Information Theory Applications
Entropy is not only abstract: it shows up in compression, feature selection, decision trees, language-model evaluation, privacy leakage bounds, and ML objectives. This chapter maps the definitions from earlier chapters onto engineering practice.
Diagram: map of applications
entropy H(X)
│
├── compression / coding length
├── decision tree split (info gain)
├── feature selection (mutual information)
├── channel capacity (noisy pipes)
├── LM evaluation (perplexity)
└── ML losses (cross-entropy / KL)
1. Compression and source coding
Shannon source coding theorem (informal): for an i.i.d. source with entropy \(H(X)\), lossless compression needs about \(H(X)\) bits per symbol in the large-block limit; you cannot do better on average, and good codes approach \(H(X)\).
| Source property | Compressibility |
|---|---|
| Low \(H(X)\) | highly predictable → shorter codes |
| Max entropy (uniform) | incompressible at symbol level |
| Dependence over time | use conditional / joint coding; rate \(\approx H(X_t\mid\text{past})\) |
Practical codecs (conceptual)
- Huffman: optimal prefix codes for known symbol frequencies (per-symbol)
- Arithmetic / ANS: approach \(H\) more closely for streams
- LZ family: universal; learn repeats without an explicit \(P\)
Worked example 1
Fair six-sided die: \(H=\log_2 6\approx 2.585\) bits/roll. No lossless scheme averages fewer bits per independent roll in the Shannon limit.
Biased coin \(P(X=1)=0.1\): \(H_2(0.1)\approx 0.469\) bits — much more compressible than a fair coin.
2. Information gain and decision trees
Split feature \(A\) to reduce uncertainty about label \(Y\):
\[ \mathrm{IG}(Y,A)=H(Y)-H(Y\mid A)=I(Y;A). \]
before split: H(Y) large (mixed labels)
│
├── A=yes → purer → small H
└── A=no → purer → small H
weighted average H(Y|A) < H(Y) ⇒ positive gain
Conditional entropy after split:
\[ H(Y\mid A)=\sum_a P(A=a)\, H(Y\mid A=a). \]
Worked example 2 — pure split
Binary label \(50/50\), \(H(Y)=1\) bit. Split yields two pure leaves: \(H(Y\mid A)=0\), so \(\mathrm{IG}=1\) bit.
Practical caveats
- Greedy IG can miss globally optimal feature sets
- High-arity features can look good by fragmentation (many pure tiny leaves) — use gain ratio / regularization / min leaf size
- Continuous features need thresholds; search is part of the algorithm
3. Mutual information for features and dependence
\[ I(X;Y)=H(X)-H(X\mid Y)=H(Y)-H(Y\mid X)=D_{\mathrm{KL}}(P_{X,Y}\|P_X P_Y). \]
- \(I(X;Y)=0\) iff \(X\perp Y\) (discrete case under standard conditions)
- Feature ranking: score features by \(\hat I(X_j;Y)\)
- Caution: estimating MI in high dimension is hard (bias, curse of dimensionality); use careful estimators or proxies
Worked example 3 — independence
If \(X\perp Y\), then \(P_{X,Y}=P_X P_Y\), so \(D_{\mathrm{KL}}=0\) and \(I(X;Y)=0\).
Correlation vs MI
Uncorrelated does not imply independent. Example: discrete \(X\in\{-1,0,1\}\) symmetric, \(Y=X^2\). Correlation can be zero while \(I(X;Y)>0\).
4. Channels and capacity (systems view)
Noisy channel: input \(X\), output \(Y\), law \(p(y\mid x)\).
Capacity
\[ C=\max_{p(x)} I(X;Y) \]
is the supremum of rates (bits per channel use) for reliable communication (channel coding theorem, informal).
| Setting | Role of capacity |
|---|---|
| Wireless / modem design | ultimate rate ceiling |
| FEC vs retransmission | operate below \(C\) with codes |
| Disk / flash | constrained coding + ECC |
Worked example 4 — BSC intuition
Binary symmetric channel flip probability \(p\): \(C=1-H_2(p)\) bits (for fair-ish optimal input). At \(p=0\), \(C=1\); at \(p=1/2\), \(C=0\).
5. Perplexity and language models
For a model \(q\) evaluated on tokens from \(p\) (or on a test corpus treated as empirical \(p\)):
Cross-entropy rate (nats or bits per token) and perplexity
\[ \mathrm{PPL}=2^{H(p,q)}\quad\text{(bits convention)} \]
or \(e^{H}\) in nats. Interpretation: effective branching factor — how many sides a fair die would need to match the model’s average surprise.
Lower perplexity ↔︎ model less surprised by the test data (better coding of the corpus), not automatically better task utility.
Worked example 5
If average CE is \(3\) bits/token, \(\mathrm{PPL}=2^3=8\).
6. ML training objectives (recap with applications)
| Objective | Information view |
|---|---|
| Cross-entropy classification | code labels with model \(q_\theta\) |
| KL to teacher (distillation) | match soft distributions |
| InfoNCE / contrastive | lower bound / proxy related to MI |
| Variational bounds | ELBO involves KL to posterior |
Training with CE is not “just a loss hack”: it is maximum likelihood under a categorical model and coding-optimal under that model class.
7. Decision systems and expected information
Before an experiment or A/B test, expected information gain about a parameter \(\theta\) from observing \(Y\):
\[ \mathrm{EIG}=H(\theta)-H(\theta\mid Y)=I(\theta;Y). \]
Bayesian optimal design picks interventions maximizing EIG (or related utilities). Same math as feature selection with a different “feature.”
8. Privacy and leakage (conceptual)
If \(S\) is a secret and \(R\) is a released message (query answer, model prediction, noisy aggregate):
\[ I(S;R) \]
measures residual dependence — a leakage metric. Perfect privacy in the MI sense would require \(I(S;R)=0\) (often too strong); differential privacy uses a different, worst-case neighboring-dataset guarantee, but MI still guides intuition about how much a release can say about individuals.
9. Hashing, randomness, and min-entropy
Shannon entropy averages surprise; min-entropy
\[ H_{\infty}(X)=-\log\max_x P(x) \]
captures worst-case guessability (passwords, cryptographic keys). High Shannon entropy with a heavy atom can still be weak for security. CS security prefers min-entropy and extractors.
10. End-to-end worked case: email spam filter features
- Labels \(Y\in\{\mathrm{spam},\mathrm{ham}\}\) with empirical \(H(Y)\).
- For each binary feature \(A_j\) (“word \(w\) present”), estimate \(\mathrm{IG}(Y,A_j)\).
- Keep top-\(k\) features or use MI filter then train logistic regression.
- Train with CE loss; report CE and calibration on a holdout.
- Compression analogy: a good model assigns short codes (high probability) to true labels.
11. Pitfalls
- Treating high entropy as “meaningful content”
- Greedy IG trees without depth/leaf constraints
- MI feature selection without correcting for finite-sample bias
- Using perplexity alone to claim product quality
- Ignoring that capacity is an asymptotic, ideal-code limit
- Confusing Shannon entropy with cryptographic strength
12. Checkpoint
- State source coding lower bound \(H(X)\)
- Compute information gain for a pure binary split
- Write \(I(X;Y)\) three ways
- Define channel capacity as max MI
- Convert CE to perplexity
- Distinguish Shannon vs min-entropy for security
Exercises
Easy
- Binary label 50/50, split yields two pure leaves. What is IG in bits?
- If \(X\perp Y\), what is \(I(X;Y)\)?
- Compute \(H\) of a fair six-sided die in bits (\(\log_2 6\)).
- Connect cross-entropy training to “coding with model \(Q\).”
- Perplexity if CE \(=2\) bits/token?
Medium
- Why is estimating \(I(X;Y)\) hard for high-dimensional continuous \(X\)?
- Name one failure mode of greedy info-gain tree splits.
- BSC: explain why \(C\to 0\) as \(p\to 1/2\).
- Feature \(A\) independent of \(Y\): what is \(\mathrm{IG}(Y,A)\)?
- Password distribution with \(H(X)=20\) bits but one password has probability \(2^{-5}\): bound \(H_{\infty}\).
Challenge
- Construct (small joint table) \(X,Y\) with \(\mathrm{Corr}=0\) but \(I(X;Y)>0\).
- Gain ratio: define \(\mathrm{IG}(Y,A)/H(A)\) and explain the high-arity bias fix.
- Show that \(I(X;Y)\le \min\{H(X),H(Y)\}\).
- Sketch how arithmetic coding achieves average length near \(H\) for a known \(P\).
- Privacy sketch: adding noise to \(R\) tends to decrease \(I(S;R)\); discuss utility tradeoff \(I(S;\hat t(R))\) for a target statistic \(t\).
Checks
- \(H(Y)=1\), \(H(Y\mid A)=0\), IG \(=1\).
- \(0\).
- \(\approx 2.585\) bits.
- \(\mathrm{PPL}=4\).
Summary
Information measures are engineering tools: \(H\) bounds compression, \(I\) scores dependence and channels, CE/KL train and evaluate probabilistic models, perplexity rephrases CE for language, and min-entropy speaks to worst-case unpredictability. Use the right functional for the job, and never forget estimation and asymptotics when moving from chalkboard to production data.