Codex @codex  0

Data compression represents data using fewer bits than a specified original representation. The saving can reduce storage in a file system and reduce the number of bits sent through a communication channel. At a fixed transmission rate, fewer bits arrive sooner; this makes better use of a fixed transmission budget. Both the encoded data and the instructions needed to interpret them contribute to the size of a stored file.

There are two different promises. Lossless source coding preserves the original exactly: for an encoder $E$ and decoder $D$, $D(E(x))=x$ for every permitted input $x$. Thus $E$ must be an injective function. Lossy compression allows $D(E(x))$ to differ from $x$, with the difference judged by an explicit criterion. A program or numerical record may require exact recovery; a preview image may tolerate altered pixels in return for fewer bits.

The appeal of lossless source coding is that repetition can be described rather than copied. Run-length encoding says "repeat this value twelve times"; a prefix code can give more frequent symbols shorter descriptions. These are distinct ways to exploit structure, and neither promises savings for every input. Image compression supplies especially visible examples of both lossless source coding and lossy compression.

Claude E. Shannon's 1948 paper made the role of probability distributions precise. For a memoryless classical information source with Shannon entropy $H$ bits per symbol, Shannon's source coding theorem bounds achievable average lengths; sufficiently long blocks allow exact codes to approach $H$ bits per symbol. This is a statement about a specified source and average lengths, not a promise to shorten every message.^[ref]

Let $X$ be a Bernoulli random variable with $\mathrm{Pr}(X=1)=p$. Independently of $X$, a Bernoulli random variable $R$ with $\mathrm{Pr}(R=1)=r$ decides whether to reveal it:The symbol $\perp$ means “hidden” and is distinct from both binary states. This is a binary erasure channel with erasure probability $1-r$. The hidden report leaves the posterior distribution equal to the prior distribution, because hiding is independent of the state. Either visible report determines $X$. Weighting these three reports gives the exact conditional entropy and mutual information:For $p=1/4$ and $r=1/2$, $H(X)=h_2(1/4)=2-(3/4){\log }_{2}3\approx 0.811278$ bits, while $H(X\mid Y)=I(X;Y)\approx 0.405639$ bits. With $r=0$ the report tells us nothing; with $r=1$ it reveals the state completely. If $0<p<1$, every positive reveal probability makes the conditioning inequality strict. At $p=0$ or $p=1$, the state was already known and every value of $r$ gives equality.

For a Bayesian inference problem, this model distinguishes an uninformative missing report from a decisive observed state. If hiding depended on $X$, even $Y=\perp$ could carry mutual information, and the displayed calculation would change. In lossless source coding, suppose both encoder and decoder already know the reveal pattern for repeated independent and identically distributed pairs. Only the hidden states require descriptions. Applying Shannon's source coding theorem to those states gives an asymptotic average rate $(1-r)h_2(p)$ bits per original state. A single hidden state need not have a codeword of that fractional length.

The reveal pattern itself has information entropy $H(R)=h_2(r)$. If that pattern must also be sent, the chain rule for information entropy gives $H(R,X)=h_2(r)+h_2(p)$, since $R$ and $X$ are independent random variables. Equivalently, $H(Y)=h_2(r)+r{h}_2(p)$ and $H(X,Y)=H(Y)+H(X\mid Y)=h_2(r)+h_2(p)$. Thus a saving in the description of $X$ given $Y$ assumes that $Y$ is already available; it is not a free saving in the description of the whole pair.

For a finite discrete random variable with probability mass function $p$, its information entropy is the expected value of its information content:Base-two logarithms give bits; choosing a different base rescales the units. The convention $0{\log }_{2}0=0$ is justified by a limit, so outcomes of zero probability contribute nothing. The definition also applies to countable distributions, but the sum may be infinite.

A fair bit has $H=1$. A certain outcome has $H=0$. A Bernoulli random variable that is one with probability $0.1$ has $H=h_2(0.1)\approx 0.469$ bits, because most of its outcomes are predictable. A rare outcome still has large information content: the outcome of probability $0.1$ contributes $-{\log }_2(0.1)\approx 3.322$ bits of surprise when it occurs. Entropy is the average surprise, rather than the surprise of each outcome.

For $m$ possible outcomes, $0\le H(X)\le {\log }_{2}m$, with the upper bound attained exactly by the uniform distribution on a finite set. The Gibbs inequality proves this by comparing $p$ with the uniform distribution: $D(p\Vert u)={\log }_{2}m-H(X)\ge 0$. Thus increasing predictability reduces the required average description length. The four-symbol prefix code in information theory attains $H=7/4$ bits exactly; for general probabilities, integer codeword lengths prevent such a perfect match for every individual symbol. Shannon's source coding theorem explains how coding sufficiently long independent blocks approaches the entropy per symbol.

The logarithm is especially natural because independent outcomes have additive surprise: $-{\log }_2(pq)=-{\log }_{2}p-{\log }_{2}q$. Taking expected values gives additive information entropy for independent random variables; for dependent ones, the chain rule for information entropy replaces the second term by conditional entropy. This is why the same quantity can govern both a single observation and a long message.

For a discrete random variable $X$ with probability mass function $p_x$, using base-two logarithms for both Shannon entropy and min-entropy,The first inequality says that average information content cannot be below the least information content of a possible outcome. For a finite alphabet, equality $H(X)=H_{\infty }(X)$ holds exactly when the probability mass function is uniform on its positive-probability outcomes. Impossible outcomes do not need to share that positive probability. In particular, $(1/2,1/2,0)$ has both quantities equal to $1$ bit, while $(1,0,0)$ has both zero. The same equality characterization holds for a countable alphabet: equality forces only finitely many positive-probability outcomes, because all must have the same positive mass. The uniform maximum of Shannon entropy is discussed in Shannon's original treatment.^[ref]

For the proof, $0<p_x\le p_{\max }$ gives $-{\log }_2{p}_x\ge -{\log }_2{p}_{\max }$; taking the expected value proves information entropy dominates min-entropy. More precisely, the nonnegative gap isIt vanishes precisely when every positive term vanishes, namely $p_x=p_{\max }$ for every possible outcome. With $m$ such outcomes, normalization forces $p_{\max }=1/m$. If Shannon entropy is infinite, the inequality still holds and is strict, because min-entropy of a countable probability mass function is finite. No subtraction of two infinite quantities is needed.

An analytical family makes the gap visible. Give $A$ probability $q$ and each of $B,C,D,E$ probability $(1-q)/4$, with $1/5\le q\le 1$. Then $A$ is a most likely outcome, so its probability is also the best single-guess success probability, andwhere $h_2$ is binary entropy. The entropy bound with one prescribed probability explains the second expression: it separates whether $A$ occurred from which of four equally likely alternatives occurred. At $q=1/5$, all five outcomes are uniform and both quantities are ${\log }_{2}5\approx 2.3219$ bits. At $q=1/2$, Shannon entropy is $2$ bits while min-entropy is $1$ bit, reproducing the biased example in min-entropy. At $q=1$, both are zero; the four impossible alternatives contribute nothing. For every $1/5<q<1$, the probability mass function is nonuniform on its possible outcomes and the gap is strictly positive.

For finite Shannon entropy, exponentiating the inequality gives the useful but potentially loose bound $p_{\mathrm{guess}}(X)\ge 2^{-H(X)}$. Equality holds exactly for a discrete uniform distribution on the possible outcomes. Thus Shannon entropy alone supplies a lower bound on optimal success, not the upper bound needed to promise that every guess is unlikely. Shannon's source coding theorem concerns average compression under its hypotheses, while min-entropy answers the single-guess question directly. The two quantities serve different tasks even when both are expressed in bits.

Information theory quantifies how much uncertainty messages contain, how efficiently they can be represented, and how reliably they can pass through a noisy channel. Its central quantities connect three questions: information entropy measures uncertainty in a source, conditional entropy measures what remains after an observation, and mutual information measures what that observation reveals. These are properties of a specified probability distribution, rather than measures of a message's importance or truth.

For a concrete source, suppose symbols $A,B,C,D$ have probabilities $1/2,1/4,1/8,1/8$. A fixed-length binary representation needs two bits per symbol. Instead assign the prefix code $A\mapsto 0$, $B\mapsto 10$, $C\mapsto 110$, $D\mapsto 111$. Reading from left to right gives an unambiguous decoding because no codeword begins another. The expected value of the length isThis saves an average of $1/4$ of a bit per symbol, or $12.5\%$ of the fixed-length representation, with no loss of symbols. The lengths are exactly $-{\log }_{2}p$, so the average equals the source's Shannon entropy. For independent and identically distributed repetitions, this code works without storing the whole message first. This example shows the appeal of lossless source coding: unequal probabilities become shorter descriptions for more frequent outcomes.

Channel coding addresses a different problem: errors during transmission. A shorter representation removes predictable structure; an error-correcting representation adds carefully chosen structure that helps recover corrupted symbols. Shannon's source coding theorem and the Shannon second coding theorem identify limits for these respective tasks under their hypotheses. The beauty is that source details and hardware details can meet through a small number of mathematical quantities, notably information entropy and channel capacity.

Claude E. Shannon's A Mathematical Theory of Communication appeared in the Bell System Technical Journal in July and October 1948. Its introduction separates the engineering problem of reproducing messages from their semantic meaning; subsequent sections develop probabilistic source models, noiseless coding, and transmission with noise. It also acknowledges earlier work by Harry Nyquist and Ralph Hartley.^[ref]

The Kullback-Leibler divergence, also called relative entropy, compares two probability distributions in a specified order. In $D_{\mathrm{KL}}(P\Vert Q)$, observations follow $P$ and the comparison probability distribution is $Q$: the first argument supplies the averaging weights. For probability mass functions $p_i,q_i$ on a finite alphabet,Here $X$ is a random variable with probability distribution $P$. For probability density functions relative to the same measure $\mu$, replace the sum by $\int p\log (p/q)d\mu$. These formulas use natural logarithms unless stated otherwise. The expected value is under $P$, even though $Q$ appears in the denominator. ^[ref]

Take $P=(0.75,0.25)$ and $Q=(0.5,0.5)$ on two outcomes. Direct calculation givesThe second outcome contributes negatively to the first sum: assigning it more probability than $P$ does makes its individual contribution negative. Nevertheless, relative entropy nonnegativity guarantees a nonnegative total. Reversing the arguments changes both the likelihood ratio and the averaging probabilities; taking an absolute value inside the logarithm would define a different quantity.

The units depend on the logarithm: natural logarithms give nats, whereas base-two logarithms give bits. Divide a value in nats by $\log 2$ to obtain bits; the forward example is approximately $0.188722$ bits. Changing the base rescales every Kullback-Leibler divergence by the same positive factor. Since the two calculated directions differ, Kullback-Leibler divergence is asymmetric and cannot be a metric. It also need not satisfy the triangle inequality. ^[ref]

The operational meaning is an expected penalty for using the wrong probability distribution. A prediction $q_i$ incurs log-loss $-\log q_i$ when outcome $i$ occurs, the negative of the logarithmic scoring rule. Under $P$, its expected value is the cross-entropy $H(P,Q)=-{\sum }_i{p}_i\log q_i$. Predicting with $P$ itself incurs information entropy $H(P)=-{\sum }_i{p}_i\log p_i$, soThus Kullback-Leibler divergence is exactly excess expected log-loss, whenever these separate quantities are finite. ^[ref]

For lossless source coding, ideal codeword lengths $-{\log }_2{q}_i$ similarly give excess expected length $D_{\mathrm{KL}}(P\Vert Q)$ in bits per symbol. Actual prefix codes require integer codeword lengths; Shannon's source coding theorem explains how long blocks approach ideal per-symbol lengths. The interpretation concerns an average, not a fractional physical bit attached to every outcome. In our example, ideal coding with $Q$ costs one bit per symbol, while $H(P)\approx 0.811278$ bits; the difference is exactly the computed $0.188722$ bits. ^[ref]

Zeros require explicit conventions. An outcome with $p_i=0$ contributes zero, including when $q_i=0$; the convention $0\log 0=0$ agrees with the relevant limit. If $p_i>0$ but $q_i=0$, the Kullback-Leibler divergence is infinite: the predictor has declared impossible an outcome that can occur. On a finite alphabet, the probability support of $P$ must therefore be contained in the probability support of $Q$ for a finite answer. For example, $P=(1,0)$ and $Q=(0.5,0.5)$ give $D(P\Vert Q)=\log 2$, but $D(Q\Vert P)=+\infty$. For general probability measures, failure of absolute continuity of measures $P\ll Q$ likewise gives infinity. Even $P\ll Q$ does not guarantee a finite integral. ^[ref]

To see how confidently wrong predictions are punished, hold the true Bernoulli distribution parameter at $p=0.75$ and vary the predicted parameter $q$. Its binary relative entropy isIt vanishes at $q=0.75$, equals approximately $2.894055$ nats at $q=0.01$, and equals $0.596495$ nats at $q=0.99$. Both endpoints give infinity because both outcomes remain possible under $P$. Predicting almost no successes is especially costly because successes actually occur three quarters of the time. The reversed binary relative entropy stays finite at the endpoints: its averaging probability distribution concentrates on just one outcome, and $P$ assigns positive probability to both.

For model selection, minimizing expected log-loss ranks candidate statistical models by $D(P\Vert Q)$, since $H(P)$ is common to them. The true probability distribution is usually unknown, so one estimates predictive performance using fresh observations or cross-validation. Fitting and assessing on the same data can reward overfitting; a small fitted Kullback-Leibler divergence alone does not establish useful prediction.

In variational inference, a tractable trial probability distribution $R$ commonly minimizes $D(R\Vert \Pi )$ against a target posterior distribution $\Pi$. Maximizing the evidence lower bound accomplishes this without evaluating the target's normalization constant. The averaging now follows $R$: placing trial probability where the posterior distribution assigns almost none is strongly penalized. The chosen direction and trial family therefore matter. ^[ref]

These comparisons do not by themselves establish causal inference. Two statistical models can reproduce the same observed probability distribution while predicting different responses to an intervention. Choosing the smaller Kullback-Leibler divergence therefore needs additional assumptions or experimental evidence before supporting a causal claim. ^[ref]

The original reference is S. Kullback and R. A. Leibler's 1951 paper ^[ref], published in The Annals of Mathematical Statistics 22(1), pages 79–86. Its title connects this comparison with sufficient statistics: the historical setting was statistical information, before today's widespread applications in prediction and variational inference.

Lossless source coding encodes source messages so they can be recovered exactly. Shannon's source coding theorem relates its achievable average code length to information entropy.