Uni versal Cl asses of Hash Funct i ons

( ext ended abst ract )

J. Lawr ence Car t er and Mar k N. Wegman
I BM Thomas J. Wat son Research Cent er
Yor k t own Hei ght s, New Yor k 10598
Ab s t r a c t :
Thi s paper gi ves an i , put i ndepende, t aver age l i near t i me al gor i t hm f or st or age
and r et r i eval on keys. The al gor i t hm makes a r andom choi ce of hash f unct i on f r om a
sui t abl e cl ass of hash f unct i ons. Gi ven any sequence of i nput s t he expect ed t i me
( aver agi ng over al l f unct i ons in t he cl ass) t o st or e and r et r i eve el ement s is l i near in
t he l engt h of ~ t he sequence. The number of ref erences t o t he dat a base r equi r ed by
t he al gor i t hm f or any i nput is ext r emel y cl ose t o t he t heor et i cal mi ni mum f or any
possi bl e hash f unct i on wi t h r andoml y di st r i but ed i nput s. We pr esent t hr ee sui t abl e
cl asses of hash f unct i ons whi ch al so may be eval uat ed r api dl y. The abi l i t y t o anal yze
t he cost of st or age and r et r i eval wi t hout wor r yi ng about t he di st r i but i on of t he i nput
al l ows as cor ol l ar i es i mpr ovement s on t he bounds of several al gor i t hms.
I n t r o d u c t i o n :
One may v i ew di f f er ent i nput s t o a pr ogr am as
el ement s f r om a cl ass of pr obl ems. The answer
gi ven by t he pr ogr am is, hopef ul l y, a cor r ect sol ut i on
t o t he pr obl em. Ordi nari l y, when one t al ks about
t he aver age per f or mance of a pr ogr am, one aver -
ages over t he cl ass of pr obl ems t he pr ogr am can
sol ve. Gi l l [ 2 ] , Ra b i n [ 7 ] , St rassen and So l o v a y [ 9 ]
have used a di f f er ent appr oach on some cl asses of
pr obl ems. They suggest t hat t he pr ogr am r andoml y
choose an al gor i t hm f r om t he cl ass of al gor i t hms t o
sol ve t he pr obl em. They are abl e t o bound t he av -
er age per f or mance of t he cl ass of al gor i t hms f or t he
wor s t case i nput . Thi s aver age on t he wor s t case
can be bet t er t han t he per f or mance of any known
si ngl e al gor i t hm on i t s wor st case. Some of t he
p r o b l e ms whi ch t hi s appr oach over comes are t he
f ol l owi ng:
1) Cl assi cal anal ysi s ( aver agi ng over t he cl ass of
i nput s) must make assumpt i ons about t he di st r i b-
ut i on of t he i nput s. These assumpt i ons may not hol d
in cer t ai n appl i cat i ons.
2) A consequence of (1) is t hat one cannot cl assi -
cal l y anal yze t he aver age per f or mance of a subr out -
i ne i ndependent l y of t he mai n r out i ne, si nce t he
mai n r out i ne may skew t he di st r i but i on of dat a.
3) I f t he pr ogr am is pr esent ed wi t h a wor s t - c as e
i nput , t her e is no way t o avoi d t he r esul t i ng poor
per f or mance. However , i f one had a cl ass of al gor -
i t hms t o choose f r om and was abl e t o real i ze t hat a
par t i cul ar al gor i t hm was r unni ng sl owl y on a g i v e n
i nput , t hen i t mi ght be possi bl e t o choose a di f f er ent
al gor i t hm.
In t hi s paper, we appl y t hese not i ons t o hash-
i ng f or st or age and r et r i eval , and suggest t hat a
cl ass of hash f unct i ons be used. We show t hat i f
t he cl ass of f unct i ons is chosen pr oper l y, t hen t he
aver age per f or mance of t he pr ogr am on any i nput
wi l l be nearl y as good as i f a si ngl e f unct i on, chosen
wi t h knowl edge of t he i nput , wer e used. W~ pr es-
ent sever al cl asses of hash f unct i ons whi ch i nsure
t hat ever y sampl e chosen f r om t he i nput space wi l l
be di st r i but ed evenl y by enough of t he f unct i ons t o
1 0 6
compensat e f or t he poor per f or mance of t he al gor -
i t hm when an unl ucky choi ce of f unct i on is made.
A br i ef out l i ne of our paper f ol l ows. Af t er i n-
t r oduci ng some not at i on, we def i ne a pr oper t y of a
cl ass of f unct i ons: uni ver sal 2. We s how t hat any
cl ass of f unct i ons t hat is uni ver sal 2 has t he pr oper t y
t hat gi ven any sampl e, a r andoml y chosen member
of t hat cl ass wi l l be expect ed t o di st r i but e t he s am-
pl e evenl y. We t hen ex hi bi t sever al uni ver sal 2
cl asses of f unct i ons whi ch can be eval uat ed easi l y.
Fi nal l y we gi ve sever al exampl es of t he use of t hese
f unct i ons.
N o t a t i o n :
I f S is a set, I SI wi l l denot e t he number of
el ement s in S. I x ] means t he l east i nt eger > x.
If x and y ar e bi t st r i ngs, t hen x ( ~ y i s t he excl usi ve-
or of x and y. Z wi l l r epr esent t he i nt eger s mod n.
Al l hash f unct i ons map a set A i nt o a set B. We wi l l
al ways assume I AI > I BI . A is somet i mes cal l ed
t h e set of possi bl e keys, and B t he set of i ndi ces. If
f is a hash f unct i on and x, y~A, we l et ~r(x,y) be 1 i f
x p y and f (x) = f(y), and 0 ot her wi se. Thus, 8f(x,y)
i s 1 i f x and y ar e di st i nct el ement s of A whi ch map
t o t he same val ue under f. I f f, x or y i s r epl aced in
8,(x,y) by a set , we sum over al l t he el ement s in t he
set . Thus, i f H is a col l ect i on of hash f unct i ons, x~A
and S e A t hen 8H(X,S) means
Z ~ 8f(x,y).
f~H y~S
Not i ce t hat t he or der of summat i on does not mat t er .
P r o p e r t i e s o f Un i v e r s a l Cl a s s e s :
Let H be a cl ass of f unct i ons f r om A t o B. We
say t hat H is universal 2 i f f or al l x, y in A,
8. ( x, y) < I HI . That is, H is uni ver sal 2 i f no pai r of
- I B I
di st i nct keys ar e mapped i nt o t he same i ndex by
mor e t han one I B I'h of t he f unct i ons. Pr oposi t i on 1
shows t hat t hi s bound on 8. ( x, y) is t i ght when I AI
i s much l ar ger t han I BI . The second pr oposi t i on
f o l l o ws al mos t i mmedi at el y f r om t he def i ni t i on of
uni ver sal 2 .
Pr oposi t i on 1: Gi ven any col l ect i on H of hash f unc-
t i ons (not necessar i l y uni versal 2), t her e exi st s x, yeA
such t hat
I HI I HI
8. ( x, y) > -
I BI I AI
Pr oof (sket ch): Let a = I AI and b = I BI . A coun-
t i ng ar gument shows t hat f or each f el l ,
a 2
(~f (A, A) > - - - a.
- b
Thus, 8H(A,A) _> a 2 1 Hl ( 1 / b - l / a ) . Ther ef or e, by
t he pi geon hol e pr i nci pl e, t her e exi st s x, yeA such
t hat 8H(x,y) > [ H I ( 1 / b - 1/ a) . [ ]
Pr oposi t i on 2: Let x be any el ement of A and S any
subset of A. Let f be a f unct i on chosen r andoml y
f r om a uni ver sal 2 cl ass of f unct i ons ( wi t h equal
pr obabi l i t i es on t he f unct i ons. ) Then. t he expect ed
number of el ement s of S t hat x col l i des wi t h, i.e.
I SI
8,(x,S), is _< - -
I BI
Pr oof :
Mean val ue of 8f(x,S)
_ 1 ~'~
8f(x,S)
I HI feB
= 1 ~_, 8H(x,y) ( by not at i on)
I HI y~s
1 ~'~ I H I
< ~ ~**~. q~- ( by def . of uni versal 2)
_ I S l [ ]
I B I "
In addi t i on t o bei ng usef ul l ater, Pr oposi t i on 2
has some di r ect appl i cat i ons. For i nst ance, an opt i -
cal char act er r eader post pr ocessi ng syst em is d e -
scr i bed in [ 8 ] . Thi s syst em is desi gned t o check i f
a wor d x is a member of a set of val i d wor ds S.
The set { f(y) l yeS } is st or ed in memor y. To t est
whet her x is in S, a check is made t o see i f f(x) is in
t he st or ed set . Si nce f(y) is gener al l y shor t er t han y,
a consi der abl e amount of space was saved. Ho we v -
er, t her e is a chance of er r or ; i f f (x)=f (y) and yeS,
t hen x may er r oneousl y be accept ed as val i d Pr op-
osi t i on 2 gi ves a bound on t he pr obabi l i t y of er r or
when f is chosen f r om a cl ass of uni versal 2 f 0nc-
t i ons.
1 0 7
We ar e i nt er est ed i n t he cost of usi ng t hese
f unct i ons in st or age and r et r i eval oper at i ons. Gi ven
a sequence R of r equest s ( i nser t i ons or r et r i eval s) t o
some dat a base, and a hash f unct i on f, we def i ne
t he cost of R wi t h r espect t o f, C(f,R), t o be t he sum
of t he cost s of t he i ndi vi dual r equest s. The cost of
an i ndi vi dual r equest r ef er r i ng t o an el ement x is one
pl us t he number of di st i nct pr evi ousl y i nser t ed ys
f or whi ch f (x) = f(y).
Thi s cost f unct i on r ef l ect s t he wor s t case cost
of i nser t i ng or f i ndi ng el ement s i n a st or age and
r et r i eval scheme in whi ch each el ement of B is as -
soci at ed wi t h a l i nked l i st, and an el ement x i s
st or ed in t he l i st associ at ed wi t h f (x) (see [ 1 ] , page
111 - 113.) Ot her col l i si on r esol ut i on schemes
wo u l d have ot her cost f unct i ons associ at ed wi t h
t hem. For exampl e, i f t he keys wi t h t he same i ndex
wer e st or ed in a bal anced t ree, t he cor r espondi ng
cost f unct i on woul d be smal l er.
The f ol l owi ng t heor em gi ves a ni ce bound on
t he ex pec t ed cost of usi ng a uni ver sal 2 cl ass of
hash f unct i ons wi t h t he l i nked l i st met hod f or r e-
sol vi ng col l i si ons.
Pr oposi t i on 3: Let R be a sequence of r r equest s
whi ch i ncl udes k i nser t i ons. Suppos e H i s a
uni ver sal 2 cl ass of hash f unct i ons. Then i f we
choose f at r andom f r om H, t he expect ed cost
k
C(f,R) is _< r(1 + ~- ~- ) .
Pr oof : The ex pec t ed cost of R is t he sum of t he
expect ed cost s of t he i ndi vi dual r equest s. Pr oposi -
t i on 2 and t he def i ni t i on of cost t el l us t hat an i ndi -
vi dual r equest has expect ed cost no gr eat er t han
k
1 + - - [ ]
I BI "
A speci al case of t hi s pr oposi t i on is t hat i f k is
r oughl y t he si ze of B t hen t he expect ed cost is 2r.
Not i ce t hat t hi s l i near bound hol ds f or any sequence
of r equest s, not j ust f or t he " a v e r a g e " r equest . For
many appl i cat i ons, t her e is an upper bound on t he
number of el ement s t o be st or ed and hence, B can
be c l ~s e n appr opr i at el y. I f t her e is no known upper
bound i t i s possi bl e t o dynami cal l y choose a size,
and rehash when t hat choi ce pr oves t o be t o o smal l .
Rehashi ng can be done in l i near t i me, and even in
r eal - t i me.
We can s how t hat t he expect ed cost ( aver agi ng
over t he hash f unct i ons) of any r equest is vi r t ual l y
t he same as t he expect ed cost ( aver agi ng over t he
possi bl e r equest s) of any si ngl e hash f unct i on when
appl i ed t o a r andom r equest af t er r andom i nser t i ons
have been made. The r eason i s as f ol l ows : Let
a = I AI and b = I BI . The count i ng ar gument
ci t ed in pr oposi t i on 1 i mpl i es t hat i f f is any hash
f unct i on and x and y ar e chosen at' r andom f r om A,
t hen t he expect at i on of 8f(x,y) is > ( 1 / b - l / a ) . It
f o l l o ws t h a t t he e x p e c t a t i o n of 8f ( x, S)
is > I S a ( 1 / b - l / a ) , wher e S is t he r andom s ub-
set of A whi ch has been pr evi ousl y st or ed. Thus,
t he r . ost of t he r e q u e s t i s at l east
1 + I S I ( 1 / b - l / a ) . When A is much l ar ger t han
B ( whi ch wi l l be t he case in most appl i cat i ons of
hashi ng), t hi s is vi r t ual l y t he same as t he cost of a
r equest ci t ed in t he pr oof of Pr oposi t i on 3.
It is al so possi bl e t o bound t he pr obabi l i t y t hat
gi ven a sequence of r equest s R, t he per f or mance of
a r andoml y chosen f unct i on wi l l be wor se t han t o l -
er abl e on R. Si nce we k now t hat C(f,R) must be at
l east r, we can concl ude t hat when k is r oughl y t he
si ze of I BI , t he pr obabi l i t y t hat C(f,R) > t . r is l ess
t han 1/ ( t - 1) . For some cl asses of hash f unct i ons
(such as t he l ast cl ass ment i oned in t hi s paper), i t is
possi bl e t o der i ve a bound on t he st andar d devi at i on
or hi gher moment s of t he cost of a r andoml y chosen
f unct i on on a par t i cul ar R. Thi s al l ows us t o get a
much bet t er est i mat e of t he pr obabi l i t y t hat C(f,R)
wi l l be l arge.
S o m e u n i v e r s a l 2 c l a s s e s :
The f i r st cl ass of uni ver sal 2 hash f unct i ons we
pr esent is sui t abl e f or appl i cat i ons wher e t he bi t
1 0 8
st r i ngs whi c h r epr es ent t he keys can c o n v e n i e n t l y
be mul t i pl i ed by t he c omput er .
Su p p o s e A={ O, 1 ..... a - l } and B={ O, 1 . . . . . b - l } .
Let p be a pr i me wi t h p>_a. Let g be any f unc t i on
f r om Zp t o B whi ch, as cl osel y as possi bl e, maps t he
same number of el ement s of Z, i nt o each el ement of
B. For mal l y, we r equi r e
I { y ~ Z p I g(y)=z} I < [ - ~ ] for all z ~ B. A
nat ur al choi ce f or g i s t he r esi due mo d u l o b. When
b = 2 k f or s ome k, t hi s a mo u n t s t o t aki ng t he l ast k
bi t s i n t he bi nar y r epr es ent at i on of y.
Let m and n be el ement s of Zp wi t h mpO. We
def i ne hm,o:A--,Z p by h (x) = ( mx+n) mo d p. No w
def i ne fm.n(X) = g(hm,,(X)). The cl ass H i s t he set
{f~,n I m, neZp m# 0 } . If desi r ed, p can be chosen
so t he mod p oper at i on can be cal cul at ed wi t h o u t a
di vi si on.
The f o l l o wi n g l emma i s usef ul i n pr ovi ng t hat
t hi s cl ass i s uni ver sal 2.
Lemma: When H i s def i ned as above, t hen f or any
x, y~A wi t h x # y , 8H(x,y) equal s t he number of o r -
der ed pai r s (r,s) wi t h r,s~Zp, r ~ s and g(r)=g(s).
Pr oof : Ther e is a nat ur al c or r es pondenc e b e t we e n
t he f unc t i ons hm, n and t he or der ed pai r s (r,s) wh e r e
r, seZp and r ~s. Speci f i cal l y, we i dent i f y t he f u n c -
t i on hm, n by t he or der ed pai r (hm,n(X),hm, n(y)). Si nce
m~O, hm,n(X)~hrn,n(y). Thi s c or r es pondenc e i s o n e -
t o - o n e and o n t o si nce t he l i near equat i ons
x m+ n ~ r ( mod p) and y m+ n - - s ( mod p) have a uni -
que sol ut i on f or m and n i n t he f i el d Zp.
If (r,s) i s t he pai r (hm,n(X),hm,n(y)), t hen
fm,n(X)=fm,n(y) i f and onl y i f g(r)=g(s). Thus, 8H(x,y)
i s t he n u mb e r of such pai r s. [ ]
Pr opos i t i on 4 : The cl ass H def i ned a b o v e i s
uni ver sal 2 .
Pr oof : Let n i be t he n u mb e r of e l e me n t s i n
[ t eZp I g( t ) =i } . The r est r i ct i on on g i s t hat f or
each i, ni < I - ~ ] . Si nce p and b ar e i nt eger s, t hi s
i mpl i es t hat n i < p - 1 +1. Thus, f or a gi ven r, t he
- b
n u mb e r of choi ces f or s such t hat r # s but g(r)=g(s)
i s < P - - I . Si nce t her e ar e p choi ces f or r,
b
P p- 1 > t he n u mb e r of or der ed pai r s (r,s) s at i s f y -
b
i ng t he c ondi t i on i n t he l emma = 8H(x,y). Recal l i ng
t hat f or x=y, 8H(x, y)=O, t hi s s hows t hat H i s
uni ver sal 2 . I-I
Fr equent l y, al gor i t hms ar e anal yzed mak i ng t he
as s umpt i on t hat mul t i pl i c at i on t akes uni t t i me. The
n u mb e r of mul t i pl i c at i ons i s s a i d , t o be t he cost of
t he al gor i t hm. Thi s model i s appr opr i at e when t her e
ar e no oper at i ons whi c h can be done an unbounded
n u mb e r of t i mes f or each mul t i pl i cat i on. When a
uni ver sal 2 cl ass of hash f unc t i ons i s used, t hen t he
n u mb e r of me mo r y r ef er ences per r equest can be
b o u n d e d wh e n av er aged ov er al l f unc t i ons i n t he
cl ass as i n Pr opos i t i on 3 ( assumi ng k i s not u n -
b o u n d e d wi t h r espect t o J BI . ) Ther ef or e t he model
i s appr opr i at e, and under i t t he hash f unc t i ons i n t he
cl ass gi ven abov e may be appl i ed i n c ons t ant t i me.
Thus i n t hi s model , f or any s equenc e of r equest s,
t he al gor i t hm t akes av er age t i me l i near i n t he n u m-
ber of r equest s.
I t may seem t hat t he addi t i on of n i n t he cl ass
of f unc t i ons gi ven abov e pl ays an uni mpor t ant rol e.
Thi s i s onl y par t l y t r oe. Suppos e f or m~Zp we d e -
f i ne hm(X) = ( mx) r ood p, and as bef or e def i ne fm(X)
as g(hm(X)). Let H = {frn I meZp m~ O} . I t can
be s h o wn t hat t hi s cl ass of f unc t i ons c omes wi t hi n a
f ac t or of t wo of bei ng uni ver sal 2, t hat is, f or any x
^ I HI
and y, 8H(x, y) <_ Z ' ~ T - . On t he ot her hand, t hi s
b o u n d c a n n o t be i mp r o v e d si gni f i cant l y. For i n-
st ance, l et b = J BI , and c hoos e k so t hat
p = k b + k + l i s pr i me ( t her e wi l l be i nf i ni t el y ma n y
such k' s. ) Let g(x) be t he f unc t i on x ( mod b). Let
x = 1 and y = b + l . Then t he 2k f unc t i ons f l ,
f 2 ..... fk f p - k f p- k. l ..... fp-1 each ma p x and y t o
t he s ame bi n. Thus, 8H(x, y) = 2k whi l e
I H_~_l_ p-_~_l _ k b + k _ ( 1 + 1 ) k .
I B I b b b
The uni ver sal 2 cl ass of f unc t i ons gi ven abov e
109
may not be conveni ent when t he keys are t oo l ong
t o be mul t i pl i ed u=~ing a si ngl e machi ne i nst r uct i on.
However , t he next pr oposi t i on gi ves a met hod of
ext endi ng a cl ass of f unct i ons f or l ong keys.
Pr oposi t i on 4: Suppose I BI is a power of t wo and
H is a cl ass of f unct i ons f r om A t o B wi t h t he p r o p -
er t y t hat f or each i~B,
I { f ~H I f ( x) ( ~f ( y) = i } I - ] HI Recal l t hat
I BI "
Q is ex c l us i v e- or . Then we can def i ne a
uni versal 2 cl ass of hash f unct i on f r om A x A t o B a s
f ol l ows. For f, geH, def i ne hf.g((x,y)) = f (x) ( ~ g(y).
Then t hi s new col l ect i on of hash f unct i ons
J = {hf.Q ] f , g~H} is uni ver sal 2 and al so sat i sf i es
t he condi t i on of t hi s pr oposi t i on.
The pr oof is qui t e si mi l ar t o t he pr oof of Pr op-
osi t i on 6, and t her ef or e omi t t ed. Pr oposi t i on 5 can
be appl i ed r epeat edl y t o ext end t he f unct i ons t o
ar bi t r ar i l y l ong keys. If t he f unct i ons in H can be
appl i ed in const ant t i me, t hen t he t i me r equi r ed t o
c omput e an ext ended f unct i on i s pr opor t i onal t o t he
l engt h of t he key.
The pr oposi t i on does not qui t e appl y t o t he
uni versal 2 cl ass of f unct i ons gi ven ear l i er bot h b e -
cause H is not a power of 2 (so I HI / I BI cannot be
an i nt eger) and because t he number of f unct i ons f or
whi ch f (x) ( ~ f(y) = 0, i.e. 6H(x,y), is act ual l y l ess
t han I HI / I B I . Bot h of t hese di f f er ences add smal l
f act or s t o 8j ( x, y) whi ch bar el y pr event J f r om bei ng
uni ver sal 2. The per cent age cont r i but ed by t hese
smal l f act or s decr ease asympt ot i cal l y t o 0 as p is
i ncr eased. Mor e det ai l s wi l l be gi ven in a f ut ur e p a -
per.
The f ol l owi ng is a cl ass of f unct i ons whi ch do
not r equi r e mul t i pl i cat i on, whi ch may be bet t er f or
many appl i cat i ons, Suppose A can be vi ewed as
t he set of i - di g; t number s wr i t t en in base a, and B
as t he set of bi nar y number s of l engt h j. Then I AI
= a ' a n d I BI = 2 j. Let M be t he cl ass of ar r ays of
l engt h i a, each of whos e el ement s ar e bi t s t r i n g s o f
l engt h j . For m~M, l et m(k) be t he bi t st r i ng whi ch
is t he k th el ement of m, and f or x t A, l et x k be t he k 'h
di gi t of x wr i t t en in base a. We def i ne
fro(x) -- m( x , +l ) O m( xl +x2+2) ( ~ ... (~) m( ~ xk+k).
k =l
The cl ass H i s t he set { f . I me M} .
Anot her way of def i ni ng t hi s cl ass i s t o gi ve a
pr ogr am f or appl yi ng a f unct i on t o an i nput x.
dcl m( i a) bi t ( j ) i ni t ( r andom) ;
dcl x(i ) di gi t s base a;
dcl val ue bi t (j );
di sp := 0;
val ue := 0;
f or k := 1 t o i do begi n
di sp := di sp + x(k) + 1;
val ue := val ue (~) m( di sp) ;
end;
r et ur n (val ue);
Pr oposi t i on 6: The c l a s s H def i ned above i s
uni ver sal 2 .
Pr oof : For x and y in A, s uppos e fm(X) i s t he
ex c l us i v e- or of t he r ows r 1, r 2 ..... r s of m, and fro(Y)
is r s. l ( ~. . . ( ~) r t . Not i ce t hat fro(x) = fro(Y) i f and onl y
i f r l ( ~) . . . ( ~) r t = 0. As s umi ng x ~ y , t her e wi l l be
some k such t hat r k is i nvol ved in t he cal cul at i on of
onl y one of fm(X) or fm(y). Then fro(x) wi l l equal fro(y)
i f and onl y i f r k is t he ex c l us i v e- or of t he ot her ri ' s.
Si nce t her e are 2 j = B possi bi l i t i es f or t hat r ow, x
and y wi l l col l i de f or one B th of t he possi bl e f unc -
t i ons fm" Thus, t he cl ass of al l frn'S i s uni ver sal 2.
[ ]
For a gi ven B, each f unct i on in H t akes t i me
l i near in t he l engt h of t he key. In addi t i on, we can
mor e accur at el y descr i be t he di st r i but i on of cost s of
a par t i cul ar sequence of r equest s under t he di f f er ent
f unct i ons. For i nst ance, Ma r k o ws k y [ 6 ] has shown
t hat f or any sequence R of r r equest s wi t h k i nser -
t i ons, and any posi t i ve t, t he pr obabi l i t y t hat
k and
C( f , R) - - r ( l + ) > r-t is l ess t han t Zl BI
7k
al so l ess t han - -
t 3 i B]
ii0
i m p or t ance:
Pr ogr ammer s somet i mes spend a consi der abl e
amount of t i me r ef i ni ng hash f unct i ons f or appl i ca-
t i ons wher e i t is cri t i cal t hat a uni f or m di st r i but i on
be achi eved ( [ 5 ] , p . 5 0 8 - 5 1 3 ) . Thi s may be di f f i cul t
because i t is necessar y t hat t he expect ed i nput set
not be bi ased in such a way as t o make tl~e hash
f unct i on per f or m poor l y. One of t he pr act i cal val ues
of a cl ass of uni versal 2 f unct i ons is t hat we k now
t hat t her e ar e many accept abl e f unct i ons in t he
cl ass. Si mpl y choosi ng a si ngl e hash f unct i on r an-
doml y f r om such a cl ass gi ves a hi gh expect at i on
t hat a uni f or m di st r i but i on wi l l be achi eved. Fur t her -
more, i f t he f unct i on is changed each t i me t he pr o-
gr am is run, t hen we can be sure of good per f or m-
ance aver aged over al l runs.
The t heor et i cal i mpor t ance is t hat i t al l ows one
t o get a good bound on t he aver age per f or mance of
an al gor i t hm whi ch uses hashi ng. The pr obl em wi t h
an or di nar y hashi ng scheme is t hat t he al gor i t hm
mi ght bi as t he i nf or mat i on bei ng st or ed and r e-
t r i eved t o wa r d s t hose cases t hat ar e di st r i but ed
unevenl y by t he par t i cul ar hash f unct i on bei ng used.
Rabi n has devel oped an al gor i t hm whi ch f i nds
t he near est nei ghbor s of a col l ect i on of poi nt s in a
pl ane, gi ven t he coor di nat es of t he poi nt s [ 7 ] . Thi s
al gor i t hm i nvol ves maki ng a r andom choi ce of
poi nt s, and i t uses hashi ng. If one al so r andoml y
chooses t he hash f unct i on f r om a uni versal 2 cl ass,
t hen t he expect ed r unni ng t i me of t he al gor i t hm wi l l
al ways be l i near t he number of poi nt s.
In [ 3 ] and [ 4 ] an al gor i t hm is suggest ed f or
mul t i pl yi ng spar se pol ynomi al s, usi ng hashi ng. We
can st r engt hen t he r esul t s of t hese paper s. Let t wo
pol ynomi al s, P and Q, have n and m non- z er o t er ms
r espect i vel y. If mul t i pl i es and aCds are vi ewed as
t aki ng const ant t i me, t hen gi ven any t wo pol y nomi -
al s P and O., we can mul t i pl y t hem in aver age t i me
O(n' m). Let CP1,CP 2 ..... CP n be t he coef f i ci ent s of
t he n t er ms of P. Let EP1,EP 2 ..... EP n be the ex p o-
nents of t hose t er ms. Let CQ i and EQ i st and f or t he
same quant i t i es of Q. The f ol l owi ng al gor i t hm wi l l
have t he per f or mance we suggest ed, assumi ng
St or e and Ret r i eve are i mpl ement ed usi ng a
uni ver sal 2 cl ass of hash f unct i ons. Thei r f i r st ar gu-
ment is a key, and t he second is t he val ue st or ed or
r et r i eved. If a val ue has not been st or ed pr evi ousl y
f or a gi ven key, a r et r i eve wi l l have a zer o val ue.
Begi n
Choose a hash f unct i on;
For i := 1 t o n do
For j := 1 t o md o
Begi n
Coef f i ci ent := CP i * CQj ;
Ret r i eve (EP i + EQj ,k);
St or e (EP i + EQi , Coef f i ci ent + k);
End;
Pri nt al l keys and val ues whi ch have been st or ed;
End;
Si nce addi t i on and mul t i pl i cat i on are vi ewed as
t aki ng const ant t i me, t he f i r st cl ass of f unct i ons we
pr esent ed woul d seem appr opr i at e f or t hi s anal ysi s.
Fut ur e r esear ch:
Ther e ar e a number of areas whi ch can be i n-
vest i gat ed, such as:
1) I mpr ove t he bounds ci t ed here on t he pr obabi l i t y
t hat a par t i cul ar f unct i on f r om t he t abl e l o o k - u p
cl ass wi l l per f or m poor l y on a par t i cul ar i nput .
2) Suppos e t he st or e and r et r i eve al gor i t hm is
changed so t hat t he l ast el ement r et r i eved is moved
t o t he f i r st posi t i on on t he l i st. Can a smal l er bound
on t he pr obabi l i t y t hat t he cost f unct i on exceeds a
speci f i ed val ue be der i ved?
3) Ext end t he anal ysi s t o ot her st or age and r et r i eval
al gor i t hms whi ch i nvol ve hashi ng, such as doubl e
hashi ng and open addr essi ng.
4) Gener al i ze t he def i ni t i on of uni ver sal 2 t o
uni ver sal n t o consi der t he act i on of t he cl ass of
f unct i ons on any col l ect i on of n el ement s of A. De-
t er mi ne i f one can obt ai n i mpr oved r esul t s wi t h such
iii
a stronger assumpti on. (One def i ni t i on of universal k
i mpl i es t hat t he expect ed number of keys f rom a
k- el ement set mappi ng i nto a gi ven el ement of B
woul d be bi nomi al l y di stri buted.)
5) When should~one deci de that a particular f unct i on
is a poor choice and it woul d be wor t h the ef f ort to
choose a new f unct i on and rehash?
Ack nowledg ements:
We woul d like to thank Ashok Chandra f or hel pi ng
suggest and f ormul at e the probl em; Dave Gl i ckman
f or suggesting we exami ne the t abl e l ook- up t ech-
ni que; Hani a Gaj ewska and Davi d Smi t h f or sug-
gesti ng revi si ons in an earl i er manuscri pt ; Wal t er
Rosenbaum f or discussions about a practical use of
this work; and George Markowsky f or hel p in un-
derstandi ng the di st ri but i on of perf ormance of the
class obt ai ned by;tabl e l ook- up.
References:
[ 1 ]
[ 2 ]
Aho, A.V., Hopcroft, J.E., and Ullman, J.D.,
The Design and Analysis of Computer Algor-
ithms, Addison-Wesley, Reading, Mass.
(1974).
Gill, J.T., III, Comput at i onal Compl exi t y of
Probabi l i sti c Turing Machines, Proceedings of
the Sixth AC,~f Symposium on the Theory of
[ 3 ]
[ 4 ]
[ 5 ]
[ 6 ]
[ 7]
[ 8 ]
[ 9]
Computing, May, 1976, Seattle, Washi ngton,
p. 91- 95.
Goto, E., and Kanada, Y., Hashing Lemmas
on Ti me Compl exi t i es wi t h Appl i cat i ons to
Formul a Mani pul at i on, Proceedings of the
1976 ACM Symposium on Symbolic and Alge-
braic Computation, August, 1976, Yorktown
Heights, New York, p. 149- 153.
Gustavson, F., and Yun, D.Y.Y., Ari t hmet i c
Compl exi t y of Unor der ed or Sparse
Pol ynomi al s Proceedings of the 1976 ACM
Symposium on Symbolic and Algebraic Computa-
tion, August, 1976, Yor kt own Heights, New
York, p. 154- 159.
Knuth, D.E., Sorting and Searching,
Addi son- Wesl ey, Reading, Mass. (1973).
Markowsky, G., pri vate communi cat i on.
Rabin, M.O., Probabilistic algorithms,
Proceedings of Symposium on New Directions and
Recent Results in Algorithms and Complexity
(1976 - to appear).
Rosenbaum, W.S., and Hilliard, J.J,, Mul t i -
f ont OCR postprocessi ng system. I BM Jour-
nal of Research and Development (Jul y 1975).
Strassen, V., and Sol ovay, R., A fast
Mont ~- Car l o test f or pri mal i ty, Sl AM Journal
on Computing (to appear).
1 1 2