"... In PAGE 8: ... 3 This speedup is mainly derived from the elimination of the program counter and by the specialization of generic instructions into instructions having the functionality of \quick quot; instructions. The last four columns of Table1 compare the perfor- mance of run-time and compile-time specialized code with the code generated by hand-optimized JIT and o -line compilers, respectively. The optimized JIT ka e produces code that is up to 4 times faster than that produced by run-time special- ization, while the optimized o -line compiler Hac produces code that is up to 10 times faster than that produced by compile-time specialization.... ..."

"... In PAGE 5: ... If the tar- get price fluctuates, bidders may alternatively bid to different targets without completing an auction. Table3 shows an exam- ple. Ping-pong bidding usually stops either by another bidder or by a completion of the auctions.... In PAGE 5: ... Applying an additional cost for retracting a bid mitigates the problem a little. However, as shown on Table3 , ping-pong bidding still happens with the cost of retracting a bid. If the cost of retracting a bid is too high the algorithm cannot adapt to the change in the environ- ment.... ..."

"... In PAGE 10: ...) Again, the new value of attribute X is obtained by computing aggregate function a(A) on the relational expression E [ Ein ins grouped on attribute B. The formulas given in Table5 don apos;t take into account the internal structure of selection predicates and update expressions. In the case of simple predicates (comparisons between an attribute and a constant3) and simple arithmetic update expressions (addition or subtraction of constants from an attribute), it is often pos- sible to eliminate some of the propagated modi cations.... In PAGE 12: ... Suppose attribute A in E is updated by an arbitrary arithmetic expression. From Table5 , the propagation through node A=k of the incoming up- date Ein upd would yield three modi cations Eout ins; Eout del ; Eout upd. However, the update operation can only cause the insertion of tuples into Q (that previously did not satisfy the selection predicate), and the deletion of tuples from Q that now do not satisfy the predicate (but did before the update).... In PAGE 13: ... Update operation M increases by 1% the rate of all San Francisco customers having accounts with balance gt; 5000 and rate lt; 3%. The input to the algorithm is: Q = balance;rate( balance lt;500^rate gt;0ACCOUNT) M = Eupd = E[rate0 = rate + 1]( balance gt;5000^rate lt;3(ACCOUNT gt; lt; ( city=0SF0CUSTOMER))) Using Table5 , the propagation of Eupd through the selection operation in Q yields insert and update operations (the delete operation is eliminated, see Table 7). We have: E0 ins = new(( balance lt;500^rate0 gt;0Eupd) gt; lt;( balance lt;500^rate gt;0Eupd)) E0 upd = ( balance lt;500^rate0 gt;0Eupd) 1 ( balance lt;500^rate gt;0Eupd) In both cases, predicates balance lt; 500 and balance gt; 5000 (the latter from Eupd) are contradictory, so both expressions E0 ins and E0 upd are unsatis able.... In PAGE 15: ... As an example, let op be a selection p performed over an arbitrary subtree S, and consider an update operation Ein upd associated with S and performed on an attribute in p. Ap- plying our propagation rules from the second line of Table5 , we obtain a triple hEout ins; Eout del ; Eout updi, corresponding to tuples added to, deleted from, and updated in the result of Q = pS. Then: E+(Q; M)(d) = new(( p0Ein upd(d)) gt; lt;( pEin upd(d)))[ new(( p0Ein upd(d)) 1 ( pEin upd(d))) E?(Q; M)(d) = old(( pEin upd(d)) gt; lt;( p0Ein upd(d)))[ old(( p0Ein upd(d)) 1 ( pEin... ..."