in DaPrato/Zabczyk's book "Second Order Partial Differential Equations in Hilbert Spaces", there is a useful proposition (Prop. 1.2.8) about a particular calculation of a Gaussian integral in Hilbert spaces: Take a symmetric operator $M$ in a Hilbert space, a centred Gaussian measure $N_Q$ on $H$ with covariance operator $Q$ and assume $ \langle Q^\frac{1}{2} MQ^\frac{1}{2}u,u\rangle < \langle u , u\rangle$ for all $u\neq 0$. Then for $b\in H$,

$$\int_H\exp\left\{\frac{1}{2}\langle M y,y\rangle + \langle b,y\rangle\right \}N_Q(d y) = \frac{\exp\left\{\frac{1}{2}|(1-Q^\frac{1}{2}MQ^\frac{1}{2})^{-\frac{1}{2}}Q^\frac{1}{2}b|^2\right\}}{\sqrt{\det (1-Q^\frac{1}{2}MQ^\frac{1}{2})}}.$$

My problem is that I need to do a calculation of the form

$$\int_H \exp\left\{\frac{1}{2}\langle My,y\rangle + \langle b_1 + ib_2, y\rangle\right\}N_Q(dy).$$

This is almost the setting where I could use the formula above, but it cannot hold in the literal sense, with replacing $b = b_1 + b_2 i$, as the left hand side of the formula can (and will be) properly complex, whereas the right hand side is always real.

Does anyone know of a generalization of this formula? I did the calculation in 1d and there I get a square on the right hand side (which can be complex) instead of the squared absolute value (which will be real).


$$\int_H \exp\left\{\frac{1}{2}\langle My,y\rangle + \langle b_1 + ib_2, y\rangle\right\}N_Q(dy)=$$ $$= \frac{\exp\left\{\frac{1}{2}\langle b_1+ib_2,Q^{1/2}(1-Q^{1/2}MQ^{1/2})^{-1}Q^{1/2}(b_1+ib_2)\rangle\right\}}{\sqrt{\det (1-Q^\frac{1}{2}MQ^\frac{1}{2})}}$$

the inner product is defined here without complex conjugation:

$$\langle b_1+ib_2,c\rangle=\langle b_1,c\rangle+i\langle b_2,c\rangle$$

so the right-hand-side will be complex if $b_2\neq 0$.

  • $\begingroup$ But how does this address my concern that the l.h.s. will have a nonzero imaginary part and the r.h.s will be purely real? $\endgroup$ – Mercury Bench Jan 25 '17 at 13:32
  • $\begingroup$ added a clarification $\endgroup$ – Carlo Beenakker Jan 25 '17 at 13:36
  • $\begingroup$ Thanks, I will take a look at it and see whether it works out. $\endgroup$ – Mercury Bench Jan 25 '17 at 13:42
  • $\begingroup$ Now I'm confused: Is it $\langle c, b_1+ib_2\rangle = \langle c, b_1 \rangle + i \langle c, b_2\rangle$ or with the other sign? Because for the formula to yield the result I want in the end, I need this. Also, I need $\langle i b, ic\rangle = -\langle b, c\rangle$. So the key is to use a non-positive inner product? How can we justify this in this context? $\endgroup$ – Mercury Bench Jan 25 '17 at 13:50
  • 1
    $\begingroup$ So on the premises stated in my previous comment, I see that your proposal works (as the calculation yields the "correct" result). My problem is that the $\langle \cdot,\cdot\rangle$ inner product is not positively definite anymore (but it needs to be as it is the inner product of a Hilbert space $H$). I suggest to slightly adapt this in the following way: Use the "normal" inner product, but write $\langle b_1 + ib_2, L\cdot (b_1 - ib_2)\rangle$, where $L$ is the complicated operator in the expression above. Then the calculation proceeds as I'd like it too. What do you think? $\endgroup$ – Mercury Bench Jan 25 '17 at 14:14

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.