key: cord-0834669-j3ew6j2z authors: Hua, Hao; Hovestadt, Ludger title: p-adic numbers encode complex networks date: 2021-01-08 journal: Sci Rep DOI: 10.1038/s41598-020-79507-4 sha: 81b6436d77193dc22b1764271dc4841dc441b8b3 doc_id: 834669 cord_uid: j3ew6j2z The Erdős-Rényi (ER) random graph G(n, p) analytically characterizes the behaviors in complex networks. However, attempts to fit real-world observations need more sophisticated structures (e.g., multilayer networks), rules (e.g., Achlioptas processes), and projections onto geometric, social, or geographic spaces. The p-adic number system offers a natural representation of hierarchical organization of complex networks. The p-adic random graph interprets n as the cardinality of a set of p-adic numbers. Constructing a vast space of hierarchical structures is equivalent for combining number sequences. Although the giant component is vital in dynamic evolution of networks, the structure of multiple big components is also essential. Fitting the sizes of the few largest components to empirical data was rarely demonstrated. The p-adic ultrametric enables the ER model to simulate multiple big components from the observations of genetic interaction networks, social networks, and epidemics. Community structures lead to multimodal distributions of the big component sizes in networks, which have important implications in intervention of spreading processes. www.nature.com/scientificreports/ • Manipulating the rules for constructing edges. The Achlioptas process picks two candidate edges each time 18 for competitive graph-evolution. ER process, Bohman and Frieze process, and product rule process were compared with one another for analyzing explosive percolation 35 . • Projecting graphs onto social, geometric, or geographic structures. Major developments in that aspect include the hyperbolic networks 36 , spatial preferential attachment 37 , inhomogeneous random graph 38, 39 , and spatial networks 30 . Findings. Common attempts to modeling the hierarchical structures 7 are reflected by various notions including subpopulation, subgraph, mixing pattern, community, and module. We postulate that hierarchical structures are naturally encrypted in a standard graph. Consequently, imposing additional structures onto a graph to enrich its behavior is not always necessary. The key is the p-adic absolute value 40 . According to Ostrowski's theorem, every non-trivial absolute value on Q is either the usual real absolute value or the p-adic absolute value. Various hierarchical structures can be represented by p-adic integers as nodal indices. The network topology and relative strengths between connections are unified as p-adic distances between numbers. The p-adic random graph (PARG), probably the simplest model of inhomogeneous networks, offers a flexible method to simulate various observations in complex networks, especially the phenomenon of multiple big components. Degree distribution is a key property for distinguishing random, free-scale, and small world networks. However, PARG indicates that two random graphs with identical degree distribution may produce significantly different component sizes. We fit PARG to the component size distribution of the genetic interaction networks, and also to the joint distributions of big components in COVID-19 outbreaks. The experiments imply that the community structures are responsible for the multimodal distributions of the sizes of big components. The largest or the second largest component could be more stable at (multiple) specific sizes. Therefore, maintaining a local peak could be valuable for intervening the spreading processes. In contrast to the celebrated ER network, another early prototype of random graph, the Rado graph, has been rarely revisited. The Rado graph employs the binary number system to encode the graph edges, using Ackermann coding of hereditarily finite sets. PARG explores the fundamental nature of integers to encode the probability of connecting a pair of nodes. The p-adic number system extends the ordinary arithmetic of rational numbers 3,40 . Our PARG model will focus on the p-adic metric on nonnegative integers. An integer's r-adic (picking a prime number r) absolute value is the reciprocal of the largest power of r that divides it. For example, |40| 2 = 1/8 (let r=2, then the 2-adic absolute value of 40 equals 1/8), |40| 3 = 1 , |40| 5 = 1/5 . Such absolute value is the most significant example of ultrametrics 40 . Each node is naturally associated with an integer, i.e., its index in the ordered set of nodes, or an arbitrary (unique) integer can be assigned to each node. The probability of connecting any pair of nodes i, j is proportional to the p-adic closeness between the two nodal indices v i , v j : . p * is a constant as the probability in the ER sense. When comparing ER and PARG, we normalize the PARG probability so that ij p ij = n(n − 1)p * /2 . As a result, the number of edges in ER equals that in PARG. The p-adic distances encode a hierarchical structure, as shown in the circular tree-map in Fig. 1a . The distances between any pair of numbers from the same small circle and the same big circle are 1/9 and 1/3, respectively. If the two numbers are from different big circles, their distance is 1. The digit format of a p-adic number is intuitive, for example, 201 3 = 2 · 3 0 + 0 · 3 1 + 1 · 3 2 = 11 . One can construct a set of integers (as the nodal indices) in their digit formats by where r is the chosen prime number. We call the model a full PARG if u 0 = u 1 = · · · = u m = r (i.e., n = r m+1 ), otherwise, we call it a regular PARG (i.e., n = i u i ). The nodal indices can also be arbitrary digits under a given prime r, which leads to a general PARG. The expression (2) facilities the enumeration of hierarchical structures. For example, 3.2.3 3 fully describes a hierarchical structure as shown in Fig. 1c . More examples are in illustrated in Fig. 1 . A notation such as G(3.2.3 3 , p * ) fully specifies a PARG. PARG implements a Bernoulli process on all pairs from a set of p-adic numbers. Let p k represent the probability of a randomly chosen node with degree k. In the ER model, p k follows a Binomial distribution. It becomes a Poisson distribution in the limit of large n. By contrast, n in PARG equals the number of individuals in observations. Because of the symmetric connectivity in regular PARG, p k can be obtained from the degree distribution of one node: where m i=0 d i = n − 1 holds. d i denotes the number of links (from the chosen node) with probability r i p * according to (1) (2) . The numbers α 0 , α 1 , · · · , α m denote all combinations satisfying k = m i=0 α i . Numeric computing of (3) indicates that p k in PARG and that in ER are almost identical. The two models can have the same degree distribution and the same number of edges. However, there could be significant differences in the size distributions of big components in PARG and ER (Fig. 2 , Supplementary Note 1, Table S1 and S2). The largest www.nature.com/scientificreports/ www.nature.com/scientificreports/ components in PARG may exhibit multimodal distribution due to the hierarchical structure. This implies that certain sizes of (multiple) big components are more statistically stable than other sizes. The relative sizes of multiple big components in complex networks were much less studied compared to the studies carried out on the giant component 12, 41 . Analytical methods 11, 42 and generating functions 10 have been widely employed for analyzing the component sizes. Rather than let n → ∞ , n in PARG is equal to the number of relevant individuals in observations. As a result, the sizes of simulated components are similar to that in ground truth. When n is finite, numeric random realizations are suited to evaluate the various probabilities about component sizes (see "Methods" section). When n is fixed, fitting ER model to empirical data only involves the single parameter p, while PARGs involve two kinds of parameters: the probability p and the hierarchical structures represented by (2) . The configuration space of (2) is very vast, even when n < 1000 , so we opted for ad hoc heuristics to choose the hierarchical structures that fit relatively well to observations. The heuristics includes: (1) scaling the distances. For example, 7.5.6 r with r = 7, 11, 13, · · · represent the same hierarchical structure, though the distances between the levels are scaled. (2) Based on the hierarchical structures in PARG, the following experiments analyze multiple big components, especially 5|C 2 | > |C 1 | , as static structures observed in networks. The topics range from microscopic networks, such as biological networks 26, 28 , to macroscopic networks, such as epidemics 21, 43 . been lately revealed 25 . The essential genetic interaction network of yeast genes (theCellmap.org) contains 1,261 mutant strains. Their interactions have been characterized by Pearson correlation coefficient (PCC). Genes with highly correlated genetic interaction profiles (PCC>0.4) form clusters of specific pathways or protein complexes 26 . We set three PCC thresholds above 0.4 to obtain three graphs with distinct big components, as shown in Fig. 3g -i. We count the components sizes falling into the predefined intervals of Fibonacci numbers . Let θ i be the number of simulated components whose sizes fall between b i and b i+1 , and θ * i be that in ground truth. The error of a random realization is given by The averaged error from many random realizations yields a relatively accurate evaluation. ER and PARG with distinct values of np lead to different errors (Fig. 3a-c) . Equipped with (configurable) hierarchical structures, PARGs fit better to observations than ER. The component size distributions of the best fits are shown in Fig. 3d −f. Exploring the protein interaction networks of proteins poses a major challenge in biomedicine. Protein-protein interaction (PPI) is crucial to understanding cellular pathways and human diseases 44 . The following experiment creates a graph from a set of 408 S. cerevisiae protein complexes as 45 . The graph nodes represent individual proteins from these complexes. An edge is constructed between two nodes (proteins) if they belong to the same complex. The graph includes 1,628 nodes and 11,249 edges. We define the similar metric S duo and S tri (see "Methods" section) to compare the simulated component sizes with the ground truth. PARG fits better ( S tri =0.307) than the ER model ( S tri =0.112) to the PPI network, as shown in Fig. 4 . The simulation data can be found in supplementary Table S3 and S4. It means that the chosen PARG has a higher probability that the sizes of its big components resemble those in the PPI network. Instrumental resource street network. There has been growing attention to the impact of social networks on health 46 . For example, homeless youth is an active research field, including analysis and interventions. Social networks of homeless youth 47, 48 are vital for understanding and intervening the observed phenomena. This experiment involves a social network about employment services utilization among homeless youth. The original research 49 queried 136 homeless youth in Los Angeles in 2008. Four distinct networks were constructed from the same population, according to instrumental, emotional, employment services use, and sociometric relationship respectively. Only the instrumental network ( |C * 1 |=30, |C * 2 |=13) satisfies 5|C * 2 | > |C * 1 | , i.e., the second largest component is large enough. Regarding the similarity metric S duo , the ER model attains the maximum similarity ( S duo =0.155) at p = 1.08/n ; while G(10.14 353 , 1.75/n) has a much higher similarity ( S duo =0.316). The community structure in the PARG corresponds to the social or geographical networks of the homeless youth; although the map between the two is still elusive. We also fit the hyperbolic networks 36 and the Achlioptas process 35 to the observed component sizes ( |C * 1 |=30, |C * 2 |=13). Details can be found in "Methods" section. The random hyperbolic graph 50,51 reaches the maximum similarity ( S duo =0.266) when C = −9.3 , α = 10 , D = 0.0684R . The Achlioptas process with product rule (PR) has the maximum similarity ( S duo =0.221) when the number of edges is equal to 100. So PARG outperforms the other two models in this case. Coronavirus 52 has spread among many Chinese cities since the end of January 2020. The incubation period 53 and possible mild symptoms 54 made the prevention more complicated. Social networking sites (or local officials) reported traces of infected people. Relationships between the infected (and those who had close contacts with them) were also investigated. From a point of view of networks, the cities Table 5 ), which consists of multiple big components. A graph is created to visualize the relationship between the infected when the outbreak was around its peak, as shown in Fig. 5 . The simulation data can be found in Supplementary Table S6 and S7. The similarity metric S tri could be biased when |C * 3 | is quite smaller than |C * 1 | , so the similarity metric T tri (see "Methods" section) is employed in this case to fit the models to the observed clusters, as shown in Fig. 6 . The PARG G(8.17 59 , 1.49/n) has the highest similarity T tri =0.00916. We also compared the results of the random hyperbolic graph and the Achlioptas process with that of PARG. The random hyperbolic graph reaches the maximum similarity ( T tri =0.0150) when C = −9.65 , α = 1 , D = 0.04294R . The Achlioptas process with Bohman Frieze (BF) rule has the maximum similarity ( T tri =0.00825) when the number of edges is equal to 93. So the random hyperbolic graph fits best to this case. www.nature.com/scientificreports/ A later investigation indicates that the 11 cases (in yellow, Fig. 5 ) are probably related to the department store as well. In this new perspective, the two big clusters form the largest component of 55 nodes. We employ the metric T duo to fit modes to this new observation ( |C * 1 |=55, |C * 2 |=17), as shown in Fig 7. The simulation data can be found in Supplementary Table S8 and S9. The great variety of hierarchical structures enable PARG to fit relatively well to observations from distinct perspectives. The size distributions of big components in complex networks are attributed to the structure of physical world; the behaviors of agents (nodes) and the information transmitting between them; and the observer (how to look at the events). The ER model offered prominent findings of component sizes in networks, however, it rarely fits the joint distribution P(|C 1 | ≈ |C * 1 | , |C 2 | ≈ |C * 2 | , · · · ) in real observations. A successful strategy is introducing inhomogeneous structures or selective rules (for constructing edges) to the random graph to increase its versatility. PARG probably provides the simplest way to fully describe a hierarchical structure in an ER-like model. PARG interprets the n in ER model as the cardinality of a set of natural numbers (nodal indices). Consequently, the probability p can be weighted by distances between the nodal indices. In number theory (Ostrowski's theorem), any non-trivial definition of absolute value on Q is either the conventional one or the p-adic absolute value. So, the p-adic ultrametric reveals the natural hierarchical structures hidden in any graph with indexed nodes. PARG blurs the boundaries between the topology approaches (e.g., multiplex networks) and the geometric approaches (e.g., hyperbolic networks). In our PARG approach, n denotes the number of observed individuals, whereas in previous ER-like models n → ∞ . The limit of n facilities analytical approaches 10, 11 , while a relatively small finite n is convenient for numerical random realizations. Random graph theories such as explosive percolation 35 and synchronization 55 , deeply revealed the dynamics of the emerged components. By contrast, this work explores the sizes of resultant (static) components from observations or simulations. The results imply that the proportions between the big component sizes are closely associated with the hierarchical structures of complex networks. This implication is in contrast to previous emphasis 10,27,56 on degree distribution as the fingerprint of network structures. We fit PARGs and other random graphs to observations from various types of networks. The PARG outperforms the ER model and the Achlioptas process. The random hyperbolic model fits better than PARG to certain cases but worse than PARG to other cases. The simulations of PARG show that the size of big components, e.g., P(|C 1 | = x) and P(|C 2 | = x) , can exhibit multimodal distribution (e.g., Figs. 2b and 7(b) ) due to their modular structures. In the case of multimodal distribution, the first peak of P(|C 2 | = x) could be very close to the last peak of P(|C 1 | = x) . Thus, one may distinguish the major mode from the minor modes when investigating the giant component. One present challenge in network epidemic modeling 57 www.nature.com/scientificreports/ P(|C i | = x) has implications in controlling the spreading processes in networks. Since P(|C i | = x) has more than one local peak, it might be possible to predict and maintain the big components' growth around a local peak. Random graph implementation. All random graphs are created through random experiments implemented in the Java programming language (Java 1.8 with Eclipse IDE). Given a connecting probability p (in the ER context), an edge is included in the graph if where rand stands for a random number between 0 and 1, generated by the method nextDouble() from the java. util.Random class. The method generates a stream of pseudorandom numbers via linear congruential generator (LCG) with modulus 2 48 . A disjoint-set data structure (Union-Find algorithm) is employed to find all components in graph. N random realizations of ER or PARG yield an ensemble of binary numbers for x = 1, 2, · · · , n . |C j i | denotes the size of the ith largest component in the jth realization. Then one can evaluate the size distribution of the big components by One can choose an a prior function to measure the similarity between the simulated components sizes and the observed sizes. For example, Poisson distribution can be used to define the similarity between |C j i | and the observation : where s is a constant (typically s = 0.1 , so that 10% deviation from the truth results in 1/2 similarity). Simulations in this work employed the later definition. The probability distribution of |C 2 | under the condition |C 1 | ≈ |C * 1 | is given by . We define the objective function for a random graph to measure whether its first two/three largest component sizes are close to that in observation. When |C * 2 | or |C * 3 | is much smaller than |C * 1 | , the following metric would be more appropriate: Genetic interaction network of yeast genes. The data of yeast genes is from https://thecellmap.org/ costanzo2016/. The networks in Fig. 3g -i are drawn from the data of the Essential × Essential network, which involves 1,261 mutant strains. So, each network in Fig. 3g -i consists of 1261 nodes. The PPC values between the mutant strains are obtained from the genetic interaction profile similarity matrices. An edge is included in the graph if the corresponding PPC value is above the predefined threshold. The graphs are projected onto squares using our Java program, as shown in Fig. 3g -i. In a dynamical process, the agents (nodes) push away from each other, while each edge drags the two end nodes into a fixed range. The color (hue) of edges indicates the size of the relevant component. Protein network of yeast genes. The data of S. cerevisiae protein complexes is obtained from the additional File 1 of 45 . We programmed a Java application to read the table, construct the 1,628 nodes (proteins) and 11,249 edges (a pair of nodes belong to a same complex), and find the components via Union-Find algorithm. Random hyperbolic graph. Our implementation follows the formulation in 50, 51 . R = 2 ln n + C denotes the radius of the disc. The probability density for the radial coordinate r of a point (r, φ) is given by We use inverse transform sampling to generate the radii of the points, i.e., r = 1 a arcosh(1 + cosh(αR)x − x) where x ∈ (0, 1) denotes a random number from the uniform distribution. Our experiments generate x by nextDouble() in Java. For each pair of nodes u, v, a link is added to the graph if d(u, v) < D where D ∈ (0, R] is a constant and d(u, v) denotes the distance between the two points in the hyperbolic space. Achlioptas process. At each time step of the Achlioptas process 18,35 , two edges e 1 and e 2 compete for addition. Suppose e 1 involves two components of size |C a |, |C b | ; e 2 involves two components of size |C c |, |C d | , we consider three types of competing rules: • add e 1 if |C a | + |C b | < |C c | + |C d | , add e 2 otherwise. www.nature.com/scientificreports/ A number theory approach to random graph is proposed. A set of n random numbers generates an n by n adjacency matrix whose binary elements follow the probability (1). Thus, a hierarchical structure is implemented through ultrametrics 40, 58 . The simplicity of the digit form (2) for hierarchical structures of random graph facilitates the enumeration of different setting of clusters and hierarchies. In contrast to mapping complex structures or rules from real world to random graph, our PARG approach explores the complex structures in numbers 2,59 which might be rich enough for modeling complicated observations. An alternative point of view suggests that a plain graph can unfold a hidden hierarchical structure, based on an indispensable definition of absolute value. The proposed PARG model is more abstract than multilayer networks, multiplex networks, and social-geographical models, but more concrete than ER-like models without community structures. Therefore, a future framework of research would consist of two interconnected levels: (1) Searching for hyper parameters, such as hierarchical structures, in PARG, given empirical data and (2) Constructing the ad hoc realistic models, for example, the biomolecular environments in cell or the social-geographical structures in a city, which the PARG model is projected onto. Received: 19 May 2020; Accepted: 8 December 2020 Divisibility patterns of natural numbers on a complex network Multiplex congruence network of natural numbers Application of p-adic wavelets to model reaction-diffusion dynamics in random porous media Protein ultrametricity and spectral diffusion in deeply frozen proteins p-adic hierarchical properties of the genetic code Communities, modules and large-scale structure in networks Modelling dynamical processes in complex socio-technical systems The statistical physics of real-world networks Random graphs with arbitrary degree distributions and their applications An old approach to the giant component problem Revealing the microstructure of the giant component in random graph ensembles From networks to optimal higher-order models of complex systems An introduction to exponential random graph (p*) models for social networks Stochastic blockmodels and community structure in networks Community detection and stochastic block models: recent developments Explosive percolation in random networks Recurrence-based time series analysis by means of complex network methods Epidemic processes in complex networks Epidemic spreading on complex networks with community structures Network analysis in public health: history, methods, and applications Chains of affection: The structure of adolescent romantic and sexual networks The human disease network Architecture of the human interactome defines protein communities and disease networks A global genetic interaction network maps a wiring diagram of cellular function Network biology: Understanding the cell's functional organization Drug-target network Structure and dynamics of molecular networks: A novel paradigm of drug discovery: A comprehensive review Spatial networks Spreading processes in multilayer networks The physics of spreading processes in multilayer networks Towards real-world complexity: an introduction to multiplex networks Hidden geometric correlations in real multiplex networks Anomalous critical and supercritical phenomena in explosive percolation Hyperbolic geometry of complex networks Spatial preferential attachment networks: Power laws and clustering coefficients General formalism for inhomogeneous random graphs Geometric inhomogeneous random graphs p-adic mathematical physics: the first 30 years. p-Adic Number Emergence and size of the giant component in clustered random graphs with a given degree distribution The size of the giant component of a random graph with a given degree sequence Networks and the epidemiology of infectious disease Fundamentals of protein interaction network mapping Markov clustering versus affinity propagation for the partitioning of protein interaction graphs Social networks and health Social networks of homeless youth in emerging adulthood Who is supporting homeless youth? predictors of support in personal networks Social networks as the context for understanding employment services utilization among homeless youth Random hyperbolic graphs: degree sequence and clustering Cliques in hyperbolic random graphs Clinical characteristics of coronavirus disease 2019 in China The incubation period of coronavirus disease 2019 (covid-19) from publicly reported confirmed cases: estimation and application Clinical features of patients infected with 2019 novel coronavirus in Wuhan Explosive transitions in complex networks' structure and dynamics: Percolation and synchronization Controllability of complex networks Eight challenges for network epidemic models Trees, ultrametrics, and noncommutative geometry Complex architecture of primes and natural numbers The authors declare no competing interests. Supplementary Information The online version contains supplementary material available at https ://doi. org/10.1038/s4159 8-020-79507 -4.Correspondence and requests for materials should be addressed to H.H. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat iveco mmons .org/licen ses/by/4.0/.