Convolution, approximate identities, and Young's inequality

On $\mathbb R^d$ with Lebesgue measure define $(f*g)(x)=\int f(x-y)g(y)\,dy$ wherever the integral exists. Young's inequality says that for exponents $1\le p,q,r\le\infty$ with $1+1/r=1/p+1/q$, an $L^p$ input and an $L^q$ kernel produce an $L^r$ convolution with norm at most the product of their norms. The relation between exponents is part of the theorem, not a menu of independently selectable spaces. The familiar integrable-kernel estimate is the case $q=1$, $r=p$.

For that case, view the convolution as an integral of translates of the input and apply the integral triangle inequality. Translation invariance keeps every translated Lᵖ norm unchanged, leaving the kernel's absolute integral as the multiplier. General Young bounds follow from Hölder-type arguments or interpolation, with endpoint cases handled carefully. A signed kernel is controlled by its absolute integral, not its signed mass. Thus a kernel integrating to one may still strongly amplify errors if its positive and negative parts are both large.

Take $k\in L^1(\mathbb R^d)$ with integral one and define $k_\varepsilon(x)=\varepsilon^{-d}k(x/\varepsilon)$ for positive $\varepsilon$. Substitution preserves both signed mass and absolute integral. For every fixed positive radius, the absolute mass outside that radius tends to zero as the scale shrinks, by integrability of the original kernel. These conditions form an approximate identity. Positivity is convenient for interpretation as an average but is not required for this finite-p norm convergence theorem.

For $f\in L^p$ with $1\le p<\infty$, subtract $f$ using the mass-one identity and bound the convolution error by an integral of translation differences. Small translations have small Lᵖ error, and the kernel mass of large translations becomes negligible. Translation continuity follows first for smooth compactly supported functions and then by density. The same argument works uniformly for bounded uniformly continuous functions, using their uniform translation continuity. It does not imply uniform recovery of every bounded measurable function.

An arbitrary integrable approximate identity has the finite-p conclusion above, but a pointwise theorem requires additional control of the kernels and a suitable local property of the input. Common positive radially decreasing kernels recover values at Lebesgue points under standard hypotheses. Do not infer that assertion merely from unit mass and Lᵖ convergence. In particular, values assigned on null sets remain invisible to integrals, and smoothing a jump must be interpreted through nearby averages rather than a demanded isolated value.

On the Fourier side, convolution becomes multiplication when justified, for example for Schwartz functions or integrable inputs with the relevant transform identities. An approximate identity has multipliers tending to one at each fixed frequency, but this pointwise multiplier limit alone does not prove convergence in every norm. Uniform operator bounds and approximation arguments supply the missing control. Conversely a large L¹ kernel norm warns of possible amplification even if its multiplier looks benign at a few sampled frequencies. Both spatial and frequency descriptions need their corresponding estimates.