In modern wireless communications and radar systems, a signal rarely carries information along just a single dimension. When a wave reflects off a moving vehicle or propagates through a 5G/6G massive MIMO channel, it arrives with a specific angle of arrival, after a specific time delay, and with a specific Doppler shift.
Traditionally, signal processing algorithms estimated these parameters either independently or by discretizing the parameter space into a fine grid. However, independent estimation discards the rich, coupled structure of the signal, while grid-based methods suffer from the "curse of dimensionality" (a 3D grid scales exponentially) and grid mismatch errors.
Multi-dimensional and multivariate ($M^2$) spectral estimation bypasses these limitations. By treating physical measurements as a continuous superposition of complex sinusoids across multiple coupled dimensions, this framework allows systems to jointly resolve parameters with super-resolution accuracy and manageable computational complexity.
Subspace Decomposition: Exploiting Shift Invariance Across Dimensions
The core intuition behind high-resolution subspace methods—such as ESPRIT, MUSIC, and the Matrix Pencil approach—is shift invariance. If you sample a complex sinusoid at equal intervals in time or space, a step forward corresponds to a fixed phase rotation. In multi-dimensional systems, this phase rotation happens simultaneously along multiple axes.
For instance, in a mobile massive MIMO system, the channel changes across both the frequency domain (due to path delays) and the time domain (due to user mobility). Weidong Li et al. extended the matrix pencil technique to a multi-dimensional framework (MDMP) to exploit this joint structure. By analyzing the channel's coupled angular-frequency and angular-time domains, the authors successfully estimated elevation/azimuth angles of departure and path delays simultaneously. In later work on fluid antenna systems, Weidong Li et al. adapted a 2D Matrix Pencil algorithm to capture spatial and temporal dimensions concurrently, transforming highly time-varying mobile channels into predictable, quasi-static structures.
Physical Dimensions (Angle, Delay, Doppler)
│
▼ [Kronecker / Tensor Structure]
High-Dimensional Coupled Measurement Matrix
│
▼ [Receive Combining / Subspace Projection]
Low-Dimensional Digital Array (Preserves Shift Invariance)
│
▼ [Matrix Pencil / Unitary ESPRIT]
Decoupled, Continuous Parameter Extraction
When the number of antennas scales up to hundreds or thousands, processing these massive matrices directly becomes computationally prohibitive. To solve this, researchers project the physical channel into a compressed digital subspace. Anwen Liao et al. demonstrated that hybrid antenna arrays can be mathematically transformed into low-dimensional digital arrays using specially designed receive combining matrices. This transformation preserves the underlying spatial angles, allowing a Multi-Dimensional Unitary ESPRIT (MDU-ESPRIT) algorithm to resolve the angles of arrival and departure directly from the low-dimensional representation.
This mathematical compression drastically reduces computational overhead without sacrificing accuracy. As Ping Zhang et al. noted, Unitary ESPRIT further optimizes this process by mapping complex-valued matrices into real-valued ones, cutting the computational complexity of joint parameter and channel estimation in massive MIMO systems. Similarly, Zhipeng Lin et al. showed that exploiting the structured invariance of a transformed antenna manifold allows high-resolution subspace methods to completely decouple the computational complexity from the total number of physical antennas. When additional signal dimensions like polarization are introduced, Frédéric Pascal et al. proved that Tensor MUSIC can successfully isolate and resolve both direction and polarization parameters simultaneously from complex, multi-dimensional data blocks.
Convex Optimization: Super-Resolution in Continuous Space
An alternative to algebraic subspace methods is framing multi-dimensional spectral estimation as a continuous optimization problem. Instead of picking the best fits from a discrete grid, the goal is to find the exact sparse frequencies in a continuous domain that match the observed data.
Reinhard Heckel demonstrated that Multiple-Input Multiple-Output (MIMO) radar can achieve perfect "super-resolution"—recovering the exact, continuous angles, delays, and Doppler shifts of targets—by solving a convex optimization problem.
Why does this work? The radar signal processing task is modeled as a multi-dimensional line spectral estimation problem. Because the physical targets are sparse in the environment, the received data is a superposition of a few distinct complex sinusoids. Reinhard Heckel proved that as long as the targets are sufficiently separated in the multi-dimensional parameter space, convex optimization can perfectly recover the continuous parameters, entirely bypassing the resolution limits imposed by traditional matched filtering and compressed sensing grids.
Multivariate Multidimensional ($M^2$) Covariance Extension
The problem becomes even more intricate when dealing with multivariate data—where multiple distinct sensors or channels each observe a multidimensional field. A prime example is an autonomous vehicle equipped with an array of overlapping radar sensors.
Bin Zhu et al. framed this automotive sensor fusion task as a multivariate multidimensional ($M^2$) spectral estimation problem. By modeling the measurements from multiple sensors as channels of a single complex signal characterized by a multivariate spectrum, the authors leveraged the cross-spectra between sensors to significantly improve target localization and velocity tracking.
Observed Covariance Lags
│
▼
Optimization Problem: Minimize Relative Entropy (Itakura-Saito / Divergence)
│
▼
Guaranteed Non-Negative, Rational Spectral Density (Finite-Dimensional Realization)
However, estimating an $M^2$ spectrum from real-world data introduces a fundamental mathematical hurdle known as the covariance extension problem. In practice, we only observe a finite number of correlation or covariance "lags." We must find a spectral density function that perfectly matches these observed lags while remaining physically valid (i.e., non-negative everywhere). Traditional multidimensional estimation methods often break down here, yielding singular spectral measures that contain mathematically problematic "shocks" or singular parts, making them impossible to implement in finite-dimensional engineering systems.
To overcome this, Bin Zhu et al. introduced a relative entropy approach using the Itakura-Saito distance as an optimization criterion, establishing a well-posed framework for $M^2$ signals. In a companion paper, Bin Zhu et al. designed a specialized family of divergences for this optimization. By minimizing these divergences subject to the observed covariance constraints, the optimization mathematically guarantees a rational spectral density solution. This rationality ensures that the resulting spectral estimate corresponds to a clean, finite-dimensional system representation, making $M^2$ spectral estimation fully viable for real-time, practical engineering applications.
Key Foundational Work
Super-Resolution MIMO Radar (2016) — contributed 1 concept elements to this direction.
Special issue on 5G wireless communication systems and technologies (2017) — contributed 1 concept elements to this direction.
Derivation of the theoretical performance of a Tensor MUSIC algorithm (2016) — contributed 1 concept elements to this direction.