An antenna array estimates the direction of an incoming signal by measuring phase differences across its elements. For narrowband signals, this works cleanly — the phase at each antenna is a linear function of direction, and solving for direction is a linear algebra problem. For ultra-wideband signals, three problems compound: the array may be spatially undersampled (elements too far apart for the highest frequency), the absolute phase at each element may be unknown (asynchronous hardware), and the beam direction shifts with frequency (beam squint). Each problem has been solved independently. Tian and colleagues solve all three simultaneously by reframing the problem geometrically.
Instead of treating the received signal as a vector of complex amplitudes, they represent it as a curve in a high-dimensional space parameterized by frequency. Different directions of arrival produce curves with different geometric shapes — different curvatures, torsions, and orientations in this space. The direction of arrival is encoded not in the absolute position of the curve (which depends on unknown phase offsets) but in its shape (which doesn't).
Note (2026-11-13; read against the paper's full text): the paper does not say one set of shape invariants absorbs all three problems. It splits its geometric parameters into two kinds. The intrinsic invariants (curvature and the like, unchanged under rigid-body transformations of the curve) are the ones it ties to robustness against uncalibrated array phase errors. Resolving the spatial ambiguity of a sparse array is done by its first framework, which uses the extrinsic vector parameters of the whole trajectory; the paper calls these a dynamic extension of conventional array processing, and by definition they are not invariant. The essay says the curve is parameterized by frequency; the abstract describes a manifold curve 'driven by temporal evolution and array topology', and I did not find a frequency parameterization stated in the passages I searched. The paper treats beam squint as frequency non-stationarity of the signal in its signal model; I did not find a statement that squint changes the parameterization while leaving invariants untouched. The validation is numerical simulation (a chapter of simulated multipath, linear and sinusoidal frequency-modulated signals), not hardware, and the paper adds that feature extraction uses high-order derivatives that amplify noise, so the framework suits high-SNR conditions or large apertures; the essay's closing claim that the three problems become one naturally solved problem leaves that out.
This geometric representation absorbs all three problems at once. Spatial undersampling creates ambiguity in phase, but not in the shape of the curve. Unknown absolute phase shifts the curve without changing its shape. Beam squint changes the parameterization of the curve but preserves its geometric invariants. The direction of arrival is recovered from shape features — curvature ratios, projection angles — that are invariant to all three sources of corruption.
The physics didn't change. The representation did. By lifting the signal from amplitudes to geometry, the three problems that were hard separately become one problem that is naturally solved.