Q1Digital Signal Processing

(a) Sampling Theorem and Necessary Condition

The sampling theorem (Nyquist-Shannon sampling theorem) states that a continuous-time signal x(t) that is band-limited (containing no frequency components above a maximum frequency fm) can be completely and uniquely represented by, and perfectly reconstructed from, a sequence of its instantaneous samples taken at a uniform sampling rate fs, provided that fs is at least twice the signal's maximum frequency component: fs >= 2fm. This minimum required sampling rate, 2fm, is called the Nyquist rate.

Derivation/justification via spectrum replication: when a continuous signal x(t) is sampled at rate fs (equivalently, sampling period T=1/fs), the resulting sampled signal's Fourier spectrum consists of the original continuous signal's spectrum X(f), replicated periodically at integer multiples of fs, i.e., the sampled signal's spectrum is (1/T)sum over all integers k of X(f-kfs). If the original signal is band-limited to fm (X(f)=0 for |f|>fm), and if fs>=2*fm, these periodic spectral replicas do not overlap with one another, meaning the original spectrum X(f) can be exactly recovered from the sampled signal's spectrum simply by applying an ideal low-pass filter with cutoff frequency fm (isolating just the central, unshifted replica, k=0, and rejecting all other replicas).

Necessary condition (avoiding aliasing): if instead fs < 2fm, the periodic spectral replicas overlap with the central (original) spectrum, and this overlap causes aliasing - high-frequency components of the original signal become irrecoverably folded down and mixed with lower-frequency components in the sampled signal's spectrum, such that no subsequent filtering operation can separate them back into the original signal, permanently corrupting the sampled representation. This is why the Nyquist condition, fs>=2fm, is not merely a convenient guideline but a strict, necessary mathematical requirement for perfect reconstruction to even be theoretically possible; in practical systems, an anti-aliasing low-pass filter is applied to the continuous signal before sampling specifically to ensure the signal is adequately band-limited relative to the chosen sampling rate, since real-world signals are rarely perfectly band-limited on their own.

(b) Continuous-Time Processing of Discrete-Time Signals

Continuous-time processing of discrete-time signals refers to a signal-processing configuration in which a discrete-time signal is first converted to a continuous-time signal, processed using an analog (continuous-time) system, and then converted back to discrete-time, effectively using continuous-time processing techniques or hardware to implement what is conceptually a discrete-time system operation.

Discrete-to-continuous (D/C) conversion: the input discrete-time sequence x(n) is converted into a continuous-time impulse train xs(t)=sum over n of x(n)*delta(t-nT), which is then passed through an ideal reconstruction (interpolation) low-pass filter to produce a genuine continuous-time signal xc(t), effectively performing ideal band-limited interpolation between the discrete sample values.

Analog processing: this reconstructed continuous-time signal xc(t) is processed by an analog (continuous-time) LTI system with frequency response Hc(jOmega), producing a continuous-time output yc(t).

Continuous-to-discrete (C/D) conversion: the continuous-time output yc(t) is then re-sampled (via an ideal C/D converter/sampler) at the same rate T to produce the final discrete-time output sequence y(n).

Equivalent discrete-time system: provided the intermediate continuous-time signal xc(t) is adequately band-limited (satisfying the sampling theorem relative to the sampling period T used for re-conversion) both before and after the analog processing stage, this entire D/C-analog-C/D chain behaves exactly as an equivalent discrete-time LTI system, with an equivalent discrete-time frequency response H(e^jw) related to the analog system's continuous-time frequency response by H(e^jw) = Hc(jw/T) for |w|<pi (within the fundamental frequency range), a relationship of major practical importance since it allows discrete-time systems to be designed and analyzed by first designing an equivalent, often more mathematically convenient or physically intuitive, continuous-time (analog) system and then converting the design to the discrete-time domain via this equivalence, a technique directly exploited in analog-prototype-based digital filter design methods such as the impulse invariance and bilinear transformation techniques discussed elsewhere in this paper.

Practical Anti-Aliasing Filter Design Considerations

In an ideal theoretical treatment, the sampling theorem simply requires fs>=2fm with equality permitted as a limiting case. In practical instrumentation and communication system design, however, a real anti-aliasing filter can never achieve the ideal brick-wall (infinitely sharp cutoff) response assumed in the theorem - every physically realizable analog low-pass filter has a finite-width transition band between its passband edge and its stopband edge, and some residual gain (however small) persists beyond the nominal cutoff frequency. Consequently, practicing engineers always design the sampling rate with a deliberate guard band (transition margin) between the signal's actual maximum frequency of interest fm and the Nyquist limit fs/2, rather than setting fs exactly equal to 2fm.

Guard-band reasoning: if the anti-aliasing filter's transition band extends from fm up to some higher frequency f_stop before reaching the required stopband attenuation, the sampling frequency must be chosen such that fs/2 >= f_stop (not merely fs/2>=fm), otherwise frequency components within the filter's imperfect transition band - which have only been partially attenuated, not eliminated - will still fold back (alias) into the baseband upon sampling. This is why practical data-acquisition systems commonly sample at rates well above the theoretical Nyquist minimum (a practice sometimes called oversampling), for example sampling audio at 44.1 kHz or 48 kHz for a nominal 20 kHz audio bandwidth, rather than at the bare minimum 40 kHz, thereby allowing a physically realizable anti-aliasing filter with a finite transition band width to still provide adequate stopband attenuation by the time the folding frequency fs/2 is reached.

• Filter order versus transition width trade-off: achieving a narrower transition band (allowing fs to be chosen closer to the theoretical 2*fm minimum) generally requires a higher-order anti-aliasing filter, increasing analog circuit complexity, cost, and the risk of introducing passband ripple, phase distortion, or component-tolerance sensitivity - so the choice of guard band width is itself an engineering trade-off between sampling-rate/data-rate economy and anti-aliasing filter complexity.
• ADC-related practical sampling constraints: beyond the anti-aliasing filter itself, a real analog-to-digital converter (ADC) introduces its own practical limitations that interact with the sampling-rate choice - the ADC's sample-and-hold circuit has finite aperture time (the interval during which the input must remain stable for an accurate conversion), and any jitter (timing uncertainty) in the sampling clock introduces additional noise that grows with the slew rate of the input signal, effectively imposing a further practical upper bound on how high a signal frequency can be accurately digitized at a given clock-jitter specification.
• Finite ADC resolution and quantization noise: since practical ADCs quantize the sampled amplitude to a finite number of bits, the guard-band/oversampling margin is sometimes deliberately made even larger than the anti-aliasing-filter argument alone would require, because oversampling followed by digital decimation filtering and averaging can be used to improve the effective signal-to-quantization-noise ratio beyond what the ADC's raw bit resolution alone would provide - a technique widely used in high-resolution audio and instrumentation ADCs (delta-sigma converters being the extreme case, sampling at rates many times the Nyquist rate).