The frontier map — territory 5 of 5

Information Theory & the Geometry of Learning

Bits as the currency of learning. Entropy, KL divergence and mutual information are where compression, generation and alignment turn out to be one subject — the territory that derives objectives (ELBO, InfoNCE, DPO) instead of guessing them.

The central picture: KL divergence is not symmetric — and that asymmetry is a design tool

Forward vs reverse KL — two ways to approximate a two-peaked truth

minimize KL(P ‖ Q) — "forward" Q stretches to cover every mode of P mode-covering · blurry but complete MLE / cross-entropy behaves like this minimize KL(Q ‖ P) — "reverse" Q commits to one mode and ignores the rest mode-seeking · sharp but partial variational inference · RLHF's KL leash
P — the true distribution (two peaks) Q — the model (one peak, dashed)
Same two distributions, opposite behaviors — purely from the direction of the KL. Choosing that direction is a real design decision in VAEs, distillation and alignment.

The roadmap of understanding — three levels deep

Level 1 — Core objects

surprise, bits, and shared information
entropy  ·  cross-entropy  ·  KL divergence  ·  mutual information

The ideas

  • Entropy: the irreducible surprise of a source
  • Cross-entropy: surprise under the wrong model
  • KL: the gap between them — extra bits paid
  • Mutual information: what knowing X tells you about Y

Where it lives

  • The LLM training objective, read as coding theory
  • Every "distance between distributions" argument

What it unlocks

  • Seeing prediction = compression: a better LLM is literally a better compressor of text

Level 2 — Working theory

bounds you can train on
variational bounds (ELBO)  ·  InfoNCE  ·  rate–distortion  ·  KL asymmetry

The ideas

  • ELBO: when the true likelihood is intractable, optimize a tractable lower bound
  • InfoNCE: a trainable lower bound on mutual information — contrastive learning, derived
  • Rate–distortion: the best possible compression at a given quality
  • Forward vs reverse KL as a behavioral dial (the figure above)

Where it lives

  • VAEs — and diffusion models, whose training loss is an ELBO in disguise
  • CLIP and every contrastive embedder
  • Distillation objectives (which KL, which temperature)

What it unlocks

  • Deriving working objectives for intractable problems — the field's most reused trick
  • Explaining why contrastive pretraining produces useful geometry

Level 3 — The frontier

geometry, bottlenecks, and alignment
Fisher information & natural gradient  ·  information bottleneck  ·  KL-regularized RL (RLHF, DPO)

The ideas

  • Fisher information: the local metric on the space of distributions — natural gradient follows it, not raw parameters
  • Information bottleneck: representation = compress the input, keep what predicts the target
  • Alignment as KL-regularized RL: maximize reward while staying close to the base model
  • DPO: solve that objective in closed form — no reward model, no RL loop

Where it lives

  • Natural-gradient lineage: K-FAC, TRPO's KL constraint
  • The RLHF pipeline of every aligned chat model
  • The DPO paper — a masterclass of this territory

What it unlocks

  • Deriving alignment methods rather than memorizing them
  • Reading "the KL term" in any paper and knowing which behavior it buys
  • A principled account of what representations should keep and discard

Cheat sheet: concept → landmark → what it's doing there

ConceptLandmark useWhat it's doing there
Cross-entropy as codingLLM pretrainingPrediction = compression; loss = bits per token
ELBOVAE; diffusion training objectivesA trainable bound when the likelihood is intractable
InfoNCECLIP, contrastive embeddersA trainable bound on mutual information
KL directionVAEs, distillation, variational inferenceMode-covering vs mode-seeking behavior, by choice
Fisher / natural gradientK-FAC, TRPOSteps measured in distribution space, not parameter space
KL-regularized RLRLHF; DPO's closed formImprove on preferences without leaving the base model
Information bottleneckRepresentation-learning theoryWhat a good representation keeps, formally
The territory's signature move: when the true objective is intractable, find a bound you can train on — ELBO, InfoNCE and DPO are the same maneuver executed three times, a decade apart.
End of the map. Five territories, one loop back: the architect chooses what to build, the engineer trains it, and this track supplies the mathematics both of them are standing on. Back to the frontier overview · all field guides.
← Territory 4: Optimization & GamesConvergence & transport