Deterministic Reweighting vs IPF

Core Difference in Approach

Fundamental distinction: Deterministic reweighting adjusts weights in a single pass per dimension, while IPF iteratively refines weights until all marginal constraints are simultaneously satisfied.

Deterministic Reweighting

Approach: Sequential 1D adjustment

# Pseudocode
for each dimension:
    for each record:
        weight *= (target_marginal / current_marginal)
  • Processes one variable at a time
  • Final weights are product of single-dimension adjustments
  • No convergence checking

Iterative Proportional Fitting (IPF)

Approach: Cyclic multi-dimensional convergence

# Pseudocode
while not converged:
    for each dimension:
        adjust weights proportionally
        check all marginal differences
  • Cycles through all dimensions repeatedly
  • Adjusts weights to gradually satisfy all constraints
  • Includes convergence threshold