Darmois-Skitovic theorem and its proof

Theorem 1 (Lévy-Cramer) Let X 1 and X 2 be two independent random variables
and Y = X 1 + X 2 . Then, if Y has a Gaussian distribution, then X 1 and
X 2 will be Gaussian, too.
Recall: The characteristic function of the random variable X is defined as:
and its second characteristic function is:
Φ X (ω) = E { e jωX} (8)
Ψ X (ω) = ln Φ X (ω) (9)
Theorem 2 (Marcinkiewics-Dugué) The only random variables which have
the characteristic functions of the form e p(ω) where p(ω) is a polynomial, are the
constant random values and Gaussian random variables (and hence the degree
of p is less than or equal to 2).
Theorem 3 (Darmois-Skitovic) Let X 1 , . . . , X N be N independent random
variables. Let: {
Y1 = a 1 X 1 + · · · + a N X N
Y 2 = b 1 X 1 + · · · + b N X N
(10)
and suppose that Y 1 and Y 2 are independent. Now, if for an i we have a i b i ≠ 0,
then X i must be Gaussian.
This theorem, which is the base of blind source separation (from it, the separability
of linear instantaneous mixtures is obvious), states that a random variable
which is not Gaussian cannot appears as a summation term in two independent
random variables.
Proof: Without losing the generality, we can assume a i b j − a j b i ≠ 0 for all
i ≠ j (otherwise, we can combine two random variables to define another one,
Guassianity of this random variable, proves the Gaussianity of both, because of
Lévy-Cramer theorem). Now, we write:
{e j(ω1Y1+ω2Y2)}
Φ Y1Y 2
(ω 1 , ω 2 ) = E
= E
{e j ∑ }
i (aiω1+biω2)Xi
= Φ X1 (a 1 ω 1 + b 1 ω 2 )Φ X2 (a 2 ω 1 + b 2 ω 2 ) · · · Φ XN (a N ω 1 + b N ω 2 )
(11)
The last equation arises from the independence of X i ’s. But, independence of
Y 1 and Y 2 implies that:
and hence:
Φ Y1Y 2
(ω 1 , ω 2 ) = Φ Y1 (ω 1 )Φ Y2 (ω 2 ) (12)
Φ X1 (a 1 ω 1 + b 1 ω 2 )Φ X2 (a 2 ω 1 + b 2 ω 2 ) · · · Φ XN (a N ω 1 + b N ω 2 ) = Φ Y1 (ω 1 )Φ Y2 (ω 2 )
(13)
taking the logarithm of the both sides gives us:
Ψ X1 (a 1 ω 1 +b 1 ω 2 )+Ψ X2 (a 2 ω 1 +b 2 ω 2 )+· · ·+Ψ XN (a N ω 1 +b N ω 2 ) = Ψ Y1 (ω 1 )+Ψ Y2 (ω 2 )
(14)
Now if we first move all the term of the left side for them a i b i = 0 to the right
side, and then apply the Lemma 1, we conclude that if for an i, a i b i ≠ 0, then
Ψ Xi must be a polynomial. Hence, from Marcinkiewics-Dugué theorem, it must
be a Gaussian random variable.
2