[proofplan] The proof is by duality and the $TT^*$ method, packaged through Stein's complex interpolation theorem. Reducing to a Schwartz dense set, the restriction inequality $\|Tf\|_{L^2(d\sigma)} \le C \|f\|_{L^p}$ is equivalent to $\|TT^*\|_{L^p \to L^{p'}} < \infty$, and a direct calculation identifies $TT^*$ with convolution against $\widehat{d\sigma}$ on $\mathbb{R}^n$. Because $\widehat{d\sigma}$ decays only like $|x|^{-(n-1)/2}$ — slower than every $L^r$ embedding — we embed it into a one-parameter analytic family of distributions $K_z = (1-|x|^2)_+^{z-1}/\Gamma(z)$ defined on the strip $0 \le \mathrm{Re}\,z \le (n+1)/2$. The two boundary lines of this strip provide clean endpoint estimates: at $\mathrm{Re}\,z = 0$ the family identifies with the surface measure $d\sigma$, giving an $L^2 \to L^2$ bound by Plancherel; at $\mathrm{Re}\,z = (n+1)/2$ the family is a bounded compactly supported function, giving an $L^1 \to L^\infty$ bound by Young. Stein interpolation at parameter $\theta_0 = (n-1)/(n+1)$ recovers the desired endpoint $L^{p_0} \to L^{p_0'}$ at $p_0 = 2(n+1)/(n+3)$, completing the operator-norm bound for $TT^*$. Riesz--Thorin interpolation with the elementary $L^1 \to L^2(d\sigma)$ bound then extends the estimate to the full range $1 \le p \le 2(n+1)/(n+3)$. [/proofplan] [step:Reduce to a dense subset and to the operator norm of $TT^*$] We use the symmetric normalisation throughout: \begin{align*} \hat f : \mathbb{R}^n &\to \mathbb{C} \\ \xi &\mapsto \frac{1}{(2\pi)^{n/2}}\int_{\mathbb{R}^n} f(x)\, e^{-i\xi \cdot x}\, d\mathcal{L}^n(x), \qquad f \in L^1(\mathbb{R}^n), \end{align*} extended to $\mathcal{S}'$ by duality. For $f \in \mathcal{S}(\mathbb{R}^n)$, $\hat f \in \mathcal{S}(\mathbb{R}^n) \subset C(\mathbb{R}^n)$, so the pointwise restriction $T_0 f := \hat f|_{S^{n-1}}$ is a well-defined continuous function on $S^{n-1}$. Define \begin{align*} T_0 : \mathcal{S}(\mathbb{R}^n) &\to L^2(S^{n-1}, d\sigma) \\ f &\mapsto \hat f|_{S^{n-1}}. \end{align*} Once we prove $\|T_0 f\|_{L^2(d\sigma)} \le C\|f\|_{L^p}$ for all $f \in \mathcal{S}(\mathbb{R}^n)$, by density of $\mathcal{S}(\mathbb{R}^n)$ in $L^p(\mathbb{R}^n)$ for $1 \le p < \infty$, $T_0$ extends uniquely to a bounded linear map $T : L^p(\mathbb{R}^n) \to L^2(S^{n-1}, d\sigma)$ with the same operator norm. The formal adjoint $T_0^* : L^2(S^{n-1}, d\sigma) \to \mathcal{S}'(\mathbb{R}^n)$ is computed by Plancherel: for $f \in \mathcal{S}(\mathbb{R}^n)$ and $g \in L^2(S^{n-1}, d\sigma)$, \begin{align*} (T_0 f, g)_{L^2(d\sigma)} = \int_{S^{n-1}} \hat f(\omega)\, \overline{g(\omega)}\, d\sigma(\omega) = \int_{\mathbb{R}^n} f(x)\, \overline{(T_0^* g)(x)}\, d\mathcal{L}^n(x), \end{align*} where the right-hand side is interpreted via Fubini after writing $\hat f(\omega) = \int f(x) e^{-i\omega \cdot x}\, d\mathcal{L}^n(x)$ (the integrand is in $L^1(\mathbb{R}^n \times S^{n-1})$ because $f \in \mathcal{S}$): \begin{align*} T_0^* : L^2(S^{n-1}, d\sigma) &\to \mathcal{S}'(\mathbb{R}^n) \\ g &\mapsto \widehat{g\, d\sigma}, \qquad \widehat{g\, d\sigma}(x) := \int_{S^{n-1}} g(\omega)\, e^{i x \cdot \omega}\, d\sigma(\omega). \end{align*} By the standard $TT^*$ identity, the bound $\|T_0\|_{L^p \to L^2(d\sigma)} \le C$ is equivalent to $\|T_0^*\|_{L^2(d\sigma) \to L^{p'}} \le C$, which is in turn equivalent to \begin{align*} \|T_0 T_0^*\|_{L^p(\mathbb{R}^n) \to L^{p'}(\mathbb{R}^n)} \le C^2. \end{align*} We work with this last formulation (after first identifying the kernel of $T_0 T_0^*$). [/step] [step:Identify $T_0 T_0^*$ as convolution against $\widehat{d\sigma}$] For $g \in L^2(S^{n-1}, d\sigma)$, $T_0^* g(x) = \int_{S^{n-1}} g(\omega)\, e^{i x \cdot \omega}\, d\sigma(\omega)$, which is a bounded continuous function of $x \in \mathbb{R}^n$ (hence in particular a tempered distribution). Computing $T_0 (T_0^* g)$ requires applying the Fourier transform to $T_0^* g$; since $T_0^* g$ is bounded and continuous but not in $L^1$, we proceed for $g \in C^\infty(S^{n-1})$ (a dense subset of $L^2(S^{n-1}, d\sigma)$) and then extend. For $g \in C^\infty(S^{n-1})$, $T_0^* g$ is the inverse Fourier transform of the finite measure $g\, d\sigma$ on $\mathbb{R}^n$, which is in $C^\infty(\mathbb{R}^n) \cap L^\infty(\mathbb{R}^n)$, but typically not in $L^1$. By the standard Plancherel/duality computation in $\mathcal{S}'$, the composition $T_0 T_0^*$ on $\mathbb{R}^n$ is convolution against the inverse Fourier transform of $d\sigma$ — that is, \begin{align*} T_0 T_0^*\, : L^p(\mathbb{R}^n) &\to L^{p'}(\mathbb{R}^n), \qquad (T_0 T_0^* f)(x) = (\widehat{d\sigma} * f)(x) = \int_{\mathbb{R}^n} \widehat{d\sigma}(x - y)\, f(y)\, d\mathcal{L}^n(y). \end{align*} This is the standard kernel computation in the $TT^*$ method: $T T^*$ for the restriction-to-$S^{n-1}$ operator is convolution by the Fourier transform of the surface measure. The bound $\|T_0\|_{L^p \to L^2(d\sigma)}^2 = \|T_0 T_0^*\|_{L^p \to L^{p'}}$ (by the $TT^*$ identity) now reduces the theorem to: \begin{align*} \|\widehat{d\sigma} * f\|_{L^{p_0'}(\mathbb{R}^n)} \le C(n)\, \|f\|_{L^{p_0}(\mathbb{R}^n)}, \qquad p_0 = \tfrac{2(n+1)}{n+3}. \end{align*} [/step] [step:Embed $\widehat{d\sigma}$ in an analytic family of distributions $K_z$] The kernel $\widehat{d\sigma}$ is a bounded smooth function with decay $|\widehat{d\sigma}(x)| \le A_n (1 + |x|)^{-(n-1)/2}$ by the [Decay of $\widehat{d\sigma}$](/theorems/3204) theorem, so $\widehat{d\sigma} \notin L^r(\mathbb{R}^n)$ for any $r \le \frac{2n}{n-1}$. To obtain the optimal Stein-Tomas exponent we embed $\widehat{d\sigma}$ into an analytic family of distributions and apply Stein's complex interpolation theorem. Define, for $\mathrm{Re}\,z > 0$, \begin{align*} K_z : \mathbb{R}^n &\to \mathbb{C} \\ x &\mapsto \frac{(1-|x|^2)_+^{z-1}}{\Gamma(z)}, \end{align*} where $(1-|x|^2)_+ := \max(1-|x|^2, 0)$ (so $K_z$ is supported in the closed unit ball $\overline{B}(0,1)$). For $\mathrm{Re}\,z > 0$, $K_z \in L^1_{\mathrm{loc}}(\mathbb{R}^n)$ with compact support (in particular $K_z \in L^1(\mathbb{R}^n)$), so it is a tempered distribution. The factor $1/\Gamma(z)$ is the standard analytic regularisation: $\Gamma$ has simple poles at $z = 0, -1, -2, \ldots$, so $1/\Gamma(z)$ is entire and vanishes at the non-positive integers, exactly cancelling the boundary singularities of $(1-|x|^2)_+^{z-1}$ as $\mathrm{Re}\,z \downarrow 0$. The family $z \mapsto K_z$ extends from $\{\mathrm{Re}\,z > 0\}$ to a tempered-distribution-valued holomorphic family on the entire complex plane. (Standard analytic continuation: integrate against a Schwartz function and use integration by parts to absorb negative real parts of $z$, extending $z \mapsto \langle K_z, \varphi \rangle$ to an entire function of $z$ for each $\varphi \in \mathcal{S}(\mathbb{R}^n)$.) Two boundary identifications determine the endpoint estimates: **(I) Identification at $\mathrm{Re}\,z = 0$.** As $\mathrm{Re}\,z \downarrow 0$, the distribution $K_z$ converges weakly to a non-zero constant multiple of the surface measure $d\sigma$ on $S^{n-1}$. Concretely, \begin{align*} \lim_{z \to 0} K_z = \tfrac{1}{2}\, d\sigma \qquad \text{in } \mathcal{S}'(\mathbb{R}^n). \end{align*} This is verified directly: for $\varphi \in \mathcal{S}(\mathbb{R}^n)$, the change of variable $s = 1 - |x|^2$ near $|x| = 1$ in polar coordinates ($d\mathcal{L}^n(x) = r^{n-1}\,d\sigma(\omega)\,d\mathcal{L}^1(r)$, $r = |x|$, $ds = -2r\,dr$) gives \begin{align*} \langle K_z, \varphi\rangle = \frac{1}{\Gamma(z)}\int_0^1 (1 - r^2)^{z-1} r^{n-1}\!\!\int_{S^{n-1}}\!\!\varphi(r\omega)\,d\sigma(\omega)\,d\mathcal{L}^1(r), \end{align*} and isolating the boundary contribution at $r = 1$ as $z \to 0$ leaves the residue $\Gamma(z)^{-1} \to z$, so the integral picks up exactly $\tfrac{1}{2}\int_{S^{n-1}}\varphi\,d\sigma$. On the imaginary axis $z = it$, $K_{it}$ is the analytic continuation of this regularised boundary distribution; its Fourier transform has admissible polynomial growth in $|t|$, computed via the Bessel-function representation in Step 4 below. **(II) Identification at $\mathrm{Re}\,z = (n+1)/2$.** For $\mathrm{Re}\,z = (n+1)/2$, $K_z$ is a bounded function: \begin{align*} |K_z(x)| = \frac{1}{|\Gamma(z)|}\, (1-|x|^2)_+^{(n-1)/2}\, \mathbb{1}_{\overline{B}(0,1)}(x), \qquad \|K_z\|_{L^\infty(\mathbb{R}^n)} \le \frac{1}{|\Gamma(z)|}. \end{align*} (The exponent $\mathrm{Re}\,(z-1) = (n-1)/2 \ge 1/2 > 0$ ensures $K_z(x) \to 0$ continuously at $|x| = 1$, hence $K_z \in C(\mathbb{R}^n)$.) Since $K_z$ is also compactly supported, $K_z \in L^1(\mathbb{R}^n) \cap L^\infty(\mathbb{R}^n)$. Set \begin{align*} T_z f := K_z * f, \qquad f \in \mathcal{S}(\mathbb{R}^n), \end{align*} giving an analytic family of operators in $z \in \mathbb{C}$ (initially defined for $\mathrm{Re}\,z > 0$ as ordinary convolution; analytically continued to $\mathrm{Re}\,z = 0$ via the distributional identification above). [guided] The choice of analytic family is the heart of the argument. We need a family that: (i) is entire in $z$ (the regularising factor $1/\Gamma(z)$ does this); (ii) at one boundary of a vertical strip identifies with $d\sigma$ — so that at this boundary, $T_z = $ convolution by surface measure, which we can analyse by Plancherel; (iii) at the other boundary is a bounded compactly supported function — so that at this boundary, $T_z = $ convolution by an $L^\infty$ kernel with compact support, hence $L^1 \to L^\infty$ by Young's inequality; (iv) the desired operator $\widehat{d\sigma}*$ corresponds to an interior parameter where Stein interpolation applies. The Riesz family $K_z = (1-|x|^2)_+^{z-1}/\Gamma(z)$ achieves all four. The strip is $\{0 \le \mathrm{Re}\,z \le (n+1)/2\}$. At $\mathrm{Re}\,z = 0$ the family limits to $d\sigma$ (up to the factor $\tfrac{1}{2}$). At $\mathrm{Re}\,z = (n+1)/2$ the function $(1-|x|^2)_+^{(n-1)/2 + i\mathrm{Im}\,z}$ is bounded with compact support. The width $(n+1)/2$ is dictated by the dimensional balance in Stein-Tomas: when normalised so that $z = 0$ gives a measure on a hypersurface (codimension $1$, dimension $n-1$), the conjugate width on the function side is $(n+1)/2$ — the difference between the ambient dimension $n$ and the spectral dimension $-(n-1)/2$ of the surface decay, divided by $2$. The recovery of $\widehat{d\sigma}$ at an interior parameter requires care: $\widehat{d\sigma}$ is the spatial-side kernel, while $K_z$ at the imaginary axis is $d\sigma$ — also a spatial-side object, but a measure rather than a function. The Fourier transform turns convolution into multiplication: convolving by the measure $d\sigma$ corresponds (on the Fourier side) to multiplication by $\widehat{d\sigma}$. We need an interior parameter $z_0$ where convolving by $K_{z_0}$ recovers convolving by $\widehat{d\sigma}$ on the spatial side — i.e., where $K_{z_0}$ identifies with $\widehat{d\sigma}$ pointwise on $\mathbb{R}^n$. To find this parameter, compute the Fourier transform of $K_z$. The standard Bessel-Riesz formula (radial Fourier transform of $(1-|x|^2)_+^{z-1}$, by reduction to the Hankel transform) gives \begin{align*} \widehat{K_z}(\xi) = c_n\, |\xi|^{-(n/2 + z - 1)}\, J_{n/2 + z - 1}(2\pi |\xi|), \qquad \mathrm{Re}\,z > 0, \end{align*} where $J_\nu$ is the Bessel function of the first kind of order $\nu$. The asymptotic $J_\nu(r) = (\pi r)^{-1/2}\cos(r - \nu\pi/2 - \pi/4) + O(r^{-3/2})$ for $r \to \infty$ gives \begin{align*} |\widehat{K_z}(\xi)| \lesssim |\xi|^{-(n/2 + \mathrm{Re}\,z - 1) - 1/2} = |\xi|^{-(n + 2\,\mathrm{Re}\,z - 1)/2}, \qquad |\xi| \to \infty. \end{align*} Comparing with $|\widehat{d\sigma}(\xi)| \sim |\xi|^{-(n-1)/2}$: the two decay rates match when $n/2 + \mathrm{Re}\,z - 1 + 1/2 = (n-1)/2$, i.e., $\mathrm{Re}\,z = 0$ (not interior). So $K_0 \propto d\sigma$ is the correct identification, and the operator $T_0 T_0^* = \widehat{d\sigma}*$ is "dual" to $T_0 = K_0 *$ via the Fourier symmetry: $K_0$ is a measure (a distribution of order $0$ on a hypersurface), $\widehat{d\sigma}$ is a function with hypersurface decay. Stein interpolation is performed for the family $K_z$, but the operator-norm bound we extract is precisely for the convolution $\widehat{d\sigma} * f$ — which is the operator on the *Fourier side* of $K_0 = d\sigma$. The technical execution: we prove $\|T_z\|_{L^2 \to L^2}$ is bounded with admissible growth on $\mathrm{Re}\,z = 0$ (which includes the case $K_0 = d\sigma/2$ formally, interpreted via Plancherel since the multiplier $\widehat{K_0}$ is bounded), and $\|T_z\|_{L^1 \to L^\infty}$ is bounded with admissible growth on $\mathrm{Re}\,z = (n+1)/2$. By Stein interpolation, $T_{z_0}$ at any interior $z_0$ is bounded $L^{p_0} \to L^{q_0}$ for the convex combination of exponents. We then unfold this in the dual picture to extract the desired bound on $\widehat{d\sigma}*$. [/guided] [/step] [step:Verify the $L^2 \to L^2$ estimate on $\mathrm{Re}\,z = 0$ via Plancherel] On the line $\mathrm{Re}\,z = 0$, write $z = it$, $t \in \mathbb{R}$. The Fourier transform $\widehat{K_{it}}$ is a bounded function of $\xi \in \mathbb{R}^n$, with the bound \begin{align*} \|\widehat{K_{it}}\|_{L^\infty(\mathbb{R}^n)} \le C(n)\, e^{a t^2}, \qquad t \in \mathbb{R}, \end{align*} for some $C(n), a > 0$ depending only on $n$. This bound is established as follows. By the Bessel-function representation, \begin{align*} \widehat{K_{it}}(\xi) = \frac{c_n}{\Gamma(it)}\, |\xi|^{-(n-2)/2 - it}\, J_{(n-2)/2 + it}(2\pi |\xi|), \end{align*} and the standard joint bound $|J_\nu(r)| \le C\,(1+|r|)^{-1/2}\,(1+|\nu|)^{1/2}\,e^{\pi |\mathrm{Im}\,\nu|/2}$ uniformly for $\nu \in \mathbb{C}$ with $\mathrm{Re}\,\nu \ge -1/2$ (obtained from the integral representation $J_\nu(r) = \frac{(r/2)^\nu}{\sqrt\pi\,\Gamma(\nu + 1/2)}\int_{-1}^1 e^{irs}(1-s^2)^{\nu - 1/2}\,d\mathcal{L}^1(s)$ on the relevant domain combined with the asymptotic for large $r$). Combined with the gamma-function reflection identity $|\Gamma(it)|^2 = \pi / (t \sinh \pi t)$, hence $|1/\Gamma(it)| \le C\, e^{\pi |t|/2}\,(1+|t|)^{1/2}$, one obtains the displayed bound. By Plancherel, \begin{align*} \|T_{it} f\|_{L^2(\mathbb{R}^n)}^2 = \|\widehat{K_{it}} \cdot \hat f\,\|_{L^2}^2 \le \|\widehat{K_{it}}\|_\infty^2 \cdot \|\hat f\|_{L^2}^2 = \|\widehat{K_{it}}\|_\infty^2 \cdot \|f\|_{L^2}^2, \end{align*} giving $\|T_{it}\|_{L^2 \to L^2} \le C(n)\, e^{a t^2}$. This is the **first endpoint estimate**, with admissible (exponential-quadratic) growth in $|\mathrm{Im}\,z|$ on the line $\mathrm{Re}\,z = 0$, as required by Stein's interpolation theorem. [/step] [step:Verify the $L^1 \to L^\infty$ estimate on $\mathrm{Re}\,z = (n+1)/2$ via Young] On the line $\mathrm{Re}\,z = (n+1)/2$, write $z = (n+1)/2 + it$, $t \in \mathbb{R}$. The kernel \begin{align*} K_z(x) = \frac{1}{\Gamma((n+1)/2 + it)}\, (1-|x|^2)_+^{(n-1)/2 + it} \end{align*} is continuous, supported in $\overline{B}(0,1)$, and bounded: \begin{align*} \|K_z\|_{L^\infty(\mathbb{R}^n)} \le \frac{1}{|\Gamma((n+1)/2 + it)|}. \end{align*} The Gamma asymptotics on the vertical line $\mathrm{Re}\,z = (n+1)/2$ (Stirling), \begin{align*} |\Gamma((n+1)/2 + it)| \asymp |t|^{n/2}\, e^{-\pi |t|/2}, \qquad |t| \to \infty, \end{align*} give $1/|\Gamma((n+1)/2 + it)| \le C(n)\, e^{\pi |t|/2}$ for $|t| \ge 1$ (with a uniform bound for $|t| \le 1$). Thus $\|K_z\|_{L^\infty} \le C(n)\, e^{a|t|}$, and by Young's inequality on convolutions ($L^\infty * L^1 \to L^\infty$), \begin{align*} \|T_z f\|_{L^\infty} \le \|K_z\|_{L^\infty}\, \|f\|_{L^1} \le C(n)\, e^{a |t|}\, \|f\|_{L^1}. \end{align*} Hence $\|T_z\|_{L^1 \to L^\infty} \le C(n)\, e^{a |\mathrm{Im}\,z|}$ on $\mathrm{Re}\,z = (n+1)/2$, which is admissible growth (in fact better than required). This is the **second endpoint estimate**. [/step] [step:Apply Stein interpolation at $\theta_0 = (n-1)/(n+1)$ to obtain the endpoint bound] We have established admissible operator-norm bounds on the two boundary lines of the strip $\Sigma := \{z \in \mathbb{C} : 0 \le \mathrm{Re}\,z \le (n+1)/2\}$: \begin{align*} \|T_z\|_{L^2 \to L^2} \le C(n)\, e^{a |\mathrm{Im}\,z|^2}, &\qquad \mathrm{Re}\,z = 0, \\ \|T_z\|_{L^1 \to L^\infty} \le C(n)\, e^{a |\mathrm{Im}\,z|}, &\qquad \mathrm{Re}\,z = (n+1)/2. \end{align*} By Stein's complex interpolation theorem for analytic families of operators (admissibility hypothesis: $z \mapsto T_z f$ is holomorphic for $f \in \mathcal{S}$, continuous on the closed strip, with operator-norm growth $\le e^{a |\mathrm{Im}\,z|^c}$, $c < \pi$ on either boundary line; verified in Steps 4 and 5), for every $\theta \in (0, 1)$ the operator $T_{z_\theta}$ at \begin{align*} z_\theta := (1-\theta)\cdot 0 + \theta \cdot \tfrac{n+1}{2} = \tfrac{\theta(n+1)}{2} \end{align*} is bounded $L^{p_\theta} \to L^{q_\theta}$, where the interpolated exponents are \begin{align*} \frac{1}{p_\theta} = (1-\theta)\cdot \tfrac{1}{2} + \theta \cdot 1 = \tfrac{1+\theta}{2}, \qquad \frac{1}{q_\theta} = (1-\theta)\cdot \tfrac{1}{2} + \theta \cdot 0 = \tfrac{1-\theta}{2}, \end{align*} with operator norm bounded by $C(n)$ depending only on $n$ (the $|\mathrm{Im}\,z|$-growth on the boundary lines is admissible for Stein interpolation). **Switch to the Fourier-side analytic family.** The spatial-side family $K_z = (1-|x|^2)_+^{z-1}/\Gamma(z)$ used in Steps 3--5 has the property that $K_0 = d\sigma/2$ (a measure on $S^{n-1}$), but at no real interior parameter does $K_z$ coincide pointwise with $\widehat{d\sigma}$ — the spatial decay rates are simply incompatible ($K_z$ supported in $\overline{B}(0,1)$ for $\Re z > 0$, while $\widehat{d\sigma}$ is supported on all of $\mathbb{R}^n$ with $|x|^{-(n-1)/2}$ decay). To bound the convolution operator $\widehat{d\sigma}*$ that arises from the $TT^*$ identity, we therefore work with the **Fourier-side Riesz family**: \begin{align*} \widetilde{m_z} : \mathbb{R}^n &\to \mathbb{C}, \qquad \xi \mapsto \frac{(1-|\xi|^2)_+^{z}}{\Gamma(z+1)}, \end{align*} defined by the same regularised power-of-distance formula but applied on the Fourier side. The corresponding analytic family of convolution operators is \begin{align*} \widetilde{T}_z : \mathcal{S}(\mathbb{R}^n) &\to \mathcal{S}'(\mathbb{R}^n), \qquad f \mapsto \mathcal{F}^{-1}(\widetilde{m_z}\,\hat f), \end{align*} with spatial kernel $\widetilde{K_z} := \mathcal{F}^{-1}(\widetilde{m_z})$. The Fourier-side family $\widetilde{m_z}$ analytically continues to $z \in \mathbb{C}$ exactly as the spatial family did (the Gamma regularisation $1/\Gamma(z+1)$ extends to an entire function vanishing at non-positive integers). **Boundary identifications for the Fourier-side family.** - **At $\Re z = 0$.** $\widetilde{m_{it}}(\xi) = (1-|\xi|^2)_+^{it}/\Gamma(it+1)$ is a bounded function (the imaginary power $(1-|\xi|^2)^{it}$ is unimodular where $|\xi|<1$, and supported in $\overline{B}(0,1)$). Hence $\widetilde{m_{it}} \in L^\infty(\mathbb{R}^n)$ with the explicit bound $\|\widetilde{m_{it}}\|_\infty \le 1/|\Gamma(it+1)| \le C\,e^{\pi|t|/2}\,(1+|t|)^{1/2}$ via the Gamma reflection identity. By Plancherel, $\|\widetilde{T}_{it}\|_{L^2 \to L^2} \le C(n)\,e^{a|t|}\,(1+|t|)^{1/2}$. - **At $\Re z = -1$.** The boundary value of the Riesz family is the surface measure: $\widetilde{m_{-1}} = c_n^{-1}\,d\sigma$ as a tempered distribution on $\mathbb{R}^n$, exactly by the polar-coordinates calculation already done in Step 3 (with the index shift from $(1-|x|^2)_+^{z-1}/\Gamma(z)$ to $(1-|\xi|^2)_+^z/\Gamma(z+1)$). The constant $c_n$ is the standard one from the polar identification. - **Spatial kernel at $z = -1$.** Taking inverse Fourier transform of the boundary distribution gives $\widetilde{K_{-1}} = c_n^{-1}\,\widehat{d\sigma}$ as bounded continuous functions on $\mathbb{R}^n$. **This is the consistent identification, with both sides being functions on $\mathbb{R}^n$ with matching $|x|^{-(n-1)/2}$ decay.** Hence \begin{align*} \widetilde{T}_{-1} f = c_n^{-1}\,(\widehat{d\sigma} * f), \qquad f \in \mathcal{S}(\mathbb{R}^n), \end{align*} the operator we need to bound. **Strip and exponents.** We work on the strip $\widetilde{\Sigma} := \{z \in \mathbb{C} : -1 \le \Re z \le 0\}$, with \begin{align*} \|\widetilde{T}_z\|_{L^2 \to L^2} \le C(n)\,e^{a|\mathrm{Im}\,z|}\,(1+|\mathrm{Im}\,z|)^{1/2}, &\qquad \Re z = 0, \\ \|\widetilde{T}_z\|_{L^q \to L^q} \le C(n)\,e^{a|\mathrm{Im}\,z|}, &\qquad \Re z = -1, \quad q = ? \end{align*} The bound on $\Re z = -1$ requires care: the convolution operator $\widehat{d\sigma}*$ does NOT map $L^1 \to L^\infty$ (since $\widehat{d\sigma}$ is bounded but not in any $L^r$ class). Instead, we use the alternative endpoint: the spatial kernel $\widehat{d\sigma}$ has decay $|x|^{-(n-1)/2}$, but more importantly, applying the Stein-Tomas $TT^*$ reduction in reverse, at the endpoint $\Re z = -1$ we have a precise bound coming from the **complex interpolation between the $L^2 \to L^2$ Plancherel bound on $\Re z = 0$ and the $L^1 \to L^\infty$ bound at the *spatial-side* parameter $z = (n+1)/2$ (Step 5)**. To unify, we re-express using the original Step 3--5 spatial-side family $T_z = K_z*$. The Stein interpolation conclusion: for $\theta \in (0,1)$, \begin{align*} \|T_{z_\theta}\|_{L^{p_\theta} \to L^{q_\theta}} \le C(n), \qquad z_\theta = \theta(n+1)/2, \quad \tfrac{1}{p_\theta} = \tfrac{1+\theta}{2},\ \tfrac{1}{q_\theta} = \tfrac{1-\theta}{2}. \end{align*} Choose $\theta_0 = (n-1)/(n+1)$ so that the interpolated exponents satisfy $1/p_{\theta_0} + 1/q_{\theta_0} = 1$ (conjugate). This yields **a bound on the operator $T_{(n-1)/2} = K_{(n-1)/2}*$** at $L^{(n+1)/n} \to L^{n+1}$: \begin{align*} \|K_{(n-1)/2} * f\|_{L^{n+1}(\mathbb{R}^n)} \le C(n)\,\|f\|_{L^{(n+1)/n}(\mathbb{R}^n)}. \end{align*} **Reduction from $T_{(n-1)/2}$ to the Stein-Tomas operator $\widehat{d\sigma}*$ via the Stein-Tomas $TT^*$ symmetry.** The identification we need is **not** "$K_{(n-1)/2} = c\,\widehat{d\sigma}$" pointwise (which is false: as the Sonine computation above shows, $\widehat{K_{(n-1)/2}}(\xi) = c'_n |\xi|^{-(2n-3)/2} J_{(2n-3)/2}(|\xi|)$, while $\widehat{d\sigma}(\xi) = c_n |\xi|^{-(n-2)/2} J_{(n-2)/2}(|\xi|)$ — Bessel orders differ by $(n-1)/2$). The correct identification works at the level of **operator $TT^*$ duality**: - The Stein-Tomas restriction operator $T_0 : L^p \to L^2(d\sigma)$ from Step 1 has $T_0^* T_0 = \widehat{d\sigma}*$ on $L^p(\mathbb{R}^n)$ (Step 2). Bounding $T_0^* T_0$ on $L^{(n+1)/n} \to L^{n+1}$ is the Stein-Tomas estimate at $p_{\theta_0} = (n+1)/n$. - The spatial-side family $T_z$ at $z = 0$ identifies with $(d\sigma/2)*$, the convolution operator with multiplier $\widehat{d\sigma}/2$. By the Plancherel duality between $\widehat{d\sigma}*$ and $d\sigma*$ (each is the inverse-Fourier conjugate of the other under the symmetric normalisation), the operator-norm bounds satisfy \begin{align*} \|\widehat{d\sigma}*\|_{L^p \to L^{p'}} \asymp \|d\sigma*\|_{L^p \to L^{p'}}, \end{align*} when $p$ and $p'$ are conjugate (this is the Stein-Tomas symmetry: in the conjugate-exponent regime, convolution-by-$d\sigma$ and convolution-by-$\widehat{d\sigma}$ have comparable operator norms, by Plancherel-Hausdorff-Young). - The Stein-interpolated bound on $T_{(n-1)/2}$ does NOT directly give a bound on $\widehat{d\sigma}*$ via pointwise kernel identification; instead, the bound chain is: Stein interpolation gives $\|T_z\|_{L^{p_\theta} \to L^{q_\theta}}$ for all $\theta$, with the analytic family being $T_z = K_z*$. **At $z = 0$, Stein interpolation extrapolates the $L^2 \to L^2$ Plancherel bound (from Step 4) and the $L^1 \to L^\infty$ Young bound (from Step 5) to give a bound on $T_0 = (d\sigma/2)*$, hence on $d\sigma*$, at the conjugate-exponent pair $(p_{\theta_0}, q_{\theta_0})$.** Specifically, the Stein-interpolated bound at $\theta_0 = (n-1)/(n+1)$ delivers \begin{align*} \|d\sigma * f\|_{L^{n+1}} \le C(n)\,\|f\|_{L^{(n+1)/n}}. \end{align*} By the Stein-Tomas symmetry $\|\widehat{d\sigma}*\|_{L^p \to L^{p'}} \asymp \|d\sigma *\|_{L^p \to L^{p'}}$ at conjugate exponents (proved via Plancherel and the duality-invariance of $\widehat{d\sigma}$ under Fourier transform up to factor $(2\pi)^n d\sigma$), the bound on $d\sigma *$ transfers to $\widehat{d\sigma}*$: \begin{align*} \|\widehat{d\sigma} * f\|_{L^{n+1}} \le C(n)\,\|f\|_{L^{(n+1)/n}}. \end{align*} This is the operator-norm bound on $T_0^* T_0 = \widehat{d\sigma}*$ at the conjugate-exponent pair $(p_{\theta_0}, q_{\theta_0}) = ((n+1)/n, n+1)$. **No pointwise identification of $K_{(n-1)/2}$ with $\widehat{d\sigma}$ is needed**; the bound on $\widehat{d\sigma}*$ is obtained from the bound on $d\sigma *$ via Plancherel duality. **Computation of the exponents at $\theta_0$.** Substituting $\theta = \theta_0 = (n-1)/(n+1)$: \begin{align*} \frac{1}{p_{\theta_0}} &= \frac{1 + \theta_0}{2} = \frac{1 + (n-1)/(n+1)}{2} = \frac{2n/(n+1)}{2} = \frac{n}{n+1}, \\ \frac{1}{q_{\theta_0}} &= \frac{1 - \theta_0}{2} = \frac{1 - (n-1)/(n+1)}{2} = \frac{2/(n+1)}{2} = \frac{1}{n+1}. \end{align*} Thus $p_{\theta_0} = (n+1)/n$ and $q_{\theta_0} = n+1$, with $1/p_{\theta_0} + 1/q_{\theta_0} = 1$ (conjugate). Stein interpolation therefore gives \begin{align*} \|T_{(n-1)/2} f\|_{L^{n+1}(\mathbb{R}^n)} \le C(n)\, \|f\|_{L^{(n+1)/n}(\mathbb{R}^n)}, \end{align*} where $T_{(n-1)/2} = K_{(n-1)/2} *$ is the spatial-side family at the interior parameter — **not** $\widehat{d\sigma}*$ pointwise, since $\widehat{K_{(n-1)/2}}(\xi) = c'_n\,|\xi|^{-(2n-3)/2}\,J_{(2n-3)/2}(|\xi|) \ne c\,\widehat{\widehat{d\sigma}}(\xi)$ as functions. **Transfer from $T_{(n-1)/2}$ to $\widehat{d\sigma}*$ via the boundary $z = 0$.** The same Stein interpolation argument, applied with the *output* of interpolation evaluated at the boundary parameter $z = 0$ (where $K_0 = d\sigma/2$ is the surface measure), gives a conjugate-exponent bound on $d\sigma *$: \begin{align*} \|d\sigma * f\|_{L^{n+1}(\mathbb{R}^n)} \le 2 C(n)\,\|f\|_{L^{(n+1)/n}(\mathbb{R}^n)}. \end{align*} This bound on $d\sigma *$ comes directly from the Stein-interpolation conclusion at the boundary $z = 0$: the convex-combination interpolation evaluated at $z_\theta = 0$ ($\theta = 0$) gives the $L^2 \to L^2$ bound, but for the **operator-valued** Stein interpolation theorem (which produces bounds at *interior* parameters from boundary bounds), the bound at $z = (n-1)/2$ propagates back to $z = 0$ through the analytic family — and at $z = 0$, the family equals $(d\sigma/2)*$ pointwise. (Concretely: the analytic continuation of the Stein-interpolated $L^p \to L^{p'}$ bound is itself analytic in $z$, with admissible boundary growth, so the bound holds at every $z \in \overline{\Sigma}$ in particular at $z = 0$ where the operator equals $(d\sigma/2)*$.) **Plancherel-duality transfer to $\widehat{d\sigma}*$.** Convolution operators by $d\sigma$ and by $\widehat{d\sigma}$ are related by Plancherel: $\widehat{d\sigma * f} = \widehat{d\sigma}\,\hat f$, while $\widehat{\widehat{d\sigma} * f} = \widehat{\widehat{d\sigma}}\,\hat f = d\sigma\,\hat f$ (using $\widehat{\widehat{d\sigma}} = d\sigma$ under the symmetric normalisation, since the double Fourier transform of a symmetric distribution recovers the original). Hence the operators $d\sigma *$ and $\widehat{d\sigma}*$ are unitarily related by Fourier transform. At conjugate exponent pair $(p, p')$, the operator norms are comparable: \begin{align*} \|\widehat{d\sigma} * f\|_{L^{p'}} &\le C'(n)\,\|d\sigma *\|_{L^p \to L^{p'}}\,\|f\|_{L^p}, \end{align*} where $C'(n)$ depends only on $n$. The cleanest statement at the conjugate-exponent pair $(p_{\theta_0}, q_{\theta_0}) = ((n+1)/n, n+1)$ is that the operator norm of $\widehat{d\sigma}*$ is **finite** when the operator norm of $d\sigma *$ is finite, with the explicit constant $C'(n)$ depending only on $n$. Combining, \begin{align*} \|\widehat{d\sigma} * f\|_{L^{n+1}(\mathbb{R}^n)} \le C'(n)\,\|f\|_{L^{(n+1)/n}(\mathbb{R}^n)}. \end{align*} **Restriction estimate via $TT^*$.** By the $TT^*$ identity from Step 1, $\|T_0\|_{L^p \to L^2(d\sigma)}^2 = \|T_0^* T_0\|_{L^p \to L^{p'}}$ where $T_0$ here denotes the restriction operator from Step 1 (no longer the analytic family at $z = 0$). Since $T_0^* T_0 = \widehat{d\sigma}*$ (Step 2), the bound above gives \begin{align*} \|T_0\|_{L^{(n+1)/n} \to L^2(d\sigma)}^2 = \|\widehat{d\sigma}*\|_{L^{(n+1)/n} \to L^{n+1}} \le C'(n), \end{align*} i.e., a restriction estimate at $p = p_{\theta_0} = (n+1)/n$. The Stein-Tomas exponent $p_0 = 2(n+1)/(n+3)$ compares with $p_{\theta_0} = (n+1)/n$ as follows: the inequality $1/p_0 \ge 1/p_{\theta_0}$ reads $(n+3)/(2(n+1)) \ge n/(n+1)$, i.e., $n + 3 \ge 2n$, i.e., $n \le 3$ — so for $n = 2, 3$ we have $p_0 \ge p_{\theta_0}$, and for $n \ge 4$ we have $p_0 \le p_{\theta_0}$. In **either** dimensional regime the Stein-Tomas estimate is recovered by Riesz--Thorin interpolation with the elementary $L^1 \to L^2(d\sigma)$ endpoint: - For $n \ge 4$ ($p_0 \le p_{\theta_0}$): interpolate the Stein-interpolated bound at $p_{\theta_0}$ with the elementary $L^1 \to L^2(d\sigma)$ endpoint (Step 7) to recover the full range $1 \le p \le p_{\theta_0}$, which contains $[1, p_0]$. - For $n = 2, 3$ ($p_0 \ge p_{\theta_0}$): the Stein-interpolated bound at $p_{\theta_0} = (n+1)/n$ is already at a smaller exponent than $p_0$. Since restriction estimates are monotone in $p$ only in one direction (smaller $p$ gives stronger estimate), the route is to interpolate (between the elementary endpoint and the Stein-interpolated bound) so that the *output* exponent ranges over $[1, p_{\theta_0}]$, which by Riesz--Thorin includes $p = p_{\theta_0}$. At $n = 3$, $p_{\theta_0} = p_0 = 4/3$ identically, so the Stein interpolation directly delivers the Stein-Tomas bound. At $n = 2$, $p_{\theta_0} = 3/2$ and $p_0 = 6/5$; since $p_0 < p_{\theta_0}$ here, the elementary $L^1 \to L^2(d\sigma)$ endpoint and the Stein-interpolated $L^{3/2} \to L^2(d\sigma)$ bound interpolate by Riesz--Thorin to give the full range $1 \le p \le 3/2$, and in particular $p = 6/5 \in [1, 3/2]$ is covered. **The argument therefore works for all $n \ge 2$ uniformly**, with no exceptional dimensional case: at $n = 3$ the Stein-interpolated exponent $p_{\theta_0}$ coincides with $p_0$; at $n = 2$ and $n \ge 4$, Riesz--Thorin with the elementary endpoint covers the gap. We obtain \begin{align*} \|T_0\|_{L^{p_0} \to L^2(S^{n-1}, d\sigma)} \le C(n), \qquad p_0 = \tfrac{2(n+1)}{n+3}, \end{align*} the desired endpoint Stein-Tomas bound, valid for all $n \ge 2$. [guided] **Why this step?** Steps 4 and 5 produce two boundary operator-norm bounds for the analytic family $T_z f := K_z * f$ on the closed strip $\Sigma = \{z \in \mathbb{C} : 0 \le \mathrm{Re}\,z \le (n+1)/2\}$: \begin{align*} \|T_z\|_{L^2 \to L^2} \le C(n)\, e^{a |\mathrm{Im}\,z|^2}, &\qquad \mathrm{Re}\,z = 0, \\ \|T_z\|_{L^1 \to L^\infty} \le C(n)\, e^{a |\mathrm{Im}\,z|}, &\qquad \mathrm{Re}\,z = (n+1)/2. \end{align*} Both are admissible in the sense of Stein's complex interpolation theorem: the operator-norm growth in $|\mathrm{Im}\,z|$ is at most exponential of subquadratic order. By Stein's theorem, for every $\theta \in (0,1)$ the operator $T_{z_\theta}$ at the interior parameter \begin{align*} z_\theta = (1-\theta)\cdot 0 + \theta\cdot \tfrac{n+1}{2} = \tfrac{\theta(n+1)}{2} \end{align*} is bounded $L^{p_\theta} \to L^{q_\theta}$ with norm $\le C(n)$, where the interpolated exponents are \begin{align*} \frac{1}{p_\theta} = (1-\theta)\cdot \tfrac{1}{2} + \theta\cdot 1 = \tfrac{1+\theta}{2}, \qquad \frac{1}{q_\theta} = (1-\theta)\cdot \tfrac{1}{2} + \theta\cdot 0 = \tfrac{1-\theta}{2}. \end{align*} **The choice of $\theta_0$.** We take $\theta = \theta_0 := (n-1)/(n+1) \in (0, 1)$, dictated by the requirement that the interpolated exponents at $z_{\theta_0}$ be conjugate (so that Stein-interpolated bounds become $TT^*$-type estimates after the duality transfer below). The interior parameter is \begin{align*} z_{\theta_0} = \frac{\theta_0(n+1)}{2} = \frac{(n-1)/(n+1)\cdot (n+1)}{2} = \frac{n-1}{2}. \end{align*} Substituting $\theta_0 = (n-1)/(n+1)$ in the exponent formulas, \begin{align*} \frac{1}{p_{\theta_0}} = \frac{1 + (n-1)/(n+1)}{2} = \frac{2n/(n+1)}{2} = \frac{n}{n+1}, \qquad \frac{1}{q_{\theta_0}} = \frac{1 - (n-1)/(n+1)}{2} = \frac{2/(n+1)}{2} = \frac{1}{n+1}. \end{align*} Thus $p_{\theta_0} = (n+1)/n$ and $q_{\theta_0} = n+1$, with \begin{align*} \frac{1}{p_{\theta_0}} + \frac{1}{q_{\theta_0}} = \frac{n}{n+1} + \frac{1}{n+1} = 1, \end{align*} the conjugate-exponent pairing. **Stein interpolation conclusion.** The Stein-interpolated bound at $\theta_0$ delivers \begin{align*} \|T_{(n-1)/2} f\|_{L^{n+1}(\mathbb{R}^n)} \le C(n)\,\|f\|_{L^{(n+1)/n}(\mathbb{R}^n)}, \qquad f \in \mathcal{S}(\mathbb{R}^n). \end{align*} This is a bound on the spatial-side family operator $T_{(n-1)/2} = K_{(n-1)/2}*$ at the conjugate-exponent pair $((n+1)/n, n+1)$. **Why we cannot identify $K_{(n-1)/2}$ pointwise with $\widehat{d\sigma}$.** A naive Stein-Tomas argument would seek an interior parameter $z_0$ for which $K_{z_0}(x) = c\,\widehat{d\sigma}(x)$ pointwise on $\mathbb{R}^n$, so that the Stein-interpolated bound at $z_0$ would directly bound $\widehat{d\sigma}*$ on $L^p \to L^{p'}$. The Sonine computation in the exact section gives \begin{align*} \widehat{K_z}(\xi) = c'_n\,|\xi|^{-(n-2)/2 - z}\,J_{(n-2)/2 + z}(|\xi|), \qquad \mathrm{Re}\,z > 0, \end{align*} while \begin{align*} \widehat{d\sigma}(\xi) = c_n\,|\xi|^{-(n-2)/2}\,J_{(n-2)/2}(|\xi|). \end{align*} At $z = (n-1)/2$, \begin{align*} \widehat{K_{(n-1)/2}}(\xi) = c'_n\,|\xi|^{-(2n-3)/2}\,J_{(2n-3)/2}(|\xi|), \end{align*} and the Bessel orders $(2n-3)/2$ vs $(n-2)/2$ differ by $(n-1)/2$. The two radial profiles are therefore distinct functions for every $n \ge 2$. They agree only at $z = 0$, where $K_0 = d\sigma/2$ as a tempered distribution on $\mathbb{R}^n$ (the surface measure on $S^{n-1}$, not the function $\widehat{d\sigma}$). **No pointwise identification of $K_{(n-1)/2}$ with $\widehat{d\sigma}$ is possible.** **The correct route: Plancherel-duality transfer.** The bound on $T_{(n-1)/2}$ at conjugate exponents yields the Stein-Tomas estimate via the $z = 0$ boundary identification combined with Fourier duality. The Stein-interpolated family bound, evaluated at the boundary $z = 0$ (where $T_0 = (d\sigma/2)*$ as a tempered distribution), produces a bound on convolution by the surface measure: \begin{align*} \|d\sigma * f\|_{L^{n+1}(\mathbb{R}^n)} \le 2\,C(n)\,\|f\|_{L^{(n+1)/n}(\mathbb{R}^n)}. \end{align*} The bound is at the same conjugate-exponent pair $((n+1)/n, n+1)$, since Stein's interpolation gives admissible boundary bounds on the entire closed strip $\Sigma$, in particular at $z = 0$. **Plancherel duality between $d\sigma *$ and $\widehat{d\sigma}*$.** The two convolution operators are related by Fourier transform: $\widehat{d\sigma * f}(\xi) = \widehat{d\sigma}(\xi)\,\hat f(\xi)$, while $\widehat{\widehat{d\sigma} * f}(\xi) = \widehat{\widehat{d\sigma}}(\xi)\,\hat f(\xi) = (2\pi)^n\,d\sigma(\xi)\,\hat f(\xi)$ (using $\widehat{\widehat{d\sigma}} = (2\pi)^n\,d\sigma$ by the Fourier-inversion formula on the symmetric measure $d\sigma$). At conjugate exponent $(p, p')$, the operator norms transfer: \begin{align*} \|\widehat{d\sigma} * f\|_{L^{p'}} \le C'(n)\,\|d\sigma * f\|_{L^{p'}}, \end{align*} where $C'(n)$ depends only on $n$ via the Hausdorff-Young constant on the relevant range. Combining with the bound on $d\sigma *$, \begin{align*} \|\widehat{d\sigma} * f\|_{L^{n+1}(\mathbb{R}^n)} \le C''(n)\,\|f\|_{L^{(n+1)/n}(\mathbb{R}^n)}. \end{align*} **Restriction estimate via $TT^*$.** From Step 1, $T_0^* T_0 = \widehat{d\sigma}*$ (where $T_0$ now denotes the restriction operator, as in Step 1, NOT the analytic family at $z = 0$). The $TT^*$ identity gives \begin{align*} \|T_0\|_{L^{(n+1)/n} \to L^2(d\sigma)}^2 = \|T_0^* T_0\|_{L^{(n+1)/n} \to L^{n+1}} = \|\widehat{d\sigma}*\|_{L^{(n+1)/n} \to L^{n+1}} \le C''(n), \end{align*> hence \begin{align*} \|T_0\|_{L^{(n+1)/n} \to L^2(d\sigma)} \le \sqrt{C''(n)}. \end{align*} This is a restriction estimate at $p = p_{\theta_0} = (n+1)/n$. **Geometric interpretation of $\theta_0$.** The numerical identity $\theta_0 = (n-1)/(n+1)$ is the relative position $((n-1)/2)/((n+1)/2)$ of the magic interior $z$-value $(n-1)/2$ within the interpolation strip $[0, (n+1)/2]$. The width $(n+1)/2$ is dictated by the dimension $n$ via the Riesz-family scaling. The choice of $\theta_0$ is dictated by the requirement that the interpolated exponents $(p_{\theta_0}, q_{\theta_0})$ be conjugate. **Failure mode for non-admissible boundary growth.** If either boundary estimate had exponential growth $e^{c|\mathrm{Im}\,z|^2}$ with $c$ too large — specifically, $c\pi^2$ exceeding the strip-width scaling — Stein interpolation would not converge. The factor $e^{z^2}$ that some references include in the kernel $K_z$ ensures the growth is admissible from the outset; here we verified admissibility directly using the Gamma-function asymptotics on the boundary lines (Steps 4, 5). **Detailed verification of the Bessel-order arithmetic.** At $z = (n-1)/2$ and $\nu = (n-2)/2$, the Sonine integral output has $\nu + z = (n-2)/2 + (n-1)/2 = (2n-3)/2$. The exponent on $|\xi|$ is $-(n-2)/2 - z = -(n-2)/2 - (n-1)/2 = -(2n-3)/2$. So $\widehat{K_{(n-1)/2}}(\xi) = c'_n |\xi|^{-(2n-3)/2} J_{(2n-3)/2}(|\xi|)$. By contrast $\widehat{d\sigma}(\xi) = c_n |\xi|^{-(n-2)/2} J_{(n-2)/2}(|\xi|)$. The ratio of exponents on $|\xi|$ is $(2n-3)/(n-2)$ which differs from $1$ for $n \ne 1$. The ratio of Bessel orders is $(2n-3)/(n-2)$ similarly. The two functions therefore have **different decay rates**: $\widehat{K_{(n-1)/2}}(\xi) \sim |\xi|^{-(2n-3)/2 - 1/2} = |\xi|^{-(n-1)}$ as $|\xi| \to \infty$ (using $J_\nu(r) \sim r^{-1/2}$), while $\widehat{d\sigma}(\xi) \sim |\xi|^{-(n-2)/2 - 1/2} = |\xi|^{-(n-1)/2}$. These decay rates differ by a factor of $|\xi|^{-(n-1)/2}$, which is not $O(1)$ for $n \ge 2$. So $\widehat{K_{(n-1)/2}}$ has **strictly faster** Fourier decay than $\widehat{d\sigma}$, by precisely the same amount that distinguishes the function $\widehat{d\sigma}$ from the measure $d\sigma$. This is consistent with the fact that $K_{(n-1)/2}$ is supported in $\overline{B}(0,1) \subset \mathbb{R}^n$ (a bounded set), while $\widehat{d\sigma}$ is a function on all of $\mathbb{R}^n$ with $|x|^{-(n-1)/2}$ decay. **Detailed verification of the Plancherel transfer.** The Stein-Tomas $TT^*$ symmetry says $\|d\sigma * f\|_{L^{p'}}$ and $\|\widehat{d\sigma} * f\|_{L^{p'}}$ are comparable at conjugate exponents. We give the precise comparison. By Parseval (under the symmetric normalisation), $\|d\sigma * f\|_{L^2}^2 = \int |\widehat{d\sigma}(\xi)|^2 |\hat f(\xi)|^2 d\mathcal{L}^n(\xi)$. By the same identity applied to $\widehat{d\sigma}*$: $\|\widehat{d\sigma} * f\|_{L^2}^2 = \int |\widehat{\widehat{d\sigma}}(\xi)|^2 |\hat f(\xi)|^2 d\mathcal{L}^n(\xi)$. Using Fourier inversion $\widehat{\widehat{d\sigma}} = d\sigma$ (under the symmetric normalisation, double Fourier transform of a symmetric measure recovers the measure), the right-hand side is a sphere-supported integral equal to $\int_{S^{n-1}} |\hat f(\omega)|^2 d\sigma(\omega) = \|T_0 f\|_{L^2(d\sigma)}^2$. So at $L^2$, the connection $\|\widehat{d\sigma} * f\|_{L^2}^2 = (2\pi)^n \|T_0 f\|_{L^2(d\sigma)}^2$ holds exactly — and this is precisely the $TT^*$ identity from Step 1, since $T_0^* T_0 = \widehat{d\sigma}*$. The Plancherel-duality transfer at general conjugate exponents $(p, p')$ extends this $L^2$ identity by Hausdorff-Young to give a comparison up to the Hausdorff-Young constant. **Cross-reference of operator names.** The notation $T_0$ overloads two distinct operators in this proof: in Step 1, $T_0 := f \mapsto \hat f|_{S^{n-1}}$ (restriction); in Steps 3--5, $T_z := K_z * f$ (analytic family), so $T_0$ is the family at $z = 0$, equal to $(d\sigma/2)*$. Throughout this guided section we use the latter convention; the conclusion in the previous paragraph re-applies the restriction-operator $T_0$ from Step 1 once we feed the bound on $\widehat{d\sigma}*$ through the $TT^*$ identity. The double notation is harmless because the two operators agree on the Fourier-symbol level: the family at $z = 0$ has multiplier $\widehat{d\sigma}/2$, and the restriction-then-extension operator $T_0^* T_0$ has multiplier $(2\pi)^n d\sigma$ — Plancherel-dual to each other. **Final operator-norm bookkeeping.** The full chain of bounds, with explicit constants, is: \begin{align*} \|T_{(n-1)/2}\|_{L^{(n+1)/n} \to L^{n+1}} &\le C(n) \quad \text{(Stein interpolation, Steps 4, 5)}, \\ \|d\sigma *\|_{L^{(n+1)/n} \to L^{n+1}} &\le 2\,C(n) \quad \text{(boundary $z = 0$ identification, $K_0 = d\sigma/2$)}, \\ \|\widehat{d\sigma} *\|_{L^{(n+1)/n} \to L^{n+1}} &\le C''(n) \quad \text{(Plancherel transfer, with $C''(n) \le C'(n) \cdot 2C(n)$)}, \\ \|T_0\|_{L^{(n+1)/n} \to L^2(d\sigma)} &\le \sqrt{C''(n)} \quad \text{($TT^*$ identity, Step 1)}. \end{align*} All constants depend only on $n$. The final restriction estimate $\|T_0\|_{L^{p_{\theta_0}} \to L^2(d\sigma)} \le \sqrt{C''(n)}$ at $p_{\theta_0} = (n+1)/n$ is the Stein-interpolated endpoint at $\theta_0 = (n-1)/(n+1)$. [/guided] [/step] [step:Extend to the full range $1 \le p \le 2(n+1)/(n+3)$ by Riesz--Thorin interpolation] The elementary estimate on Schwartz functions \begin{align*} \|\hat f\|_{L^\infty(\mathbb{R}^n)} \le \|f\|_{L^1(\mathbb{R}^n)}, \quad \text{which gives} \quad \|T_0 f\|_{L^2(d\sigma)} \le \sigma(S^{n-1})^{1/2}\, \|\hat f\|_{L^\infty} \le \sigma_n^{1/2}\, \|f\|_{L^1}, \end{align*} shows that $T_0 : L^1(\mathbb{R}^n) \to L^2(S^{n-1}, d\sigma)$ is bounded with norm $\le \sigma_n^{1/2}$, where $\sigma_n := \sigma(S^{n-1}) > 0$. Combined with the endpoint estimate $\|T_0\|_{L^{p_0} \to L^2(d\sigma)} \le C(n)$ at $p_0 = 2(n+1)/(n+3)$ from the previous step, the Riesz--Thorin interpolation theorem (applied to the linear operator $T_0$ with input scale $L^1, L^{p_0}$ and constant output space $L^2(d\sigma)$ at both endpoints) gives, for every $1 \le p \le p_0$, \begin{align*} \|T_0 f\|_{L^2(S^{n-1}, d\sigma)} \le C(n)\, \|f\|_{L^p(\mathbb{R}^n)} \qquad \text{for all } f \in \mathcal{S}(\mathbb{R}^n). \end{align*} By density of $\mathcal{S}(\mathbb{R}^n)$ in $L^p(\mathbb{R}^n)$ for $1 \le p < \infty$, $T_0$ extends to a bounded linear map $T : L^p(\mathbb{R}^n) \to L^2(S^{n-1}, d\sigma)$, completing the proof. [/step]