Published November 10, 2025 | Version v2

Universal Interior Optimality for Two-Input Jensen-Shannon Divergence Channels

Authors/Creators

Description

We establish through exhaustive computational verification that Jensen-Shannon 
divergence (JSD) contraction coefficients for two-input channels universally 
achieve their maximum in the interior of the probability simplex, never at 
vertices. This challenges the implicit assumption that vertex optimality 
(proven for total variation distance) extends to other divergence measures.

Key Results:
• Universal interior optimality: 100% of two-input JSD channels 
  (115,000/115,000 tested, including 100,000 for minimal 2×2 case)
• Fast detection criterion: O(n²) curvature parameter L with perfect 
  prediction accuracy (200/200 channels)
• Quantitative improvement: 20-35% over vertex bounds (mean 31%, median 34%)
• Minimal counterexample: Even 2×2 channels show interior optimality
• Classical channels: BSC, BEC, Z-channel all confirmed

Practical Implications:
• Differential privacy: Vertex-based bounds may underestimate leakage by 20-35%
• Quantum cryptography: QKD security bounds require re-evaluation
• Machine learning: Information bottleneck theory may need revision
• Channel capacity: Optimal signaling strategies may require mixed inputs

Complete package includes: analytical proofs, 115K experimental validation, 
benchmark test suite, and full source code for independent verification.

Files

benchmark_results.txt

Files (479.4 kB)

Name Size
md5:0f9f736d7b2e1775786c8525bffa45f8
5.5 kB Preview Download
md5:57d3446269e8ad90554d2ffea048ddaa
18.2 kB Download
md5:6de3df378c0deb77da5063640c77d93c
7.4 kB Preview Download
md5:ae4e71228e056843e02b5e504faca3a6
79.3 kB Preview Download
md5:6a0d6589808a58c8b0e7f02cd8f62f3f
83.5 kB Preview Download
md5:0bf907c48b91691081348d61d9e5ea26
215.1 kB Preview Download
md5:67ae982e837afb6c3439fd2eaae4b177
8.3 kB Preview Download
md5:92d539e9317bb782e2fc935ea27fb0b6
3.7 kB Preview Download
md5:b02e612a112321240e9ad3efcc17c239
30.0 kB Preview Download
md5:7f70dc2df878fbd2edf8091187bd8458
3.0 kB Preview Download
md5:8af38be3098ae57e5a7051ee6140163a
25.2 kB Preview Download