Fourier analysis Study Guide

📖 Core Concepts Fourier analysis – decomposes a function into sums of sinusoids or complex exponentials; the reverse process is Fourier synthesis. Fourier transform (FT) – operator that maps a time‑domain signal \(x(t)\) to its frequency‑domain representation \(X(\omega)\). Inverse Fourier transform – reconstructs the original signal from \(X(\omega)\). Frequency‑domain interpretation – magnitude \(|X(\omega)|\) = amplitude of the \(\omega\)‑frequency component; angle \(\arg X(\omega)\) = its initial phase. Linearity – FT\([a f + b g] = a\,\text{FT}[f] + b\,\text{FT}[g]\). Unitarity (Parseval/Plancherel) – energy is preserved: \(\int |x(t)|^2 dt = \frac{1}{2\pi}\int |X(\omega)|^2 d\omega\). Eigenfunction property – differentiating \(x(t)\) multiplies its FT by \(i\omega\): \(\mathcal{F}\{x'(t)\}= i\omega X(\omega)\). Convolution theorem – convolution in time ↔ pointwise multiplication in frequency: \(\mathcal{F}\{f g\}=F(\omega)G(\omega)\). Time‑invariant (LTI) systems – each exponential \(e^{i\omega t}\) is an eigenfunction, so the system’s frequency response fully characterizes its behavior. CTFT, CTFS, DTFT, DFT – four common “flavors” (continuous‑time transform, continuous‑time series, discrete‑time transform, discrete Fourier transform) with paired forward/backward formulas. Fast Fourier Transform (FFT) – algorithm that computes the DFT in \(O(N\log N)\) time. --- 📌 Must Remember CTFT: \(X(\omega)=\displaystyle\int{-\infty}^{\infty} x(t)e^{-i\omega t}\,dt\). Inverse CTFT: \(x(t)=\frac{1}{2\pi}\displaystyle\int{-\infty}^{\infty} X(\omega)e^{i\omega t}\,d\omega\). DTFT: \(X(\omega)=\displaystyle\sum{n=-\infty}^{\infty} x[n]e^{-i\omega n}\); periodic with \(2\pi\). Inverse DTFT: \(x[n]=\frac{1}{2\pi}\displaystyle\int{-\pi}^{\pi} X(\omega)e^{i\omega n}\,d\omega\). DFT: \(X[k]=\displaystyle\sum{n=0}^{N-1} x[n]e^{-i2\pi kn/N}\). Inverse DFT: \(x[n]=\frac{1}{N}\displaystyle\sum{k=0}^{N-1} X[k]e^{i2\pi kn/N}\). Sampling frequency spacing: \( \Delta f = \frac{1}{NT}\) when sampling interval is \(T\) and there are \(N\) points. Convolution ↔ multiplication: \(f g \xleftrightarrow{\mathcal{F}} F(\omega)G(\omega)\). Differentiation ↔ \(i\omega\) multiplication: \(\mathcal{F}\{x'(t)\}=i\omega X(\omega)\). Parseval: \(\displaystyle\int |x(t)|^2 dt = \frac{1}{2\pi}\int |X(\omega)|^2 d\omega\). --- 🔄 Key Processes Compute a CTFT Write the integral \( \int x(t)e^{-i\omega t} dt\). Evaluate analytically (look‑up tables) or numerically (FFT after sampling). From time‑domain to frequency‑domain (sampling → DFT) Sample \(x(t)\) uniformly: \(x[n]=x(nT)\). Form the length‑\(N\) vector \(\{x[0],…,x[N-1]\}\). Apply FFT to obtain \(X[k]\). Filtering via convolution theorem Take FT of signal \(X(\omega)\) and filter \(H(\omega)\). Multiply: \(Y(\omega)=X(\omega)H(\omega)\). Inverse FT → filtered signal \(y(t)\). Design a band‑pass filter Choose frequency window \(W(\omega)\) that is 1 inside passband, 0 elsewhere. Multiply \(X(\omega)\) by \(W(\omega)\) → filtered spectrum. Zero‑padding to increase spectral resolution Append zeros to \(\{x[n]\}\) to length \(M>N\). Compute DFT; frequency bins become \(\Delta f = 1/(MT)\). --- 🔍 Key Comparisons CTFT vs. DTFT Domain: CTFT – continuous time; DTFT – discrete‑time sequence. Frequency: CTFT – continuous \(\omega\); DTFT – periodic (period \(2\pi\)). DTFT vs. DFT Signal length: DTFT – infinite/aperiodic; DFT – finite, length‑\(N\) block. Result: DTFT – continuous periodic function; DFT – \(N\) discrete samples. CTFS vs. DTFS Original signal: CTFS – continuous periodic; DTFS – discrete periodic. Coefficients: CTFS – integrals over one period; DTFS – finite sum \(\frac{1}{N}\sum\). FFT vs. Direct DFT Complexity: FFT – \(O(N\log N)\); Direct – \(O(N^2)\). --- ⚠️ Common Misunderstandings “Fourier transform of a real signal is always real.” – Wrong; real signals yield Hermitian symmetry: \(X(-\omega)=X^(\omega)\). “Zero‑padding adds new frequency content.” – It only interpolates the spectrum; it does not create new information. “Convolution in time always equals multiplication in frequency, regardless of scaling.” – You must respect normalization (e.g., \(1/2\pi\) factors depending on FT convention). “DFT and DTFT are interchangeable.” – DFT samples the DTFT at \(N\) equally spaced frequencies; they coincide only when the underlying sequence is periodic with period \(N\). --- 🧠 Mental Models / Intuition Spectrum as a fingerprint – every signal’s unique “color” is its frequency magnitude pattern; phases tell you how those colors line up in time. Differentiation = high‑pass filter – multiplying by \(i\omega\) boosts high frequencies, so the derivative emphasizes rapid changes. Convolution = blending paints – convolving two signals mixes their spectral “colors”; in the frequency domain you simply blend by multiplication. --- 🚩 Exceptions & Edge Cases Dirichlet conditions – Fourier series converges to the original periodic function only if the function is piecewise‑continuous, has a finite number of extrema, and is absolutely integrable over a period. Non‑unitary normalization – some textbooks omit the \(1/2\pi\) factor; always check the convention before applying Parseval or inversion formulas. Aliasing – decimating (down‑sampling) without anti‑alias filtering causes overlapping of frequency components. --- 📍 When to Use Which | Situation | Best Transform | Why | |-----------|----------------|-----| | Continuous‑time signal, analytic expression | CTFT | Direct integral gives closed‑form spectrum. | | Periodic continuous‑time signal | CTFS | Provides discrete set of harmonic coefficients. | | Discrete‑time (sampled) signal, need full spectrum | DTFT | Captures periodic frequency behavior of infinite sequence. | | Finite‑length block, computational efficiency required | DFT / FFT | Gives discrete spectrum; FFT makes it fast. | | Design a filter in frequency domain | Multiply FT by window (use whichever FT matches signal type). | | Need time–frequency localization | Short‑time FT / Gabor (beyond core outline). | | Signal is real‑valued | Exploit Hermitian symmetry to halve computation/storage. | --- 👀 Patterns to Recognize \(i\omega\) factor appears whenever a time‑domain derivative is present. Multiplication ↔ Convolution: if a problem mentions “filtering” or “smoothing”, look for a frequency‑domain product. Periodicity in frequency → the underlying signal is discrete‑time (DTFT). Discrete set of coefficients → the original signal is periodic (Fourier series). Zero‑padding → expect finer frequency grid but same underlying spectral shape. --- 🗂️ Exam Traps Choosing wrong normalization – mixing conventions (e.g., missing \(1/2\pi\)) leads to incorrect inverse results. Confusing DFT with DTFT – DFT gives \(N\) samples; DTFT is continuous. Selecting the wrong one yields “off‑by‑N” frequency spacing errors. Assuming real‑signal FT is real – remember Hermitian symmetry; imaginary parts may be non‑zero. Neglecting periodicity of DTFT – forgetting the \(-\pi\) to \(\pi\) wrap‑around causes aliasing mistakes. Applying convolution theorem without scaling – the theorem holds only under the same transform convention; missing scaling factors produces amplitude errors. ---