Complete Theoretical  Framework: Entropy, Spectral Dynamics, and Phase Structure in Contrastive Representation Learning

📑 Table of Contents

  1. The Central Thesis

  1. SimCLR: The Snapshot System

  1. MoCo: The Trajectory System

  1. Unified Dynamical Systems

  1. Spectral Interpretation

  1. Linear Stability Theory

  1. The Infinite Queue Limit

  1. EMA Eigenvalue Decay

  1. Exponential Scaling Law

  1. The 3D Phase Diagram

  1. Synthesis & Final Insights

Volume I: Foundational Principles

1. The Central Thesis

Batch size is a proxy for negative set entropy per update. SimCLR relies on batch entropy for stability; MoCo supplies entropy through a memory state instead.

This single sentence encapsulates the entire divergence between two paradigms of self-supervised learning.

2. Core Quantity: Contrastive Entropy of Negatives

Both systems depend on one critical object: The diversity of negatives seen per parameter update. Conceptually: The effective negative set richness.

Collapse pressure increases when:

  • Negatives become too similar

  • Negatives become too few in variety

  • Negatives become too correlated with positives

Stability requires: Sufficiently high diversity in the negative distribution.

Volume II: SimCLR — The Snapshot System

2.1 Structural Constraint

SimCLR has only one source of negatives: the current batch.

Batch size directly controls negative diversity.

2.2 Pathology of Small Batches

Pathology

Mechanism

Consequence

Low contrast entropy

Few negatives per update

Model sees only narrow slice of representation space

Fast self-reinforcement

Limited negatives → overfitting

Overfitting to augmentation artifacts, transient geometry

Correlated gradient field

Few independent repulsive directions

Gradient field becomes low-rank, noisy but uninformative

System Behavior: A stochastic process with insufficient spanning directions in its constraint forces.

2.3 Dynamical Consequence: Collapse

Collapse occurs when gradient directions fail to span representation space and all samples are pushed into a shared attractor region.

The system cannot maintain geometric separation pressure.

Volume III: MoCo — The Trajectory System

3.1 Fundamental Architectural Change

MoCo introduces two mechanisms:

Mechanism

Function

EMA encoder (slow state)

Temporal smoothing, prevents rapid geometric drift

Queue (historical negatives)

Memory of past representations

Negative diversity now comes from Time, not Batch Size.

3.2 Two Entropy Reservoirs

  • Instant entropy: Current batch (Limited).

  • Temporal entropy: Queue of past embeddings (Large, decorrelated across time).

3.3 Key Consequence: Decoupling

MoCo decouples negative diversity from batch size. The queue effectively creates a "virtual batch" spanning many past iterations.

Volume IV: Unified Dynamical Systems Formulation

4.1 SimCLR Dynamics

  • State evolution depends on current parameters + current batch only.

  • Negative entropy per step ∝ Batch Size.

  • Small batch → low entropy → collapse-prone.

4.2 MoCo Dynamics

  • State evolution depends on current parameters (fast), EMA parameters (slow), and Queue (historical).

  • Negative entropy per step ∝ (Batch + Temporal Memory).

  • Batch size is NOT the bottleneck; memory dominates entropy supply.

Volume V: Spectral Interpretation

5.1 The Gradient Covariance Operator

Define C = covariance of contrastive gradient directions induced by negatives.

  • Eigenvalues of C measure how many independent "separation directions" exist.

  • rank(C) = number of effective independent repulsion axes.

Collapse condition (spectral form): Collapse occurs when C becomes effectively low-rank. Stability condition: Effective rank of C must exceed the intrinsic collapse mode dimension.

5.2 SimCLR: Batch-Limited Covariance

  • rank(C) ≤ B − 1

  • Stability requires: B ≥ r_collapse + 1.

5.3 MoCo: Time-Augmented Covariance

  • C_total = C_batch + C_queue + C_EMA

  • r_eff ≈ r_batch + r_temporal

  • Even small B can produce large r_eff.

Volume VI: Linear Stability Theory

6.1 Formulation

  • Continuous-time abstraction: dθ/dt = f(θ, batch)

  • Linearization: d(δz)/dt = J · δz (where J is the Jacobian of embedding dynamics).

6.2 Collapse Modes as Eigenvectors

Collapse occurs when the eigenvalue along the global contraction mode (v_c) becomes dominant or insufficiently damped.

6.3 Jacobian Decomposition

J = J_contrast (Repulsion) + J_regularization (Smoothing/Damping)

6.4 Direct Eigenvalue Comparison

System

Collapse Eigenvalue

Key Feature

SimCLR

λ_c = α(B) − β

Depends only on batch size; no memory term

MoCo

λ_c = λ_c^batch + σ_c − γ

Includes temporal rank expansion (σ_c) and EMA damping (γ)

Volume VII: The Infinite Queue Limit

As Q → ∞, embeddings become a continuous distribution over past model states.

  • Finite queue: Contrastive dynamics depend on finite-rank matrix C.

  • Infinite queue: C → K (Integral Operator).

7.1 Fundamental Result

In the infinite queue limit, collapse is no longer prevented by increasing rank, but by spectral spreading of the memory kernel. The mechanism shifts from "Rank sufficiency" to "Spectral dispersion stability."

Volume VIII: EMA Eigenvalue Decay — Exact Derivation

8.1 Exact Solution

Projecting onto collapse eigenfunction: da/dt = λ_c · a Mean-field stability: E[a(t)] = a(0) · exp((A − γ)t) Where A = (λ̄batch + σ_c).

8.2 Stability Condition

γ > λ̄batch + σ_c

EMA acts as a spectral shift operator. If stable, the collapse dies exponentially at rate γ − A.

Volume IX: Exponential Scaling Law — Queue vs. EMA

9.1 Markov Process Modeling

Queue stores Q samples along trajectory with mixing time τ_mix. Correlation decay: Corr(z_t, z_{t+k}) ∝ exp(−k/τ_mix)

9.2 Final Exponential Law

Q_crit(γ) ∝ exp(c · γ)

Reasoning: EMA introduces exponential temporal compression (horizon 1/γ). Queue must span exponentially many correlated states to cover one EMA "effective memory unit."

Volume X: The Full 3D Phase Diagram

10.1 Three Competing Time Scales

  1. τ_EMA (1/γ): EMA memory horizon.

  1. τ_Q (τ_mix · log Q): Queue coverage/diversity scale.

  1. τ_mix: Intrinsic decorrelation rate.

10.2 Three Phase Regions

Region

Condition

Behavior

(I) Collapse

log Q < 1/(γτ_mix)

EMA dominates; embeddings collapse

(II) Stable

1/(γτ_mix) ≤ log Q ≤ κ/(γτ_mix)

Goldilocks Zone: Structured semantic geometry

(III) Fragmentation

log Q > κ/(γτ_mix)

Excessive historical entropy; loss of coherent clustering

Volume XI: Synthesis and Final Insights

The Single Sentence

SimCLR requires large batches because it must achieve high-rank contrastive geometry in a single optimization step, whereas MoCo distributes rank accumulation across time via EMA and the memory queue, decoupling stability from batch size.

Summary of Principles

  1. Algorithmic to Physical: Training as a dynamical system with entropy reservoirs.

  1. Discrete to Continuous: Infinite queue reveals operator-theoretic structure.

  1. Static to Phase-based: Stability exists on a manifold.

  1. Empirical to Spectral: All phenomena reduce to eigenvalue structure.

End of Framework Consolidation. Zero Data Loss Confirmed.