1
$\begingroup$

Say I have two zero-mean vectors, with a size of N and the translation between them is k. Say the image signal std is $\sigma_s$ and the noise std is given by $\sigma_n$. What is the variance of the correlation value in case the correlation is estimated by Fourier transform, without zero-padding? Is there a way to find the variance for the Pearson-correlation too?

$\endgroup$
  • $\begingroup$ I have suggested an answer but I do not feel comfortable about it $\endgroup$ – Gideon Genadi Kogan Mar 3 at 14:59
0
$\begingroup$

Presented are two vectors of length $l$ with a displacement of $k>0$ between them

$A_i=a_i+ω_i$

$B_i=b_i+ω_i$

where $ω_i\sim N(0,σ_n^2)$ and $a_i\sim N(0,σ_s^2)$. The displacement means that for $k<i<l-k$, $b_{i+k}=a_i$

Our estimation of the circular correlation, without zero-padding, takes the form of

$C_k=\frac{∑_{i=1}^NA_i B_{mod(i-k+m,m)}}{\sqrt{∑_{i=1}^NA_i^2∑_{i=1}^NB_i^2}} \frac{l}{l-k}$

Following derivation

$C_k=\frac{∑_{i=1}^N (a_i+ω_i)(b_{mod(i-k+m,m)}+ω_{mod(i-k+m,m)})}{\sqrt{∑_{i=1}^N(a_i+ω_i)^2∑_{i=1}^N(b_i+ω_i)^2}} \frac{l}{l-k}$

Expected value

The expected value for square of one sample is

$E[(a_i+ω_i )^2 ]=E[a_i^2+a_i ω_i+ω_i^2 ]=σ_s^2+σ_n^2$

Expected value for the circular correlation at the true peak location $E[C_k]=\frac{σ_s^2}{σ_s^2+σ_n^2}$

which means that we use biased estimation of the true peak correlation.

Correlation value not at the true peak location is $E[C_c]=0$

Variance

General expression for the variance of multiplication of two samples is given by $var((a_i+ω_i)(b_i+ω_j))=var(a_i b_i+ω_i b_i+a_i ω_i+ω_i ω_j )$

For the correlated samples

$var(a_ib_i+ω_ib_i+a_iω_i+ω_iω_j)=var(2a_iω_i+ω_iω_j)=4σ_s^2 σ_n^2+σ_n^4$

For the uncorrelated samples

$var(a_ib_i+ω_ib_i+a_iω_i+ω_iω_j)=var(a_ib_i+2a_iω_i+ω_iω_j)=σ_s^4+4σ_s^2σ_n^2+σ_n^4$

Lets neglect the variance of the dynamometer. This leads to the variance of the true peak correlation value

$var(C_k )≈\left(k(σ_s^4+4σ_s^2 σ_n^2+σ_n^4 )+(l-k)(4σ_s^2 σ_n^2+σ_n^4)\right)\left(\frac{l}{l-k}\frac{1}{l(σ_s^2+σ_n^2)}\right)^2$

simplification leads to

$var(C_k)=\frac{kσ_s^4+l(4σ_s^2 σ_n^2+σ_n^4)}{(l-k)^2(σ_s^2+σ_n^2 )^2}$

and for a location which is not the true peak v≠k

$var(C_v)≈l(σ_s^4+4σ_s^2σ_n^2+σ_n^4)\left(\frac{l}{l-v}\frac{1}{l(σ_s^2+σ_n^2)}\right)^2$

here simplification leads to

$var(C_v)=\frac{l(σ_s^4+4σ_s^2σ_n^2+σ_n^4)}{(l-v)^2(σ_s^2+σ_n^2)^2}$

Therefore, the true peak value

$C_k\sim\left(\frac{σ_s^2}{σ_s^2+σ_n^2},\frac{kσ_s^4+l(4σ_s^2 σ_n^2+σ_n^4)}{(l-k)^2(σ_s^2+σ_n^2 )^2}\right)$

and the false peak value

$C_v\sim\left(0,\frac{l(σ_s^4+4σ_s^2σ_n^2+σ_n^4)}{(l-v)^2(σ_s^2+σ_n^2)^2}\right)$

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.