Eigenvalue Estimates and Mutual Information for the Linear Time-Varying Channel

Brendan Farrell, Thomas Strohmer

Abstract

We consider linear time-varying channels with additive white Gaussian noise. For a large class of such channels we derive rigorous estimates of the eigenvalues of the correlation matrix of the effective channel in terms of the sampled time-varying transfer function and, thus, provide a theoretical justification for a relationship that has been frequently observed in the literature. We then use this eigenvalue estimate to derive an estimate of the mutual information of the channel. Our approach is constructive and is based on a careful balance of the trade-off between approximate operator diagonalization, signal dimension loss, and accuracy of eigenvalue estimates.

Introduction

In classical information theory, Shannon’s capacity formula for linear time-invariant (LTI) channels provides a foundational mathematical convenience: the eigenvalues of the channel matrix are directly related to uniformly spaced samples of the channel's transfer function. This relationship allows for a straightforward characterization of channel capacity based on the power spectrum.

However, real-world wireless environments—particularly those involving high mobility, such as vehicle-to-vehicle communication—are fundamentally linear time-varying (LTV). In these scenarios, the channel operator does not easily diagonalize, making it difficult to characterize the channel's spectrum (eigenvalues) via its symbol (transfer function). This paper addresses this longstanding gap by providing a rigorous mathematical framework for estimating the eigenvalues and mutual information of a broad class of LTV channels, specifically those whose spreading functions exhibit exponential decay in both time and frequency.

Overcoming the Balian-Low Roadblock

To approximate the eigenvalues of an LTV channel, one must identify a set of transmission signals that "approximately diagonalize" the channel. This requires signals that are highly localized in both time and frequency. However, the Balian-Low theorem presents a fundamental obstacle: it is impossible to construct a complete, orthonormal basis of signals that are perfectly localized in both domains.

To navigate this, the authors employ a two-pronged mathematical strategy:

  1. Weyl-Heisenberg (Gabor) Systems: To achieve the necessary exponential localization and force the off-diagonal entries of the channel matrix to decay rapidly, the authors utilize Weyl-Heisenberg signals. To circumvent the Balian-Low theorem, they use an incomplete set of these signals at the transmitter and an overcomplete tight frame at the receiver. This ensures strict localization at the cost of some signal dimensionality.
  2. Prolate Spheroidal Wave Functions (PSWFs): Because the Weyl-Heisenberg set is incomplete, there is an inherent "dimension loss" in the signal space. To quantify and bound this loss, the authors map the signal space to PSWFs. These functions are optimally localized within a finite time-bandwidth region, providing the necessary tools for dimension-counting, even though they lack the exponential temporal decay required for eigenvalue estimation.

By balancing the diagonalization properties of Weyl-Heisenberg systems with the dimension-counting capabilities of PSWFs, the authors model the channel as a pseudodifferential operator and rigorously bound its behavior.

Key Theoretical Insights

The Twisted Autocorrelation

In LTI channels, capacity is tied to the power spectrum (the Fourier transform of the impulse response's autocorrelation). For LTV channels, this paper proves that the eigenvalues are closely approximated by samples of the twisted autocorrelation of the time-varying transfer function's symbol. This provides a formal LTV analogue to the LTI power spectrum.

Rigorous Trade-offs

A central contribution of this work is the explicit quantification of the trade-off between three competing factors:

  • The precision of the channel matrix diagonalization.
  • The loss of signal dimensions due to the use of incomplete sets.
  • The accuracy of the resulting eigenvalue and mutual information estimates.

Principal Results

Eigenvalue Estimate (Theorem 3.5)

The authors prove that the eigenvalues of the effective channel's correlation matrix $A^*A$ can be approximated by samples of the twisted autocorrelation $S$. They establish a strict error bound that decays exponentially: $$O(e^{-\frac{\rho}{2}(\beta+\alpha)} + \frac{1}{(\alpha\beta D)^2})$$ where $\alpha$ and $\beta$ represent the decay rates of the channel's spreading function, and $D$ relates to the signal dimensions.

Mutual Information Estimate (Theorem 4.1)

Building on the eigenvalue bounds, the authors derive a rigorous approximation for the mutual information of the LTV channel. The mutual information is shown to be approximately: $$\sum \log(1 + \frac{S^+}{\eta^2})$$ This result mirrors the classic Shannon formula while incorporating explicit, bounded error terms that isolate the inaccuracies caused by eigenvalue approximation, the transition to PSWFs, and the dimension loss from the Weyl-Heisenberg construction.

Limitations and Future Directions

While this framework provides a rigorous foundation, it relies on the assumption that the channel's spreading function decays exponentially. While the proofs could be adapted for polynomial decay, the resulting error bounds would converge significantly slower.

Furthermore, the paper estimates the mutual information for a specific, constructed signaling scheme rather than the absolute channel capacity (the supremum over all possible input distributions). Finally, the mathematical construction relies on a redundancy parameter $\rho > 1$. Optimizing the balance between $\rho$ (to minimize dimension loss) and the accuracy of the eigenvalue approximation remains an open challenge for practical system design.

Summary updated

Abstract

We consider linear time-varying channels with additive white Gaussian noise. For a large class of such channels we derive rigorous estimates of the eigenvalues of the correlation matrix of the effective channel in terms of the sampled time-varying transfer function and, thus, provide a theoretical justification for a relationship that has been frequently observed in the literature. We then use this eigenvalue estimate to derive an estimate of the mutual information of the channel. Our approach is constructive and is based on a careful balance of the trade-off between approximate operator diagonalization, signal dimension loss, and accuracy of eigenvalue estimates.
BibTeX
@article{farrell2011eigenvalue,
  author = {Brendan Farrell and Thomas Strohmer},
  title = {{Eigenvalue Estimates and Mutual Information for the Linear Time-Varying Channel}},
  journal = {IEEE Trans. Information Theory},
  year = {2011},
  url = {https://openreview.net/forum?id=FMYTA-QLj1},
}

Loading related content...

Figure