---
title: 'Standard-Deviation Vector: A Multifaceted Approach'
url: https://www.emergentmind.com/topics/standard-deviation-vector
type: topic
---

# Standard-Deviation Vector: A Multifaceted Approach

A standard-deviation vector is not a single universally standardized object. In the cited literature, the term and its underlying idea denote several related constructions: a vector whose components are standard deviations and whose ratios encode allometric structure; a local sample vector whose standard deviation is used as a decision statistic; a vector of componentwise quantum uncertainties for a superposition; a per-feature activation-dispersion vector used to score weights for pruning; and matrix-level extensions of standard deviation in noncommutative probability. A related development replaces a componentwise spread vector by a single distance-based scalar for multivariate data [2004.01385] [1304.6379] [1607.08000] [2503.22451] [1208.4072] [1705.05777].

## 1. Terminological scope and recurring structures

Across these sources, the phrase refers to distinct but structurally analogous objects. The common theme is that dispersion is first attached to components, neighborhoods, or observables, and is then used to induce slopes, thresholds, bounds, pruning scores, or seminorms [2004.01385] [1304.6379] [1607.08000] [2503.22451] [1208.4072] [1705.05777].

| Context | Mathematical object | Role |
|---|---|---|
| Urban allometry | \(S=(s_1,\dots,s_n)\) or \(\Sigma=(\sigma_1,\dots,\sigma_n)\) | Ratios \(s_i/s_j\) or \(\sigma_i/\sigma_j\) generate scaling exponents |
| Edge detection | \(\mathbf{x}=[x_{11},x_{12},x_{21},x_{22}]^T\) | \(\sigma(\mathbf{x})\) is thresholded to classify edges |
| Quantum superposition | \(\vec{\Delta}(A)=(\Delta_{\psi_1}A,\dots,\Delta_{\psi_N}A)\) | Bounds the SD of an observable on the superposed state |
| LLM pruning | \(\sigma=(\sigma_1,\dots,\sigma_M)\) | Scales per-weight pruning scores |
| Noncommutative probability | \(\{\sigma_\mu^{(n)}\}_{n\in\mathbb{N}}\) | Gives matrix-level strongly Leibniz seminorms |
| Multivariate spread | \(\mathcal{V}(X)\) | Acts as a scalar alternative to a componentwise SD vector |

A common misconception is that the expression names a single canonical vector-space object. In these sources, it does not. Sometimes it is literally a vector of standard deviations; sometimes it is a vector of raw values on which a standard deviation is computed; sometimes it is a family of seminorms indexed by matrix size; and sometimes it is contrasted with a scalar alternative that summarizes multivariate spread without componentwise decomposition.

## 2. Allometric scaling, standard-deviation ratios, and scaling matrices

In urban allometry, the standard-deviation vector is introduced for log-transformed measures \(Q_i(t)\) that obey a power-law relation. If the allometric law holds exactly, the scaling exponent equals the ratio of the standard deviations of the corresponding logarithmic measures; empirically, the exponent equals the product of that standard-deviation ratio and the Pearson correlation coefficient. In the notation of the paper, the theoretical relation is \(a_{ij}=\sigma_i/\sigma_j\), whereas the empirical relation is \(a_{ij}=r_{ij}\, s_i/s_j\), with \(a_{ij}a_{ji}=r_{ij}^2\) [2004.01385].

For \(n\) variables, the construction becomes explicitly vectorial. The paper defines the population SD vector \(\Sigma\) and its reciprocal vector \(\Sigma'\), and forms the matrix
\[
M_{ij}=\frac{\sigma_i}{\sigma_j}.
\]
This is the matrix of theoretical allometric exponents under perfect correlation. The paper then shows
\[
M\Sigma=n\Sigma,
\]
so the standard-deviation vector is an eigenvector of the scaling matrix with eigenvalue \(n\). In empirical work, population SDs \(\sigma_i\) are replaced by sample SDs \(s_i\), yielding the analogous relation \(MS=nS\). This makes the standard-deviation vector a characteristic-length vector in log space, while the exponent matrix is generated by component ratios [2004.01385].

The same paper extends the dispersion-ratio logic to city fractals and city size distribution. It states that scaling exponents can be interpreted as ratios of characteristic lengths across time, space, and probability dimensions, and records the relation
\[
a_{ij}=\frac{r_i}{r_j}=\frac{D_i}{D_j}=\frac{\sigma_i}{\sigma_j}.
\]
A plausible implication is that the “vector” viewpoint is not restricted to urban GDP or population series; it is presented as a unifying formalism for relative growth rates, fractal dimensions, and Zipf-type exponents [2004.01385].

The empirical illustration uses GDP time series from 1998–2018 for Beijing, Tianjin, Shanghai, and Chongqing. The paper reports that exponents computed by standard-deviation ratio plus correlation are completely identical to those computed by conventional OLS regression in log space, with goodness-of-fit values \(r^2\) around \(0.99\) to \(0.9988\). From the resulting exponent matrix, it derives the geometric-mean vector
\[
W=[0.9600,\,1.1181,\,0.8445,\,1.0898]^T
\]
and the normalized eigenvector
\[
w=[0.2393,\,0.2787,\,0.2105,\,0.2716]^T,
\]
interpreted as allometric scaling indexes for the four cities [2004.01385].

## 3. Local intensity vectors and statistical edge detection

In image processing, the relevant object is a local intensity vector rather than a vector of precomputed standard deviations. A \(2\times 2\) neighborhood is written as
\[
\mathbf{x}=[x_{11},x_{12},x_{21},x_{22}]^T,
\]
with mean
\[
\mu=\frac{1}{4}(x_{11}+x_{12}+x_{21}+x_{22})
\]
and simple statistical standard deviation
\[
\sigma=\sqrt{\frac{1}{4}\sum_{k=1}^4(x_k-\mu)^2}.
\]
If \(\sigma\) exceeds a threshold \(T\), the upper-left pixel of the window is marked as an edge pixel. The paper reports an empirical threshold range of \(4\) to \(9\), with \(7\) used in the experiments [1304.6379].

This construction treats each neighborhood as a short vector of samples and compresses it to a scalar dispersion statistic. In uniform regions, local variation is low and the standard deviation is low; at edges, intensities differ sharply across the \(2\times 2\) block and the standard deviation is high. The detector therefore uses the standard deviation of a local vector as an “edge-likeness” statistic rather than using derivative masks or a gradient vector [1304.6379].

The method is explicitly median-prefiltered. First, a median filter is applied to suppress Salt-and-Pepper noise; then the \(2\times 2\) standard deviation is computed on the denoised image; finally, thresholding determines the edge map. The paper states that applying the detector without median filtering yields many false edges because noisy pixels create high local dispersion. With median filtering, the standard deviation is a cleaner indicator of true boundaries. The method is described as computationally light, with \(O(MN)\) complexity for an \(M\times N\) image, and it is qualitatively compared with Sobel and Canny. The paper reports that the visual results are close to Canny but the edges are thicker; for text images, the boundaries are often clearer than those produced by Sobel and Canny [1304.6379].

## 4. Quantum superposition, coherence, and incompatibility vectors

In the quantum setting, the standard-deviation vector is explicitly defined for an observable \(A\) and a superposition
\[
|\psi\rangle=\sum_{i=1}^N \alpha_i |\psi_i\rangle.
\]
For each component state \(|\psi_i\rangle\), the standard deviation is \(\Delta_{\psi_i}A\), and the paper packages these into
\[
\vec{\Delta}(A)=\bigl(\Delta_{\psi_1}A,\Delta_{\psi_2}A,\dots,\Delta_{\psi_N}A\bigr).
\]
The central result is that the standard deviation of \(A\) on the normalized superposition is bounded above and below by expressions built from the weighted sum \(\sum_i |\alpha_i|^2\Delta_{\psi_i}A\) together with interference terms \(E_\pm(A)\) and \(F(A)\). The bounds show that the superposed-state uncertainty is not merely a convex average of component uncertainties; it is shifted by interference contributions [1607.08000].

The paper interprets \(\vec{\Delta}(A)\) geometrically as a point in \(\mathbb{R}^N\). The scalar uncertainty of the superposed state is then confined to an interval determined by a weighted linear functional of that vector plus pairwise geometric-mean-type corrections involving both \(\Delta_{\psi_i}A\) and \(\langle A\rangle_i\). This suggests a projection-like viewpoint: the SD of the full superposition is a constrained scalar function of an underlying SD vector and the superposition coefficients [1607.08000].

For pure states, the paper uses the skew-information-based K-coherence and states
\[
\Delta_\psi A=\sqrt{I(\varrho,A)}
\]
when \(\varrho=|\psi\rangle\langle\psi|\). Consequently, the same bounds also serve as bounds on coherence. It further defines a state-dependent incompatibility of observables \(A\) and \(B\) by
\[
U_\psi(A,B)=\Delta_{\tilde{\psi}}^2 A+\Delta_{\tilde{\psi}}^2 B,
\]
and derives analogous superposition bounds using the componentwise incompatibilities \(U_{\psi_i}(A,B)\). This yields an “incompatibility vector” \(\vec{U}(A,B)\) alongside the SD vector \(\vec{\Delta}(A)\) [1607.08000].

The numerical illustration uses a randomly generated \(4\times 4\) Hermitian observable and two normalized states, combined into the family \( |\psi_\pm\rangle = x|\psi_1\rangle \pm \sqrt{1-x^2}|\psi_2\rangle \). The reported behavior of the bounds depends on interference: for \(|\psi_+\rangle\), the upper bound is quite tight and the lower bound is looser; for \(|\psi_-\rangle\), the lower bound becomes much tighter while the upper bound is looser. The same analytic form therefore changes effective tightness with the interference structure of the superposition [1607.08000].

## 5. Activation-dispersion vectors in pruning large language models

In STADE, the standard-deviation vector is a per-feature dispersion summary of layer inputs. A linear layer is modeled with input random vector \(X\in\mathbb{R}^M\), weight matrix \(\mathbb{W}\), and bias vector \(\mathbb{B}\), and pruning is formulated as setting one weight to zero while minimizing the expected squared change in the layer output. The analysis shows that, when bias adjustment is allowed, the optimal criterion for choosing which coordinate to prune depends on \(\sigma_j^2 W_j^2\), where \(\sigma_j^2=\mathbb{V}[X_j]\). This naturally yields the standard-deviation vector
\[
\sigma=(\sigma_1,\dots,\sigma_M)
\]
for the layer [2503.22451].

Under the specific assumptions \(B=0\) and \(\mu_j=\mathbb{E}[X_j]=0\), the paper proves that Wanda is optimal. In that case,
\[
S_{i,j}^{\text{Wanda}}=\|X_{:,j}\|_2\,|\mathbb{W}_{j,i}|
\]
is equivalent to the optimal criterion because \(\sigma_j\) is approximated by the RMS norm of centered inputs. For biased or uncentered inputs, however, the correct scale factor is standard deviation rather than raw RMS magnitude. STADE therefore uses the mean-subtracted score
\[
S_{i,j}^{\text{STD}}=
\left\|X_{:,j}-\frac{1}{N}\sum_{n=1}^N X_{n,j}\right\|_2\,|W_{j,i}|,
\]
which is proportional to \(\sigma_j |W_{j,i}|\). The “STADE w/o bias” variant uses a score corresponding to \((\sigma_j^2+\mu_j^2)W_j^2\) when the pruned layer must remain bias-free [2503.22451].

Operationally, the method computes the mean and standard deviation of each input feature from activations collected on a calibration dataset, here C4, and uses the same activation statistics for all output columns of the layer. The resulting vector is thus per feature or channel, not per token. In centered layers such as the first MLP layer and Q/K/V projections, STADE reduces to Wanda; in biased layers such as self-attention output projection or MLP down projection, the mean subtraction changes the ranking of weights [2503.22451].

The empirical results cover Llama and OPT families, unstructured pruning at \(50\%\) sparsity, and structured \(2\!:\!4\) and \(4\!:\!8\) pruning. The paper reports that Wanda and STD perform similarly in centered layers, as predicted by the theory, whereas in uncentered layers the STD metric yields lower perplexity than Wanda, especially at high sparsity. Across whole-model pruning, STADE matches or slightly improves over Wanda in perplexity at \(50\%\) sparsity, and zero-shot accuracy generally matches or slightly outperforms Wanda, especially in \(N\!:\!M\) pruning regimes [2503.22451].

## 6. Seminorm, matricial, and distance-based generalizations

A more abstract interpretation treats standard deviation as a seminorm rather than as a statistic attached to a finite vector. In the commutative setting \(A=L^\infty(X,\mu)\), the paper defines
\[
\sigma(f)=\|f-\mu(f)\|_{L^2(\mu)}
       =\big(\mu(|f|^2)-|\mu(f)|^2\big)^{1/2},
\]
and proves the Leibniz inequality
\[
\sigma(fg)\le \sigma(f)\|g\|_\infty+\|f\|_\infty\sigma(g).
\]
It then shows that standard deviation is strongly Leibniz, and extends the result to noncommutative probability spaces, where
\[
\sigma_\mu(A)=\max\Big\{\|A-\mu(A)1_A\|_{2,\mu},\ \|A^*-\mu(A^*)1_A\|_{2,\mu}\Big\}.
\]
At the matrix level, the paper defines
\[
\sigma_\mu^{(n)}(X)=
\max\Big\{\|X-E_\mu^{(n)}(X)\|_{E_\mu^{(n)}},\ \|X^*-E_\mu^{(n)}(X^*)\|_{E_\mu^{(n)}}\Big\},
\]
and proves that the family \(\{\sigma_\mu^{(n)}\}_{n\in\mathbb{N}}\) is a strongly Leibniz matricial *-seminorm on \(A\). Although the paper does not use the phrase “standard-deviation vector” explicitly, it provides a rigorous matrix-level generalization of joint deviation [1208.4072].

A different generalization replaces a componentwise SD vector by a single scalar spread functional on multivariate data. The distance standard deviation is defined by
\[
\mathcal{V}(X)=\sqrt{\mathcal{V}^2(X)},
\]
where \(\mathcal{V}^2(X)\) is built from expectations of pairwise Euclidean distances. For \(X=(X_1,\dots,X_p)\), the paper contrasts \(\mathcal{V}(X)\) with classical practice based on the componentwise standard deviations \((\sigma(X_1),\dots,\sigma(X_p))\) or the covariance matrix \(\Sigma_X\). It proves that \(\mathcal{V}(X)\) is bounded above by the classical standard deviation and by Gini’s mean difference in the scalar case, and shows that \(\mathcal{V}\) satisfies the axiomatic properties of a measure of spread [1705.05777].

The asymptotic and applied results sharpen that contrast. The paper derives the asymptotic normality of the empirical distance standard deviation under finite second moments and reports asymptotic relative efficiencies for heavy-tailed models. For example, at \(t_3\),
\[
\mathrm{ARE}(\mathcal{V}_n;t_3)=0.965,\qquad
\mathrm{ARE}(\widehat{\sigma}_n;t_3)=0,
\]
whereas at \(N(0,1)\),
\[
\mathrm{ARE}(\mathcal{V}_n;N(0,1))=0.784,\qquad
\mathrm{ARE}(\widehat{\sigma}_n;N(0,1))=1.
\]
In multivariate statistical quality control, the proposed \(\mathcal{V}\)-chart can outperform the generalized-variance \(|S|\)-chart under heavy tails. This suggests a complementary perspective: instead of organizing spread into a vector of componentwise standard deviations, one may summarize the joint distribution by a geometry-aware scalar built from inter-point distances [1705.05777].

Taken together, these lines of work show that “standard-deviation vector” is best understood as a family resemblance rather than a unique definition. In one setting it is a vector of dispersions whose component ratios generate scaling exponents; in another it is a local data vector whose dispersion is thresholded; in another it is a componentwise uncertainty vector subject to interference corrections; in another it is a per-feature activation-dispersion vector used for pruning; and in another it becomes a matrix-level seminorm or is displaced by a scalar alternative to componentwise spread. The invariant idea is the same: standard deviation is treated not only as a scalar summary, but as a structured object that can organize relations among multiple components.

Source: https://www.emergentmind.com/topics/standard-deviation-vector