"... In PAGE 12: ... We end this section on a general comparative note. Table1 summarizes qualitatively different aspects of the three schemes we studied here. On the space optimality issue, replication is clearly very inefficient compared to the other two.... ..."

"... In PAGE 37: ... We have access to the budget of each school, although data quality here is somewhat suspect. In Table9 , we present per capita expenditure by school type, monthly teacher salaries and cash incentives, as well as shares of school income from various sources. Table 10 shows the changes in teacher attendance between 1995 and 1997, by school type.... In PAGE 39: ... But, in an environment, where resources are scarce, the central government may want to rethink its financing policy concerning private schools if the goal is to maximize learning in sciences and languages. [ Table9 about here] In Table 10, we take a look at teacher attendance, and compare the levels in 1995 to those in 1997 for each school type.... ..."

"... In PAGE 13: ... In the remainder of this section, we highlight our results analyzing and comparing the effects of applying the various methods in Table 1, and determining the optimal number of chunks. Table2 summarizes our analysis on the use of Chebychev for Equation 2 in conjunction with various values for p, with and without chunking. This table shows, for various desired confidences p, the optimal choice for the number of 5If k is even, take either of the two medians.... In PAGE 14: ... The bounds are shown for Chebychev (known ) . Alternatively, as in Table 1, we can obtain bounds for Chebychev (estimated ) by plugging in ^ for in Table2 , where ^ is computed over all the sample points, not just those in one chunk. We can also obtain bounds for Chebychev (conservative) by plugging in (MAX ? MIN)=2 for .... In PAGE 14: ...1) on the ineffectiveness of chunking when analyzed using bounds whose dependency on k is the kth root of the no chunking bounds. Although we indicated above that one can obtain bounds for Chebychev (conservative) by plugging in (MAX ? MIN)=2 for in Table2 , this bound is strictly worse than the Hoeffding bound for all probabilities at which chunking is useful for Chebychev (conservative). For example, for :76 lt; p :99, the Hoeffding bound with no chunking is smaller than any Chebychev (conservative) bound obtained with or without chunking.... In PAGE 15: ... Within the general chunking framework, we proposed and explored a number of alternative procedures for reporting an estimate and an error bound based on the chunks. Our results (see Table2 ) showed that for confidence probabilities above 96%, the best bounds are obtained by taking a small number of chunks and applying Chebychev (known ) or Chebychev (estimated ) to the chunk estimators and reporting the median of these estimators. For smaller probabilities, the best bounds are obtained by reporting an overall estimate ignoring the chunks (which is equivalent to taking an average of the chunk estimators6), and then either applying Hoeffding for guaranteed bounds (see Table 3), applying Chebychev (estimated ) for large sample bounds, or using the chunk estimators for empirical bounds.... In PAGE 20: ... The bar expands with increasing chunks but always includes the real answer. Finally, we also plot the median, which as shown in Table2 leads to tighter guaranteed bounds when the desired confidence is above 96%. Since we use a lower confidence bound (90%) in this case, we do not plot the error bar for the median.... ..."

"... In PAGE 13: ... In the remainder of this section, we highlight our results analyzing and comparing the effects of applying the various methods in Table 1, and determining the optimal number of chunks. Table2 summarizes our analysis on the use of Chebychev for Equation 2 in conjunction with various values for a182 , with and without chunking. This table shows, for various desired confidences a182 , the optimal choice for the number of 5If a44 is even, take either of the two medians.... In PAGE 14: ... The bounds are shown for Chebychev (known a165 ) . Alternatively, as in Table 1, we can obtain bounds for Chebychev (estimated a165 ) by plugging in a164 a165 for a165 in Table2 , where a164 a165 is computed over all the sample points, not just those in one chunk. We can also obtain bounds for Chebychev (conservative) by plugging in a29 MAX a90 MINa32 a10 a14 for a165 .... In PAGE 14: ...1) on the ineffectiveness of chunking when analyzed using bounds whose dependency on a24 is the a24 th root of the no chunking bounds. Although we indicated above that one can obtain bounds for Chebychev (conservative) by plugging in a29 MAX a90 MINa32 a10 a14 for a165 in Table2 , this bound is strictly worse than the Hoeffding bound for all probabilities at which chunking is useful for Chebychev (conservative). For example, for a22 a4 a143 a76 a182 a80 a22 a190a148a190 , the Hoeffding bound with no chunking is smaller than any Chebychev (conservative) bound obtained with or without chunking.... In PAGE 15: ... Within the general chunking framework, we proposed and explored a number of alternative procedures for reporting an estimate and an error bound based on the chunks. Our results (see Table2 ) showed that for confidence probabilities above 96%, the best bounds are obtained by taking a small number of chunks and applying Chebychev (known a165 ) or Chebychev (estimated a165 ) to the chunk estimators and reporting the median of these estimators. For smaller probabilities, the best bounds are obtained by reporting an overall estimate ignoring the chunks (which is equivalent to taking an average of the chunk estimators6), and then either applying Hoeffding for guaranteed bounds (see Table 3), applying Chebychev (estimated a165 ) for large sample bounds, or using the chunk estimators for empirical bounds.... In PAGE 20: ... The bar expands with increasing chunks but always includes the real answer. Finally, we also plot the median, which as shown in Table2 leads to tighter guaranteed bounds when the desired confidence is above a190a145a143 a149 . Since we use a lower confidence bound (a190a148a23 a149 ) in this case, we do not plot the error bar for the median.... ..."

"... In PAGE 21: ... Bailout Responsibilities The relationship between lagged profitability and bank finance variables varies over the period. Our simplest specifications (models 1 and 3, Table3 ) show no significant relationship between bank loans and SOE profitability at the start of the period. However, the significant positive coefficient on the interaction between profitability and year indicates that banks increasingly allocated credit to profitable SOEs.... In PAGE 22: ...21 By contrast, there were almost no significant relationships between past profitability and government finance at any point during the period (models 4 through 8, Table3 ). We interpret these results as evidence that, despite the constraints and limitations placed on banks, they lent to SOEs that were relatively profitable.... In PAGE 22: ... Signals and Separation The signaling variables perform largely as expected. SOEs that adopted performance contracts with a relatively high elasticity of wages with respect to profits tended to receive more bank finance than did others ( Table3 ). Those that had low base profit retention rates and those not under performance contracts received less bank finance.... In PAGE 22: ... Firms with high base retention rates had better access to government transfers; those with high wage elasticities had less access. The wage elasticity coefficient, however, does not achieve significance at conventionally accepted levels in the tobit specifications ( Table3 ). Performance contracts did not affect either access to or the share of government finance in the 1980s.... In PAGE 23: ...22 relationship between wage elasticity and bank finance in the 1990s. There is, however, a strong negative relationship between the performance contract dummy and government finance variables in the 1990s ( Table3 ). Base profit retention rates were phased out as policy instruments, and so we cannot test whether the strong negative correlation between those rates and bank finance persisted into the 1990s.... In PAGE 23: ... Base profit retention rates were phased out as policy instruments, and so we cannot test whether the strong negative correlation between those rates and bank finance persisted into the 1990s. The negative and significant coefficient for base profit retention in Table3 captures only the 1980s relationship.36 On the basis of the wage elasticity and the performance contract results, we conclude that performance contracts became less effective signals as their use became more widespread.... In PAGE 23: ... C. Other Reforms In the 1980s, SOEs whose managers had discretion over wages had better access to bank finance ( Table3 ), although the coefficient is insignificant (t=1.62) in the tobit specifications.... In PAGE 23: ...62) in the tobit specifications. Discretion over wages is negatively associated with direct government transfers in the 1980s, but positively so in the 1990s ( Table3 ). By contrast, the correlation between government financing and our index of production autonomy is positive and significant for the 1980s, and negative in the 1990s.... In PAGE 25: ... That said, they explain some variation in our dependent variables and thus warrant brief mention. SOEs with relatively high capital to labor ratios had greater access to government transfers than did others ( Table3 ). There is no significant relationship between that ratio and bank finance.... ..."

"... In PAGE 115: ...Table 1: performance guarantees: BV C4CB D1CPDC BPC7C8 CC same machine, then the swap neighborhood is empty; therefore, we define the swap neighborhood as one that consists of all possible jumps and all possible swaps. As can be seen in Table1 , the jump and swap neighborhoods have no constant performance guarantee for C9CZBV D1CPDC . Therefore, we introduce a push neighborhood, for which any local optimum is at most a factor BE A0 BE D1B7BD of optimal for C9CZBV D1CPDC .... In PAGE 115: ... When pushing all jobs on the critical machines is unsuccessful, we are in a push optimal solution. In Table1 the performance guarantees for the various local optima and scheduling problems are given. UB = AQ denotes that AQ is a performance guarantee and LB = AQ denotes that the performance guarantee cannot be less than AQ; AQ denotes that UB = LB = AQ.... In PAGE 121: ...Empty Out-tree To approximate solution 0,079 0,005 Tolower bound 0,115 0,318 Table1 : Average relative errors of approximate solution of algorithm based on y jt -formulation to approximate solution of algorithm based on x jt -formulation and lower bound ( = 1). The graph of precedence constraints Empty Out-tree To approximate solution 0,048 0,001 Tolower bound 0,073 0,309 Table 2: Average relative errors of approximate solution of algorithm based on y jt -formulation to approximate solution of algorithm based on x jt -formulation and lower bound ( =1= p 2).... ..."

"... In PAGE 8: ... The first case examined is that where the regulator chooses to never allow firms to return to a higher group when compliance is confirmed by inspection. Table1 shows the compliance which can be achieved subject to the inspection constraint. Subject to this constraint, it is clear that the addition of reputation groups has important positive compliance returns.... ..."

"... In PAGE 12: ... Since the outcome of the actual policy (b) is not a point on STCP Qc , , government has not combined policy instruments optimally. The optimal policy instrument combination for the present producer support level is summarized in Table1 . Firstly, it would be optimal to abandon the co-responsibility levy.... ..."

"... In PAGE 13: ... One of the purposes of the comparisons presented in this section is to analyze the characteristics of the production process that affect the relative gain of the optimal over the heuristic so- lution. Table9 presents the ratio of the heuristic policy cost to the optimal cost (see Table 4), for the three policies for a batch size of 500 units and for all 120 combinations of the cost and probability scenarios. Examining this table we can make following observations: a114 As the inspection cost grows, the Inspect-All Heuristic becomes more expensive relative to the optimal policy (compare scenarios J, I and H).... ..."

"... In PAGE 5: ... 4. RESULTS AND DISCUSSIONS The optimal operating policy function l * = l( k,i,t), obtained by solving SLP model, is given in Table3 . This policy is the steady state operating policy.... ..."