[proofplan] The strategy is to verify that $T_m$ is a [Calderón–Zygmund operator](/theorems/???) and then invoke the [Calderón–Zygmund theorem](/theorems/???) to upgrade $L^2$ boundedness to weak-$(1,1)$ and $L^p$ boundedness for $1 < p < \infty$. The $L^2$ bound is immediate from [Plancherel's theorem](/theorems/???) since $|m(\xi)| \le A$ from the $|\alpha| = 0$ case of the Mihlin condition. The kernel size and Hörmander smoothness conditions are verified by dyadic decomposition: write $m = \sum_{j \in \mathbb{Z}} m_j$ where each $m_j$ is supported in a dyadic annulus and inherits from the Mihlin condition the scale-invariant bound $|D^\alpha m_j(\xi)| \le A 2^{-j|\alpha|}$ on its support. The $L^\infty$ and Lipschitz norms of the inverse Fourier transform $K_j = \mathcal{F}^{-1} m_j$ are then estimated using the [Sobolev embedding](/theorems/???) $W^{\lfloor n/2 \rfloor + 1, 2} \hookrightarrow L^\infty$ on the unit ball, transported back to scale $2^j$ by Plancherel and chain rule. Summing the dyadic pieces and splitting the resulting series at $|y|$-dependent thresholds yields the Hörmander integral bound. [/proofplan] [step:Establish $L^2$ boundedness via Plancherel] The $|\alpha| = 0$ case of the Mihlin condition asserts $|m(\xi)| \le A$ for all $\xi \in \mathbb{R}^n \setminus \{0\}$. Since the singleton $\{0\}$ is $\mathcal{L}^n$-null, this is equivalent to $\|m\|_{L^\infty(\mathbb{R}^n)} \le A$. For $f \in \mathcal{S}(\mathbb{R}^n)$, by [Plancherel's theorem](/theorems/???) under the symmetric normalisation ($\|\hat{h}\|_{L^2} = \|h\|_{L^2}$), \begin{align*} \|T_m f\|_{L^2(\mathbb{R}^n)}^2 &= \|\widehat{T_m f}\|_{L^2(\mathbb{R}^n)}^2 \\ &= \|m \cdot \hat{f}\|_{L^2(\mathbb{R}^n)}^2 \\ &\le \|m\|_{L^\infty(\mathbb{R}^n)}^2 \|\hat{f}\|_{L^2(\mathbb{R}^n)}^2 \\ &\le A^2 \|f\|_{L^2(\mathbb{R}^n)}^2. \end{align*} The first equality is Plancherel applied to $T_m f$. The second uses the definition $\widehat{T_m f}(\xi) = m(\xi) \hat{f}(\xi)$. The third applies the pointwise bound $|m(\xi)| \le \|m\|_{L^\infty}$. The fourth applies Plancherel to $f$. By density of $\mathcal{S}(\mathbb{R}^n)$ in $L^2(\mathbb{R}^n)$ and the bound just proved, $T_m$ extends uniquely to a bounded linear operator on $L^2(\mathbb{R}^n)$ with $\|T_m\|_{\mathcal{L}(L^2)} \le A$. [/step] [step:Decompose $m$ dyadically and define $T_m = \sum_j T_{m_j}$] Pick a smooth bump $\hat{\psi}$ as in the [Littlewood–Paley decomposition](/theorems/3186): $\hat{\psi} \in C_c^\infty(\mathbb{R}^n)$ supported in $\{1/2 \le |\xi| \le 2\}$, with $\hat{\psi}(\xi) := \hat{\varphi}(\xi/2) - \hat{\varphi}(\xi)$ for a $\hat{\varphi} \in C_c^\infty(\mathbb{R}^n)$ equal to $1$ on $\{|\xi| \le 1\}$ and $0$ on $\{|\xi| \ge 2\}$. By the Littlewood–Paley identity, $\sum_{j \in \mathbb{Z}} \hat{\psi}(2^{-j} \xi) = 1$ for all $\xi \neq 0$. Define \begin{align*} m_j: \mathbb{R}^n &\to \mathbb{C}, \\ \xi &\mapsto m(\xi) \, \hat{\psi}(2^{-j} \xi), \quad j \in \mathbb{Z}. \end{align*} Note that $m$ is smooth on $\mathbb{R}^n \setminus \{0\}$ from the Mihlin hypothesis (the existence of $D^\alpha m$ is part of the assumption for $|\alpha| \le \lfloor n/2 \rfloor + 1$, and we treat $m$ as $C^{\lfloor n/2 \rfloor + 1}$ on the punctured space). The cutoff $\hat{\psi}(2^{-j}\xi)$ vanishes near $0$, so the product $m_j$ extends smoothly across $\xi = 0$ by setting $m_j(0) := 0$. Hence $m_j \in C_c^\infty(\mathbb{R}^n)$ if $m$ is smooth, or at least $m_j \in C^{\lfloor n/2 \rfloor + 1}$ with compact support, and $\operatorname{supp}(m_j) \subseteq \{2^{j-1} \le |\xi| \le 2^{j+1}\}$. For all $\xi \in \mathbb{R}^n \setminus \{0\}$, \begin{align*} m(\xi) = m(\xi) \cdot 1 = m(\xi) \sum_j \hat{\psi}(2^{-j}\xi) = \sum_j m_j(\xi). \end{align*} The convergence holds pointwise (only finitely many terms are non-zero at each $\xi$) and in $\mathcal{S}'(\mathbb{R}^n)$ (the partial sums approach $m$ as a tempered distribution). Define the kernel of $T_{m_j}$ as \begin{align*} K_j: \mathbb{R}^n &\to \mathbb{C}, \\ x &\mapsto \mathcal{F}^{-1}(m_j)(x) = \frac{1}{(2\pi)^{n/2}} \int_{\mathbb{R}^n} m_j(\xi) e^{i \xi \cdot x} \, d\mathcal{L}^n(\xi). \end{align*} Since $m_j$ has compact support and is at least $C^{\lfloor n/2 \rfloor + 1}$, $K_j \in C^\infty(\mathbb{R}^n)$ with $K_j$ and all its derivatives in $L^\infty(\mathbb{R}^n) \cap L^2(\mathbb{R}^n)$ (by Plancherel and the Riemann–Lebesgue lemma). The candidate kernel of $T_m$ is the formal sum \begin{align*} K(x) := \sum_{j \in \mathbb{Z}} K_j(x), \quad x \in \mathbb{R}^n \setminus \{0\}. \end{align*} We will verify in subsequent steps that this sum converges away from the origin and defines a [Calderón–Zygmund kernel](/theorems/???) with the required size and Hörmander conditions. [/step] [step:Estimate $\|K_j\|_{L^\infty}$ and $\|\nabla K_j\|_{L^\infty}$ via Sobolev embedding on the rescaled symbol] Set $N := \lfloor n/2 \rfloor + 1$, so $N > n/2$. We work with the rescaled symbol $\widetilde{m}_j(\eta) := m_j(2^j \eta) = m(2^j \eta)\,\hat{\psi}(\eta)$, which is supported in the fixed annulus $\{1/2 \le |\eta| \le 2\}$ uniformly in $j$. \textbf{Scale-invariant Mihlin estimate.} Apply the Leibniz rule to compute $D^\alpha m_j$: \begin{align*} D^\alpha m_j(\xi) = D^\alpha [m(\xi) \hat{\psi}(2^{-j}\xi)] = \sum_{\beta + \gamma = \alpha} \binom{\alpha}{\beta} D^\beta m(\xi) \cdot D^\gamma[\hat{\psi}(2^{-j}\xi)]. \end{align*} For each term, $|D^\beta m(\xi)| \le A |\xi|^{-|\beta|}$ by the Mihlin condition (valid since $|\beta| \le |\alpha| \le N$), and $|D^\gamma[\hat{\psi}(2^{-j}\xi)]| = 2^{-j|\gamma|} |(D^\gamma \hat{\psi})(2^{-j}\xi)| \le 2^{-j|\gamma|} \|D^\gamma \hat{\psi}\|_{L^\infty}$. On the support of $\hat{\psi}(2^{-j}\xi)$, $|\xi| \asymp 2^j$, so $|\xi|^{-|\beta|} \le C \cdot 2^{-j|\beta|}$. Combining, \begin{align*} |D^\alpha m_j(\xi)| \le C_{n, \alpha} A \cdot 2^{-j|\alpha|} \cdot \mathbb{1}_{\operatorname{supp}(m_j)}(\xi). \end{align*} This is the **scale-invariant Mihlin estimate**: the magnitude of $D^\alpha m_j$ scales as $2^{-j|\alpha|}$ on its support of $\mathcal{L}^n$-measure $\asymp 2^{nj}$. By the chain rule applied to $\widetilde{m}_j(\eta) = m_j(2^j \eta)$, transferring the estimate to the rescaled symbol, \begin{align*} |D^\alpha \widetilde{m}_j(\eta)| \le 2^{j|\alpha|} \cdot |D^\alpha m_j|(2^j \eta) \le 2^{j|\alpha|} \cdot C_{n, \alpha} A \cdot 2^{-j|\alpha|} = C_{n, \alpha} A. \end{align*} Hence $\|\widetilde{m}_j\|_{C^N(\overline{B(0, 2)})} \le C_{n, N} A$ uniformly in $j$, where the norm is the maximum of $\|D^\alpha \widetilde{m}_j\|_{L^\infty}$ over $|\alpha| \le N$. \textbf{$L^\infty$ bound on $\widetilde{K}_j$ via Sobolev embedding.} The kernel of $\widetilde{m}_j$ is \begin{align*} \widetilde{K}_j: \mathbb{R}^n &\to \mathbb{C}, \\ y &\mapsto \mathcal{F}^{-1}(\widetilde{m}_j)(y) = \frac{1}{(2\pi)^{n/2}} \int_{\mathbb{R}^n} \widetilde{m}_j(\eta)\,e^{i\eta \cdot y}\, d\mathcal{L}^n(\eta). \end{align*} By [scaling of the Fourier transform](/theorems/???), $K_j(x) = 2^{nj} \widetilde{K}_j(2^j x)$. Apply the $L^2$-Sobolev embedding $W^{N, 2}(\mathbb{R}^n) \hookrightarrow L^\infty(\mathbb{R}^n)$, valid for $N > n/2$ (which holds since $N = \lfloor n/2 \rfloor + 1 > n/2$). Using [Plancherel's theorem](/theorems/???) and the relation $D^\alpha \widetilde{K}_j \leftrightarrow (i\eta)^\alpha \widetilde{m}_j$, \begin{align*} \|\widetilde{K}_j\|_{W^{N, 2}(\mathbb{R}^n)}^2 = \sum_{|\alpha| \le N} \|D^\alpha \widetilde{K}_j\|_{L^2(\mathbb{R}^n)}^2 = \sum_{|\alpha| \le N} \|\eta^\alpha \widetilde{m}_j\|_{L^2(\mathbb{R}^n)}^2, \end{align*} Since $\widetilde{m}_j$ is supported in the fixed annulus $\{1/2 \le |\eta| \le 2\}$ with $\|\widetilde{m}_j\|_{L^\infty} \le A$, $|\eta^\alpha| \le 2^{|\alpha|}$ on the support, and the support has bounded $\mathcal{L}^n$-measure independent of $j$, so $\|\eta^\alpha \widetilde{m}_j\|_{L^2}^2 \le C_{n, N} A^2$. Hence \begin{align*} \|\widetilde{K}_j\|_{L^\infty(\mathbb{R}^n)} \le C_n \|\widetilde{K}_j\|_{W^{N, 2}(\mathbb{R}^n)} \le C_{n, N} A. \end{align*} Translating back via the scaling $K_j(x) = 2^{nj} \widetilde{K}_j(2^j x)$, \begin{align*} \|K_j\|_{L^\infty(\mathbb{R}^n)} = 2^{nj} \|\widetilde{K}_j\|_{L^\infty(\mathbb{R}^n)} \le C_{n, N} A \cdot 2^{nj}. \end{align*} The same procedure applied to $i\eta \widetilde{m}_j$ — whose $C^N$ norm on the bounded annulus is $\le C_{n, N} A$ since multiplication by $\eta$ on the bounded support adds a bounded factor — gives $\|\nabla \widetilde{K}_j\|_{L^\infty} \le C_{n, N} A$. By the chain rule $\nabla K_j(x) = 2^{(n+1) j} (\nabla \widetilde{K}_j)(2^j x)$, so \begin{align*} \|\nabla K_j\|_{L^\infty(\mathbb{R}^n)} \le C_{n, N} A \cdot 2^{(n+1) j}. \end{align*} [/step] [step:Refine the kernel bounds with rapid decay using higher Sobolev embedding] The bounds in Step 3 are uniform in $x$ but do not capture the decay of $K_j(x)$ as $|x| \to \infty$. We refine them. Recall $\widetilde{m}_j$ is compactly supported in $\{1/2 \le |\eta| \le 2\}$ with $\|\widetilde{m}_j\|_{C^N} \le C_{n, N} A$. By the same Sobolev embedding argument applied to $D^\beta \widetilde{m}_j$ for any multi-index $\beta$ (where the bound on $\|D^\beta \widetilde{m}_j\|_{C^N}$ requires $|\beta| + N$ derivatives of $\widetilde{m}_j$ — but Mihlin only gives $N$ derivatives), we obtain a polynomial-decay bound on $\widetilde{K}_j$ only up to a finite order. Specifically, from $\widetilde{K}_j = \mathcal{F}^{-1}(\widetilde{m}_j)$ and integration by parts in the Fourier integral, for any multi-index $\beta$ with $|\beta| \le N$, \begin{align*} |y^\beta \widetilde{K}_j(y)| = \left| \mathcal{F}^{-1}(D^\beta \widetilde{m}_j)(y) \right| \le \|D^\beta \widetilde{m}_j\|_{L^1(\mathbb{R}^n)}. \end{align*} Since $\widetilde{m}_j$ is supported in the bounded annulus $\{1/2 \le |\eta| \le 2\}$ of finite measure, $\|D^\beta \widetilde{m}_j\|_{L^1} \le C_n \|D^\beta \widetilde{m}_j\|_{L^\infty} \le C_{n, N} A$. Hence for $|\beta| \le N$, \begin{align*} |y^\beta \widetilde{K}_j(y)| \le C_{n, N} A. \end{align*} Choose $|\beta| = N$ to obtain $|\widetilde{K}_j(y)| \le C_{n, N} A \cdot |y|^{-N}$ (where the bound is uniform in directions, taking $\beta$ to be a multi-index with $|y^\beta| = |y_i|^N$ in the largest coordinate direction $i$). Combined with the uniform bound from Step 3, $|\widetilde{K}_j(y)| \le C_{n, N} A \cdot \min(1, |y|^{-N})$, equivalently, \begin{align*} |\widetilde{K}_j(y)| \le \frac{C_{n, N} A}{(1 + |y|)^N}, \quad y \in \mathbb{R}^n. \end{align*} Translating back to $K_j$: \begin{align*} |K_j(x)| = 2^{nj} |\widetilde{K}_j(2^j x)| \le \frac{C_{n, N} A \cdot 2^{nj}}{(1 + 2^j |x|)^N}, \quad x \in \mathbb{R}^n. \end{align*} For $\nabla K_j$: by the same argument applied to the function $i\eta \widetilde{m}_j$ (whose $C^N$ norm on the bounded annulus is $\le C_{n, N} A$), $|\nabla \widetilde{K}_j(y)| \le C_{n, N} A (1 + |y|)^{-N}$, and \begin{align*} |\nabla K_j(x)| \le \frac{C_{n, N} A \cdot 2^{(n+1) j}}{(1 + 2^j |x|)^N}, \quad x \in \mathbb{R}^n. \end{align*} [/step] [step:Bound the kernel size $|K(x)| \le C A |x|^{-n}$ by Plancherel and dyadic summation] We show that the formal sum $K(x) = \sum_{j \in \mathbb{Z}} K_j(x)$ converges absolutely for $x \neq 0$ and satisfies the integrated size estimate $\int_{|x| \asymp r} |K(x)|\,d\mathcal{L}^n(x) \le C_{n, N} A$ uniformly in $r > 0$. The argument follows Stein, *Singular Integrals and Differentiability Properties of Functions*, Chapter IV, Section 3.2. The key technical lemma is a Plancherel-based integral estimate on dyadic shells, which avoids the need for the pointwise bound $|K(x)| \le C A |x|^{-n}$ that the naive integration-by-parts cannot deliver when $N \le n$. [claim:Dyadic-shell integral estimate] For all $r > 0$ and all $j \in \mathbb{Z}$, \begin{align*} \int_{r/2 \le |x| \le 2r} |K_j(x)|\,d\mathcal{L}^n(x) \le C_{n, N} A \cdot \min\bigl((2^j r)^{n},\, (2^j r)^{-(N - n/2)}\bigr). \end{align*} Both exponents are positive — the first because $n > 0$ and the second because $N - n/2 > 0$ by the choice $N = \lfloor n/2 \rfloor + 1 > n/2$. [proof] Let $E_r := \{x \in \mathbb{R}^n : r/2 \le |x| \le 2r\}$, of $\mathcal{L}^n$-measure $\asymp r^n$. **Small-$r$ regime ($2^j r \le 1$).** By Step 3 (uniform $L^\infty$ bound) and Cauchy–Schwarz on $E_r$, \begin{align*} \int_{E_r} |K_j(x)|\,d\mathcal{L}^n(x) \le \|K_j\|_{L^\infty(\mathbb{R}^n)} \cdot \mathcal{L}^n(E_r) \le C_{n, N} A \cdot 2^{nj} \cdot r^n = C_{n, N} A \cdot (2^j r)^n. \end{align*} **Large-$r$ regime ($2^j r > 1$).** By Cauchy–Schwarz on $E_r$, \begin{align*} \int_{E_r} |K_j(x)|\,d\mathcal{L}^n(x) \le \mathcal{L}^n(E_r)^{1/2} \cdot \|K_j \mathbb{1}_{E_r}\|_{L^2(\mathbb{R}^n)}. \end{align*} Estimate $\|K_j \mathbb{1}_{E_r}\|_{L^2}$ by introducing the radial weight $|x|^N$: \begin{align*} \|K_j \mathbb{1}_{E_r}\|_{L^2(\mathbb{R}^n)} \le (2r)^{-N} \cdot \| |x|^N K_j \|_{L^2(\mathbb{R}^n)}. \end{align*} For $|\alpha| = N$, $|x^\alpha| \le |x|^N$, so by [Plancherel's theorem](/theorems/???) and the Fourier-derivative correspondence $x^\alpha K_j \leftrightarrow i^{|\alpha|} D^\alpha m_j$, \begin{align*} \| |x|^N K_j \|_{L^2(\mathbb{R}^n)}^2 \le \sum_{|\alpha| = N} \|x^\alpha K_j\|_{L^2}^2 = \sum_{|\alpha| = N} \|D^\alpha m_j\|_{L^2}^2. \end{align*} Using the scale-invariant Mihlin estimate from Step 3, $|D^\alpha m_j(\xi)| \le C_{n, N} A \cdot 2^{-jN}$ on a support of $\mathcal{L}^n$-measure $\asymp 2^{nj}$, \begin{align*} \|D^\alpha m_j\|_{L^2(\mathbb{R}^n)}^2 \le C_{n, N} A^2 \cdot 2^{-2jN} \cdot 2^{nj} = C_{n, N} A^2 \cdot 2^{j(n - 2N)}. \end{align*} Combining, \begin{align*} \| |x|^N K_j \|_{L^2(\mathbb{R}^n)} \le C_{n, N} A \cdot 2^{j(n/2 - N)}, \end{align*} and therefore \begin{align*} \|K_j \mathbb{1}_{E_r}\|_{L^2} \le C_{n, N} A \cdot r^{-N} \cdot 2^{j(n/2 - N)}. \end{align*} Inserting into the Cauchy–Schwarz bound, \begin{align*} \int_{E_r} |K_j(x)|\,d\mathcal{L}^n(x) \le C_{n, N} A \cdot r^{n/2} \cdot r^{-N} \cdot 2^{j(n/2 - N)} = C_{n, N} A \cdot (2^j r)^{n/2 - N} = C_{n, N} A \cdot (2^j r)^{-(N - n/2)}. \end{align*} The exponent $N - n/2 > 0$ is positive by $N = \lfloor n/2 \rfloor + 1 > n/2$, completing the proof. [/proof] [/claim] \textbf{Integrated size bound.} Fix $r > 0$ and let $j_0 \in \mathbb{Z}$ be the unique integer with $2^{j_0 - 1} < r^{-1} \le 2^{j_0}$, so $2^{j_0} r \asymp 1$. By the claim, \begin{align*} \int_{r/2 \le |x| \le 2r} |K(x)|\,d\mathcal{L}^n(x) &\le \sum_{j \le j_0} \int_{E_r} |K_j(x)|\,d\mathcal{L}^n(x) + \sum_{j > j_0} \int_{E_r} |K_j(x)|\,d\mathcal{L}^n(x) \\ &\le C_{n, N} A \sum_{j \le j_0} (2^j r)^n + C_{n, N} A \sum_{j > j_0} (2^j r)^{-(N - n/2)} \\ &\le C_{n, N} A \cdot \frac{(2^{j_0} r)^n}{1 - 2^{-n}} + C_{n, N} A \cdot \frac{(2^{j_0+1} r)^{-(N - n/2)}}{1 - 2^{-(N - n/2)}} \\ &\le C'_{n, N} A. \end{align*} Both geometric series converge because the exponents $n$ and $N - n/2$ are positive. The integrated size bound on each dyadic shell is uniform in $r$ — equivalently, $K$ is integrable on $\{|x| \asymp r\}$ at rate $\le C A$ — which is a standard restatement of $|K(x)| \le C A |x|^{-n}$ in averaged form, and suffices for the Calderón–Zygmund machinery (Stein, Chapter II, Section 4.2). [/step] [step:Verify the Hörmander kernel condition via integration by parts on dyadic shells] We verify the [Hörmander condition](/theorems/???): \begin{align*} \sup_{y \neq 0} \int_{|x| > 2|y|} |K(x - y) - K(x)| \, d\mathcal{L}^n(x) \le C_{n, N} A. \end{align*} Fix $y \in \mathbb{R}^n \setminus \{0\}$ and let $j_1 \in \mathbb{Z}$ be the unique integer with $2^{j_1 - 1} < |y|^{-1} \le 2^{j_1}$. Decompose \begin{align*} \int_{|x| > 2|y|} |K(x - y) - K(x)| \, d\mathcal{L}^n(x) \le \sum_{j \le j_1} I_j^{\mathrm{low}} + \sum_{j > j_1} I_j^{\mathrm{high}}, \end{align*} with $I_j^{\mathrm{low}} := \int_{|x| > 2|y|} |K_j(x - y) - K_j(x)| \, d\mathcal{L}^n(x)$ for $j \le j_1$, and similarly for $I_j^{\mathrm{high}}$. \textbf{Low-frequency pieces ($j \le j_1$, so $2^j |y| \le 2$).} Apply the [mean value theorem](/theorems/???) along the segment from $x - y$ to $x$: \begin{align*} |K_j(x - y) - K_j(x)| \le |y| \sup_{0 \le t \le 1} |\nabla K_j(x - ty)|. \end{align*} For $|x| > 2|y|$ and $t \in [0,1]$, $|x - ty| \ge |x| - |y| \ge |x|/2$. Step 3 gives the rapid-decay bound on $\nabla \widetilde{K}_j$, hence $|\nabla K_j(z)| \le C_{n, N} A \cdot 2^{(n+1)j}\,(1 + 2^j |z|)^{-N}$. Therefore \begin{align*} \int_{\mathbb{R}^n} |\nabla K_j(z)| \, d\mathcal{L}^n(z) = 2^j \int_{\mathbb{R}^n} |\nabla \widetilde{K}_j(\eta)|\,d\mathcal{L}^n(\eta) \le C_{n, N} A \cdot 2^j, \end{align*} where the equality is the chain-rule rescaling and the $L^1$ bound on $\nabla\widetilde K_j$ requires justification depending on whether $N > n$ or $N \le n$. We give it as a separate sub-claim. [claim:$\|\nabla\widetilde K_j\|_{L^1(\mathbb{R}^n)} \le C_{n, N} A$ uniformly in $j$] [proof] The Fourier symbol of $\nabla\widetilde K_j$ is $i\eta\,\widetilde m_j(\eta)$, supported in the fixed annulus $\{1/2 \le |\eta| \le 2\}$ with $C^N$-norm $\le C_{n,N} A$ (Step 3). When $N > n$ the rapid-decay bound $|\nabla\widetilde K_j(\eta)| \le C_{n,N}A(1+|\eta|)^{-N}$ from Step 4 gives $L^1$ integrability immediately, since $\int(1+|\eta|)^{-N}\,d\mathcal{L}^n(\eta) < \infty$ for $N > n$. When $N \le n$ we cannot use rapid decay alone; instead we apply Cauchy–Schwarz on dyadic shells. By exactly the argument of the dyadic-shell integral estimate above, applied to the symbol $i\eta\,\widetilde m_j(\eta)$ in place of $m_j$ (the $C^N$-norm bound carries over with the same constant $C_{n, N}A$ since the support is bounded and multiplication by $\eta$ on a bounded set increases each derivative bound by at most a factor of $C_{n,N}$), \begin{align*} \int_{r/2 \le |\eta| \le 2r} |\nabla\widetilde K_j(\eta)|\,d\mathcal{L}^n(\eta) \le C_{n, N}A \cdot \min\bigl(r^n,\, r^{-(N - n/2)}\bigr) \qquad \text{for all } r > 0. \end{align*} Summing the geometric series in dyadic shells $\{2^\ell \le |\eta| \le 2^{\ell+1}\}$ for $\ell \in \mathbb{Z}$, and using that both exponents $n$ and $N - n/2$ are positive (since $N = \lfloor n/2 \rfloor + 1 > n/2$), \begin{align*} \|\nabla\widetilde K_j\|_{L^1(\mathbb{R}^n)} &= \sum_{\ell \in \mathbb{Z}}\int_{2^\ell \le |\eta| \le 2^{\ell+1}}|\nabla\widetilde K_j(\eta)|\,d\mathcal{L}^n(\eta)\\ &\le C_{n, N}A\Bigl(\sum_{\ell \le 0} 2^{\ell n} + \sum_{\ell > 0} 2^{-\ell(N - n/2)}\Bigr) \le C'_{n, N}A. \end{align*} This is the desired bound, valid for all $N \ge 1$. [/proof] [/claim] Hence \begin{align*} I_j^{\mathrm{low}} \le |y| \cdot \|\nabla K_j\|_{L^1(\mathbb{R}^n)} \le C_{n, N} A \cdot |y| \cdot 2^j. \end{align*} Summing the geometric series, \begin{align*} \sum_{j \le j_1} I_j^{\mathrm{low}} \le C_{n, N} A \cdot |y| \cdot \sum_{j \le j_1} 2^j = C_{n, N} A \cdot |y| \cdot \frac{2^{j_1 + 1}}{1} \le C'_{n, N} A, \end{align*} since $2^{j_1} \asymp |y|^{-1}$. \textbf{High-frequency pieces ($j > j_1$, so $2^j |y| > 1$).} Use the triangle inequality and bound each term separately: \begin{align*} I_j^{\mathrm{high}} \le \int_{\mathbb{R}^n} |K_j(x - y)| \, d\mathcal{L}^n(x) + \int_{\mathbb{R}^n} |K_j(x)| \, d\mathcal{L}^n(x) = 2\,\|K_j\|_{L^1(\mathbb{R}^n)}. \end{align*} By the chain-rule rescaling $K_j(x) = 2^{nj} \widetilde{K}_j(2^j x)$, \begin{align*} \|K_j\|_{L^1(\mathbb{R}^n)} = \|\widetilde{K}_j\|_{L^1(\mathbb{R}^n)} \le C_{n, N} A, \end{align*} where the last inequality uses the same dyadic-shell argument as for $\nabla\widetilde K_j$ in the previous claim: the symbol $\widetilde m_j$ is supported in the bounded annulus with $C^N$-norm $\le C_{n,N}A$, and the dyadic-shell estimate gives \begin{align*} \int_{2^\ell \le |\eta| \le 2^{\ell+1}}|\widetilde K_j(\eta)|\,d\mathcal{L}^n(\eta) \le C_{n, N}A \cdot \min(2^{\ell n}, 2^{-\ell(N - n/2)}), \end{align*} which sums geometrically over $\ell \in \mathbb{Z}$ to $\|\widetilde K_j\|_{L^1} \le C'_{n,N} A$ (both exponents positive by $N > n/2$). This holds whether $N > n$ or $N \le n$. The naive summation $\sum_{j > j_1} \|K_j\|_{L^1}$ does not converge, since each term is bounded by $C A$ uniformly. We must extract additional decay in $j$. From the dyadic-shell decay claim applied to the integration region $\{|x| > 2|y|\}$: setting $u = 2^j x$, \begin{align*} \int_{|x| > 2|y|} |K_j(x)| \, d\mathcal{L}^n(x) = \int_{|u| > 2^{j+1} |y|} |\widetilde{K}_j(u)| \, d\mathcal{L}^n(u) \le C_{n, N} A \cdot (2^{j+1}|y|)^{-(N - n/2)}, \end{align*} where the last inequality uses the dyadic-shell integration of the claim with the exponent $N - n/2 > 0$ ensured by $N = \lfloor n/2 \rfloor + 1 > n/2$. Setting $\delta := N - n/2 > 0$, \begin{align*} I_j^{\mathrm{high}} \le 2 C_{n, N} A \cdot (2^j |y|)^{-\delta}. \end{align*} Summing the geometric series, \begin{align*} \sum_{j > j_1} I_j^{\mathrm{high}} \le 2 C_{n, N} A \cdot |y|^{-\delta} \sum_{j > j_1} 2^{-j\delta} = 2 C_{n, N} A \cdot |y|^{-\delta} \cdot \frac{2^{-(j_1 + 1) \delta}}{1 - 2^{-\delta}} \le C'_{n, N} A, \end{align*} using $2^{-j_1} \asymp |y|$ to cancel $|y|^{-\delta} \cdot 2^{-j_1 \delta} \asymp 1$. \textbf{Combining both halves.} The Hörmander condition is established: \begin{align*} \int_{|x| > 2|y|} |K(x - y) - K(x)| \, d\mathcal{L}^n(x) \le C_{n, N} A, \end{align*} uniformly in $y \neq 0$. The technical core — the dyadic-shell integral estimate combining Plancherel with the scale-invariant Mihlin derivative bound — is precisely the content of Stein, Chapter IV, Section 3.2. [/step] [step:Apply the Calderón–Zygmund theorem to conclude $L^p$ and weak-$(1,1)$ bounds] We have established: - $T_m: L^2(\mathbb{R}^n) \to L^2(\mathbb{R}^n)$ is bounded with $\|T_m\|_{\mathcal{L}(L^2)} \le A$ (Step 1). - The convolution kernel $K = \sum_j K_j = \mathcal{F}^{-1}(m)$ (in the principal-value sense, away from the origin) satisfies the integrated size bound from Step 5 and the Hörmander smoothness condition from Step 6, with constant $C_{n, N} A$. - Off the diagonal, $T_m$ acts by convolution against $K$: for Schwartz $f$ with compact support and $x \notin \operatorname{supp}(f)$, $T_m f(x) = (K * f)(x)$ in the principal-value sense. This follows because $T_m f = \mathcal{F}^{-1}(m \hat{f})$ and the multiplier–kernel duality. Apply the [Calderón–Zygmund theorem](/theorems/???): a convolution operator with kernel $K$ satisfying $L^2$ boundedness and the Hörmander condition extends to: - A bounded operator $T_m: L^p(\mathbb{R}^n) \to L^p(\mathbb{R}^n)$ for all $1 < p < \infty$, with $\|T_m\|_{\mathcal{L}(L^p)} \le C_{n, p}(A + \|T_m\|_{\mathcal{L}(L^2)}) \le C_{n, p} A$. - A weak-type-$(1, 1)$ operator $T_m: L^1(\mathbb{R}^n) \to L^{1, \infty}(\mathbb{R}^n)$, with $\|T_m f\|_{L^{1, \infty}} \le C_n (A + \|T_m\|_{\mathcal{L}(L^2)}) \|f\|_{L^1} \le C_n A \|f\|_{L^1}$. The constants $C_{n, p}, C_n$ depend only on $n$ and (where indicated) $p$. We verify the hypotheses of the Calderón–Zygmund theorem: it requires (i) $L^2$ boundedness, (ii) the kernel size $|K(x)| \lesssim A\,|x|^{-n}$, and (iii) the Hörmander condition. Items (i) and (iii) are Steps 1 and 6. We establish (ii) directly from the dyadic estimates already in hand. Fix $x \ne 0$ and let $j_0 \in \mathbb{Z}$ be the unique integer with $2^{j_0 - 1} < |x|^{-1} \le 2^{j_0}$, so $2^{j_0} |x| \asymp 1$. From Step 4 we have the pointwise bound \begin{align*} |K_j(x)| = 2^{nj}\,|\widetilde K_j(2^j x)| \le \frac{C_{n, N}A\,2^{nj}}{(1 + 2^j|x|)^N}. \end{align*} Split the sum at $j_0$: \begin{align*} \sum_{j \le j_0}|K_j(x)| &\le C_{n, N}A\sum_{j \le j_0} 2^{nj} \cdot 1 = C_{n, N}A \cdot \frac{2^{n(j_0 + 1)} - \cdots}{2^n - 1} \le C'_{n, N}A\,2^{nj_0} \asymp C'_{n, N}A\,|x|^{-n},\\ \sum_{j > j_0}|K_j(x)| &\le C_{n, N}A\sum_{j > j_0}\frac{2^{nj}}{(2^j|x|)^N} = C_{n, N}A\,|x|^{-N}\sum_{j > j_0} 2^{j(n - N)}. \end{align*} By choice of $N = \lfloor n/2 \rfloor + 1 \ge 1$: if $n \ge 2$ then $N \le n$ may occur, in which case $n - N \ge 0$ and the second series diverges; we therefore choose to use a *higher-order* integration by parts. Specifically, repeat the integration-by-parts argument of Step 4 with $|\beta| = N + n$ derivatives. Since $\widetilde m_j \in C^N$ we cannot directly take $N + n$ derivatives, but we can iterate the *dyadic-shell* $L^1$ argument: from the proof of the previous claim, $\|\widetilde K_j\|_{L^1(\{|\eta| \ge R\})} \le C_{n,N}A\,R^{-(N - n/2)}$ for all $R > 0$. Applied to $|\eta| \ge 2^j|x|$ (i.e. $|y| \ge |x|$ in original coordinates after rescaling), \begin{align*} \sum_{j > j_0}|K_j(x)| &= \sum_{j > j_0} 2^{nj}|\widetilde K_j(2^j x)|. \end{align*} Instead we use the integrated form: for each fixed $x \ne 0$, \begin{align*} \sum_{j > j_0}|K_j(x)| \le \sum_{j > j_0}\|K_j\|_{L^\infty} \cdot \mathbb{1}_{\{2^j|x| > 1\}} \le \sum_{j > j_0} C_{n,N}A\,2^{nj}\,(2^j|x|)^{-N}, \end{align*} which is a geometric series with ratio $2^{n - N}$. When $N > n$ (i.e. $\lfloor n/2 \rfloor + 1 > n$, which holds only for $n = 1$), it converges to $C\,A\,|x|^{-n}$. For $n \ge 2$, we replace the pointwise $L^\infty$ bound by an averaged one: by the dyadic-shell estimate (claim of Step 5) applied at radius $r = |x|$ to the shell $\{|x|/2 \le |z| \le 2|x|\}$, \begin{align*} \frac{1}{\mathcal{L}^n(\{|z| \asymp |x|\})}\int_{\{|z| \asymp |x|\}}|K(z)|\,d\mathcal{L}^n(z) \le \frac{C_{n,N}A}{|x|^n}, \end{align*} by the bound $\int_{\{|z| \asymp |x|\}}|K(z)|\,d\mathcal{L}^n(z) \le C_{n,N}A$ (Step 5) and $\mathcal{L}^n(\{|z| \asymp |x|\}) \asymp |x|^n$. The standard formulation of the Calderón–Zygmund theorem (Stein, *Singular Integrals*, Chapter II, §4.2; or Grafakos, *Classical Fourier Analysis*, Theorem 5.3.3) requires only this *averaged* size condition together with $L^2$ boundedness and the Hörmander condition — not a pointwise $|K(x)| \le C|x|^{-n}$. We have established all three, so the theorem applies. This proves the Mihlin Multiplier Theorem. [/step]