REVIEW 3 major objections 5 minor
Analysis of random telegraph noise in resistive memories: The case of unstable filaments
T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read An adaptive moving-average detrending filter can flatten unstable RTN signals in ReRAM devices, making trap capture and emission time constants measurable where standard threshold-based analysis fails.
desk verdict Useful detrending method with a solid synthetic test, but the experimental validation is circular and needs independent confirmation before the accuracy claim can be taken at face value. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the adaptive moving-average detrending filter, defined by $y_M(k) = x(k) - (x*G_M)(k) + \overline{x(k)}$, where $G_M$ is an $M$-tap uniform averaging mask. The detrended signal's usefulness is steered by the cost criterion $C = a(\sigma_c+\sigma_e)/|\mu_c-\mu_e| + b\,|f_0 - (1/2\pi)(1/\tau_c+1/\tau_e)|/N$, which combines histogram separation and spread of the two RTN levels with consistency with the Lorentzian corner frequency; minimization over mask size is performed with a global optimizer (SHGO). The known identity $f_0 = (1/2\pi)(1/\tau_c+1/\tau_e)$ for single-trap RTN is what connects the time-domain estimates to the frequency-domain cross-check.
What would settle it
A direct test would be to synthesize unstable RTN signals with known $\tau_c$ and $\tau_e$, add a slow drift whose corner frequency is close to the RTN corner frequency (violating the stated separation), run the adaptive filter, and check whether the recovered time constants match the ground truth; if they do, the band-separation assumption is not necessary, and if they do not, the method's range of validity is bounded by that assumption.
Extended reading notes
Core claim
The central claim is that an adaptive moving-average detrending filter can flatten unstable RTN signals well enough that a simple threshold separates the two levels and the trap time constants can be measured accurately. The filter subtracts a moving average of the signal and adds back the global mean, and the mask size is not chosen by trial but by a minimization algorithm: the cost function combines the separation and spread of the two histogram lobes with a term that penalizes deviation of the estimated corner frequency from the Lorentzian prediction. On a real unstable RTN measurement from an SiN$_x$ ReRAM cell tuned to an intermediate resistance state, the filter produces a detrended signal whose histogram shows two clean Gaussian lobes, and the extracted $\tau_c$ and $\tau_e$ satisfy the consistency relation $f_0 = (1/2\pi)(1/\tau_c + 1/\tau_e)$ with the corner frequency $f_0=985$ Hz obtained from the power spectral density. The paper claims this cross-validation demonstrates the accuracy of the proposed method.
Load-bearing premise
The method assumes that the slow instability in the measured signal lives at a corner frequency at least an order of magnitude below the RTN corner frequency and with at least three orders of magnitude less power, so that moving-average detrending removes the drift without touching the two-level statistics.
Editorial extensions
If this is right
- Unstable RTN signals, previously unusable for time-domain analysis, become analyzable for single-trap capture and emission time constants.
- The method gives a criterion for choosing detrending filter width without knowing the ground-truth signal, using only measurable histogram and PSD parameters.
- The extracted $\tau_c$ and $\tau_e$ satisfy the Lorentzian consistency relation, providing a built-in validation that the measured two-level signal is from a single trap.
- The approach is intended to support multi-level cell (MLC) tuning, where a failed tuning protocol leaves the resistance state drifting, by allowing trap parameters to be read from the noisy state.
Reading between the lines
- Because the method assumes the instability lies at much lower frequencies and much lower power than the RTN corner frequency, it will likely fail when filament drift is comparable in timescale to the trap switching; a testable extension is to map the failure boundary in the $(f_0'/f_0, P'/P)$ plane on simulated signals.
- The same detrending-then-fit logic could be applied to other two-level fluctuators with slow baseline drift, such as qubit charge noise or molecular conductance switches, where the Lorentzian relation plays the same validation role.
- A direct comparison against Hidden Markov Model fitting on the same unstable signals would clarify whether the advantage comes from detrending or from the specific cost criterion; the paper cites HMM results but does not benchmark its own output against them.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes an adaptive moving-average detrending filter to flatten unstable random telegraph noise (RTN) signals in silicon nitride ReRAM devices, where conductive-filament instabilities cause slow drift that obscures two-level switching. The filter optimizes the moving-average mask size by minimizing a composite criterion C (Eq. 4) that includes separation of histogram lobes (term A) and consistency with the RTN Lorentzian corner frequency f0 (term B). The method is validated on a simulated RTN signal with injected 1/f^3 instability and on one measured unstable RTN signal; the authors report that the estimated capture/emission time constants satisfy the Lorentzian relation f0 = (1/2π)(1/τc + 1/τe).
Significance. The problem addressed is real: RTN analysis in ReRAM multi-level cells is often corrupted by slow filament instabilities, and standard threshold-based time-domain analysis fails when levels drift. The synthetic validation with known ground truth is a strength; the paper shows that the optimizer recovers injected time constants and that the adaptive filter produces a histogram with separable Gaussian lobes. However, the experimental validation is weakened by circularity: term B of the optimization criterion (Eq. 4) directly enforces the Lorentzian relation that is later used as the 'confirmation criterion' (Eq. 5 and Fig. 5(f)), so the agreement is a constrained outcome rather than an independent cross-check. The method may well be useful, but the current experimental evidence is a self-consistency check, not a true cross-validation.
major comments (3)
- [Section 3, Eq. (4)-(5), Fig. 5(f)] The experimental 'cross-validation' is not independent. The optimization criterion C in Eq. (4) includes term B = |f0 - (1/2π)(1/τc+1/τe)|/N, with f0 obtained from Lorentzian fitting of the PSD of the same unstable signal. Section 3 then confirms the estimated τc and τe by checking Eq. (5), which is exactly the relation minimized in term B. Consequently, Fig. 5(f) does not provide independent evidence that the method is accurate; it merely shows that the optimizer satisfied its own constraint. To support the claimed cross-validation, the authors should either (i) compare the time-domain estimates with an independently fitted Lorentzian to the stabilized signal's PSD, or (ii) use an alternative independent estimation method, or (iii) explicitly rephrase the result as a self-consistency check rather than a proof of accuracy.
- [Section 2.2] The method rests on a timescale-separation assumption: the instability must have a corner frequency f0' at least an order of magnitude below the RTN corner frequency f0 and a power at least three orders of magnitude lower. This condition is verified for the simulated signal by construction, but for the measured device the paper does not quantitatively verify that the experimental drift satisfies these bounds. The bottom panel of Fig. 5(b) shows that subtracted spectral content lies below f0, but no power-ratio quantification is provided. Since the moving-average filter is not a sharp high-pass filter, partial overlap between the instability band and the RTN Lorentzian band would bias the estimated τc and τe. The authors should quantify the instability contribution (e.g., from the PSD of the subtracted signal) and show that the separation condition holds for the measured data, or state the limitation explicitly.
- [Section 3 and Fig. 4] The paper does not report quantitative accuracy metrics for the time-constant estimates. In the synthetic experiment (Fig. 4), the authors state that the time constants are 'calculated accurately,' but no numerical comparison with the injected values (e.g., relative errors or confidence bounds) is given. For the experimental signal (Fig. 5), no error bars or comparison with any independent estimate are provided. Given that the central claim is that the method increases accuracy, quantitative error reporting on the simulated ground truth is essential and would also make the synthetic validation more convincing.
minor comments (5)
- [Abstract] The abstract contains typographical errors: 'tunning protocol' should be 'tuning protocol,' and 'The te and tc emission/capture time constants' should use τe and τc.
- [Eq. (4)] The scaling parameters are introduced as 'α and b' in the text but written as 'a' and 'b' in Eq. (4); please unify the notation.
- [Section 2.2] The notation '1/f 2' and '1/f 3' in the text and Fig. 2 should be typeset as superscripts (1/f^2, 1/f^3) to avoid confusion.
- [Section 2.1] There are grammatical errors in the sentence 'The presence of RTN is the present SiNx RRAM MIS devices have been already investigated and demonstrated' (sic); please rephrase.
- [Fig. 5] The caption of Fig. 5(b) says 'Lower figure of (b) validates that all subtracted frequencies were below corner frequency f0,' but it is helpful to state explicitly how the subtracted signal was computed and what metric is shown.
Circularity Check
Experimental cross-validation is circular: Eq. (4)'s term B encodes the Eq. (5) relation later reported as confirmation.
-
fitted input called prediction
[Section 2.2 (Eq. 4) and Section 3 (Fig. 5(f))]
"In order to reduce the searching space area for the minimization algorithm evenmore, an extra term B in the criterion ’s C eq. (4) has been used. Now, the algorithm will try to minimize only near corner frequency f0. ... Our algorithm continues to perform well and the confirmation criterion -relation (5)- is satisfied for the estimated τc and τe parameters, as clearly shown in Fig. 5(f)."
Term B of Eq. (4) is |f0 - (1/2π)(1/τc+1/τe)|/N, and Eq. (5) is exactly f0 = (1/2π)(1/τc+1/τe). The f0 value is fixed by Lorentzian fitting of the same unstable signal, and the optimizer explicitly searches for τc and τe that minimize B. The agreement with Eq. (5) shown in Fig. 5(f) is therefore a constrained outcome of the objective function, not an independent measurement of accuracy. The simulated test with known ground truth provides partial independent evidence, but it uses the same single-trap exponential and Lorentzian model, so the experimental cross-validation claimed in the abstract remains circular.
full rationale
The moving-average detrending derivation itself is self-contained, and the synthetic experiment in Fig. 4 gives a genuine ground-truth check that the filter can separate a simulated 1/f^3 instability from a simulated two-level RTN signal. However, the experimental validation of the central claim is circular. The optimization criterion C in Eq. (4) contains term B, the absolute deviation between f0 and (1/2π)(1/τc+1/τe), with f0 obtained from a Lorentzian fit to the same unstable input PSD. The algorithm therefore searches for τc and τe that make Eq. (5) hold. When Section 3 reports that Eq. (5) is satisfied for the estimated time constants, it is reporting a property the optimizer was constructed to enforce. This is a fitted-input-called-prediction issue, not a self-citation issue: the authors' prior RTN publications are background, and the Lorentzian relation itself is standard and externally referenced. Because the synthetic experiment supplies partial independent support but the experimental cross-validation reduces to the objective function, the circularity score is 6.
Assumptions & free parameters
free parameters (5)
- a (criterion scaling parameter) =
near 1 (rule of thumb)
- b (criterion scaling parameter) =
near 1 (rule of thumb)
- N (quantization factor) =
220 for the measured signal
- f0 (RTN corner frequency) =
985 Hz for the measured signal
- M (moving-average mask size) =
optimized per signal, not reported explicitly
assumptions (5)
- domain assumption Single-trap two-level RTN model
- standard math Exponential dwell-time distribution
- domain assumption Lorentzian corner relation f0 = (1/2*pi)*(1/tau_c+1/tau_e)
- ad hoc to paper Timescale separation of instability
- domain assumption Histogram lobes are Gaussian
Cite this review
Pith. "Pith review of Analysis of random telegraph noise in resistive memories: The case of unstable filaments." pith.science (2026). https://pith.science/paper/XPHSILKP
@misc{pith2026250202117,
author = {Pith},
title = {Pith review of: Analysis of random telegraph noise in resistive memories: The case of unstable filaments},
year = {2026},
howpublished = {\url{https://pith.science/paper/XPHSILKP}},
note = {Machine review of arXiv:2502.02117}
}
read the original abstract
Through Random Telegraph Noise (RTN) analysis, valuable information can be provided about the role of defect traps in fine tuning and reading of the state of a nanoelectronic device. However, time domain analysis techniques exhibit their limitations in case where unstable RTN signals occur. These instabilities are a common issue in Multi-Level Cells (MLC) of resistive memories (ReRAM), when the tunning protocol fails to find a perfectly stable resistance state, which in turn brings fluctuations to the RTN signal especially in long time measurements and cause severe errors in the estimation of the distribution of time constants of the observed telegraphic events, i.e., capture/emission of carriers from traps. In this work, we analyze the case of the unstable filaments in silicon nitride-based ReRAM devices and propose an adaptive filter implementing a moving-average detrending method in order to flatten unstable RTN signals and increase sufficiently the accuracy of the conducted measurements. The te and tc emission/capture time constants of the traps, respectively, are then calculated and a cross-validation through frequency domain analysis (Lorentzian fitting) was performed proving that the proposed method is accurate.
Figures
Figures from the paper (2 more)
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.