Abstract
Many complex networks expose global hub structures: for some nodes, the number of incident edges far exceeds the average, leading to a small average shortest path length. Such ‘smallworld properties’ are often guided by a scalefree powerlaw distribution of the node degrees, and selforganization inside the network has been identified as a reason driving the emergence of this structure. Smallworld networks have recently raised lots of interest, because they capture the global topology of the WorldWide Web, metabolic, and social networks. While smallworld networks reflect global structures, little attention is paid to the local structure of complex networks. In this article neighbourhoods are demonstrated to share a common local structure in many real complex networks, manifested by a polynomial volume law. This law can, in case of networks that are embedded in space, be explained in terms of the embedding and the properties of Euclidean space. A model of hierarchical spatial networks is introduced to examine the effect of global structures, in particular of hierarchies, on the polynomial volume law. It turns out that the law is robust against the coexistence of such global structures. The local structure of space and global optimization can both be found in transport, brain, and communication networks, which suggests the polynomial volume law, often in combination with hierarchies or other global optimization principles, to be a generic property inherent to many networks.
Introduction
Networks are used to describe and analyse systems that expose relations between objects. The increasing availability of network datasets – e.g., communication and social networks^{1,2}, transport and road networks^{3,4}, computer networks, and the WorldWide Web^{5} – has rendered possible a structural comparison of their topological features, which can be shared across different topics and types of network. The distribution P(k) of the node degree k has, e.g., been shown to follow a power law P(k) ~ k^{−γ} in many networks, a property that is also called scalefree^{6,7}: many nodes have a low number of incident edges, while only very few nodes have a high number of these and thus act as ‘hubs’, connecting many different nodes. Shortest paths contain only very few nodes in average, when many hubs are present in a network or the network exposes a strong hierarchy; a network is said to expose smallworld properties^{2,8} in case of such a short average shortest path length.
Many networks, such as road networks, do not expose such smallworld properties. Instead of a short average shortest path length, the average shortest path length of a road network grows much faster for an increasing size of the network; and the node degree is usually limited by 4 or 5. This is despite the fact that road networks are optimized for travel time or travelled distance. Smallworld networks, in contrast, expose a low average number of nodes on shortest paths without taking some metric, e.g., travel time or travelled distance, into account. Many different factors constrain the generation of road networks, but there is a major principle that governs their generation: nodes that are close in space are much more often adjacent in the network than distant nodes. This autocorrelation of the road network is also known as Tobler’s first law of geography when applied to geographical information in general^{9,10,11}.
The principle of nodes in a neighbourhood being connected by edges has a local optimization effect in respect to shortest paths in the network, resembling the realm of Zipf’s principle of least effort^{12}. Here, local optimization refers to the fact that the optimization can be executed on smaller neighbourhoods inside the network, while global optimization on the contrary always refers to the entire network. The global effect of hubs has, e.g., extensively been discussed in literature, among others by Watts and Strogatz^{2} and by Barabási and Albert^{6}. The following investigates the coexistence of local optimization, caused by the principle of connected neigbourhoods, and global optimization, here exemplified at hierarchies. In particular, I introduce a novel hierarchical spatial network model, which prototypically resembles the coexistence of local spatial structure – the local constellation of edges – and global hierarchies. The volume law^{13,14}, which is a result of the local spatial structure, can easily be traced in a network and allows for a reconstruction of the dimension of space. The resulting concept of the dimension of a network is compared to other concepts of fractal dimension. Thereby, the influence of the spatial layout of the network on the evaluation of the dimension is discussed. It is examined at the introduced hierarchical spatial network model in how far local optimization masks the polynomial volume law and impairs the estimation of the dimension. The results are set into context by the analysis of a number of realworld networks, which expose both local and global optimization.
Results
Space shaping networks: the polynomial volume law
Networks are in many cases shaped by space: the edges of the network, which represent thematic information, relate in some way to distance in space. If this relation between network and space is strong, both expose similar characteristics. In this case, the network is called to be spatial. We can accordingly expect the size of a neighbourhood in the network to depend on its ‘radius’ in a similar way than the volume of a ball does in Euclidean space. Next, we concretize the concept of a neighbourhood in a network, which is called ball in this case. The ball B_{n}(r) of radius r centred at a node n is defined as the nodes within distance r of the node n, i.e., as the nodes that can be reached by traversing at most r edges or, in a weighted network, as the nodes that can be reached by traversing edges of total weight not more than r. The volume B_{n}(r), in turn, is defined as the number of nodes contained in B_{n}(r). Now assume a network that is embedded in an Euclidean space. The volume of a ball in Euclidean space scales as r^{d}, where r denotes the radius and d the dimension of the embedding space. The volume of a ball in the network can be expected to scale in a similar way, in case of the edges of the network being related to the distance between the nodes. In fact, many realworld networks statistically expose the polynomial volume law (Fig. 1a–p):
where k and d are some positive real numbers. The left side of the polynomial volume law (Equation 1) refers to the volume in the network, while the right side refers to the volume of a ball in Euclidean space incremented by 1, reflecting that the ball of radius 0 in the network contains exactly one node. This law has been discussed previously by Song et al.^{13} and Shanker^{14} but has to my knowledge never been examined in detail with respect to hierarchies inside the network.
When statistically fitting the volume of a ball for different radii r in a network by Equation 1, the parameter d can, in contrast to Euclidean space, be a noninteger. This is in particular the case if the network does not equally ‘extend’ to every dimension of the embedding space. While a meaningful embedding in space is able to explain why many networks expose the polynomial volume law, such an embedding is not needed to compute the volumes B_{n}(r) for different nodes n and different radii r, and, in turn, for determining whether a network follows the law and which real number d – called the dimension of the network – fits best.
Comparison to the fractal dimension
The dimension derived by the polynomial volume law is in many aspects similar to other approaches that relate a network to the dimension of the space it is embedded in. Most notable, the box counting dimension, also called MinkowskiBouligand dimension or fractal dimension, establishes a relation between the embedding space and the network by comparing the complexity at different scales^{13,15}. Thereby, space is tessellated with a grid of boxes and the number of boxes containing at least one node (or alternatively, the number of boxes intersecting at least one edge of the network) is determined. As a result, one is able to conclude the dimension by the relation between the number of such boxes and their side lengths. The box counting dimension has been discussed in various articles, among others, in respect to self similarity in networks^{16,17,18}. Efficient algorithms for the computation of this dimension have been published^{19}. A comparison of such algorithms has been provided by Song et al.^{20}. Even the idea of the box counting dimension has been subject to advancements^{17,21}. The box counting dimension has been discussed in various contexts, among others, in the geographical context^{22,23}.
Approaches similar to the dimension defined by the polynomial volume law have been discussed in literature. For instance, Daqing et al.^{24} have considered the average Euclidean distance in space E_{n}(r) from a centre node n to all nodes inside a ball B_{n}(r), i.e., to all nodes that can be reached by traversing at most r edges of the network. This average Euclidean distance has been compared to the volume of the ball in the network, as defined previously^{24}. Thereby, a number referred to as the dimension is assigned to the network, much similar as in case of the polynomial volume law. The comparison of the volume of a ball B_{n}(r) to the average distance E_{n}(r) instead as to 1 + k · r^{d} has two major consequences. First, the average distance E_{n}(r) explicitly includes the concept of Euclidean distance, which presumes the network to be explicitly embedded in an Euclidean space. The comparison to 1 + k · r^{d} can though also be performed for an abstract network, without any knowledge about the potential location of a node. Secondly, the comparison of the volume of a ball B_{n}(r) to the average distance E_{n}(r) examines how topological and Euclidean aspects of the very network relate, while the comparison to 1 + k · r^{d} how the topological aspects of the network relate to the universal polynomial law that describes the Euclidean volume of a ball in general. In short, the considerations of Daqing et al.^{24} include an explicit Euclidean embedding of the network, while the polynomial volume law 1 + k · r^{d} only compares to Euclidean spaces in general.
Further approaches exist to characterize networks by their dimension. Daqing et al.^{24} examine the root mean square displacement by a random walk. Song et al.^{13} have pointed out that the different estimations of the dimension of space do not coincide in some cases, e.g., in case of smallworld networks.
Figure 2 compares different types of network dimensions for two realworld networks, the Bus network of Manhattan and the Metro North Railroad in NY. The figure shows the dimensions resulting from the polynomial volume law in four variants. First, the volumes by the distance in the network are determined by the distance in an unweighted network. Secondly, the distance in Euclidean space between two adjacent nodes is used as weight, and the volumes are computed for the weighted network. Thirdly, the distances between adjacent nodes is computed in the embedded network, i.e., the weights correspond to the distance a bus or train needs to travel. Fourthly, travel times are used as weights. In addition to these dimensions resulting from the polynomial volume law, the box counting dimension is computed by counting the boxes that contain a node of the network, or by counting the boxes that intersect an edge of the network.
The different concepts of dimension result in different values, as can be seen in Fig. 2. The estimated dimensions by the volumes in the weighted network are very similar, both in the example of the bus network as well as of the railroad network. In the unweighted network however, the estimation of the dimension is higher and is subject to a large standard deviation. In the case of the Metro North Railroad, a fit can hardly be made because the effective diameter of the network is small, which is reflected by the small range of the fit. The box counting dimension provides lower estimations when referring to nodes compared to when referring to edges. Both, variants of the box counting dimension provide lower estimates than the polynomial volume law, which is much likely an artefact of the dimensions to reflect different concepts: the box counting dimension compares complexity at different scales while the polynomial volume law carries over concepts from Euclidean space to the network. Despite of this difference, the box counting dimension is higher in case of the Bus network compared to the Metro North Railroad, which is consistent with the polynomial volume law.
Local and global optimization principles
Many generation principles are known to guide the emergence of networks. Among them are principles that avoid edges between distant nodes in space, leading to a large diameter of the network, as well as principles that minimize the average distance between the nodes of the network and thus lead to smallworld networks. In the following, we discuss factors that lead to these principles and how they relate.
The polynomial volume law is often the result of a local optimization principle: assuming that the costs of an edge depend on its length, how can a node be adjacent to as many nodes of the network as possible? This principle is of local nature because it can be answered independently for each node. In the resulting network, a node is obviously adjacent to the nodes of its neighbourhood in space while being nonadjacent to more distant nodes. This local optimization principle has been resembled by different models. An approach is to introduce edges with a probability that depends on the distance of the nodes in space, e.g., with a probability of P(l) = α exp(−l/l_{0}) with positive values α and l_{0}^{25}, or with a probability of P(l) = 1 if l < l_{0} and P(l) = 0 otherwise^{26}. Another model, which we refer to as the spatial network model or Mocnik model, has been proposed by Mocnik^{27,28}. Assume a number of nodes being embedded in space. We then introduce a directed edge (n_{1}, n_{2}) if and only if
where dist denotes the Euclidean distance and ρ > 1 a parameter that influences the density of the network (Fig. 3a). The model prototypically resembles Tobler’s first law of geography: ‘everything is related to everything else, but near things are more related than distant things’^{9,10,11}. Despite this, the model applies to other scales than the geographical scale as well.
The Mocnik model follows the polynomial volume law. When the nodes are randomly distributed in space with a uniform distribution, the edges introduced by the model reflect properties of space, e.g., the existence of proximity. As a consequence, the number of edges is expected to be linear in the number of nodes^{27}; and the dimension of space has an impact on the configuration of the edges. In fact, the volume of the undirected network associated to this model follows the polynomial volume law (Fig. 3b), in which the exponent d resembles the dimension of space. The parameter ρ determines the density of the network, i.e., the ratio of the number of actual edges to the maximal number of edges in a simple network. Thereby, ρ has an impact on the configuration of edges as well, but in a model with an infinite number of nodes ρ does not influence the exponent d when fitting by the polynomial volume law. Even in middle and large size networks, the influence of ρ does practically not mask the impact of the dimension^{28}. The Mocnik model – a network embedded in space with only shortdistance edges – can thus serve as an explanation of the polynomial volume law by local optimization.
In contrast to the local optimization principle that maximizes the number of adjacent nodes, global optimization principles often play a major role: assuming that a network shall only contain a limited number of edges, how can the average distance between pairs of nodes statistically be minimized for the entire network? This principle does not refer to some nodes only rather than to the entire network. If a network complies with this optimization principle, most shortest paths between two randomly chosen nodes are, in fact, very short, but single nodes may suffer from a longer distance to large parts of the network. Among the models that create such smallworld networks are the WattsStrogatz model^{2} and the BarabásiAlbert model^{6}.
Realworld networks are often organized by both local and global optimization principles. Local optimization principles naturally occur when the costs of an edge positively correlate to its length, which is the case for physical networks (road and railway networks, etc.) but also for many types of communication networks (network of postal delivery services, the telephone network, etc.). Global optimization, in contrast, often minimizes the average length of shortest paths by introducing edges between distant nodes in space. Such global principles naturally occur for networks that are, at least in large parts, of virtual nature, e.g., to friend networks in social media. Most networks are though guided by a combination of local and global optimization to achieve a balance between costs and the length of shortest paths in the network. In the remainder of the article, we explore the interaction between local and global optimization and discuss its effect on the polynomial volume law.
A model of hierarchical spatial networks
The Mocnik model^{27} is guided by a local optimization principle, as becomes apparent by Equation 2. In order to study the interaction between local and global optimization principles, we extend the Mocnik model in the following to a hierarchical Mocnik model. Thereby, the hierarchical model aims at including a global optimization principle by introducing different layers in the network. This hierarchy is, as we show later, to some extent compatible with local optimization principles. If the layers of the hierarchy share nodes, i.e., if they are connected, shortest paths in the network become shorter in comparison to the nonhierarchical model, because shortest paths often traverse higher layers of the hierarchy, which are more efficient in bridging space.
Hierarchies and the principle of layered networks can be found in many transport networks. For instance, many road networks expose layers: motorways, primary, secondary and tertiary roads, residential roads, etc. Railway networks often consist of longdistance and of local trains, the former which usually have less stops and are much faster than the latter. The shortest route in a railway network incorporates thus often a local train to a larger station, then longdistance trains, and potentially another local train. The universal nature of this principle has been widely recognized, and important routing algorithms take thus advantage of hierarchies inside the data^{29}.
The hierarchical Mocnik model makes use of the nonhierarchical model in every layer of the hierarchy. Assume a number of node sets \({N}_{l}\subset {N}_{l1}\subset \ldots \subset {N}_{0}\) to be embedded in space, which each correspond to one layer of the network. Then, for each layer consisting of nodes N_{i}, edges E_{i} are created in accordance to Equation 2. The nodes N_{i} of a layer together with the corresponding edges E_{i} are, accordingly, a Mocnik model. The lowest layer N_{0} will in the following be referred to as the base layer of the network.
Local and global optimization coexist in the hierarchical Mocnik model. The base layer is guided by local optimization in the same way as the nonhierarchical Mocnik model: in each neighbourhood, the constellation of edges is optimized for a high number of adjacent nodes. In addition to the nonhierarchical model, the layers of the hierarchical variant expose different degrees of local and global optimization. The less nodes a layer contains, i.e., the higher the layer in the hierarchy, the more global the optimization becomes. The optimization in a higher layer of the hierarchy only involves some nodes of the network while ignoring many other ones, which means that the optimization is not any longer performed in spatial neighbourhoods.
Even a weighted variant of the hierarchical Mocnik model can be introduced, in which the edges are complemented by weights. Thereby, the weight of an edge corresponds to its lengths in Euclidean space. The introduction of a new layer to an existing Mocnik model makes shortest paths potentially shorter even in case of the weighted model. While the nodes stay untouched, new edges are introduced in each layer of the network but none is removed. Accordingly, some nodes are directly connected in some layer E_{i} while the shortest path in E_{j} with j < i is potentially longer – triangle inequality applies.
The weights of the weighted variant of the hierarchical Mocnik model can even be systematically adjusted in respect to the hierarchy. The weights w_{i} for each layer of the hierarchy reflect that the layers are of different speed, require different communication costs, etc. In the weighted hierarchical Mocnik model, the weight of an edge in layer i is defined as the length of the edge in Euclidean space, multiplied by w_{i}. If all w_{i} are equal to 1, the weights are, accordingly, equal to the length of the edges. The resulting network usually consists of many edges with low weights in the base layer, and only some edges with slightly greater weights in higher layers (Fig. 3d). Such a weighted hierarchical model is very similar to transport networks, in which local transport connects adjacent places, and more distant places are connected by motorways or longdistance trains operating at a higher speed.
Synthesis of local and global optimization principles
The hierarchical Mocnik model is characterized by local optimization in each layer and global optimization by hierarchies. In case of one layer only, the model is prototypically characterized by local optimization. In case of several layers there exist shortcuts, which expose characteristics of global optimization and lead to smallworld properties in the network. Here, we examine the impact of coexisting local and global optimization principles on the polynomial volume law.
The volume of a nonweighted hierarchical network is always larger than the volume in the base layer of a network. In fact, the volume increases when a layer with 3000 nodes (flat hierarchy) is introduced, or when a layer with 100 nodes (steep hierarchy) is introduced on top of the base layer in the example of Fig. 3e. For smaller radii, the volume is larger for flatter hierarchies, because more nodes are adjacent to the higher layer and shortest paths between two nodes of the same small spatial neighbourhood more often traverse nodes of a higher layer. At larger radii, the increase in volume is, though, larger for steeper hierarchies, because the shortcuts introduced by the hierarchy are more efficient. If several layers are added on top of each other, the increase of volume at smaller radii is guided by the lower layers of the hierarchy, and the increase of volume at larger radii is guided by the higher layers of the hierarchy.
While the influence of the hierarchies are obvious in case that the base layer of the network is known, such comparisons can hardy be drawn in general. Instead, we may ask how the hierarchies affect the measured volumes in comparison to the fit (to the polynomial volume law), because the difference between the fit and the actual data can be examined without any knowledge about prevailing layers. In fact, the fit underestimates the volume and overestimates the exponent d, the dimension, in different ways. For a steep hierarchy with much less nodes in a higher layer than in the base layer, the fit underestimates the volume at smaller radii (Fig. 3e). This effect is independent of whether there exist additional layers in the hierarchy in case of a nonweighted model, i e., the number of nodes in the highest layer of the hierarchy has a major impact on the underestimation. At the same time, the rate of growth is higher in case of a steep hierarchy for larger radii, leading to higher estimates of the dimension. If the hierarchy is flatter, the estimated dimension is lower than in case of a steep hierarchy but higher than for the base layer alone.
The underestimation of the volume and overestimation of the dimension can also be observed in case of the weighted hierarchical Mocnik model (Fig. 3f). The effect is though less significant because the lengths of the edges is taken into account, and higher layers provide less effective shortcuts than in the nonweighted model. The presence of a layer with more nodes can even obfuscate the effect of a layer with much less nodes in case of the weighted hierarchical Mocnik model. The fact that both kinds of hierarchical Mocnik models follow a polynomial volume law, despite being layered networks with several hierarchies, suggests that the polynomial power law is robust and not necessarily masked by other structures inside the network.
Discussion
The examination of the Mocnik model has demonstrated that global optimization leads to an underestimation of the volume at lower radii and an overestimation of the dimension when fitting to the polynomial volume law. Despite having only examined hierarchical structures as an example of global optimization, the underestimation of the volume and the overestimation of the dimension origins from the existence of shortcuts in the network, which suggests that other global optimization principles have similar effects. Here we study realworld networks in order to trace the effects of coexisting local and global optimization at real examples.
Examples of spatial networks
The considered transport networks follow a polynomial volume law with an exponent d between 1.2 and 2.5 (Figs 1a–i and 3c). The bus network of Manhattan, e.g., is highlybranched, the exponent d = 2.13 ± 0.38 reflecting that the network exhausts the twodimensional space of Manhattan (Fig 1a, b, d). The Metro North Railroad in NY is, in contrast, a lowlybranched railway line with exponent d = 1.26 ± 0.15, which reflects that it does not exhaust space (Fig. 1c, e). The fact that the estimation of the dimension is biased towards a larger number than 2 in case of the bus network of Manhattan can be explained by the hierarchies inherent to the network. When the radius approaches the travel time from the west to the east of Manhattan (≈20 min), the volume increases more linearly due to the elongated shape of Manhattan (Fig. 1a, b, d). The road networks of California, Pennsylvania, and Texas follow the polynomial volume law as well, with exponent d = 2.47 ± 0.26, d = 2.50 ± 0.24, and d = 2.47 ± 0.25 respectively (Fig. 1g–i). These road networks are, in fact, strongly hierarchically, which is why the estimated dimension is higher than 2 and the volume is underestimated for lower radii. The same applies to the public transport network of Sweden (Fig. 3c), which incorporates many modes of transport with different service areas and travel speeds.
The polynomial law also applies to examples of nongeographical networks. Human connectomes^{30}, which describe neural connections in the brain, can be represented as networks^{31,32} with small brain regions, containing a collection of neurones, as nodes; and fibres between these regions, as edges. An exemplary brain network follows the polynomial volume law with exponent d = 2.69 ± 0.76 (Fig. 1j). A reason for why the human connectome exposes this law is, at least in parts, its existence in threedimensional space. In addition, the folding of the more twodimensional cerebral cortex, the ‘outer shell’ of the brain, might be a reason for the dimension of the fit to be smaller than 3.
Other networks following the polynomial volume law
Some networks follow the polynomial volume law despite having no obvious embedding into space. Among them are the metabolic network of Caenorhabditis elegans^{33,34}, a social friendshipnetwork, an email communication network, a collaboration network, and the BarabásiAlbert model^{6}, but often with a much higher exponent (Fig. 1k–p). The volume can though hardly be examined in many of these and other smallworld networks because their diameter is low by definition – global optimization principles play a role. Hence saturation effects limit the increase of volume when the volume is computed for increasing radii. The characterization of the structure of balls in a network by their volume is thus only meaningful if the network provides some notion of locality.
Networks guided by other principles
Many networks expose a power law, often indicating a local organization principle. In fact, such a law is prototypically met if all nodes have the same node degree. The volume in a tree with constant branching factor b, e.g., relates to the radius r essentially by the power law b^{r}. Balls in such networks are organized differently compared to those of spatial networks, and they can thus often not naturally be embedded in space. Volumes in the ErdősRényi model^{35,36}, in case of a low number of edges, follow an exponential law (Fig. 1q). Examples of a peertopeer computer network and a product copurchasing network follow an exponential law as well (Fig. 1r, s). There exist though networks that follow neither a polynomial nor a power or exponential law, e.g., two web graphs (Fig. 1t, u).
Conclusion and Future Work
Space shapes networks, a fact that manifests itself in a polynomial volume law. This law has been demonstrated for several networks, and many other networks that are naturally embedded in space can be expected to expose this law as well. The characterization of networks by the volume of balls has been shown to reveal generic design principles, e.g., a local optimization in regards to short average shortest path lengths in the network. Such local optimization is often accompanied by global optimization principles that create shortcuts in the network and often lead to a much smaller diameter of the network. As an example, such global optimization is inherent to hierarchical networks. We have extended the Mocnik model to incorporate hierarchies. The resulting network prototypically demonstrates the influence of global optimization on spatial networks: the effect of local optimization becomes visible at smaller radii (polynomial volume law) and global optimization at the existence or nonexistence of larger radii (small diameter of the network). If the volumes are fitted, the volume is underestimated for low radii, and the dimension is overestimated. Besides this effect, global optimization principles do, in many cases, not destroy the polynomial volume law in its core. Local and global optimization principles, here studied at the example of hierarchies, thus complement each other. Such observations can not only be made for the hierarchical network model but also for many realworld networks.
The exploration of how local and global organization principles and their mechanisms relate allows for advancements in the understanding of geographical, brain, social, and other complex networks, and it may decipher some of the principles that guide the emergence of such networks. As an example of such advancements, algorithms can be optimized to take advantage of the fact that many networks expose a polynomial volume law at least heuristically. Dijkstra’s algorithm^{37}, used for graph traversal, does not take advantage of the embedding of a network in Euclidean space. This is in contrast to the A* algorithm^{38}, which takes, as an extension of Dijkstra’s algorithm, advantage of such an embedding when choosing the Euclidean distance as heuristic cost function. It is subject to further investigations to examine how such improvements can be made for other algorithms. As soon as the polynomial volume law is met in a network, it exposes some notion of locality, similar to a network embedded in Euclidean space. This allows networkrelated algorithms to take advantage of locality, at least heuristically, and thus to be improved in terms of efficiency.
The introduced hierarchical Mocnik model can serve as a general model of objects and relations in between, in case that both can or are naturally embedded in space, i.e., of spatial data or information. The model does though not reflect all characteristics of real examples. Further research might tailor the model to reflect even more accurately the characteristics of road networks, brain networks, or communications in space. Thereby, the characteristics of these networks need to be formally understood and translated to the Mocnik model. In particular, a better understanding is needed of how the number of layers and their size can be estimated in a realworld network, which allows for a more realistic modelling. It might even be explored how the choice of the layers of the hierarchical Mocnik model can be optimized. Many spatial networks even evolve over time. Principles that guide the evolution of a network can, in many cases, be used to understand the emergence of a network, e.g., in case of the BarabásiAlbert model. Further investigations of the Mocnik model might reveal similar principles that generate the model iteratively and explain its properties in terms of its evolution.
Methods
The transport networks considered in this article consist of stops and stations as nodes, and pairs of successive stops of the same trip as edges. The weights of the edges refer to the travel time, where the start point in time of the travel is defined as the arithmetic mean of the preceding arrival and the departure of the travel represented by the edge, and the end point as the mean of the arrival and the subsequent departure. The public transport network of Sweden (Fig. 3c) contains data from almost all public transport providers in Sweden. The brain network (Fig. 1j) refers to the network GROUP_MATRIX_HD_gr2 of the referenced dataset. All other datasets are adopted from the respective references provided in the captions.
The volume is computed for 10,000 randomly chosen nodes of the largest connected component of each network, or of the associated undirected network in case of a directed network (Fig. 4a, d). The deviation of the volume roughly follows a normal distribution (Fig. 4b). A regression analysis is performed by the method of least squares for the arithmetic means, in consideration of the standard deviations computed for each radius (Fig. 4c, f). The regression analysis excludes very small and larger radii for some networks, because the computed volumes for these radii sometimes differ from the theoretic expectations: First, the finiteness of the network sets an upper limit to the volume, leading to lower volumes for balls near the boundary of the network. Secondly, volumes in a network are discrete and differ thus for small radii from the volumes in Euclidean space.
The box counting method is applied to 500 grids of boxes. These boxes have been randomly translated in space to minimize effects of the tessellation of space. For determining the boxes that intersect an edge, the edges have been discretized by points (25 points per side length of the smallest examined box).
Data availability
All sources of data have been cited. An implementation of the described Mocnik model is published as part of NetworKit^{39} (https://github.com/kitparco/networkit), an opensource toolkit for largescale network analysis.
References
 1.
Newman, M. E. J. The structure of scientific collaboration networks. Proc. Natl. Acad. Sci. United States Am. 98, 404–409 (2001).
 2.
Watts, D. J. & Strogatz, S. H. Collective dynamics of smallworld networks. Nat. 393, 440–442 (1998).
 3.
Metropolitan Transportation Authority. MTA data, http://web.mta.info/developers/developerdataterms.html  accessed on April 14, 2016 (2016).
 4.
Leskovec, J. & Krevl, A. Stanford large network dataset collection. http://snap.stanford.edu/data  accessed on June 14, 2018 (2018).
 5.
Albert, R., Jeong, H. & Barabási, A.L. Diameter of the WorldWide Web. Nat. 401, 130–131 (1999).
 6.
Barabási, A.L. & Albert, R. Emergence of scaling in random networks. Sci. 286, 509–512 (1999).
 7.
Barabási, A.L. & Bonabeau, E. Scalefree networks. Sci. Am. 288, 50–59 (2003).
 8.
Barthélémy, M. & Amaral, L. A. N. Smallworld networks: Evidence for a crossover picture. Phys. Rev. Lett. 82, 3180–3183 (1999).
 9.
Tobler, W. R. A computer movie simulating urban growth in the Detroit region. Econ. Geogr. 46, 234–240 (1970).
 10.
Tobler, W. R. On the first law of geography: A reply. Annals Assoc. Am. Geogr. 94, 304–310 (2004).
 11.
Miller, H. J. Tobler’s first law and spatial analysis. Annals Assoc. Am. Geogr. 94, 284–289 (2004).
 12.
Zipf, G. K. The hypothesis of the minimum equation as a unifying social principle: with attempted synthesis. Am. Sociol. Rev. 12, 627–650 (1947).
 13.
Song, C., Havlin, S. & Makse, H. A. Selfsimilarity of complex networks. Nat. 433, 392–395 (2005).
 14.
Shanker, O. Defining dimension of a complex network. Mod. Phys. Lett. B 21, 321–326 (2007).
 15.
Mandelbrot, B. B. The fractal geometry of nature (Freeman, New York, 1982).
 16.
Gallos, L. K., Song, C. & Makse, H. A. A review of fractality and selfsimilarity in complex networks. Phys. A 386, 686–691 (2007).
 17.
Zhou, W.X., Jiang, Z.Q. & Sornette, D. Exploring selfsimilarity of complex cellular networks: The edgecovering method with simulated annealing and logperiodic sampling. Phys. A 375, 741–752 (2007).
 18.
Li, B.G., Yu, Z.G. & Zhou, Y. Fractal and multifractal properties of a family of fractal networks. J. Stat. Mech. Theory Exp. 2014 (2014).
 19.
Schneider, C. M., Kesselring, T. A., Andrade, J. S. Jr. & Herrmann, H. J. Boxcovering algorithm for fractal dimension of complex networks. Phys. Rev. E 86, 016707 (2012).
 20.
Song, C., Gallos, L. K., Havlin, S. & Makse, H. A. How to calculate the fractal dimension of a complex network: the box covering algorithm. J. Stat. Mech. Theory Exp. 2007 (2007).
 21.
Wei, D.J. et al. Boxcovering algorithm for fractal dimension of weighted networks. Sci. Reports 3 (2013).
 22.
Li, R. et al. Simple spatial scaling rules behind complex cities. Nat. Commun. 8 (2017).
 23.
Batty, M. & Longley, P. A. Fractal cities. A geometry of form and function (Academic Press, London, 1994).
 24.
Daqing, L., Kosmidis, K., Bunde, A. & Havlin, S. Dimension of spatially embedded networks. Nat. Phys. 7, 481–484 (2001).
 25.
Waxman, B. M. Routing of multipoint connections. IEEE J. on Sel. Areas Commun. 6, 1617–1622 (1988).
 26.
Huson, M. L. & Sen, A. Broadcast scheduling algorithms for radio networks. Proc. Mil. Commun. Conf. (MILCOM) 2, 647–651 (1995).
 27.
Mocnik, F.B. & Frank, A. U. Modelling spatial structures. Proc. 12th Conf. on Spatial Inf. Theory (COSIT), 44–64 (2015).
 28.
Mocnik, F.B. A scaleinvariant spatial graph model. Ph.D. thesis, Vienna University of Technology (2015).
 29.
Geisberger, R., Sanders, P., Schultes, D. & Delling, D. Contraction hierarchies: faster and simpler hierarchical routing in road networks. Proc. 7th Int. Conf. on Exp. Algorithms (WEA), 319–333 (2008).
 30.
Sporns, O., Tonoi, G. & Kӧtter, R. The human connectome: A structural description of the human brain. PLoS Comput. Biol. 1, e42 (2005).
 31.
van den Heuvel, M. P., Kahn, R. S., Goñi, J. & Sporns, O. Highcost, highcapacity backbone for global brain communication. Proc. Natl. Acad. Sci. United States Am. 109, 11372–11377 (2012).
 32.
Bullmore, E. & Sporns, O. Complex brain networks: Graph theoretical analysis of structural and functional systems. Nat. Rev. Neurosci. 10, 186–198 (2009).
 33.
Jeong, H., Tombor, B., Albert, R., Oltvai, Z. N. & Barábasi, A.L. The largescale organization of metabolic networks. Nat. 407, 651–654 (2000).
 34.
Jeong, H., Tombor, B., Albert, R., Oltvai, Z. N. & Barabási, A.L. Data as used in the paper ‘The largescale organization of metabolic networks’. http://www3.nd.edu/~networks/resources/metabolic/CE.dat.gz  accessed on April 14, 2016 (2016).
 35.
Erdős, P. & Rényi, A. On random graphs I. Publ. Math. Debrecen 6, 290–297 (1959).
 36.
Gilbert, E. N. Random graphs. Annals Math. Stat. 30, 1141–1144 (1959).
 37.
Dijkstra, E. W. A note on two problems in connexion with graphs. Numer. Math. 1, 269–271 (1959).
 38.
Hart, P. E., Nilsson, N. J. & Raphael, B. A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions Syst. Sci. Cybern. 4, 100–107 (1968).
 39.
Staudt, C., Sazonovs, A. & Meyerhenke, H.: NetworKit: a tool suite for largescale complex network analysis. Netw. Sci. 4(4), 508–530 (2016).
 40.
van den Heuvel, M. P., Kahn, R. S., Goñi, J. & Sporns, O. Data as used in the paper ‘Highcost, highcapacity backbone for global brain communication’. http://www.dutchconnectomelab.nl/wordpress/wpcontent/uploads/GR_Dataset_n2x40.mat (accessed on June 14, 2018) (2018).
 41.
Trafiklab. GTFS Sverige. https://api.trafiklab.se/samtrafiken/gtfs/sweden.zip  accessed on April 14, 2016 (2016).
Acknowledgements
The author would like to thank Andrew U. Frank for critical discussions of the Mocnik model and Christian Freksa for his support, as well as Bernhard Höfle, Barnaby Walters, and René Westerholt for critical proofreading. The author would further like to thank the anonymous reviewers for their valuable and helpful comments. The author has been funded by Deutsche Forschungsgemeinschaft as part of the project A framework for measuring the fitness for purpose of OpenStreetMap data based on intrinsic quality indicators (FA 1189/31). The publication has also financially been supported by Deutsche Forschungsgemeinschaft within the funding programme Open Access Publishing, by the BadenWürttemberg Ministry of Science, Research and the Arts, and by Heidelberg University.
Author information
Ethics declarations
Competing Interests
The author declares no competing interests.
Additional information
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
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 license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license 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 license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Mocnik, F. The Polynomial Volume Law of Complex Networks in the Context of Local and Global Optimization. Sci Rep 8, 11274 (2018) doi:10.1038/s41598018291310
Received:
Accepted:
Published:
Further reading

Triterpenoids Extracted from Rhus chinensis Mill Act Against Colorectal Cancer by Inhibiting Enzymes in Glycolysis and Glutaminolysis: Network Analysis and Experimental Validation
Nutrition and Cancer (2019)

Shared Data Sources in the Geographical Domain—A Classification Schema and Corresponding Visualization Techniques
ISPRS International Journal of GeoInformation (2019)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.