更改

跳到导航 跳到搜索
添加52字节 、 2025年7月6日 (日) 12:51
第129行: 第129行:     
该模型族能有效解释真实网络连接性的关键特征,包括异质度分布<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref><ref name="Krioukov_Hyperbolic_geometry">{{cite journal|author1=Krioukov D|author2=Papadopoulos F|author3=Kitsak M|author4=Vahdat A|author5=Boguñá M.|title=Hyperbolic geometry of complex networks|journal=Phys. Rev. E|year=2010|volume=82|url=}}</ref><ref name="Gugelmann_Random_hyperbolic_graphs">{{cite journal|author1=Gugelmann L|author2=Panagiotou K|author3=Peter U.|title=Random hyperbolic graphs: degree sequence and clustering|journal=Proc. Automata, Languages and Programming|year=2012|pages=573–585|url=}}</ref>、显著集聚性<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref><ref name="Krioukov_Hyperbolic_geometry">{{cite journal|author1=Krioukov D|author2=Papadopoulos F|author3=Kitsak M|author4=Vahdat A|author5=Boguñá M.|title=Hyperbolic geometry of complex networks|journal=Phys. Rev. E|year=2010|volume=82|url=}}</ref><ref name="Gugelmann_Random_hyperbolic_graphs">{{cite journal|author1=Gugelmann L|author2=Panagiotou K|author3=Peter U.|title=Random hyperbolic graphs: degree sequence and clustering|journal=Proc. Automata, Languages and Programming|year=2012|pages=573–585|url=}}</ref><ref name="Candellero_Clustering_hyperbolic">{{cite journal|author1=Candellero E|author2=Fountoulakis N.|title=Clustering and the hyperbolic geometry of complex networks|journal=Internet Math.|year=2016|volume=12|url=}}</ref><ref name="Fountoulakis_Clustering_hyperbolic_model">{{cite journal|author1=Fountoulakis N|author2=Hoorn P|author3=Müller T|author4=Schepers M.|title=Clustering in a hyperbolic model of complex networks|journal=Electron. J. Probab.|year=2021|volume=26|url=}}</ref>、小世界现象<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref><ref name="Abdullah_Typical_distances">{{cite journal|author1=Abdullah M. A|author2=Fountoulakis N|author3=Bode M.|title=Typical distances in a geometric model for complex networks|journal=Internet Math.|year=2017|url=}}</ref><ref name="Friedrich_Diameter_hyperbolic">{{cite journal|author1=Friedrich T|author2=Krohmer A.|title=On the diameter of hyperbolic random graphs|journal=SIAM J. Discrete Math.|year=2018|volume=32|url=}}</ref><ref name="Müller_Diameter_KPKVB">{{cite journal|author1=Müller T|author2=Staps M.|title=The diameter of KPKVB random graphs|journal=Adv. Appl. Probab.|year=2019|volume=51|url=}}</ref>、渗流性质<ref name="Serrano_Percolation_self_similar">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Percolation in self-similar networks|journal=Phys. Rev. Lett.|year=2011|volume=106|url=}}</ref><ref name="Fountoulakis_Law_large_numbers">{{cite journal|author1=Fountoulakis N|author2=Müller T.|title=Law of large numbers for the largest component in a hyperbolic model of complex networks|journal=Ann. Appl. Probab.|year=2018|volume=28|url=}}</ref><ref name="Bianconi_Topological_percolation">{{cite journal|author1=Bianconi G|author2=Ziff R.|title=Topological percolation on hyperbolic simplicial complexes|journal=Phys. Rev. E|year=2018|volume=98|url=}}</ref>、谱特征<ref name="Kiwi_Spectral_gap">{{cite journal|author1=Kiwi M|author2=Mitsche D.|title=Spectral gap of random hyperbolic graphs and related parameters|journal=Ann. Appl. Probab.|year=2018|volume=28|url=}}</ref>及自相似性<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref>。该框架还可以进一步扩展至生长网络的优先连接机制<ref name="Papadopoulos_Popularity_similarity">{{cite journal|author1=Papadopoulos F|author2=Kitsak M|author3=Serrano M. Á.|author4=Boguñá M|author5=Krioukov D.|title=Popularity versus similarity in growing networks|journal=Nature|year=2012|volume=489|url=}}</ref>、加权网络<ref name="Allard_Geometric_nature_weights">{{cite journal|author1=Allard A|author2=Serrano M. Á.|author3=García-Pérez G|author4=Boguñá M.|title=The geometric nature of weights in real complex networks|journal=Nat. Commun.|year=2017|volume=8|url=}}</ref>、二部网络<ref name="Serrano_Metabolic_networks">{{cite journal|author1=Serrano M. Á.|author2=Boguñá M|author3=Sagués F.|title=Uncovering the hidden geometry behind metabolic networks|journal=Mol. Biosyst.|year=2012|volume=8|url=}}</ref><ref name="Kitsak_Latent_geometry_bipartite">{{cite journal|author1=Kitsak M|author2=Papadopoulos F|author3=Krioukov D.|title=Latent geometry of bipartite networks|journal=Phys. Rev. E|year=2017|volume=95|url=}}</ref>、互补性驱动网络<ref name="Budel_Complementarity_networks">{{cite journal|author1=Budel G|author2=Kitsak M.|title=Complementarity in complex networks|journal=Preprint at arXiv|year=2023|url=}}</ref>、多层网络<ref name="Kleineberg_Hidden_geometric_correlations">{{cite journal|author1=Kleineberg K.-K|author2=Boguñá M|author3=Serrano M. Á.|author4=Papadopoulos F.|title=Hidden geometric correlations in real multiplex networks|journal=Nat. Phys.|year=2016|volume=12|url=}}</ref><ref name="Kleineberg_Geometric_correlations_multiplex">{{cite journal|author1=Kleineberg K.-K|author2=Buzna L|author3=Papadopoulos F|author4=Boguñá M|author5=Serrano M. Á.|title=Geometric correlations mitigate the extreme vulnerability of multiplex networks against targeted attacks|journal=Phys. Rev. Lett.|year=2017|volume=118|url=}}</ref>、具有社区结构的网络<ref name="Zuev_Soft_communities">{{cite journal|author1=Zuev K|author2=Boguñá M|author3=Bianconi G|author4=Krioukov D.|title=Emergence of soft communities from geometric preferential attachment|journal=Sci. Rep.|year=2015|volume=5|url=}}</ref><ref name="García-Pérez_Soft_communities_similarity">{{cite journal|author1=García-Pérez G|author2=Serrano M. Á.|author3=Boguñá M.|title=Soft communities in similarity space|journal=J. Stat. Phys.|year=2018|volume=173|url=}}</ref><ref name="Muscoloni_nPSO_model">{{cite journal|author1=Muscoloni A|author2=Cannistraci C. V.|title=A nonuniform popularity-similarity optimization (nPSO) model to efficiently generate realistic complex networks with communities|journal=New J. Phys.|year=2018|volume=20|url=}}</ref>、有向网络<ref name="Allard_Geometric_description_directed">{{cite journal|author1=Allard A|author2=Serrano M. Á.|author3=Boguñá M.|title=Geometric description of clustering in directed networks|journal=Nat. Phys.|year=2024|volume=20|url=}}</ref>以及节点含关联特征的网络<ref name="Aliakbarisani_Feature_enriched_hyperbolic">{{cite journal|author1=Aliakbarisani R|author2=Ángeles Serrano M|author3=Boguñá M.|title=Feature-enriched hyperbolic network geometry|journal=Preprint at arXiv|year=2023|url=}}</ref>。
 
该模型族能有效解释真实网络连接性的关键特征,包括异质度分布<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref><ref name="Krioukov_Hyperbolic_geometry">{{cite journal|author1=Krioukov D|author2=Papadopoulos F|author3=Kitsak M|author4=Vahdat A|author5=Boguñá M.|title=Hyperbolic geometry of complex networks|journal=Phys. Rev. E|year=2010|volume=82|url=}}</ref><ref name="Gugelmann_Random_hyperbolic_graphs">{{cite journal|author1=Gugelmann L|author2=Panagiotou K|author3=Peter U.|title=Random hyperbolic graphs: degree sequence and clustering|journal=Proc. Automata, Languages and Programming|year=2012|pages=573–585|url=}}</ref>、显著集聚性<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref><ref name="Krioukov_Hyperbolic_geometry">{{cite journal|author1=Krioukov D|author2=Papadopoulos F|author3=Kitsak M|author4=Vahdat A|author5=Boguñá M.|title=Hyperbolic geometry of complex networks|journal=Phys. Rev. E|year=2010|volume=82|url=}}</ref><ref name="Gugelmann_Random_hyperbolic_graphs">{{cite journal|author1=Gugelmann L|author2=Panagiotou K|author3=Peter U.|title=Random hyperbolic graphs: degree sequence and clustering|journal=Proc. Automata, Languages and Programming|year=2012|pages=573–585|url=}}</ref><ref name="Candellero_Clustering_hyperbolic">{{cite journal|author1=Candellero E|author2=Fountoulakis N.|title=Clustering and the hyperbolic geometry of complex networks|journal=Internet Math.|year=2016|volume=12|url=}}</ref><ref name="Fountoulakis_Clustering_hyperbolic_model">{{cite journal|author1=Fountoulakis N|author2=Hoorn P|author3=Müller T|author4=Schepers M.|title=Clustering in a hyperbolic model of complex networks|journal=Electron. J. Probab.|year=2021|volume=26|url=}}</ref>、小世界现象<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref><ref name="Abdullah_Typical_distances">{{cite journal|author1=Abdullah M. A|author2=Fountoulakis N|author3=Bode M.|title=Typical distances in a geometric model for complex networks|journal=Internet Math.|year=2017|url=}}</ref><ref name="Friedrich_Diameter_hyperbolic">{{cite journal|author1=Friedrich T|author2=Krohmer A.|title=On the diameter of hyperbolic random graphs|journal=SIAM J. Discrete Math.|year=2018|volume=32|url=}}</ref><ref name="Müller_Diameter_KPKVB">{{cite journal|author1=Müller T|author2=Staps M.|title=The diameter of KPKVB random graphs|journal=Adv. Appl. Probab.|year=2019|volume=51|url=}}</ref>、渗流性质<ref name="Serrano_Percolation_self_similar">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Percolation in self-similar networks|journal=Phys. Rev. Lett.|year=2011|volume=106|url=}}</ref><ref name="Fountoulakis_Law_large_numbers">{{cite journal|author1=Fountoulakis N|author2=Müller T.|title=Law of large numbers for the largest component in a hyperbolic model of complex networks|journal=Ann. Appl. Probab.|year=2018|volume=28|url=}}</ref><ref name="Bianconi_Topological_percolation">{{cite journal|author1=Bianconi G|author2=Ziff R.|title=Topological percolation on hyperbolic simplicial complexes|journal=Phys. Rev. E|year=2018|volume=98|url=}}</ref>、谱特征<ref name="Kiwi_Spectral_gap">{{cite journal|author1=Kiwi M|author2=Mitsche D.|title=Spectral gap of random hyperbolic graphs and related parameters|journal=Ann. Appl. Probab.|year=2018|volume=28|url=}}</ref>及自相似性<ref name="Serrano_Self_similarity_hidden_metric">{{cite journal|author1=Serrano M. Á.|author2=Krioukov D|author3=Boguñá M.|title=Self-similarity of complex networks and hidden metric spaces|journal=Phys. Rev. Lett.|year=2008|volume=100|url=}}</ref>。该框架还可以进一步扩展至生长网络的优先连接机制<ref name="Papadopoulos_Popularity_similarity">{{cite journal|author1=Papadopoulos F|author2=Kitsak M|author3=Serrano M. Á.|author4=Boguñá M|author5=Krioukov D.|title=Popularity versus similarity in growing networks|journal=Nature|year=2012|volume=489|url=}}</ref>、加权网络<ref name="Allard_Geometric_nature_weights">{{cite journal|author1=Allard A|author2=Serrano M. Á.|author3=García-Pérez G|author4=Boguñá M.|title=The geometric nature of weights in real complex networks|journal=Nat. Commun.|year=2017|volume=8|url=}}</ref>、二部网络<ref name="Serrano_Metabolic_networks">{{cite journal|author1=Serrano M. Á.|author2=Boguñá M|author3=Sagués F.|title=Uncovering the hidden geometry behind metabolic networks|journal=Mol. Biosyst.|year=2012|volume=8|url=}}</ref><ref name="Kitsak_Latent_geometry_bipartite">{{cite journal|author1=Kitsak M|author2=Papadopoulos F|author3=Krioukov D.|title=Latent geometry of bipartite networks|journal=Phys. Rev. E|year=2017|volume=95|url=}}</ref>、互补性驱动网络<ref name="Budel_Complementarity_networks">{{cite journal|author1=Budel G|author2=Kitsak M.|title=Complementarity in complex networks|journal=Preprint at arXiv|year=2023|url=}}</ref>、多层网络<ref name="Kleineberg_Hidden_geometric_correlations">{{cite journal|author1=Kleineberg K.-K|author2=Boguñá M|author3=Serrano M. Á.|author4=Papadopoulos F.|title=Hidden geometric correlations in real multiplex networks|journal=Nat. Phys.|year=2016|volume=12|url=}}</ref><ref name="Kleineberg_Geometric_correlations_multiplex">{{cite journal|author1=Kleineberg K.-K|author2=Buzna L|author3=Papadopoulos F|author4=Boguñá M|author5=Serrano M. Á.|title=Geometric correlations mitigate the extreme vulnerability of multiplex networks against targeted attacks|journal=Phys. Rev. Lett.|year=2017|volume=118|url=}}</ref>、具有社区结构的网络<ref name="Zuev_Soft_communities">{{cite journal|author1=Zuev K|author2=Boguñá M|author3=Bianconi G|author4=Krioukov D.|title=Emergence of soft communities from geometric preferential attachment|journal=Sci. Rep.|year=2015|volume=5|url=}}</ref><ref name="García-Pérez_Soft_communities_similarity">{{cite journal|author1=García-Pérez G|author2=Serrano M. Á.|author3=Boguñá M.|title=Soft communities in similarity space|journal=J. Stat. Phys.|year=2018|volume=173|url=}}</ref><ref name="Muscoloni_nPSO_model">{{cite journal|author1=Muscoloni A|author2=Cannistraci C. V.|title=A nonuniform popularity-similarity optimization (nPSO) model to efficiently generate realistic complex networks with communities|journal=New J. Phys.|year=2018|volume=20|url=}}</ref>、有向网络<ref name="Allard_Geometric_description_directed">{{cite journal|author1=Allard A|author2=Serrano M. Á.|author3=Boguñá M.|title=Geometric description of clustering in directed networks|journal=Nat. Phys.|year=2024|volume=20|url=}}</ref>以及节点含关联特征的网络<ref name="Aliakbarisani_Feature_enriched_hyperbolic">{{cite journal|author1=Aliakbarisani R|author2=Ángeles Serrano M|author3=Boguñá M.|title=Feature-enriched hyperbolic network geometry|journal=Preprint at arXiv|year=2023|url=}}</ref>。
−
<br>
+
[[文件:D26cc9a6-ef2d-4022-8c1b-8080bbc9e67b.png|替代=|无|缩略图|600x600像素|图1 真实网络上几何重整化的直接与逆向方法及其结果]]
−
[[文件:D26cc9a6-ef2d-4022-8c1b-8080bbc9e67b.png|替代=|无框|640x640像素]]
     −
</br>
      
图1 真实网络上几何重整化的直接与逆向方法及其结果。 a. 几何重整化(GR)变换示意图。在层级 <math>\ell = 0</math> 中,沿相似性圆定义 <math>r = 2</math> 个连续节点组成的非重叠蓝色块,从而将原 <math>N_0</math> 个节点划分为 <math>N_1 = N_0/r</math> 个块,形成分区 <math>\Omega_0</math>。这些块被粗粒化为层级 <math>\ell = 1</math> 中的超节点,每个超节点的角坐标由其 <math>\ell = 0</math> 层构成节点所在区域确定。如方程(2)所述连接 <math>\ell = 1</math> 层的超节点,该过程可迭代生成 <math>\ell = 2</math> 层。由于半群性质, <math>\ell = 2</math> 层亦可直接通过 <math>r = 4</math> 的块从 <math>\ell = 0</math> 层生成。b. GR 复现人脑连接组的多尺度连通结构。基于解剖数据重建五级长度尺度的多尺度层级表示(上图)。各尺度 <math>\ell</math> 下连接组的拓扑特性(符号)与基于最高分辨率几何映射的 GR 预测(曲线)高度吻合(下图)。从左至右依次为:互补累积度分布 <math>P_c(k_{\text{res}})</math>、聚类系数 <math>\bar{c}(k_{\text{res}})</math>、归一化最近邻平均度 <math>\bar k_{\text{nn}} (k_{\text{res}}) = \bar k_{\text{nn}}(k_\text{res}) \langle k \rangle/\langle k \rangle^2</math> 及富俱乐部系数 <math>r(k_{\text{res}})</math>。计算中采用重整化度 <math>k_{\text{res}} = k/\langle k \rangle</math> 以消除各层级平均度差异的影响。 c. 几何分支生长过程(GBG)以自相似方式演化网络。d. GBG 是 GR 的统计逆过程:对互联网重整化层级应用 GBG 可恢复原始网络的统计等效结构;反之,对 GBG 生成的人体代谢网络替代体应用 GR 可复原原始架构。e. GR 与 GBG 的结合为单实例真实网络的有限尺寸标度研究提供方法,可数值估算临界指数。以互联网链路随机失效的键逾渗模拟为例,网络尺寸 <math>N</math>、临界键占据概率 <math>p_c</math> 及磁化率最大值 <math>\xi_{\text{max}}</math> 均符合幂律关系(黑色符号为原始网络数据)。GBG 与 GR 壳层参数为分支率 <math>b = 2</math>、块尺寸 <math>r = 2</math>,结果表明互联网图存在逾渗阈值消失现象,符合无标度网络预期。图a改编自文献<ref name="Zheng_Geometric_renormalization_weighted">{{cite journal|author1=Zheng M|author2=García-Pérez G|author3=Boguñá M|author4=Serrano M. Á.|title=Geometric renormalization of weighted networks|journal=Commun. Phys.|year=2024|volume=7|url=}}</ref>(Springer Nature Limited);图b转载自文献<ref name="Zheng_Geometric_renormalization_connectome">{{cite journal|author1=Zheng M|author2=Allard A|author3=Hagmann P|author4=Alemán-Gómez Y|author5=Serrano M. Á.|title=Geometric renormalization unravels self-similarity of the multiscale human connectome|journal=Proc. Natl Acad. Sci. USA|year=2020|volume=117|url=}}</ref>(PNAS);图c、d、e改编自文献<ref name="Zheng_Geometric_branching_growth">{{cite journal|author1=Zheng M|author2=García-Pérez G|author3=Boguñá M|author4=Serrano M. Á.|title=Scaling up real networks by geometric branching growth|journal=Proc. Natl Acad. Sci. USA|year=2021|url=}}</ref>(PNAS)。
 
图1 真实网络上几何重整化的直接与逆向方法及其结果。 a. 几何重整化(GR)变换示意图。在层级 <math>\ell = 0</math> 中,沿相似性圆定义 <math>r = 2</math> 个连续节点组成的非重叠蓝色块,从而将原 <math>N_0</math> 个节点划分为 <math>N_1 = N_0/r</math> 个块,形成分区 <math>\Omega_0</math>。这些块被粗粒化为层级 <math>\ell = 1</math> 中的超节点,每个超节点的角坐标由其 <math>\ell = 0</math> 层构成节点所在区域确定。如方程(2)所述连接 <math>\ell = 1</math> 层的超节点,该过程可迭代生成 <math>\ell = 2</math> 层。由于半群性质, <math>\ell = 2</math> 层亦可直接通过 <math>r = 4</math> 的块从 <math>\ell = 0</math> 层生成。b. GR 复现人脑连接组的多尺度连通结构。基于解剖数据重建五级长度尺度的多尺度层级表示(上图)。各尺度 <math>\ell</math> 下连接组的拓扑特性(符号)与基于最高分辨率几何映射的 GR 预测(曲线)高度吻合(下图)。从左至右依次为:互补累积度分布 <math>P_c(k_{\text{res}})</math>、聚类系数 <math>\bar{c}(k_{\text{res}})</math>、归一化最近邻平均度 <math>\bar k_{\text{nn}} (k_{\text{res}}) = \bar k_{\text{nn}}(k_\text{res}) \langle k \rangle/\langle k \rangle^2</math> 及富俱乐部系数 <math>r(k_{\text{res}})</math>。计算中采用重整化度 <math>k_{\text{res}} = k/\langle k \rangle</math> 以消除各层级平均度差异的影响。 c. 几何分支生长过程(GBG)以自相似方式演化网络。d. GBG 是 GR 的统计逆过程:对互联网重整化层级应用 GBG 可恢复原始网络的统计等效结构;反之,对 GBG 生成的人体代谢网络替代体应用 GR 可复原原始架构。e. GR 与 GBG 的结合为单实例真实网络的有限尺寸标度研究提供方法,可数值估算临界指数。以互联网链路随机失效的键逾渗模拟为例,网络尺寸 <math>N</math>、临界键占据概率 <math>p_c</math> 及磁化率最大值 <math>\xi_{\text{max}}</math> 均符合幂律关系(黑色符号为原始网络数据)。GBG 与 GR 壳层参数为分支率 <math>b = 2</math>、块尺寸 <math>r = 2</math>,结果表明互联网图存在逾渗阈值消失现象,符合无标度网络预期。图a改编自文献<ref name="Zheng_Geometric_renormalization_weighted">{{cite journal|author1=Zheng M|author2=García-Pérez G|author3=Boguñá M|author4=Serrano M. Á.|title=Geometric renormalization of weighted networks|journal=Commun. Phys.|year=2024|volume=7|url=}}</ref>(Springer Nature Limited);图b转载自文献<ref name="Zheng_Geometric_renormalization_connectome">{{cite journal|author1=Zheng M|author2=Allard A|author3=Hagmann P|author4=Alemán-Gómez Y|author5=Serrano M. Á.|title=Geometric renormalization unravels self-similarity of the multiscale human connectome|journal=Proc. Natl Acad. Sci. USA|year=2020|volume=117|url=}}</ref>(PNAS);图c、d、e改编自文献<ref name="Zheng_Geometric_branching_growth">{{cite journal|author1=Zheng M|author2=García-Pérez G|author3=Boguñá M|author4=Serrano M. Á.|title=Scaling up real networks by geometric branching growth|journal=Proc. Natl Acad. Sci. USA|year=2021|url=}}</ref>(PNAS)。
第187行: 第185行:     
需注意 <math>\zeta^{(0)}(\tau = 0)</math>是 <math>N_0 \times N_0</math>的单位矩阵,且对所有 <math>i_0, j_0</math>有 <math>\zeta_{i_0 j_0}^{(0)}(\tau \to \infty) = 1</math>。对固定的 <math>\tau^* > 0</math>,矩阵 <math>\zeta^{(0)}(\tau^*)</math>可视为原网络的超图的邻接矩阵,其包含 <math>N_1</math>个在尺度 <math>\tau^*</math>上通过扩散连接的连通分支。此过程将原来的 <math>N_0</math>个节点划分为 <math>N_1 < N_0</math>个超节点(图2a)。 <math>\tau^*</math>越大,连通分支数量越少且每个连通分支的规模越大。 <math>\tau^* \to \infty</math>意味着仅存在单一分支,也就是说,因均匀扩散场最终所有网络节点均连接在一起。基于方程(3)的粗粒化加权邻接矩阵 <math>W^{(1)}</math>可通过定义超节点间的加权超边来构造,其权重等于连接两对应细尺度簇的微边权重经适当归一化后的总和(图2b)。亦可采用统计物理粗粒化中的一致性裁剪方法获取二值邻接矩阵 <math>A^{(1)}</math>(例如通过方程(1))。整个过程可迭代执行以定义更高层级 <math>\ell > 1</math>(图2c,d)。
 
需注意 <math>\zeta^{(0)}(\tau = 0)</math>是 <math>N_0 \times N_0</math>的单位矩阵,且对所有 <math>i_0, j_0</math>有 <math>\zeta_{i_0 j_0}^{(0)}(\tau \to \infty) = 1</math>。对固定的 <math>\tau^* > 0</math>,矩阵 <math>\zeta^{(0)}(\tau^*)</math>可视为原网络的超图的邻接矩阵,其包含 <math>N_1</math>个在尺度 <math>\tau^*</math>上通过扩散连接的连通分支。此过程将原来的 <math>N_0</math>个节点划分为 <math>N_1 < N_0</math>个超节点(图2a)。 <math>\tau^*</math>越大,连通分支数量越少且每个连通分支的规模越大。 <math>\tau^* \to \infty</math>意味着仅存在单一分支,也就是说,因均匀扩散场最终所有网络节点均连接在一起。基于方程(3)的粗粒化加权邻接矩阵 <math>W^{(1)}</math>可通过定义超节点间的加权超边来构造,其权重等于连接两对应细尺度簇的微边权重经适当归一化后的总和(图2b)。亦可采用统计物理粗粒化中的一致性裁剪方法获取二值邻接矩阵 <math>A^{(1)}</math>(例如通过方程(1))。整个过程可迭代执行以定义更高层级 <math>\ell > 1</math>(图2c,d)。
−
 
+
[[文件:图2 实空间中的拉普拉斯网络重整化方法的构造.png|图2 实空间中的拉普拉斯网络重整化方法的构造|替代=|无|缩略图|600x600像素]]
−
 
  −
[[文件:图2 实空间中的拉普拉斯网络重整化方法的构造.png|有框|图2 实空间中的拉普拉斯网络重整化方法的构造|替代=|无]]
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
      
图2 实空间中的拉普拉斯网络重整化方法的构造。a. 底层 ( <math>\ell = 0</math>) 表示原始网络 <math>\mathcal{A}^{(0)}</math>(此处为 Barabási-Albert 网络,通过每次添加一个带 <math>m = 1</math> 条连接的节点演化而成,最终规模为 <math>N_0 = 24</math> 个节点),上层展示扩散时间尺度 <math>\tau^* = 1.96</math> 对应的划分 <math>\Omega_0</math>。不同颜色标识 <math>N_1</math> 个 Kadanoff 超节点。 b. 基于 <math>\Omega_0</math>,每个块被聚合为单个超节点 <math>i</math>( <math>i = 1, \cdots, N_1</math>),保留与原节点连接的边,从而在下一层级 <math>\ell = 1</math> 构建粗粒化网络 <math>\mathcal{A}^{(1)}</math>。 c. 实空间拉普拉斯重整化群(LRG)过程的三个步骤。 d. 随机树的 LRG 粗粒化。图转载自文献<ref name="Caldarelli_heterogeneous_2024">{{cite journal|author1=Caldarelli, G.|author2=Gabrielli, A.|author3=Gili, T.|author4=Villegas, P.|title=Laplacian renormalization group: an introduction to heterogeneous coarse-graining|journal=J. Stat. Mech.|year=2024|url=}}</ref>,遵循 CCBY 4.0 协议。
 
图2 实空间中的拉普拉斯网络重整化方法的构造。a. 底层 ( <math>\ell = 0</math>) 表示原始网络 <math>\mathcal{A}^{(0)}</math>(此处为 Barabási-Albert 网络,通过每次添加一个带 <math>m = 1</math> 条连接的节点演化而成,最终规模为 <math>N_0 = 24</math> 个节点),上层展示扩散时间尺度 <math>\tau^* = 1.96</math> 对应的划分 <math>\Omega_0</math>。不同颜色标识 <math>N_1</math> 个 Kadanoff 超节点。 b. 基于 <math>\Omega_0</math>,每个块被聚合为单个超节点 <math>i</math>( <math>i = 1, \cdots, N_1</math>),保留与原节点连接的边,从而在下一层级 <math>\ell = 1</math> 构建粗粒化网络 <math>\mathcal{A}^{(1)}</math>。 c. 实空间拉普拉斯重整化群(LRG)过程的三个步骤。 d. 随机树的 LRG 粗粒化。图转载自文献<ref name="Caldarelli_heterogeneous_2024">{{cite journal|author1=Caldarelli, G.|author2=Gabrielli, A.|author3=Gili, T.|author4=Villegas, P.|title=Laplacian renormalization group: an introduction to heterogeneous coarse-graining|journal=J. Stat. Mech.|year=2024|url=}}</ref>,遵循 CCBY 4.0 协议。
第250行: 第236行:     
所求的聚合不变随机图概念在思想上类似于稳定随机变量<ref name="Uchaikin_chance_2011">{{cite book|author1=Uchaikin, V. V.|author2=Zolotarev, V. M.|title=Chance and Stability: Stable Distributions and their Applications|publisher=Walter de Gruyter|year=2011|url=}}</ref><ref name="Samorodnitsky_stable_1994">{{cite book|author1=Samorodnitsky, G.|author2=Taqqu, M. S.|title=Stable Non-Gaussian Random Processes: Stochastic Models with Infinite Variance|publisher=Chapman & Hall|year=1994|url=}}</ref><ref name="Lévy_addition_1939">{{cite journal|author1=Lévy, P.|title=L'addition des variables aléatoires définies sur une circonférence|journal=Bull. Soc. Math. Fr.|year=1939|volume=67|url=}}</ref>,即组合后仍服从相同分布的随机变量(例如求和<ref name="Samorodnitsky_stable_1994">{{cite book|author1=Samorodnitsky, G.|author2=Taqqu, M. S.|title=Stable Non-Gaussian Random Processes: Stochastic Models with Infinite Variance|publisher=Chapman & Hall|year=1994|url=}}</ref><ref name="Lévy_addition_1939">{{cite journal|author1=Lévy, P.|title=L'addition des variables aléatoires définies sur une circonférence|journal=Bull. Soc. Math. Fr.|year=1939|volume=67|url=}}</ref>或取极大值后<ref name="Balkema_max-infinite_1977">{{cite journal|author1=Balkema, A. A.|author2=Resnick, S. I.|title=Max-infinite divisibility|journal=J. Appl. Probab.|year=1977|volume=14|url=}}</ref><ref name="Giné_max-infinitely_1990">{{cite journal|author1=Giné, E.|author2=Hahn, M. G.|author3=Vatan, P.|title=Max-infinitely divisible and max-stable sample continuous processes|journal=Probab. Theory Relat. Fields|year=1990|volume=87|url=}}</ref>)。此处使用在粗粒化方法小节末尾所引入的符号:考虑以概率 <math>P_\ell\left(A^{(\ell)} \mid \Theta_\ell\right)</math>生成具体实现 <math>A^{(\ell)}</math>的随机图系综,其中 <math>\Theta_\ell</math>是模型所有参数的集合。给定概率分布 <math>P_\ell\left(A^{(\ell)} \mid \Theta_\ell\right)</math>和任意划分 <math>\Omega_\ell</math>,可在高一层级 <math>\ell+1</math>导出结果一个图系综由特定概率分布 <math>P_{\ell}\left(A^{(\ell-1)} \mid \Theta_\ell\right)</math>描述(图3a)。多尺度重整化方法<ref name="Garlaschelli_multispecies_2006">{{cite journal|author1=Garlaschelli, D.|author2=Loffredo, M. I.|title=Multispecies grand-canonical models for networks with reciprocity|journal=Phys. Rev. E|year=2006|volume=73|url=}}</ref>的核心思想是:识别在所有层次和所有可能的划分下保持不变的图概率分布函数的形式(因此可从 <math>P_{\ell}</math>中省略下标 <math>\ell</math>)。只有(重整化)参数 <math>\Theta_{\ell+1}</math>允许依赖层级 <math>\ell+1</math>,且通常是通过 <math>\Omega_\ell</math>与 <math>\Theta_\ell</math>关联。该条件对应于通过在图概率空间中识别重整化群流不动点来实现参数的精确重整化。
 
所求的聚合不变随机图概念在思想上类似于稳定随机变量<ref name="Uchaikin_chance_2011">{{cite book|author1=Uchaikin, V. V.|author2=Zolotarev, V. M.|title=Chance and Stability: Stable Distributions and their Applications|publisher=Walter de Gruyter|year=2011|url=}}</ref><ref name="Samorodnitsky_stable_1994">{{cite book|author1=Samorodnitsky, G.|author2=Taqqu, M. S.|title=Stable Non-Gaussian Random Processes: Stochastic Models with Infinite Variance|publisher=Chapman & Hall|year=1994|url=}}</ref><ref name="Lévy_addition_1939">{{cite journal|author1=Lévy, P.|title=L'addition des variables aléatoires définies sur une circonférence|journal=Bull. Soc. Math. Fr.|year=1939|volume=67|url=}}</ref>,即组合后仍服从相同分布的随机变量(例如求和<ref name="Samorodnitsky_stable_1994">{{cite book|author1=Samorodnitsky, G.|author2=Taqqu, M. S.|title=Stable Non-Gaussian Random Processes: Stochastic Models with Infinite Variance|publisher=Chapman & Hall|year=1994|url=}}</ref><ref name="Lévy_addition_1939">{{cite journal|author1=Lévy, P.|title=L'addition des variables aléatoires définies sur une circonférence|journal=Bull. Soc. Math. Fr.|year=1939|volume=67|url=}}</ref>或取极大值后<ref name="Balkema_max-infinite_1977">{{cite journal|author1=Balkema, A. A.|author2=Resnick, S. I.|title=Max-infinite divisibility|journal=J. Appl. Probab.|year=1977|volume=14|url=}}</ref><ref name="Giné_max-infinitely_1990">{{cite journal|author1=Giné, E.|author2=Hahn, M. G.|author3=Vatan, P.|title=Max-infinitely divisible and max-stable sample continuous processes|journal=Probab. Theory Relat. Fields|year=1990|volume=87|url=}}</ref>)。此处使用在粗粒化方法小节末尾所引入的符号:考虑以概率 <math>P_\ell\left(A^{(\ell)} \mid \Theta_\ell\right)</math>生成具体实现 <math>A^{(\ell)}</math>的随机图系综,其中 <math>\Theta_\ell</math>是模型所有参数的集合。给定概率分布 <math>P_\ell\left(A^{(\ell)} \mid \Theta_\ell\right)</math>和任意划分 <math>\Omega_\ell</math>,可在高一层级 <math>\ell+1</math>导出结果一个图系综由特定概率分布 <math>P_{\ell}\left(A^{(\ell-1)} \mid \Theta_\ell\right)</math>描述(图3a)。多尺度重整化方法<ref name="Garlaschelli_multispecies_2006">{{cite journal|author1=Garlaschelli, D.|author2=Loffredo, M. I.|title=Multispecies grand-canonical models for networks with reciprocity|journal=Phys. Rev. E|year=2006|volume=73|url=}}</ref>的核心思想是:识别在所有层次和所有可能的划分下保持不变的图概率分布函数的形式(因此可从 <math>P_{\ell}</math>中省略下标 <math>\ell</math>)。只有(重整化)参数 <math>\Theta_{\ell+1}</math>允许依赖层级 <math>\ell+1</math>,且通常是通过 <math>\Omega_\ell</math>与 <math>\Theta_\ell</math>关联。该条件对应于通过在图概率空间中识别重整化群流不动点来实现参数的精确重整化。
−
[[文件:Dda84141-3417-4cfe-810e-2ab0c477ca48.png|200x150像素|有框|左|图3 多尺度网络重整化方法]]
+
[[文件:Dda84141-3417-4cfe-810e-2ab0c477ca48.png|600x600px|图3 多尺度网络重整化方法|替代=|无|缩略图]]
−
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
         
图3 多尺度网络重整化方法。 a. 给定邻接矩阵为 <math>A^{(\ell)}</math> 、参数集为 <math>\Theta_{\ell}</math> 的图概率分布 <math>P_{\ell}(A^{(\ell)} | \Theta_{\ell})</math> ,节点划分 <math>\Omega_{\ell}</math> 将节点集映射到粗粒化邻接矩阵 <math>A^{(\ell+1)}</math> 的“块节点”上。若存在从 <math>i_{\ell}</math> 到 <math>j_{\ell}</math> 的连接(其中 <math>i_{\ell} \in I_{\ell+1}, j_{\ell} \in J_{\ell+1}</math> ),则绘制该连接。需注意,层级 <math>\ell</math> 的多个图可能在层级 <math>\ell+1</math> 合并为同一图,因此粗粒化将诱导下一层级图的新概率分布 <math>P_{\ell+1}(A^{(\ell+1)} | \Theta_{\ell+1})</math> 。多尺度方法旨在寻找该概率的尺度不变形式。 b. 基于多尺度模型预测重整化国际贸易网络在不同地理聚合层级的局域拓扑性质。首行展示三个代表性层级( <math>\ell_1, \ell_2, \ell_3</math> ,对应 <math>N_{\ell_1} = 183</math> (左)、 <math>N_{\ell_2} = 100</math> (中)、 <math>N_{\ell_3} = 50</math> (右))中所有 <math>N_{\ell}</math> 个节点的观测与预测度值 <math>k_{i_{\ell}}</math> 随 <math>\ln(\text{GDP}_{i_{\ell}})</math> 的变化。中间行对应相同层级下所有节点的观测与预测最近邻平均度 <math>k_{i_{\ell}}^{\text{nn}}</math> 。末行展示相同条件下局域聚类系数 <math>c_{i_{\ell}}</math> 的观测与预测值。GDP 表示国内生产总值。图转载自文献<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref>,遵循 CC BY 4.0 协议。
 
图3 多尺度网络重整化方法。 a. 给定邻接矩阵为 <math>A^{(\ell)}</math> 、参数集为 <math>\Theta_{\ell}</math> 的图概率分布 <math>P_{\ell}(A^{(\ell)} | \Theta_{\ell})</math> ,节点划分 <math>\Omega_{\ell}</math> 将节点集映射到粗粒化邻接矩阵 <math>A^{(\ell+1)}</math> 的“块节点”上。若存在从 <math>i_{\ell}</math> 到 <math>j_{\ell}</math> 的连接(其中 <math>i_{\ell} \in I_{\ell+1}, j_{\ell} \in J_{\ell+1}</math> ),则绘制该连接。需注意,层级 <math>\ell</math> 的多个图可能在层级 <math>\ell+1</math> 合并为同一图,因此粗粒化将诱导下一层级图的新概率分布 <math>P_{\ell+1}(A^{(\ell+1)} | \Theta_{\ell+1})</math> 。多尺度方法旨在寻找该概率的尺度不变形式。 b. 基于多尺度模型预测重整化国际贸易网络在不同地理聚合层级的局域拓扑性质。首行展示三个代表性层级( <math>\ell_1, \ell_2, \ell_3</math> ,对应 <math>N_{\ell_1} = 183</math> (左)、 <math>N_{\ell_2} = 100</math> (中)、 <math>N_{\ell_3} = 50</math> (右))中所有 <math>N_{\ell}</math> 个节点的观测与预测度值 <math>k_{i_{\ell}}</math> 随 <math>\ln(\text{GDP}_{i_{\ell}})</math> 的变化。中间行对应相同层级下所有节点的观测与预测最近邻平均度 <math>k_{i_{\ell}}^{\text{nn}}</math> 。末行展示相同条件下局域聚类系数 <math>c_{i_{\ell}}</math> 的观测与预测值。GDP 表示国内生产总值。图转载自文献<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref>,遵循 CC BY 4.0 协议。
−
   
=== 带独立边的多尺度模型 ===
 
=== 带独立边的多尺度模型 ===
   第292行: 第267行:  
=== 真实世界网络重整化的淬火变体 ===
 
=== 真实世界网络重整化的淬火变体 ===
   −
在多尺度模型的“淬火”变体中,可以将节点适应度解释为某种确定性的属性,例如观测量或潜在的节点特征等<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Lalli_geometry-free_2024">{{cite journal|author1=Lalli, M.|author2=Garlaschelli, D.|title=Geometry-free renormalization of directed networks: scale-invariance and reciprocity|journal=Preprint at arXiv|year=2024|url=}}</ref><ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref><ref name="Milocco_embeddings">{{cite journal|author1=Milocco R|author2=Jansen F|author3=Garlaschelli D.|title=Multi-scale node embeddings for graph modeling and generation|journal=Preprint at arXiv|year=2024}}</ref>。因此该变体可用于对真实世界的网络进行基于经验量的显式建模(如适应度模型族<ref name="Caldarelli_fitness">{{cite journal|author1=Caldarelli G|author2=Capocci A|author3=Los Rios P|author4=Munoz M. A.|title=Scale-free networks from varying vertex intrinsic fitness|journal=Phys. Rev. Lett.|year=2002|volume=89}}</ref>),或基于潜在变量的建模(如节点嵌入算法族<ref name="Dehghan_embeddings">{{cite journal|author1=Dehghan-Kooshkghazi A|author2=Kamiński B|author3=Kraiński Ł|author4=Prałat P|author5=Théberge F.|title=Evaluating node embeddings of complex networks|journal=J. Complex Netw.|year=2022|volume=10}}</ref>)。然而,这些算法族的通用模型是在固定分辨率层级构建的,而多尺度模型要求通过重整化参数在所有聚合层级一致描述同一系统,即保持相同连接概率。因此,这是唯一在聚合后仍保持适应度模型特征的适应度可加性模型<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref>,同时也是在任意粗粒化下生成一致节点嵌入的方法<ref name="Milocco_embeddings">{{cite journal|author1=Milocco R|author2=Jansen F|author3=Garlaschelli D.|title=Multi-scale node embeddings for graph modeling and generation|journal=Preprint at arXiv|year=2024}}</ref>。\
+
在多尺度模型的“淬火”变体中,可以将节点适应度解释为某种确定性的属性,例如观测量或潜在的节点特征等<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Lalli_geometry-free_2024">{{cite journal|author1=Lalli, M.|author2=Garlaschelli, D.|title=Geometry-free renormalization of directed networks: scale-invariance and reciprocity|journal=Preprint at arXiv|year=2024|url=}}</ref><ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref><ref name="Milocco_embeddings">{{cite journal|author1=Milocco R|author2=Jansen F|author3=Garlaschelli D.|title=Multi-scale node embeddings for graph modeling and generation|journal=Preprint at arXiv|year=2024}}</ref>。因此该变体可用于对真实世界的网络进行基于经验量的显式建模(如适应度模型族<ref name="Caldarelli_fitness">{{cite journal|author1=Caldarelli G|author2=Capocci A|author3=Los Rios P|author4=Munoz M. A.|title=Scale-free networks from varying vertex intrinsic fitness|journal=Phys. Rev. Lett.|year=2002|volume=89}}</ref>),或基于潜在变量的建模(如节点嵌入算法族<ref name="Dehghan_embeddings">{{cite journal|author1=Dehghan-Kooshkghazi A|author2=Kamiński B|author3=Kraiński Ł|author4=Prałat P|author5=Théberge F.|title=Evaluating node embeddings of complex networks|journal=J. Complex Netw.|year=2022|volume=10}}</ref>)。然而,这些算法族的通用模型是在固定分辨率层级构建的,而多尺度模型要求通过重整化参数在所有聚合层级一致描述同一系统,即保持相同连接概率。因此,这是唯一在聚合后仍保持适应度模型特征的适应度可加性模型<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref>,同时也是在任意粗粒化下生成一致节点嵌入的方法<ref name="Milocco_embeddings">{{cite journal|author1=Milocco R|author2=Jansen F|author3=Garlaschelli D.|title=Multi-scale node embeddings for graph modeling and generation|journal=Preprint at arXiv|year=2024}}</ref>。
    
该模型在多个层级成功复现了国际贸易网络<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Lalli_geometry-free_2024">{{cite journal|author1=Lalli, M.|author2=Garlaschelli, D.|title=Geometry-free renormalization of directed networks: scale-invariance and reciprocity|journal=Preprint at arXiv|year=2024|url=}}</ref>(适应度对应于各国经验GDP,二元因子对应于 <math>f(d)=d^{-1}</math>描述的地理距离)、企业间网络<ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref>和投入产出网络<ref name="Milocco_embeddings">{{cite journal|author1=Milocco R|author2=Jansen F|author3=Garlaschelli D.|title=Multi-scale node embeddings for graph modeling and generation|journal=Preprint at arXiv|year=2024}}</ref>(外向/内向适应度分别定义为公司/行业总产出/总投入, <math>f\equiv1</math>)的特性。当二元和/或适应度量从数据输入时,仅剩全局参数作为自由参数。该参数可在特定分辨率层级拟合,同时提供所有其他层级的预测。这些多尺度预测与多层级经验网络特性高度吻合<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Lalli_geometry-free_2024">{{cite journal|author1=Lalli, M.|author2=Garlaschelli, D.|title=Geometry-free renormalization of directed networks: scale-invariance and reciprocity|journal=Preprint at arXiv|year=2024|url=}}</ref><ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref>(图3b)。
 
该模型在多个层级成功复现了国际贸易网络<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Lalli_geometry-free_2024">{{cite journal|author1=Lalli, M.|author2=Garlaschelli, D.|title=Geometry-free renormalization of directed networks: scale-invariance and reciprocity|journal=Preprint at arXiv|year=2024|url=}}</ref>(适应度对应于各国经验GDP,二元因子对应于 <math>f(d)=d^{-1}</math>描述的地理距离)、企业间网络<ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref>和投入产出网络<ref name="Milocco_embeddings">{{cite journal|author1=Milocco R|author2=Jansen F|author3=Garlaschelli D.|title=Multi-scale node embeddings for graph modeling and generation|journal=Preprint at arXiv|year=2024}}</ref>(外向/内向适应度分别定义为公司/行业总产出/总投入, <math>f\equiv1</math>)的特性。当二元和/或适应度量从数据输入时,仅剩全局参数作为自由参数。该参数可在特定分辨率层级拟合,同时提供所有其他层级的预测。这些多尺度预测与多层级经验网络特性高度吻合<ref name="Garuccio_multiscale_2023">{{cite journal|author1=Garuccio, E.|author2=Lalli, M.|author3=Garlaschelli, D.|title=Multiscale network renormalization: scale-invariance without geometry|journal=Phys. Rev. Res.|year=2023|volume=5|url=}}</ref><ref name="Lalli_geometry-free_2024">{{cite journal|author1=Lalli, M.|author2=Garlaschelli, D.|title=Geometry-free renormalization of directed networks: scale-invariance and reciprocity|journal=Preprint at arXiv|year=2024|url=}}</ref><ref name="Ialongo_multi-scale_2024">{{cite journal|author1=Ialongo, L. N.|author2=Bangma, S.|author3=Jansen, F.|author4=Garlaschelli, D.|title=Multi-scale reconstruction of large supply networks|journal=Preprint at arXiv|year=2024|url=}}</ref>(图3b)。
第333行: 第308行:     
物理系统中的参数相关性概念在场论及其流体动力学极限下是通过幂次计算可重整化概念体现的,该概念大致追踪了描述系统动力学所有可能项的导数展开。即,在理论空间中进行有效的泰勒展开,使得即时特征显现。在足够大尺度下,含高阶导数的项对物理观测量的贡献递减。从信息论视角看,这意味观测系综的结果更敏感地依赖于导数展开中主导阶项的系数。因此,参数相关性概念可直接转译为信息论量,且能推广至物理系统之外。这种转译不仅具有直接概念用途,更为识别和理解决定任意系统在大尺度行为的参数开辟通道——即使缺乏微观模型描述,或处理任意几何与系综随机性时亦然。
 
物理系统中的参数相关性概念在场论及其流体动力学极限下是通过幂次计算可重整化概念体现的,该概念大致追踪了描述系统动力学所有可能项的导数展开。即,在理论空间中进行有效的泰勒展开,使得即时特征显现。在足够大尺度下,含高阶导数的项对物理观测量的贡献递减。从信息论视角看,这意味观测系综的结果更敏感地依赖于导数展开中主导阶项的系数。因此,参数相关性概念可直接转译为信息论量,且能推广至物理系统之外。这种转译不仅具有直接概念用途,更为识别和理解决定任意系统在大尺度行为的参数开辟通道——即使缺乏微观模型描述,或处理任意几何与系综随机性时亦然。
−
[[文件:图4 信息论参数流。.png|800*300px|有框|左|图4 信息论参数流]]
+
[[文件:图4 信息论参数流。.png|600x600px|图4 信息论参数流|替代=|无|缩略图]]
−
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
  −
 
         
图4 信息论参数流。 a. MNIST(Modified National Institute of Standards and Technology)数据集中的手写数字“7”实例。 b. 成功分类 MNIST 数据集中左侧图像的30个训练神经网络的 Fisher 信息矩阵特征值(左图)及其分箱密度直方图(右图)。特征值呈现指数型层级结构,表明仅少量集体权重对训练网络精度起关键作用。图由 Leone Luzzatto 提供。
 
图4 信息论参数流。 a. MNIST(Modified National Institute of Standards and Technology)数据集中的手写数字“7”实例。 b. 成功分类 MNIST 数据集中左侧图像的30个训练神经网络的 Fisher 信息矩阵特征值(左图)及其分箱密度直方图(右图)。特征值呈现指数型层级结构,表明仅少量集体权重对训练网络精度起关键作用。图由 Leone Luzzatto 提供。
−
   
=== 信息几何与参数流 ===
 
=== 信息几何与参数流 ===
  
5,443

个编辑

导航菜单