3  Variance Stabilization and Normalization

Chapter Aim

Chapter 02 treated distributions as generative models. Chapter 03 asks what these models imply for preprocessing. Raw high-throughput count data are not only noisy; their noise is structured. Library sizes differ across observations, rows of sequencing data often contain relative rather than absolute information, and the variance of a feature often changes with its mean. Normalization and transformation are therefore part of the statistical model, because they put observations and features on a scale where distances, PCA, clustering, heatmaps, and regression behave more meaningfully. The purpose of this chapter is to make the link between measured counts and transformed analysis scales explicit.

TipNotation convention

Notation follows the shared convention in the notation chapter. Vectors are columns by default, row vectors are written with transposes, and dimensions should be checked whenever matrices are multiplied, transformed, modeled, or decomposed.

NoteMain external reference

The main book reference is MSMB VST Delta Method, where variance-stabilizing transformations are derived from the delta method. For RNA-seq size factors, count models, and DESeq2 transformations, see MSMB High-Throughput Count Data, MSMB Count Data Normalization, and MSMB RNA-seq VST. For the broader experimental-design view of mean-variance relationships, see MSMB Mean-Variance Relationships.

The empirical consequences of these choices are worked out in Exercise Sheet 02, where raw count matrices are compared with median-of-ratios normalization, variance-stabilizing transformations, PCA, and single-cell normalization strategies (Exercise Sheet 02 2026). The compositional continuation appears in Exercise Sheet 03 through 16S normalization, log-ratio geometry, and ordination (Exercise Sheet 03 2026).

Learning Goals

After this chapter, students should be able to:

  1. Explain why raw sequencing counts cannot usually be compared directly across observations.
  2. Distinguish normalization from variance-stabilizing transformation.
  3. Interpret total-count scaling, common-scale normalization, rarefying, median-of-ratios normalization, and log-ratio transformations as different answers to different modeling assumptions.
  4. Derive the variance-stabilizing transformation associated with a mean-variance relationship using the delta method.
  5. Connect the Poisson, negative-binomial, multinomial, Dirichlet-multinomial, and logistic-normal models from Chapter 02 to concrete preprocessing choices.
  6. Explain why transformed data may be useful for exploratory analysis even when formal testing is done with a count model.
  7. Recognize that simple transformations can perform well in practice, but that their behavior should still be understood through the model they approximate.

3.1 What Normalization Is Trying To Estimate

For count data, a first model is that the observed count \(C_{ij}\) is generated from an expected count \(\mu_{ij}\). In sequencing data this expectation often factors into an observation-specific exposure and an abundance-like quantity:

\[\begin{equation} \mu_{ij}=s_i q_{ij}. \label{eq:ch03-mean-factorization} \end{equation}\]

Here \(s_i>0\) is a size factor or exposure for observation \(i\), and \(q_{ij}\geq 0\) is the quantity we would like to compare across observations after accounting for exposure. The size factor can represent sequencing depth, sampling efficiency, total captured molecules, or another multiplicative technical effect. The abundance \(q_{ij}\) is not necessarily an absolute molecular count in the biological system; in many sequencing assays it is still a relative or assay-specific quantity.

Normalization is the act of estimating or specifying the nuisance scale \(s_i\) and removing it from the count scale. A simple normalized count is

\[\begin{equation} \tilde{x}_{ij}=\frac{c_{ij}}{\hat{s}_i}. \label{eq:ch03-size-factor-normalized-count} \end{equation}\]

This equation encodes a model. It assumes that the dominant unwanted difference between observations is multiplicative and row-specific. It also assumes that the same \(\hat{s}_i\) is appropriate for all features in row \(i\). In RNA-seq count models, size factors are used to model sampling-depth bias before dispersion estimation and testing (see MSMB Count Data Normalization).

Normalization and transformation should be kept conceptually separate. Normalization estimates and removes nuisance scale. Transformation changes the measurement scale, often to reduce heteroscedasticity or make distances more interpretable. In practice they are often chained:

\[\begin{equation} x_{ij}=g\left(\frac{c_{ij}}{\hat{s}_i}\right), \label{eq:ch03-normalize-then-transform} \end{equation}\]

where \(g\) might be a logarithm, a square-root transform, a variance-stabilizing transform, a residual transform, or a log-ratio transform.

3.2 Library Size And Total-Count Scaling

The simplest size factor is proportional to the row total

\[\begin{equation} N_i=\sum_{j=1}^p C_{ij}. \label{eq:ch03-library-size} \end{equation}\]

If \(N_i\) is interpreted as sequencing depth, then total-count scaling uses

\[\begin{equation} \hat{s}_i=\frac{N_i}{L}, \label{eq:ch03-total-count-size-factor} \end{equation}\]

where \(L>0\) is a reference depth, such as the mean library size, the median library size, \(10^4\), or \(10^6\). The resulting scaled count is

\[\begin{equation} \tilde{x}_{ij}=L\frac{c_{ij}}{N_i}. \label{eq:ch03-total-count-scaled-count} \end{equation}\]

If \(L=10^6\), this is counts per million. For single-cell workflows one often sees \(L=10^4\). The constant \(L\) does not change relative differences within an observation, but it changes the numerical scale on which pseudocounts and logarithms operate. This is why the scaling constant is not completely harmless when the next step is \(\log(\tilde{x}_{ij}+\kappa)\).

Total-count scaling is appropriate when library size is the main nuisance effect and the row total does not itself carry biological signal. It can be misleading under strong compositional effects. If a small number of features dominate one observation, then all other scaled features are pushed downward even if their absolute molecule counts did not change. This is the same compositional constraint that appeared in Chapter 02 in the multinomial model.

3.3 Common-Scale Counts And Rarefying

In microbiome and amplicon sequencing, one tempting strategy is to force all observations to have the same library size. Let

\[\begin{equation} N_{\min}=\min_{i=1,\ldots,n}N_i. \label{eq:ch03-minimum-library-size} \end{equation}\]

A deterministic common-scale version of the counts would be

\[\begin{equation} \tilde{x}_{ij}=N_{\min}\frac{c_{ij}}{N_i}. \label{eq:ch03-common-scale-normalization} \end{equation}\]

Rarefying instead creates a random subsample of size \(N_{\min}\) from each row. Conditional on the observed count vector \(c_i=(c_{i1},\ldots,c_{ip})^\top\) and its total \(N_i\), this can be written as

\[\begin{equation} C_i^{\mathrm{rare}}\mid c_i \sim \operatorname{Multinomial} \left( N_{\min}, \frac{c_i}{N_i} \right). \label{eq:ch03-rarefying-model} \end{equation}\]

This equation makes the problem visible. Rarefying adds another multinomial sampling step and discards observed counts from deeper samples. It may make library sizes equal, but it does so by injecting avoidable randomness. McMurdie and Holmes’ “Waste Not, Want Not” argument is the main warning against treating rarefying as a default normalization method (McMurdie and Holmes 2014). Weiss and colleagues make the complementary point that normalization and differential-abundance strategies depend on the data characteristics, especially sparsity, library-size variation, and compositionality (Weiss et al. 2017).

3.4 Median-Of-Ratios Normalization

Bulk RNA-seq workflows often use a more robust size-factor estimate than the library total. DESeq2’s median-of-ratios normalization constructs a pseudo-reference feature profile by taking the geometric mean of each feature across observations. For features that are positive in all observations, the geometric mean is

\[\begin{equation} g_j= \exp\left( \frac{1}{n}\sum_{i=1}^n \log c_{ij} \right) = \left( \prod_{i=1}^n c_{ij} \right)^{1/n}. \label{eq:ch03-deseq-geometric-mean} \end{equation}\]

The logarithms in Equation \(\ref{eq:ch03-deseq-geometric-mean}\) are not decoration. They make clear why features with zeros need special handling: if any \(c_{ij}=0\), then \(\log c_{ij}\) is undefined and the product form gives \(g_j=0\). In the basic median-of-ratios construction, only features with \(g_j>0\) contribute to the ratio calculation. For those features, each observation is compared with the pseudo-reference:

\[\begin{equation} r_{ij}=\frac{c_{ij}}{g_j}. \label{eq:ch03-deseq-ratio} \end{equation}\]

The size factor for observation \(i\) is then the median of these ratios:

\[\begin{equation} \hat{s}_i=\operatorname{median}_{j:g_j>0} r_{ij}. \label{eq:ch03-deseq-size-factor} \end{equation}\]

Equivalently, this can be written on the log scale as

\[\begin{equation} \log \hat{s}_i = \operatorname{median}_{j:g_j>0} \left( \log c_{ij}-\frac{1}{n}\sum_{\ell=1}^n \log c_{\ell j} \right). \label{eq:ch03-deseq-log-size-factor} \end{equation}\]

The key assumption is that most features are not systematically changing, or that changes are roughly balanced between up- and down-regulation. Under that assumption, the median ratio estimates the sample-specific multiplicative effect without being dominated by a small number of highly expressed or differentially expressed features. Exercise Sheet 02 works through this idea in the pasilla data, comparing raw counts, median-of-ratios normalization, heatmaps, VST-transformed data, and PCA (Exercise Sheet 02 2026).

In the notation of Chapter 02, a feature-wise negative-binomial model can be written as

\[\begin{equation} C_{ij}\sim \operatorname{NB}(\mu_{ij},\phi_j), \qquad \mu_{ij}=s_i q_{ij}, \qquad \operatorname{Var}(C_{ij})=\mu_{ij}+\phi_j\mu_{ij}^2. \label{eq:ch03-nb-with-size-factor} \end{equation}\]

This model makes clear why size-factor estimation and dispersion estimation belong together. The size factor affects the mean, and the mean affects the variance. Normalization is therefore not independent of modeling; it is one part of the same mean-variance story.

3.5 Mean-Variance Relationships

A central reason for transforming high-throughput data is heteroscedasticity: features with larger means often have larger variances. Suppose that, after accounting for exposure, a random count \(Y\) has mean \(\mu\) and variance

\[\begin{equation} \mathbb{E}(Y)=\mu, \qquad \operatorname{Var}(Y)=v(\mu). \label{eq:ch03-mean-variance-function} \end{equation}\]

Different models imply different variance functions. For a Poisson model,

\[\begin{equation} v(\mu)=\mu. \label{eq:ch03-poisson-variance-function} \end{equation}\]

For the mean-dispersion negative-binomial model used in Chapter 02,

\[\begin{equation} v(\mu)=\mu+\phi\mu^2. \label{eq:ch03-nb-variance-function} \end{equation}\]

For a measurement with approximately constant coefficient of variation, one has

\[\begin{equation} v(\mu)=a\mu^2, \qquad a>0. \label{eq:ch03-quadratic-variance-function} \end{equation}\]

These equations are not just descriptive. They tell us what a useful transformation should look like. If the variance grows linearly with the mean, a square-root-like transformation is natural. If the variance grows quadratically with the mean, a log-like transformation is natural. If both terms are present, as in the negative-binomial model, a good transformation behaves like a square root for small counts and like a logarithm for large counts. This point is developed through the delta method in the main reference (see MSMB VST Delta Method and MSMB Mean-Variance Relationships).

3.6 The Delta Method

The delta method is the mathematical bridge between a model’s variance function and a variance-stabilizing transformation. Let \(Y\) be a random variable with mean \(\mu\) and variance \(v(\mu)\), and let \(g\) be a differentiable transformation. A first-order Taylor expansion around \(\mu\) gives

\[\begin{equation} g(Y) \approx g(\mu)+g'(\mu)(Y-\mu). \label{eq:ch03-delta-taylor} \end{equation}\]

Taking variances in this approximation yields

\[\begin{equation} \operatorname{Var}(g(Y)) \approx \left[g'(\mu)\right]^2\operatorname{Var}(Y) = \left[g'(\mu)\right]^2v(\mu). \label{eq:ch03-delta-variance} \end{equation}\]

To make the transformed variance approximately constant, choose \(g\) so that

\[\begin{equation} \left[g'(\mu)\right]^2v(\mu)=c^2, \label{eq:ch03-vst-condition} \end{equation}\]

where \(c>0\) is an arbitrary constant. Equivalently,

\[\begin{equation} g'(\mu)=\frac{c}{\sqrt{v(\mu)}}. \label{eq:ch03-vst-derivative} \end{equation}\]

Thus a variance-stabilizing transformation is obtained by integration:

\[\begin{equation} g(\mu)= \int^\mu \frac{c}{\sqrt{v(t)}}\,dt. \label{eq:ch03-vst-integral} \end{equation}\]

The arbitrary multiplicative and additive constants do not change the variance-stabilizing property. They only change the scale and origin of the transformed data. This is why different software implementations can use slightly different constants while representing the same underlying idea.

3.7 Important Variance-Stabilizing Transformations

For a Poisson model, \(v(\mu)=\mu\). Taking \(c=1\) gives

\[\begin{equation} g(\mu)=\int^\mu \frac{1}{\sqrt{t}}\,dt = 2\sqrt{\mu}. \label{eq:ch03-poisson-vst} \end{equation}\]

Up to scale, the square-root transform is therefore the Poisson variance-stabilizing transform. Anscombe’s classical correction refines this for small counts:

\[\begin{equation} g(y)=2\sqrt{y+\frac{3}{8}}. \label{eq:ch03-anscombe-transform} \end{equation}\]

For a quadratic variance function \(v(\mu)=a\mu^2\), one obtains

\[\begin{equation} g(\mu)=\int^\mu \frac{1}{\sqrt{a}t}\,dt = \frac{1}{\sqrt{a}}\log(\mu). \label{eq:ch03-log-vst} \end{equation}\]

This explains why logarithms are so common: they stabilize variance when the standard deviation is proportional to the mean. In real count data, however, zeros make \(\log(y)\) undefined and small counts make the log extremely steep. For a normalized count \(\tilde{x}_{ij}=c_{ij}/\hat{s}_i\), the shifted logarithm is

\[\begin{equation} x_{ij} = \log\left(\tilde{x}_{ij}+\kappa\right) = \log\left(\frac{c_{ij}}{\hat{s}_i}+\kappa\right), \qquad \kappa>0, \label{eq:ch03-shifted-log} \end{equation}\]

and is therefore a pragmatic transformation after size-factor normalization. The pseudocount \(\kappa\) is not just a numerical trick. It determines how strongly low counts are compressed or separated.

For the negative-binomial variance function \(v(\mu)=\mu+\phi\mu^2\), the delta-method integral gives

\[\begin{equation} g_\phi(\mu) = \int^\mu \frac{1}{\sqrt{t+\phi t^2}}\,dt. \label{eq:ch03-nb-vst-integral} \end{equation}\]

One convenient scaled form is

\[\begin{equation} g_\phi(y) = \frac{1}{\sqrt{\phi}} \operatorname{acosh}(2\phi y+1), \qquad \phi>0. \label{eq:ch03-nb-vst-acosh} \end{equation}\]

The same transformation can be written with the inverse hyperbolic sine as

\[\begin{equation} g_\phi(y) = \frac{2}{\sqrt{\phi}} \operatorname{asinh}\left(\sqrt{\phi y}\right), \qquad \phi>0. \label{eq:ch03-nb-vst-asinh} \end{equation}\]

Equivalently,

\[\begin{equation} g_\phi(y) = \frac{1}{\sqrt{\phi}} \log\left( 2\phi y+1+ \sqrt{(2\phi y+1)^2-1} \right). \label{eq:ch03-nb-vst-log-form} \end{equation}\]

The identity \(\operatorname{acosh}(2u+1)=2\operatorname{asinh}(\sqrt{u})\) shows that Equations \(\ref{eq:ch03-nb-vst-acosh}\) and \(\ref{eq:ch03-nb-vst-asinh}\) are the same. The logarithm is present because

\[\begin{equation} \operatorname{acosh}(z) = \log\left(z+\sqrt{z^2-1}\right), \qquad z\geq 1. \label{eq:ch03-acosh-log-identity} \end{equation}\]

The important qualitative behavior is more important than the exact constant: for small \(y\), the transform is square-root-like, and for large \(y\) with \(\phi>0\), it is log-like. Guan gives a more detailed theoretical treatment of variance-stabilizing transformations for Poisson, binomial, and negative-binomial distributions (Yu 2009). The same negative-binomial/gamma-Poisson idea is derived with the delta method in MSMB VST Delta Method.

3.8 DESeq2 VST, rlog, And Model-Based Transformations

DESeq2 uses the negative-binomial model to estimate size factors and dispersions, and it provides transformations for exploratory analysis (Love et al. 2014). In the notation of this script, the conceptual model is

\[\begin{equation} C_{ij}\sim \operatorname{NB}(\mu_{ij},\phi_j), \qquad \mu_{ij}=s_i q_{ij}. \label{eq:ch03-deseq-model} \end{equation}\]

The variance-stabilizing transformation uses the estimated mean-dispersion relationship to define a transformation whose variance is approximately stable across the range of expected counts. This is why the VST should not be understood as a generic formula applied blindly to counts. It is a transformation derived from an estimated count model.

The regularized logarithm, or rlog, has a different motivation. It fits a log-scale model with shrinkage of sample effects, producing log-like transformed values that are less sensitive to very different size factors. The shifted logarithm, VST, and rlog are contrasted for RNA-seq data in MSMB RNA-seq VST. For this course, the practical distinction is:

Transformation Model idea Typical use
shifted log approximate log-scale stabilization after size-factor normalization fast exploratory analysis
VST delta-method transform using estimated mean-variance trend PCA, clustering, heatmaps
rlog regularized log-scale model with shrinkage exploratory analysis when size factors vary strongly
Pearson residuals standardized residuals from a count model single-cell normalization and feature selection

The table is only a guide. The right choice depends on the assay, library-size variation, sparsity, downstream method, and whether the goal is visualization, clustering, network analysis, or formal inference.

3.9 Single-Cell RNA-seq: Simple Logs And Residuals

Single-cell RNA-seq makes the normalization problem sharper. The count matrix is sparse, library sizes vary strongly across cells, and many downstream methods operate on Euclidean distances in a low-dimensional embedding. A common preprocessing step is

\[\begin{equation} x_{ij}= \log\left( \frac{c_{ij}}{\hat{s}_i}+1 \right). \label{eq:ch03-single-cell-log1p} \end{equation}\]

Transformation choices for single-cell RNA-seq have been compared across delta-method transformations, model residuals, inferred latent expression states, and factor-analysis approaches (Ahlmann-Eltze and Huber 2023). In these benchmarks, a simple shifted logarithm followed by PCA performed as well as or better than many more elaborate alternatives. This does not mean that modeling is unnecessary. It means that preprocessing should be judged both by theoretical behavior and by downstream performance.

SCTransform takes a residual-based route. It fits regularized negative-binomial regression models with sequencing depth as a covariate and uses Pearson residuals as transformed values (Hafemeister and Satija 2019). In simplified notation, suppose

\[\begin{equation} C_{ij}\sim \operatorname{NB}(\mu_{ij},\phi_j), \qquad \log\mu_{ij}=\beta_{0j}+\beta_{1j}\log N_i. \label{eq:ch03-sctransform-model} \end{equation}\]

The Pearson residual is

\[\begin{equation} r_{ij} = \frac{c_{ij}-\hat{\mu}_{ij}} {\sqrt{\hat{\mu}_{ij}+\hat{\phi}_j\hat{\mu}_{ij}^2}}. \label{eq:ch03-pearson-residual} \end{equation}\]

This is not a simple monotone transformation of each count. It subtracts a fitted mean and divides by a fitted standard deviation. The connection to variance stabilization is direct: both approaches use a mean-variance model to put features on a more comparable scale. The difference is that VST transforms the observed count scale, while Pearson residuals measure deviations from a fitted model.

3.10 Compositional Normalization And Log-Ratios

For microbiome and amplicon data, a row often carries primarily relative information. The multinomial model from Chapter 02 wrote

\[\begin{equation} C_i\mid N_i,\pi_i \sim \operatorname{Multinomial}(N_i,\pi_i), \qquad \pi_i\in\Delta^{p-1}. \label{eq:ch03-multinomial-compositional-model} \end{equation}\]

Total-count scaling estimates \(\pi_i\) by

\[\begin{equation} \hat{\pi}_{ij}=\frac{c_{ij}}{N_i}. \label{eq:ch03-relative-abundance} \end{equation}\]

But ordinary Euclidean distances between relative abundance vectors ignore the geometry of the simplex. In addition, Equation \(\ref{eq:ch03-relative-abundance}\) can contain exact zeros, and log-ratio transformations require positive entries. A simple pseudocount-and-closure estimate is

\[\begin{equation} \hat{\pi}_{ij}^{(\kappa)} = \frac{c_{ij}+\kappa}{N_i+p\kappa}, \qquad \kappa>0. \label{eq:ch03-pseudocount-closure} \end{equation}\]

The choice of \(\kappa\) affects low-count features and should be reported. As introduced in Chapter 02, the centered log-ratio transform of a positive composition is

\[\begin{equation} \operatorname{clr}(\pi_i)_j = \log\frac{\pi_{ij}}{g(\pi_i)}, \qquad g(\pi_i)= \left(\prod_{j=1}^p \pi_{ij}\right)^{1/p}. \label{eq:ch03-clr-transform} \end{equation}\]

The CLR transform returns a \(p\)-dimensional vector whose components sum to zero. It is symmetric in the components and useful for defining compositional geometry, but the transformed coordinates are constrained rather than free coordinates in \(\mathbb{R}^p\).

The additive log-ratio transform chooses one component as a reference. If component \(r\) is the reference, then for \(j\neq r\),

\[\begin{equation} \operatorname{alr}_{r}(\pi_i)_j = \log \left( \frac{\pi_{ij}}{\pi_{ir}} \right), \qquad \operatorname{alr}_{r}(\pi_i)\in\mathbb{R}^{p-1}. \label{eq:ch03-alr-transform} \end{equation}\]

ALR coordinates are unconstrained, but they depend on the chosen reference component. This can be useful when the reference has scientific meaning or is stable, but it also means that interpretation is relative to that choice.

The isometric log-ratio transform uses an orthonormal basis of the CLR subspace. Let \(H\in\mathbb{R}^{p\times(p-1)}\) satisfy

\[\begin{equation} H^\top H=I_{p-1}, \qquad H^\top\mathbf{1}=0. \label{eq:ch03-ilr-basis} \end{equation}\]

Then

\[\begin{equation} \operatorname{ilr}(\pi_i) = H^\top\operatorname{clr}(\pi_i) \in \mathbb{R}^{p-1}. \label{eq:ch03-ilr-transform} \end{equation}\]

ILR coordinates are orthonormal coordinates for the simplex geometry, but their individual coordinates depend on the chosen basis \(H\). In practice, this basis may be chosen through balances, trees, or other scientifically meaningful contrasts. Distances induced by these log-ratio transformations are treated in Chapter 04, where distances and embeddings are the central topic.

Exact zeros require special handling before applying logarithms. This is not a minor technicality, because sparse microbiome and single-cell matrices can contain many zeros. Exercise Sheet 03 develops this direction by using normalization, log-ratio transformations, and MDS for 16S sequencing data (Exercise Sheet 03 2026). The main conceptual point is that compositional normalization is not simply another size-factor method. It changes the coordinate system of the data analysis problem.

3.11 How Transformation Choices Affect Downstream Analysis

Many downstream algorithms implicitly assume that Euclidean geometry is meaningful. PCA, \(k\)-means, hierarchical clustering with Euclidean distance, and many regression methods are sensitive to scale. If raw counts are used, high-count features and high-depth observations can dominate the analysis. If log-transformed counts are used, low counts and pseudocount choices can matter strongly. If VST or residuals are used, the transformed scale is tied to an estimated mean-variance model.

This is why Chapter 3 is placed before distances, embeddings, and clustering. A distance between rows of \(X\) is not just a property of the biological samples; it is a property of the preprocessing map

\[\begin{equation} C \longmapsto X. \label{eq:ch03-preprocessing-map} \end{equation}\]

For exploratory analysis, transformed matrices are often appropriate:

\[\begin{equation} X_{ij}=g\left(\frac{C_{ij}}{\hat{s}_i}\right). \label{eq:ch03-exploratory-transformed-matrix} \end{equation}\]

For differential expression or differential abundance testing, it is often better to keep the count likelihood and include size factors in the model:

\[\begin{equation} C_{ij}\sim \operatorname{NB}(\mu_{ij},\phi_j), \qquad \log \mu_{ij} = \log s_i + z_i^\top\beta_j. \label{eq:ch03-count-regression-with-offset} \end{equation}\]

Here \(z_i^\top\) is row \(i\) of the design matrix \(Z\in\mathbb{R}^{n\times r}\), and \(\log s_i\) is an offset. This equation is the bridge to the testing and regression chapters: normalization can enter the model as an offset rather than as a preliminary division.

3.12 Practical Summary

Normalization answers the question: which part of the observed scale is nuisance exposure? Variance stabilization answers the question: which transformation makes the noise approximately comparable across the mean range? Compositional log-ratio analysis answers a different question: which geometry is appropriate when only relative information is observed?

The main take-home message is that modeling and transformations are tied together. The Poisson model leads to a square-root transform. A quadratic mean-variance relationship leads to a logarithm. The negative-binomial model leads to a square-root-to-log VST. A fitted negative-binomial regression model leads naturally to Pearson residuals. A multinomial or logistic-normal view of compositions leads to log-ratio transformations. These are not unrelated tricks; they are different faces of the same generative thinking from Chapter 02.

References

Ahlmann-Eltze, Constantin, and Wolfgang Huber. 2023. “Comparison of Transformations for Single-Cell RNA-seq Data.” Nature Methods, ahead of print. https://doi.org/10.1038/s41592-023-01814-1.
Exercise Sheet 02. 2026. Exercise Sheet 02 Solutions: Normalization and Variance-Stabilizing Transformations of Sequencing Data. https://bio-datascience.github.io/ahbd/assets/exercises/AHDBD_SS26_Exercise02_Solution.html.
Exercise Sheet 03. 2026. Exercise Sheet 03 Solutions: Normalization and Low Dimensional Representation of 16S Sequencing. https://bio-datascience.github.io/ahbd/assets/exercises/AHDBD_SS26_Exercise03_Solution.html.
Hafemeister, Christoph, and Rahul Satija. 2019. “Normalization and Variance Stabilization of Single-Cell RNA-seq Data Using Regularized Negative Binomial Regression.” Genome Biology 20 (296). https://doi.org/10.1186/s13059-019-1874-1.
Love, Michael I, Wolfgang Huber, and Simon Anders. 2014. “Moderated Estimation of Fold Change and Dispersion for RNA-Seq Data with DESeq2.” Genome Biology 15 (12). https://doi.org/10.1186/s13059-014-0550-8.
McMurdie, Paul J., and Susan Holmes. 2014. “Waste Not, Want Not: Why Rarefying Microbiome Data Is Inadmissible.” PLoS Computational Biology 10 (4): e1003531. https://doi.org/10.1371/journal.pcbi.1003531.
Weiss, Sophie, Zhenjiang Zech Xu, Shyamal Peddada, et al. 2017. “Normalization and Microbial Differential Abundance Strategies Depend Upon Data Characteristics.” Microbiome 5 (1). https://doi.org/10.1186/s40168-017-0237-y.
Yu, Guan. 2009. “Variance Stabilizing Transformations of Poisson, Binomial and Negative Binomial Distributions.” Statistics &Amp; Probability Letters 79 (14): 1621–29. https://doi.org/10.1016/j.spl.2009.04.010.