[proofplan] We prove the three parts of the Trace Theorem in sequence. First, we establish the trace estimate $\|u\|_{L^p(\partial U)} \le C\|u\|_{W^{1,p}(U)}$ for smooth functions $u \in C^1(\overline{U})$ by applying the Divergence Theorem to the vector field $v_\varepsilon V$, where $v_\varepsilon = (u^2 + \varepsilon)^{p/2}$ is a smooth regularisation of $|u|^p$ and $V$ is a $C^1$ transversal vector field satisfying $V \cdot \nu \ge c_0 > 0$ on $\partial U$. We then extend $T$ from $C^1(\overline{U})$ to $W^{1,p}(U)$ by density and the completeness of $L^p(\partial U)$. Finally, we prove the kernel characterisation $\ker T = W_0^{1,p}(U)$ in both directions: the forward direction follows from continuity of $T$ and the definition of $W_0^{1,p}$; the reverse uses boundary flattening, zero-extension across $\{y_n = 0\}$, and translation-mollification to approximate boundary-vanishing functions by $C_c^\infty(U)$ functions. [/proofplan] [step:Construct a $C^1$ transversal vector field $V$ with $V \cdot \nu \ge c_0 > 0$ on $\partial U$] Since $\partial U$ is $C^1$, for every $z \in \partial U$ the local graph representation provides an orthogonal rotation $R_z$, a radius $r_z > 0$, and a $C^1$ function $\gamma_z: \mathbb{R}^{n-1} \to \mathbb{R}$ such that in the rotated coordinates $\tilde{x} = R_z(x - z)$: \begin{align*} \partial U \cap B(z, r_z) &= \{x \in B(z, r_z) : \tilde{x}_n = \gamma_z(\tilde{x}_1, \ldots, \tilde{x}_{n-1})\}. \end{align*} In these coordinates, the outward unit normal has $n$-th component bounded away from zero: there exists $\delta_z > 0$ such that $\tilde{e}_n \cdot \tilde{\nu}(\tilde{x}) \ge \delta_z$ for all $x \in \partial U \cap B(z, r_z)$. Setting $W_z := R_z^\top \tilde{e}_n \in \mathbb{R}^n$, the invariance of the inner product under orthogonal transformations gives $W_z \cdot \nu(x) \ge \delta_z$. By compactness of $\partial U$, select a finite subcover $\{B(z_i, r_i)\}_{i=1}^N$ and a smooth [partition of unity](/page/Partition%20of%20Unity) $\{\theta_i\}_{i=1}^N$ subordinate to this cover on a neighbourhood of $\partial U$. Define \begin{align*} V: \mathbb{R}^n &\to \mathbb{R}^n, \\ x &\mapsto \sum_{i=1}^N \theta_i(x)\, W_{z_i}. \end{align*} Then $V$ is $C^1$, and for $x \in \partial U$: \begin{align*} V(x) \cdot \nu(x) &= \sum_{i=1}^N \theta_i(x)\, W_{z_i} \cdot \nu(x) \ge \sum_{i=1}^N \theta_i(x)\, \delta_{z_i} \ge c_0 := \min_{1 \le i \le N} \delta_{z_i} > 0. \end{align*} [/step] [step:Prove the trace estimate $\|u\|_{L^p(\partial U)} \le C\|u\|_{W^{1,p}(U)}$ for $u \in C^1(\overline{U})$] Let $u \in C^1(\overline{U})$. For $\varepsilon > 0$, define \begin{align*} v_\varepsilon: \overline{U} &\to \mathbb{R}_{>0}, \\ x &\mapsto (u(x)^2 + \varepsilon)^{p/2}. \end{align*} The function $v_\varepsilon$ is strictly positive and $C^1$ on $\overline{U}$. Applying the Divergence Theorem to the vector field $v_\varepsilon V$ over $U$: \begin{align*} \int_{\partial U} v_\varepsilon\, (V \cdot \nu) \, d\mathcal{H}^{n-1}(x) &= \int_U \nabla \cdot (v_\varepsilon V) \, d\mathcal{L}^n(x) = \int_U \bigl(\nabla v_\varepsilon \cdot V + v_\varepsilon\, \nabla \cdot V\bigr) \, d\mathcal{L}^n(x). \end{align*} Using $V \cdot \nu \ge c_0$ on the left-hand side: \begin{align*} c_0 \int_{\partial U} v_\varepsilon \, d\mathcal{H}^{n-1}(x) &\le \int_U |\nabla v_\varepsilon \cdot V| \, d\mathcal{L}^n(x) + \int_U v_\varepsilon\, |\nabla \cdot V| \, d\mathcal{L}^n(x). \end{align*} Computing $\nabla v_\varepsilon = p(u^2 + \varepsilon)^{(p-2)/2} u\, \nabla u$ and using $|u| \le (u^2 + \varepsilon)^{1/2}$: \begin{align*} c_0 \int_{\partial U} (u^2 + \varepsilon)^{p/2} \, d\mathcal{H}^{n-1}(x) &\le p\|V\|_\infty \int_U (u^2 + \varepsilon)^{(p-1)/2} |\nabla u| \, d\mathcal{L}^n(x) + \|\nabla \cdot V\|_\infty \int_U (u^2 + \varepsilon)^{p/2} \, d\mathcal{L}^n(x). \end{align*} Let $\varepsilon \to 0$. By the [Dominated Convergence Theorem](/theorems/4) (the integrands are bounded by $C(|u|^{p-1}|\nabla u| + |u|^p) \in L^1(U)$ since $u \in C^1(\overline{U})$ and $U$ is bounded): \begin{align*} c_0 \int_{\partial U} |u|^p \, d\mathcal{H}^{n-1}(x) &\le C_1 \int_U \bigl(|u|^{p-1} |\nabla u| + |u|^p\bigr) \, d\mathcal{L}^n(x). \end{align*} Apply Young's inequality with conjugate exponents $q = p/(p-1)$ and $r = p$ to the mixed term: \begin{align*} |u|^{p-1}|\nabla u| &\le \frac{p-1}{p}\,|u|^p + \frac{1}{p}\,|\nabla u|^p. \end{align*} Substituting: \begin{align*} \int_{\partial U} |u|^p \, d\mathcal{H}^{n-1}(x) &\le C_2 \int_U \bigl(|u|^p + |\nabla u|^p\bigr) \, d\mathcal{L}^n(x) = C_2 \|u\|_{W^{1,p}(U)}^p. \end{align*} Taking the $p$-th root: $\|u\|_{L^p(\partial U)} \le C\|u\|_{W^{1,p}(U)}$. [guided] The goal is to bound the $L^p$ norm of $u$ on the boundary $\partial U$ by the $W^{1,p}$ norm of $u$ on $U$. The idea is to apply the Divergence Theorem to convert a boundary integral into a volume integral, which naturally introduces the gradient. Let $u \in C^1(\overline{U})$. We would like to apply the Divergence Theorem to $|u|^p V$, but $|u|^p$ may not be $C^1$ at zeros of $u$ when $p$ is not an even integer. To handle this, introduce a regularisation: for $\varepsilon > 0$, define $v_\varepsilon: \overline{U} \to \mathbb{R}_{>0}$ by $v_\varepsilon(x) = (u(x)^2 + \varepsilon)^{p/2}$. This is strictly positive and $C^1$, and $v_\varepsilon \to |u|^p$ pointwise as $\varepsilon \to 0$. The Divergence Theorem applied to $v_\varepsilon V$ over $U$ gives: \begin{align*} \int_{\partial U} v_\varepsilon\, (V \cdot \nu) \, d\mathcal{H}^{n-1}(x) &= \int_U \bigl(\nabla v_\varepsilon \cdot V + v_\varepsilon\, \nabla \cdot V\bigr) \, d\mathcal{L}^n(x). \end{align*} Since $V \cdot \nu \ge c_0 > 0$ on $\partial U$, the left-hand side is bounded below by $c_0 \int_{\partial U} v_\varepsilon \, d\mathcal{H}^{n-1}$. On the right-hand side, compute $\nabla v_\varepsilon = p(u^2 + \varepsilon)^{(p-2)/2} u\, \nabla u$. Using $|u| \le (u^2 + \varepsilon)^{1/2}$, we get $|\nabla v_\varepsilon| \le p(u^2 + \varepsilon)^{(p-1)/2}|\nabla u|$. Since $V$ and $\nabla \cdot V$ are bounded on the compact set $\overline{U}$: \begin{align*} c_0 \int_{\partial U} (u^2 + \varepsilon)^{p/2} \, d\mathcal{H}^{n-1}(x) &\le C_1 \int_U \bigl((u^2 + \varepsilon)^{(p-1)/2}|\nabla u| + (u^2 + \varepsilon)^{p/2}\bigr) \, d\mathcal{L}^n(x). \end{align*} Now let $\varepsilon \to 0$. The integrands converge pointwise to $|u|^{p-1}|\nabla u| + |u|^p$, and are dominated by the $\varepsilon = 1$ version (or by a fixed $L^1$ function, since $u \in C^1(\overline{U})$ on bounded $U$). The [Dominated Convergence Theorem](/theorems/4) gives: \begin{align*} c_0 \int_{\partial U} |u|^p \, d\mathcal{H}^{n-1}(x) &\le C_1 \int_U \bigl(|u|^{p-1}|\nabla u| + |u|^p\bigr) \, d\mathcal{L}^n(x). \end{align*} The mixed term $|u|^{p-1}|\nabla u|$ involves different powers of $|u|$ and $|\nabla u|$. We separate them using Young's inequality with exponents $q = p/(p-1)$ and $r = p$ (conjugate: $1/q + 1/r = 1$): \begin{align*} |u|^{p-1}|\nabla u| &\le \frac{1}{q}(|u|^{p-1})^q + \frac{1}{r}|\nabla u|^r = \frac{p-1}{p}\,|u|^p + \frac{1}{p}\,|\nabla u|^p. \end{align*} Substituting and absorbing constants: \begin{align*} \int_{\partial U} |u|^p \, d\mathcal{H}^{n-1}(x) &\le C_2 \int_U (|u|^p + |\nabla u|^p) \, d\mathcal{L}^n(x) = C_2\|u\|_{W^{1,p}(U)}^p. \end{align*} Taking the $p$-th root gives $\|u\|_{L^p(\partial U)} \le C\|u\|_{W^{1,p}(U)}$ with $C = C_2^{1/p}$. [/guided] [/step] [step:Extend the trace operator from $C^1(\overline{U})$ to $W^{1,p}(U)$ by density] Since $\partial U$ is $C^1$, the subspace $C^1(\overline{U})$ is dense in $W^{1,p}(U)$. For $u \in W^{1,p}(U)$, choose a sequence $\{u_m\}_{m=1}^\infty \subset C^1(\overline{U})$ with $u_m \to u$ in $W^{1,p}(U)$. By linearity and the trace estimate: \begin{align*} \|u_m|_{\partial U} - u_l|_{\partial U}\|_{L^p(\partial U)} &\le C\|u_m - u_l\|_{W^{1,p}(U)}. \end{align*} Since $\{u_m\}$ is Cauchy in $W^{1,p}(U)$, the sequence $\{u_m|_{\partial U}\}$ is Cauchy in the complete [Banach space](/page/Banach%20Space) $L^p(\partial U)$, and hence converges. Define \begin{align*} Tu &:= \lim_{m \to \infty} u_m|_{\partial U} \quad \text{in } L^p(\partial U). \end{align*} This definition is independent of the approximating sequence: if $\{u_m'\}$ is another sequence converging to $u$ in $W^{1,p}(U)$, then $\|u_m|_{\partial U} - u_m'|_{\partial U}\|_{L^p(\partial U)} \le C\|u_m - u_m'\|_{W^{1,p}(U)} \to 0$. The boundedness estimate $\|Tu\|_{L^p(\partial U)} \le C\|u\|_{W^{1,p}(U)}$ follows by passing to the limit, and consistency with the classical restriction holds by construction. [/step] [step:Prove the forward kernel characterisation: $u \in W_0^{1,p}(U)$ implies $Tu = 0$] Suppose $u \in W_0^{1,p}(U)$. By definition, there exists a sequence $\{u_m\}_{m=1}^\infty \subset C_c^\infty(U)$ converging to $u$ in $W^{1,p}(U)$. Since each $u_m$ has compact support strictly inside $U$, the classical restriction satisfies $u_m|_{\partial U} = 0$, so $Tu_m = 0$. By the boundedness of $T$: \begin{align*} \|Tu\|_{L^p(\partial U)} &= \|Tu - Tu_m\|_{L^p(\partial U)} \le C\|u - u_m\|_{W^{1,p}(U)} \to 0. \end{align*} Hence $Tu = 0$. [/step] [step:Prove the reverse kernel characterisation: $Tu = 0$ implies $u \in W_0^{1,p}(U)$ via boundary flattening and translation-mollification] Suppose $Tu = 0$ in $L^p(\partial U)$. Choose a smooth partition of unity $\{\theta_i\}_{i=0}^N$ on $\overline{U}$ with $\operatorname{supp}(\theta_0) \subset\subset U$ and $\operatorname{supp}(\theta_i) \subseteq B(z_i, r_i)$ for $i \ge 1$ (boundary charts). The interior piece $u_0 := \theta_0 u$ has compact support in $U$. A standard [mollification](/page/Standard%20Mollifier) with sufficiently small parameter produces $C_c^\infty(U)$ approximations, so $u_0 \in W_0^{1,p}(U)$. For each boundary piece $u_i := \theta_i u$ ($i \ge 1$), flatten the boundary via the $C^1$ diffeomorphism $\Phi_i$ (as in the local graph representation) mapping $W_i \cap U$ to $\mathbb{R}^n_+ := \{y \in \mathbb{R}^n : y_n > 0\}$. Define $w_i: \mathbb{R}^n_+ \to \mathbb{R}$ by $w_i(y) := u_i(\Phi_i^{-1}(y))$. Since $Tu_i = 0$ on $\partial U$ and $\Phi_i$ maps $\partial U$ to $\{y_n = 0\}$, the trace of $w_i$ on $\{y_n = 0\}$ vanishes. Extend $w_i$ by zero to $\overline{w}_i: \mathbb{R}^n \to \mathbb{R}$. Because the trace on $\{y_n = 0\}$ is zero, [integration by parts](/theorems/210) produces no boundary terms across the hyperplane, so $\overline{w}_i \in W^{1,p}(\mathbb{R}^n)$. Now translate: for $\varepsilon > 0$, define \begin{align*} w_{i,\varepsilon}: \mathbb{R}^n &\to \mathbb{R}, \\ y &\mapsto \overline{w}_i(y_1, \ldots, y_{n-1}, y_n - \varepsilon). \end{align*} The support of $w_{i,\varepsilon}$ is contained in $\{y_n > \varepsilon\}$, strictly away from $\{y_n = 0\}$. Mollifying $w_{i,\varepsilon}$ with a parameter smaller than $\varepsilon$ produces a function in $C_c^\infty(\mathbb{R}^n_+)$ that converges to $w_i$ in $W^{1,p}(\mathbb{R}^n_+)$ as $\varepsilon \to 0$ (using $L^p$ continuity of translations). Hence $w_i \in W_0^{1,p}(\mathbb{R}^n_+)$. Mapping back via $\Phi_i^{-1}$ preserves the closure-of-$C_c^\infty$ property, so $u_i \in W_0^{1,p}(U)$. Since $W_0^{1,p}(U)$ is a linear subspace: \begin{align*} u &= \sum_{i=0}^N u_i \in W_0^{1,p}(U). \end{align*} [guided] We must show that if $Tu = 0$, then $u$ can be approximated by $C_c^\infty(U)$ functions in the $W^{1,p}$ norm. The idea is to reduce to the half-space case using boundary charts, exploit the zero trace to extend by zero, and then use translations to shift the support away from the boundary. Start with a partition of unity $\{\theta_i\}_{i=0}^N$ on $\overline{U}$: $\theta_0$ is supported compactly inside $U$, and each $\theta_i$ ($i \ge 1$) is supported in a boundary chart neighbourhood $B(z_i, r_i)$. The interior piece $u_0 = \theta_0 u$ has compact support in $U$, so standard mollification immediately gives $u_0 \in W_0^{1,p}(U)$. For boundary pieces, fix $i \ge 1$ and set $u_i = \theta_i u$. The $C^1$ diffeomorphism $\Phi_i$ maps the boundary portion $W_i \cap \partial U$ to $\{y_n = 0\}$ and $W_i \cap U$ to $\{y_n > 0\}$. Define $w_i(y) = u_i(\Phi_i^{-1}(y))$. Since $Tu = 0$ on $\partial U$, and $T(\theta_i u) = \theta_i \cdot Tu = 0$ (as $\theta_i$ is smooth), the trace of $w_i$ on $\{y_n = 0\}$ vanishes: $w_i(y', 0) = 0$ in $L^p$. Now extend $w_i$ by zero to define $\overline{w}_i$ on all of $\mathbb{R}^n$. Why does this extension belong to $W^{1,p}(\mathbb{R}^n)$? The key is the zero trace. For any test function $\phi \in C_c^\infty(\mathbb{R}^n)$ and any $j < n$, the weak derivative $\partial_{y_j}\overline{w}_i$ equals $\overline{\partial_{y_j} w_i}$ (extended by zero) because no boundary term arises. For $j = n$, [integration by parts](/theorems/210) on $\{y_n > 0\}$ gives a boundary term $\int_{\{y_n = 0\}} w_i \phi \, d\mathcal{L}^{n-1}$, which vanishes because $w_i = 0$ on $\{y_n = 0\}$. Therefore $\overline{w}_i \in W^{1,p}(\mathbb{R}^n)$. To approximate $w_i$ by $C_c^\infty(\mathbb{R}^n_+)$ functions, translate downward: define $w_{i,\varepsilon}(y) = \overline{w}_i(y', y_n - \varepsilon)$. The support of $w_{i,\varepsilon}$ lies in $\{y_n > \varepsilon\}$, which is bounded away from $\{y_n = 0\}$. Mollifying $w_{i,\varepsilon}$ with parameter smaller than $\varepsilon$ gives a function in $C_c^\infty(\{y_n > 0\})$ that converges to $w_{i,\varepsilon}$ in $W^{1,p}$. By the $L^p$ continuity of translations, $w_{i,\varepsilon} \to w_i$ in $W^{1,p}(\mathbb{R}^n_+)$ as $\varepsilon \to 0$. Combining these two limits shows that $w_i$ lies in the $W^{1,p}$ closure of $C_c^\infty(\mathbb{R}^n_+)$, i.e., $w_i \in W_0^{1,p}(\mathbb{R}^n_+)$. Mapping back via $\Phi_i^{-1}$ (a $C^1$ diffeomorphism) preserves the property of being approximable by $C_c^\infty$ functions, so $u_i \in W_0^{1,p}(U)$. Since $W_0^{1,p}(U)$ is a vector subspace and $u = \sum_{i=0}^N u_i$, we conclude $u \in W_0^{1,p}(U)$. [/guided] [/step]