Auto- and Cross-Correlation and LTI Systems
Autocorrelation, cross-correlation, Wiener-Khinchin, vibroseis, LTI properties, impulse response, convolution, stability and random inputs for CSIR NET.
- 19 min
On this page · 11 sections
Prerequisites
- Continuous and Discrete Signals, Fourier Series and Transforms — impulses, energy, transform pairs and the convolution theorem.
- Joint, Marginal and Conditional Distributions; Characteristic Functions — expectation, covariance and correlation for random variables, used here for noise.
- Earthquakes: Causes, Measurement and Seismicity — the seismometer records that are processed as signals.
The Hook: Shaking the Ground for Seconds
How CSIR NET asks it. Part C Geophysics: compute a short autocorrelation or cross-correlation by hand, find a delay from a correlation peak, decide whether a system is linear, time-invariant, causal or stable, find the output of a system from its impulse response, and use the Wiener–Khinchin relation or the white-noise trick for system identification.
Correlation
Definitions
For real energy signals (transients), with lag \( \tau \):
$$R_{xy}(\tau) = \int_{-\infty}^{\infty} x(t)\,y(t + \tau)\,dt, \qquad R_{xx}(\tau) = \int_{-\infty}^{\infty} x(t)\,x(t + \tau)\,dt .$$For sequences, with integer lag \( k \):
$$r_{xy}[k] = \sum_{n} x[n]\,y[n + k], \qquad r_{xx}[k] = \sum_{n} x[n]\,x[n + k].$$For power signals and stationary random processes the sum or integral is replaced by a time average or an expectation, \( R_{xx}(\tau) = E[x(t)\,x(t + \tau)] \).
Correlation is convolution with one signal reversed:
$$R_{xy}(\tau) = x(-\tau) * y(\tau), \qquad \text{so} \qquad R_{xy} \;\leftrightarrow\; X^{*}(f)\,Y(f)\ \text{(real } x\text{)}.$$Properties of the autocorrelation
| Property | Statement | Why |
|---|---|---|
| Even | \( R_{xx}(-\tau) = R_{xx}(\tau) \) | swap the order of the product |
| Maximum at zero lag | \( \lvert R_{xx}(\tau)\rvert \le R_{xx}(0) \) | Cauchy–Schwarz |
| Zero lag = energy or power | \( R_{xx}(0) = E \) (energy signal), \( = \overline{x^2} \) (power signal) | definition |
| Zero phase | its transform \( \lvert X(f)\rvert^2 \) is real and non-negative | phase of \( x \) cancels |
| Periodic in, periodic out | \( A\cos(\omega t + \phi) \to \tfrac{A^2}{2}\cos\omega\tau \) | keeps frequency, loses phase |
| White noise | \( R(\tau) = \sigma^2\delta(\tau) \) | no memory between samples |
| Uncorrelated sum | \( R_{s+n} = R_{s} + R_{n} \) if \( s \), \( n \) uncorrelated and zero-mean | cross terms vanish |
For the cross-correlation: \( R_{yx}(\tau) = R_{xy}(-\tau) \), it is not even in general, and \( \lvert R_{xy}(\tau)\rvert \le \sqrt{R_{xx}(0)\,R_{yy}(0)} \). That bound gives the normalised correlation coefficient
$$\rho_{xy}(\tau) = \frac{R_{xy}(\tau)}{\sqrt{R_{xx}(0)\,R_{yy}(0)}}, \qquad -1 \le \rho_{xy} \le 1,$$which equals \( \pm 1 \) only when one signal is an exact scaled (and shifted) copy of the other.
The autocorrelation throws away phase
Because \( R_{xx} \leftrightarrow \lvert X\rvert^2 \), every signal with the same amplitude spectrum has the same autocorrelation: a zero-phase wavelet, its minimum-phase equivalent and its time reverse all share one autocorrelation. This is why spiking deconvolution, which uses only the trace autocorrelation, must assume the wavelet’s phase (minimum phase). See Filters, Deconvolution and Fractal Analysis.
Wiener–Khinchin theorem
For a wide-sense stationary random process (constant mean, autocorrelation depending only on lag), the power spectral density is the Fourier transform of the autocorrelation:
$$\boxed{S_{xx}(f) = \int_{-\infty}^{\infty} R_{xx}(\tau)\,e^{-i2\pi f\tau}\,d\tau}, \qquad R_{xx}(0) = \int S_{xx}(f)\,df = \text{total power}.$$The cross-spectrum \( S_{xy}(f) \) is likewise the transform of \( R_{xy}(\tau) \). A random signal has no ordinary Fourier transform (infinite energy), so its spectrum is defined this way. In practice we also assume ergodicity: time averages over one long record equal ensemble averages.
| Process | \( R_{xx}(\tau) \) | \( S_{xx}(f) \) |
|---|---|---|
| White noise | \( \sigma^2\delta(\tau) \) | flat, \( \sigma^2 \) |
| Sinusoid \( A\cos(2\pi f_0 t + \phi) \), random \( \phi \) | \( \tfrac{A^2}{2}\cos 2\pi f_0\tau \) | two lines at \( \pm f_0 \), each \( A^2/4 \) |
| First-order (exponentially correlated) | \( \sigma^2 e^{-a\lvert\tau\rvert} \) | \( \sigma^2\,2a/[a^2 + (2\pi f)^2] \): “red” |
| Band-limited white | sinc-shaped | flat inside the band, zero outside |
Uses in geophysics
| Task | Operation | What you read |
|---|---|---|
| Vibroseis | cross-correlate record with the sweep | sweep autocorrelation (the zero-phase Klauder wavelet) placed at each reflection time |
| Time delay between stations or traces | cross-correlate, find peak lag | delay; with spacing, apparent velocity |
| Residual statics | cross-correlate each trace with a pilot trace | static shift per trace |
| Hidden periodicity (multiples, cyclic beds) | autocorrelation | repeated side peaks at the period |
| Ambient-noise studies | cross-correlate long noise records at two stations | an estimate of the wave propagating between them |
| Well-to-well or log matching | cross-correlate log curves | depth shift |
Vibroseis in one line. Recorded trace \( x = s * e \) (sweep \( s \), Earth response \( e \)). Correlating with \( s \):
$$s \star x = (s \star s) * e = K * e,$$where \( K = R_{ss} \) is the Klauder wavelet, zero-phase, with amplitude spectrum \( \lvert S(f)\rvert^2 \) confined to the sweep band. A wider sweep band gives a narrower \( K \) and better resolution.
Systems
A system maps an input signal to an output signal: a seismometer (ground motion to voltage), an amplifier, a digital filter, or the Earth itself (source pulse to seismogram).
System properties and how to test them
| Property | Definition | Quick test |
|---|---|---|
| Linear | \( a x_1 + b x_2 \to a y_1 + b y_2 \) (superposition) | Zero input must give zero output; no powers, products or added constants of \( x \) |
| Time-invariant | \( x(t - t_0) \to y(t - t_0) \) | Coefficients do not depend on \( t \) or \( n \); no \( x(2t) \), no \( x(-t) \) |
| Causal | output at \( t \) depends on input at \( \le t \) only | no future samples \( x[n + 1] \) |
| Memoryless | output depends only on input at the same instant | \( y = 3x \) yes; \( y[n] = x[n - 1] \) no |
| BIBO stable | every bounded input gives a bounded output | for LTI: \( \sum\lvert h[n]\rvert < \infty \) |
| Invertible | input recoverable from output | \( y = x^2 \) is not (sign lost) |
| System | Linear | Time-invariant | Causal | Stable |
|---|---|---|---|---|
| \( y[n] = 3x[n - 2] \) | yes | yes | yes | yes |
| \( y[n] = n\,x[n] \) | yes | no | yes | no |
| \( y[n] = x[n]^2 \) | no | yes | yes | yes |
| \( y[n] = x[n] + 2 \) | no (affine) | yes | yes | yes |
| \( y(t) = x(2t) \) | yes | no | no | yes |
| \( y[n] = x[-n] \) | yes | no | no | yes |
| \( y[n] = \sum_{k \le n} x[k] \) (accumulator) | yes | yes | yes | no |
| \( y[n] = x[n + 1] - x[n] \) | yes | yes | no | yes |
The accumulator is unstable because a bounded step input gives an output that grows without limit (\( h[n] = u[n] \), \( \sum\lvert h\rvert = \infty \)).
The impulse response and convolution
For an LTI system let \( h[n] \) be the response to \( \delta[n] \). Write the input as a sum of shifted impulses (sifting property), then use linearity and time invariance:
$$x[n] = \sum_k x[k]\,\delta[n - k] \;\Longrightarrow\; \boxed{y[n] = \sum_k x[k]\,h[n - k] = (x * h)[n]}$$$$y(t) = \int_{-\infty}^{\infty} x(\tau)\,h(t - \tau)\,d\tau .$$The impulse response completely characterises an LTI system. Everything else (step response, frequency response, stability, causality) follows from \( h \).
graph LR
X["Input x(t)"] --> H["LTI system
impulse response h(t)
transfer function H(f)"]
H --> Y["Output
y = x * h
Y(f) = X(f) H(f)"]Convolution properties:
- Commutative \( x * h = h * x \); associative \( (x * h_1) * h_2 = x * (h_1 * h_2) \); distributive over addition.
- Identity \( x * \delta = x \); shift \( x * \delta(t - t_0) = x(t - t_0) \).
- Cascade: systems in series have \( h = h_1 * h_2 \), and their order can be swapped. Parallel: \( h = h_1 + h_2 \).
- Length: an \( N \)-sample signal convolved with an \( M \)-sample filter gives \( N + M - 1 \) samples.
- Step response \( s(t) = \int_{-\infty}^{t} h(\tau)\,d\tau \); conversely \( h = ds/dt \).
Causality and stability from \( h \):
$$\text{causal} \iff h[n] = 0 \text{ for } n < 0, \qquad \text{BIBO stable} \iff \sum_n\lvert h[n]\rvert < \infty .$$Example: \( h[n] = (0.5)^n u[n] \) is causal and stable (\( \sum = 2 \)); its steady response to a unit step is \( \sum_{n\ge0}0.5^n = 2 \). \( h[n] = (1.2)^n u[n] \) is causal but unstable.
Frequency response: why sinusoids are special
Feed \( x[n] = e^{i\Omega n} \) into an LTI system:
$$y[n] = \sum_k h[k]\,e^{i\Omega(n - k)} = \underbrace{\left(\sum_k h[k]e^{-i\Omega k}\right)}_{H(e^{i\Omega})} e^{i\Omega n}.$$Complex exponentials are eigenfunctions of LTI systems; the eigenvalue is the frequency response \( H \), the Fourier transform of \( h \). A real sinusoid comes out at the same frequency, scaled by \( \lvert H\rvert \) and shifted in phase by \( \arg H \). An LTI system can only reweight frequencies already present; it cannot create new ones. If a system’s output contains harmonics that were not in the input, it is non-linear.
By the convolution theorem \( Y(f) = X(f)\,H(f) \): in the frequency domain, filtering is multiplication. Amplitude spectra multiply, phase spectra add; in decibels the gains add.
The convolutional model of a seismic trace
graph LR
S["Source wavelet"] --> E["Earth
reflectivity r(t)
plus absorption"]
E --> R["Receiver and
instrument response"]
R --> T["Recorded trace
x = w * r + n"]
N["Noise n(t)"] --> TAll the linear stages combine into one wavelet \( w = s * a * i \) (source, absorption, instrument), so the trace is \( x(t) = w(t) * r(t) + n(t) \). Treating the Earth as LTI is an approximation: absorption makes the wavelet change with travel time (non-stationary), and strong ground motion near a source is non-linear. Seismic processing still begins from this model, and deconvolution tries to undo the \( w * \) part.
Random signals through an LTI system
Let a stationary random input \( x \) (mean \( \mu_x \), autocorrelation \( R_{xx} \)) pass through \( h \). With the correlation convention above:
| Quantity | Relation |
|---|---|
| Output mean | \( \mu_y = \mu_x\sum h[n] = \mu_x H(0) \) |
| Input–output cross-correlation | \( R_{xy} = h * R_{xx} \) |
| Output autocorrelation | \( R_{yy} = R_{hh} * R_{xx} \) |
| Output PSD | \( S_{yy}(f) = \lvert H(f)\rvert^2 S_{xx}(f) \) |
| Cross-spectrum | \( S_{xy}(f) = H(f)\,S_{xx}(f) \) |
White-noise input (\( R_{xx} = \sigma^2\delta \)):
- \( R_{xy}[k] = \sigma^2 h[k] \): cross-correlating output with input measures the impulse response. This is system identification by noise, used for instrument calibration.
- Output variance \( R_{yy}[0] = \sigma^2\sum h[n]^2 \).
- Output autocorrelation \( R_{yy}[k] = \sigma^2\sum_n h[n]\,h[n + k] \): filtering makes white noise coloured (correlated between samples).
Memory Tricks
Don’t Confuse
| Pair | First | Second |
|---|---|---|
| Convolution vs correlation | flip one signal, then slide (\( x * h \)) | slide without flipping (\( x \star y \)); \( \leftrightarrow X^{*}Y \) |
| Autocorrelation vs cross-correlation | even, peak at zero lag | not even; peak at the delay; \( R_{yx}(\tau) = R_{xy}(-\tau) \) |
| Energy spectrum vs PSD | \( \lvert X\rvert^2 \) of a transient | transform of \( R_{xx} \) of a stationary process |
| Linear vs affine | \( y = 3x \) | \( y = 3x + 2 \) (not linear: zero in gives 2 out) |
| Causal vs stable | \( h = 0 \) for \( n < 0 \) | \( \sum\lvert h\rvert < \infty \); independent properties |
| Time-invariant vs memoryless | shift commutes with the system | output depends on the present input only |
| Vibroseis sweep vs Klauder wavelet | long chirp sent into the ground | its autocorrelation, short and zero-phase, seen after correlation |
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 does the autocorrelation of a real energy signal equal at zero lag?
- What is the autocorrelation of zero-mean white noise of variance \( \sigma^2 \)?
- State the condition on the impulse response for an LTI system to be BIBO stable.
Answers
- Its total energy, \( R_{xx}(0) = \int x^2dt \); it is also the maximum of \( R_{xx} \).
- \( \sigma^2\delta(\tau) \) (\( \sigma^2\delta[k] \) for sequences): uncorrelated at every non-zero lag, flat spectrum.
- \( \sum_n\lvert h[n]\rvert < \infty \) (or \( \int\lvert h(t)\rvert dt < \infty \)).
Level 2: Part B Style
Practice Q1. For \( x[n] = [2, -1, 1] \), the normalised autocorrelation at lag 2 is:
(a) \( 1/3 \) (b) \( -1/2 \) (c) \( -1/3 \) (d) \( 2/3 \)
Answer: (a) 1/3
\( r[0] = 4 + 1 + 1 = 6 \); \( r[2] = x[0]x[2] = 2 \). \( \rho[2] = 2/6 = 1/3 \). (b) is the value at lag 1: \( r[1] = 2(-1) + (-1)(1) = -3 \), and \( -3/6 = -1/2 \). (c) has the wrong sign. (d) divides by 3, the number of samples, instead of \( r[0] \).
Practice Q2. Which of the following systems is linear and time-invariant?
(a) \( y[n] = n\,x[n] \) (b) \( y[n] = x[n] + 2 \) (c) \( y[n] = 3x[n - 2] \) (d) \( y[n] = x[n]^2 \)
Answer: (c)
(c) is a scaled delay: superposition holds and the coefficient does not depend on \( n \). (a) is linear but time-varying (the gain \( n \) changes). (b) is affine: zero input gives 2. (d) is non-linear: doubling \( x \) quadruples \( y \).
Practice Q3. Two seismometers 300 m apart on a line record the same arrival. With \( R_{12}(\tau) = \int x_1(t)\,x_2(t + \tau)\,dt \), the cross-correlation peaks at \( \tau = +50 \) ms. The apparent velocity along the line, and which station the wave reached first, are:
(a) 6.0 km/s, station 1 (b) 6.0 km/s, station 2 (c) 0.6 km/s, station 1 (d) 3.0 km/s, station 1
Answer: (a)
A positive peak lag means \( x_2 \) is a delayed copy of \( x_1 \): the wave arrived at station 1 first. Apparent velocity \( = 300/0.050 = 6{,}000 \) m/s \( = 6.0 \) km/s. (b) reverses the sign convention. (c) slips a decimal. (d) treats the lag as a two-way time and doubles it to 100 ms.
Practice Q4. An LTI system has \( h[n] = (0.5)^n u[n] \). Its output after a long time for a unit-step input is:
(a) 0.5 (b) 1 (c) 2 (d) it grows without bound
Answer: (c) 2
The step response is the running sum of \( h \): \( s[n] = \sum_{k=0}^{n}0.5^k \to 1/(1 - 0.5) = 2 \). The system is stable because \( \sum\lvert h\rvert = 2 \) is finite. (d) would be the answer for \( \lvert a\rvert \ge 1 \). (b) is \( h[0] \); (a) is \( h[1] \).
Level 3: Part C Style
Practice Q5. Zero-mean white noise of variance 4 is passed through an FIR filter \( h = [1, 0.5, 0.25] \). Find (i) the output variance, (ii) the input–output cross-correlation \( R_{xy}[1] \), and (iii) the normalised output autocorrelation at lag 1.
(a) 5.25, 2.0, 0.48 (b) 7.0, 2.0, 0.36 (c) 5.25, 0.5, 0.48 (d) 12.25, 2.0, 0.20
Answer: (a) 5.25, 2.0, 0.48
(i) \( \sigma_y^2 = \sigma^2\sum h^2 = 4(1 + 0.25 + 0.0625) = 4 \times 1.3125 = 5.25 \).
(ii) For white input \( R_{xy}[k] = \sigma^2 h[k] \), so \( R_{xy}[1] = 4 \times 0.5 = 2.0 \).
(iii) \( R_{yy}[1] = \sigma^2\sum_n h[n]h[n + 1] = 4(1\times0.5 + 0.5\times0.25) = 4 \times 0.625 = 2.5 \). Normalised: \( 2.5/5.25 \approx 0.476 \approx 0.48 \).
(b) uses \( \sigma^2\sum h = 4 \times 1.75 = 7 \) for the variance. (c) forgets to multiply \( h[1] \) by \( \sigma^2 \). (d) uses \( \sigma^2(\sum h)^2 = 12.25 \). The non-zero lag-1 value shows that filtering has coloured the white noise.
Practice Q6. A reference trace is \( x = [1, 2, 1] \) (samples 0 to 2). A second trace is \( y = [0, 0, 2, 4, 2] \). Using \( r_{xy}[k] = \sum_n x[n]\,y[n + k] \), which statement is correct?
(a) The peak is at \( k = 2 \) with \( \rho = 1 \): \( y \) is \( x \) delayed by 2 samples and doubled (b) The peak is at \( k = 2 \) with \( \rho = 0.5 \), because \( y \) has twice the amplitude (c) The peak is at \( k = -2 \) with \( \rho = 1 \) (d) The peak is at \( k = 3 \), the position of the largest sample of \( y \)
Answer: (a)
\( r_{xy}[2] = 1\cdot2 + 2\cdot4 + 1\cdot2 = 12 \); other lags are smaller (\( r_{xy}[1] = 0 + 4 + 4 = 8 \), \( r_{xy}[3] = 4 + 4 = 8 \)). Energies: \( r_{xx}[0] = 6 \), \( r_{yy}[0] = 4 + 16 + 4 = 24 \). \( \rho = 12/\sqrt{6 \times 24} = 12/12 = 1 \).
The correlation coefficient ignores scale, so a doubled copy still gives \( \rho = 1 \) (b is wrong). (c) uses the opposite lag convention. (d) confuses the peak sample of \( y \) (at index 3, the centre of the pulse) with the lag of best alignment.
Practice Q7. Consider these statements:
- The autocorrelation of \( A\cos(2\pi f_0 t + \phi) \) is \( \tfrac{A^2}{2}\cos(2\pi f_0\tau) \), independent of \( \phi \).
- After vibroseis correlation, each reflection is represented by the zero-phase autocorrelation of the sweep.
- The cascade of two LTI systems gives the same output whichever system comes first.
- A minimum-phase wavelet and its time-reversed version have different autocorrelations.
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. The autocorrelation keeps frequency and power but loses phase.
- True. \( s \star (s * e) = R_{ss} * e \); \( R_{ss} \) (the Klauder wavelet) is symmetric, hence zero-phase.
- True. Convolution is commutative and associative: \( h_1 * h_2 = h_2 * h_1 \).
- False. Time reversal conjugates the spectrum, leaving \( \lvert W\rvert^2 \), and therefore the autocorrelation, unchanged.
Quick Revision Table
| Idea | Fact or formula |
|---|---|
| Cross-correlation | \( R_{xy}(\tau) = \int x(t)\,y(t + \tau)\,dt = x(-\tau) * y(\tau) \leftrightarrow X^{*}Y \) |
| Autocorrelation | even, max at 0, \( R(0) \) = energy or power, zero-phase, \( \leftrightarrow \lvert X\rvert^2 \) |
| Correlation coefficient | \( \rho = R_{xy}/\sqrt{R_{xx}(0)R_{yy}(0)} \in [-1, 1] \) |
| Wiener–Khinchin | PSD = Fourier transform of autocorrelation (WSS process) |
| White noise | \( R = \sigma^2\delta \), flat PSD |
| Vibroseis | correlate with sweep; Klauder wavelet = sweep autocorrelation |
| LTI | superposition + shift invariance; fully described by \( h \) |
| Convolution | \( y[n] = \sum_k x[k]h[n - k] \); length \( N + M - 1 \) |
| Causal / stable | \( h[n] = 0 \) for \( n < 0 \) / \( \sum\lvert h\rvert < \infty \) |
| Eigenfunction | \( e^{i\Omega n} \to H(e^{i\Omega})\,e^{i\Omega n} \); no new frequencies |
| Cascade / parallel | \( h_1 * h_2 \) / \( h_1 + h_2 \) |
| Random input | \( S_{yy} = \lvert H\rvert^2 S_{xx} \); \( R_{xy} = h * R_{xx} \) |
| White input | \( R_{xy}[k] = \sigma^2 h[k] \); \( \sigma_y^2 = \sigma^2\sum h^2 \) |
| Trace model | \( x = w * r + n \) |
Key Takeaways
What’s Next
- Sampling, FFT, Z-Transform and Convolution: how continuous records become samples, how the DFT is computed fast, and how the Z-transform turns convolution into polynomial multiplication.
- Filters, Deconvolution and Fractal Analysis: designing \( h \) on purpose, and undoing the wavelet.
- Noise, Filtering, Shielding and Grounding: the instrument-side view of noise and filters from Physical Sciences.
- Signal Processing: back to the chapter overview.