Mixed Modulation Spaces

• Mixed modulation spaces are function spaces defined by measuring the short-time Fourier transform in mixed phase-space norms to capture time-frequency concentration.
• They include variants such as permutation-dependent, quasi-Banach, and α-modulation spaces that refine classical modulation frameworks through anisotropic and order-sensitive norms.
• These spaces facilitate discrete decompositions via Gabor frames and underpin sharp kernel theorems and operator characterizations in advanced time-frequency analysis.

Mixed modulation spaces are time–frequency function spaces defined by measuring the short-time Fourier transform (STFT) in mixed phase-space norms. In the standard setting, for a nonzero window $g$, the STFT is

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$

and the classical modulation space $M^{p,q}(\mathbb R^d)$ is obtained by requiring $V_g f\in L^{p,q}(\mathbb R^{2d})$. In broader usage, the subject includes permutation-dependent spaces $M(c)^{p_1,\dots,p_{2d}}$, multi-index mixed quasi-norm spaces $M_w^{\mathbf p}$, mixed-norm $\alpha$-modulation spaces, and modulation spaces associated to tensor products of amalgam spaces; all of them retain the same basic principle that time–frequency concentration is quantified by anisotropic or order-sensitive mixed norms on phase space (Cordero et al., 2017, Toft, 2014, Han et al., 2011, Feichtinger et al., 2020).

1. Foundational definitions and the STFT framework

The core object throughout the theory is the STFT, which identifies a distribution $f$ with a phase-space function $V_g f(x,\xi)$. In the classical mixed-norm case,

$$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$

with the usual modifications when $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$0 or $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$1. The window $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$2 can be changed without changing the space, provided $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$3 is chosen in the admissible class specified in the relevant theory; this window-independence is a basic structural property in both the Banach and quasi-Banach settings (Cordero et al., 2017, Cordero et al., 2023, Toft, 2014).

A more flexible construction replaces the two exponents $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$4 by a full multi-index and, in some formulations, a permutation of the phase-space coordinates. In the permutation-based setting of kernel theorems, one fixes a permutation $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$5 of $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$6, lets $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$7 reorder the coordinates, and defines

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$8

When $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$9 is the identity and all exponents coincide, one recovers the standard space $M^{p,q}(\mathbb R^d)$0; when the exponents split into spatial and frequency blocks, one recovers $M^{p,q}(\mathbb R^d)$1. A distinguished permutation $M^{p,q}(\mathbb R^d)$2 satisfies

$M^{p,q}(\mathbb R^d)$3

so the mixed modulation formalism also contains Wiener amalgam spaces (Cordero et al., 2017).

The quasi-Banach generalization replaces $M^{p,q}(\mathbb R^d)$4 by iterated mixed quasi-norms $M^{p,q}(\mathbb R^d)$5, with $M^{p,q}(\mathbb R^d)$6, possibly after a prescribed permutation of coordinates. In that setting,

$M^{p,q}(\mathbb R^d)$7

and the analysis is formulated in Gelfand–Shilov classes when the weights are more general than polynomially moderate. This extends the classical $M^{p,q}(\mathbb R^d)$8 scale to a genuinely multi-parameter mixed-norm theory (Toft, 2014).

2. Variants of mixed modulation spaces

Across the literature, mixed modulation spaces appear in several structurally distinct but closely related forms. One major variant is the $M^{p,q}(\mathbb R^d)$9-modulation scale $V_g f\in L^{p,q}(\mathbb R^{2d})$0, $V_g f\in L^{p,q}(\mathbb R^{2d})$1, built from an $V_g f\in L^{p,q}(\mathbb R^{2d})$2-covering of frequency space. With associated projections $V_g f\in L^{p,q}(\mathbb R^{2d})$3, the norm is

$V_g f\in L^{p,q}(\mathbb R^{2d})$4

For $V_g f\in L^{p,q}(\mathbb R^{2d})$5 one recovers modulation-space geometry, while $V_g f\in L^{p,q}(\mathbb R^{2d})$6 is treated by convention as the Besov space $V_g f\in L^{p,q}(\mathbb R^{2d})$7. The space is therefore a mixed decomposition scale interpolating geometrically between uniform and dyadic frequency tilings (Han et al., 2011).

A further anisotropic refinement is the mixed-norm $V_g f\in L^{p,q}(\mathbb R^{2d})$8-modulation space

$V_g f\in L^{p,q}(\mathbb R^{2d})$9

where $M(c)^{p_1,\dots,p_{2d}}$0. Here the mixedness is spatially anisotropic: each coordinate direction may carry a different integrability exponent, while the global $M(c)^{p_1,\dots,p_{2d}}$1-summation remains tied to the $M(c)^{p_1,\dots,p_{2d}}$2-dependent frequency covering (Nielsen, 2023).

Another variant starts from a Banach space $M(c)^{p_1,\dots,p_{2d}}$3 on phase space and defines

$M(c)^{p_1,\dots,p_{2d}}$4

When $M(c)^{p_1,\dots,p_{2d}}$5, one recovers $M(c)^{p_1,\dots,p_{2d}}$6. When $M(c)^{p_1,\dots,p_{2d}}$7 is a completed projective or injective tensor product of Wiener amalgam spaces, the resulting modulation space can itself be identified with a Wiener amalgam space. For example, one of the paper’s principal special cases is

$M(c)^{p_1,\dots,p_{2d}}$8

together with the injective analogue

$M(c)^{p_1,\dots,p_{2d}}$9

This identifies a broad family of apparently different mixed modulation constructions with concrete amalgam spaces (Feichtinger et al., 2020).

The special affine Fourier transform generates yet another realization. The SAFT-based spaces $M_w^{\mathbf p}$0 are defined through $M_w^{\mathbf p}$1-convolution and $M_w^{\mathbf p}$2-modulation, but the key structural theorem is that they are chirp-conjugated classical modulation spaces: $M_w^{\mathbf p}$3 This shows that the SAFT theory does not introduce an unrelated scale; it re-encodes mixed modulation spaces in a transform-adapted form (Biswas et al., 2022).

3. Discretization, Gabor frames, and molecular decompositions

Discretization by Gabor analysis is one of the defining technical mechanisms of the subject. If $M_w^{\mathbf p}$4 is a Gabor frame with dual window $M_w^{\mathbf p}$5, then the coefficient and synthesis maps

$M_w^{\mathbf p}$6

satisfy $M_w^{\mathbf p}$7, and the resulting coefficient sequence yields an equivalent discrete description of the modulation norm. For mixed modulation spaces $M_w^{\mathbf p}$8, the Gabor coefficients characterize the norm in mixed $M_w^{\mathbf p}$9 spaces, with unconditional reconstruction when all exponents are finite (Cordero et al., 2017).

This discretization persists in the broad quasi-Banach setting. For $\alpha$0, the analysis operator $\alpha$1 and synthesis operator $\alpha$2 extend continuously between modulation spaces and the corresponding mixed sequence spaces $\alpha$3, and dual Gabor frames provide the reconstruction formula

$\alpha$4

with unconditional convergence when $\alpha$5 and weak-$\alpha$6 convergence otherwise. This result places mixed-norm modulation spaces within a fully reconstructible frame theory rather than a merely abstract norm definition (Toft, 2014).

For mixed-norm $\alpha$7-modulation spaces, the analogous tool is the $\alpha$8-transform, which produces a discrete sequence space $\alpha$9 and a tight frame adapted to the $f$0-covering. The boundedness of the analysis map $f$1, the boundedness of the synthesis map $f$2, and the identity $f$3 provide a discrete decomposition theory parallel to Frazier–Jawerth constructions (Nielsen, 2023).

The molecular formulation refines this further. A family $f$4 is a system of $f$5-molecules if it obeys simultaneous spatial and frequency localization estimates of the form

$f$6

$f$7

The change-of-frame matrix between two such molecular systems is almost diagonal, and the associated almost diagonal matrices form an algebra under composition. This algebraic closure is the basis for perturbation theory, compactly supported frame constructions, and multiplier estimates in mixed-norm $f$8-modulation spaces (Nielsen, 2023).

4. Kernel theorems and operator-theoretic characterizations

The sharpest operator-theoretic use of mixed modulation spaces is the kernel theorem for operators on $f$9. Let $V_g f(x,\xi)$0 be a linear continuous operator on $V_g f(x,\xi)$1 with distribution kernel $V_g f(x,\xi)$2. For $V_g f(x,\xi)$3,

$V_g f(x,\xi)$4

and

$V_g f(x,\xi)$5

From these endpoint statements one deduces

$V_g f(x,\xi)$6

The result completely characterizes bounded operators on $V_g f(x,\xi)$7 by placing the kernel in carefully permuted mixed modulation spaces (Cordero et al., 2017).

The theorem is quantitative: the operator norm is equivalent to the corresponding kernel norm. In particular,

$V_g f(x,\xi)$8

The proof identifies the kernel with a Gabor matrix

$V_g f(x,\xi)$9

and translates operator boundedness into mixed $$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$0 estimates for this matrix. Feichtinger’s kernel theorem appears as the special case

$$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$1

so the mixed formulation is a genuine extension of the earlier theorem (Cordero et al., 2017).

A persistent point of interpretation is that the clean iff-characterization is specific to the diagonal scale $$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$2. For operators on $$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$3, the same paper states that a full kernel characterization analogous to the $$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$4 case is not expected. Instead, it proves sufficient conditions such as

$$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$5

$$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$6

and, by interpolation,

$$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$7

implies boundedness on $$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$8 for all $$\|f\|_{M^{p,q}}=\left(\int_{\mathbb R^d}\left(\int_{\mathbb R^d}|V_g f(x,\xi)|^p\,dx\right)^{q/p}d\xi\right)^{1/q},$$9. This distinction is central: mixed modulation spaces provide robust kernel criteria beyond the classical kernel theorem, but the diagonal and off-diagonal cases behave differently (Cordero et al., 2017).

The same kernel perspective underlies recent boundedness theory for Fourier integral operators. In the study of Schrödinger-type propagators, the kernel

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$00

is estimated in permutation-dependent mixed modulation spaces such as $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$01, while the phase derivatives are measured in working spaces of Wiener amalgam type. The resulting theorems give sharp $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$02-boundedness thresholds in low, mild, critical, and high growth regimes for the phase, with necessity shown by discrete embedding counterexamples. This suggests that mixed modulation spaces have become a primary language for sharp time–frequency kernel estimates of oscillatory operators (Guo et al., 7 Jul 2025).

5. Embeddings, symmetries, and comparison with neighboring scales

Embedding theory makes explicit how mixed norms interact with smoothness and frequency geometry. For classical modulation spaces $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$03, one may write

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$04

which isolates two different mechanisms: local $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$05-control on uniform frequency boxes and global $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$06-summation over the box index. The sharp comparison with Besov, Triebel–Lizorkin, Sobolev, and Fourier $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$07 spaces is governed by the thresholds

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$08

These determine the exact smoothness needed for embeddings into modulation spaces and the maximal smoothness compatible with embeddings out of modulation spaces (Lu, 2021).

Within the $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$09-modulation scale, the corresponding sharp embedding criterion is

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$10

where

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$11

The same paper proves exact complex interpolation and duality, including the formula

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$12

It also shows that the heuristic identification of $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$13-modulation spaces as interpolation spaces between modulation and Besov scales fails in general. This is a notable correction to an oversimplified view of the subject (Han et al., 2011).

Symmetry properties are equally sensitive to mixed structure. For a metaplectic operator $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$14 projecting to $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$15, the covariance formula

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$16

reduces boundedness on $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$17 to boundedness of the pullback $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$18 on $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$19. The precise criterion is that $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$20 is bounded if and only if either $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$21, or $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$22 and $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$23 is upper triangular. The mixed-norm geometry is therefore not invariant under arbitrary symplectic changes of variables when $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$24; only the diagonal case $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$25 enjoys the full metaplectic flexibility stated in that theorem (Cordero et al., 2023).

The same metaplectic viewpoint yields alternative norm characterizations. Shift-invertible metaplectic Wigner distributions $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$26 satisfy

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$27

under the upper-triangularity hypothesis on the relevant linear map, with the triangularity condition dropping in the case $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$28. This places STFTs, $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$29-Wigner distributions, and related phase-space representations within a common mixed-norm invariance theory (Cordero et al., 2023).

6. Operator classes, compactness, and analytical applications

The mixed-norm viewpoint has extensive operator-theoretic consequences. On the functional-analytic side, weighted mixed-norm Lebesgue spaces $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$30 have the metric approximation property under the factorization hypothesis

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$31

and this transfers to weighted modulation spaces $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$32 via Gabor frame representations. As a consequence, Grothendieck’s nuclearity theory applies, and an operator $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$33 is $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$34-nuclear precisely when its kernel admits a decomposition

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$35

with

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$36

For $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$37, the associated trace formula is

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$38

This ties mixed modulation norms directly to trace and spectral theory (Delgado et al., 2014).

Compactness questions can also be phrased in mixed modulation terms. For localization operators $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$39 acting on weighted spaces $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$40 of $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$41-tempered distributions, boundedness follows from $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$42, while compactness follows from a vanishing-at-infinity condition on the STFT of the symbol. In particular, if $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$43, then

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$44

is compact for $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$45. The proof passes through a kernel expansion using a tight Gabor frame and then approximates general symbols by test symbols in $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$46 (Boiti et al., 2019).

Spectral multiplier theory provides another major application. For the Hermite operator

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$47

the multiplier $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$48 is bounded on $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$49 under a localized Sobolev condition of order $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$50 when

$$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$51

and on the diagonal spaces $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$52 under the sharper condition $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$53. The same work states that these mixed-norm conditions are not sharp in the off-diagonal case and uses Heisenberg-group transference and torus multiplier theory to prove the diagonal improvement. It also deduces boundedness of Hermite Riesz transforms in the same mixed range and shows that the wave and Schrödinger propagators for $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$54 preserve modulation-space regularity (Bhimani et al., 2017).

Nonlinear dispersive theory likewise exploits mixed modulation spaces. For the mixed fractional Hartree equation and the Hartree equation with harmonic potential, local and global well-posedness are established in $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$55 and in Fourier amalgam spaces $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$56, based on trilinear Hartree estimates and Strichartz estimates. The paper emphasizes that this yields $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$57- and $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$58-regularity for all $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$59 in the stated regimes, extending earlier Sobolev-based results and admitting low-regularity data (Bhimani et al., 2023).

A final analytical direction is variational. The uncertainty-principle framework based on modulation spaces formulates minimization problems on constraint sets defined by mixed-norm modulation spaces $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$60, proves a mixed-norm extension of Lieb’s inequality for ambiguity functions, and identifies the optimal constant with an extremal eigenvalue of an inverse compact localization operator in the Hilbert case. This connects mixed modulation norms to compact embeddings, localization operators, and Euler–Lagrange equations, including the harmonic-oscillator ground-state equation in the classical one-dimensional case (Dias et al., 2022).

Taken together, these developments show that mixed modulation spaces are not merely a notational extension of $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$61. They form a family of STFT-based spaces in which anisotropy, coordinate ordering, frequency geometry, and operator structure can be encoded with high precision. The literature also makes clear that this precision comes with sharp constraints: full kernel theorems are currently specific to certain scales, metaplectic invariance is restricted in the genuinely mixed case $$V_g f(x,\xi)=\int_{\mathbb R^d} f(t)\,\overline{g(t-x)}\,e^{-2\pi i \xi\cdot t}\,dt,$$62, and interpolation heuristics can fail outside special parameter regimes (Cordero et al., 2017, Cordero et al., 2023, Han et al., 2011).