Modulation space

Modulation spaces are a family of Banach spaces defined by the behavior of the short-time Fourier transform with respect to a test function from the Schwartz space. They were originally proposed by Hans Georg Feichtinger and are recognized to be the right kind of function spaces for time-frequency analysis. Feichtinger's algebra, while originally introduced as a new Segal algebra, is identical to a certain modulation space and has become a widely used space of test functions for time-frequency analysis.

Modulation spaces are defined as follows. For $1\leq p,q\leq \infty$, a non-negative function $m(x,\omega )$ on $\mathbb {R} ^{2d}$ and a test function $g\in {\mathcal {S}}(\mathbb {R} ^{d})$, the modulation space $M_{m}^{p,q}(\mathbb {R} ^{d})$ is defined by

$M_{m}^{p,q}(\mathbb {R} ^{d})=\left\{f\in {\mathcal {S}}'(\mathbb {R} ^{d})\ :\ \left(\int _{\mathbb {R} ^{d}}\left(\int _{\mathbb {R} ^{d}}|V_{g}f(x,\omega )|^{p}m(x,\omega )^{p}dx\right)^{q/p}d\omega \right)^{1/q}<\infty \right\}.$

In the above equation, $V_{g}f$ denotes the short-time Fourier transform of $f$ with respect to $g$ evaluated at $(x,\omega )$, namely

Failed to parse (unknown function "\mathscr"): {\displaystyle V_gf(x,\omega)=\int_{\mathbb{R}^d}f(t)\overline{g(t-x)}e^{-2\pi it\cdot \omega}dt=\mathscr{F}^{-1}_{\xi}(\overline{\hat{g}(\xi)}\hat{f}(\xi))).}

In other words, $f\in M_{m}^{p,q}(\mathbb {R} ^{d})$ is equivalent to $V_{g}f\in L_{m}^{p,q}(\mathbb {R} ^{2d})$. The space $M_{m}^{p,q}(\mathbb {R} ^{d})$ is the same, independent of the test function $g\in {\mathcal {S}}(\mathbb {R} ^{d})$ chosen. The canonical choice is a Gaussian.

Feichtinger's algebra

For $p=q=1$ and $m(x,\omega )=1$, the modulation space $M_{m}^{1,1}(\mathbb {R} ^{d})=M^{1}(\mathbb {R} ^{d})$ is known by the name Feichtinger's algebra and often denoted by $S_{0}$ for being the minimal Segal algebra invariant under time-frequency shifts, i.e. combined translation and modulation operators. $M^{1}(\mathbb {R} ^{d})$ is a Banach space embedded in $L^{1}(\mathbb {R} ^{d})\cap C_{0}(\mathbb {R} ^{d})$, and is invariant under the Fourier transform. It is for these and more properties that $M^{1}(\mathbb {R} ^{d})$ is a natural choice of test function space for time-frequency analysis.