Cramér-Rao Bound Analysis and Near-Optimal Performance of the Synchronous Nyquist-Folding Generalized Eigenvalue Method (SNGEM) for Sub-Nyquist Multi-Tone Parameter Estimation
Huiguang Zhang · 2026
Read Cramér-Rao Bound Analysis and Near-Optimal Performance of the Synchronous Nyquist-Folding Generalized Eigenvalue Method (SNGEM) for Sub-Nyquist Multi-Tone Parameter Estimation by Huiguang Zhang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-20866.txt.
Introduction
Huiguang Zhang analyzes the accuracy limits of a method for reading signals sampled too slowly. The 2026 paper concerns the synchronous Nyquist-folding generalized eigenvalue method, or SNGEM. The method estimates frequency, amplitude, and phase for multi-tone signals. It works at extreme sub-Nyquist rates, far below the usual sampling rule. It does so by processing a signal together with its time derivative.
Zhang derives the exact Cramér-Rao bound for one key parameter. That bound is the lowest error any estimator can reach. The parameter is an amplitude ratio tied to the tone's frequency. The paper then checks how close SNGEM comes to that limit. Near-optimal performance would certify the method for hard sensing jobs. Such jobs include radar and spectrum monitoring with cheap, slow hardware.
Catalog record
| CREATOR | Huiguang Zhang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 5 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-20866
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Cramér-Rao Bound Analysis and Near-Optimal Performance of the Synchronous Nyquist-Folding Generalized Eigenvalue Method (SNGEM) for Sub-Nyquist Multi-Tone Parameter Estimation
Huiguang Zhang
Abstract
The Synchronous Nyquist-folding Generalized Eigenvalue Method (SNGEM) achieves full frequency/amplitude/phase estimation for multi-tone signals at extreme sub-Nyquist rates by jointly processing the original signal and its time-derivative (differential channel). This paper derives the exact Cramér-Rao bound for the amplitude-ratio parameter R = A / B = 1 / ( 2 π f ) R=A/B=1/(2\pi f) under equal-SNR dual-channel Gaussian noise. We prove that frequency can be indirectly estimated from the ratio with relative variance only 2× worse than direct single-channel estimation—an astonishingly small penalty that explains SNGEM’s remarkable robustness. Monte-Carlo simulations confirm that, even at 10–20× compression, SNGEM reaches machine-precision accuracy in noise-free conditions and tightly approaches the derived CRB across all SNR levels, while classical compressive-sensing OMP exhibits an irreducible error floor due to DFT grid bias and aliasing noise. These results establish SNGEM as statistically near-optimal for deterministic sub-Nyquist parametric spectral analysis.
I
Introduction
The ever-increasing demand for wideband spectral monitoring in radar, 6G communications, and instrumentation has driven sampling rates toward tens of GS/s, far exceeding ADC capabilities. Classical compressive sensing (CS) recovers sparse spectra via
\ell_{1}
or greedy methods but suffers from basis mismatch, off-grid bias, and poor amplitude/phase accuracy. The Synchronous Nyquist-folding Generalized Eigenvalue Method (SNGEM)
[
1
,
2
]
circumvents random measurement matrices by synchronously acquiring the signal
x[n]
and its discrete derivative
\dot{x}[n]
, enabling deterministic generalized eigenvalue solutions at compression ratios
>
10× while preserving full parameter accuracy.
This paper provides the missing statistical efficiency analysis via closed-form Cramér-Rao bounds for the dual-channel amplitude-ratio parameter.
II
Signal Model and Dual-Channel Observation
Consider a single real tone embedded in dual-channel additive white Gaussian noise:
\displaystyle x[n]
\displaystyle=A\cos(2\pi fnT_{s}+\phi)+w[n],
(1)
\displaystyle\dot{x}[n]
\displaystyle=-A(2\pi f)\sin(2\pi fnT_{s}+\phi)+\dot{w}[n],
(2)
where
B=2\pi fA
exactly (ideal differentiator). Both channels have identical noise variance
\sigma^{2}
, hence identical SNR:
\text{SNR}_{x}={SNR}_{\dot{x}}=\frac{A^{2}}{2\sigma^{2}}.
(3)
III
Closed-Form Cramér-Rao Bound for Amplitude-Ratio Frequency Estimation
The Fisher information for amplitude estimates
A
and
B
in each channel is classical
[
3
,
4
]
:
\text{Var}(\hat{A})\geq\frac{2\sigma^{2}}{N},\quad\text{Var}(\hat{B})\geq\frac{2\sigma^{2}}{N}\quad\Rightarrow\quad\frac{\text{Var}(\hat{A})}{A^{2}}\geq\frac{4}{N\cdot\text{SNR}_{x}}.
(4)
Define the amplitude-ratio parameter
R=A/B=1/(2\pi f)
. For independent Gaussian estimates
\hat{A}
,
\hat{B}
, the relative variance of ratio
R
is
\frac{\text{Var}(\hat{R})}{R^{2}}=\frac{\text{Var}(\hat{A})}{A^{2}}+\frac{\text{Var}(\hat{B})}{B^{2}}.
(5)
Substituting the individual CRBs and using
B=2\pi fA
yields
\frac{\text{Var}(\hat{R})}{R^{2}}\geq\frac{8}{N\cdot\text{SNR}}\quad\Rightarrow\quad\text{CRB}(\hat{R})=\frac{2}{N\cdot\text{SNR}}R^{2}.
(6)
Since
f=1/(2\pi R)
, the relative frequency CRB is
\boxed{\frac{\text{Var}(\hat{f})}{f^{2}}\geq\frac{2}{N\cdot\text{SNR}}}
(7)
— exactly 2× (3 dB) worse than the classical single-channel deterministic CRB
\text{Var}(\hat{f})/f^{2}\geq 1/(N\cdot\text{SNR})
[
3
]
.
Remark 1
.
This astonishingly small 3 dB penalty explains why SNGEM remains highly accurate even under extreme compression: frequency is extracted from a ratio of two equally noisy but highly correlated measurements.
Fig. 1
:
Relative frequency RMSE vs. SNR for
f=100
MHz, compression ratio 15× (
f_{s}=133
MS/s). SNGEM tightly follows the derived dual-channel CRB (solid black), while OMP exhibits irreducible floor
>10^{-4}
due to grid bias and aliasing noise.
IV
Monte-Carlo Validation and Comparison with Compressive Sensing
5000 independent runs were performed with 5–15 real tones, random phases, compression ratios 8–20× via synchronized random undersampling + ideal differentiation. Key findings:
•
Noise-free case: SNGEM achieves frequency/amplitude/phase RMSE
<10^{-14}
(machine precision) for all tested rates.
•
Noisy case (Fig.
1
): SNGEM tracks the theoretical dual-channel CRB within 1.1–1.3× across -10 to 50 dB SNR.
•
Orthogonal Matching Pursuit (OMP) on the same measurements shows error floor
>10^{-5}
–
10^{-4}
independent of SNR, confirming inherent bias predicted by theory.
V
Conclusion