Limiting Stimulus Estimation

• Limiting Stimulus Estimation is a procedure using sequential binary-outcome tests to identify the minimal or threshold stimulus needed to elicit a response.
• It is highly dependent on experimental parameters (e.g., stimulus set, step size, and stopping rules like UN 1-in-6), which can lead to instability in the estimated sensitivity.
• Robust alternatives such as up-and-down designs, BCD+CIR, and RMJ stochastic approximation offer improved quantification with valid uncertainty intervals and enhanced inferential properties.

Limiting stimulus estimation encompasses a family of procedures in which the “minimal” or “threshold” stimulus required to elicit a response is determined using binary-outcome experiments. These procedures have been central to safety testing protocols in energetic materials, to quantification of sensory thresholds in neurophysiology, and to the theoretical study of discrimination limits in neural coding. However, both empirical practice and formal analysis demonstrate substantial pitfalls in the use of limiting stimuli as summary metrics, as well as the evolution toward more robust and inferentially well-founded alternatives.

1. Formal Definition of Limiting Stimuli

Limiting stimulus (LS) estimation arises in binary-outcome testing settings where an experimenter sequentially applies stimulus levels $\{s_1 > s_2 > \cdots > s_m\}$ and records outcomes $y_i \in \{0,1\}$ for trial $i$ at stimulus $x_i \in S$. The dominant protocol, particularly in energetic materials, is the United Nations (UN) 1-in-6 test. In this test, the experiment starts at the highest level $x_1 = s_1$ and proceeds as follows:

• If $y_i = 1$ (e.g., explosion), the next stimulus $x_{i+1}$ is the next lower level in $S$.
• If $y_i = 0$, decrement $x_{i+1}$ only after $K$ consecutive zeros are observed ($K = 6$ in the “1-in-6”).

Define $\tau = \inf\{ n: y_{n-5} = \cdots = y_n = 0 \}$. The test terminates at trial $\tau$. Two conventions exist for the LS:

• Type I: LS = penultimate stimulus (last level at which a “1” was observed, $x_{\tau-1}$).
• Type II: LS = terminal stimulus ($x_\tau$).

The procedure is sometimes summarized as "stop when 1 of 6 fails" within a block of 6 trials. Crucially, the entire approach is predicated on a deterministic design-dependent stopping rule, not on a fixed target quantile or true threshold of a response function $F(x) = P(y = 1 | x)$ (Christensen et al., 18 Jan 2026).

2. Theoretical Foundations and Critique

Christensen & Novik provide four principal arguments demonstrating that limiting stimulus, as defined in these protocols, does not constitute a well-defined sensitivity parameter:

• Dependence on Experimental Parameters: LS depends on choices such as the set $S$, step size between stimulus levels, the value of $K$, the “bracketing” strategy, and whether Type I or Type II convention is used. These setup parameters, not the intrinsic properties of the sample, directly impact the LS distribution.
• Lack of Functional Relationship $F \to \text{LS}$: Unlike quantiles $\xi_{100p} = F^{-1}(p)$, the mapping from response probability curve $F$ to LS is convoluted; LS is not a deterministic or invariant property of $F$. Any changes in the testing design break this relationship.
• No Inferential Uncertainty: Because no fixed true parameter underlies LS, classical interval estimation or confidence interval construction is invalid. Any uncertainty interval would implicitly depend on design parameters, not material sensitivity alone.
• Empirical Counterexamples: Changes in stimulus-level definitions or terminus conventions result in large shifts in reported LS, even for identical response curves $F$. Simulations under a probit model $P(y=1 \mid x) = \Phi(\alpha + \beta \log x)$ reveal this instability. In PETN friction tests, the default UN protocol with discrete loads yields LS = 80 N (“insensitive”), but including intermediate loads produces LS = 48 N (“sensitive”), with no principled method for error quantification (Christensen et al., 18 Jan 2026).

3. Alternative Frameworks for Sensitivity and Threshold Estimation

Contemporary methodology targets estimation of well-defined functionals of the sensitivity curve $F$. Three classes of alternatives enable direct quantile estimation, likelihood-based inference, and well-defined confidence intervals.

3.1 Up-and-Down Design with Parametric Maximum Likelihood

Based on Dixon’s rule, after each trial at $x_i$:

• $x_{i+1} = x_i - d$ if $y_i = 1$
• $x_{i+1} = x_i + d$ if $y_i = 0$

Assuming a probit link: $P(y=1 \mid x) = \Phi(\alpha + \beta \log x)$.

• Maximum likelihood estimates $(\hat{\alpha}, \hat{\beta})$ are obtained from the log-likelihood

$$\ell(\alpha, \beta) = \sum_{i} [ y_i \log \Phi(\alpha + \beta \log x_i) + (1-y_i) \log(1 - \Phi(\alpha + \beta \log x_i)) ].$$

• The quantile $\xi_{50}$ (median sensitivity) is estimated as $\log \xi_{50} = -\hat{\alpha}/\hat{\beta}$.
• Confidence intervals employ Fieller’s theorem, using the observed Fisher information matrix.

3.2 Biased-Coin Design (BCD) with Centered Isotonic Regression (CIR)

BCD (Durham & Flournoy) is an adaptive procedure to target a specific quantile $\xi_{100p}$ where $p < 1/2$. After each trial:

• If $y_i = 1$, move up one level as in up-and-down.
• If $y_i = 0$, move down with probability $p/(1-p)$, or remain at the same level.

Data $(x_i, y_i)$ are aggregated and analyzed via centered isotonic regression, under the constraint that $F$ is nondecreasing:

• $F$ is estimated by maximizing the binomial log-likelihood over monotone functions.
• The quantile is $\xi_{100p} = \inf \{ x : \hat{F}(x) \geq p \}$.
• Nonparametric confidence intervals are available using the delta method on the inversion of the piecewise-constant $\hat{F}$ (Christensen et al., 18 Jan 2026).

3.3 Robbins–Monro–Joseph (RMJ) Stochastic Approximation

RMJ operates via stochastic approximation on the log scale:

$\log x_{i+1} = \log x_i - a_i (y_i - b_i),$

with $a_i$, $b_i$ tuned for near-optimal convergence properties. $x_{n+1}$ is the estimator for $\xi_{100p}$, and under smooth $F$, asymptotic normality and variance formulas permit confidence interval construction (Christensen et al., 18 Jan 2026).

4. Simulation Studies and Methodological Comparison

The performance characteristics of these alternatives were benchmarked across six analytic sensitivity curves ($F$: normal, uniform, logistic, Gumbel, skewed-logistic, Cauchy), quantile targets $p \in \{0.10,0.25,0.50,0.75,0.90\}$, and sample sizes $n \in \{30, 100\}$ through $10^4$ replications:

• After $n=30$ trials, up-and-down+MLE exhibits the highest mean squared error (MSE), often $2$–$3\times$ that of BCD or RMJ. Both BCD and RMJ achieve substantially lower MSEs; RMJ yields slightly narrower average 90% interval widths.
• All methods under-cover for $n=30$, but for $n=100$: coverage for up-and-down+Fieller is strong when $F$ is normal but degrades quickly for heavy-tailed or skewed models. BCD+CIR consistently achieves empirical coverage between $87$–$92$%.
• In PETN friction tests, BCD+CIR targeting $\xi_{10}$ provides explicit confidence intervals, with estimates overlapping across runs of different step sizes and total trials (e.g., $\hat{\xi}_{10}=58.95$ N with 90% CI $[23.47, 61.87]$ by $n=30$, and $\hat{\xi}_{10}=38.17$ N with 90% CI $[18.83, 61.79]$ by $n=50$), as opposed to design-dependent, unquantified LS values (Christensen et al., 18 Jan 2026).

Method	Target	Estimation	CIs Available	Robustness (p, F)
UN 1-in-6	—	Stopping rule LS	No	Design-dependent
Up-and-Down	$\xi_{100p}$	Parametric MLE	Yes (Fieller)	Sensitive to model
BCD+CIR	$\xi_{100p}$	Nonparametric (CIR)	Yes	High
RMJ	$\xi_{100p}$	Stochastic approx.	Yes	Needs smoothness

5. Limiting Stimulus Estimation in Neurophysiology and Sensory Systems

In neurophysiological and clinical contexts, the conceptually analogous problem is determining the minimum stimulus intensity (the sensory threshold) that evokes a detectable neural response (Schilling et al., 2018). Classical threshold setting schemes, such as those positing a criterion relative to background noise (e.g., a response $2\sigma_{noise}$ above baseline), are equally ill-defined: estimates are highly sensitive to sample size, background variability, and subjective disease criteria.

A rigorous alternative models the full stimulus–response curve using a noise-incorporated hard sigmoid:

$$f_0(I) = \begin{cases} 0, & I \leq t \ s (I - t), & t < I < t + h/s \ h, & I \geq t + h/s \end{cases}$$

with fitted parameters: threshold $t$, slope $s$, saturation $h$; and observed data modeled as $f(I) = \sqrt{f_0(I)^2 + \sigma^2}$ for RMS-type signals. Random subsampling and repeated model fitting provide a distribution of threshold ($t$) estimates, from which robust median and percentile-based uncertainty intervals are derived (Schilling et al., 2018). This method eliminates systematic dependence on noise, sample size, and arbitrary detection criteria.

6. Non-Euclidean Distances and Fundamental Sensitivity Limits in Neural Populations

The theoretical limit for stimulus discrimination is not determined by classic thresholding, but by the distinguishability of stimulus-induced response distributions. “Retinal distance,” defined as the symmetrized Kullback–Leibler (KL) divergence between the distributions $P(\sigma|s_1)$ and $P(\sigma|s_2)$ of neural population responses, quantifies this limit:

$$D_{ret}(s_1, s_2) = \frac{1}{2} \left[\sum_\sigma P(\sigma|s_1) \log_2\frac{P(\sigma|s_1)}{P(\sigma|s_2)} + (s_1 \leftrightarrow s_2)\right]$$

This metric is generally non-Euclidean—projection onto a static quadratic form fails to capture the actual discriminability structure—which reflects the biological sensitivity manifold of the system (Tkačik et al., 2012). In practice, minimal discriminable stimulus change corresponds to $D_{ret} \approx 1$ bit. The Fisher information matrix $J(s)$ locally relates to $D_{ret}$ through

$$D_{ret}(s, s+\delta s) \approx \frac{1}{2} \delta s^\top J(s) \delta s$$

leading directly to fundamental Cramér–Rao bounds on estimation precision (Tkačik et al., 2012).

7. Implications and Recommended Practice

Limiting stimulus estimation via stopping rule conventions (e.g., UN 1-in-6) is fundamentally ill-posed as an intrinsic sensitivity metric. Any valid approach must estimate a well-defined parameter of the sensitivity (or threshold) function $F(x)$ and provide rigorous uncertainty quantification. For energetic materials, the choice of BCD+CIR is robust across response curves, nonparametric, and provides valid confidence intervals; when strong parametric assumptions are justified, up-and-down with Fieller interval is optimal for small sample regimes (Christensen et al., 18 Jan 2026). In sensory physiology, hard sigmoid regression with principled subsampling achieves robust, unbiased threshold estimation with meaningful uncertainty intervals even under high noise or limited sampling (Schilling et al., 2018). For neural population coding, distance-to-discrimination measures built upon KL-divergence between response distributions encode the fundamental limits that cannot be resolved by thresholding alone (Tkačik et al., 2012).