Paul Beame University of Washington Outline Branching program basics Space size lower bounds Multioutput functions TimeSpace tradeoff lower bounds for general BPs Singleoutput functions ID: 637748
Download Presentation The PPT/PDF document "Branching Programs Part 3" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Branching ProgramsPart 3
Paul
Beame
University of WashingtonSlide2
Outline
Branching program basics
Space (size) lower bounds
Multi-output functions
Time-Space tradeoff lower bounds for general BPs
Single-output functions
Restricted classes of BPs
OBDDs, Read-once (FBDDs), Oblivious,
Read-k
Lower bound methods for restricted classes
Lower bound methods for general BPs
Applying tradeoffs: BPs and static data structures
Multi-output
functions using single-output techniques
Lower bound for encoding good codesSlide3
Limited Branching Program Forms
Structure-based
Oblivious
For each BP level, all the nodes on that level have the same variable name
e.g. Parity BP
Time-Based
Read-OnceOn every path through the BP each variable is queried at most oncee.g. Parity BPRead-πOn every path through the BP each variable is queried at most π timesTime-BoundedEvery path in the BP has length at most
Β Slide4
Oblivious vs Read-k and Best Bounds
Recall argument for oblivious BPs:
If
length
then
variables are read at most
timesSo oblivious length ππ β Read-
2πExponential Read-
π
size lower bounds for simple explicit Boolean functions for π=π(log π)Inspired by 2-party communication complexity [Borodin-Razborov-Smolensky 1989][Okolβnishnikova 1989]Exponential size lower bound for an explicit function over large domain for π=π(log2 π)Inspired by multiparty communication complexity [B-Vee 2002]Drawback: Function is not known to be in NP. No larger π possible until the oblivious case is improved
Β Slide5
Read-k BPs
On every path through the BP each variable is queried at most
π
times
Unlike read-once, there may be paths that are not consistent with any input. We assume that those paths are restricted, too.Lower bound methods for read-
π BPs often also apply to nondeterministic read-π BPsthe βevery pathβ constraint is essential thereDefn: Nondeterministic BPs (NBPs) generalize BPs by allowing many out-edges from a vertex with the same label:An NBP outputs 1 on input π iff there is some path that π can take that leads to the 1-sink node.Slide6
Read-k BPs
We can also separate the levels of the Read-
π
BP hierarchy:
e.g., a small Read-
2 BP can tell whether an
π x π binary matrix is a permutation matrix, unlike small Read-Once BPsThere are explicit Boolean functions with small Read-π+1 BPs that require exponential size Read-π BPs for π β€
/
2
, even allowing nondeterminism or randomization
[Jayram S. Thathachar 1998] Techniques are easier but are similar enough to those for general time-bounded BPs in the case of non-Boolean inputs that we do them togetherΒ Slide7
Limited Branching Program Forms
Structure-based
Oblivious
For each BP level, all the nodes on that level have the same variable name
e.g. Parity BP
Time-Based
Read-OnceOn every path through the BP each variable is queried at most oncee.g. Parity BPRead-πOn every path through the BP each variable is queried at most π timesTime-BoundedEvery path in the BP has length at most
Β Slide8
Breaking up a BP via its Traces
Split BP
π·
with input set
π«
π
into layersLet traces(π·)={trace(π)| πβπ«π}
For π
β
traces(
π·) let ππ be the function that has value 1 on input π iff trace(π)=π and π(π)=1Since π· computes π π=
The
π
π
are disjoint and
|
traces
(π·)|
β€
2
πΊ
Can extend this to nondeterministic BPs:
Each
π
may have multiple tracesThe ππ are no longer disjoint
Β
1
0Slide9
Read-k BPs and Traces
Split BP
π·
with input
set
π«
π into layers π=
and
|
traces
(π·)| β€ 2πΊIf π· is a read-π BP then w.l.o.g. for every pair of nodes π, π in π· the same set of variables is read on every path from π to πOnly must avoid variables read π times on some pair of paths above π and below
πSo, each trace
π
yields a
fixed sequence
of
π³
sets of variables read, each of size
β€ ππ/π³
Can assign the layers for each
π
π
as we did for oblivious BPs
Get 2πΊ assignments, one for each πΒ
1
0Slide10
Recall: Strategy for Assigning Layers
Assign each of
π³
layers to either Alice or Bob for
Goal:
maximize # of bits per player π, while minimizing π³.Flip an independent coin for each layer: π=π/2π+1, π³
=8π
2
2π equal length layers [Borodin-Razborov-Smolensky 1989, B-Jayram-Saks 2001]Β Slide11
Recall: Strategy for Assigning Layers
Assign each of
π³
layers to either Alice or Bob for
Goal:
maximize # of bits per player π, while minimizing π³.Flip an independent coin for each layer: π=π/2π+1, π³
=8π
2
2π equal length layers [Borodin-Razborov-Smolensky 1989, B-Jayram-Saks 2001]Use 4π2 equal layers. Give a random subset of 2π of them to Bob. π=π/(2ππ)2π, π³=4π2 [Okolβnishnikova 1989,
Ajtai 1999]
Β Slide12
Alice
Input
Β
Bob
Input
Β
β¦ β¦
010
Β
11
0
= # of bits Alice and Bob need to
exchange
to compute
on
Β
Communication ComplexitySlide13
Communication Complexity
Defn
:
A
(combinatorial) rectangle in πΏππ is a subset
πΌππ½ where πΌβπΏ and π½βπ.Lemma: Any deterministic π-bit protocol for π:πΏππβ{0,
1}
yields a
partition of πΏππ into 2π rectangles on which π is constant Lemma: Nondeterministic π-bit protocols correspond to coverings of π-1(1) by 2π rectangles.To prove that π requires large (non)deterministic communication complexity it suffices to prove π-1
(1
)
is large
Any rectangle in
π
-
1
(
1
)
is
small
πΏ
π
πΌ
π½Slide14
Best Partition and Fixed Variables
For BP lower bounds, we donβt know
a priori
how the input
{
0,
1}π or [π]π to π is partitioned into πΏππ Need to analyze rectangles for all possible partitions of [π]βbest partitionβ communication complexityAlso, in the case of oblivious BPs, most input variables (ones seen by both parties) were fixed ahead of time
Implicitly, rectangle size was only important relative to the space of unfixed variables.For Read-
π
and general time-bounded BPs this aspect is even more important so we need to make it explicit.Slide15
Embedded Rectangles
Defn
:
For disjoint sets
π¨
, π© β [π] and
πΆ β π«[π]-π¨-π©, the embedded rectangle πΉ β π«π with footprint (π¨,π©), tail πΆ,
and body
πΉ
π¨
ππΉπ© for πΉπ¨βπ«π¨, πΉπ©βπ«π© is the set πΉ={πβπ«π | ππ¨βπΉπ¨, ππ©βπΉπ©, π
[π]-π¨-π© = πΆ}
.
Defn
:
The
density
πΉ
πΉ
of
πΉ
is
|πΉπ¨ππΉπ©|/ |π«π¨ππ«π©| The footsize ππΉ of πΉ is min{|π¨|,|π©|}.For oblivious BP lower bounds, the footprint
(π¨
,
π©) is the same for all of the embedded rectangles associated with the partition of the input sequence π
Lets us use communication lower bounds over the reduced variable set π¨
βͺ π©.Slide16
Lower Bounds from Embedded Rectangles
Strategy:
Write
where each
π
π-1
(1) is a union of embedded rectangles with footsize
π
,
the same footprint (π¨π,π©π), but different tailsFor Read-π BPs π¬ β€ 2π³πΊ (one per trace).For general length ππ BPs?Show that any embedded rectangle in π-1(1) with footsize
β₯ π
has density
β€
πΉ
.
Implies that
π¬
β
πΉ β₯ |π
-
1
(1
)|/|π«π|Β Slide17
Decomposition for Length
ππ
Recall that
π
=
. Fix one the 2πΊ
traces π
.
Apply layer assignment separately for each π with trace π to the sequence of variables queried on input π.One of β€ 2π³ possible layer assignments π=layers(π)Let private(π) be the pair consisting of the first π variables in the private inputs for Alice and Bob, respectively under πAt most
choices
(π¨, π©)
for
private
(π
)
Claim:
For disjoint
π¨,π©β[π]
with
|π¨|=|π©|=π
,
values πΆβπ«[π]-π¨-π©, trace π and layer assignment π, πΉ={πβπ«π
| private
(π)=(π¨,π©), π
[π]-π¨-π©=πΆ,
trace(π)=π,
layers(π)=π} is an embedded rectangle
with footprint
(π¨,π©) in
π-
1(1
).
Β
Claim
β¨ we can choose π¬β€
2πΊ
2
π³
Β Slide18
Proving the Claim
Claim
:
For
disjoint
π¨,π©β[π]
with
|π¨|=|π©|=π
,
values
πΆβπ«
[π]-π¨-π©
,
trace
π
and
layer
assignment
π
πΉ={πβπ«
π
|
private
(π)=(π¨,π©), π
[π]-π¨-π©
=πΆ,
trace(π)=π,
layers(π)=π}
is an embedded rectangle
with footprint (π¨,π©)
in
π-
1(1).
1
0Slide19
Lower Bounds from Embedded Rectangles
Strategy:
Write
where each
π
π
-1(1)
is a union of embedded rectangles with footsize
π
,
the same footprint (π¨π,π©π), but different tailsFor Read-π BPs: π¬ β€ 2π³πΊ (one per trace).For general length ππ BPs: π¬β€ 2(πΊ+1)
Show that any embedded rectangle in
π
-
1
(1)
with
footsize
β₯
π
has density β€ πΉ.Implies that π¬βπΉ β₯ |π-1(1)|/|π«π|
Β Slide20
Functions with Embedded Rectangle Tradeoffs
Show that any embedded rectangle in
π
-
1
(1)
with footsize = π has density β€ πΉ:Cannot be smaller than πΉ=|π«|-
2π
(just one point)
Functions π with πΉ=|π«|-πΊ π: Hamming separation HAMπΈ : [Ajtai 2002] e.g. π«=[π6]={0,1}6log π
Output is
1
iff
π«(π
π
,π
π
)β₯5 log
2
π
for all
πβ π
Membership in linear codes over finite field π½ [Jukna 2009]Middle bit of integer multiplication of numbers with π«={0,1}π
, i.e.
π
π-bit blocks. [Sauerhoff-Woelfel
2003]All have |π
-1
(1)|/|π«π
| β₯ 1
/|π«|.
Β Slide21
Lower Bounds for BPs/NBPs/RAMs for large
π«
Suppose that
and BRS layer assignment based on independent coin-flips is used. Then for
99%
of
π π=π/2π+1 π³=8π
2
2
π
π¬ β πΉ β₯ 0.99|π-1(1)|/|π«π| β₯ 0.99/|π«| π¬ β€ 2(πΊ+1)
and
πΉ
=|π«|
-πΊ π
For these values
< 2
(2
π
+6)
π
.
If
log
2
|π«| > 4π/πΊ
then
2πΊ
β₯
|π«|
πΊ π/4
Taking logs we get πΊ
β₯
πΊβ π log
2 |
π«|
Plugging in as before yields π»=π(π log((πlog π«
)/πΊ))
[B-Jayram
-Saks 2001, B-Saks-Sun-Vee 2003]
Β Slide22
Boolean Domains and
Β
For
we only have
πΉ=
2-πΊπ
Over
π½23π Ajtai defined an explicit cubic form π(π,π)=ππ§ππ π that requires πΉ=2-πΊπ
Alternatively:
π(π)=
1
iff
# of
(π,π)
pairs
s.t.
π<π
,
ππ=ππ=ππ+π=1 is oddΒ
0
π
1
π
2n-1
π
2n-2
π
n+2
π
n+1
π
n
π
4
π
3
π
2
M
πSlide23
BP Lower Bound Technology for
πΉ
=2
-πΊπ
Much more complicated argument that holds only up to small amounts of nondeterminism.
[Ajtai 2005]Uses correlations between private(π) values for related inputs. an independent layer assignment that leaves most layers unassigned to either playera probability of assigning a layer for input π to one of Alice or Bob that depends on the typical # of different layers in which input variables are read on input π.Theorem:[Ajtai 2005,B-Saks-Sun-Vee 2003]
and
π
π§ππ π both require
.
Β Slide24
BPs and Static Data Structures
Theorem:
[
Miltersen
-Nisan-Safra-
Wigderson 1998]With query set {0,1}π, time lower bounds of π(π) for size
ππ(
1
)
static cell-probe data structures require non-trivial time-space tradeoffs (i.e. π(log π) space requires superlinear time.)Proof:View the query πβ{0,1}π as the input vectorFor each fixed dataset π, have a different branching program π·π. Can use each cell of the cell-probe data structure to hold a node of π·πValues are the index of the variable and the names of the two pointers Size
ππ(
1
)
BP implies
π
π(
1
)
size cell-probe data structure with
π=3log π
-bit words (simply follow the branching program)
Time is preserved.
Can extend each step to full tree of height π at cost of 2π factor larger word-size π. Saves a factor π in time.Slide25
A Converse
Theorem:
[B-
Vee
2002]Static data structure:
2πΊ cells + extra work space at most πΊ time π» query algorithm that reads β€ π consecutive bits of the query in a one step yields a
2π-
way BP
π·
π of time π(π») and space π(πΊ+log π») for every dataset π.Proof: Each BP node corresponds to a cell name + configuration of the extra storage. Memory contents are fixed by π. The input bits accessed are determined by the algorithm and the fixed memory cell contents just read. Slide26
Application to
π
-Near Neighbor
[B-
Vee
2002]
Hamming separation HAMπΈ : e.g. π«=[π6]={0
,1}
6
log π
. Output is 1 iff π«(ππ,ππ)β₯5 log2 π for all πβ πCan solve HAMπΈ using a
π-Near Neighbor data structure:
Encode each coordinate
π
π
in
π«
as
π
π
using twice the bits so distance from
0
is fixed
Choose π to be set of all possible strings of the form 0π-1π 0π-π-1 π 0π-π HAMπΈ(π)=0
iff
π contains a close string to
π
So...BP lower bound for HAMπΈ
implies:
Theorem: Any π-Near Neighbor
data structure for Hamming distance on {0
,1
}π that reads
π(log π)
consecutive bits per time step and
memory cells requires time
π(π)
.
Β Slide27
Larger bounds for Huge Domains
Inspired by multiparty NOF communication complexity
Uses
embedded cylinder intersections
instead of embedded rectangles
Theorem: There is an explicit function over a huge domain for which
π»=π(π log2 (πlog|π«|/πΊ)) is needed. [B-Vee 2002]Drawbacks: Domain size |π«| requires
π―(log3
π)
bits to encode
Function, which is based on tensored, interleaved Reed-Solomon codes, is not known to be in NP.Slide28
Single-Output Methods for Multi-output ProblemsSlide29
Open Problems
Prove general BP lower bounds for out-degree 2 (arbitrary) directed graph reachability
Savitchβs
Theorem implies
πΊ=π(log
2
π) and we donβt expect πΊ=π(log π) is possible at all. (Would imply NL/poly=L/poly.)Prove that πΊ=π(log
π)
implies
π»=π(π2) or π»= π(π1+πΊ)At least match oblivious BP bound of π»=π(π log2(π/πΊ)) for out-degree 1.Improve best lower bound for Boolean functions from
to
π»=π(π
log(π/
πΊ
))
to match the large domain and oblivious BP bounds.
Generalize embedded rectangle techniques for Boolean inputs to embedded cylinder intersections.
Β Slide30
Open Problems
Prove any oblivious BP lower bound for an explicit single-output function that holds for time
π»= π log
π(1)
π
or even
π»=π(π log2 π).Prove π»=π(π log2 (π/πΊ)) oblivious BP lower bound for a wider range of natural functions.Slide31
Open Problems
Prove
an
π(π
2
)
size lower bound for an explicit Boolean function Find better time-space tradeoff lower bounds for other multi-output functions, e.g.Encoding asymptotically good error-correcting codes. [Bazzi-Mitter 2005] conjectured
Element distinctness
in sliding windows
[B-Clifford-Machmouchi 2013] ?Β Slide32
Okolβnishnikova Strategy for Layers
Use
4
π
2
equal-length
layers. Give a random subset of 2π of them to Bob π β€ 4π+1 follows automatically If length
then
variables
are read at most
timesEach such variable appears in at most layersProbability all of those layers given to Bob is β₯ (2ππ)-2πso expected size for Bobβs set is
β₯ (
2
ππ)
-
2
π
π/
2
Choose some subset that achieves the average
Bob only sees
β€
2
π(ππ/(
4π2)) = π/2 inputs so Aliceβs set size is always β₯ π/2Β