Discrepancy Minimization by Walking
on the EdgesRaghu Meka (IAS/DIMACS)
Shachar Lovett (IAS)
1 2 3 4 5
Discrepancy• Subsets • Color with or - to minimize imbalance
1 * 1 1 ** 1 1 * 1
1 1 1 1 1
* * * 1 1
1 * 1 * 1
1 2 3 4 51 * 1 1 ** 1 1 * 1
1 1 1 1 1
* * * 1 1
1 * 1 * 1
3
1
1
0
1
Discrepancy Examples• Fundamental combinatorial
conceptArithmetic Progressions
Roth 64: Matousek, Spencer 96: {1,3,5 ,⋯ }, {1,4,7 ,⋯ },⋯
Discrepancy Examples• Fundamental combinatorial
conceptHalfspaces
Alexander 90: Matousek 95:
Discrepancy Examples• Fundamental combinatorial
conceptAxis-aligned boxes
Beck 81: Srinivasan 97:
Why Discrepancy?Complexity theory
Communication Complexity
Computational Geometry
PseudorandomnessMany more!
Spencer’s Six Sigma Theorem
• Central result in discrepancy theory.
• Beats random:• Tight: Hadamard.
Spencer 85: System with n sets has discrepancy at most .
“Six standard deviations suffice”
Conjecture (Alon, Spencer): No efficient algorithm can find one.
Bansal 10: Can efficiently get discrepancy .
A Conjecture and a Disproof
• Non-constructive pigeon-hole proof
Spencer 85: System with n sets has discrepancy at most .
This Work
• Truly constructive• Algorithmic partial coloring lemma• Extends to other settings
Main: Can efficiently find a coloring with discrepancy
New elemantary constructive proof of Spencer’s result
EDGE-WALK: New algorithmic tool
Outline1. Partial coloring Method
2. EDGE-WALK: Geometric picture
• Focus on m = n case.Lemma: Can do this in randomized
time.
Partial Coloring MethodInput:
Output:
Outline1. Partial coloring Method
2. EDGE-WALK: Geometric picture
1 * 1 1 ** 1 1 * 11 1 1 1 1* * * 1 11 * 1 * 1
Discrepancy: Geometric View• Subsets
• Color with or - to minimize imbalance
1-111-1
3
1101
31101
1 2 3 4 5
1 * 1 1 ** 1 1 * 11 1 1 1 1* * * 1 11 * 1 * 1
Discrepancy: Geometric View
1-111-1
31101
1 2 3 4 5
• Vectors • Want
Discrepancy: Geometric View• Vectors
• Want
Goal: Find non-zero lattice points in
Polytope view used earlier by Gluskin’ 88.
Claim: Will find good partial coloring.
Edge-Walk
• Start at origin• Gaussian walk
until you hit a face• Gaussian walk
within the face
Goal: Find non-zero lattice point in
Edge-Walk: AlgorithmGaussian random walk in subspaces
• Subspace V, rate • Gaussian walk in V
Standard normal in V:Orthonormal basis
change
Edge-Walk AlgorithmDiscretization issues: hitting faces
• Might not hit face• Slack: face hit if
close to it.
1. For
2. Cube faces nearly hit by .
Disc. faces nearly hit by .
Subspace orthongal to
Edge-Walk: Algorithm• Input: Vectors • Parameters:
Pr [𝑊𝑎𝑙𝑘h𝑖𝑡𝑠𝑎𝑑𝑖𝑠𝑐 . 𝑓𝑎𝑐𝑒 ]≪ Pr [𝑊𝑎𝑙𝑘 h𝑖𝑡𝑠𝑎𝑐𝑢𝑏𝑒′ 𝑠 ]
Edge-Walk: Intuition
1100 Hit cube more often!
Discrepancy faces much farther than cube’s
Summary
1. Edge-Walk: Algorithmic partial coloring lemma
2. Recurse on unfixed variables
Spencer’s Theorem
Open Problems
Q: Other applications?General IP’s, Minkowski’s theorem?
• Some promise: our PCL “stronger” than Beck’s
Q: Beck-Fiala Conjecture 81: Discrepancy for degree t.
Thank you
Main Partial Coloring Lemma
Algorithmic partial coloring lemmaTh: Given thresholds
Can find with 1. 2.