Signal Processing Interview Questions

A signal is a function that conveys information about a phenomenon. Classification: Continuous-time (CT) vs Discrete-time (DT) based on time variable, Analog vs Digital based on amplitude values, Deterministic vs Random based on predictability, and Periodic vs Aperiodic based on repetition. Examples: speech is CT analog, digital audio is DT digital, and power line frequency is periodic deterministic. Signal processing analyzes, modifies, or synthesizes signals for communication, control, and analysis applications.

LTI systems have two fundamental properties: Linearity means the system satisfies superposition - if x1->y1 and x2->y2, then ax1+bx2 -> ay1+by2. Time-invariance means behavior does not change over time - if x(t)->y(t), then x(t-t0)->y(t-t0). LTI systems are completely characterized by their impulse response h(t) or frequency response H(jw). Output is the convolution of input with impulse response: y(t) = x(t) * h(t). Most classical signal processing assumes LTI behavior.

The Fourier Transform decomposes a signal into its constituent frequencies, converting time-domain representation to frequency-domain: X(f) = integral of x(t)*e^(-j2*pi*f*t)dt. It reveals frequency content, bandwidth, and spectral characteristics. Important for: filter design (frequency-selective systems), modulation analysis, spectral analysis, and understanding system frequency response. The inverse transform reconstructs time signal from spectrum. Variants include DTFT (discrete-time), DFT (discrete frequency), and FFT (efficient algorithm).

Convolution is a mathematical operation that combines two signals to produce a third: y(t) = x(t) * h(t) = integral of x(tau)*h(t-tau)d(tau). For LTI systems, output equals input convolved with impulse response. Properties: commutative, associative, and distributive. Convolution in time equals multiplication in frequency domain (and vice versa). Applications: Filter implementation (signal convolved with filter impulse response), Image processing (2D convolution with kernels), and System analysis (cascade systems multiply frequency responses).

The sampling theorem (Nyquist-Shannon) states: A bandlimited signal with maximum frequency fmax can be perfectly reconstructed from samples taken at rate fs > 2*fmax. The minimum sampling rate (2*fmax) is the Nyquist rate. Aliasing occurs when fs < 2*fmax: high-frequency components fold back into lower frequencies, causing distortion that cannot be removed after sampling. Prevention: Use anti-aliasing filter (lowpass) before sampling to remove frequencies above fs/2. Practical systems use fs significantly above Nyquist for margin.

FIR (Finite Impulse Response) filters have impulse response of finite duration, implemented as weighted sum of input samples. Advantages: always stable, linear phase possible, no feedback. IIR (Infinite Impulse Response) filters have infinite-duration impulse response using feedback (recursive). Advantages: achieve sharp cutoffs with fewer coefficients (more efficient for same specification). Trade-offs: FIR needs more computation but is inherently stable; IIR is efficient but requires stability analysis and has nonlinear phase. FIR preferred for linear phase requirements; IIR for efficiency.

Lowpass filters pass frequencies below cutoff, attenuate above; used for anti-aliasing, smoothing, and removing high-frequency noise. Highpass filters pass frequencies above cutoff; used for removing DC offset and low-frequency interference. Bandpass filters pass a range of frequencies; used in communication receivers and audio equalization. Bandstop (notch) filters attenuate a frequency range; used for removing interference like power line hum. Allpass filters pass all frequencies equally but alter phase; used for phase equalization.

The Z-transform converts discrete-time signals to the complex z-domain: X(z) = sum of x[n]*z^(-n). It is the discrete-time equivalent of the Laplace transform. Uses: Analyzing discrete-time systems and stability (poles inside unit circle = stable), Designing digital filters (transfer function H(z)), and Solving difference equations. The unit circle z = e^(jw) corresponds to the frequency response. Z-transform properties (linearity, time shift, convolution) parallel Laplace transform properties, enabling systematic discrete-time system analysis.

Quantization maps continuous amplitude values to a finite set of discrete levels during analog-to-digital conversion. With B bits, there are 2^B levels. Quantization error is the difference between actual and quantized values, appearing as quantization noise. For uniform quantization of full-scale sinusoid, SNR approximately equals 6.02B + 1.76 dB (each bit adds ~6dB). Quantization noise is modeled as uniform random noise. Dithering (adding small noise before quantization) can decorrelate quantization error from signal, improving perceived quality in audio.

The DFT (Discrete Fourier Transform) computes the frequency content of a finite-length discrete signal: X[k] = sum(x[n]*e^(-j2*pi*k*n/N)). It transforms N time samples to N frequency samples. The FFT (Fast Fourier Transform) is an efficient algorithm for computing DFT, reducing complexity from O(N^2) to O(N*log N). Common FFT algorithms: Radix-2 (N is power of 2), Radix-4, and split-radix. FFT enables practical spectral analysis, fast convolution (multiply in frequency domain), and many signal processing applications.

The impulse response h[n] (or h(t)) is the output of a system when the input is a unit impulse. For LTI systems, impulse response completely characterizes the system - output to any input is convolution with h. In frequency domain, Fourier transform of h gives frequency response H(f). From h, determine: Stability (absolutely summable for BIBO stability), Causality (h[n]=0 for n<0), System type (FIR has finite h, IIR has infinite h), and Filter characteristics. Measuring impulse response is a standard system identification technique.

Frequency response H(jw) or H(e^jw) describes how a system affects each frequency component. It has magnitude |H| (gain at each frequency) and phase angle(H) (phase shift at each frequency). Bode plots show magnitude (dB) and phase vs log frequency. For filters: Passband (|H| near 1), Stopband (|H| near 0), Cutoff frequency (typically -3dB point), and Transition band (between pass and stop). Linear phase (constant group delay) preserves waveform shape. Reading frequency response reveals filter selectivity, bandwidth, and phase distortion.

Decimation reduces sampling rate by factor M: keep every Mth sample, discard others. Requires anti-aliasing filter before decimation to prevent aliasing of high frequencies into lower bands. Interpolation increases sampling rate by factor L: insert L-1 zeros between samples, then lowpass filter to remove spectral images. Sample rate conversion by rational factor L/M combines interpolation by L then decimation by M. Efficient implementation uses polyphase filters. Applications: Multi-rate signal processing, audio resampling, and communication systems.

Window functions taper signal at edges before DFT to reduce spectral leakage from finite-length analysis. Without windowing (rectangular window), sharp truncation causes sinc-like spectral spreading. Common windows: Hanning, Hamming (low sidelobes, moderate mainlobe), Blackman (very low sidelobes, wide mainlobe), and Kaiser (adjustable parameters). Trade-off: wider mainlobe (reduced frequency resolution) for lower sidelobes (reduced leakage). Also used in FIR filter design (window method) to truncate ideal infinite impulse response. Window selection depends on application requirements.

Correlation measures similarity between signals. Autocorrelation compares a signal with time-shifted versions of itself, revealing periodicity and power spectral density (via Wiener-Khinchin theorem). Cross-correlation compares two different signals, used for time delay estimation and pattern matching. Mathematical form similar to convolution but without time reversal. Applications: Radar/sonar (range detection), GPS (code synchronization), Speech recognition, Communication (synchronization), and Image matching. Efficient computation via FFT (correlation theorem: correlation in time = conjugate multiplication in frequency).

FIR design methods: Window method - design ideal filter (infinite), window with finite function, simple but limited control. Frequency sampling - specify frequency response at DFT points, inverse DFT gives coefficients. Parks-McClellan (Remez) - optimal equiripple design, minimizes maximum error in pass/stop bands, most common for precision filters. Least squares - minimizes total squared error, good for some applications. Design parameters: filter order (length), passband/stopband edges, ripple specifications. Higher order = sharper transition but more computation. Parks-McClellan is standard for communications; window method for less critical applications.

IIR filter design starts with analog prototype (Butterworth, Chebyshev, Elliptic) then transforms to digital. Transformation methods: Bilinear transform - maps s-plane to z-plane nonlinearly, frequency warping requires pre-warping critical frequencies, most common method. Impulse invariance - samples analog impulse response, can cause aliasing, suitable for lowpass. Design steps: Determine digital specifications, pre-warp frequencies, design analog prototype, apply bilinear transform. Butterworth is maximally flat; Chebyshev has ripple but sharper cutoff; Elliptic has ripple in both bands with sharpest cutoff for given order.

A digital system is BIBO (bounded-input, bounded-output) stable if all poles of H(z) lie inside the unit circle |z| < 1. Analysis methods: Pole location - find roots of denominator polynomial, check magnitudes. Jury stability test - algebraic test without finding roots (analog of Routh-Hurwitz). For IIR filters, check each second-order section (biquad) independently. FIR filters are always stable (no poles except at origin). Marginally stable if poles on unit circle. Coefficient quantization can move poles, potentially causing instability in fixed-point implementations. Use scaled structures and proper word lengths.

STFT analyzes how frequency content changes over time by computing DFT of successive windowed segments: X(m,k) = sum(x[n]*w[n-m]*e^(-j2*pi*k*n/N)). Parameters: window length (frequency resolution), hop size (time resolution), and window type. Spectrogram displays |STFT|^2 as time-frequency image. Trade-off: longer window = better frequency resolution, worse time resolution (uncertainty principle). Applications: Speech analysis, music processing, biomedical signals, and vibration analysis. Variants: reassigned spectrogram (sharper localization), multi-resolution STFT.

LMS (Least Mean Squares) adapts filter coefficients to minimize mean squared error between output and desired signal. Algorithm: y[n] = w^T * x[n] (filter output), e[n] = d[n] - y[n] (error), w[n+1] = w[n] + mu * e[n] * x[n] (update). Parameter mu (step size) controls convergence speed and stability (mu < 2/(lambda_max) for convergence). Trade-offs: larger mu = faster adaptation but more noise, smaller mu = slower but steadier. Applications: Echo cancellation, noise cancellation, channel equalization, and system identification. Variants: NLMS (normalized), RLS (faster but more complex).

Polyphase decomposition splits a filter into multiple subfilters (phases) that operate at reduced rate. For M-fold interpolation: H(z) = sum(z^(-k) * E_k(z^M)), each E_k processes at output rate, combined with upsamplers. Efficiency: Avoids computing outputs that will be discarded (decimation) or processing zero-valued samples (interpolation). Reduces computation by factor M. Used in: Efficient multirate systems, filter banks, channelizers, and sample rate converters. Implementation: Commutator structure switches between phase filters. Critical for practical multirate signal processing.

Fixed-point considerations: Word length selection (bits for coefficients, data, accumulator), Overflow handling (saturation vs wrap-around), Quantization effects (coefficient quantization affects filter response, signal quantization adds noise), Scaling strategy (prevent overflow while maximizing dynamic range), and Precision analysis (error propagation, limit cycles in IIR). Analysis: Fixed-point simulation, comparison with floating-point reference, SNR measurement. Techniques: Block floating-point, guard bits, double-precision accumulator, and careful coefficient scaling. Trade-offs between hardware cost (fewer bits) and performance (more bits). Use tools like MATLAB Fixed-Point Designer for analysis.

Both methods use FFT for efficient convolution of long signals with shorter filters. Overlap-Add: Divide input into non-overlapping blocks, FFT-convolve each with filter, results overlap and are added. Overlap-Save: Divide input into overlapping blocks, FFT-convolve, discard corrupted samples from circular convolution, save valid portion. Both achieve O(N*log(N)) complexity vs O(N*M) direct convolution. Choice depends on implementation: overlap-add has simpler input handling, overlap-save avoids explicit addition. Block size selection balances FFT efficiency vs latency. Essential for real-time processing of long signals.

DCT transforms signal to frequency domain using only cosine basis functions (real-valued, unlike complex DFT). DCT-II (most common): X[k] = sum(x[n]*cos(pi*k*(2n+1)/(2N))). Properties: Energy compaction (most energy in few low-frequency coefficients), Real coefficients (efficient for real signals), and Reduced boundary artifacts vs DFT. Applications: JPEG image compression (2D DCT of 8x8 blocks), video compression (MPEG, H.264), audio compression (MP3 uses modified DCT), and speech coding. DCT enables lossy compression by quantizing/discarding small high-frequency coefficients.

The Hilbert transform shifts all frequency components by 90 degrees (pi/2): H{x(t)} produces the quadrature component. Combined with original signal, forms the analytic signal: xa(t) = x(t) + jH{x(t)}, which has zero negative frequency components. Applications: Envelope detection (|xa(t)| gives instantaneous envelope), Instantaneous frequency (derivative of phase), Single-sideband modulation (generate SSB from DSB), and Creating quadrature signals for I/Q processing. Implemented via FIR filter approximation or via FFT (zero negative frequencies, double positive, inverse FFT).

Filter banks split signals into multiple frequency subbands (analysis) and reconstruct from subbands (synthesis). Perfect reconstruction (PR) requires synthesis filters exactly undo analysis effects. Types: Uniform filter banks (equal bandwidth subbands), Non-uniform (different bandwidths), and Tree-structured (cascaded for octave bands). Design conditions: Aliasing cancellation and Magnitude distortion elimination. QMF (Quadrature Mirror Filter) banks use H1(z) = H0(-z) relationship. Applications: Audio coding, subband adaptive filtering, wavelet transforms, and transmultiplexers. Modern designs use lattice structures or modulated filter banks for efficiency.

PSD describes signal power distribution across frequencies. For random signals, PSD S(f) = Fourier transform of autocorrelation (Wiener-Khinchin theorem). Estimation methods: Periodogram (|FFT|^2/N), high variance but simple; Welch method (average periodograms of overlapping windowed segments), reduced variance; Blackman-Tukey (Fourier transform of windowed autocorrelation estimate); and Parametric methods (AR, ARMA model fitting, good frequency resolution with short data). Trade-offs: frequency resolution vs variance vs bias. Welch is most common for general use.

Goertzel algorithm efficiently computes single DFT bins without full FFT. Uses recursive filter: s[n] = x[n] + 2*cos(2*pi*k/N)*s[n-1] - s[n-2], then X[k] = s[N] - e^(-j2*pi*k/N)*s[N-1]. Complexity: O(N) per bin vs O(N*log(N)) for full FFT. Efficient when computing fewer than log2(N) DFT bins. Applications: DTMF tone detection (detect 8 specific frequencies), frequency monitoring, and narrow-band analysis. Numerical stability good for large N. Second-order section structure fits DSP hardware well. Preferred for real-time detection of specific frequencies.

Sigma-delta modulation uses oversampling and feedback to achieve high resolution with simple quantizer (often 1-bit). Structure: Integrator accumulates difference between input and quantized output, quantizer produces output, feedback subtracts quantized output. Noise shaping: Loop filter shapes quantization noise spectrum, pushing noise to high frequencies where it is removed by digital decimation filter. Higher-order loops provide more aggressive noise shaping but stability is challenging. Result: High SNR in band of interest despite coarse quantization. Used in audio ADCs, DACs, and instrumentation converters.

Linear phase filters have phase response that is linear function of frequency: phi(w) = -alpha*w, causing constant group delay. All frequency components are delayed equally, preserving waveform shape. FIR filters achieve linear phase with symmetric (Type I, II) or antisymmetric (Type III, IV) coefficients. IIR filters cannot have exactly linear phase. Importance: Critical in audio (phase distortion audible), communications (ISI in data transmission), imaging (edge preservation), and measurement systems. When phase distortion is unacceptable, use linear phase FIR despite higher order requirement.

Noise reduction approaches: Filtering (lowpass, bandpass for in-band signal, reject out-of-band noise), Averaging (multiple measurements average, reduces random noise by sqrt(N)), Adaptive filtering (LMS, NLMS estimate and subtract correlated noise), Spectral subtraction (estimate noise spectrum, subtract from noisy signal spectrum), Wiener filtering (optimal linear filter minimizing MSE), and Wavelet denoising (threshold small wavelet coefficients, preserve large ones). Choice depends on: Signal and noise characteristics, availability of noise reference, real-time requirements, and acceptable artifacts.

Laplace transform analyzes continuous-time systems: X(s) = integral(x(t)*e^(-st)dt), where s = sigma + jw. The s-plane reveals system behavior: Pole locations determine stability (left half-plane = stable), Natural frequencies and damping from pole positions, and s = jw (imaginary axis) gives frequency response. Uses: Solving differential equations (convert to algebra), Transfer function analysis H(s), Stability analysis (root locus, Routh-Hurwitz), and System design (compensator design). Related to Fourier transform when sigma = 0. Foundation for control theory and analog filter design.

All-pass filters have unity magnitude response at all frequencies but non-constant phase: |H(e^jw)| = 1 for all w. Transfer function has poles and zeros in conjugate reciprocal pairs (z0 and 1/z0*). Applications: Phase equalization (compensate nonlinear phase of minimum-phase systems), Group delay equalization (flatten varying group delay), Phase shifters (specific phase shifts at target frequencies), and Building blocks for filter design (lattice/ladder structures). Cascade of all-pass sections can create arbitrary phase response while maintaining magnitude.

Half-band filter is a lowpass FIR filter with cutoff at fs/4 (half of Nyquist frequency). Special property: every other coefficient (except center) is zero, reducing computation by nearly half. Design: For symmetric FIR with odd length, impulse response h[n] = 0 for n = even (except n=0). Results from frequency response symmetry: H(e^jw) + H(e^j(pi-w)) = 1. Applications: Efficient 2x decimation/interpolation, cascade for higher rate changes, and quadrature mirror filter banks. Half-band filters are fundamental building blocks in multirate systems.

Pitch detection estimates fundamental frequency of periodic signals (voice, music). Methods: Autocorrelation (peak in autocorrelation at period lag, robust but octave errors), AMDF (Average Magnitude Difference Function, dip at period lag, similar to autocorrelation), Cepstrum (log spectrum of spectrum reveals period as quefrency peak), YIN algorithm (modified autocorrelation with better accuracy), and PYIN (probabilistic YIN for polyphonic). Challenges: Noise, multiple harmonics, varying amplitude, and real-time requirements. Modern systems often combine methods with post-processing for robustness.

Wavelet transform decomposes signals using scaled and shifted wavelets (localized oscillations) instead of infinite sinusoids. Continuous WT: W(a,b) = integral(x(t)*psi*((t-b)/a)dt). Discrete WT uses dyadic scales: a=2^j, b=k*2^j. Advantages over STFT: Multi-resolution analysis (fine time resolution at high frequencies, fine frequency resolution at low frequencies), Compact support (good for transients), and Efficient computation (fast wavelet transform via filter banks). Applications: Image compression (JPEG2000), denoising, edge detection, and biomedical signals. Wavelet choice (Haar, Daubechies, etc.) affects properties.

Kalman filter is optimal recursive estimator for linear systems with Gaussian noise. Model: State equation x[k] = A*x[k-1] + B*u[k-1] + w[k-1], Observation z[k] = H*x[k] + v[k]. Predict: x_pred = A*x_est + B*u, P_pred = A*P*A' + Q. Update: K = P_pred*H'*(H*P_pred*H' + R)^-1, x_est = x_pred + K*(z - H*x_pred), P = (I - K*H)*P_pred. Design requires: System model (A, B, H), Process noise covariance Q, Measurement noise covariance R. Tuning Q/R ratio affects responsiveness vs smoothing. Extensions: EKF (nonlinear), UKF (unscented), particle filters.

Compressed sensing recovers sparse signals from fewer measurements than Nyquist requires. Theory: If signal x is K-sparse in some basis (K << N nonzero coefficients), it can be recovered from M measurements where M ~ K*log(N/K). Requirements: Sparsity (or compressibility), Incoherent sensing matrix (random measurements), and L1-minimization or matching pursuit for recovery. Applications: MRI (faster scans), radar (reduced data), single-pixel camera, and spectrum sensing. Limitations: Sparsity assumption must hold, computational complexity of recovery algorithms, and noise sensitivity. Active research area bridging information theory and optimization.

BSS recovers original source signals from mixtures without knowing mixing parameters. Model: x = A*s (observations = mixing matrix * sources). Methods: ICA (Independent Component Analysis) maximizes statistical independence using non-Gaussianity measures (kurtosis, negentropy), FastICA algorithm is efficient. PCA-based (maximizes variance, not independence). NMF (Non-negative Matrix Factorization) for non-negative sources. Applications: Cocktail party problem (audio separation), EEG artifact removal, image separation, and communications (MIMO detection). Assumptions: Sources are independent, at most one Gaussian, sufficient observations. Limitations: Order and scaling ambiguity.

Sparse signal processing exploits sparsity (few significant components) for efficiency and performance. Techniques: Matching pursuit and OMP (greedy selection of dictionary atoms), Basis pursuit/LASSO (L1-regularized optimization for sparse solutions), Iterative thresholding (shrinkage for sparse recovery), and Dictionary learning (learn optimal sparse representation). Implementation: Efficient algorithms like ADMM, proximal gradient methods. Applications: Compressed sensing recovery, sparse coding for features, image inpainting, and audio source separation. Trade-offs: Computational complexity vs optimality vs sparsity level. Modern deep learning also learns sparse representations.

Non-parametric (classical): Periodogram, Welch, Blackman-Tukey - make no assumptions about signal, general but limited resolution, variance trade-offs. Parametric: Assume signal model (AR, ARMA), estimate model parameters, derive spectrum from model. AR (all-pole) models: Yule-Walker, Burg, covariance methods for parameter estimation. Advantages: Superior frequency resolution for short data, smooth spectra, explicit poles for system identification. Disadvantages: Model order selection critical (AIC, MDL criteria), wrong model causes artifacts, computationally more complex. Parametric preferred for: spectral line detection, system identification, and limited data scenarios.

Beamforming combines signals from sensor arrays to enhance desired direction and suppress others. Delay-and-sum: Align signals by steering delays, sum constructively for target direction. MVDR/Capon: Minimize output power while maintaining unity gain for desired direction, adapts to interference. LCMV: Multiple linear constraints for nulling specific directions. MUSIC/ESPRIT: Subspace methods for high-resolution DOA estimation. Implementation: Spatial covariance estimation, weight calculation, and real-time processing. Challenges: Calibration errors, near-field effects, wideband signals, and computational complexity. Applications: Radar, sonar, communications, microphone arrays, and medical imaging.

FPGA DSP optimization: Parallelism (unroll loops, parallel filter sections, SIMD-like processing), Pipelining (insert registers for high clock rates, increases latency), Resource sharing (time-multiplex expensive operations like multipliers), Fixed-point optimization (minimize bit widths, saturating arithmetic), Memory architecture (block RAM vs distributed, ping-pong buffers), and DSP primitive utilization (MAC units in DSP48 blocks). Trade-offs: Throughput vs latency vs resources. High-level synthesis (HLS) helps, but manual optimization often needed for critical paths. Verification: bit-accurate simulation matching hardware.

Cepstrum is the spectrum of the log spectrum: c[n] = IDFT(log(|DFT(x)|)). Separates convolved components (e.g., source and filter) that appear multiplicative in spectrum but additive in log domain. The 'quefrency' domain reveals periodicities in log spectrum. Applications: Speech analysis (vocal tract and excitation separation), MFCC extraction (mel-frequency cepstral coefficients for speech recognition), Pitch detection (cepstral peak at fundamental period), Echo detection (impulse in cepstrum at echo delay), and Deconvolution (homomorphic filtering). Real cepstrum (real part only) vs complex cepstrum (preserves phase) for different applications.

Advanced TF methods address STFT resolution limitations: Wavelet Transform (multi-resolution, better for transients), Wigner-Ville Distribution (fine resolution but cross-term interference for multi-component signals), Smoothed pseudo-Wigner (reduced cross-terms via smoothing), Cohen's class (generalized TF distributions), Chirplet transform (for chirp signals), Empirical Mode Decomposition (data-driven, no fixed basis), and Synchrosqueezing (sharpened TF representations). Selection criteria: Signal type (transients, chirps, multi-component), cross-term tolerance, computational complexity, and interpretability requirements. Often multiple methods provide complementary insights.

Subspace methods exploit eigenstructure of signal covariance matrix. Principle: Covariance matrix R = signal subspace + noise subspace (orthogonal). MUSIC: Find noise subspace eigenvectors, search for frequencies/angles where steering vector is orthogonal to noise subspace (pseudospectrum peaks). ESPRIT: Uses shift-invariance structure, estimates frequencies from signal subspace rotation. Advantages: Super-resolution (resolves closely spaced components), statistical consistency. Limitations: Requires number of sources known (or estimated via MDL/AIC), assumes uncorrelated sources, and noise must be white. Applications: Radar DOA, spectral analysis, and communications channel estimation.

Real-time audio challenges: Latency (< 10-20ms for live monitoring, FFT block size vs latency), Determinism (consistent processing time, avoid jitter), Buffer management (handle input/output asynchrony, prevent underruns/overruns), Computational efficiency (optimize algorithms, use SIMD, avoid dynamic allocation), Sample rate conversion (between device rates), Thread safety (audio callback vs control threads), and Platform-specific (audio APIs, driver latency). Solutions: Double/triple buffering, lock-free queues, optimized libraries (FFTW, IPP), profiling for worst-case timing. Embedded constraints add memory and power limits.

Higher-order statistics (HOS) extend beyond mean and variance to 3rd order (skewness/bispectrum), 4th order (kurtosis/trispectrum), and beyond. Properties: Preserve phase information (unlike power spectrum), Suppress Gaussian noise (HOS of Gaussian is zero), and Detect nonlinear coupling (quadratic phase coupling in bispectrum). Applications: Non-Gaussianity detection, nonlinear system identification, phase retrieval, texture analysis, and ICA source separation. Estimation challenges: Large variance requiring substantial data, computational complexity. Used when Gaussian/linear assumptions fail and nonlinear/non-Gaussian analysis is needed.

Graph signal processing extends DSP to signals on irregular graph structures. Graph signal: Values on nodes, graph encodes relationships. Graph Fourier Transform: Eigendecomposition of graph Laplacian defines frequency, Graph Laplacian eigenvectors are basis functions. Graph filtering: Multiplication in graph spectral domain. Applications: Social network analysis, sensor networks (irregular placement), brain connectivity, point cloud processing, and recommendation systems. Challenges: Non-uniform sampling on graphs, computational complexity of large graphs, and graph uncertainty. Active research area extending classical DSP concepts to networked and irregular data structures.

Deep learning in signal processing: End-to-end learning (replace traditional pipelines, learn directly from raw signals), Learned representations (autoencoders, embeddings instead of hand-crafted features), Sequence modeling (RNN, LSTM, Transformer for temporal patterns), CNNs for spectrograms (treat as images), and Model-based deep learning (unrolled optimization algorithms with learned parameters). Applications: Speech recognition, audio synthesis, noise reduction, super-resolution, radar processing, and biomedical signal analysis. Advantages: Adapts to data, captures complex patterns. Challenges: Interpretability, data requirements, computational cost, and real-time constraints. Hybrid approaches combine domain knowledge with learning.