# inverse

*Entry of the [Dictionary of Applied Machine Learning](https://dictionaryofml.org) (course edition), by Alexander Jung, Aalto University. CC BY 4.0. Updated 2026-10-29. [HTML](https://dictionaryofml.org/terms/inverse.html) · [typeset PDF](https://dictionaryofml.org/terms/inverse.pdf) · [LaTeX](https://dictionaryofml.org/terms/inverse.tex) · [Python demo](https://dictionaryofml.org/terms/inverse.py) · [notebook](https://dictionaryofml.org/terms/inverse.ipynb)*

Math is LaTeX with the book's macros expanded, so no notation package is needed to read it.

## Abstract

An inverse of a function is a second function that undoes it. For $f: \mathcal{U} \to \mathcal{V}$ it is the function $g: \mathcal{V} \to \mathcal{U}$ with $g(f(u)) = u$ and $f(g(v)) = v$, and it exists precisely when $f$ is injective and every element of $\mathcal{V}$ has a nonempty preimage. The inverse of a linear map of $\mathbb{R}^{n}$ is again a linear map, represented by the inverse matrix ${\mathbf A}^{-1}$ with ${\mathbf A} {\mathbf A}^{-1} = {\mathbf A}^{-1} {\mathbf A} = \mathbf{I}$, which exists exactly when the square matrix ${\mathbf A}$ is of full-rank, equivalently when its determinant is nonzero. Where no inverse exists, a left or right inverse or the Moore–Penrose pseudoinverse takes its place. In machine learning, inverses appear in the closed-form solution of linear regression, in the pair formed by the sigmoid function and the logit, which turns a score into a probability and back, and in drawing a realization through the inverse of a cumulative distribution function.

## Definition

A frost warning for Krems an der Donau delivers a real-valued score for the day, and the sigmoid function turns that score into the probability of frost: the score $0$ becomes the probability $0.5$, the score $2.2$ becomes $0.9$. Reading the score back off a reported probability undoes this step, and the function that does so is the logit $\ln\big( p / (1-p) \big)$. The logit is the inverse of the sigmoid function.

An inverse of a function $f: \mathcal{U} \to \mathcal{V}$ is a function $g: \mathcal{V} \to \mathcal{U}$ that undoes it in both directions: $g\big(f(u)\big) = u$ for every $u \in \mathcal{U}$, and $f\big(g(v)\big) = v$ for every $v \in \mathcal{V}$. Such a $g$ exists precisely when $f$ is injective and every $v \in \mathcal{V}$ has a nonempty preimage. It is then unique and written $f^{-1}$. If only the first identity holds, $g$ is a left inverse, and if only the second holds, a right inverse. Fig. 1 shows both cases for a function on a domain of four shapes.

![Figure 1 of the entry inverse](https://dictionaryofml.org/terms/fig/inverse-fig1.svg)

*Figure 1: Two functions on the same domain $\mathcal{U}$ of four shapes, with $\mathcal{U}$ and the set $\mathcal{V}$ of values each drawn as a thin ellipse. In (a) the function $f$ sends a shape to its number of corners, and every value of $\mathcal{V} = \{0, 3, 4, 5\}$ receives exactly one solid arrow, so following the arrows backwards defines the inverse $f^{-1}$ (dashed arrows), which returns the shape with that number of corners. In (b) the function $f'$ records only whether a shape has a straight edge. Three shapes share the value yes, so no function undoes $f'$: the value yes does not say which shape produced it*

The best known instance is the inverse of a linear map. A square matrix ${\mathbf A} \in \mathbb{R}^{n \times n}$ represents a linear map of $\mathbb{R}^{n}$ into itself, and that map has an inverse exactly when ${\mathbf A}$ is of full-rank, meaning its columns are linearly independent. Such a matrix is said to be invertible, and the matrix representing the inverse linear map is the inverse matrix ${\mathbf A}^{-1}$, which satisfies

$$
{\mathbf A} {\mathbf A}^{-1} = {\mathbf A}^{-1} {\mathbf A} = \mathbf{I} \text{.}
$$

A square matrix is invertible if and only if its determinant is nonzero (Horn and Johnson, 1991; G. Strang, 2007). Fig. 2 shows the two linear maps acting on the standard basis of $\mathbb{R}^{2}$.

![Figure 2 of the entry inverse](https://dictionaryofml.org/terms/fig/inverse-fig2.svg)

*Figure 2: Matrix ${\mathbf A}$ representing a linear transformation of $\mathbb{R}^{2}$; inverse matrix ${\mathbf A}^{-1}$ representing the inverse transformation*

For a matrix that is not square or not of full-rank, the one-sided inverses are what remains: a left inverse ${\mathbf B}$ satisfies ${\mathbf B} {\mathbf A} = \mathbf{I}$, and a right inverse ${\mathbf C}$ satisfies ${\mathbf A} {\mathbf C} = \mathbf{I}$. The Moore–Penrose pseudoinverse ${{\mathbf A}}^{+}$ selects one generalized inverse for every matrix (Golub and Loan, 2013).

Inverses are used throughout machine learning. Solving a system of linear equations is one task, and ridge regression is another: its closed-form solution inverts the matrix ${\mathbf X}^{\top} {\mathbf X} + \alpha \mathbf{I}$, which is invertible for every positive regularization strength $\alpha$. The sigmoid function and the logit of the opening paragraph are an invertible pair, which allows translating between the score of a linear classifier and a probability value. A feature transformation that is invertible loses no information, since the raw feature vector can be recovered from the transformed one. The inverse of the cumulative distribution function of a random variable allows generating realizations of that random variable from a realization of another, simpler random variable, one uniformly distributed on $[0,1]$ (Ross, 2014).

A counterfactual explanation for a learned hypothesis can be constructed from a right inverse of it, which returns, for a given prediction, one feature vector of its preimage. A learned hypothesis usually maps many feature vectors to the same prediction, so it has no inverse and no left inverse; a right inverse is what remains, and choosing which element of the preimage it returns is what makes one counterfactual more useful than another.

**See also:** function, injective, preimage, invertible, [matrix](https://dictionaryofml.org/terms/matrix.html), linear map, determinant, pseudoinverse, [linear regression](https://dictionaryofml.org/terms/linreg.html), [sigmoid function](https://dictionaryofml.org/terms/sigmoid.html), logit.

## References

- Horn and Johnson (1991). Topics in Matrix Analysis. *Cambridge Univ. Press*.
- G.~Strang (2007). Computational Science and Engineering. *Wellesley-Cambridge Press*.
- Golub and Loan (2013). Matrix Computations. *The Johns Hopkins Univ. Press*.
- Ross (2014). A First Course in Probability. *Pearson Education*.

## Cite this entry

```bibtex
@misc{dictml_inverse,
  author = {Jung, Alexander},
  editor = {Olioumtsevits, Konstantina and Schnoor, Ekkehard},
  title = {inverse},
  howpublished = {Dictionary of Applied Machine Learning (course edition)},
  year = {2026},
  doi = {10.5281/zenodo.21569296},
  note = {ISBN 978-952-64-3013-3, CC BY 4.0, retrieved 2026-10-29},
  url = {https://dictionaryofml.org/terms/inverse.html}
}
```