It Takes Two Moments: Rigidity and Limits of Continuous Encoders for Over-the-Air Computation

Let K\ge4. For every node k, let \phi_{1,k},\phi_{2,k}\in C(\mathbb I)\cap C^1(\mathbb I^\circ) and suppose that 1,\phi_{1,k},\phi_{2,k} are linearly independent. For j=1,2, let \psi_j\in C(\mathbb J_j)\cap C^1(\mathbb J_j^\circ) and suppose that \psi_j is either affine on \mathbb J_j or nonaffine on every nonempty open subinterval of \mathbb J_j^\circ. If both receiver postprocessing functions \psi_1 and \psi_2 are nonconstant and at least one is nonaffine, a collection of continuous encoders permits identifiable computation if and only if

where c_1,\ldots,c_K,u,v\in\mathbb C and u,v are linearly independent over \mathbb R.

Let K\ge2 and P>0. Assume that all local preprocessing functions \phi_{j,k} and both receiver postprocessing functions \psi_1,\psi_2 are continuous on their prescribed domains and that f is nonconstant on \mathcal M_K. Then

where \lambda_\star is the attained optimum of the convex program

Here, G=(H_1)_{2:3,2:3} is the common lower-right block. Moreover, \lambda_\star>0 if and only if f is Lipschitz on \mathcal M_K. Under the assumptions of Theorem 1, we also have \Lambda^\star(P;f)=K^2\lambda_\star, so the same program optimizes robustness over the full class \mathcal E_P.

Let K\ge1, let \mathscr{E} be a collection of continuous encoders, and let f:\mathbb I^K\to\mathbb R be continuous and nonconstant. Then

with 0^{-1/2}=+\infty. A continuous decoder attains the infimum whenever it is finite.

Under the assumptions of Theorem 2, fix an encoder collection of the form (1) with \Lambda(\mathscr{E};f)>0. Using c and \mathsf U from Section ?, the decoder

satisfies

It therefore attains the optimal uniform noise-amplification constant L_\mathscr{E}=\Lambda(\mathscr{E};f)^{-1/2}.

For a fixed continuous encoder collection \mathscr{E}, a continuous desired function f, and a known noise bound |z|\le r_n with r_n\ge0, the minimum worst-case absolute error is \omega_\mathscr{E}(2r_n)/2, where

An optimal decoder outputs the midpoint between the smallest and largest values of f consistent with the observation and noise bound.

Assume \mathcal M_K=\overline{\operatorname{int}\mathcal M_K}. For every fixed encoder collection \mathscr{E}\in\mathcal B_P,

If, in addition, 1,\phi_{1,k},\phi_{2,k} are linearly independent for every node k and the vectors \{\boldsymbol w(\boldsymbol m):\boldsymbol m\in \operatorname{int}\mathcal M_K\} span \mathbb R^2, then

The minimum is attained. Both asymptotic coefficients are achievable using encoders and decoders that are independent of the noise power.

Fix \mathscr{E}\in\mathcal B_P and suppose that \pi is a regular moment distribution. Then

The decoder { \mathscr{D}_\mathscr{E}} of Theorem 3 attains this coefficient.

Let K\ge4. Assume that \phi\in C(\mathbb I)\cap C^1(\mathbb I^\circ) and \psi\in C^1(\mathbb I) are each strictly convex or strictly concave. For the desired function (?), every encoder collection \mathscr{E}_k(s)=c_k+us+v\phi(s) with G\succ0 satisfies

where \boldsymbol w_\pm=(-\psi'(\pm1),1)^\mathsf T. The infimum defining \Lambda(\mathscr{E};f_{\phi,\psi}) in (?) is approached by attainable pairs whose noiseless aggregate distance tends to zero.

Let K\ge4, P>0, and let \psi\in C^1(\mathbb I) be even and strictly convex or strictly concave. Write \chi_p(s)=s^p and M=|\psi'(1)|. For (?) with \phi=\chi_2 and \phi=\chi_4, respectively, the shared encoders

and

maximize robustness over all continuous encoder collections satisfying (?). The quadratic coefficients are b=\frac{2\sqrt{M^2+8}+2M}{\sqrt{M^2+8}+3M} and a=\sqrt{b(2-b)}. For the quartic encoder, \rho\in(0,1/3) is the unique solution of

and d_\rho=1+2\rho+9\rho^2, a_\rho=\sqrt{\frac{8\rho}{d_\rho}}, b_\rho=\frac{2}{(1+\rho)\sqrt{d_\rho}}, and c_\rho=-\frac{1+2\rho+3\rho^2}{(1+\rho)\sqrt{d_\rho}}.

Let K\ge1, P>0, and \phi_{1,k},\phi_{2,k}\in C^2(\mathbb I) for every node k. Assume that 1,\phi_{1,k},\phi_{2,k} are linearly independent for each k, and that f is continuous and nonconstant on \mathcal M_K. Let \lambda_\star be the optimum of the exact secant program (2). For each node, choose finite grids \mathcal T_{k,h} containing both endpoints and three fixed inputs whose vectors (1,\phi_{1,k}(s),\phi_{2,k}(s))^\mathsf T span \mathbb R^3; each grid has maximum adjacent spacing at most h. Replace only the peak constraints in (2) by those at \mathcal T_{k,h}, retain every pairwise received-separation constraint, and denote an optimizer by (H_{1,h},\ldots,H_{K,h},{ \lambda_h}). Write G_h=(H_{1,h})_{2:3,2:3} for its common block. Such an optimizer exists. Define

Set

Then (\rho_hH_{1,h},\ldots,\rho_hH_{K,h},\rho_h{ \lambda_h}) is feasible for (2), and

The reconstruction in Section ? yields peak-feasible encoders \mathscr{E} with robustness score \Lambda(\mathscr{E};f)\ge K^2\rho_h\lambda_h. Under the assumptions of Theorem 1, the bounds also certify the optimum over all continuous peak-feasible encoders.