Oscillatory integral

An oscillatory integral uses cancellation to define a distribution even when its integral is not absolutely convergent. Repeated integration by parts transfers derivatives to the amplitude and exposes decay.

An exponential damping factor makes many otherwise conditionally convergent or divergent oscillatory integrals absolutely convergent. Removing the damping after integration defines an Abel limit when it exists. For the Bessel function $J_1$, ${\int }_0^{\infty }{e}^{-ak}{J}_1(kR)dk=R^{-1}(1-a/\sqrt{a^2+R^2})$, giving the Mestel disk radial field as $a\downarrow 0$. This does not remove an unrelated additive divergence of the potential at low wavenumber.

For an amplitude of symbol order $N$ in $k$ frequency variables, applying a phase integration operator $r>N+k$ times makes the pairing with a test function absolutely integrable. Derivatives of a large-frequency cutoff are supported where $|\theta |\asymp R$ and give an error bounded by $C{R}^{N-r+k}$ times finitely many test derivatives. That error tends to zero, making the resulting distribution independent of the cutoff and of the admissible number of integrations.

For a homogeneous phase function, let $q=|{\nabla }_x\Phi {|}^2+|\theta {|}^2|{\nabla }_{\theta }\Phi {|}^2$ and $L=(iq{)}^{-1}({\nabla }_x\Phi \cdot {\nabla }_x+|\theta {|}^2{\nabla }_{\theta }\Phi \cdot {\nabla }_{\theta })$. Then $L{e}^{i\Phi }=e^{i\Phi }$. The formal transpose of a differential operator $L^t$ lowers the symbol class order by one: the $x$ coefficients have order $-1$, while the order-zero frequency coefficients accompany a frequency derivative.

An oscillatory integral with a degree-one homogeneous phase function, $k$ frequency variables, and oscillatory integral amplitude in fixed symbol class order $N$ has order of a distribution at most any nonnegative integer $r>N+k$. The integer works on every compact set; the continuity constant can vary with the compact set.

An oscillatory integral amplitude multiplies $e^{i\Phi }$ in an oscillatory integral with phase function $\Phi$. A finite symbol class order controls its frequency growth and the order of a distribution obtained by integration by parts.

A closed conic enlargement of an oscillatory integral amplitude support is obtained by taking the closure of its frequency-direction projection in $X\times S^{k-1}$ and lifting that set radially to nonzero frequencies. It records limiting spatial locations and frequency directions, including limits reached only at arbitrarily large frequency. Ordinary support in $X\times {\mathbb{R}}^k$ need not record those limits. This distinction is necessary in singular support bounds for an oscillatory integral.

For a homogeneous phase function $\Phi$ and finite-order symbol class amplitude $a$, the singular support of $I_{\Phi }(a)$ lies in the spatial projection of the closed conic amplitude support intersected with ${\nabla }_{\theta }\Phi =0$. Away from that set, the frequency gradient is uniformly nonzero on compact spatial sets and supported directions. Repeated frequency integration by parts lowers the amplitude order until every desired spatial derivative is absolutely integrable. If the ordinary amplitude support is already conic, it gives the same bound.

In one base and one phase variable, take $\Phi =x\theta$ and $a=x\eta (\theta )/|\theta |$, with an even smooth cutoff function equal to zero near zero and one for large $|\theta |$. This is a symbol of order $-1$. Its oscillatory integral is $-2x\log |x|$ plus a smooth function near zero. Thus it is singular there even though the restricted function $a(0,\cdot )$ has empty support. The joint closed conic support of an oscillatory amplitude retains the limiting base point and its stationary directions. A support-sensitive regularity theorem must use that conic neighborhood information, not merely vanishing of an amplitude at one base point.

For $\Phi (s,t,\theta )=s\theta$, take a smooth symbol that turns on increasingly high frequencies in disjoint spatial bumps centered at $t_j\to 0$. Superpolynomially small bump coefficients keep all symbol class estimates valid. The amplitude vanishes near every finite-frequency point over $t=0$, yet its oscillatory integral has delta singularities at $(0,t_j)$. Closedness of singular support forces a singularity at $(0,0)$ as well. A bound using only ordinary, rather than closed conic, amplitude support can consequently fail.

A phase function is a real smooth function $\Phi (x,\theta )$ on $X\times ({\mathbb{R}}^k\setminus \{0\})$ that is positively homogeneous of degree one in $\theta$ and has nonvanishing total differential. Its critical points in the frequency variable determine where an associated oscillatory integral may be singular.