Fourier Multipliers and the Hörmander-Mikhlin Theorem

Intuition Beginner

Every function can be taken apart into pure frequencies, the way a chord splits into its individual notes. A multiplier is a frequency-by-frequency volume control: a knob setting attached to each frequency that says how much to scale that note before reassembling. Turn the high-frequency knobs down and you blur the function; turn them up and you sharpen it. The collection of all the knob settings is one fixed gain profile, and applying it — scale every frequency by its knob, then add the notes back together — is the operation this unit studies.

Some gain profiles are harmless and some are dangerous. If the loudest knob setting is finite — no frequency gets infinitely amplified — then the operation is safe in the energy sense: it never increases the total energy of a square-integrable signal by more than that loudest setting. That is the easy half of the story, and it is exact. The hard and useful question is different: when does a gain profile stay safe under other ways of measuring size, not just energy? Many natural operations in analysis — recovering a derivative, smoothing, picking out a direction — are multipliers, and we want them to behave well on every reasonable measurement of size, not only the energy one.

The answer is a smoothness rule for the knob profile. If the gain knobs vary gently as you move along the frequency axis — gently in a precise sense that allows the profile to be large or oscillating but forbids it from jerking around faster and faster at high frequencies — then the operation is safe under all the standard size measurements at once. A jerky profile can be catastrophic; a smoothly-varying one is universally tame.

The one-sentence takeaway: a multiplier scales each frequency by a fixed gain knob, it is automatically safe for energy whenever the loudest knob is finite, and a gentle-variation rule on the knob profile upgrades that to safety under every standard way of measuring a function's size.

Visual Beginner

Picture the frequency axis as a long horizontal ruler, with low frequencies near the center and higher ones running outward in both directions. Above the ruler, draw the gain profile: a curve whose height over each frequency is the knob setting there. A flat curve at height one is the do-nothing operation. A curve that climbs is a sharpener; one that sags toward the edges is a smoother. The whole drama is in how this curve behaves far out along the ruler.

The top panel is a good profile. It stays between two horizontal lines, so the loudest knob is finite and the operation is energy-safe at once. More importantly, inside each doubling octave the curve changes by only a bounded amount, and it does not wiggle faster as you go out — this is the gentle-variation rule in picture form. The bottom panel is a bad profile: bounded in height, so still energy-safe, but its wiggles bunch up tighter and tighter at high frequency. That jerkiness is exactly what the smoothness rule forbids, and it is what can break safety under the non-energy size measurements.

Worked example Beginner

We check the easy half — the energy bound — for a concrete gain profile that flips sign.

Step 1. Take the profile that sets the knob to plus one on all positive frequencies and minus one on all negative frequencies. Every knob has size exactly one. The loudest knob setting is therefore one. This is the profile of a famous operation that rotates each frequency's phase, the same shape that underlies recovering a signal's companion.

Step 2. Recall the energy rule for frequencies: the total energy of a function equals the total energy stored across its frequencies (this is the faithfulness of the frequency picture for energy). So if we want the output energy, we can add up the energy in each output frequency.

Step 3. Apply the knobs. At each frequency the output is the input frequency times its knob. Squaring to get energy, the output energy at that frequency is the input energy there times the knob's size squared. Here every knob has size one, so size squared is one: the energy at each frequency is unchanged.

Step 4. Add up. Since no single frequency's energy changed, the total output energy equals the total input energy exactly. So this operation neither grows nor shrinks energy — the energy gain is exactly one, the loudest knob.

What this tells us: when the gain profile is bounded by some number — here that number is one — the operation cannot increase energy by more than that number, because energy multiplies frequency-by-frequency by the knob size squared. This is the whole content of the easy half: a finite loudest knob guarantees an energy-safe operation, with the energy gain equal to that loudest knob. The hard part of the theory is getting the other size measurements to behave, and that needs the profile to be not just bounded but gently varying.

Check your understanding Beginner

Formal definition Intermediate+

Throughout, $\hat{\cdot }$ is the Fourier transform on ${\mathbb{R}}^n$ of 02.10.04, normalized by $\hat{f}(\xi )=\int f(x)e^{-2\pi ix\cdot \xi }dx$, with inverse $g^{\vee }(x)=\int g(\xi )e^{2\pi ix\cdot \xi }d\xi$, and $f$ ranges over the Schwartz class $\mathcal{S}({\mathbb{R}}^n)$ unless stated otherwise.

Definition (Fourier multiplier operator). Given $m\in L^{\infty }({\mathbb{R}}^n)$, the Fourier multiplier operator with symbol $m$ is $$T_{m}f=(m\hat{f}{)}^{\vee },\text{equivalently}\widehat{T_{m}f}(\xi )=m(\xi )\hat{f}(\xi ).$$ $T_m$ is the convolution $T_{m}f=K\ast f$ with kernel $K=m^{\vee }$ (a tempered distribution); $T_m$ is the prototypical translation-invariant operator, commuting with every translation ${\tau }_{h}f(x)=f(x-h)$. The symbol $m$ is also called the multiplier.

Proposition ($L^2$ characterization). $T_m$ extends to a bounded operator on $L^2({\mathbb{R}}^n)$ if and only if $m\in L^{\infty }$, and then $$\Vert T_m{\Vert }_{L^2\to L^2}=\Vert m{\Vert }_{L^{\infty }}.$$ Indeed, by Plancherel 02.10.04, $\Vert T_{m}f{\Vert }_{L^2}=\Vert m\hat{f}{\Vert }_{L^2}\le \Vert m{\Vert }_{\infty }\Vert \hat{f}{\Vert }_{L^2}=\Vert m{\Vert }_{\infty }\Vert f{\Vert }_{L^2}$; equality is approached by concentrating $\hat{f}$ near a point where $|m|$ nears its essential supremum. Boundedness of every translation-invariant $L^2$ operator forces this form: such an operator is a multiplier, and the $L^2$ norm is the sup-norm of its symbol.

Definition (Mikhlin condition). A symbol $m\in C^{\lfloor n/2\rfloor +1}({\mathbb{R}}^n\setminus \{0\})$ satisfies the Mikhlin condition with constant $A$ if $$|{\partial }^{\alpha }m(\xi )|\le A|\xi {|}^{-|\alpha |}\text{for all multi-indices }|\alpha |\le \lfloor \frac{n}{2}\rfloor +1\text{ and all }\xi \ne 0.$$ The bound is scale-invariant: $m$ and its dilate $m(\lambda \cdot )$ satisfy it with the same $A$. The $\alpha =0$ case is plain boundedness $\Vert m{\Vert }_{\infty }\le A$.

Definition (Hörmander condition). A symbol $m$ satisfies the Hörmander condition of order $s>n/2$ if, with $\eta$ a fixed smooth bump supported in the annulus $\{\frac{1}{2}\le |\xi |\le 2\}$, $$\underset{R>0}{\sup }\Vert m(R\cdot )\eta {\Vert }_{H^s({\mathbb{R}}^n)}^2=\underset{R>0}{\sup }\int (1+|\zeta {|}^2{)}^s\left|\right(m(R\cdot )\eta {)}^{\wedge }(\zeta ){|}^{2}d\zeta \le A^2<\infty .$$ This *localized $L^2$-Sobolev* form asks only that each dyadic rescaling of $m$ lie in a fixed Sobolev space $H^s$, $s>n/2$, uniformly in the scale $R$. It is strictly weaker than the Mikhlin condition: pointwise derivative bounds up to order $\lfloor n/2\rfloor +1$ imply membership in $H^s$ for $n/2<s\le \lfloor n/2\rfloor +1$, but the Hörmander condition also admits symbols with only fractional smoothness.

Counterexamples to common slips Intermediate+

• Boundedness of $m$ is not enough for $L^p$, $p\ne 2$. The symbol $m={\chi }_{\{{\xi }_1>0\}}$ (a sharp half-space cutoff) is bounded, so $T_m$ is fine on $L^2$, but the ball multiplier $m={\chi }_{\{|\xi |\le 1\}}$ is unbounded on $L^p({\mathbb{R}}^n)$ for every $p\ne 2$ when $n\ge 2$ (Fefferman's theorem). A bounded symbol can fail every non-energy measurement; the derivative decay is what rescues it.

• The order $\lfloor n/2\rfloor +1$ is dimension-dependent and not an off-by-one cosmetic. In ${\mathbb{R}}^1$ one needs derivatives up to order $1$; in ${\mathbb{R}}^2$ and ${\mathbb{R}}^3$ up to order $2$. Using "up to order $1$" universally fails the kernel estimate in high dimensions — the Sobolev embedding $H^s\hookrightarrow L^{\infty }$ that controls the kernel needs $s>n/2$.

• Scale-invariance is the point, not homogeneity. The Mikhlin bound $|{\partial }^{\alpha }m|\le A|\xi {|}^{-|\alpha |}$ does not say $m$ is homogeneous of degree $0$; it says $m$ behaves like a degree-$0$ symbol under differentiation. Homogeneous degree-$0$ symbols (Riesz transforms) satisfy it, but so do non-homogeneous ones like $(1+|\xi {|}^2{)}^{is/2}$.

• The endpoints $p=1,\infty$ genuinely fail. Even the smoothest non-constant Mikhlin multiplier (the Riesz transform) is unbounded on $L^1$ and $L^{\infty }$; the theorem is for the open range $1<p<\infty$, with the operator norm degenerating like $\max \{p,(p-1{)}^{-1}\}$ at the ends, exactly as for the Calderón-Zygmund theory 02.19.03 it specializes.

Key theorem with proof Intermediate+

Theorem (Hörmander-Mikhlin multiplier theorem). Let $m$ satisfy the Mikhlin condition with constant $A$: $|{\partial }^{\alpha }m(\xi )|\le A|\xi {|}^{-|\alpha |}$ for $|\alpha |\le \lfloor n/2\rfloor +1$. Then $T_m$ is of weak type $(1,1)$ and bounded on $L^p({\mathbb{R}}^n)$ for every $1<p<\infty$, with $$\Vert T_m{\Vert }_{L^p\to L^p}\le C(n)A\max \{p,(p-1{)}^{-1}\}.$$

Proof. The strategy is to verify that the kernel $K=m^{\vee }$ is a Calderón-Zygmund kernel and then invoke the master theorem of 02.19.03. The $\alpha =0$ Mikhlin bound gives $\Vert m{\Vert }_{\infty }\le A$, so by Plancherel 02.10.04 $T_m$ is bounded on $L^2$ with $\Vert T_m{\Vert }_{2\to 2}\le A$ — this is the $L^2$ input. The content is the Hörmander regularity of $K$.

Dyadic decomposition of the symbol. Fix a Littlewood-Paley bump $\eta$ with $\hat{\eta }$ supported in $\{\frac{1}{2}\le |\xi |\le 2\}$ and ${\sum }_{j\in \mathbb{Z}}\hat{\eta }(2^{-j}\xi )=1$ for $\xi \ne 0$ (the partition of 02.20.03). Write $m_j(\xi )=m(\xi )\hat{\eta }(2^{-j}\xi )$, so $m={\sum }_j{m}_j$ off the origin, and let $K_j=m_j^{\vee }$ be the corresponding kernel piece, $K={\sum }_j{K}_j$. Each $m_j$ is supported in $\{|\xi |\sim 2^j\}$.

Per-block kernel estimates. Fix $j$ and rescale to unit frequency: set $\mu (\xi )=m_j(2^j\xi )=m(2^j\xi )\hat{\eta }(\xi )$, supported in the fixed annulus $\{\frac{1}{2}\le |\xi |\le 2\}$. The Mikhlin bounds are scale-invariant, so $|{\partial }^{\alpha }\mu (\xi )|\le CA$ for $|\alpha |\le \lfloor n/2\rfloor +1$ uniformly in $j$ (on the annulus $|\xi |\sim 1$ the weights $|\xi {|}^{-|\alpha |}$ are $\sim 1$, and the Leibniz expansion against $\hat{\eta }$'s bounded derivatives costs only a dimensional constant). Hence $\mu \in C^{\lfloor n/2\rfloor +1}$ with compact support and norm $\le CA$. Two bounds on $\check{\mu }$ follow from Sobolev embedding $H^s\hookrightarrow L^{\infty }$, $s=\lfloor n/2\rfloor +1>n/2$: $$|\check{\mu }(x)|\le CA(1+|x|{)}^{-(\lfloor n/2\rfloor +1)},|\nabla \check{\mu }(x)|\le CA(1+|x|{)}^{-(\lfloor n/2\rfloor +1)},$$ the second because $\xi \mapsto {\xi }_k\mu (\xi )$ has the same regularity and support. Undoing the rescaling, $K_j(x)=2^{jn}\check{\mu }(2^{j}x)$, so by the chain rule $$\int |K_j(x-y)-K_j(x)|dx\le CA\min \{1,2^j|y|\}\cdot \min \{1,(2^j|y|{)}^{-(\lfloor n/2\rfloor +1-n)}\}\le CA\min \{2^j|y|,(2^j|y|{)}^{-1}\}.$$ The mean value theorem gives the factor $2^j|y|$ at small scales (from $|\nabla K_j|$), and the decay of $\check{\mu }$ — using $\lfloor n/2\rfloor +1>n$ when $n=1$ and the $L^1$-integrability of the difference more carefully for $n\ge 2$, where one splits $|x|\lessgtr 2|y|$ — gives the reciprocal factor at large scales.

Summing the Hörmander integral. For fixed $y\ne 0$, sum the per-block estimates over $j\in \mathbb{Z}$, splitting at the scale $2^j\sim |y{|}^{-1}$: $${\int }_{|x|>2|y|}|K(x-y)-K(x)|dx\le \sum _j\int |K_j(x-y)-K_j(x)|dx\le CA\sum _j\min \{2^j|y|,(2^j|y|{)}^{-1}\}\le CA.$$ The two geometric series — ${\sum }_{2^j|y|\le 1}{2}^j|y|$ and ${\sum }_{2^j|y|>1}(2^j|y|{)}^{-1}$ — each sum to a dimensional constant, independent of $y$. So $K$ satisfies the Hörmander regularity condition (2) of 02.19.03 with constant $B=C(n)A$.

Conclusion. $K$ is a Calderón-Zygmund kernel: the $L^2$ bound supplies the cancellation/multiplier hypothesis, and the dyadic sum supplies Hörmander regularity. The Calderón-Zygmund master theorem 02.19.03 yields weak type $(1,1)$ and, by Marcinkiewicz interpolation 02.10.04 against the $(2,2)$ endpoint and duality across $p=2$, strong type $(p,p)$ for $1<p<\infty$ with the stated constant. $\square$

Bridge. This theorem builds toward the entire symbol calculus of pseudodifferential operators and the $L^p$ theory of constant-coefficient PDE, and it appears again in the boundedness of the Riesz transforms, the imaginary powers $(-\Delta {)}^{is}$, and every Fourier-integral estimate that decomposes a symbol dyadically. The foundational reason it works is that the Mikhlin derivative decay, rescaled inside each octave by the scale-invariance of the bound, becomes a uniform $H^s$ estimate on the localized symbol, and through Sobolev embedding $H^s\hookrightarrow L^{\infty }$ this is exactly the kernel regularity the Calderón-Zygmund machinery of 02.19.03 consumes; this is the mechanism by which "smoothness of the symbol" converts into "smoothness of the kernel". The central insight is that a multiplier is a singular integral whose kernel smoothness is read off the symbol one frequency-octave at a time, so the Littlewood-Paley decomposition of 02.20.03 is the bridge: it splits the symbol into pieces each of which is a single-scale bump, and putting these together with the geometric summation generalises the scalar Plancherel $L^2$ bound into the full $L^p$ multiplier criterion.

Exercises Intermediate+

Advanced results Master

Theorem 1 (Hörmander-Mikhlin; Mikhlin 1956 Dokl. Akad. Nauk 109, 701; Hörmander 1960 Acta Math. 104, 93). A symbol with $|{\partial }^{\alpha }m|\le A|\xi {|}^{-|\alpha |}$ for $|\alpha |\le \lfloor n/2\rfloor +1$ defines $T_m$ bounded on $L^p$, $1<p<\infty$, weak type $(1,1)$, with norm $\le C(n)A\max \{p,(p-1{)}^{-1}\}$. The proof feeds $\Vert m{\Vert }_{\infty }\le A$ as the $L^2$ input and converts the scale-invariant derivative decay, dyadically localized, into the Hörmander kernel regularity of 02.19.03 ^{[Mikhlin 1956]}.

Theorem 2 (Hörmander's $L^2$-Sobolev sharpening; Hörmander 1960 Acta Math. 104, 93). The pointwise Mikhlin bounds may be relaxed to the localized Sobolev condition ${\sup }_R\Vert m(R\cdot )\eta {\Vert }_{H^s}\le A$ for any $s>n/2$. This is the minimal smoothness that closes the per-block kernel estimate: $H^s\hookrightarrow L^1$-weighted-kernel needs $s>n/2$, and the integer order $\lfloor n/2\rfloor +1$ is merely the smallest integer exceeding $n/2$. The fractional form admits symbols with no classical derivatives, extending the reach to rough multipliers ^{[Hörmander 1960]}.

Theorem 3 (the Riesz transforms and the Laplacian calculus). The Riesz symbols $-i{\xi }_k/|\xi |$ are degree-$0$ Mikhlin multipliers, hence $L^p$-bounded; they satisfy ${\sum }_k{R}_k^2=-I$ and represent second derivatives, ${\partial }_i{\partial }_j=-R_i{R}_j\Delta$, giving the elliptic estimate $\Vert {\partial }_i{\partial }_{j}u{\Vert }_{L^p}\le C_p\Vert \Delta u{\Vert }_{L^p}$. The whole first-order symbol calculus ${\xi }_k/|\xi |$ and the fractional Laplacians $|\xi {|}^{is}$, $(1+|\xi {|}^2{)}^{-\sigma /2}$ are multiplier-theoretic consequences of Theorem 1 ^{[Stein 1970]}.

Theorem 4 (imaginary powers and Stein analytic interpolation). The family $(-\Delta {)}^{is}$, $s\in \mathbb{R}$, are Mikhlin multipliers with $\Vert (-\Delta {)}^{is}{\Vert }_{L^p\to L^p}\le C(n,p)(1+|s|{)}^{\lfloor n/2\rfloor +1}$, polynomial in $s$. This controlled growth across the imaginary axis is the input to Stein's interpolation theorem for analytic families $(-\Delta {)}^z$, $\mathrm{Re}z\in [0,1]$, which underlies the $L^p$ theory of complex powers, heat and Schrödinger propagators, and the functional calculus of $-\Delta$ ^{[Stein 1970]}.

Theorem 5 (Bochner-Riesz: the multiplier at the edge of the theory). The Bochner-Riesz means $S^{\delta }$, symbol $(1-|\xi {|}^2{)}_{+}^{\delta }$, are $L^p$-bounded for all $1<p<\infty$ when $\delta >(n-1)/2$ (kernel in $L^1$, Young), and conjecturally for $\delta >\delta (p)=n|1/p-1/2|-1/2$ when $p\ne 2$, $n\ge 2$. The sharp range is open for $n\ge 3$ and equivalent to the Fourier restriction conjecture; Hörmander-Mikhlin handles the symbol away from $\{|\xi |=1\}$, isolating the difficulty at the sphere ^{[Fefferman 1970]}.

Theorem 6 (the Hilbert transform and the one-dimensional model). On $\mathbb{R}$ the Hilbert transform has symbol $-i\mathrm{sgn}\xi$, a degree-$0$ Mikhlin multiplier (constant $\pm i$ on each half-line, smooth off the origin). It is the prototype: $L^2$-bounded by $\Vert -i\mathrm{sgn}{\Vert }_{\infty }=1$, $L^p$-bounded by Theorem 1, with sharp Pichorides norm $\tan (\pi /2p)$ for $1<p\le 2$. The Marcinkiewicz theorem (Exercise 6) is the one-dimensional refinement, asking only bounded dyadic variation rather than smoothness ^{[Marcinkiewicz 1939]}.

Synthesis. The Hörmander-Mikhlin theorem is the foundational reason that smoothness of a Fourier symbol, measured scale-by-scale, certifies $L^p$-boundedness on the whole open range: the central insight is that the scale-invariant Mikhlin bound makes every dyadic localization $m(2^j\cdot )\eta$ a single fixed bump in $H^s$, $s>n/2$, so the symbol's regularity becomes the kernel's regularity through Sobolev embedding, and this is exactly what the Calderón-Zygmund theory of 02.19.03 consumes. Putting these together, three reformulations carry the same content — the pointwise Mikhlin derivative bounds (Theorem 1), the localized $L^2$-Sobolev form (Theorem 2), and the dyadic Littlewood-Paley splitting that the proof actually runs on (Theorem 1's proof) — and the bridge between them is that each $H^s$ piece is a unit-scale Calderón-Zygmund kernel summed geometrically. This generalises in two directions at once: vertically it is dual to itself, the weak-$(1,1)$ endpoint pairing with $L^{\infty }\to \mathrm{BMO}$ exactly as in 02.20.01, so duality across $p=2$ fills the upper range; horizontally it specializes to every named operator of the subject — Riesz transforms, imaginary powers $(-\Delta {)}^{is}$, the Hilbert transform, and, at the smooth end, Bochner-Riesz — each a Mikhlin multiplier whose $L^p$ boundedness is one application of this theorem. The deepest reading is that a translation-invariant operator is a singular integral precisely when its symbol is gently varying, and the gentleness needed is exactly the half-derivative more than $n/2$ that Sobolev embedding demands to turn an $L^2$ symbol bound into an $L^1$ kernel-difference bound.

Full proof set Master

Proposition 1 ($L^2$ characterization). $T_m$ is bounded on $L^2({\mathbb{R}}^n)$ iff $m\in L^{\infty }$, and $\Vert T_m{\Vert }_{L^2\to L^2}=\Vert m{\Vert }_{L^{\infty }}$.

Proof. If $m\in L^{\infty }$, Plancherel 02.10.04 gives $\Vert T_{m}f{\Vert }_2=\Vert m\hat{f}{\Vert }_2\le \Vert m{\Vert }_{\infty }\Vert f{\Vert }_2$, so $\Vert T_m{\Vert }_{2\to 2}\le \Vert m{\Vert }_{\infty }$. Conversely, fix ${\xi }_0$ a Lebesgue point of $|m|$ and let $\widehat{f_r}=|B({\xi }_0,r){|}^{-1/2}{\chi }_{B({\xi }_0,r)}$, a unit-$L^2$ bump; then $\Vert T_m{f}_r{\Vert }_2^2=|B({\xi }_0,r){|}^{-1}{\int }_{B({\xi }_0,r)}|m{|}^2\to |m({\xi }_0){|}^2$ as $r\to 0$. Taking the supremum over Lebesgue points recovers $\Vert T_m{\Vert }_{2\to 2}\ge \Vert m{\Vert }_{\infty }$. If $m\notin L^{\infty }$ the same construction makes $\Vert T_m{f}_r{\Vert }_2$ unbounded, so $T_m$ is not $L^2$-bounded. $\square$

Proposition 2 (Mikhlin $\Rightarrow$ Hörmander kernel regularity). If $|{\partial }^{\alpha }m|\le A|\xi {|}^{-|\alpha |}$ for $|\alpha |\le \lfloor n/2\rfloor +1$, then $K=m^{\vee }$ satisfies ${\int }_{|x|>2|y|}|K(x-y)-K(x)|dx\le C(n)A$.

Proof. Decompose $m={\sum }_j{m}_j$, $m_j=m\hat{\eta }(2^{-j}\cdot )$, and set ${\mu }_j(\xi )=m_j(2^j\xi )=m(2^j\xi )\hat{\eta }(\xi )$, supported in $\{\frac{1}{2}\le |\xi |\le 2\}$. By scale-invariance (Proposition 4) and the Leibniz rule against $\hat{\eta }$, $|{\partial }^{\alpha }{\mu }_j|\le CA$ for $|\alpha |\le N:=\lfloor n/2\rfloor +1$, uniformly in $j$. Since $N>n/2$, the weighted estimate of Exercise 7 gives $\int (1+|x|)|{\check{\mu }}_j(x)|dx\le CA$ and likewise for $\nabla {\check{\mu }}_j$. With $K_j(x)=2^{jn}{\check{\mu }}_j(2^{j}x)$, the mean value theorem yields $\int |K_j(x-y)-K_j(x)|dx\le CA{2}^j|y|$ and the decay yields $\le CA(2^j|y|{)}^{-1}$; the minimum, summed over $j$ via two geometric series split at $2^j\sim |y{|}^{-1}$, gives ${\sum }_j\int |K_j(x-y)-K_j(x)|dx\le C(n)A$. Restricting to $|x|>2|y|$ only decreases the integral. $\square$

Proposition 3 (Hörmander-Mikhlin $L^p$ bound). Under the hypothesis of Proposition 2, $T_m$ is weak type $(1,1)$ and bounded on $L^p$, $1<p<\infty$, with $\Vert T_m{\Vert }_{p\to p}\le C(n)A\max \{p,(p-1{)}^{-1}\}$.

Proof. By Proposition 1 with the $\alpha =0$ bound, $T_m$ is $L^2$-bounded with norm $\le A$. By Proposition 2 its kernel satisfies the Hörmander regularity condition with constant $C(n)A$. These are exactly the two hypotheses of the Calderón-Zygmund master theorem 02.19.03, which gives weak type $(1,1)$ with constant $C(n)(A+C(n)A)=C(n)A$, then strong $(p,p)$ for $1<p<2$ by Marcinkiewicz interpolation 02.10.04 against the $(2,2)$ endpoint, and $2<p<\infty$ by duality (the transpose $T_m^t=T_{m(-\cdot )}$ satisfies the same bounds). The constant degenerates as $\max \{p,(p-1{)}^{-1}\}$. $\square$

Proposition 4 (scale-invariance). If $m$ satisfies the Mikhlin condition with constant $A$, so does $m(\lambda \cdot )$ for every $\lambda >0$, with the same $A$.

Proof. ${\partial }^{\alpha }[m(\lambda \cdot )](\xi )={\lambda }^{|\alpha |}({\partial }^{\alpha }m)(\lambda \xi )$, so $|{\partial }^{\alpha }[m(\lambda \cdot )](\xi )|\le {\lambda }^{|\alpha |}A|\lambda \xi {|}^{-|\alpha |}=A|\xi {|}^{-|\alpha |}$, the powers of $\lambda$ cancelling. $\square$

Proposition 5 (Riesz transforms reproduce second derivatives). With $R_k$ the $k$-th Riesz transform, ${\partial }_i{\partial }_j=-R_i{R}_j\Delta$ on $\mathcal{S}$, hence $\Vert {\partial }_i{\partial }_{j}u{\Vert }_{L^p}\le C_p\Vert \Delta u{\Vert }_{L^p}$, $1<p<\infty$.

Proof. On the Fourier side $\widehat{R_{k}f}=-i({\xi }_k/|\xi |)\hat{f}$, and $\widehat{\Delta u}=-4{\pi }^2|\xi {|}^2\hat{u}$, so $$\widehat{R_i{R}_j\Delta u}(\xi )=(-i\frac{{\xi }_i}{|\xi |})(-i\frac{{\xi }_j}{|\xi |})(-4{\pi }^2|\xi {|}^2)\hat{u}(\xi )=4{\pi }^2{\xi }_i{\xi }_j\hat{u}(\xi )=-\widehat{{\partial }_i{\partial }_{j}u}(\xi ),$$ using $\widehat{{\partial }_i{\partial }_{j}u}=-4{\pi }^2{\xi }_i{\xi }_j\hat{u}$. Hence ${\partial }_i{\partial }_{j}u=-R_i{R}_j\Delta u$. Each $R_k$ is a Mikhlin multiplier (Exercise 3), $L^p$-bounded by Proposition 3, so $\Vert {\partial }_i{\partial }_{j}u{\Vert }_p=\Vert R_i{R}_j\Delta u{\Vert }_p\le \Vert R_i{\Vert }_{p\to p}\Vert R_j{\Vert }_{p\to p}\Vert \Delta u{\Vert }_p=C_p\Vert \Delta u{\Vert }_p$. $\square$

Proposition 6 (imaginary-power norm growth). $\Vert (-\Delta {)}^{is}{\Vert }_{L^p\to L^p}\le C(n,p)(1+|s|{)}^{\lfloor n/2\rfloor +1}$.

Proof. The symbol $m_s(\xi )=(2\pi |\xi |{)}^{2is}=e^{2is\log (2\pi |\xi |)}$ has $|m_s|=1$. By Faà di Bruno, ${\partial }^{\alpha }{m}_s$ is $m_s$ times a polynomial in the derivatives of $2is\log (2\pi |\xi |)$; the order-$|\alpha |$ term carries at most $(2|s|{)}^{|\alpha |}$ and homogeneity weight $|\xi {|}^{-|\alpha |}$, giving $|{\partial }^{\alpha }{m}_s(\xi )|\le C_{\alpha }(1+|s|{)}^{|\alpha |}|\xi {|}^{-|\alpha |}$. Thus $m_s$ is Mikhlin with $A=C(n)(1+|s|{)}^{\lfloor n/2\rfloor +1}$, and Proposition 3 gives the stated $L^p$ bound. $\square$

Connections Master

• Littlewood-Paley theory and the square function 02.20.03. The proof engine. The dyadic decomposition $m={\sum }_{j}m\hat{\eta }(2^{-j}\cdot )$ that splits the symbol into single-scale bumps is exactly the Littlewood-Paley partition of that unit, and the Khintchine-randomized reduction of the Marcinkiewicz refinement (Exercise 6) is its square-function machinery; this unit is the multiplier-side payoff of the square function, with the per-octave control assembled by the same geometric summation.

• Calderón-Zygmund singular integral operators 02.19.03. The master theorem this specializes. The Hörmander-Mikhlin theorem is the statement that a Mikhlin symbol has a Calderón-Zygmund kernel: the $\alpha =0$ bound supplies the $L^2$ input and the dyadic derivative decay supplies the Hörmander regularity, so the weak-$(1,1)$/interpolation/duality machinery proved there transfers verbatim. Every multiplier bound in this unit is one invocation of that scalar singular-integral theorem.

• Fourier transform on ${\mathbb{R}}^n$ and the Plancherel theorem 02.10.04. The source of the $L^2$ characterization. Plancherel makes multiplication by $m$ on the frequency side isometric up to $\Vert m{\Vert }_{\infty }$, giving $\Vert T_m{\Vert }_{2\to 2}=\Vert m{\Vert }_{\infty }$ exactly, and the inverse transform is what converts the symbol's Sobolev regularity into the kernel's weighted-$L^1$ decay in the per-block estimates.

• BMO and the John-Nirenberg inequality 02.20.01. The endpoint partner. A Mikhlin multiplier fails strong $(\infty ,\infty )$ but maps $L^{\infty }\to \mathrm{BMO}$, and by Fefferman-Stein duality the weak-$(1,1)$ endpoint is dual to a strong $H^1\to L^1$ estimate; interpolation against the $\mathrm{BMO}$ endpoint reproves the open $L^p$ range, exactly as for the Calderón-Zygmund theory.

• Oscillatory integrals and stationary phase 02.20.05. The downstream sharpening. Where Hörmander-Mikhlin handles smooth symbols, the symbols at the edge — Bochner-Riesz $(1-|\xi {|}^2{)}_{+}^{\delta }$ near the sphere — require van der Corput and stationary-phase estimates on the oscillatory kernel; that unit supplies the decay of $\widehat{d\sigma }$ on the sphere that decides the critical Bochner-Riesz index this theorem cannot reach.

• Strichartz estimates via the TT* method 02.21.02. The dispersive-PDE application. Frequency-localized propagator estimates are proved by treating the Schrödinger/wave symbols $e^{it|\xi {|}^2}$, $e^{it|\xi |}$ as multipliers on each dyadic block; the Mikhlin calculus and the imaginary powers $(-\Delta {)}^{is}$ of this unit are the constant-coefficient operators that the $T{T}^{\ast }$ machinery reassembles into space-time estimates.

Historical & philosophical context Master

The $L^p$ multiplier problem was settled by Solomon Mikhlin in 1956 in a Doklady note On the multipliers of Fourier integrals ^{[Mikhlin 1956]}, which gave the derivative criterion $|{\partial }^{\alpha }m|\le A|\xi {|}^{-|\alpha |}$ for $|\alpha |\le n$ as sufficient for $L^p$-boundedness, building on his earlier work on singular integral equations. The hypothesis was sharpened almost immediately by Lars Hörmander in his 1960 Acta Mathematica paper Estimates for translation invariant operators in $L^p$ spaces ^{[Hörmander 1960]}, which lowered the required number of derivatives to $\lfloor n/2\rfloor +1$ and, more significantly, replaced the pointwise bounds by the localized $L^2$-Sobolev condition ${\sup }_R\Vert m(R\cdot )\eta {\Vert }_{H^s}<\infty$ for $s>n/2$, isolating the exact smoothness the kernel argument consumes.

The one-dimensional precursor is the multiplier theorem of Józef Marcinkiewicz, published in 1939 in Studia Mathematica ^{[Marcinkiewicz 1939]}, which asked only bounded variation of the symbol on each dyadic block and was proved by the Littlewood-Paley square function. The modern synthesis — that a multiplier is a Calderón-Zygmund operator whose kernel regularity is read off the symbol through a dyadic decomposition — is due to Elias Stein, who codified it in his 1970 monograph Singular Integrals and Differentiability Properties of Functions ^{[Stein 1970]}. The limits of the smooth theory are marked by Charles Fefferman's 1970 work on the Bochner-Riesz problem and his 1971 disproof of the ball-multiplier conjecture ^{[Fefferman 1970]}, which showed that a merely bounded but non-smooth symbol can fail every $L^p$ with $p\ne 2$, placing the smoothness hypothesis of Hörmander-Mikhlin at the boundary of what is true.

Bibliography Master

@article{Mikhlin1956,
  author  = {Mikhlin, Solomon G.},
  title   = {On the multipliers of Fourier integrals},
  journal = {Doklady Akademii Nauk SSSR},
  volume  = {109},
  year    = {1956},
  pages   = {701--703}
}

@article{Hormander1960,
  author  = {H\"ormander, Lars},
  title   = {Estimates for translation invariant operators in $L^p$ spaces},
  journal = {Acta Mathematica},
  volume  = {104},
  year    = {1960},
  pages   = {93--140}
}

@article{Marcinkiewicz1939,
  author  = {Marcinkiewicz, J\'ozef},
  title   = {Sur les multiplicateurs des s\'eries de Fourier},
  journal = {Studia Mathematica},
  volume  = {8},
  year    = {1939},
  pages   = {78--91}
}

@article{Fefferman1970,
  author  = {Fefferman, Charles},
  title   = {Inequalities for strongly singular convolution operators},
  journal = {Acta Mathematica},
  volume  = {124},
  year    = {1970},
  pages   = {9--36}
}

@article{Fefferman1971Ball,
  author  = {Fefferman, Charles},
  title   = {The multiplier problem for the ball},
  journal = {Annals of Mathematics},
  volume  = {94},
  year    = {1971},
  pages   = {330--336}
}

@book{Stein1970,
  author    = {Stein, Elias M.},
  title     = {Singular Integrals and Differentiability Properties of Functions},
  publisher = {Princeton University Press},
  year      = {1970}
}

@book{Stein1993,
  author    = {Stein, Elias M.},
  title     = {Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals},
  publisher = {Princeton University Press},
  year      = {1993}
}

@book{Grafakos2014Modern,
  author    = {Grafakos, Loukas},
  title     = {Modern Fourier Analysis},
  edition   = {3},
  publisher = {Springer},
  year      = {2014}
}