Under the Linear Blurring Model - $ f = H \ast u $.
I'm trying to calculate the Euler Lagrange of with respect to $ u $ of the functional:

$$ E \left( u \right) = {\left\| f - H \ast u \right\|}^{2}_{2} $$

I know the answer should be $ H \left( -x \right) \ast \left( f - H \ast u \right) $.
Since is must be equivalent to the discrete solution of Linear Operator.
Namely for $ J \left( u \right) = {\left\| f - A u \right\|}^{2}_{2} $ the functional derivative is given by $ {J}' \left( u \right) = {A}^{T} \left( f - A u \right) $.

This is result of the Adjoint of the Convolution is the mirroring the function - $ \bar{f} \left( x \right) = f \left( - x \right) $.

Still I don't understand how to apply it and how to calculate the derivative with respect to $ u $, a function, under the Convolution Operator.

Thank You.

  • $\begingroup$ What is the domain? How is the convolution $H \ast u$ defined. $\endgroup$ – RRL Dec 14 '14 at 7:38
  • $\begingroup$ Hi, Those are images, hence the domain is $ {R}^{2} $ you can even assume $ {R}^{1} $ as it won't change anything. This is the classic convolution (Like in Filtering by Linear Spatial / Time Invariant Kernel). Thank You. $\endgroup$ – Royi Dec 14 '14 at 7:47


$$G(x) := (H \ast u)(x)=\int_{-\infty}^{\infty}H(x-s)u(s)\,ds,\\\hat{H}(x) := H(-x).$$


$$E(u) = {\left\| f - H \ast u \right\|}^{2}_{2}=\int_{-\infty}^{\infty}\left|f(x)-G(x)\right|^2\,dx\\ = \int_{-\infty}^{\infty}\left|f(x)-\int_{-\infty}^{\infty}H(x-s)u(s)\,ds\right|^2\,dx.$$

The variational (Gateaux) derivative for some admissible function $v$ is

$$D_vE(u)= \left.\frac{d}{d \epsilon}E(u + \epsilon v)\right|_{\epsilon = 0} \\ = \int_{-\infty}^{\infty}2\left[f(x)-G(x)\right]\left[\int_{-\infty}^{\infty}H(x-s)v(s)\,ds\right]\,dx \\ = \int_{-\infty}^{\infty}2\left[\int_{-\infty}^{\infty}H(x-s)\left[f(x)-G(x)\right]\,dx\right]v(s)\,ds \\ = \int_{-\infty}^{\infty}2\left[\int_{-\infty}^{\infty}\hat{H}(s-x)\left[f(x)-G(x)\right]\,dx\right]v(s)\,ds \\ = \int_{-\infty}^{\infty}2\hat{H} \ast (f - H \ast u)(s)v(s)\,ds.$$

A necessary condition for an extremum is that the variational derivative vanish for all admissible $v$:

$$\int_{-\infty}^{\infty}2\hat{H} \ast (f - H \ast u)(s)v(s)\,ds = 0 \\ \implies\hat{H} \ast (f - H \ast u) \equiv 0$$

As an aside, the variational derivative is analagous to the directional derivative of a function defined on a finite-dimensional linear space. A function, $f: \mathbb{R^n}\to \mathbb{R^m}$, is differentiable at a point $a$ if there exists a linear operator $Df[a]: \mathbb{R^n}\to \mathbb{R^m}$ such that for $h \in \mathbb{R^n}$

$$\lim_{||h|| \to 0}\frac{||f(a+h) - f(a) - Df[a]\cdot h||}{||h||} =0.$$

The directional derivative, in the direction $v \in \mathbb{R^n}$, is defined as the element $D_vf(a) \in \mathbb{R^m}$ such that for each component

$$\lim_{t \to 0}t^{-1}[f_i(a+tv) - f_i(a)] = D_vf_i(a).$$

The existence of the directional derivative is necessary but not sufficient for the existence of the derivative operator. However, when both exist

$$D_vf(a) = Df[a] \cdot v.$$

In the setting of this problem, the variational derivative of the functional $E$ is an element of the ambient space: $D_vE[u] \in L_2(\mathbb{R})$ with $u,v \in L_2(\mathbb{R})$. This represents a directional derivative in the "direction" $v$. The variation of $E$ is the linear operator $\delta E[u]: L_2(\mathbb{R}) \to L_2(\mathbb{R}) $ defined as

$$\delta E[u]: v \mapsto \int_{-\infty}^{\infty}2\hat{H} \ast (f - H \ast u)(s)v(s)\,ds = D_vE[u].$$

| cite | improve this answer | |
  • $\begingroup$ @Drazick: To be precise, the variation of the functional $E$ is the linear operator $\delta E[u]: v \mapsto \int_{-\infty}^{\infty}2\hat{H} \ast (f - H \ast u)(s)v(s)\,ds$. The variational derivative is the value $\delta E[u](v).$ $\endgroup$ – RRL Dec 15 '14 at 7:28

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.