Filters, Deconvolution and Fractal Analysis
Filter types, Butterworth, FIR and IIR, phase, Wiener filters, spiking and predictive deconvolution, fractal dimension and power-law spectra for CSIR NET.
- 25 min
On this page · 12 sections
Prerequisites
- Auto- and Cross-Correlation and LTI Systems — impulse response, frequency response and the convolutional model \( x = w * r + n \).
- Sampling Theorem, FFT, Z-Transform and Convolution — Z-transforms, poles and zeros, and minimum phase.
- Earthquakes: Causes, Measurement and Seismicity — the Gutenberg–Richter law, which reappears here as a fractal relation.
The Hook: Ringing in Shallow Water
How CSIR NET asks it. Part C Geophysics: filter attenuation in decibels, the response of a running average, filter phase types, the inverse of a short wavelet, the assumptions of spiking deconvolution, the role of prediction lag and prewhitening, and fractal dimension from a box count, a divider count or a spectral slope.
Filters
A filter is an LTI system designed to change the spectrum of a signal on purpose: keep the signal band, reject noise. In the time domain it convolves; in the frequency domain it multiplies, \( Y(f) = H(f)X(f) \).
By pass band
| Type | Passes | Geophysical use |
|---|---|---|
| Low-pass | below \( f_c \) | anti-alias before sampling; regional field in potential-field maps |
| High-pass | above \( f_c \) | removing drift, swell noise, very low-frequency ground roll |
| Band-pass | \( f_1 \)–\( f_2 \) | standard seismic processing step |
| Band-reject / notch | all except a narrow band | 50 Hz power-line interference (mains frequency in India) |
Decibels and specifications
$$\text{gain (dB)} = 20\log_{10}\frac{A_{out}}{A_{in}} = 10\log_{10}\frac{P_{out}}{P_{in}}.$$| dB | Amplitude ratio | Power ratio |
|---|---|---|
| 0 | 1 | 1 |
| −3 | 0.707 | 0.5 (the usual cut-off definition) |
| −6 | 0.5 | 0.25 |
| −20 | 0.1 | 0.01 |
| −40 | 0.01 | \( 10^{-4} \) |
A filter is specified by its pass band (and allowed ripple), stop band (and required attenuation), the transition band between them, and its roll-off, in dB per octave (factor 2 in frequency) or dB per decade (factor 10).
The ideal filter cannot be built
An ideal low-pass filter with \( H = 1 \) for \( \lvert f\rvert < f_c \) and 0 beyond has impulse response
$$h(t) = 2f_c\,\mathrm{sinc}(2f_c t),$$which extends to \( \pm\infty \) and is non-zero before \( t = 0 \): non-causal and infinitely long. Truncating it to a finite operator causes Gibbs ripple in the pass and stop bands. Practical designs taper the response: a trapezoidal pass band specified by four corner frequencies in frequency-domain filtering, or a window applied to the truncated sinc in time-domain (FIR) design.
Butterworth filters
The Butterworth response is maximally flat in the pass band:
$$\boxed{\lvert H(f)\rvert^2 = \frac{1}{1 + (f/f_c)^{2n}}}$$- At \( f = f_c \) the gain is \( -3 \) dB for every order \( n \).
- Far beyond \( f_c \), \( \lvert H\rvert \approx (f_c/f)^n \): roll-off of \( 6n \) dB per octave or \( 20n \) dB per decade.
- Higher order means a sharper cut-off but more phase distortion and longer ringing near the edge.
FIR and IIR filters
| FIR (non-recursive) | IIR (recursive) | |
|---|---|---|
| Equation | \( y[n] = \sum_{k=0}^{M-1} b_k x[n-k] \) | \( y[n] = \sum b_k x[n-k] - \sum_{k\ge1} a_k y[n-k] \) |
| Impulse response | finite (M samples) | infinite |
| Poles | only at \( z = 0 \) | away from origin |
| Stability | always stable | stable only if all poles inside unit circle |
| Phase | exactly linear if coefficients are symmetric | generally non-linear |
| Cost for sharp cut-off | many coefficients | few coefficients |
| Typical example | running average, windowed sinc | digital Butterworth, analog recording filters |
Phase of a filter
| Phase type | Impulse response | Effect on a wavelet |
|---|---|---|
| Zero phase | symmetric about \( t = 0 \), non-causal | no time shift; peaks stay in place. Achieved digitally by filtering forward then backward |
| Linear phase | symmetric about \( t_0 \) | constant delay \( t_0 \) for all frequencies; shape preserved. A symmetric \( M \)-point FIR delays by \( (M-1)/2 \) samples |
| Minimum phase | causal, energy front-loaded | least delay for its amplitude response, but different frequencies are delayed differently: phase distortion. Analog field filters are of this kind |
Two filters to compute by hand
Running average (\( M \)-point, low-pass): \( h = [1, 1, \dots, 1]/M \). For \( M = 3 \):
$$H(\Omega) = \frac{1}{3}\left(e^{i\Omega} + 1 + e^{-i\Omega}\right) = \frac{1 + 2\cos\Omega}{3}\quad\text{(centred form).}$$Gain is 1 at DC and zero at \( \Omega = 2\pi/3 \), that is at \( f = f_s/3 \). An \( M \)-point average has zeros at multiples of \( f_s/M \); between them its side lobes let some high-frequency energy through, and the gain changes sign, so it is a crude low-pass.
First difference (high-pass): \( h = [1, -1] \), \( \lvert H(\Omega)\rvert = 2\lvert\sin(\Omega/2)\rvert \): zero at DC, maximum (2) at Nyquist. It approximates a derivative.
Wiener (least-squares) filters
Suppose we want a filter \( f \) that turns an input \( x \) into something as close as possible to a desired output \( d \), minimising \( \sum_n (d[n] - (f * x)[n])^2 \). Setting derivatives to zero gives the normal equations:
$$\boxed{\sum_{j=0}^{M-1} f_j\,r_{xx}[i - j] = g[i], \quad i = 0, \dots, M-1}$$where \( r_{xx} \) is the autocorrelation of the input and \( g \) the cross-correlation of the desired output with the input. The matrix is Toeplitz (constant along diagonals) and symmetric, so it is solved by the Levinson recursion in about \( M^2 \) operations instead of the \( M^3 \) of general elimination. Choosing \( d \) gives different tools:
| Desired output \( d \) | Filter |
|---|---|
| a spike at lag 0 | spiking (inverse) deconvolution |
| a spike at a chosen delay | delayed-spike filter (useful for non-minimum-phase wavelets) |
| a chosen wavelet | shaping filter (for example, to a zero-phase wavelet) |
| the input itself, \( \alpha \) samples ahead | prediction filter (predictive deconvolution) |
Deconvolution
Deconvolution undoes an earlier convolution. For the trace model
$$x(t) = w(t) * r(t) + n(t),$$the aims are to compress the wavelet \( w \) towards a spike (better temporal resolution) and to remove predictable reverberations and short-period multiples.
graph TD
A["Recorded trace x = w * r + n"] --> B{"Is the wavelet known?"}
B -- "Yes: recorded signature,
vibroseis sweep" --> C["Deterministic deconvolution
design inverse or shaping filter from w"]
B -- "No" --> D["Statistical deconvolution
estimate from trace autocorrelation"]
D --> E["Spiking: lag 1
whitens spectrum"]
D --> F["Predictive: lag α
removes periodic multiples"]
D --> G["Homomorphic:
separate in the cepstrum"]
C --> H["Band-pass after deconvolution
to suppress boosted noise"]
E --> H
F --> HInverse filtering and why phase matters
In the frequency domain the ideal inverse is \( F(f) = 1/W(f) \). Where \( \lvert W\rvert \) is small (outside the source band) this blows up and amplifies noise, so a stabilised form is used:
$$F(f) = \frac{W^{*}(f)}{\lvert W(f)\rvert^2 + \varepsilon}.$$In the time domain, take the dipole \( w = (1, -0.5) \), written \( W(Z) = 1 - 0.5Z \) with \( Z \) a unit delay. Its inverse is a geometric series:
$$\frac{1}{1 - 0.5Z} = 1 + 0.5Z + 0.25Z^2 + 0.125Z^3 + \cdots$$It converges because \( \lvert 0.5\rvert < 1 \): the dipole is minimum phase and its inverse is causal and stable. Truncating to three terms, \( (1, 0.5, 0.25) * (1, -0.5) = (1, 0, 0, -0.125) \): a spike plus a small error at the end.
For the reversed dipole \( (-0.5, 1) \), the series in \( Z \) is \( -2(1 + 2Z + 4Z^2 + \cdots) \), which grows without limit. A maximum-phase wavelet has no causal, stable inverse; its stable inverse runs backwards in time. This is why statistical deconvolution must assume a phase.
Spiking (Wiener) deconvolution
When the wavelet is unknown, spiking deconvolution estimates the inverse from the trace itself. It relies on four assumptions:
- Reflectivity is white (random, uncorrelated), so the trace autocorrelation has the shape of the wavelet autocorrelation.
- The wavelet is minimum phase, so the autocorrelation (which carries no phase) is enough to fix it.
- Noise is small.
- The wavelet is stationary (does not change with time), which absorption violates; in practice the operator is designed in time windows.
The normal equations use the trace autocorrelation on the left and, for a zero-lag spike and minimum-phase wavelet, the right-hand side \( (1, 0, 0, \dots) \) (up to a scale factor). The output has an approximately flat (white) amplitude spectrum within the band.
Prewhitening. A small percentage is added to the zero-lag autocorrelation value, \( r[0] \to r[0](1 + \varepsilon) \), equivalent to adding a little white noise. It keeps the Toeplitz system well conditioned and stops the filter boosting frequencies where the signal is weak. Larger \( \varepsilon \) gives a more stable but less sharp result.
Predictive deconvolution
A prediction filter uses past samples to predict the trace \( \alpha \) samples ahead (\( \alpha \) = prediction lag). Whatever is predictable (the wavelet tail, periodic reverberations) is predicted; subtracting the prediction leaves the unpredictable part, the reflectivity. The prediction-error filter has the form
$$(1, \underbrace{0, \dots, 0}_{\alpha - 1}, -a_0, -a_1, \dots, -a_{M-1}),$$with the \( a_j \) from normal equations whose right-hand side is the autocorrelation shifted by \( \alpha \): \( \sum_j a_j\,r[i - j] = r[i + \alpha] \).
| Prediction lag | Effect |
|---|---|
| \( \alpha = 1 \) sample | identical to spiking deconvolution: maximum whitening, maximum noise boost |
| \( \alpha \) = a few samples (around the first zero crossing of the autocorrelation) | keeps the first part of the wavelet, shortens its tail, less noise boost |
| \( \alpha \) and operator length spanning the multiple period | attacks periodic reverberations and short-period multiples |
Water-layer reverberation. For a water layer of depth \( d \) and velocity \( v \), the reverberation period is the two-way time \( \tau = 2d/v \). With \( d = 150 \) m and \( v = 1{,}500 \) m/s, \( \tau = 0.2 \) s, which is 50 samples at 4 ms. Backus treated this case analytically: with sea-floor reflection coefficient \( R \) and a sea surface of reflection coefficient −1, the reverberation train is undone by the three-point operator \( (1 + RZ^n)^2 = 1 + 2RZ^n + R^2Z^{2n} \), where \( n \) is the two-way water time in samples.
Homomorphic (cepstral) deconvolution
Taking the logarithm of the spectrum turns the product \( W(f)R(f) \) into a sum, \( \log W + \log R \). Its inverse transform, the cepstrum (independent variable called quefrency), often separates the smooth wavelet (low quefrency) from the spiky reflectivity or echoes (high quefrency), which can then be removed by “liftering”. It needs no minimum-phase assumption but is sensitive to noise and to phase unwrapping.
Fractal Analysis
Many geophysical quantities have no characteristic scale: a coastline looks equally wiggly on a map at any zoom, topography is rough at every wavelength, fractures come in all sizes, and earthquakes follow a power law in size. The natural description is a power law, and its exponent is a fractal dimension.
The word “fractal” was coined by Mandelbrot. His paper asking how long the coast of Britain is built on Richardson’s earlier observation that the measured length of a coastline grows as the measuring ruler shrinks.
Self-similar and self-affine
- Self-similar: a part, magnified equally in all directions, looks statistically like the whole. Coastlines, fracture networks, the Koch curve.
- Self-affine: a part looks like the whole only if the axes are stretched by different factors. A topographic profile or a time series: height \( h(x) \) with \( h(\lambda x) \) statistically like \( \lambda^H h(x) \). The exponent \( H \) is the Hurst exponent.
Similarity dimension of exact fractals
If an object is made of \( N \) copies of itself, each scaled by a ratio \( r \):
$$\boxed{D = \frac{\log N}{\log(1/r)}}$$| Object | \( N \) | \( r \) | \( D \) |
|---|---|---|---|
| Line segment | 2 | 1/2 | 1 |
| Square | 4 | 1/2 | 2 |
| Cantor set | 2 | 1/3 | \( \log 2/\log 3 \approx 0.631 \) |
| Koch curve | 4 | 1/3 | \( \log 4/\log 3 \approx 1.262 \) |
| Sierpinski triangle | 3 | 1/2 | \( \log 3/\log 2 \approx 1.585 \) |
| Sierpinski carpet | 8 | 1/3 | \( \log 8/\log 3 \approx 1.893 \) |
| Menger sponge | 20 | 1/3 | \( \log 20/\log 3 \approx 2.727 \) |
A fractal curve has \( 1 < D < 2 \); a fractal surface \( 2 < D < 3 \). The rougher, the higher.
Measuring D on real data
| Method | Measure | Power law | Read D from |
|---|---|---|---|
| Divider (ruler) | length \( L \) with ruler \( r \) | \( L(r) \propto r^{1 - D} \) | slope of \( \log L \) vs \( \log r \) \( = 1 - D \) |
| Box counting | number of boxes \( N \) of size \( r \) that contain the object | \( N(r) \propto r^{-D} \) | slope of \( \log N \) vs \( \log(1/r) \) |
| Number–size | number of objects larger than \( r \) (fragments, faults) | \( N(>r) \propto r^{-D} \) | slope on a log–log plot |
Two scales give \( D \) directly: for box counting, \( D = \dfrac{\log(N_2/N_1)}{\log(r_1/r_2)} \).
Self-affine signals: Hurst exponent and spectral slope
For a self-affine profile or time series, the standard model is fractional Brownian motion (fBm) with Hurst exponent \( 0 < H < 1 \):
- Variance of increments grows as \( \lvert\Delta x\rvert^{2H} \); the semivariogram is \( \gamma(h) \propto h^{2H} \).
- The power spectrum is a power law, \( S(f) \propto f^{-\beta} \), with \( \beta = 2H + 1 \) for a 1-D profile.
- Fractal dimension of the profile (graph of the function): \( D = 2 - H \). For a surface, \( D = 3 - H \).
Combining for a profile:
$$\boxed{D = 2 - H = \frac{5 - \beta}{2}}$$| \( H \) | \( \beta \) | \( D \) | Character |
|---|---|---|---|
| \( \to 0 \) | \( \to 1 \) | \( \to 2 \) | very rough, strongly antipersistent |
| 0.5 | 2 | 1.5 | ordinary Brownian motion: uncorrelated increments |
| 0.8 | 2.6 | 1.2 | persistent: rises tend to be followed by rises; smoother |
| \( \to 1 \) | \( \to 3 \) | \( \to 1 \) | nearly smooth curve |
Reference spectra: white noise \( \beta = 0 \) (flat); flicker or pink noise \( \beta = 1 \); Brownian (red) noise \( \beta = 2 \). White noise itself is not an fBm profile; it is the increment of one (fractional Gaussian noise with \( H = 0.5 \)).
Rescaled range (R/S). For a series of length \( n \), the range of cumulative departures from the mean, divided by the standard deviation, grows as \( n^H \). The method is named after the hydrologist H. E. Hurst, who used it on long records of Nile flows.
Where fractal analysis is used in geophysics
| Data | What is measured | Why it matters |
|---|---|---|
| Topography and bathymetry | spectral slope, \( D \) of profiles | roughness, process signatures, interpolation |
| Well logs | power-law spectra of logs | heterogeneity models for simulating rock properties |
| Fracture and fault maps | box-counting or number–size \( D \) | connectivity, permeability, seismic hazard |
| Earthquake catalogues | \( b \)-value; spatial clustering dimension | \( D = 2b \) link; clustering of epicentres |
| Magnetic and gravity sources | power-law (scaling) source spectra | correcting depth estimates that assume uncorrelated sources |
Memory Tricks
Don’t Confuse
| Pair | First | Second |
|---|---|---|
| −3 dB vs −6 dB | half power (amplitude 0.707) | half amplitude (power 0.25) |
| Zero phase vs minimum phase filter | symmetric, no shift, non-causal | causal, least delay, phase distortion |
| FIR vs IIR | finite \( h \), always stable, can be linear phase | recursive, poles, can be unstable |
| Deterministic vs statistical deconvolution | wavelet known (signature, sweep) | wavelet estimated from trace autocorrelation |
| Spiking vs predictive deconvolution | lag 1: whitens, compresses wavelet | lag \( \alpha > 1 \): removes periodic multiples, keeps wavelet front |
| Prewhitening vs whitening | adding a small constant to \( r[0] \) for stability | flattening the output spectrum (the goal of spiking deconvolution) |
| Self-similar vs self-affine | same scaling in all directions | different scaling for height and distance; described by \( H \) |
| Hurst \( H \) vs dimension \( D \) | smoothness exponent (higher = smoother) | roughness measure (\( D = 2 - H \), higher = rougher) |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper.
Level 1: Foundation
- What is the similarity dimension of the Koch curve?
- Which type of filter phase leaves the time of every peak unchanged?
- Why is a small percentage added to the zero-lag autocorrelation before solving the deconvolution normal equations?
Answers
- \( \log 4/\log 3 \approx 1.262 \): four copies, each scaled by 1/3.
- Zero phase (symmetric impulse response about \( t = 0 \)).
- Prewhitening: it stabilises the Toeplitz system and stops the filter amplifying frequencies where the signal is weak, as if a little white noise were added.
Level 2: Part B Style
Practice Q1. A 3rd-order Butterworth low-pass filter has a cut-off of 40 Hz. Its attenuation at 160 Hz is closest to:
(a) 3 dB (b) 18 dB (c) 36 dB (d) 72 dB
Answer: (c) 36 dB
\( f/f_c = 4 \), so \( \lvert H\rvert^2 = 1/(1 + 4^{6}) = 1/4097 \); attenuation \( = 10\log_{10}4097 \approx 36.1 \) dB. This matches \( 6n = 18 \) dB per octave over two octaves (40 to 80 to 160 Hz). (a) is the value at cut-off. (b) counts only one octave. (d) uses \( 20\log_{10} \) on the power ratio, doubling the answer.
Practice Q2. A centred 3-point running average \( (1, 1, 1)/3 \) is applied to data sampled every 1 ms. The frequency it removes completely is about:
(a) 167 Hz (b) 250 Hz (c) 333 Hz (d) 500 Hz
Answer: (c) 333 Hz
\( H(\Omega) = (1 + 2\cos\Omega)/3 = 0 \) when \( \cos\Omega = -1/2 \), \( \Omega = 2\pi/3 \), i.e. \( f = f_s/3 = 1000/3 \approx 333 \) Hz. (d) is the Nyquist frequency, where the gain is \( (1 - 2)/3 = -1/3 \), not zero. (a) is \( f_s/6 \), where the gain is \( (1 + 1)/3 = 2/3 \).
Practice Q3. The first four coefficients of the causal inverse of the wavelet \( (1, -0.5) \), and its phase type, are:
(a) \( (1, 0.5, 0.25, 0.125) \); minimum phase (b) \( (1, -0.5, 0.25, -0.125) \); minimum phase (c) \( (1, 2, 4, 8) \); maximum phase (d) \( (-2, -4, -8, -16) \); minimum phase
Answer: (a)
\( 1/(1 - 0.5Z) = 1 + 0.5Z + 0.25Z^2 + 0.125Z^3 + \cdots \); it converges because the larger coefficient comes first, so the dipole is minimum phase. (b) is the inverse of \( (1, +0.5) \). (c) and (d) come from expanding the wrong dipole; for \( (-0.5, 1) \) the causal series is \( -2(1 + 2Z + 4Z^2 + \cdots) \) and diverges.
Practice Q4. A fault map is covered with square grids. With 4 km boxes, 150 boxes contain a fault trace; with 2 km boxes, 360 do. The box-counting dimension is closest to:
(a) 0.79 (b) 1.26 (c) 1.58 (d) 2.40
Answer: (b) 1.26
\( D = \dfrac{\log(360/150)}{\log(4/2)} = \dfrac{\log 2.4}{\log 2} \approx 1.263 \). (d) is the ratio 2.4 itself. (a) inverts the formula, \( \log 2/\log 2.4 \approx 0.79 \). (c) is the Sierpinski triangle value, a distractor. A dimension between 1 and 2 is expected for a network of lines that partly fills the plane.
Level 3: Part C Style
Practice Q5. A minimum-phase wavelet \( (1, 0.4) \) is to be compressed by a 2-point least-squares spiking filter (no prewhitening). Which is correct?
(a) Filter \( \approx (0.978, -0.337) \); output \( \approx (0.978, 0.054, -0.135) \) (b) Filter \( = (1, -0.4) \); output \( = (1, 0, -0.16) \), the least-squares optimum (c) Filter \( \approx (0.862, -0.345) \); output \( \approx (0.862, 0, -0.138) \), the least-squares optimum (d) Filter \( \approx (-0.337, 0.978) \); output \( \approx (-0.337, 0.843, 0.391) \)
Answer: (a)
Autocorrelation: \( r[0] = 1^2 + 0.4^2 = 1.16 \), \( r[1] = (1)(0.4) = 0.4 \). Desired output a spike at zero lag, so the right-hand side is \( (w_0, 0) = (1, 0) \).
$$\begin{pmatrix} 1.16 & 0.4 \\ 0.4 & 1.16 \end{pmatrix}\begin{pmatrix} a_0 \\ a_1 \end{pmatrix} = \begin{pmatrix} 1 \\ 0 \end{pmatrix}$$Determinant \( 1.3456 - 0.16 = 1.1856 \); \( a_0 = 1.16/1.1856 \approx 0.978 \), \( a_1 = -0.4/1.1856 \approx -0.337 \).
Output: \( (0.978,\; 0.391 - 0.337,\; -0.135) \approx (0.978, 0.054, -0.135) \). Error energy against the spike \( (1, 0, 0) \): \( 0.022^2 + 0.054^2 + 0.135^2 \approx 0.022 \).
(b) is the truncated exact inverse; its error energy is \( 0.16^2 = 0.0256 \), larger, so it is not the least-squares optimum. (c) scales the truncated inverse by \( 1/r[0] \); its error energy is about 0.038. (d) swaps \( a_0 \) and \( a_1 \).
Practice Q6. The power spectrum of a long topographic profile follows \( S(k) \propto k^{-2.4} \) over two decades of wavenumber. Assuming fractional Brownian motion, which statement is correct?
(a) \( H = 0.7 \), \( D = 1.3 \); the profile is persistent (b) \( H = 0.2 \), \( D = 1.8 \); the profile is antipersistent (c) \( H = 1.2 \), \( D = 0.8 \); the profile is smoother than a line (d) \( H = 0.7 \), \( D = 2.3 \); the profile is a fractal surface
Answer: (a)
For a 1-D fBm profile, \( \beta = 2H + 1 \Rightarrow H = (2.4 - 1)/2 = 0.7 \), and \( D = 2 - H = 1.3 \) (check: \( (5 - 2.4)/2 = 1.3 \)). \( H > 0.5 \) means persistent increments.
(b) uses \( H = (\beta - 2)/2 \). (c) uses \( H = \beta/2 \), which exceeds 1 and is impossible for fBm. (d) uses the surface relation \( D = 3 - H \) for a profile.
Practice Q7. A coastline is measured with dividers. With a 10 km setting the walk takes 100 steps; with a 5 km setting it takes 238 steps. The fractal dimension of the coastline, and its length with a 2.5 km setting if the power law holds, are closest to:
(a) 1.25 and 1,416 km (b) 1.25 and 1,190 km (c) 2.38 and 2,380 km (d) 0.75 and 1,000 km
Answer: (a)
Lengths: \( L(10) = 1{,}000 \) km, \( L(5) = 1{,}190 \) km. Since \( L \propto r^{1 - D} \): \( 1{,}190/1{,}000 = 2^{D - 1} \Rightarrow D - 1 = \log_2 1.19 \approx 0.251 \), so \( D \approx 1.25 \).
Halving the ruler again multiplies the length by the same factor 1.19: \( L(2.5) \approx 1{,}190 \times 1.19 \approx 1{,}416 \) km.
(b) gives the 5 km length again. (c) takes the step ratio as \( D \). (d) inverts the sign of the slope.
Practice Q8. A marine trace sampled at 4 ms shows strong reverberations. The sea floor is at 150 m depth and the water velocity is 1,500 m/s. Consider:
- The reverberation period is 0.2 s, or 50 samples.
- The trace autocorrelation will show repeated peaks near multiples of 0.2 s.
- Predictive deconvolution with a 1-sample lag is the same operation as spiking deconvolution.
- Zero-padding the trace before the FFT removes the reverberations.
Which are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2, 3 and 4 only (d) all four
Answer: (b) 1, 2 and 3 only
- True. \( \tau = 2d/v = 300/1500 = 0.2 \) s; \( 0.2/0.004 = 50 \) samples.
- True. A periodic train produces autocorrelation peaks at the period and its multiples.
- True. With \( \alpha = 1 \), the prediction-error filter is the spiking deconvolution operator.
- False. Zero padding only interpolates the spectrum; it does nothing to periodic multiples.
Quick Revision Table
| Idea | Fact or formula |
|---|---|
| Decibels | amplitude \( 20\log_{10} \), power \( 10\log_{10} \); −3 dB = half power |
| Ideal low-pass | \( h = 2f_c\,\mathrm{sinc}(2f_c t) \): non-causal, infinite; truncation → Gibbs |
| Butterworth | \( \lvert H\rvert^2 = 1/[1 + (f/f_c)^{2n}] \); \( 6n \) dB/octave |
| FIR / IIR | finite, always stable, linear phase if symmetric / recursive, poles, cheaper |
| Phase types | zero (no shift), linear (constant delay), minimum (causal, phase distortion) |
| 3-point average | \( (1 + 2\cos\Omega)/3 \); zero at \( f_s/3 \) |
| Wiener normal equations | \( \sum f_j r[i - j] = g[i] \); Toeplitz; Levinson recursion |
| Trace model | \( x = w * r + n \) |
| Stabilised inverse | \( W^{*}/(\lvert W\rvert^2 + \varepsilon) \) |
| Dipole inverse | \( (1, -a) \to (1, a, a^2, \dots) \) if \( \lvert a\rvert < 1 \) |
| Spiking assumptions | white reflectivity, minimum-phase wavelet, low noise, stationarity |
| Prediction lag | \( \alpha = 1 \): spiking; larger \( \alpha \): multiples, keeps wavelet front |
| Water reverberation | period \( 2d/v \); Backus operator \( (1 + RZ^n)^2 \) |
| Similarity dimension | \( D = \log N/\log(1/r) \): Koch 1.262, Sierpinski 1.585, Cantor 0.631 |
| Divider / box count | \( L \propto r^{1 - D} \) / \( N \propto r^{-D} \) |
| Self-affine profile | \( \beta = 2H + 1 \), \( D = 2 - H = (5 - \beta)/2 \); surface \( D = 3 - H \) |
| Gutenberg–Richter link | \( D = 2b \) |
Key Takeaways
What’s Next
- Signal Processing: Quick Revision: every formula and trap from the chapter on one page.
- Noise, Filtering, Shielding and Grounding: analog filters and noise sources from the instrument side.
- Seismology Basics and Earth’s Internal Structure: the wave physics behind the traces these methods process.
- Signal Processing: back to the chapter overview.