Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_Hiring_from_a_pool_of_workers.txt | Notice that in this example, worker w2, who is part of the minority, has a higher score than w3, who is not a minority. Worker w3 is hired, while w2 is not. Given that the affirmative action rules were designed with the intent of increasing the access that minorities have to these jobs, this type of lack of fairness is especially undesirable, since if the hiring process was purely merit-based, worker w2 would have been hired. |
Processed_Hiring_from_a_pool_of_workers.txt | Aygün and Bo (2013) describe the implementation of affirmative action in the admis- sion to Brazilian public universities. There, as here, the problems arise from the fact that positions (in that case seats) and workers are partitioned between those reserved for minorities and the open positions. Differently from there, however, unfair outcomes may not be prevented by workers even if they strategically manipulate their minority status. In the example above, even if w2 applied as a non-minority he would not be hired. |
Processed_Hiring_from_a_pool_of_workers.txt | Proposition 3. The Brazilian rule is aggregation independent and respects minority rights. However, it is not minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | One may conjecture that the scenario above is very unexpected, since the affirmative action law must have been enacted in response to minority workers not being hired based solely on scores. As shown in Aygün and Bo (2013), however, this conjecture may be misleading. For example, even if the average score obtained by minority workers is lower, one can have situations in which the preferences of the higher achieving minority workers are correlated, leading to the top minority workers in the entire population applying to a specific job. |
Processed_Hiring_from_a_pool_of_workers.txt | Example 5. Let W = {w1, w2, w3, w4}, and W F = {w1, w2} and W M = {w3, w4} be the set of female and male workers, respectively. Suppose that the scores are sW = (100, 90, 80, 50). Let q = 2. Then ϕN SW (W, q) = {w1, w3}. Since w2 is not hired but sw2 > sw3, the NSW rule is not fair. Moreover, it is easy to see that if either gender is considered a minority, the rule is also not minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | Proposition 4. The NSW rule is aggregation independent but not fair. If one of the genders is deemed as a minority, then it respects minority rights but is not minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | Until now, we evaluated rules from the perspective of whether they satisfy the de- sirable properties we introduced in the previous sections: aggregation independence, respecting minority rights, and minority fairness. While our analysis focuses on hir- ings potentially involving multiple rounds, one might wonder whether some of these problems would be present when the hiring is made in a single round. |
Processed_Hiring_from_a_pool_of_workers.txt | The property of aggregation independence does not imply anything regarding a sin- gle round of hiring. As we mentioned in section 5, since the properties of respecting minority rights and minority fairness are only satisfied when they are satisfied for any number of rounds, the rules for which they are satisfied will also satisfy them for a single round of hiring. For the French policies, the results do not change. |
Processed_Hiring_from_a_pool_of_workers.txt | Remark 3. Suppose, as in Proposition 2, that M ∗ ⊆ W ∗, and that for every w, w′ ∈ M ∗, w > sO sO w′. Policy 1 of the French assignment rule does not respect static minority rights, and Policy 2 respects static minority rights and is static minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | Next, consider the (sequential) use of minority reserves. When only a single round of hiring occurs, the rule satisfies all the desirable characteristics. Moreover, as we mentioned in section 5, when there is a single round of hiring, it is equivalent to the sequential adjusted minority reserves. |
Processed_Hiring_from_a_pool_of_workers.txt | Remark 4. The (sequential) use of minority reserves respects static minority rights and is static minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | The problems with minority reserves are not present under single hiring. As shown in Example 1, the issues reside on the fact that minority fairness requires that an asymmetric priority is given to minority candidates only when their proportion among those hired is below m. The sequential use of minority rights “lacks memory”, in the sense that it always gives this asymmetric priority to minority workers, regardless of how much it is still needed given past hires. When only one round of hiring occurs, that is not a problem. |
Processed_Hiring_from_a_pool_of_workers.txt | Remark 5. Let q1 be the number of hires that occur in a single round using the Brazilian rule. If k = q1, then the rule respects static minority rights and is static minority fair. If k > q1, then the rule respects static minority rights, but is not static minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | Remark 6. When only one round of hiring occurs, the NSW rule is not fair. If one of the genders is deemed as a minority, then it respects static minority rights but is not static minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | As we noted in section 1, most of the hiring in the public sector in France and Brazil, for example, requires the use of order of merit lists and reserve lists. While for some positions, such as police officers, it is natural to expect that the workers might be matched to different locations, many other positions are more specific and do not result in a pool of candidates shared by more than one job. Examples include the hiring of doctors with a specific specialty for a municipality with a single hospital, the role of economist in state companies, who only work at the headquarters, the entry-level diplomatic career, etc. |
Processed_Hiring_from_a_pool_of_workers.txt | At the end of each round r ≥ 1, we define the matching of workers to institutions as a function µr. |
Processed_Hiring_from_a_pool_of_workers.txt | A plural sequence of hires Λ is a list of pairs (i, q), where i ∈ I, and q is the number of workers hired. A plural sequence of hires Λ = h(i1, 3) , (i3, 2) , (i1, 2)i , for example, represents the case in which in the first round institution i1 hires three workers, in the second round institution i3 hires two workers, and then in the third round institution i1 hires two workers. |
Processed_Hiring_from_a_pool_of_workers.txt | When considering hiring with multiple institutions, a rule, therefore, can be general- ized to produce matchings instead of allocations. Given a pool of workers W , an initial matching µ0, and a plural sequence of hires Λ = h(i1, q1), (i2, q2), . . . , (ik, qk)i, a hiring rule Φ is derived from a set of institutional rules (Φi)i∈I by returning the matching combining all institutional rules. That is, if Φ (W, µ0, Λ) = µ, then µ(i) = Φi (W, µ0, Λ). |
Processed_Hiring_from_a_pool_of_workers.txt | We abuse notation and consider µ (w) to be an element of I, instead of a set with an element of I. |
Processed_Hiring_from_a_pool_of_workers.txt | Let µ∅ be a matching in which for all i ∈ I, µ∅(i) = ∅. We denote by Φi (W, Λ) and Φ (W, Λ) the values of Φi . Finally, we abuse notation and if Λ = h(i1, q1), (i2, q2), . . . , (ik, qk)i, we can append the plural sequence of hires with the (cid:0) notation hΛ, (i∗, q∗)i ≡ h(i1, q1), (i2, q2), . . . , (ik, qk), (i∗, q∗)i. |
Processed_Hiring_from_a_pool_of_workers.txt | Definition 7. A rule Φ satisfies common top if there exists a worker w∗ ∈ W such that, for every institution i ∈ I, w∗ ∈ Φi (W, h(i, 1)i). |
Processed_Hiring_from_a_pool_of_workers.txt | Definition 9. A rule Φ is aggregation independent if for any q ≥ q1 ≥ 0, sets of workers W , institution i ∈ I, and matching µ, Φi (W, µ, h(i, q)i) = Φi (W, µ, h(i, q1), (i, q − q1)i). |
Processed_Hiring_from_a_pool_of_workers.txt | Theorem 3. A rule satisfies common top, aggregation independence and permutation independence if and only if it is a single priority rule. |
Processed_Hiring_from_a_pool_of_workers.txt | Notice, however, that if the compositional objectives are interpreted as being applied to the entire set of workers hired by these institutions, as a whole, then we can simply use the sequential adjusted minority reserves, as defined in section 5, every time an institution wants to hire a given number of workers. This procedure satisfies aggregation independence and is also permutation independent. Moreover, it satisfies a natural adaptation of what it means to respect minority rights and minority fairness. Instead of applying to whether workers are hired by a specific institution, it applies to being hired at some institution. |
Processed_Hiring_from_a_pool_of_workers.txt | Take, for example, I = {i1, i2}, W = {w1, w2, w3, w4, w5}, M = {w1, w2, w5}, sw1 > sw2 > sw3 > sw4 > sw5, and the value of m for the two institutions being m1 = 0.5 and m2 = 0. If both institutions use the sequential adjusted minority reserves and i1 hires two workers before i2 also hires two, worker w3 is hired and w5 is not. If the order is that i2 hires first, then w5 is hired and w3 is not. A violation of permutation independence. |
Processed_Hiring_from_a_pool_of_workers.txt | In this paper, we evaluate a hiring method that is widely used around the world, especially for public sector jobs, where institutions select their workers over time from a pool of eligible workers. While the simple and natural rule of sequential priority satisfies all desirable characteristics, the addition of compositional objectives such as affirmative action policies increases the complexity of the procedures. We show that the rules being used in practical hiring processes, as well as the direct application of minority reserves, fail fairness or aggregation independence. When the compositional objectives can be modeled as affirmative action for minorities, the sequential adjusted minority reserves, which we introduced, is therefore the unique solution that satisfies those desirable properties. |
Processed_Hiring_from_a_pool_of_workers.txt | If multiple institutions hire from the same pool of applicants, however, we show that the space for different hiring criteria between institutions, is highly restricted when a minimal requirement of independence is imposed. |
Processed_Hiring_from_a_pool_of_workers.txt | Formal description of the rules. For the descriptions in this section, consider as given a set W of workers, a set M ⊆ W of minority workers, a set A of workers previously hired, a sequence of hires qr = hq1, q2, . . . , qki, and a score profile sW. |
Processed_Hiring_from_a_pool_of_workers.txt | Round 1: Let W1 = W . The highest scoring q1 workers in W1 are selected. Let A1, be the set of selected workers, where for each w ∈ A1 and each w′ ∈ W1 \ A1 we have sw > sw′, and |A1| = q1. |
Processed_Hiring_from_a_pool_of_workers.txt | Round k > 1: Let Wk = Wk−1 \ Ak−1. The highest scoring qk workers in Wk are selected. Let Ak be the set of selected workers, where for each w ∈ Ak and each w′ ∈ Wk \ Ak we have sw > sw′, and |Ak| = qk. |
Processed_Hiring_from_a_pool_of_workers.txt | Step 1.1: Let W1,1 = W , M1,1 = M ∩ W1,1 and q1,1 = ⌈m × q1⌉. The highest scoring min{q1,1, |M1,1|} workers in M1,1 are selected. Let A1,1 be the set of selected workers, where A1,1 ⊆ M1,1. |
Processed_Hiring_from_a_pool_of_workers.txt | Step 1.2: Let W1,2 = W1,1\A1,1, M1,2 = M ∩W1,2 and q1,2 = q1−|A1,1|. The highest scoring q1,2 workers in W1,2 are selected. Let A1,2 be the set of selected workers. Round k > 1: Step k.1: Let Wk,1 = Wk−1,2 \Ak−1,2, Mk,1 = M ∩(Wk−1,2 \Ak−1,2) and qk,1 = ⌈min{max{m − ω(A1,2)+...+ω(Ak−1,2) , 0} × qk, |Mk,1|}⌉. The highest scoring qk,1 workers in Mk,1 are selected. Let Ak,1 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Step k.2: Let Wk,2 = Wk,1\Ak,1, Mk,2 = M ∩Wk,2 and qk,2 = qk−|Ak,1|. The highest scoring qk,2 workers are selected from Wk,2. Let Ak,2 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Step 1.1: Let W1,1 = W , M1,1 = M ∩ W1,1 and q1,1 = ⌈m × q1⌉. The highest scoring min{q1,1, |M1,1|} workers are selected from M1,1. Let A1,1 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Step 1.2: Let W1,2 = W \ A1,1, M1,2 = M ∩ W1,2 and q1,2 = q1 − |A1,1|. The highest scoring q1,2 workers are selected from W1,2. Let A1,2 be the set of selected workers. Round k > 1: Step k.1: Let Wk,1 = Wk−1,2 \ Ak−1,2, Mk,1 = M ∩ (Wk−1,2 \ Ak−1,2) and qk,1 = ⌈m × qk⌉. The highest scoring min{qk,1, |Mk,1|} workers from Mk,1 are selected. Let Ak,1 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Step k.2: Let Wk,2 = W \ Ak,1, Mk,2 = M ∩ Wk,2 and qk,2 = qk − |Ak,1|. The highest scoring qk,2 workers from Wk,2 are selected. Let Ak,2 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | workers: T M ⊆ M with |T M| = ⌈k × m⌉ such that for each w ∈ T M and each w′ ∈ M \ T M, we have sw > sw′ and (ii) O, which is the set with the top k(1 − m) workers among those who were not chosen in (i), that is: O ⊆ W \ T M such that |O| = ⌊k(1 − m)⌋ and for each w ∈ O and w′ ∈ W \ (O ∪ T M), we have sw > sw′. Within each round a ≤ r, we have two steps. |
Processed_Hiring_from_a_pool_of_workers.txt | Step 1.1: Let O1,1 = O, T M1,1 = T M and q1,1 = ⌈m × q1⌉. The high- est scoring min{q1,1, |T M1,1|} minority workers are selected from T M1,1. LetA1,1 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Step 1.2: Let O1,2 = O1,1, T M1,2 = T M \ A1,1 and q1,2 = q1 − |A1,1|. The highest scoring q1,2 workers are selected from O1,2. Let A1,2 be the set of selected workers. Round k > 1: Step k.1: Let Ok,1 = Ok−1,2 \ Ak−1,2, T Mk,1 = T Mk−1,2 and qk,1 = ⌈m × qk⌉. The highest scoring min{qk,1, |T Mk,1|} minority workers are selected from T Mk,1. Let Ak,1 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Step k.2: Let Ok,2 = Ok,1, T Mk,2 = T M \ Ak,1 and qk,2 = qk − |Ak,1|. The highest scoring qk,2 workers are selected from Ok,2. LetAk,2 be the set of selected workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Let m be the target ratio of people with disabilities, sO profile for workers in the open competition and sD for workers in the competition for workers with disabilities. |
Processed_Hiring_from_a_pool_of_workers.txt | Policy 2: Let W1,1 = W , M1,1 = M. The highest scoring min{⌊(1 − m) × q1,1⌋, |M1,1|} workers, with respect to sD W , are selected from M1,1. Let A1,1 be the set of workers selected in this step. Then, the highest scoring min{⌈m × q1,1⌉, |W1,1 \ A1,1|} workers, with respect to sO W , are selected from W1,1 \ A1,1. Let A1,2 be the set of selected workers in this step, and let A1 = A1,1 ∪ A1,2. |
Processed_Hiring_from_a_pool_of_workers.txt | min ing qk,1 workers, with respect to sD (cid:16)P the set of workers selected in this step. |
Processed_Hiring_from_a_pool_of_workers.txt | scoring qk,2 workers, with respect to sO be the set of selected workers, and Ak = Ak,1 ∪ Ak,2. |
Processed_Hiring_from_a_pool_of_workers.txt | Proof of Theorem 1. First, we show that the SA rule respects minority rights and is minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | By definition, the SA rule respects minority rights, at the step k.1 of each round k, selects minority workers to satisfy the minimum requirement up to that round. Note that when there are not enough minority workers, SA selects all the available minority workers. |
Processed_Hiring_from_a_pool_of_workers.txt | Now, we show that the rule is minority fair. Let A ≡ ϕSA(W, hq1, . . . , qri) be the selection made for the problem. We want to show that (i) for each w, w′ ∈ W \ M, if w ∈ A and w′ /∈ A, then sw > sw′, (ii) for each w, w′ ∈ M, if w ∈ A and w′ /∈ A, then sw > sw′. (iii) for each w ∈ W \ M and w′ ∈ M, if sw < sw′ and w ∈ A, then w′ ∈ A, (iv) if there is w ∈ W \ M and w′ ∈ M with sw > sw′, w /∈ A and w′ ∈ A, then ω(A)/|A| ≤ m. |
Processed_Hiring_from_a_pool_of_workers.txt | First note that cases (i), (ii) and (iii) hold trivially as at step k.1 of each round k, the rule selects the highest scoring workers in M, and in step k.2 it selects the highest scoring workers. |
Processed_Hiring_from_a_pool_of_workers.txt | r ∩ M = ∅. Thus, we obtain |A ∩ M| = m × q which contradicts our assumption. For the next part, we need to introduce some concepts. First, we will say that given a set of workers W , minority workers M ⊆ W , score profile sW, q ≥ 0 and m ≥ 0, a set W ∗ ⊆ W respects minority rights if W ∗ satisfies the same conditions that ϕ (W, hqi) must satisfy when ϕ respects static minority rights (definition 5). Similarly, a set W ∗ ⊆ W is minority fair if W ∗ satisfies the same conditions that ϕ (W, hqi) must satisfy when ϕ is static minority fair (definition 6). |
Processed_Hiring_from_a_pool_of_workers.txt | To show that if a rule is minority fair and respects minority rights it is the SA rule, we show that, for any given number of workers to be hired q, there is only one set of workers of that size that is minority fair and respects minority rights. |
Processed_Hiring_from_a_pool_of_workers.txt | Lemma 1. For any given set of workers W , minority workers M ⊆ W , score profile sW, q ≥ 0 and m ≥ 0, there exists only one set W ∗ ⊆ W that respects minority rights, is minority fair and such that |W ∗| = q. |
Processed_Hiring_from_a_pool_of_workers.txt | Proof. First, note that property (i) of minority fairness implies that if a set W ∗ is minority fair, it contains the set topω(W ∗)(M), that is, the top ω(W ∗) highest scoring workers in M, and the set topq−ω(W ∗)(W \M), the q − ω(W ∗) highest scoring workers in W \M, both with respect to sW. |
Processed_Hiring_from_a_pool_of_workers.txt | Suppose, for contradiction, that there are W 1 ⊆ W and W 2 ⊆ W , where both W 1 and W 2 respect minority rights and are minority fair, |W 1| = |W 2| = q, and W 1 6= W 2. Note first that if |M| < m × q, respecting minority rights implies that M ⊂ W 1 and M ⊂ W 2. Minority fairness implies, moreover, that topq−|M |(W \M) ⊆ W 1 and topq−|M |(W \M) ⊆ W 2. But then W 1 = W 2, a contradiction. It must be, therefore, that |M| ≥ m × q. |
Processed_Hiring_from_a_pool_of_workers.txt | W 1 being minority fair, since m(W 1) > m × q and w∗ sw∗ 2 > sw∗ contradiction. |
Processed_Hiring_from_a_pool_of_workers.txt | Proof of Theorem 2. Let ϕ be a rule that is static minority fair, satisfies static minority rights, and is aggregation independent. Let λ∗ be any sequence of hires. |
Processed_Hiring_from_a_pool_of_workers.txt | We will follow by induction on the rounds in λ∗. First, the base hq1i: from lemma 1, there is a unique set W 1 ⊆ W that is minority fair and respects minority rights. Both ϕ and ϕSA are static minority fair and respect static minority rights. Therefore, ϕ(W, hq1i) = ϕSA(W, hq1i) = W 1. |
Processed_Hiring_from_a_pool_of_workers.txt | For the induction step, assume that ϕ(W, hq1, q2, . . . , qℓi) = ϕSA(W, hq1, q2, . . . , qℓi). |
Processed_Hiring_from_a_pool_of_workers.txt | claim above implies that ϕ (W, hqi) = ϕSA (W, hqi) and ϕ (W, hq + qℓ+1i) = ϕSA (W, hq + qℓ+1i). Since workers cannot be hired more than once, ϕ (W, hqi) ∩ ϕ (W, H, hqℓ+1i) = ∅. |
Processed_Hiring_from_a_pool_of_workers.txt | Proof of Proposition 2. Let W ∗ = {w1, w2, w3, w4, w5} with scores sW = (50, 40, 30, 20, 10). For simplicity, we will use m = 0.5. |
Processed_Hiring_from_a_pool_of_workers.txt | Consider now the case M ∗ = {w4, w5}. Consider two possibilities: q1 = q2 = 2 and q = 4. Then ϕF1 ({W ∗, M ∗} , hq1, q2i) = {w1, w2, w4, w5} but ϕF1 ({W ∗, M ∗} , q) = {w1, w2, w3, w4}, a violation of aggregation independence. |
Processed_Hiring_from_a_pool_of_workers.txt | Proof of Proposition 3. We will show that the Brazilian rule respects minority rights and is aggregation independent, but fails to be minority fair. |
Processed_Hiring_from_a_pool_of_workers.txt | Proof of Theorem 3. The single priority rule satisfying common top, aggregation inde- pendence and permutation independence is straightforward to see. |
Processed_Hiring_from_a_pool_of_workers.txt | Denote by sequence of single hirings a plural sequence of hires of the form Λ = h(i1, 1), (i2, 1), (i3, 1), . . .i. That is, every hire made by any institution in any round consists of only one worker. |
Processed_Hiring_from_a_pool_of_workers.txt | Claim 1. Let W be a set of workers, and Φ be a rule that satisfies common top and permutation independence. There exists a ranking ≻∗ over W such that for any sequence of single hirings Λ, Φ(W, Λ) = Φ≻∗ is the single priority rule that uses ≻∗. |
Processed_Hiring_from_a_pool_of_workers.txt | Proof. We will prove by induction on the number of hires in a plural sequence of hires. That is, we will show that there exists a ranking ≻∗, independent of the sequence of hires, that is followed by Φ as a single priority. |
Processed_Hiring_from_a_pool_of_workers.txt | In the remaining steps of the proof, the set W and the rule Φ are given, and so for any plural sequence of hires Λ, we will use the notation {Λ} to represent the set i∈I Φi (W, Λ). That is, {Λ} is the set of workers in W hired by some institution under Φ after the sequence of hires Λ. Since we will only look at single hirings, we will S represent plural sequences of hires as sequences of institutions and use hi1, i2, . . .i to represent h(i1, 1), (i2, 1), . . .i. |
Processed_Hiring_from_a_pool_of_workers.txt | We will use (PI) to indicate that we used the property of permutation independence of Φ, (AI) to indicate that we used aggregation independence, and (CT) for common top. |
Processed_Hiring_from_a_pool_of_workers.txt | If Λ1, Λ2, and i ∈ I are such that {Λ1} = {Λ2} and Φi (W, Λ1) = Φi (W, Λ2), then Φi (W, hΛ1, ii) = Φi (W, hΛ2, ii). That is, if Λ1 and Λ2 are such that institution i hires the same set of workers, and the set of workers remaining after all of the hires in both plural sequences of hires is the same, then i would hire the same worker after both Λ1 and Λ2. This comes directly from the definition of a hiring rule Φi. |
Processed_Hiring_from_a_pool_of_workers.txt | Induction Base The induction base is the case where the smallest number of hires is made while still having at least two institutions hiring. Therefore |Λ| = 2. Suppose that the claim is not true. That is, there might be plural sequences of hires with two hires that cannot be explained by a ranking ≻∗ over W . That implies that there are Λ1 6= Λ2, where Λ1 = hi1, i2i, Λ2 = hi3, i4i, and {Λ1} 6= {Λ2}. |
Processed_Hiring_from_a_pool_of_workers.txt | Since the sequences of hires involve at least two institutions, i1 6= i2 and i3 6= i4. Since Λ1 6= Λ2, there are two cases to consider: (i) i1 6= i3, and (ii) i2 6= i4. Consider (i). By (PI), {hi1, i2i} = {hi2, i1i}. By (P*), (CT) and the fact that i1 6= i3, {hi2, i1i} = {hi2, i3i}. By (PI), {hi2, i3i} = {hi3, i2i}. By (P*), (CT) and the fact that i3 6= i4, {hi3, i2i} = {hi3, i4i}. But then {hi1, i2i} = {hi3, i4i}, a contradiction. For case (ii), (PI) implies that {hi1, i2i} = {hi2, i1i} and {hi3, i4i} = {hi4, i3i}, which makes this case equivalent to (i). |
Processed_Hiring_from_a_pool_of_workers.txt | Induction Step We now assume that for every sequence of single hirings Λ such that |Λ| ≤ k, the rule Φ hires according to the ranking ≻∗. We will use (IA) to indicate that we are using this induction assumption. |
Processed_Hiring_from_a_pool_of_workers.txt | Suppose now that the claim is not true. That is, there are sequences of single hirings Λ1, Λ2, such that |Λ1| = |Λ2| = k, and institutions i1, i2 ∈ I, for which {hΛ1, i1i} 6= {hΛ2, i2i}. There are two cases to consider. |
Processed_Hiring_from_a_pool_of_workers.txt | Every round in which i3 hires in Λ2, the hire is made by i1 instead, • Denote by t∗ the first round in which i3 does not hire in Λ2. Note that this must exist, since Λ2 has hirings from at least two institutions. Let i2 hire in round t∗ in Λc 1 instead. |
Processed_Hiring_from_a_pool_of_workers.txt | Notice, therefore, that there is no hire from i3 in Λc hΛb Next, let Λd 1, except that the i2 in round t∗ is replaced by i3. By (cid:9) (cid:8) (PI), {hΛc 1 are exactly 1, i3i} = the same as in Λ2, and as a result, (P*) and (IA) imply that the last hire made by i2 in hΛd 1, i2i is the same as in hΛ2, i2i. Not only that, (IA) implies that the set of workers hired in the first k hires are the same, and therefore = {hΛ2, i2i}, implying that {hΛ1, i1i} = {hΛ2, i2i}, a contradiction. |
Processed_Hiring_from_a_pool_of_workers.txt | That is, aggregation independence implies that each hire from an institution can be split into single hires without changing the workers that are chosen, round by round.30 Our claim above implies, therefore, that the rule Φ∗ must be single priority, finishing our proof. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | Abstract: The Square Kilometre Array and its pathfinders ASKAP and MeerKAT will produce prodigious amounts of data that necessitate automated source finding. The performance of automated source finders can be improved by pre-processing a dataset. In preparation for the WALLABY and DINGO surveys, we have used a test HI datacube constructed from actual Westerbork Telescope noise and WHISP HI galaxies to test the real world improvement of linear smoothing, the Duchamp source finder’s wavelet de-noising, iterative median smoothing and mathematical morphology subtraction, on intensity threshold source finding of spectral line datasets. To compare these pre-processing methods we have generated completeness-reliability performance curves for each method and a range of input parameters. We find that iterative median smoothing produces the best source finding results for ASKAP HI spectral line observations, but wavelet de-noising is a safer pre-processing technique. In this paper we also present our implementations of iterative median smoothing and mathematical morphology subtraction. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | Source finding reduces a dataset to a manageable ab- stract representation that is a collection of objects with physically meaningful properties. When a dataset becomes too large the dataset is virtually impossi- ble to work with directly, and the catalogue is the only method of data exploration. This is the case for the two Australian Square Kilometre Array Pathfinder (ASKAP) HI surveys, Wide-field ASKAP Legacy L- band All-sky Blind surveY (WALLABY) (Koribalski et al. (2009); Koribalski, B., Staveley-Smith, L. et al., in preparation) and Deep Investigation of Neutral Gas Origins (DINGO). Individual ASKAP spectral line ob- servations will be at least 2,048 by 2,048 by 16,384 vox- els, which is 256GB (512GB) in float (double) preci- sion and only directly accessible using supercomputing facilities. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | ness that a source finder achieves on a given dataset. It is a common practise to pre-process a dataset before applying a source finding method. The goal of pre-processing is to improve both the completeness and reliability of the source finder. This is achieved by ‘correcting’ the dataset. In a ‘corrected’ dataset the noise behaves as your source finder assumes, the dataset is free from background structure and sources have maximised signal-to-noise ratios. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | It should be noted that the term ‘signal-to-noise ra- tio’ in this context does not account for the Jy/beam units of radio observations. Technically a radio obser- vation should be re-scaled for the new beam size when an observation is smoothed. This involves reversing the initial beam scaling, which in some circumstances increases the noise level more than it is minimised by smoothing. For the purposes of pre-processing and source finding though units are irrelevant. The signal- to-noise ratios discussed here therefore refer to the unscaled signal-to-noise ratios of radio observations, which are always enhanced by smoothing. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | methods for ASKAP HI datacubes: smoothing, mathematical morphology subtraction, wavelet de-noising and linear smoothing. We will compare these pre-processing methods by examining the effect they have on the performance curve of a simple inten- sity threshold source finder. We analyse the effect on an intensity threshold performance curve, because in- tensity thresholding is at the core of most source find- ers eg. SExtractor (Bertin & Arnouts 1996) and SFind (Hopkins et al. 2002). |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | most commonly used pre-processing methods. These two methods take contrasting approaches. Linear smooth- ing uses averaging or convolution to re-distribute the flux within the datacube so that noise fluctuations are reduced more than source signal, which results in increased source signal-to-noise ratios. Wavelet de- noising however tries to directly subtract noise from the datacube. A wavelet transform decomposes a dat- acube into signal on different scales at all positions within the datacube. Signal on scales smaller and larger than the expected size of sources can then be re- moved. Alternatively, the noise level on different scales can be measured from the wavelet transform, and only ‘significant’ signal (as defined in some way by a user) on each scale is retained. We use the Duchamp source finder (Whiting 2008, 2012) to implement both, be- cause Duchamp is not only a commonly used state-of- the-art source finder, but it is also the default ASKAP source finder. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | wavelet transform kernel from matching an astronom- ical source, which is either a positive feature (emission) or a negative feature (absorption). Consequently, any astronomical source will necessarily exist on multiple scales (and probably multiple locations). Depending on how the wavelet transform information is filtered to de-noise a datacube, this can have a negative effect on source finder performance. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | We will carry out our comparison of these source finder pre-processing methods using the Westerbork Telescope (WSRT) test datacube in Serra et al. (2012). The WSRT test datacube was created by injecting WHISP sources (van der Hulst 2002) into a datacube of real WSRT spectral datacube noise. This test dat- acube not only contains real sources embedded in real noise, but the resolution (both angular and spectral) and noise level closely matches that expected of the APERTIF and ASKAP telescopes. In particular, the ′′ 30 channels and 1.86 mJy/beam noise level of this test datacube, is designed to match WALLABY observations. This al- lows us to test the ‘real world’ performance of the var- ious pre-processing methods. We illustrate the scale of the WHISP sources and the test datacube using a channel map in Figure 1. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | comparison, because Arias-Castro & Donoho (2009) has shown that iterative median smoothing produces a larger gain in source signal-to-noise ratio than linear smooth- ing methods. The key is that calculating a median is a non-linear process that preserves source ‘edges’. Source edges are preserved because median calcula- tions are insensitive to sample outliers. Crucially, Arias-Castro & Donoho (2009) found that only two iterations are required, so long as the first iteration uses the smallest smooth- ing kernel possible. This minimal number of iterations results in a reasonable computational load even when large smoothing kernels are used for the second itera- tion. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | We chose to test mathematical morphology sub- traction, because it is a proven technique for size fil- tering images. We can use mathematical morphology to filter out the small-scale information in a dataset to identify large scale structure in the image (Rudnick 2002). Subtracting this large scale structure can po- tentially improve reliability by re-normalising the dataset noise properties, so that the mean of the noise distri- bution is constant throughout the dataset. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | There are distinct disadvantages common to all of these pre-processing methods. First, a poor choice of smoothing kernel can actually decrease source signal- to-noise ratios when using linear smoothing, iterative median smoothing and mathematical morphology sub- traction. This is dealt with by using multiple smooth- ing kernels. This increases the computational load though, and the results of the multiple smoothing ker- nels need to be combined intelligently. Second, all of these pre-processing methods need to account for datasets having different types of dimensions. Fl¨oer & Winkel (2012) is a good example of a wavelet transform for HI datacubes that accounts for the difference between the RA, Dec angular dimensions and the frequency dimen- sion. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | Figure 1: This is a channel map of the WSRT test datacube used in this paper. The source in the centre of this channel map is one of the most spatially resolved sources. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | The rest of this paper is organised in the follow- ing way. We begin by presenting the implementation of iterative median smoothing and mathematical mor- phology that we used for our comparison in Section 2. Next we compare and analyse the performance impact of the various source finder pre-processing methods in Section 3. Then we finish in Section 4 with our con- clusions and recommendations. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | Iterative median smoothing is reviewed and analysed in depth in Arias-Castro & Donoho (2009). Here we present a brief overview of iterative median smoothing and our implementation. Iterative median smooth- ing is the process of repeatedly replacing each ele- ment of a dataset with the median of a region cen- tered on the element, using progressively larger re- gions. Arias-Castro & Donoho (2009) found that with the right choice of region size, only two iterations are needed to obtain near maximal performance from it- erative median smoothing. To do so, the first itera- tion needs to measure the median of the smallest re- gion possible, and the second iteration needs to mea- sure the median of a region matching the size and shape of the signal being optimised. The first pass re- moves elements that are outliers and the second pass re-distributes the flux, while preserving source edges, to improve source signal-to-noise ratio by averaging noise. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | In this paper we use a two-pass implementation of iterative median smoothing. The first pass uses only the element being processed and the six neighbouring voxels that share a face with it. This is a 3-D extension of ‘four-connected’ pixels. We chose a 3-D version of ‘four-connected’ pixels for the first iteration, because this matches the pixel size of the beam’s central com- ponent. This is sufficiently large to filter out individual noisy voxels in the presence of beam convolution (con- firmed by us visually). The second pass uses either a rectangular parallelepiped or an ellipsoid extending along the frequency axis as a smoothing element. The ellipsoid (rectangular parallelepiped) is defined using separate radii (lengths) for the frequency axis and an- gular axes, RA and Dec. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | We have developed software that efficiently applies two-pass iterative median smoothing using an initial six-connected voxel element followed by a n-channel rectangular parallelepiped or ellipsoid element. This software deals with large datacubes using a two-pronged approach. First, the software uses a ‘buffer-and-shuffle’ approach to minimise the memory overheads associ- ated with multiply smoothing the input datacube. Sec- ond, sufficiently large datacubes are broken up into manageable ‘chunks’, and processed sequentially. The use of a buffer-and-shuffle approach allows processing of files as large as three gigabytes on a 32-bit laptop before efficiency requires segmentation of a datacube. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | and ‘closing’ an object by dilating then eroding. The easiest way to think of the open and close operations is the effect that they have on text. The open op- eration sharpens the characters by filtering out small scale structure. A consequence of the open operation is that it ‘rounds’ the remaining structure. The close operation by contrast blurs the characters. It ampli- fies small scale structure using the large scale structure as a guide. Unfortunately, sufficiently close characters will be merged into each other. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | Monochromatic images, such as HI datacubes, are processed using ‘structuring element’ kernels. Using a structuring element dilation is achieved by replac- ing an element with the maximum value in the region around it (specified by the kernel). Similarly, erosion is achieved by replacing an element with the minimum of the surrounding region. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | In this paper we use the approach developed in Rudnick (2002). We use an open operation to filter the small scale structure out of the image and obtain an open image of the large scale structure. By subtract- ing this large scale structure from the original image, we can obtain a residual image of the small scale struc- ture. We use the approach in Rudnick (2002) because combining the open and residual images preserves the flux of the original image. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | We compare the effectiveness of various source finder pre-processing methods by constructing completeness- reliability performance curves for a simple intensity threshold source finder, after applying the various pre- processing methods. We have chosen to compare the performance curves instead of completeness and relia- bility for a given threshold, because each pre-processing method will alter the datacube and its noise distribu- tion in different ways. An arbitrary threshold (in units of noise level) will therefore not be consistent across the outputs of the various pre-processing methods. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | The simple intensity threshold source finder that we use here is a two step process. In the first step a C++ code robustly measures the datacube’s standard deviation from the interquartile range, and then se- lects all voxels greater than a user specified multiple of the standard deviation as source voxels. These flagged voxels are then combined into objects, merged, size filtered and turned into a catalogue using the object generating library presented in Jurek (2012). Every catalogue used merging lengths of 1, 1 and 3 empty voxels along the RA, Dec and frequency axes. We size filtered every catalogue to only include object’s con- taining 14 voxels, occupying 5 lines of sight and whose extent in the RA, Dec and frequency dimensions is at least 3 voxels. We chose these merging and size filter- ing parameters, because they are representative of the values that would be chosen by a user when trying to maximise completeness. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | we use these sparse representations to match the ob- ject’s in the various output catalogues to the input cat- alogue on a voxel-to-voxel basis. From these matches we calculate the source finder performance metrics. |
Processed_Examining_alternatives_to_wavelet_de-noising_for_a.txt | we measured for the various pre-processing methods are plotted in Figure 2. We measured performance curves for all of the sources and for a subset of ‘de- tectable’ objects, which are defined as having peak signal-to-noise ratios greater than or equal to three. We refer to these two types of performance curves as the total and detectable performance curves. The input parameters we used with these pre-processing methods are listed in Table 1 in the Appendix. To obtain a meaningful comparison of the pre-processing methods, we used similar parameter values across the pre-processing methods. The choice of mathematical morphology opening subtraction kernels ranges from the the beam size to 20 times the beam size. This en- sures that our choice of opening kernels brackets the optimal size of three times the source size, that was found by Rudnick (2002). |
Processed_The_lunar_Askaryan_technique_with_the_Square_Kilom.txt | event rate will be presented, along with the predicted energy and directional resolution. Prospects for directional studies with phase 1 of the SKA will be discussed, as will the major technical challenges to be overcome to make full use of this powerful instrument. Finally, we show how phase 2 of the SKA could provide a vast increase in the number of detected cosmic rays at the highest energies, and thus to provide new insight into their spectrum and origin. |
Processed_The_lunar_Askaryan_technique_with_the_Square_Kilom.txt | c(cid:13) Copyright owned by the author(s) under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike Licence. |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.