5

Let $f\in L_1(-\infty,\infty)$ be a Lebesgue-summable function on $\mathbb{R}$ and let $x\mapsto e^{\delta|x|}f(x)$ also be Lebesgue-summable on all the real line. I have added the condition that $f\in L_1(-\infty,\infty)$, but I think that if $e^{\delta|x|}f(x)$ is Lebesgue-summable (the only condition explicitly stated in the book, Kolmogorov-Fomin's, p. 430 here) on $\mathbb{R}$ then $f$ also is.

Kolmogorov-Fomin's Элементы теории функций и функционального анализа states (p. 430) that the complex function defined by $$g(z)=\int_{\mathbb{R}}f(x)e^{-ixz}d\mu_x$$where the integral is Lebesgue's, $\mu_x$ is Lebesgue linear measure and $z=\lambda+i\nu$, $\lambda,\nu\in\mathbb{R}$ is a complex variable, is analytic in $\{z\in\mathbb{C}:|\nu|<\delta\}$. I clearly see that the integral is the Fourier transform of $f(x)e^{\nu x}$.

By trying to prove the analiticity of $g$ I have convinced myself that, if $x\mapsto e^{\delta|x|}f(x)$ is Lebesgue-summable on $\mathbb{R}$, then $x\mapsto xf(x)$ also is. If that is correct, then I think we could use the Cauchy-Riemann equations and calculate the two partial derivatives with respect to $\lambda$ by using$^1$ the fact, where I call $F$ the Fourier transform, that $\frac{d}{d\lambda}F[e^{\nu x}f(x)](\lambda)=-iF[xe^{\nu x}f(x)]$. Nevertheless, I have no idea about how to calculate $\frac{\partial}{\partial\nu}\text{Re }g$ and $\frac{\partial}{\partial\nu}\text{Im }g$ (I do not even see why such derivatives exist).

Can anybody prove the analiticity of $g$ in $\{z\in\mathbb{C}:|\nu|<\delta\}$ either by using the Cauchy-Riemann equations or in some other way? I $\infty$-ly thank you!


$^1$To complete "what I have tried" I should say that I get the two following partial derivatives:$$\frac{\partial}{\partial\lambda}\text{Re }g(\lambda)=\int_{\mathbb{R}}xe^{\nu x}[\text{Im }f(x)\cos(x\lambda)-\text{Re }f(x)\sin(x\lambda)]d\mu_x$$ $$\frac{\partial}{\partial\lambda}\text{Im }g(\lambda)=\int_{\mathbb{R}}xe^{\nu x}[\text{Re }f(x)\cos(x\lambda)+\text{Im }f(x)\sin(x\lambda)]d\mu_x$$

2 Answers2

4

Yes, the fact that $h(x) = e^{\delta\lvert x\rvert}f(x)$ is Lebesgue integrable implies that $f$ is also Lebesgue integrable. Since $x\mapsto e^{\delta \lvert x\rvert}$ is continuous and everywhere $\geqslant 1$, $f$ is (Lebesgue) measurable if and only if $h$ is measurable, and since $\lvert f(x)\rvert \leqslant \lvert h(x)\rvert$, the finiteness of $\int \lvert f\rvert\,d\mu$ follows from $\int \lvert h\rvert\,d\mu < +\infty$.

One nice way to prove $g$ analytic uses Morera's theorem. First we note that $\lvert e^{-ixz}\rvert = e^{\operatorname{Re} (-ixz)} = e^{x\cdot \operatorname{Im} z} < e^{\delta\lvert x\rvert}$, so the integrand in the definition of $g(z)$ is dominated by $\lvert h\rvert$ for all eligible $z$. The integrand also depends continuously on $z$, so by the dominated convergence theorem, $g$ is continuous. Now we use Morera's theorem to conclude that $g$ is analytic: Let $\Delta$ be a closed triangle contained in the strip $S = \{ z : \lvert \operatorname{Im} z\rvert < \delta\}$. Then

\begin{align} \int_{\partial \Delta} g(z)\,dz &= \int_{\partial\Delta} \int_\mathbb{R} f(x) e^{-ixz}\,d\mu_x\,dz\\ &= \int_\mathbb{R} f(x) \int_{\partial\Delta} e^{-ixz}\,dz\,d\mu_x \tag{Fubini}\\ &=\int_\mathbb{R} f(x)\cdot 0\,d\mu_x \tag{Goursat}\\ &= 0, \end{align}

so Morera's theorem says that $g$ is analytic in $S$.

Another way uses differentiation under the integral sign. For that, we must see that for every $0 < \delta_0 < \delta$, there is a constant $C < +\infty$ such that

$$\lvert x\cdot e^{-ixz}\rvert \leqslant C\cdot e^{\delta\lvert x\rvert}$$

for all $x\in\mathbb{R}$ and all $z$ with $\lvert \operatorname{Im} z\rvert \leqslant \delta_0$. That follows from the observation that $\lvert x\rvert\cdot e^{-\alpha\lvert x\rvert}$ is bounded for every $\alpha > 0$, where we use $\alpha = \delta - \delta_0$. Then the dominated convergence theorem (or a variant or corollary thereof, depends on how the text is organised) allows differentiation under the integral. One can now verify the Cauchy-Riemann equations in the real form, but it is less cumbersome to check them in the complex form,

$$\frac{\partial g}{\partial \overline{z}} \equiv 0,$$

since we can also move the $\frac{\partial}{\partial\overline{z}}$ operator under the integral sign, and obtain

\begin{align} \frac{\partial g}{\partial\overline{z}}(z) &= \frac{\partial}{\partial\overline{z}}\int_{\mathbb{R}} f(x) e^{-ixz}\,d\mu_x\\ &= \int_{\mathbb{R}} \frac{\partial}{\partial\overline{z}}\left(f(x) e^{-ixz}\right)\,d\mu_x\\ &= \int_{\mathbb{R}} f(x) \frac{\partial}{\partial\overline{z}} e^{-ixz}\,d\mu_x\\ &= \int_\mathbb{R} f(x)\cdot 0\,d\mu_x\\ &= 0, \end{align}

since $z\mapsto e^{-ixz}$ is holomorphic for every fixed $x$.

Daniel Fischer
  • 206,697
  • I know some very very basic facts of complex analysis, including Morera's theorem and Goursat's, but I didn't know that a complex line integral on $\partial\Delta$ can be considered as a Lebesgue integral (which I guess is what you do when you apply Fubini's theorem, by taking the fact -which I'm not able to verify- that $\int_{\partial\Delta}\int_{\mathbb{R}}|f(x)e^{\nu x}|d\mu_x dz$ exists into account). Can these facts be "elementary seen"? And how can we see that, if the integrand depends continuously on $z$, $g$ is continuous by the dominated convergence theorem? – Self-teaching worker Nov 21 '14 at 21:19
  • ...I know it (the dominated convergence theorem) in this form: http://librarum.org/book/10022/156 Could you give me a reference for the dominated convergence theorem allowing differentiation under the integral? It sounds like a very interesting theorem. My textbook states a lot of things and proves only a part of them, often implicitly using unproved non-trivial results that are often considered fundamental... I $\aleph_1$-ly thank you!!! ;-) – Self-teaching worker Nov 21 '14 at 21:20
  • 1
    Complex line integrals are also/can also be interpreted as Lebesgue integrals. You can pull it back to an integration over a real interval via a parametrisation of the contour, $\int_\gamma h(z),dz = \int_0^1 h(\gamma(t))\cdot \gamma'(t),dt$, or you can view it as a Lebesgue integral with respect to the complex measure $dz$. For the application of Fubini's theorem, you need the measurability of the integrand, and the finiteness of $$\int_{\partial\Delta}\int_\mathbb{R} \lvert f(x)e^{-ixz}\rvert,d\mu_x,\lvert dz\rvert,$$ but since $\int_{\mathbb{R}}\lvert f(x)e^{-ixz}\rvert,d\mu_x$ is ... – Daniel Fischer Nov 21 '14 at 21:57
  • 1
    ... bounded independently of $z\in S$ since the integrand is dominated by $\lvert h(x)\rvert$, the finiteness of the length of $\partial\Delta$ yields that. Since the integrand depends continuously on $z$, we have pointwise convergence of the integrands as $z_k\to z$, and since we have the global domination by $\lvert h(x)\rvert$, the dominated convergence theorem asserts the continuity of the integral. For the differentiation under the integral sign, we need that the difference quotients are dominated (by a constant multiple of $\lvert h(x)\rvert$), then the DCT shows that the difference ... – Daniel Fischer Nov 21 '14 at 21:58
  • 1
    ... quotient of the integrals, which is the integral of the difference quotients by linearity, converges to the integral of the partial derivative. Sometimes the conditions for differentiation under the integral sign are explicitly given as a corollary or a second theorem, or even as examples in the statement of the main theorem, since it's an important application of that theorem. – Daniel Fischer Nov 21 '14 at 21:58
  • 1
    @DavideZena Oops, forgot to ping you. – Daniel Fischer Nov 21 '14 at 22:07
  • @DanielFischer I heartily thank you for your answer(s)! I've tried to word a lemma to write as a side note: let $f:X\times[c,d]\to\mathbb{C}$, let $\forall y\in [c,d]\quad f(-,y)\in L_1(X,\mu_x)$, let $\varphi\in L_1(X,\mu_x)$ with $\varphi (X)\in\mathbb{R}_{\ge 0}$ and let $f(x,-)$ be differentiable in $y_0$ for all $x\in X$; if $\exists \delta>0:$$\forall|h|<\delta$ $\forall x\in X\quad$$|\frac{f(x,y_0+h)-f(x,y_0)}{h}|\le\varphi (x)$ then $\frac{\partial}{\partial y}\int_X f(x,y_0)d\mu_x=\int_X\frac{\partial}{\partial y} f(x,y_0)d\mu_x$. Sound correct? $\aleph_1$ thanks! – Self-teaching worker Nov 22 '14 at 13:11
  • 1
    Yes, it's correct. Usually, one cares about the function being differentiable not only at $y_0$ but on a full neighbourhood of $y_0$, and then, rather than dominating the difference quotients, one posits that $$\left\lvert \frac{\partial f}{\partial y}(x,y)\right\rvert \leqslant \varphi(x)$$ for [almost] all $x$, for all $y\in (y_0-\delta,y_0+\delta)$, which is typically simpler to express and check. But if only differentiability of $f(x,,\cdot,)$ in $y_0$ is given, that of course isn't possible. – Daniel Fischer Nov 22 '14 at 13:25
  • Oh, well, I see that "for almost all $x\in X$" can substitute "for all $x\in X$" and "$\forall x\in X$" in my previous statement too. Thank you so much again!!! – Self-teaching worker Nov 22 '14 at 15:02
3

You can directly compute the derivative of $g$ by dominated convergece

$\lim_{h\to 0}\dfrac{g(z+h) - g(z)}{h}=\lim_{h\to 0}\int_{\mathbb{R}}f(x)\dfrac{e^{-ix(z+h)}- e^{-ixz}}{h}d\mu_x$

And we have

$$\dfrac{e^{-ix(z+h)}- e^{-ixz}}{h} = e^{-ixz} \dfrac{e^{-ixh}- 1}{h}$$

Let $z = \lambda + i\nu$ with $|\nu| < \delta$

$$\left|e^{-ixz} \dfrac{e^{-ixh}- 1}{h}\right| \leq e^{|\nu| |x|} \left|\dfrac{e^{-ixh}- 1}{h}\right| \leq e^{|\nu| |x|} \left|\dfrac{e^{|x||h|}- 1}{h}\right| \leq e^{|\nu| |x|} \dfrac{e^{(\delta - |\nu|)|x|}-1}{\delta-|\nu|}, \forall |h| < \delta -|\nu|$$

i.e.

$$\left|e^{-ixz} \dfrac{e^{-ixh}- 1}{h}\right| \leq \dfrac{e^{\delta|x|}-e^{|\nu||x|}}{\delta-|\nu|}, \forall |h| < \delta -|\nu|$$

Then $$\left|f(x)\dfrac{e^{-ix(z+h)}- e^{-ixz}}{h}\right| \leq |f(x)|\dfrac{e^{\delta|x|}-e^{|\nu||x|}}{\delta-|\nu|} \leq |f(x)|\dfrac{e^{\delta|x|}}{\delta-|\nu|}, \forall |h| < \delta -|\nu|$$

The last term at RHS is integrable by your assumption, so you can apply dominated convergence to conclude