\section{Proofs of the enclosure lemmas}\label{app:enclosures} We use the notation of \S\ref{ssec:enclosures}: $M$ is the resolution, $x_n = n/M$, $K_n = [x_n, x_{n+1}]$ and $K_n^\circ$ is its interior. A cellwise affine function $F$ has end values $a_n = F(x_n+)$ and $b_n = F(x_{n+1}-)$ on $K_n$ and vanishes outside $[0, 1)$. Its distributional derivative is \[ F' = \sum_n M\,\Delta a_n\, \mathbf 1_{K_n} + \sum_n J^F_n\, \delta_{x_n}, \qquad \Delta a_n = b_n - a_n, \quad J^F_n = a_n - b_{n-1}, \] with $b_{-1} = 0$, and $J^F_M = -b_{M-1}$ the jump at $1$. We use the interpolation remainder: if $\varphi$ is continuous on $K_k$ and twice differentiable on $K_k^\circ$, and $\ell$ is its chord on $K_k$, then for every $x \in K_k$ there is $\xi \in K_k^\circ$ with \begin{equation}\label{eq:chord} \varphi(x) - \ell(x) = -\tfrac12 (x - x_k)(x_{k+1} - x)\, \varphi''(\xi), \qquad 0 \le (x - x_k)(x_{k+1} - x) \le \frac{1}{4M^2}. \end{equation} It follows from Rolle's theorem applied twice to $t \mapsto \varphi(t) - \ell(t) - c\,(t - x_k)(t - x_{k+1})$, with $c$ chosen so that this function vanishes at $x$. \begin{proof}[Proof of Lemma~\ref{lem:enc-kernel}] On $K_i$, $1/u$ is decreasing, which gives the first pair of bounds, and convex, so it lies below its chord $\ell_i$. Its second derivative is $2/u^3 \le 2M^3/i^3$ on $K_i$, so by \eqref{eq:chord} $1/u \ge \ell_i(u) - \frac{1}{8M^2}\cdot\frac{2M^3}{i^3} = \ell_i(u) - \frac{M}{4i^3}$. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:enc-conv}] $F$ and $G$ are bounded with bounded support, so $s = F * G$ is continuous. (a) Only pairs of cells with $n + m = k - 1$ contribute to $s(x_k)$. Put $y = x_n + \theta/M$ with $\theta \in [0, 1]$; then $x_k - y$ lies in $K_m$ at relative position $1 - \theta$, and \[ \int_{K_n} F(y)\, G(x_k - y)\, dy = \frac1M \int_0^1 \bigl(a_n + \Delta a_n\theta\bigr)\bigl(d_m - \Delta c_m\theta\bigr)\, d\theta = \frac{1}{6M}\bigl(a_n c_m + b_n d_m + 2a_n d_m + 2b_n c_m\bigr). \] (b) We have $s'' = F' * G'$. The term $M^2 \Delta a_n \Delta c_m\, \mathbf 1_{K_n} * \mathbf 1_{K_m}$ is a hat function: it is affine on $K_{n+m}$, rising from $0$ to $M\Delta a_n\Delta c_m$, and affine on $K_{n+m+1}$, falling back to $0$. The term $M J^F_n \Delta c_m\, \delta_{x_n} * \mathbf 1_{K_m} = M J^F_n\Delta c_m\, \mathbf 1_{K_{n+m}}$ is constant on $K_{n+m}$, and likewise $M\Delta a_n J^G_m\,\mathbf 1_{K_{n+m}}$. The terms $J^F_n J^G_m\,\delta_{x_{n+m}}$ are carried by grid points. So $s''$ is affine on $K_k^\circ$; at $x_k+$ the hats with $n + m = k - 1$ contribute their peaks and those with $n + m = k$ contribute $0$, and at $x_{k+1}-$ the reverse. This gives $M X^-_k$ and $M X^+_k$. Jumps at $1$ act only on cells beyond $1$. (c) $s$ is continuous on $K_k$ and, by (b), twice differentiable on $K_k^\circ$ with $s''$ affine there, so the extreme values of $s''$ on $K_k^\circ$ are its one-sided end values. By \eqref{eq:chord}, $s - s_k \ge -\frac{1}{8M^2}\max(0, MX^-_k, MX^+_k)$ and $s - s_k \le \frac{1}{8M^2}\max(0, -MX^-_k, -MX^+_k)$. For the correlation $r(x) = \int F(u)\, G(u - x)\, du$ the same computation, with $u = x_n + \theta/M$ and $u - x_k$ at relative position $\theta$ in $K_{n-k}$, gives the stated knot values, and two integrations by parts give $r'' = -\int F'(u)\, G'(u - x)\, du$; for $0 \le x < 1$ the jump of $F$ at $1$ meets $G$ at $1 - x \in (0, 1]$, so it is included. Finally, if $0 \le F^- \le f \le F^+$ and $0 \le G^- \le g \le G^+$ on open cells, then $F^- * G^- \le f * g \le F^+ * G^+$ everywhere, since the integrands are ordered almost everywhere; (c) applied to the two outer convolutions encloses $f * g$, and outward rounding of the knot values and of the slack keeps it enclosed. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:enc-sign}] Fix an open cell. If both end values of $F^-$ are $\ge 0$, then $F^- \ge 0$ on the cell, so $f^+ = f \ge F^-$; otherwise $f^+ \ge 0$ is the claimed lower bound. Next, $f^+ = \max(f, 0) \le \max(F^+, 0)$, and $\max(F^+, 0)$ is convex on the cell, being the maximum of two affine functions, so it lies below its chord. The statements for $f^- = (-f)^+$ follow from these, applied to $-F^+ \le -f \le -F^-$. Since $f = f^+ - f^-$ with $f^\pm \ge 0$, a product or convolution of signed functions is a signed sum of four products or convolutions of nonnegative functions, each enclosed by Lemma~\ref{lem:enc-conv} or~\ref{lem:enc-cell}; its lower bound uses the lower enclosures of the terms with sign $+$ and the upper enclosures of the terms with sign $-$, and its upper bound the reverse. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:enc-cell}] Fix $\tau \in [0, 1]$. (a) The map $(p, q) \mapsto pq$ is affine in each variable, so on the box $[p^-(\tau), p^+(\tau)] \times [q^-(\tau), q^+(\tau)]$ its extreme values are taken at corners. Each corner function $r = p^{\pm} q^{\pm}$ is quadratic in $\tau$ with $r'' = 2\Delta p^\pm \Delta q^\pm$, so by \eqref{eq:chord} (in the variable $\tau$, with cell length $1$) it lies within $|\Delta p^\pm\Delta q^\pm|/4$ of its chord. Hence $pq \le \max_r(\ell_r + \sigma_r)$, where $\ell_r$ is the chord of $r$ and $\sigma_r$ its slack. The right side is a maximum of affine functions of $\tau$, hence convex, hence below its own chord, which is the chord through the end values $\max_r(r(0) + \sigma_r)$ and $\max_r(r(1) + \sigma_r)$. The lower bound is symmetric: a minimum of affine functions is concave, so it lies above its chord. (b) If $q^- > 0$ throughout the closed cell or $q^+ < 0$ throughout it, then on the box the map $(p, q) \mapsto -p/q$ is monotone in $p$ and in $q$, so again its extreme values are at corners. For a corner $r = -p/q$ with $p$, $q$ affine in $\tau$, $r' = -(\Delta p\, q - p\, \Delta q)/q^2$ and $\Delta p\, q - p\,\Delta q = \Delta p\, q_0 - p_0\, \Delta q$ is constant, so $r'' = 2\Delta q\,(\Delta p\, q_0 - p_0\,\Delta q)/q^3$. As $q$ is affine and, by the hypothesis, does not vanish on the closed cell, $|q| \ge \min(|q_0|, |q_1|)$ there, which gives the stated bound, and the argument of (a) applies with the slack $\max|r''|/8$. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:enc-adjoint}] (a) By induction downwards from $j = m^*$: the coupling terms $C_{j,i}[h_{j+i}]$ with $i \ge 1$ involve only levels already enclosed, and they are built from convolutions, correlations, cell products and block bounds of enclosed functions (Lemmas~\ref{lem:enc-kernel} to~\ref{lem:enc-cell} and~\ref{lem:enc-blocks}). So the numerator of $h_j = -(A_j\cdot c + \sum_{i \ge 1} C_{j,i}[h_{j+i}])/B_j$ is enclosed on every cell, and Lemma~\ref{lem:enc-cell}(b) encloses the quotient, because the enclosure of $B_j$ excludes zero on the closed cell. (b) Write $\langle u, v\rangle = \int u v$ over the forcing range and let $C^*_{j,i}$ be the transpose of $C_{j,i}$, so that $\langle \lambda, C_{j,i}[g]\rangle = \langle C^*_{j,i}[\lambda], g\rangle$ (Fubini; all functions are bounded with bounded support). The recursion says $B_m\lambda_m + \sum_{j < m} C^*_{j,m-j}[\lambda_j] = \kappa_m$ for every $m$, and the forcing says $B_j h_j + \sum_{i \ge 1} C_{j,i}[h_{j+i}] = -A_j\cdot c$ for every $j$. Hence \begin{align*} \sum_m \langle \kappa_m, h_m\rangle &= \sum_m \langle B_m\lambda_m, h_m\rangle + \sum_{j < m} \langle C^*_{j,m-j}[\lambda_j], h_m\rangle = \sum_j \Bigl\langle \lambda_j,\ B_j h_j + \sum_{i \ge 1} C_{j,i}[h_{j+i}]\Bigr\rangle\\ &= -\sum_j \langle \lambda_j, A_j\cdot c\rangle . \end{align*} The recursion runs upwards from $m = 1$, and each step is enclosed as in (a), with the enclosure of $B_m$ excluding zero on every closed cell; so the exact $\lambda_m$ are enclosed, and then so is the right side. The left side is also enclosed directly from the enclosures of $h_m$. Both enclosures contain the exact value, so their intersection does. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:enc-blocks}] (a) For $\tau_0 \le \tau \le \tau_1$, $\mathbf 1_{[1/2 - \tau_0, 1/2)} \le \mathbf 1_{[1/2 - \tau, 1/2)} \le \mathbf 1_{[1/2 - \tau_1, 1/2)}$, and integrating against a nonnegative factor preserves the order. (b) The difference $X_\tau(x) - X_{\tau_0}(x)$ is the integral of $r(1/2 - t)\, g(x - 1/2 + t)$ over $t \in [\tau_0, \tau]$, and $r \ge 0$; replacing $g$ by its minimum and its maximum over the arguments that occur gives the bounds. When the correction is integrated against a signed function $H$, the bounds are applied separately to the integrals against $H^+$ and $H^-$ (Lemma~\ref{lem:enc-sign}), since $g$ depends on the integration variables. (c) is immediate. Under the additional hypotheses stated in the lemma, bounded integrated density proves the rate in (a), and the Lipschitz range bound proves it in (c). In (b), bounded $r$ gives $\rho(\tau)=O(\tau_1-\tau_0)$ and the Lipschitz bound gives variation $O(\tau_1-\tau_0+1/M)$ on the window. Multiplying proves the stated additional block loss. These rate arguments are not used on windows crossing a jump. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:enc-pos}] (a) On the cell, $a\,\Phi(a) \ge a\,L(a) =: q(a)$, as $a > 0$, and $q$ is quadratic with $q'' = 2M(L_1 - L_0)$. If $L_1 \le L_0$, $q$ is concave and its minimum is at an end of the cell. Otherwise \eqref{eq:chord} gives $q \ge \ell - \frac{1}{8M^2}\cdot 2M(L_1 - L_0)$, and the chord $\ell$ of $q$ is at least the smaller of its end values $x_n L_0$ and $x_{n+1}L_1$. (b) For fixed $a$, minimising over $t$ and over the interval for $f_g$ uses its four corners. Replacing $f_c$ by $L_c$ gives the stated minimum of affine functions, hence a concave function above its end chord. (c) Pass to the limit from within a cell. The assertion about the assigned point value then follows from the stated convention; it does not follow without that convention. The remaining collapsed strata are checked directly from the class constants and the support of the forced terms. \end{proof} \begin{proof}[Proof of Lemma~\ref{lem:enc-binary64}] (a) Round to nearest returns a binary64 number $y$ with $|y - z| \le \frac12 \mathrm{ulp}$ for the exact result $z$, where the unit in the last place is that of the binade in which the rounding takes place, unless the operation overflows. So $z$ lies between the neighbours of $y$ below and above. For (b), expand each computed term and each addition along its path to the final sum as multiplication by factors $1+\delta$, with $|\delta|\le u$. A path has at most $k-1+r$ factors, and the standard product bound \cite[Lemma~3.1]{Higham} gives $|\prod(1+\delta)-1|\le\gamma_{k-1+r}$ when $(k-1+r)u<1$. Taking absolute values and summing proves the claim. For additions of binary64 inputs an exact subnormal result is representable, so such an addition is exact; the relative-error hypothesis holds for products and quotients with normal results. Source-specific removal of true-zero terms in the raw nonnegative accumulators is described after the lemma. Individually widened operations instead use part (a), including for subnormal results. (c) Write $x = k\,2^e$ with integers $|k| < 2^{53}$ and $e \ge -1074$. Then $x\,2^S = k\,2^{e + S}$ with $e + S \ge -1074$ and $|x\,2^S| < 2^{52}$, so it is a binary64 number and the multiplication is exact. The integers $\lfloor x2^S\rfloor$ and $\lceil x2^S\rceil$ have absolute value at most $2^{52}$, so they are exactly representable, and they bound $x2^S$ below and above. Convolutions of such integer images are exact integer polynomial products. \end{proof}