Download presentation
Presentation is loading. Please wait.
Published byΦοίβος Μοσχοβάκης Modified over 6 years ago
1
Sketching and Embedding are Equivalent for Norms
Alex Andoni, Columbia Robert Krauthgamer, Weizmann Institute Ilya Razenshteyn, MIT HALG 2016 in Paris [paper in STOC 2015, arXiv: ] TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: AAA
2
Sketching and Embedding are Equivalent for Norms
“Compressing” large object to short summary Example: dimension reduction JL [Johnson-Lindenstrauss,1984]: approximating ℓ 2 -distances PCA (truncated SVD): improve signal-to-noise, easier to visualize high-dimensional vectors, matrices, graphs, … lossy, functional n d When is sketching possible? Sketching and Embedding are Equivalent for Norms
3
Sketching and Embedding are Equivalent for Norms
Sketching Metrics 1 … Fix a metric space 𝑀 (e.g. Euclidean) Alice and Bob hold points 𝑥,𝑦∈𝑀 Each sends 𝑠-bit sketch to Charlie Who should distinguish If 𝑑 𝑀 (𝑥, 𝑦)≤𝑟 or 𝑑 𝑀 𝑥, 𝑦 >𝐷𝑟 (for threshold 𝑟>0, approx. 𝐷>1) Shared randomness, error probability ≤1/3 Alice Bob Charlie 𝑥 𝑦 sketch(𝑥) sketch(𝑦) Output: “close” or “far” Expect tradeoff between sketch size s and distortion D? Sketching and Embedding are Equivalent for Norms
4
Sketching and Embedding are Equivalent for Norms
Why Sketch Metrics? Useful algorithmic building block Near-Neighbor Search algorithm with approximation 𝐷 and storage 𝑛 𝑂(𝑠) Fast estimation of distances (filtering) Linear sketches are powerful for streaming (and sparse recovery) To identify current bottlenecks Some algorithms for 𝑙 1 and 𝑙 2 use this approach (implicitly) Based on information theory Less sensitive to representation and transformations In general, to measure the “complexity” of the metric Sketching and Embedding are Equivalent for Norms
5
Example: Sketching for ℝ
Distinguish |𝑥–𝑦|≤1 vs. |𝑥–𝑦|>1+𝜀 Break line into pieces of width 𝑤=1+𝜀 w/random shift and color Red/Blue at random Pr[color(𝑥)=color(𝑦)] ? if far: 1/2 if close: 𝜖+(1−𝜖)⋅1/2=1/2+𝜖/2 Repeat 𝑂(1/ 𝜀 2 ) times Take (1/2+𝜖/4)-th percentile Overall: 𝐷=1+𝜀 and size 𝑠=𝑂(1/ 𝜀 2 ) B R x y Sketching and Embedding are Equivalent for Norms
6
Sketching and Embedding are Equivalent for Norms
Sketching ℓ 𝑝 norms In Hamming space ( ℓ 1 ) Sample bits randomly + hashing [Indyk-Motwani, Kushilevitz-Ostrovsky-Rabani’98] ℓ 2 : Reduces it to the real line case via dimension reduction (JL lemma) For Gaussian 𝑔 , then 𝑔 𝑇 𝑥− 𝑔 𝑇 𝑦 is distributed as ||𝑥−𝑦| | 2 times N(0,1) ℓ 𝑝 for 0<𝑝≤2: Projection using 𝑝-stable distributions [Indyk’00] Achieves 𝐷=1+𝜀 using 𝑠=𝑂(1/𝜀2) tight [Woodruff’04] ℓ 𝑝 for 𝑝>2: sketching is harder [BarYossef-Jayram-Kumar-Sivakumar’02, Indyk-Woodruff’05] Achieving 𝐷=𝑂(1) requires sketch size 𝑠= Θ 𝑑 1−2/𝑝 Other metrics/norms beyond ℓ 𝒑 ? Sketching and Embedding are Equivalent for Norms
7
Reductions between Geometries
An embedding is a map 𝑓:𝑀→𝑁 of metric 𝑀 into 𝑁 It has distortion 𝐶>0 if ∀𝑥,𝑦∈𝑀, 1≤ 𝑑 𝑁 𝑓 𝑥 ,𝑓 𝑦 𝑑 𝑀 𝑥,𝑦 ≤𝐶 𝑀 𝑁 𝑓(𝑥) 𝑓(𝑦) 𝑓 𝑥 𝑦 Sketching of size s and approximation CD for M Sketching of size s and approximation D for N Sketching and Embedding are Equivalent for Norms
8
Goal: Efficient Sketching
Efficient = constant sketch size 𝑠 and approximation 𝐷 Classification? Known: Metrics 𝑀 that admit efficient sketching are 𝑀 is ℓ 𝑝 for 𝑝≤2, and 𝑀 embeds into ℓ 𝑝 for 𝑝≤2 with distortion 𝑂(1). Other metrics with efficient sketches? Essentially NO! Sketching and Embedding are Equivalent for Norms
9
Efficient Sketching vs. Embedding
Our Theorem: Every normed space 𝑋 with sketches of size 𝑠 and approximation 𝐷, embeds into ℓ 1−𝜖 with distortion 𝑂(𝑠𝐷/𝜀) (for every 0<𝜀<1) Normed space: ℝ𝑑 equipped with “length” ⋅ 𝑋 Examples: ℓ 𝑝 for 𝑝≥1, matrix norms, Earthmover distance embedding into ℓ 𝑝 , 𝑝≤2 [Kushilevitz-Ostrosvksy-Rabani’98] [Indyk’00] for norms efficient sketching Sketching and Embedding are Equivalent for Norms
10
Application: Sketching Lower Bounds
Non-embeddability implies lower bounds for sketches In a black-box manner Yields new results No embedding with distortion O(1) into ℓ 1−𝜖 No sketches* of size and approximation O(1) *in fact, no communication protocols (any number of rounds) Sketching and Embedding are Equivalent for Norms
11
Example 1: Earth-Mover’s Distance
For 𝑥∈ ℝ 𝑛×𝑛 that sums to zero, ‖𝑥‖ 𝐸𝑀𝐷 is the minimum cost of moving positive part of 𝑥 to the negative part Upper bounds: 𝐷-approximation with space 𝑠= 𝑛 𝑂(1/𝐷) [Charikar’02, Indyk-Thaper’03, Naor-Schechtman’05, Andoni-DoBa-Indyk-Woodruff’09] Lower bound extends to the min-cost matching on finite subsets in 𝑛 ×[𝑛] No embedding with distortion 𝑂(1) into ℓ 1−𝜖 [Naor-Schechtman’05] No sketches with 𝑠=𝑂(1) and 𝐷=𝑂(1) Sketching and Embedding are Equivalent for Norms
12
Sketching and Embedding are Equivalent for Norms
Example 2: Trace Norm For matrix 𝐴∈ ℝ 𝑛×𝑛 , the trace norm ‖𝐴‖ is the sum of the singular values aka nuclear norm or Schatten-1 norm Previous lower bounds: Only for restricted sketching algorithms [Li-Nguyen-Woodruff’14] Recently, 𝐷=1+𝜖 requires 𝑠≈𝑛 [Li-Woodruff’16] Embedding into ℓ 1−𝜖 requires distortion 𝑂( 𝑛 ) [Pisier’78] Sketching requires 𝑠𝐷=Ω( 𝑛 / log 𝑛 ) Sketching and Embedding are Equivalent for Norms
13
Sketching and Embedding are Equivalent for Norms
Proof Outline Linear embedding of 𝑋 into ℓ 1−𝜖 Fourier analysis [Aharoni-Maurey-Mityagin’85] Good sketches for 𝑋 Good sketches for ℓ ∞ 𝑘(𝑋) Uses that 𝑋 is a norm 𝐿 || 𝑥 1 − 𝑥 2 | | 𝑋 ≤||𝑔 𝑥 1 −𝑔 𝑥 2 ||≤𝑈(|| 𝑥 1 − 𝑥 2 | | 𝑋 ) 𝐿 and 𝑈 are non-decreasing, 𝐿(𝑡)>0 for 𝑡>0 𝑈(𝑡)→0 as 𝑡→0 || 𝑥1, …, 𝑥𝑘 | | ∞ = max i || 𝑥 𝑖 | | 𝑋 Uniform embedding 𝑔:𝑋→ ℓ 2 Lipschitz extension [Johnson-Randrianarivony’06] Absence of certain Poincaré-type inequalities on 𝑋 Direct sum for Information Complexity [Andoni-Jayram-Pătraşcu’10] || 𝑥 1 − 𝑥 2 | | 𝑋 ≤1 ⇒ ||𝑓 𝑥 1 −𝑓 𝑥 2 ||≤1 || 𝑥 1 − 𝑥 2 | | 𝑋 ≥𝑠𝐷⇒||𝑓 𝑥 1 −𝑓 𝑥 2 ||≥10 Weak embedding 𝑓:𝑋→ ℓ 2 Convex duality + compactness Sketching and Embedding are Equivalent for Norms
14
Sketching and Embedding are Equivalent for Norms
Almost a Shortcut Good sketches for 𝑋 Use [Andoni-K.’07] to get 1-bit sketch with “advantage” 2 −𝑠 , i.e. random 𝑓′:𝑋→{0,1} s.t. || 𝑥 1 − 𝑥 2 | | 𝑋 ≤1 ⇒𝔼 |𝑓′ 𝑥 1 −𝑓′ 𝑥 2 |≤1/2− 2 −𝑠 || 𝑥 1 − 𝑥 2 | | 𝑋 ≥𝐷 ⇒𝔼 |𝑓′ 𝑥 1 −𝑓′ 𝑥 2 |≥1/2+ 2 −𝑠 Define 𝑓:𝑋→ ℓ 1 by “enumerating” randomness || 𝑥 1 − 𝑥 2 | | 𝑋 ≤1 ⇒ ||𝑓 𝑥 1 −𝑓 𝑥 2 ||≤1 || 𝑥 1 − 𝑥 2 | | 𝑋 ≥𝑠𝐷⇒||𝑓 𝑥 1 −𝑓 𝑥 2 ||≥10 Weak embedding 𝑓:𝑋→ ℓ 2 Sketching and Embedding are Equivalent for Norms
15
Sketching and Embedding are Equivalent for Norms
Further Questions Extension to ℓ 1 ? Do sketches with 𝑠,𝐷=𝑂(1) imply embedding into ℓ 1 with distortion 𝑂(1)? Equivalent to an old open problem in Functional Analysis [Kwapien’69] Extension to all metrics? Here, the ℓ 1 version is false: Heisenberg group [Lee-Naor’06, Cheeger-Kleiner’10, Cheeger-Kleiner-Naor’11] Tradeoff between 𝑠,𝐷? Large approximation? 1+𝜖 approximation? Sketches → Linear sketches? “Complexity” of metric spaces? Thank You! Sketching and Embedding are Equivalent for Norms
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.