beyond set disjointness: the communication complexity of finding the intersection grigory...

Beyond Set Disjointness: The Communication Complexity of Finding the

Intersection

Grigory Yaroslavtsevhttp://grigory.us

Joint with Brody, Chakrabarti, Kondapally and Woodruff

Communication Complexity [Yao’79]

Alice: Bob:

𝒇 (𝒙 ,𝒚 )=?

Shared randomness

…

𝒇 (𝒙 ,𝒚 )• = min. communication (error ) • min. -round communication (error )

Set Intersection

𝒙=𝑺 ,𝒚=𝑻 , 𝒇 (𝒙 , 𝒚 )=𝑺∩𝑻𝑺⊆ [𝑛 ] ,|𝑆|≤𝒌 𝑻 ⊆ [𝑛 ] ,|𝑇|≤𝒌 = ?

(-Intersection) = ?

is big, n is huge, where huge big

Our results

Let

• (-Intersection) = [Brody, Chakrabarti, Kondapally, Woodruff, Y.; PODC’14]• (-Intersection) = [Saglam-Tardos FOCS’13; Brody, Chakrabarti, Kondapally, Woodruff, Y.’; RANDOM’14]

{

times

(-Intersection) = for

Applications

• Exact Jaccard index (for -approximate use MinHash [Broder’98; Li-Konig’11; Path-Strokel-Woodruff’14])• Rarity, distinct elements, joins,…• Multi-party set intersection (later)• Contrast:

1-round -protocol

𝒉 : [𝒏 ]→[𝒌3]

𝑺 𝑻

𝒉(𝑺) 𝒉(𝑻 )

[𝒏 ] [𝒏 ]

[𝒌3] [𝒌3]

Hashing

log 𝒌

=# of buckets

𝒉 : [𝒏 ]→[𝒌 / log𝒌]

Expected # of elements

Secondary Hashing

= # of hash functions

log 3𝒌 where

2-Round -protocol

log 3𝒌

log 3𝒌

|h𝑖 (𝑺 )|,|h𝑖 (𝑻 )|=𝑂 ( log𝒌 log log𝒌 )

Total communication = = O()

Collisions

𝒌log𝒌

log 3𝒌Pr [𝑐𝑜𝑙𝑙𝑖𝑠𝑖𝑜𝑛 ]=𝑂( 1log𝒌 )

Collisions

log 3𝒌

log 3𝒌

Key fact: If then also =

Collisions

• Second round: – For each bucket send -bit equality check (total -

communication)– Correct intersection computed in buckets where

– Expected # items in incorrect buckets – Use 1-round protocol for incorrect buckets– Total communication

Main protocol

𝑂 (1)

=# of buckets

𝒉 : [𝒏 ]→[𝒌]

Expected # of elements

Verification tree -degree

…i log𝑟 −1𝒌

buckets = leaves of the verification tree

Verification bottom-up

𝑺𝟏❑ ,𝐓𝟏

❑ 𝑺𝟐❑ ,𝐓𝟐

❑

𝑺𝟏❑∪𝑺𝟐 ,𝐓𝟏

❑∪𝑻 𝟐

𝑺𝟏❑∩𝐓𝟏

❑𝑺𝟐❑∩𝐓𝟐

❑

(𝑺𝟏❑∪𝑺𝟐 )∩(𝐓 ¿¿𝟏❑∪𝑻 𝟐)¿

EQUALITY CHECK

Verification bottom-up

𝑺𝟏❑∩𝐓𝟏

❑𝑺𝟐❑∩𝐓𝟐

❑

(𝑺𝟏❑∪𝑺𝟐 )∩(𝐓 ¿¿𝟏❑∪𝑻 𝟐)¿

Correct Incorrect

Incorrect

𝑺𝟏❑∩𝐓𝟏

❑𝑺𝟐❑∩𝐓𝟐

❑

(𝑺𝟏❑∪𝑺𝟐 )∩(𝐓 ¿¿𝟏❑∪𝑻 𝟐)¿

Correct Incorrect

Correct

Verification bottom-up

𝑺𝟏❑∩𝐓𝟏

❑𝑺𝟐❑∩𝐓𝟐

❑

(𝑺𝟏❑∪𝑺𝟐 )∩(𝐓 ¿¿𝟏❑∪𝑻 𝟐)¿

Correct Incorrect

EQUALITY CHECK FAILS =>RESTART THE SUBTREE

𝑺𝟏❑∩𝐓𝟏

❑𝑺𝟐❑∩𝐓𝟐

❑

(𝑺𝟏❑∪𝑺𝟐 )∩(𝐓 ¿¿𝟏❑∪𝑻 𝟐)¿

Correct Incorrect

Correct

Verification bottom-up

𝒑𝒓 −𝟐

…𝒑𝟏

𝑺𝟏𝟏 ,𝐓𝟏

𝟏 … 𝑺𝒊𝟏 ,𝐓 𝐢

𝟏𝑺𝟐𝟏 ,𝐓𝟐

𝟏 𝑺𝒌𝟏 ,𝐓𝒌

𝟏…

𝒑𝒓 −𝟏

Analysis of Stage

• = [node at stage computed correctly]• Set = – Run equality checks and basic intersection

protocols with success probability – Key lemma: [# of restarts per leaf => Cost of

Intersection in leafs = – Cost of Equality =

• [protocol succeeds] =

Multi-party extensions

players: , where

• Boost error probability of 2-player protocol to • Average per player (using coordinator):

in rounds• Worst-case per player (using a tournament)

in rounds

Open Problems

• (-Intersection) = ?• Better protocols for the multi-party setting?

-Disjointness• , iff • [Razborov’92; Hastad-Wigderson’96] • [Folklore + Dasgupta, Kumar, Sivakumar; Buhrman’12, Garcia-Soriano, Matsliah, De Wolf’12]

• [Saglam, Tardos’13]• [Braverman, Garg, Pankratov, Weinstein’13]