1: \begin{abstract}

2: This paper presents \proposed, a Decentralized Gradient Descent (DGD) algorithm that 

3: combines local gradient descent with a novel Non-Coherent Over-The-Air (NCOTA) consensus scheme to solve 

4: distributed machine-learning problems over wirelessly-connected systems.

5: \proposed\ leverages the

6: waveform superposition properties of the wireless channels:

7: it enables simultaneous transmissions under half-duplex constraints, by mapping local optimization signals to a mixture of preamble sequences, and consensus

8: via  non-coherent combining at the receivers.

9: \proposed\ operates without channel state information at transmitters and receivers, and leverages the average channel pathloss to mix signals,

10:  without explicit knowledge of the mixing weights (typically known in

11:  consensus-based optimization algorithms).

12: It is shown both theoretically and numerically that, for smooth and strongly-convex problems with fixed consensus and learning stepsizes, the updates of \proposed\ converge in Euclidean distance to the global optimum with rate

13: $\mathcal O(K^{-1/4})$ for a target of $K$ iterations.

14: \proposed\ is evaluated numerically  over a logistic regression problem, showing

15: faster convergence vis-\`a-vis running time than  implementations of the classical DGD algorithm over  digital and analog orthogonal channels.

16: \end{abstract}

17: