Theory of Stochastic Signals/Power-Spectral Density

1. Theory of Stochastic Signals
2. Random Variables with Statistical Dependence
3. Power-Spectral Density

• Previous page
• Next page

• Previous page
• Next page

Wiener-Khintchine Theorem

In the remainder of this paper we restrict ourselves to ergodic processes.  As was shown in the  "last chapter"  the following statements then hold:

• Each individual pattern function  $x_i(t)$  is representative of the entire random process  $\{x_i(t)\}$.
• All time means are thus identical to the corresponding coulter means.
• The auto-correlation function,  which is generally affected by the two time parameters  $t_1$  and  $t_2$,  now depends only on the time difference  $τ = t_2 - t_1$:

$$\varphi_x(t_1,t_2)={\rm E}\big[x(t_{\rm 1})\cdot x(t_{\rm 2})\big] = \varphi_x(\tau)= \int^{+\infty}_{-\infty}x(t)\cdot x(t+\tau)\,{\rm d}t.$$

The auto-correlation function provides quantitative information about the  (linear)  statistical bindings within the ergodic process  $\{x_i(t)\}$  in the time domain.  The equivalent descriptor in the frequency domain is the  "power-spectral density",  often also referred to as the  "power-spectral density".

$\text{Definition:}$  The  »power-spectral density«  $\rm (PSD)$  of an ergodic random process  $\{x_i(t)\}$  is the Fourier transform of the auto-correlation function  $\rm (ACF)$:

$${\it \Phi}_x(f)=\int^{+\infty}_{-\infty}\varphi_x(\tau) \cdot {\rm e}^{- {\rm j\hspace{0.05cm}\cdot \hspace{0.05cm} \pi}\hspace{0.05cm}\cdot \hspace{0.05cm} f \hspace{0.05cm}\cdot \hspace{0.05cm}\tau} {\rm d} \tau. $$

This functional relationship is called the  "Theorem of  $\text{Wiener}$  and  $\text{Khinchin}$".

Similarly,  the auto-correlation function can be computed as the inverse Fourier transform of the power-spectral density  (see section  "Inverse Fourier transform"  in the book  "Signal Representation"):

$$ \varphi_x(\tau)=\int^{+\infty}_{-\infty} {\it \Phi}_x \cdot {\rm e}^{- {\rm j\hspace{0.05cm}\cdot \hspace{0.05cm} \pi}\hspace{0.05cm}\cdot \hspace{0.05cm} f \hspace{0.05cm}\cdot \hspace{0.05cm}\tau} {\rm d} f.$$

• The two equations are directly applicable only if the random process contains neither a DC component nor periodic components.
• Otherwise,  one must proceed according to the specifications given in section  "Power-spectral density with DC component".

Physical interpretation and measurement

The lower chart shows an arrangement for  (approximate)  metrological determination of the power-spectral density  ${\it \Phi}_x(f)$.  The following should be noted in this regard:

• The random signal  $x(t)$  is applied to a  (preferably)  rectangular and  (preferably)  narrowband filter with center frequency  $f$  and bandwidth  $Δf$  where  $Δf$  must be chosen sufficiently small according to the desired frequency resolution.
• The corresponding output signal  $x_f(t)$  is squared and then the mean value is formed over a sufficiently long measurement period  $T_{\rm M}$.  This gives the  "power of  $x_f(t)$"  or the  "power components of  $x(t)$  in the spectral range from  $f - Δf/2$  to  $f + Δf/2$":

To measure the power-spectral density

$$P_{x_f} =\overline{x_f(t)^2}=\frac{1}{T_{\rm M}}\cdot\int^{T_{\rm M}}_{0}x_f^2(t) \hspace{0.1cm}\rm d \it t.$$

• Division by  $Δf$  leads to the power-spectral density  $\rm (PSD)$:

$${{\it \Phi}_{x \rm +}}(f) =\frac{P_{x_f}}{{\rm \Delta} f} \hspace {0.5cm} \Rightarrow \hspace {0.5cm} {\it \Phi}_{x}(f) = \frac{P_{x_f}}{{\rm 2 \cdot \Delta} f}.$$

• ${\it \Phi}_{x+}(f) = 2 \cdot {\it \Phi}_x(f)$  denotes the one-sided PSD defined only for positive frequencies.   For  $f<0$   ⇒   ${\it \Phi}_{x+}(f) = 0$.  In contrast,  for the commonly used two-sided power-spectral density:

$${\it \Phi}_x(-f) = {\it \Phi}_x(f).$$

• While the power  $P_{x_f}$  tends to zero as the bandwidth  $Δf$  becomes smaller,  the power-spectral density remains nearly constant above a sufficiently small value of  $Δf$.  For the exact determination of  ${\it \Phi}_x(f)$  two boundary crossings are necessary:

$${{\it \Phi}_x(f)} = \lim_{{\rm \Delta}f\to 0} \hspace{0.2cm} \lim_{T_{\rm M}\to\infty}\hspace{0.2cm} \frac{1}{{\rm 2 \cdot \Delta}f\cdot T_{\rm M}}\cdot\int^{T_{\rm M}}_{0}x_f^2(t) \hspace{0.1cm} \rm d \it t.$$

$\text{Conclusion:}$ 

• From this physical interpretation it further follows that the power-spectral density is always real and can never become negative.  
• The total power of the random signal  $x(t)$  is then obtained by integration over all spectral components:

$$P_x = \int^{\infty}_{0}{\it \Phi}_{x \rm +}(f) \hspace{0.1cm}{\rm d} f = \int^{+\infty}_{-\infty}{\it \Phi}_x(f)\hspace{0.1cm} {\rm d} f .$$

Reciprocity law of ACF duration and PSD bandwidth

All the  $\text{Fourier transform theorems}$  derived in the book  "Signal Representation"  for deterministic signals can also be applied to

On the  "Reciprocity Theorem"  of ACF and PSD

• the  auto-correlation function  $\rm (ACF)$,  and
• the  power-spectral density  $\rm (PSD)$. 

However,  not all laws yield meaningful results due to the specific properties

• of auto-correlation function  (always real and even)
• and power-spectral density  (always real, even, and non–negative).

We now consider as in the section  "Interpretation of the auto-correlation function"  two different ergodic random processes  $\{x_i(t)\}$  and  $\{y_i(t)\}$  based on

1. two pattern signals  $x(t)$  and  $y(t)$   ⇒   upper sketch,
2. two auto-correlation functions  $φ_x(τ)$  and  $φ_y(τ)$   ⇒   middle sketch,
3. two power-spectral densities  ${\it \Phi}_x(f)$  and  ${\it \Phi}_y(f)$   ⇒   bottom sketch.

Based on these exemplary graphs,  the following statements can be made:

• The areas under the PSD curves are equal   ⇒   the processes  $\{x_i(t)\}$  and  $\{y_i(t)\}$  have the same power:

$${\varphi_x({\rm 0})}\hspace{0.05cm} =\hspace{0.05cm} \int^{+\infty}_{-\infty}{{\it \Phi}_x(f)} \hspace{0.1cm} {\rm d} f \hspace{0.2cm} = \hspace{0.2cm}{\varphi_y({\rm 0})} = \int^{+\infty}_{-\infty}{{\it \Phi}_y(f)} \hspace{0.1cm} {\rm d} f .$$

• The from classical  (deterministic)  system theory well known   $\text{Reciprocity Theorem of time duration and bandwidth}$  also applies here:   A narrow ACF corresponds to a broad PSD and vice versa.
• As a descriptive quantity,  we use here the  »equivalent PSD bandwidth«   $∇f$  $($one speaks  "Nabla-f"$)$, 
  similarly defined as the equivalent ACF duration  $∇τ$  in chapter  "Interpretation of the auto-correlation function":

$${{\rm \nabla} f_x} = \frac {1}{{\it \Phi}_x(f = {\rm 0})} \cdot \int^{+\infty}_{-\infty}{{\it \Phi}_x(f)} \hspace{0.1cm} {\rm d} f, $$

$${ {\rm \nabla} \tau_x} = \frac {\rm 1}{ \varphi_x(\tau = \rm 0)} \cdot \int^{+\infty}_{-\infty}{\varphi_x(\tau )} \hspace{0.1cm} {\rm d} \tau.$$

• With these definitions,  the following basic relationship holds:

$${{\rm \nabla} \tau_x} \cdot {{\rm \nabla} f_x} = 1\hspace{1cm}{\rm resp.}\hspace{1cm} {{\rm \nabla} \tau_y} \cdot {{\rm \nabla} f_y} = 1.$$

$\text{Example 1:}$  We start from the graph at the top of this section:

• The characteristics of the higher frequency signal  $x(t)$  are  $∇τ_x = 0.33\hspace{0.08cm} \rm µs$   and  $∇f_x = 3 \hspace{0.08cm} \rm MHz$.
• The equivalent ACF duration of the signal  $y(t)$  is three times:   $∇τ_y = 1 \hspace{0.08cm} \rm µs$.
• The equivalent PSD bandwidth of the signal  $y(t)$  is thus only  $∇f_y = ∇f_x/3 = 1 \hspace{0.08cm} \rm MHz$.

$\text{General:}$  The product of equivalent ACF duration  ${ {\rm \nabla} \tau_x}$  and equivalent PSD bandwidth  $ { {\rm \nabla} f_x}$  is always "one":

$${ {\rm \nabla} \tau_x} \cdot { {\rm \nabla} f_x} = 1.$$

$\text{Proof:}$  According to the above definitions:

$${ {\rm \nabla} \tau_x} = \frac {\rm 1}{ \varphi_x(\tau = \rm 0)} \cdot \int^{+\infty}_{-\infty}{ \varphi_x(\tau )} \hspace{0.1cm} {\rm d} \tau = \frac { {\it \Phi}_x(f = {\rm 0)} }{ \varphi_x(\tau = \rm 0)},$$

$${ {\rm \nabla} f_x} = \frac {1}{ {\it \Phi}_x(f = {\rm0})} \cdot \int^{+\infty}_{-\infty}{ {\it \Phi}_x(f)} \hspace{0.1cm} {\rm d} f = \frac {\varphi_x(\tau = {\rm 0)} }{ {\it \Phi}_x(f = \rm 0)}.$$

Thus,  the product is equal to  $1$.

q.e.d.

$\text{Example 2:}$  A limiting case of the reciprocity theorem represents the so-called  "White Noise":

• This includes all spectral components  (up to infinity).
• The equivalent PSD bandwidth  $∇f$  is infinite.

The theorem given here states that for the equivalent ACF duration  $∇τ = 0$  must hold   ⇒   »white noise has a Dirac-shaped ACF«.

For more on this topic, see the three-part  (German language)  learning video  "The AWGN channel",  especially the second part.

Power-spectral density with DC component

We assume a DC–free random process  $\{x_i(t)\}$.  Further,  we assume that the process also contains no periodic components.  Then holds:

• The auto-correlation function  $φ_x(τ)$ vanishes  for  $τ → ∞$.
• The power-spectral density  ${\it \Phi}_x(f)$  –  computable as the Fourier transform of  $φ_x(τ)$  –  is both continuous in value and continuous in time,  i.e.,  without discrete components.

We now consider a second random process  $\{y_i(t)\}$,  which differs from the process  $\{x_i(t)\}$  only by an additional DC component  $m_y$:

$$\left\{ y_i (t) \right\} = \left\{ x_i (t) + m_y \right\}.$$

The statistical descriptors of the mean-valued random process  $\{y_i(t)\}$  then have the following properties:

• The limit of the ACF for  $τ → ∞$  is now no longer zero,  but  $m_y^2$.   Throughout the  $τ$–range from  $-∞$  to  $+∞$  the ACF  $φ_y(τ)$  is larger than  $φ_x(τ)$  by  $m_y^2$:

$${\varphi_y ( \tau)} = {\varphi_x ( \tau)} + m_y^2 . $$

• According to the elementary laws of the Fourier transform,  the constant ACF contribution in the PSD leads to a Dirac delta function  $δ(f)$  with weight  $m_y^2$:

$${{\it \Phi}_y ( f)} = {\Phi_x ( f)} + m_y^2 \cdot \delta (f). $$

• More information about the  $\delta$–function can be found in the chapter  "Direct current signal - Limit case of a periodic signal"  of the book "Signal Representation".  Furthermore,  we would like to refer you here to the  (German language)  learning video  "Herleitung und Visualisierung der Diracfunktion"   ⇒   "Derivation and visualization of the Dirac delta function".

Numerical PSD determination

Auto-correlation function and power-spectral density are strictly related via the  $\text{Fourier transform}$.  This relationship also holds for discrete-time ACF representation with the sampling operator  ${\rm A} \{ \varphi_x ( \tau ) \} $,  thus for

$${\rm A} \{ \varphi_x ( \tau ) \} = \varphi_x ( \tau ) \cdot \sum_{k= - \infty}^{\infty} T_{\rm A} \cdot \delta ( \tau - k \cdot T_{\rm A}).$$

The transition from the time domain to the spectral domain can be derived with the following steps:

• The distance  $T_{\rm A}$  of two samples is determined by the absolute bandwidth  $B_x$  $($maximum occurring frequency within the process$)$  via the sampling theorem:

$$T_{\rm A}\le\frac{1}{2B_x}.$$

• The Fourier transform of the discrete-time  (sampled)  auto-correlation function yields an with  ${\rm 1}/T_{\rm A}$  periodic power-spectral density:

$${\rm A} \{ \varphi_x ( \tau ) \} \hspace{0.3cm} \circ\!\!-\!\!\!-\!\!\!-\!\!\bullet\, \hspace{0.3cm} {\rm P} \{{{\it \Phi}_x} ( f) \} = \sum_{\mu = - \infty}^{\infty} {{\it \Phi}_x} ( f - \frac {\mu}{T_{\rm A}}).$$

$\text{Conclusion:}$  Since both  $φ_x(τ)$  and  ${\it \Phi}_x(f)$  are even and real functions,  the following relation holds:

$${\rm P} \{ { {\it \Phi}_x} ( f) \} = T_{\rm A} \cdot \varphi_x ( k = 0) +2 T_{\rm A} \cdot \sum_{k = 1}^{\infty} \varphi_x ( k T_{\rm A}) \cdot {\rm cos}(2{\rm \pi} f k T_{\rm A}).$$

• The power-spectral density  $\rm (PSD)$  of the continuous-time process is obtained from  ${\rm P} \{ { {\it \Phi}_x} ( f) \}$  by bandlimiting to the range  $\vert f \vert ≤ 1/(2T_{\rm A})$.
• In the time domain,  this operation means interpolating the individual ACF samples with the  ${\rm sinc}$ function, where  ${\rm sinc}(x)$  stands for  $\sin(\pi x)/(\pi x)$.

$\text{Example 3:}$  A Gaussian ACF  $φ_x(τ)$  is sampled at distance  $T_{\rm A}$  where the sampling theorem is satisfied:

Discrete-time auto-correlation function,  periodically continued power-spectral density

• The Fourier transform of the discrete-time ACF   ⇒   ${\rm A} \{φ_x(τ) \}$  be the periodically continued PSD   ⇒   ${\rm P} \{ { {\it \Phi}_x} ( f) \}$. 

• This with  ${\rm 1}/T_{\rm A}$  periodic function  ${\rm P} \{ { {\it \Phi}_x} ( f) \}$  is accordingly infinitely extended  (red curve).

• The PSD  ${\it \Phi}_x(f)$  of the continuous-time process  $\{x_i(t)\}$  is obtained by band-limiting to the frequency range  $\vert f \cdot T_{\rm A} \vert ≤ 0.5$,  highlighted in blue in the figure.

Accuracy of the numerical PSD calculation

For the following analysis,  we make the following assumptions:

1. The discrete-time ACF  $φ_x(k \cdot T_{\rm A})$  was determined numerically from  $N$  samples.  
2. As already shown in section  "Accuracy of the numerical ACF calculation",  these values are in error and the errors are correlated if  $N$  was chosen too small.
3. To calculate the periodic power-spectral density  $\rm (PSD)$,  we use only the ACF values  $φ_x(0)$, ... , $φ_x(K \cdot T_{\rm A})$:

$${\rm P} \{{{\it \Phi}_x} ( f) \} = T_{\rm A} \cdot \varphi_x ( k = 0) +2 T_{\rm A} \cdot \sum_{k = 1}^{K} \varphi_x ( k T_{\rm A})\cdot {\rm cos}(2{\rm \pi} f k T_{\rm A}).$$

$\text{Conclusion:}$  The accuracy of the power-spectral density calculation is determined to a strong extent by the parameter  $K$:

• If  $K$  is chosen too small,  the ACF values actually present  $φ_x(k - T_{\rm A})$  with  $k > K$  will not be taken into account.
• If  $K$  is too large,  also such ACF values are considered,  which should actually be zero and are finite only because of the numerical ACF calculation.
• These values are only errors  $($due to a small  $N$  in the ACF calculation$)$ and impair the PSD calculation more than they provide a useful contribution to the result.

$\text{Example 4:}$  We consider here a zero mean process with statistically independent samples.  Thus,  only the ACF value  $φ_x(0) = σ_x^2$  should be different from zero.

Accuracy of numerical PSD calculation

• But if one determines the ACF numerically from only  $N = 1000$  samples,  one obtains finite ACF values even for  $k ≠ 0$.

• The upper figure shows that these erroneous ACF values can be up to  $6\%$  of the maximum value.

• The numerically determined PSD is shown below.  The theoretical  (yellow) curve should be constant for  $\vert f \cdot T_{\rm A} \vert ≤ 0.5$.

• The green and purple curves illustrate how by  $K = 3$  resp.  $K = 10$,  the result is distorted compared to  $K = 0$.

• In this case  $($statistically independent random variables$)$  the error grows monotonically with increasing $K$. 

In contrast,  for a random variable with statistical bindings,  there is an optimal value for  $K$  in each case.

1. If this is chosen too small,  significant bindings are not considered.
2. In contrast,  a too large value leads to oscillations that can only be attributed to erroneous ACF values.

Exercises for the chapter

Exercise 4.12: Power-Spectral Density of a Binary Signal

Exercise 4.12Z: White Gaussian Noise

Exercise 4.13: Gaussian ACF and PSD

Exercise 4.13Z: AMI Code