is it related to the DTFT?
X(z) = Σ x[n] z⁻ⁿ, summed over all n, with z a complex number. Write z = r e^(jω): the sum becomes the DTFT of x[n] r⁻ⁿ. So the z-transform is a DTFT with an extra exponential weight, and on the unit circle (r = 1), if the ROC reaches it, it is the DTFT itself. The weight is the point. 1.1ⁿu[n] grows, so it has no DTFT, but on |z| = 1.5 its terms (1.1/1.5)ⁿ shrink and the sum is 1/(1 − 0.733) = 3.75. Convolution then becomes multiplication, and difference equations become algebra.
Why must every z-transform come with its region of convergence?
Because different sequences share one formula. 0.8ⁿu[n] gives Σ (0.8/z)ⁿ = z/(z − 0.8), which converges only for |z| > 0.8. The left-sided −0.8ⁿu[−n−1] gives the same z/(z − 0.8), but converges only for |z| < 0.8. The formula is identical; the ROC tells them apart. Check at z = 2: Σ 0.4ⁿ = 1.6667 = 2/(2 − 0.8). In the interactive, put the test circle inside and outside the shaded ROC and watch the terms of the sum die away or blow up.
What shapes can a region of convergence take?
Convergence depends only on |z|, so the ROC is a ring centred on the origin, and it never contains a pole. A finite-length sequence converges everywhere except perhaps z = 0 or z = ∞. A right-sided sequence converges outside its outermost pole, a left-sided one inside its innermost pole, and a two-sided one in the ring between, if a ring exists. With poles at 0.5 and 2 there are three choices: |z| < 0.5 (left-sided), 0.5 < |z| < 2 (two-sided, stable) and |z| > 2 (causal, unstable). Stable means the unit circle lies in the ROC.
Which z-transform properties do the most work in problems?
Linearity. Time shift: x[n − k] ↔ z⁻ᵏX(z), so one sample of delay is a factor z⁻¹. Scaling: aⁿx[n] ↔ X(z/a), which moves the poles out by a factor |a|. Differentiation: n x[n] ↔ −z dX/dz, so n 0.5ⁿu[n] ↔ 0.5z/(z − 0.5)², and at z = 1 both sides give 2. Convolution ↔ product, which is why systems are analysed in z. For causal x[n], x[0] = lim X(z) as z → ∞ (initial value), and the final value is (z − 1)X(z) at z = 1, provided all poles of (z − 1)X(z) are inside the unit circle.
How do you invert a z-transform?
Divide X(z) by z, split X(z)/z into partial fractions, multiply back by z, and you have a sum of z/(z − p) terms. The ROC then decides each term: a pole at or inside the ROC's inner edge gives pⁿu[n], a pole at or beyond its outer edge gives −pⁿu[−n−1]. Long division in powers of z⁻¹ gives the first few samples of a causal sequence directly and is a good check. Residues of X(z)zⁿ⁻¹ give the same answer for the causal case.
X(z) = z/(z − 0.5) with ROC |z| < 0.5. The sequence is
An ROC inside the pole means a left-sided sequence. Here x[−10] = −0.5⁻¹⁰ = −1024, and the samples keep growing to the left, so the ROC misses the unit circle and there is no DTFT.
Worked examples
ROC of a two-sided sequence · x[n] = 0.8ⁿu[n] − 2ⁿu[−n−1] (illustrative)
- Right part: 0.8ⁿu[n] ↔ z/(z − 0.8), converges for |z| > 0.8
- Left part: −2ⁿu[−n−1] ↔ z/(z − 2), converges for |z| < 2
- Both hold in the ring 0.8 < |z| < 2. X(z) = z(2z − 2.8)/((z − 0.8)(z − 2)): poles 0.8 and 2, zeros 0 and 1.4
- The ring contains |z| = 1, so the DTFT exists and the sequence is absolutely summable (stable); it is not causal
- At ω = 0: X(1) = 1/0.2 + 1/(−1) = 5 − 1 = 4. Direct check: Σ 0.8ⁿ = 5 and −(½ + ¼ + …) = −1
- At ω = π: X(−1) = 1/1.8 + 1/3 = 0.8889
Answer: ROC 0.8 < |z| < 2; stable but two-sided; X(e^j0) = 4, X(e^jπ) = 0.8889
Inverse by partial fractions, three ROCs · X(z) = (z² + 0.2z)/((z − 0.8)(z + 0.4)) (illustrative)
- X(z)/z = (z + 0.2)/((z − 0.8)(z + 0.4)) = A/(z − 0.8) + B/(z + 0.4)
- A = (0.8 + 0.2)/(0.8 + 0.4) = 0.8333 = 5/6, B = (−0.4 + 0.2)/(−0.4 − 0.8) = 0.1667 = 1/6
- ROC |z| > 0.8 (causal): x[n] = (5/6)0.8ⁿ + (1/6)(−0.4)ⁿ for n ≥ 0, so x[0] = 1, x[1] = 0.60, x[2] = 0.56, x[3] = 0.416
- Long-division check: x[n] = 0.4x[n−1] + 0.32x[n−2] after the first two samples gives 1, 0.6, 0.56, 0.416
- ROC 0.4 < |z| < 0.8: x[n] = (1/6)(−0.4)ⁿu[n] − (5/6)0.8ⁿu[−n−1] (two-sided)
- ROC |z| < 0.4: x[n] = −[(5/6)0.8ⁿ + (1/6)(−0.4)ⁿ]u[−n−1] (left-sided)
Answer: Only the causal choice |z| > 0.8 contains the unit circle, so only it is stable
Time shift and a delayed start · x[n] = 0.9ⁿ u[n − 3] (illustrative)
- Write x[n] = 0.9³ · 0.9ⁿ⁻³u[n − 3] = 0.729 · (0.9ⁿu[n] delayed by 3)
- Shift property: X(z) = 0.729 z⁻³ · z/(z − 0.9) = 0.729/(z²(z − 0.9)), ROC |z| > 0.9
- Check at z = 1.5: 0.729/(2.25 × 0.6) = 0.54; direct sum Σ from n = 3 of 0.6ⁿ = 0.216/0.4 = 0.54
- Initial value: X(z) → 0 as z → ∞, matching x[0] = 0
Answer: X(z) = 0.729/(z²(z − 0.9)), |z| > 0.9; the delay adds poles at z = 0, which do not change the ROC of a causal sequence
DT systems in the z-domain: transfer function, pole-zero plot, stability, FIR and IIR filters
Which syllabus covers thisAU EC3354 Unit VVTU BEC402 Module 5SPPU not in syllabusMU ECC404 Modules 5-6KTU PCECT402 Module 4
How do you get a transfer function from a difference equation?
Take the z-transform of every term, using x[n − k] ↔ z⁻ᵏX(z) with zero initial conditions, and solve for H(z) = Y(z)/X(z). For y[n] = 1.2y[n−1] − 0.72y[n−2] + x[n]: Y(1 − 1.2z⁻¹ + 0.72z⁻²) = X, so H(z) = z²/(z² − 1.2z + 0.72). Its poles are the roots of the denominator, 0.6 ± j0.6, and its zeros are the roots of the numerator, a double zero at z = 0. Coefficients on x make zeros; coefficients on past y make poles.
What do the radius and the angle of a pole tell you?
A pole pair r e^(±jθ) puts rⁿcos(θn + φ) into the impulse response. The radius sets the decay: a causal system is stable only if every pole has r < 1; at r = 1 it rings forever and for r > 1 it grows. The angle sets the oscillation: θ radians per sample is a period of 2π/θ samples. The poles 0.6 ± j0.6 have r = 0.849 and θ = 45°, so h[n] repeats every 8 samples and falls below 1% of its start after 29 samples.
How can you read a frequency response straight off the pole-zero plot?
Put z = e^(jω), a point going round the unit circle as frequency rises from 0 to π. Each factor (e^(jω) − zₖ) is a vector from a zero to that point, and each (e^(jω) − pₖ) a vector from a pole. So |H| is the product of zero-vector lengths over the product of pole-vector lengths. A zero on the circle makes a null; a pole just inside makes a peak. A 50 Hz notch at fs = 1 kHz (illustrative) puts zeros at ±0.1π and poles at radius 0.95 behind them: the null is exact and only about 16 Hz wide.
What separates FIR filters from IIR filters?
An FIR filter uses only present and past inputs, so h[n] has a finite number of taps and every pole sits at z = 0: it is always stable, and symmetric taps give exactly linear phase (a pure delay). An IIR filter feeds back past outputs, so h[n] never quite ends. It gets a sharp response from very few coefficients, but it can go unstable and its phase is not linear. At a quarter of the sampling rate the 3-tap FIR [0.25, 0.5, 0.25] passes 0.50, and the 2-coefficient IIR y[n] = 0.6y[n−1] + 0.4x[n] passes 0.343.
How do systems in series and in parallel combine?
In series the output of one feeds the next: H = H₁H₂, and h = h₁ * h₂. In parallel both see the same input and their outputs add: H = H₁ + H₂, h = h₁ + h₂. Take H₁ = 1/(1 − 0.5z⁻¹) and H₂ = 1/(1 + 0.3z⁻¹). In series, H = 1/(1 − 0.2z⁻¹ − 0.15z⁻²), with h[n] = 0.625(0.5)ⁿ + 0.375(−0.3)ⁿ, so h[1] = 0.2. In parallel, h[n] = 0.5ⁿ + (−0.3)ⁿ and h[0] = 2. The poles of the combination are the poles of the parts.
A pole pair at angle 60° (zeros at the origin) moves from radius 0.9 to 0.99. The response near 60°
|H| at ω = 60° rises from 6.1 to 58: the pole vector shrinks from 0.1 to 0.01. The impulse response falls to 1% after 44 samples at r = 0.9 but 458 samples at r = 0.99.
Worked examples
Poles, stability and the impulse response · y[n] = 1.2y[n−1] − 0.72y[n−2] + x[n] (illustrative)
- H(z) = 1/(1 − 1.2z⁻¹ + 0.72z⁻²) = z²/(z² − 1.2z + 0.72)
- Poles: z = 0.6 ± √(0.36 − 0.72) = 0.6 ± j0.6, radius 0.8485, angle 45°: inside the unit circle, so stable
- Recursion: h[0] = 1, h[1] = 1.2, h[2] = 1.44 − 0.72 = 0.72, h[3] = 0.864 − 0.864 = 0, h[4] = −0.5184
- Closed form: h[n] = 0.8485ⁿ sin((n + 1)π/4)/sin(π/4); n = 4 gives 0.5184 × (−1) = −0.5184
- Period 2π/(π/4) = 8 samples; envelope below 1% when 0.8485ⁿ < 0.01, n > 28.0
Answer: Stable, oscillating every 8 samples, decayed to 1% after about 29 samples
A 50 Hz mains-hum notch · Remove 50 Hz from a signal sampled at fs = 1 kHz (an illustrative design), keeping other frequencies
- ω0 = 2π × 50/1000 = 0.1π rad/sample. Zeros at e^(±j0.1π) on the unit circle; poles at 0.95e^(±j0.1π) just behind them
- b = [1, −2cos 18°, 1] = [1, −1.9021, 1]; a = [1, −1.9cos 18°, 0.9025] = [1, −1.8070, 0.9025]
- y[n] = 1.8070y[n−1] − 0.9025y[n−2] + x[n] − 1.9021x[n−1] + x[n−2]
- DC gain |H(1)| = 1.0251; scale b by 0.9755 for unity gain at 0 Hz
- Gains after scaling: 50 Hz 0, 40 Hz 0.792, 60 Hz 0.793, 150 Hz 1.022, 500 Hz 1.026
- −3 dB points by bisection: 42.2 Hz and 57.8 Hz, a width of 15.6 Hz (rule of thumb 2(1 − r)fs/2π = 15.9 Hz)
Answer: A second-order IIR notch with a 16 Hz wide null at 50 Hz; moving the poles closer to the circle makes it narrower
FIR against IIR low-pass · FIR h = [0.25, 0.5, 0.25] and IIR y[n] = 0.6y[n−1] + 0.4x[n], both with unity DC gain, fs = 1 kHz (illustrative)
- FIR: H = 0.25(1 + z⁻¹)², a double zero at z = −1; |H(e^jω)| = cos²(ω/2), phase −ω (a constant 1-sample delay, 1 ms)
- IIR: H = 0.4/(1 − 0.6z⁻¹), a pole at 0.6; h[n] = 0.4 × 0.6ⁿ never ends
- At 250 Hz (ω = π/2): FIR 0.50 (−6.02 dB, phase −90°); IIR 0.4/√1.36 = 0.343 (−9.29 dB, phase −31.0°)
- At 500 Hz (ω = π): FIR 0 (the zero on the circle), IIR 0.4/1.6 = 0.25
Answer: The FIR has linear phase and a true null at 500 Hz; the IIR cuts harder at 250 Hz with only two coefficients but never blocks completely and its delay varies with frequency
Probability and random variables: PDF, CDF, mean and variance, uniform and Gaussian
Which syllabus covers thisAU not in syllabusVTU not in syllabusSPPU 204191 Unit VIMU not in syllabusKTU not in syllabus
What are conditional probability and Bayes' theorem, in plain terms?
An experiment has a set of possible outcomes; an event is a group of them, with a probability between 0 and 1. P(A or B) = P(A) + P(B) − P(A and B). P(A | B) is the chance of A once you know B happened: P(A and B)/P(B). Bayes' theorem turns it round: P(A | B) = P(B | A)P(A)/P(B). It is how a receiver reasons backwards from what arrived to what was sent: in the binary-channel example below, a received 1 means a 1 was sent with probability 0.9668, a little below the 0.97 chance that a sent 1 arrives intact.
What are the CDF and the PDF of a random variable?
A random variable gives each outcome a number. Its CDF F(x) = P(X ≤ x) climbs from 0 to 1 and never falls. For a continuous variable the PDF is the slope, f(x) = dF/dx: it is not a probability itself, but the area under it is. P(a < X ≤ b) = F(b) − F(a) = ∫ f dx from a to b, and the total area is 1. For an exponential variable with rate 0.5 per second (mean wait 2 s), P(X ≤ 1 s) = 1 − e^(−0.5) = 0.3935.
How are the mean, the moments and the variance defined?
The mean μ = E[X] = ∫ x f(x) dx is the balance point of the PDF. The n-th moment is E[Xⁿ]. The variance σ² = E[(X − μ)²] = E[X²] − μ² measures spread, and σ, its square root, is in the units of X. For X uniform on 0 to 4: μ = 2, E[X²] = 16/3 = 5.333, so σ² = 5.333 − 4 = 1.333, which is the width²/12 rule. For a noise voltage, E[X²] is its mean power in a 1 Ω resistor.
Why do the uniform and Gaussian distributions matter so much in signals?
An ideal ADC's rounding error is uniform between −Δ/2 and +Δ/2 when the signal spans many steps, so its power is Δ²/12. Thermal noise in a resistor is very nearly Gaussian: f(x) = e^(−(x − μ)²/2σ²)/(σ√(2π)), fixed by μ and σ alone. A Gaussian falls inside μ ± σ 68.27% of the time, inside ±2σ 95.45% and inside ±3σ 99.73%. In the interactive, the same band μ ± kσ holds a different share for each shape.
How close are a sample's mean and variance to the true values?
The sample mean x̄ = Σxᵢ/n is itself random. For independent samples it scatters round μ with a standard deviation of σ/√n: 0.1σ for 100 samples, 0.01σ for 10,000. To halve the error you need four times the data. The sample variance s² = Σ(xᵢ − x̄)²/(n − 1) divides by n − 1, not n, because x̄ was fitted to the same data, which makes the deviations slightly too small. Press 'New sample' in the interactive to watch both estimates wander.
X is Gaussian with mean 5 and standard deviation 2. P(1 < X < 9) is
1 and 9 are μ − 2σ and μ + 2σ, so the probability is Φ(2) − Φ(−2) = 95.45%. About 68% is the ±1σ band, and a Gaussian has no hard edge, so it never reaches 100%.
Worked examples
Bayes on a noisy binary channel · A link sends 1 with probability 0.6 and 0 with 0.4. A 1 is received as 0 with probability 0.03; a 0 is received as 1 with probability 0.05 (illustrative)
- P(1 received) = 0.6 × 0.97 + 0.4 × 0.05 = 0.582 + 0.020 = 0.602
- P(1 sent | 1 received) = 0.582/0.602 = 0.9668
- P(0 sent | 0 received) = 0.4 × 0.95/0.398 = 0.9548
- P(bit error) = 0.6 × 0.03 + 0.4 × 0.05 = 0.038
Answer: A received 1 is right 96.68% of the time, a received 0 95.48%; overall 3.8% of bits are wrong
Quantisation error as a uniform random variable · A 10-bit ADC over 0 to 3.3 V, signal spanning many steps (illustrative)
- Step Δ = 3.3/2¹⁰ = 3.2227 mV; error uniform on ±Δ/2 = ±1.6113 mV, mean 0
- Variance Δ²/12 = 0.8655 mV², σ = Δ/√12 = 0.9303 mV
- P(|error| < 1 mV) = 2 × 1/3.2227 = 0.6206
- Full-scale sine, amplitude 1.65 V: power 1.65²/2 = 1.3612 V²; SQNR = 10 log₁₀(1.3612/8.6546e-07) = 61.97 dB
Answer: σ = 0.930 mV and SQNR 62.0 dB, the familiar 6.02N + 1.76 dB for N = 10
Gaussian noise on a sensor line · Noise voltage Gaussian with mean 0 and σ = 2 mV (illustrative)
- P(|v| > 5 mV) = 2[1 − Φ(2.5)] = 2 × 0.006210 = 0.01242
- P(1 < v < 3 mV) = Φ(1.5) − Φ(0.5) = 0.93319 − 0.69146 = 0.24173
- Threshold exceeded only 0.1% of the time: 2[1 − Φ(k)] = 0.001 gives k = 3.2905 by bisection, t = 6.58 mV
- Mean power E[v²] = σ² = 4 mV² (Simpson check of ∫v² f dv)
Answer: 1.24% of samples exceed ±5 mV; set a detection threshold of about ±6.6 mV for a 0.1% false-alarm rate
Keep going
- GATE Electronics and Communication (EC) preparation
- Control Systems interactive book
- Digital Electronics interactive book
- Microcontrollers interactive book
- Power Electronics interactive book
- Power Systems interactive book
- Analog Electronics interactive book
- Circuit Theory and Network Analysis interactive book
- All Skill-Lync interactive books
Recommended reading
- Signals and Systems, A. V. Oppenheim, A. S. Willsky and S. H. Nawab
- Signals and Systems, S. Haykin and B. Van Veen
- Linear Systems and Signals, B. P. Lathi
- Signals and Systems (Schaum's Outline), H. P. Hsu
- Fundamentals of Signals and Systems, M. J. Roberts
- Signals and Systems, A. Nagoor Kani
- Signals and Systems, A. Anand Kumar
How this book was made
Chapters follow the official syllabi listed above. Every interactive's formulas and every worked-example number were checked by script. Explanations, animations and examples are original to Skill-Lync. Topics informed by MIT OpenCourseWare 6.003 (Signals and Systems), CC BY-NC-SA. All explanations, animations and examples are original to Skill-Lync; MIT does not endorse this page.