Last modified: 2026-11-12 01:08:55 (PST)
A model that fits the data well gives the data a short code, but a model that is itself long to describe costs bits of its own. This page counts both costs in bits, and chooses the model that makes the sum shortest.
Definition 1 (Two-part description length) Let \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) be data, and let \(p_1, \ldots, p_m\) be models, each of which gives \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) a probability \(p_j(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}})\). Let a prefix code on the models give each model \(p_j\) a codeword of \({\ell\mathopen{}\left(p_j\right)\mathclose{}}\) bits.
The two-part description length of \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) under a model \(p_j\) is
\[{\ell\mathopen{}\left(p_j\right)\mathclose{}} + {\mathbb{s}_{p_j}\mathopen{}\left(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\right)\mathclose{}}.\]
The first part is the length of the codeword that names the model. The second part is the surprisal \(\log_2 \frac{1}{p_j(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}})}\) of \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) under the model, which is the length in bits of a code for \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) built from the model.
Example 1 (The two-part description lengths of coin flips)
A coin is flipped \(100\) times, and it lands heads \(70\) times. The chunk below follows Definition 1 with nine models, the Bernoulli models with the probability of heads \(0.1, 0.2, \ldots, 0.9\). It gives each model a codeword of the same length, and it prints the two parts and their sum for each model.
n_flips <- 100
n_heads <- 70
success_prob <- seq(0.1, 0.9, by = 0.1)
model_bits <- log2(length(success_prob))
heads_bits <- n_heads * log2(success_prob)
tails_bits <- (n_flips - n_heads) * log2(1 - success_prob)
data_bits <- -(heads_bits + tails_bits)
total_bits <- model_bits + data_bits
data.frame(success_prob, model_bits, data_bits, total_bits)Each model takes \(3.17\) bits to name, because nine equally long codewords need \(\log_2 9\) bits. The data part is smallest, at \(88.1\) bits, for the model with probability of heads \(0.7\).
Definition 2 (Minimum description length principle) Given data \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) and models \(p_1, \ldots, p_m\), the minimum description length principle chooses the model with the smallest two-part description length of \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\). The models may have codewords of different lengths.
Example 2 (The minimum description length model for coin flips)
The chunk below applies Definition 2 to the two-part description lengths of Example 1. It prints the model with the smallest two-part description length, and it prints the length of writing the \(100\) flips, with \(70\) heads, out as plain heads and tails.
The principle chooses the model with the probability of heads \(0.7\), with a two-part length of \(91.3\) bits. That is shorter than the \(100\) bits of the flips themselves, because the flips are far from fair.
Theorem 1 (Minimum description length and maximum likelihood) Let \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) be data, and let \(p_1, \ldots, p_m\) be models that give \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) the probabilities \(p_j(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}) > 0\), with the same codeword length \({\ell\mathopen{}\left(p_j\right)\mathclose{}} = \ell_0\) for every model. Then a model has the smallest two-part description length of \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\) exactly when it is a maximum likelihood estimate of the model given \(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\).
Proof. Every model has the first part \(\ell_0\), so the models differ only in the second part \({\mathbb{s}_{p_j}\mathopen{}\left(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}}\right)\mathclose{}}\). The surprisal \(\log_2 \frac{1}{u}\) decreases as the probability \(u\) increases. So the second part is smallest for the model with the largest \(p_j(\mathord{\smash[b]{\underset{\raise.3ex\hbox{$\scriptstyle\sim$}}{x}}\vphantom{x}})\), which is a maximum likelihood estimate.
Example 3 (Equal codeword lengths for the coin flips)
The \(9\) models of Example 1 have codewords of equal length, so Theorem 1 applies to the data of \(70\) heads in \(100\) flips. The chunk below compares the model with the shortest two-part description length to the model that gives the flips the highest probability.
Both models have the probability of heads \(0.7\), as the theorem says.
Remark 1 (Where the minimum description length principle comes from). Cover and Thomas (2006, sec. 14.13) state the principle as the choice of the model such that the model’s length plus the length of the data given the model is as short as possible. They give the model’s length as its Kolmogorov complexity, not as the length of a codeword from a prefix code.