Signals & Systems
Continuous-time and discrete-time signals, LTI systems, convolution, Fourier series, Fourier transform, Laplace transform, and Z-transform.
Signals and Systems: An Introduction
What is a Signal? A signal is a function that carries information. Examples: Audio: air pressure vs time Video: pixel intensity vs (x, y, t) Sensor: voltage vs time Image: brightness vs (x, y) Math...
What is a Signal?
A signal is a function that carries information. Examples:
- Audio: air pressure vs time
- Video: pixel intensity vs (x, y, t)
- Sensor: voltage vs time
- Image: brightness vs (x, y)
Mathematically, a signal is a function of one or more independent variables.
Continuous vs Discrete
Continuous-time (CT): x(t) defined for all t. Example: analog audio.
Discrete-time (DT): x[n] defined only for integer n. Example: sampled audio at 44.1 kHz.
Almost all modern signal processing is digital (DT), but CT analysis is the foundation.
Common Signals
Unit impulse δ(t): infinite at t=0, zero elsewhere, area = 1.
Unit step u(t): 0 for t<0, 1 for t>0.
Sinusoid: x(t) = A cos(ωt + φ). Frequency ω (rad/s), phase φ.
Complex exponential: x(t) = e^(st), where s = σ + jω. The most fundamental signal — from it, all others are built.
Unit ramp: x(t) = t × u(t).
Signal Operations
Time shift: x(t - t₀) shifts right by t₀.
Time scale: x(at) compresses by a (a > 1) or stretches (a < 1).
Time reversal: x(-t) flips around t=0.
Amplitude scaling: A × x(t).
Addition: (x₁ + x₂)(t) = x₁(t) + x₂(t).
Signal Properties
Even: x(-t) = x(t). Symmetric about 0.
Odd: x(-t) = -x(t). Antisymmetric.
Periodic: x(t) = x(t + T) for all t. Smallest T is the fundamental period.
Energy signal: finite total energy ∫|x(t)|² dt < ∞.
Power signal: finite average power lim(T→∞) (1/T) ∫|x(t)|² dt < ∞.
Causal: x(t) = 0 for t < 0.
What is a System?
A system transforms an input signal into an output signal. Examples:
- Amplifier: multiplies by gain
- Filter: removes frequency components
- Sampler: converts CT to DT
- Modulator: shifts frequency
System Properties
Linear: superposition holds. S(ax₁ + bx₂) = aS(x₁) + bS(x₂).
Time-invariant: S(x(t - t₀)) = y(t - t₀). Same behavior at all times.
Causal: output depends only on present and past inputs.
Stable (BIBO): bounded input → bounded output.
Memoryless: output depends only on current input.
Invertible: distinct inputs give distinct outputs.
LTI Systems — The Big Idea
Linear Time-Invariant (LTI) systems are the most important class. They are completely characterized by their impulse response h(t).
For any input x(t), the output is:
y(t) = x(t) * h(t) = ∫ x(τ) h(t - τ) dτ
This convolution integral is the foundation of signals and systems.
Key Takeaways
- Signals are functions carrying information
- CT vs DT, energy vs power, causal vs noncausal
- Systems transform signals
- Linear + time-invariant = LTI — the fundamental class
- LTI systems described by impulse response
- Convolution gives the output for any input
Convolution and LTI Systems
Convolution Convolution is the operation that combines two signals to produce a third: y(t) = (x * h)(t) = ∫_{-∞}^{∞} x(τ) h(t - τ) dτ For discrete-time signals: y[n] = (x * h)[n] = Σ_{k=-∞}^{∞} ...
Convolution
Convolution is the operation that combines two signals to produce a third:
y(t) = (x * h)(t) = ∫_{-∞}^{∞} x(τ) h(t - τ) dτ
For discrete-time signals:
y[n] = (x * h)[n] = Σ_{k=-∞}^{∞} x[k] h[n - k]
Convolution is the fundamental operation of LTI systems.
Properties
Commutative: x * h = h * x
Associative: (x * h₁) * h₂ = x * (h₁ * h₂)
Distributive: x * (h₁ + h₂) = x * h₁ + x * h₂
These properties make it easy to combine systems.
Impulse Response
If x(t) = δ(t), then y(t) = h(t). The impulse response is the output when the input is an impulse.
Everything about an LTI system is determined by its impulse response:
- Cascade of LTI systems → convolve their impulse responses
- Parallel combination → sum impulse responses
Step Response
If x(t) = u(t), then y(t) = ∫h(τ)dτ (integral of impulse response).
The step response is often easier to measure than the impulse response.
Convolution with Simple Signals
Impulse: x(t) * δ(t - t₀) = x(t - t₀). Just a shift.
Step: x(t) * u(t) = ∫_{-∞}^{t} x(τ) dτ. Integrator.
Exponential: x(t) * e^(-at)u(t) gives a smoothed version of x.
Convolution as a Filter
Convolution smooths, sharpens, or shifts a signal, depending on h:
- Moving average → low-pass filter
- Differentiator → high-pass filter
- Gaussian → smoothing
Every FIR filter is a convolution with its coefficient sequence.
Visual Example
Convolving a rectangle (width T) with itself gives a triangle (width 2T). This is the classic example — it's why pulse shapes change as they pass through filters.
Convolution in Matrix Form
For discrete-time, convolution can be written as matrix multiplication:
y = H × x
where H is a Toeplitz matrix containing shifted copies of h.
This is why convolution is a linear operator — it's a matrix.
Deconvolution
Given y and h, recover x:
X = Y / H (in frequency domain)
Often unstable due to noise amplification. Regularization (Tikhonov) helps.
Applications
- Audio filtering (EQ, reverb)
- Image blurring and sharpening
- Radar and sonar pulse compression
- Communication channel equalization
- Seismic deconvolution
- Neural network convolution layers
Key Takeaways
- Convolution: y = x * h
- LTI systems are fully described by h(t)
- Properties: commutative, associative, distributive
- Cascade = convolve, parallel = add
- Convolution is a linear operator (matrix form)
- The fundamental operation for filtering
Fourier Series and Fourier Transform
Fourier Series Any periodic signal x(t) with period T can be written as: x(t) = Σ_{k=-∞}^{∞} c_k × e^(j2πkt/T) where the coefficients are: c_k = (1/T) × ∫₀^T x(t) × e^(-j2πkt/T) dt The signal ...
Fourier Series
Any periodic signal x(t) with period T can be written as:
x(t) = Σ_{k=-∞}^{∞} c_k × e^(j2πkt/T)
where the coefficients are:
c_k = (1/T) × ∫₀^T x(t) × e^(-j2πkt/T) dt
The signal is decomposed into complex exponentials at frequencies k/T.
Trigonometric Form
Equivalent form using sine and cosine:
x(t) = a₀ + Σ a_k cos(2πkt/T) + Σ b_k sin(2πkt/T)
where a₀ is the DC component.
Example: Square Wave
A square wave of amplitude A and period T:
x(t) = (4A/π) × [sin(ωt) + sin(3ωt)/3 + sin(5ωt)/5 + ...]
Only odd harmonics. Amplitudes 1, 1/3, 1/5...
Fourier Transform
For aperiodic signals, take T → ∞. The Fourier transform:
X(jω) = ∫_{-∞}^{∞} x(t) × e^(-jωt) dt
Inverse:
x(t) = (1/2π) × ∫_{-∞}^{∞} X(jω) × e^(jωt) dω
X(jω) tells us the "amount" of each frequency in x(t).
Transform Pairs
| x(t) | X(jω) |
|------|-------|
| δ(t) | 1 |
| u(t) | 1/(jω) + πδ(ω) |
| e^(-at) u(t) | 1/(a + jω) |
| cos(ω₀t) | π[δ(ω-ω₀) + δ(ω+ω₀)] |
| sin(ω₀t)/πt | rect(ω) |
Properties
Linearity: a x₁ + b x₂ ↔ a X₁ + b X₂
Time shift: x(t - t₀) ↔ e^(-jωt₀) X(jω)
Frequency shift: e^(jω₀t) x(t) ↔ X(j(ω - ω₀))
Scaling: x(at) ↔ (1/|a|) X(jω/a)
Differentiation: dx/dt ↔ jω X(jω)
Integration: ∫x dt ↔ X(jω)/(jω) + π X(0)δ(ω)
Convolution: x * h ↔ X × H
Multiplication: x × h ↔ (1/2π) X * H
Parseval's Theorem
∫|x(t)|² dt = (1/2π) × ∫|X(jω)|² dω
Energy in time domain = energy in frequency domain.
Frequency Response
For an LTI system with impulse response h(t), the frequency response is H(jω) = FT{h(t)}.
For input x(t), output spectrum:
Y(jω) = H(jω) × X(jω)
The system multiplies the input spectrum by H. This is why frequency-domain analysis is so powerful.
Applications
- Filter design (specify H(jω), find h(t))
- Modulation (shift signal to a new frequency band)
- Sampling theory (Nyquist rate)
- Power spectral density
- Image processing (2D FT)
- Solving differential equations
- Quantum mechanics (wave functions)
Key Takeaways
- Fourier series decomposes periodic signals into harmonics
- Fourier transform extends to aperiodic signals
- X(jω) shows the frequency content of x(t)
- Convolution in time = multiplication in frequency
- Parseval: energy conservation between domains
- LTI systems act as frequency filters
- Every signal can be analyzed in the frequency domain
Laplace Transform and Transfer Functions
The Laplace Transform Generalizes the Fourier transform to complex frequencies: X(s) = ∫₀^∞ x(t) × e^(-st) dt where s = σ + jω is a complex variable. The Laplace transform handles signals that grow...
The Laplace Transform
Generalizes the Fourier transform to complex frequencies:
X(s) = ∫₀^∞ x(t) × e^(-st) dt
where s = σ + jω is a complex variable.
The Laplace transform handles signals that grow exponentially (Fourier doesn't), and is the natural tool for differential equations.
Region of Convergence
The set of s values for which X(s) converges is the ROC. It's usually a vertical strip in the s-plane.
For a causal signal like e^(-at)u(t), ROC is Re(s) > -a.
Common Transforms
| x(t) | X(s) |
|------|------|
| δ(t) | 1 |
| u(t) | 1/s |
| t u(t) | 1/s² |
| e^(-at) u(t) | 1/(s + a) |
| sin(ωt) u(t) | ω/(s² + ω²) |
| cos(ωt) u(t) | s/(s² + ω²) |
Properties
Linearity: a x₁ + b x₂ ↔ a X₁ + b X₂
Time shift: x(t - T) u(t - T) ↔ e^(-sT) X(s)
Scaling: x(at) ↔ (1/a) X(s/a)
Differentiation: dx/dt ↔ sX(s) - x(0⁻)
Integration: ∫₀^t x dt ↔ X(s)/s
Initial value: x(0⁺) = lim(s→∞) sX(s)
Final value: lim(t→∞) x(t) = lim(s→0) sX(s)
Convolution: x * h ↔ X × H
The derivative and convolution properties are what make Laplace so useful for circuit and control analysis.
Transfer Functions
For an LTI system with input u(t) and output y(t):
H(s) = Y(s) / U(s) (zero initial conditions)
This is the transfer function. It fully describes the system.
For a linear ODE:
a_n y^(n) + ... + a₁ y' + a₀ y = b_m u^(m) + ... + b₁ u' + b₀ u
Taking Laplace:
H(s) = (b_m s^m + ... + b₀) / (a_n s^n + ... + a₀)
The numerator gives the zeros, the denominator gives the poles.
Poles and Stability
A system is stable if all poles are in the left half-plane (Re(s) < 0).
Poles on the imaginary axis: marginally stable (constant amplitude oscillation).
Poles in the right half-plane: unstable (exponential growth).
This is why we compute poles for stability analysis.
Example: RC Circuit
Circuit equation: RC × dV/dt + V = Vin
Laplace: (RC s + 1) V(s) = Vin(s)
Transfer function: H(s) = 1 / (RCs + 1)
Pole at s = -1/RC. Time constant τ = RC.
Step response: V(t) = Vin(1 - e^(-t/τ)).
State-Space to Transfer Function
Given state-space equations:
ẋ = A x + B u
y = C x + D u
Transfer function:
H(s) = C(sI - A)^(-1) B + D
This bridges the state-space and transfer-function views.
Applications
- Circuit analysis (impedance method)
- Control system design
- Filter design
- Solving ODEs
- Process modeling
- Signal processing
Key Takeaways
- Laplace generalizes Fourier to complex s
- Converts ODEs to algebraic equations
- Transfer function H(s) = Y(s)/U(s)
- Poles determine stability (LH plane = stable)
- Properties: derivative → sX, convolution → multiplication
- Foundation for circuit and control analysis
Sampling and the Discrete-Time World
Why Sample? Digital computers can't process continuous signals directly. We must sample: take instantaneous values at regular intervals. Sampling rate fs determines everything about the resulting...
Why Sample?
Digital computers can't process continuous signals directly. We must sample: take instantaneous values at regular intervals.
Sampling rate fs determines everything about the resulting digital signal.
Sampling Theorem (Nyquist-Shannon)
A bandlimited signal with maximum frequency fmax can be perfectly reconstructed from samples taken at:
fs > 2 × fmax
2 × fmax is the Nyquist rate. fs/2 is the Nyquist frequency.
If fs < 2 fmax, aliasing occurs — high frequencies appear as low frequencies.
Aliasing
When sampled below Nyquist, a frequency f appears as:
f_alias = |f - k × fs| (closest to f)
Example: fs = 1000 Hz, f = 1200 Hz → alias = 200 Hz (indistinguishable from a real 200 Hz signal).
Aliasing is why:
- Audio is sampled at 44.1 kHz (human hearing to 20 kHz)
- Anti-aliasing filters are essential before ADCs
- "Wagon wheel effect" in movies
Sampling in Frequency Domain
Sampling in time = periodic replication in frequency:
X_s(jω) = (1/T) × Σ X(j(ω - k ωs))
The spectrum is repeated every ωs = 2π fs. If the original signal is not bandlimited, the replicas overlap → aliasing.
Reconstruction
Ideal reconstruction uses a sinc filter:
x(t) = Σ x[n] × sinc((t - nT)/T)
This is a perfect interpolation if the sampling theorem holds.
Real DACs use zero-order hold (staircase) followed by a smoothing filter.
Discrete-Time Processing
Once sampled, we work in discrete time:
y[n] = Σ h[k] × x[n - k]
The DTFT (discrete-time Fourier transform):
X(e^(jΩ)) = Σ x[n] × e^(-jΩn)
Where Ω is digital frequency (rad/sample).
The DFT and FFT
For finite-length sequences, use the Discrete Fourier Transform (DFT):
X[k] = Σ_{n=0}^{N-1} x[n] × e^(-j2πnk/N)
Computes N frequency samples. Time: O(N²).
The FFT (Fast Fourier Transform) does it in O(N log N). This is one of the most important algorithms ever invented.
Nyquist Zones
Above fs/2, frequencies fold back. The range 0 to fs/2 is the first Nyquist zone, fs/2 to fs is the second, and so on.
Undersampling (deliberately sampling below Nyquist) is used in radio receivers to down-convert signals to lower frequencies.
Quantization
Real ADCs quantize amplitudes to a finite number of levels. This adds quantization noise.
For an N-bit ADC, the SNR due to quantization is:
SNR = 6.02 N + 1.76 dB
A 12-bit ADC gives ~74 dB SNR.
Oversampling
Sampling faster than needed allows:
- Higher SNR (noise spreads over wider band)
- Simpler anti-aliasing filter
- 4× oversampling gives 6 dB SNR improvement (1 extra bit)
Used in audio (delta-sigma converters) to achieve 24-bit effective resolution from 1-bit quantizers.
Applications
- Digital audio (CD: 44.1 kHz, 16-bit)
- Digital video (multiple sampling rates)
- Software-defined radio (direct RF sampling)
- Medical imaging (CT, MRI)
- Industrial control (sensor sampling)
Key Takeaways
- Sampling: fs > 2 fmax for perfect reconstruction
- Aliasing: high frequencies appear as low frequencies
- Anti-aliasing filters essential before ADC
- FFT: O(N log N) for frequency analysis
- Quantization adds noise: 6 dB per bit
- Oversampling improves SNR
- Nyquist zones and undersampling in RF
📝 Problem Set 1: Signals, Convolution, LTI Systems
Q1. Is the system y(t) = x(t)cos(2πt) linear? Time-invariant? Justify.
Show Solution
Linear: y = cos(2πt)(ax1+bx2) = a cos(2πt)x1 + b cos(2πt)x2 ✓. Time-variant: S(x(t-t0)) = x(t-t0)cos(2πt), but y(t-t0) = x(t-t0)cos(2π(t-t0)) ≠. So it's time-varying.
Q2. Compute the convolution of x[n] = {1, 2, 3} and h[n] = {1, 1} (both starting at n=0).
Show Solution
y[0] = 1×1 = 1. y[1] = 2×1 + 1×1 = 3. y[2] = 3×1 + 2×1 = 5. y[3] = 3×1 = 3. So y = {1, 3, 5, 3}.
Q3. Find the Fourier series coefficients of a periodic square wave with amplitude A, period T, 50% duty cycle.
Show Solution
For a 50% duty-cycle square wave, a_0 = 0, a_n = 0. b_n = (2A/(nπ))(1-cos(nπ)) = 0 for even n, 4A/(nπ) for odd n. So x(t) = (4A/π)[sin(ωt) + sin(3ωt)/3 + sin(5ωt)/5 + ...].
🎯 Midterm 1: Signals, LTI, Fourier
Q1. Compute the Fourier transform of x(t) = e^(-3t)u(t). Plot the magnitude spectrum.
Show Solution
X(jω) = ∫₀^∞ e^(-3t)e^(-jωt)dt = 1/(3+jω). |X(jω)| = 1/√(9+ω²). At ω=0, |X|=1/3. At ω→∞, |X|→0. -3dB cutoff at ω=3 rad/s.
Q2. A system has impulse response h(t) = e^(-2t)u(t). Find the output for input x(t) = u(t).
Show Solution
y(t) = ∫₀^t e^(-2τ)dτ = (1/2)(1 - e^(-2t)) for t ≥ 0. So y(t) = 0.5(1 - e^(-2t))u(t).
Q3. Is the system y(t) = x(2t) causal? Stable? Time-invariant? Justify.
Show Solution
Causal: y(t) depends on x(2t), which for t>0 uses x at future time >t. So NOT causal. Stable: |x|≤M → |y|≤M ✓ BIBO stable. Time-varying: S(x(t-t0)) = x(2t-t0), but y(t-t0) = x(2(t-t0)) = x(2t-2t0). Not equal → time-varying.