Home On Decay Centrality
Article
Licensed
Unlicensed Requires Authentication

On Decay Centrality

  • Nikolas Tsakas EMAIL logo
Published/Copyright: January 16, 2018

Abstract

We establish a relationship between decay centrality and two widely used measures of centrality, namely degree and closeness. We show that for low values of the decay parameter the nodes with maximum decay centrality also have maximum degree, whereas for high values of the decay parameter they also maximize closeness. For intermediate values, we provide sufficient conditions that allow the comparison of decay centrality of different nodes and we show via numerical simulations that in the vast majority of networks, the nodes with maximum decay centrality are characterized by a threshold on the decay parameter below which they belong to the set of nodes with maximum degree and above which they belong to the set of nodes with maximum closeness. We also propose a simple rule of thumb that ensures a nearly optimal choice with very high probability.

JEL Classification: C15; C63; D85

Acknowledgement

Part of this project was carried out while the author was at SUTD–MIT International Design Center at Singapore University of Technology and Design supported by grant IDG31300110.

Appendix

A Proofs

Proof of Proposition 1

Notice that decay centrality can be rewritten as DCi(δ)=δDi+l=2n1δlDil and consider the limit limδ0DCi(δ)DCj(δ)δ=00DiDj>0, where the last inequality is implied by Di>Dj. The expression that appears inside the limit is continuous in δ, therefore exists δ_ such that for all δ(0,δ_) it holds that DCi(δ)DCj(δ)δ>0 or equivalently DCi(δ)>DCj(δ).

Proof of Proposition 2

If Di>Dj then the result holds immediately by Proposition 1. Otherwise, let l~2 be the first instance such that Dil~>Djl~. In this case, the difference between decay centralities can be written as DCi(δ)DCj(δ)=l=l~n1δl(DilDjl), because all previous terms of the sum are equal to zero. Hence, limδ0DCi(δ)DCj(δ)δl~=00Dil~Djl~>0 and by continuity with respect to δ, there is δ_ such that for all δ(0,δ_) it holds that DCi(δ)DCj(δ)δl~>0 or equivalently DCi(δ)>DCj(δ).

Proof of Proposition 3

We need three direct observations to obtain the result. First, note that Ci>Cjkid(i,k)<kjd(j,k). Second, recall the alternative formulation of decay centrality as DCi(δ)=δDi+l=2n1δlDil and third to observe that kid(i,k)=Di+l=2n1lDil. Therefore,

limδ1DCi(δ)DCj(δ)1δ=limδ1(δDi+l=2n1δlDil)(δDj+l=2n1δlDjl)1δ=00=limδ1(Di+l=2n1lδl1Dil)(Dj+l=2n1lδl1Djl)1==(kid(i,k)kjd(j,k))>0

where the last inequality is implied by Ci>Cj. Hence, by continuity with respect to δ, exists δ¯ such that for all δ(δ¯,1) holds that DCi(δ)DCj(δ)1δ>0 or equivalently DCi(δ)>DCj(δ).

Proof of Proposition 4

If Ci>Cj the result holds immediately by Proposition 3. Otherwise, let l~2 be the first instance such that Cil~>Cjl~, hence also the first instance in which Fil~<Fjl~. Hence:

(1)limδ1DCi(δ)DCj(δ)(1δ)l~=limδ1(l=1n1δlDil)(l=1n1δlDjl)(1δ)l~=00=00limδ1(l=1n1lδl1Dil)(l=1n1lδl1Djl)l~(1δ)l~1==00limδ1(l=2n1l(l1)δl2Dil)(l=2n1l(l1)δl2Djl)l~(l~1)(1δ)l~2==00=00limδ1(l=l~n1l!(ll~)!δll~Dil)(l=l~n1l!(ll~)!δll~Djl)l~!==limδ1(l=l~n1l!l~!(ll~)!δll~Dil)(l=l~n1l!l~!(ll~)!δll~Djl)=(Fil~Fjl~)>0

and by continuity with respect to δ, there is δ¯ such that for all δ(δ¯,1) it holds that DCi(δ)DCj(δ)(1δ)l~>0 or equivalently DCi(δ)>DCj(δ).

Proof of Proposition 5

DCi(δ)=δDi+l=2n1δlDil, which gives DCi(1)=n1 for all i. Therefore, the polynomial DCi(δ)DCj(δ) has a root for δ=1 and another for δ=0, for any pair i,jN. Having observed that, we obtain the following expression.

DCi(δ)DCj(δ)=δ[A1+A2δ++An1δn2]==δ(1δ)[l=2n1All=3n1Alδ(An1+An2)δn4An1δn3]

where A1=DiDj and Al=DilDjl for l{2,,n1}, with the crucial observation being that l=1n1Al=0. This last equation allows us to rewrite the above expression as follows:

(2)DCi(δ)DCj(δ)=δ(1δ)[A1+l=12Alδ++l=1n3Alδn4+l=1n2Alδn3]

From the last expression is apparent that a sufficient condition, though not necessary, to satisfy DCi(δ)>DCj(δ) is that l=1kAl0 for all k{1,,n1}, with at least one inequality being strict. This in turn is equivalent to l=1kDill=1kDjl for all k, which by definition means that Di>UDDj.

Proof of Proposition 6

Let ϵ=1δ and restate decay centrality as follows: DCi(ϵ)=(1ϵ)Di+l=2n1(1ϵ)lDil which gives DCi(ϵ=0)=n1 for all i. Therefore, the polynomial DCi(ϵ)DCj(ϵ) has a root for ϵ=1 (equivalent to δ=0) and another for ϵ=0 (equivalent to δ=1), for any pair i,jN. In addition to this, we consider the binomial identity: (1ϵ)l=k=0l(1)k(lk)ϵk, which allows us to rewrite DCi(ϵ)DCj(ϵ) as follows:

DCi(ϵ)DCj(ϵ)=A1k=01(1)k(1k)ϵk+A2k=02(1)k(2k)ϵk++A1k=0n1(1)k(n1k)ϵk==l=1n1Al=0+l=1n1Al(l1)(ϵ)+l=2n1Al(l2)(ϵ)2++l=n1n1Al(ln1)(ϵ)n1==ϵ[l=1n1Al(l1)+l=2n1Al(l2)(ϵ)++l=n1n1Al(ln1)(ϵ)n2]==ϵ[B1+B2ϵ++Bn1ϵn2]

where B1=FiFj and Bl=FilFjl for l{2,,n1}. Having observed that and knowing that ϵ=1 is also a root of the polynomial, we obtain the following expression via Euclidean division:

DCi(ϵ)DCj(ϵ)=ϵ(1ϵ)[l=2n1Bll=3n1Blϵ(Bn1+Bn2)ϵn4Bn1ϵn3]

with the crucial observation being that l=1n1Bl=0, which is the remainder of the division of the polynomial by 1ϵ, which we know that it should be equal to zero. Hence, similarly to Proposition 5, we obtain the following expression:

(3)DCi(ϵ)DCj(ϵ)=ϵ(1ϵ)[B1+l=12Blϵ++l=1n3Blδn4+l=1n2Blδn3]

It is apparent from the last expression that a sufficient condition, though not necessary, to satisfy DCi(ϵ)>DCj(ϵ) for all ϵ(0,1), hence equivalently for all δ(0,1), is that l=1kBl0 for all k{1,,n1}, with at least one inequality being strict. This in turn is equivalent to l=1kFill=1kFjl for all k, which by definition means that Fj>UDFi.

Proof of Proposition 7

1. DCi(δ)DCj(δ)DiδDjδ+[(n1)Dj]δ2δA1n1Dj. The last inequality holds for all δ(0,1/2] if 2A1(n1)Dj.

2.DCi(δ)DCj(δ)Diδ+Di2δ2Djδ+Dj2δ2+[(n1)DjDj2]δ3δA2+A22+4A1[(n1)DjDj2]2[(n1)DjDj2]. The last inequality holds for all δ(0,1/2] if A1+2A2(n1)DjDj2.

3. The result can also be implied by Rouché’s Theorem, but it is presented here with an independent proof, which follows a similar process to the one used for obtaining Cauchy’s Bound of Polynomial Roots. More specifically, observe that by Proposition 1 together with A1>0 it must hold true that DCi(δ)DCj(δ)=δ(A1+A2δ++An1δn2)>0 for δ close to zero and obviously has a root for δ=0. We also know that it has another root for δ=1. Then it is sufficient to show that DCi(δ)DCj(δ) has no other root for δ(0,1/2) as long as A1|Al| for all l{2,,n1}. To prove this, it is sufficient to focus on the term (A1+A2δ++An1δn2).

|A1+A2δ++An1δn2||A1|(|A2|δ++|An1|δn2)==|A1|(1|A2||A1|δ|An1||A1|δn2)|A1|(1δδn2)=|A1|(1l=1n2δl)==|A1|(1δδn11δ)=|A1|12δ+δn11δ

Therefore, the polynomial has the same number of roots in (0,1) as P(δ)=12δ+δn1, which by Descartes’ Rule of Signs has either zero or two positive roots. One of them is obsviously for δ=1 and and the other is for some δ^<1, because the polynomial has a unique minimum for some δ<1 and is positive for δ=0. Moreover, note that P(1/2)>0 for any n, and in fact P(1/2)0 as n. Hence, the expression is always positive in (0,1/2], which turn means that A1+A2δ++An1δn2 has no root for δ(0,1/2] and in fact it is always positive in that region, by continuity.

4. The proof is identical to that of condition 3, if one considers the alternative expression of difference between decay centralities, DCi(δ)DCj(δ), that is provided by eq. (2).

Proof of Proposition 8

1. Consider again the reformulated expression with ϵ=1δ, for which we know that DCi(ϵ)DCj(ϵ)=ϵ[B1+B2ϵ++Bn1ϵn2], for which it is sufficient to show that it has no root for ϵ(0,1/2] as long as |B1||Bl| for all l{2,,n1}. This is proven identically to condition 3 of Proposition 7. The obtained result, together with the fact that B1<0 ensures that DCi(ϵ)>DCj(ϵ) for all ϵ(0,1/2], which is identical to saying that DC(δ)>DCj(δ) for all δ[1/2,1).

2. The proof is identical to that of condition 1, if one considers the alternative expression of difference between decay centralities, DCi(ϵ)DCj(ϵ), that is provided by eq. (3).

References

Bala, V., and S. Goyal. 2000. “A Noncooperative Model of Network Formation.” Econometrica 68 (5):1181–1230.10.1111/1468-0262.00155Search in Google Scholar

Ballester, C., A. Calvó–Armengol, and Y. Zenou. 2006. “Who’s Who in Networks. Wanted: The Key Player.” Econometrica 74:1403–1417.10.1111/j.1468-0262.2006.00709.xSearch in Google Scholar

Banerjee, A., A.G. Chandrasekhar, E. Duflo, and M. Jackson. 2013a. “The Diffusion of Microfinance.” Science 341 (6144): 1236498.10.1126/science.1236498Search in Google Scholar

Banerjee, A., A.G. Chandrasekhar, E. Duflo, and M. Jackson. 2013b. “The Diffusion of Microfinance.” hdl:1902.1/21538, Harvard Dataverse, V9.10.3386/w17743Search in Google Scholar

Bloch, F., M.O. Jackson, and P. Tebaldi. 2016. “Centrality Measures in Networks.” SSRN Working Paper.10.2139/ssrn.2749124Search in Google Scholar

Bonacich, P. 1987. “Power and Centrality: A Family of Measures.” American Journal of Sociology 92:1170–1182.10.1086/228631Search in Google Scholar

Brandes, U., and T. Erlebach. 2005. Network Analysis: Methodological Foundations. Lecture Notes in Computer Science. Vol. 3418. New York, NY: Springe.10.1007/b106453Search in Google Scholar

Chatterjee, K., and B. Dutta. 2016. “Credibility and Strategic Learning in Networks.” International Economic Review 57 (3): 759–786.10.1111/iere.12175Search in Google Scholar

Dequiedt, V., and Y. Zenou. 2014. “Local and Consistent Centrality Measures in Networks.” CEPR Discussion Paper No. DP10031.Search in Google Scholar

Dijkstra, E.W. 1959. “A Note on Two Problems in Connexion with Graphs.” Numerische Mathematik 1:269–227.10.1007/BF01386390Search in Google Scholar

Erdős, P., and A. Rényi. 1959. “On Random Graphs I.” Publicationes Mathematicae Debrecen 6:290–291.10.5486/PMD.1959.6.3-4.12Search in Google Scholar

Faust, K. 1997. “Centrality in Affiliation Networks.” Social Networks 19:157–191.10.1016/S0378-8733(96)00300-0Search in Google Scholar

Galeotti, A., and S. Goyal. 2009. “Influencing the Influencers: A Theory of Strategic Diffusion.” RAND Journal of Economics 40 (3):509–532.10.1111/j.1756-2171.2009.00075.xSearch in Google Scholar

Gofman, M. 2017 . “Efficiency and Stability of a Financial Architecture with Too–Interconnected–To–Fail Institutions.” Journal of Financial Economics 124 (1): 113– 146.10.1016/j.jfineco.2016.12.009Search in Google Scholar

Jackson, M.O. 2008. Social and Economic Networks. Princeton, NJ: Princeton University Press.10.1515/9781400833993Search in Google Scholar

Jackson, M.O. 2016. “The Friendship Paradox and Systematic Biases in Perceptions and Socal Norms.” SSRN Working Paper.10.2139/ssrn.2780003Search in Google Scholar

Jackson, M.O., and A. Wolinsky. 1996. “A Strategic Model of Social and Economic Networks.” Journal of Economic Theory 71:44–74.10.1006/jeth.1996.0108Search in Google Scholar

Katz, L. 1953. “A New Status Index Derived from Sociometric Analysis.” Psychometrica 18:39–43.10.1007/BF02289026Search in Google Scholar

Kim, J.Y. 2010. “Information Diffusion and δ-Closeness Centrality.” Sociological Theory and Methods 25(1):95–106.Search in Google Scholar

König, M., C.J. Tessone, and Y. Zenou. 2014. “Nestedness in Networks: A Theoretical Model and Some Applications.” Theoretical Economics 9:695–752.10.3982/TE1348Search in Google Scholar

Liu, X., Patacchini E., Y. Zenou, and L. Lung–Fei. 2015. “Criminal Networks: Who is the Key Player?” Working paper.10.2139/ssrn.2089267Search in Google Scholar

Palacios–Huerta, I., and O. Volij. 2004. “The Measurement of Intellectual Influence.” Econometrica 72:963–977.10.1111/j.1468-0262.2004.00519.xSearch in Google Scholar

Pastor–Satorras, R., and A. Vespignani. 2002. “Immunization of Complex Networks.” Physical Review E 65:036104.10.1103/PhysRevE.65.036104Search in Google Scholar

Rothenberg, R., J.J. Potterat, D.E. Woodhouse, W.W. Darrow, S.Q. Muth, and A.S. Klovdahl. 1995. “Choosing a Centrality Measure: Epidemiologic Correlates in the Colorado Springs Study of Social Networks.” Social Networks 17:273–297.10.1016/0378-8733(95)00267-RSearch in Google Scholar

Tsakas, N. 2016. “Optimal Influence under Observational Learning.” SSRN Working Paper.10.2139/ssrn.2449420Search in Google Scholar

Valente, T.W., Corognes K., Lako C., and Costenbader E. 2008. “How Correlated are Network Centrality Measures.” Connect (Tor) 28:16–26.Search in Google Scholar


Supplementary Material

The online version of this article offers supplementary material (DOI:https://doi.org/10.1515/bejte-2017-0010).


Published Online: 2018-01-16

© 2019 Walter de Gruyter GmbH, Berlin/Boston

Downloaded on 10.9.2025 from https://www.degruyterbrill.com/document/doi/10.1515/bejte-2017-0010/html
Scroll to top button