PaperlayerView exact revision on arXiv

Source-linked AI summary

Reconstruction methods for networks: the case of economic and financial systems

Tiziano Squartini, Guido Caldarelli, Giulio Cimini, Andrea Gabrielli, Diego Garlaschelli

arXiv:1806.06941v1physics.soc-phphysics.app-phphysics.data-anq-fin.GN

TL;DR

Limited access to network structure—especially financial interconnections—motivates this review of reconstruction methods and their evaluation. It presents a unifying framework for the methods and reports comparative findings in which the fitness-induced ERG model is a strong performer, with the density of available information remaining an important constraint.

  • Problem

    Incomplete or imprecise network data, including privacy-protected financial interconnections, complicate realistic estimates of contagion and resilience.

    Source: §1 · p. 8

  • Method

    The review organizes reconstruction methods, testing indicators, and model-selection criteria in a framework focused mainly on economic and financial networks.

    Source: §1 · p. 9

  • Results

    The review reports that fitness-induced ERG and Minimum-Density methods perform strongly on several statistical measures, while fitness-induced ERG is the clear winner among ensemble methods across measures of interest.

    Source: §6.1 · p. 63

  • Takeaways & Limitations

    The review argues that basic instruments for dealing with partial network information will be increasingly needed as collecting complete information remains difficult.

    Source: §6 · p. 62

    Network density cannot be deduced from marginals alone, so it must be known or estimated from carefully selected network entries.

    Source: §6.2 · p. 65

Abstract

from arXiv · show

When studying social, economic and biological systems, one has often access to only limited information about the structure of the underlying networks. An example of paramount importance is provided by financial systems: information on the interconnections between financial institutions is privacy-protected, dramatically reducing the possibility of correctly estimating crucial systemic properties such as the resilience to the propagation of shocks. The need to compensate for the scarcity of data, while optimally employing the available information, has led to the birth of a research field known as network reconstruction. Since the latter has benefited from the contribution of researchers working in disciplines as different as mathematics, physics and economics, the results achieved so far are still scattered across heterogeneous publications. Most importantly, a systematic comparison of the network reconstruction methods proposed up to now is currently missing. This review aims at providing a unifying framework to present all these studies, mainly focusing on their application to economic and financial networks.

1. Introduction

Network reconstruction addresses incomplete or imprecise observations of real-world networks, a problem especially consequential in finance, where privacy limits access to interconnections. This review organizes reconstruction methods and evaluation tools, with emphasis on economic and financial networks.

Networks: the why and how of a theory

  • Network theory developed from graph theory and now describes relationships across domains including online communication, biology, production, and finance.The review situates this use within a broader history spanning random graph theory and sociometry.Source: §1 · p. 7

The problem of missing information

  • Privacy restrictions and partial accounting often obscure financial interconnections, limiting realistic estimates of contagion and network resilience.The review notes that incomplete network instances appear unavoidable and identifies information theory as a framework for addressing the problem.Source: §1 · p. 8

Approaching network reconstruction

  • Statistical homogeneity supports inference from observed network portions, but limits accuracy when real networks have strong structural heterogeneity.The assumption treats observed structure as representative of statistical properties across the whole network.Source: §1 · pp. 8–9
  • Entropy maximization provides the review’s unifying principle for selecting the least-biased probability distribution consistent with imposed constraints.The principle aims to avoid encoding information beyond the constraints.Source: §1 · p. 9
  • The review organizes reconstruction methods into dense, density-tunable, exact-density, and alternative approaches, then surveys accuracy tests and model-selection criteria.Its evaluation framework covers statistical, topological, and dynamical indicators, including shock-propagation resilience.Source: §1 · p. 9

2. Information theory as a basis for network reconstruction

The section frames reconstruction as inference from partial data: compatible network configurations form an ensemble, and information theory supplies a probabilistic description. Constrained entropy maximization and likelihood matching lead to the Exponential Random Graph framework.

2. Information theory as a basis for network reconstruction

  • When network data are partial, reconstruction considers all configurations compatible with the available information and assigns probabilities across that ensemble.The review frames this as making optimal use of available information while limiting unsupported assumptions.Source: §2 · pp. 9–10

2.1. Setting the problem: constraining Shannon entropy

  • Shannon entropy maximization selects a distribution that uses the imposed constraints while minimizing unjustified assumptions about unknown network information.The review describes the resulting distribution as maximally non-committal with respect to unknown information.Source: §2.1 · p. 11
  • With only O(N) available data for an adjacency matrix containing N^2 real numbers, reconstruction is under-determined and requires a probabilistic approach.When only in- and out-strengths are known, the passage gives 2N available data.Source: §2.1 · p. 13

2.2. Exponential Random Graphs

  • Exponential Random Graphs assign probabilities to network configurations by maximizing constrained entropy, with Lagrange multipliers set through constraint equations.The resulting probability distribution has an exponential form over the configuration ensemble.Source: §2.2 · p. 13
  • The ERG recipe combines entropy maximization with likelihood maximization so ensemble-average constraints match their observed values.The review also notes that ERGs can test whether a fully known network is typical under constraints chosen to define the model.Source: §2.2 · p. 14

3. Reconstruction methods

The review classifies reconstruction methods by the link density of their output networks. Figure 1 illustrates a MaxEnt reconstruction of the eMID network in 2003 alongside the observed network.

3. Reconstruction methods

  • Reconstruction methods produce fully connected or very dense networks, networks with tunable density, or networks matching the observed link count exactly.Source: §3 · p. 14
  • Figure 1 compares observed and MaxEnt-reconstructed eMID adjacency matrices for 2003, showing binary matrices at left and weighted matrices at right.Black and white encode link presence and absence in the binary panels; color intensity encodes link weight in the weighted panels.Source: Figure 1 · p. 15

3.1. Dense reconstruction methods

Dense reconstruction methods infer network weights or link probabilities from observed constraints, but MaxEnt and DWCM typically produce dense networks. IPF permits absent links, while MECAPM combines MaxEnt and ERG ideas to generate an ensemble.

3.1.1. The MaxEnt algorithm

  • MaxEnt yields a fully connected matrix unless some nodes have zero marginals, and dense reconstructions can underestimate systemic risk despite accurately estimating weight magnitudes.It produces one deterministic configuration.Source: §3.1.1 · p. 16

3.1.2. Overcoming the problem of missing connections: the IPF algorithm

  • IPF distributes known marginals across non-zero entries, allowing specified zeros while finding a matrix close to the MaxEnt estimate.With an irreducible initial matrix, the procedure yields a unique matrix satisfying the marginals.Source: §3.1.2 · pp. 16–17

3.1.3. The Directed Weighted Configuration Model

  • DWCM constrains node out-strengths and generates an ensemble by treating links as independent variables, but typically produces link probabilities near one.For non-negative integer weights, link weights follow geometric distributions.Source: §3.1.3 · pp. 17–18

3.1.4. Combining MaxEnt and ERG frameworks

  • MECAPM maximizes Shannon entropy subject to expected link weights rather than expected matrix marginals, producing a dense ensemble with link probabilities typically near one.Its Lagrange multipliers can be estimated as link probabilities when expected weights depend only on matrix marginals.Source: §3.1.4 · pp. 18–19

3.2. Density-tunable reconstruction methods

Density-tunable methods address the tendency of MaxEnt, DWCM, and MECAPM to produce networks that are too dense. They vary topology or sample compatible configurations, while their assumptions and constraints shape which networks can be reconstructed.

3.2. Density-tunable reconstruction methods

  • MaxEnt, DWCM, and MECAPM often produce configurations too dense to describe real-world networks faithfully, motivating sparser reconstruction methods.Source: §3.2 · p. 19

3.2.1. The IPF algorithm: generic formulation

  • IPF adjusts positive entries to match known marginals but requires zero-entry positions to be known or estimated beforehand.It can also account for specified subsets of entries by subtracting them from the marginals and excluding them from the update sums.Source: §3.2.1 · p. 19

3.2.2. The Drehmann & Tarashev approach

  • Drehmann and Tarashev tune density by zeroing random off-diagonal entries, randomizing remaining weights around MaxEnt estimates, then restoring marginals with IPF.The method can generate arbitrarily sparse configurations, but their topology is completely random.Source: §3.2.2 · pp. 19, 21

3.2.3. The Mastromatteo, Zarinelli & Marsili approach

  • Mastromatteo, Zarinelli, and Marsili sample configurations compatible with constraints at a chosen density, treating above-threshold entries as known and rescaling unknown entries to [0, 1].A message-passing algorithm estimates marginal link probabilities; the sampled binary topology is then weighted with IPF.Source: §3.2.3 · pp. 21–22
  • Their compatibility analysis finds no allowable configurations below a minimum density, while sparser configurations may provide lower bounds on robustness relative to MaxEnt configurations.Tunable density permits consideration of structures with potentially different robustness to financial shocks.Source: §3.2.3 · p. 22

3.2.4. The Moussa & Cont approach

  • Moussa and Cont’s exact approach generates binary priors with shared degree distributions, then uses IPF to assign weights meeting constraints in each configuration.The prior topologies are generated by a directed preferential-attachment algorithm.Source: §3.2.4 · p. 22
  • The exact approach’s IPF step may fail to converge on too-sparse matrices unless the network is strongly connected; its assigned weights are not guaranteed to follow empirical distributions.IPF has a unique solution if and only if the matrix is irreducible.Source: §3.2.4 · p. 23
  • Their average approach instead weights an ensemble of weighted networks with specified degree and weight distributions so constraints hold on average.Source: §3.2.4 · p. 23

3.3. Exact density methods

This section examines exact-density ERG reconstruction methods that incorporate observed link density, including density for subsets of nodes, as well as degree-aware, fitness-based, and gravity-inspired variants for different information constraints. It also considers how accessible topological information and degree estimates from aggregate data can improve reconstruction, while requiring suitable fitness and density estimates.

3.3.1. The density-corrected DWCM

  • Density-corrected DWCM adds the total link count to the strength constraints, making very dense configurations less probable while retaining link independence.The standard DWCM cannot tune link density independently of the weight distribution.Source: §3.3.1 · p. 24

3.3.2. The Directed Enhanced Configuration Model

  • DECM incorporates degree heterogeneity alongside strengths; its failure mode in the World Trade Network highlights the importance of degrees for faithful reconstruction.Unlike DWCM, it distinguishes establishing a first link from adding weight to an existing link.Source: §3.3.2 · p. 25

3.3.3. Simplifying the DECM: a two-step model

  • The two-step DECM simplifies reconstruction by using the DBCM’s out- and in-degree constraints, whose connection probabilities correlate strongly with those of DECM [95].Source: §3.3.3 · p. 26

3.3.5. Combining fitness-induced DBCM and IPF

  • A two-stage method estimates links with fitness-induced DBCM and assigns weights using IPF, while bootstrapped DECM uses strengths first to estimate degrees and then as additional constraints [104, 92].Source: §3.3.5 · p. 28, §3.3.6 · p. 28

3.3.7. The degree-corrected gravity model

  • The degree-corrected gravity model places MaxEnt weights only with fitness-induced DBCM link probabilities and rescales them, reproducing marginals and density on average when diagonal entries are included.Without diagonal entries, an extra correction is needed to redistribute diagonal weights; a small number of IPF-like iterations is often sufficient.Source: §3.3.7 · pp. 28, 30

3.3.8. Reconciling ERG and gravity models

  • Network-based gravity models require only global marginal information and perform much better than zero-inflated gravity models, which require the full adjacency matrix [106].Source: §3.3.8 · p. 31

3.3.9. A remark on the ensemble methods

  • Ensembles generate distinct topologies satisfying the same weighted constraints, enabling separation of marginal effects from structural effects on processes such as financial-distress spreading.Source: §3.3.9 · p. 32
  • Preliminary degree-sequence estimation enhances reconstruction performance, helping explain why algorithm [98] outperforms [84] despite both using the same information.The passage links realistic degree estimates to a good fitness ansatz and a realistic estimate of network density, including the parameter z in equation (45).Source: §3.3.9 · p. 32
  • The optimal case uses a network’s full degree sequence alongside weighted marginals in the DECM or its two-step version; aggregate topological information requires an additional assumption.Aggregate information may include the total number of links or degrees for a subset of nodes.Source: §3.3.9 · p. 32

3.4. Shannon-like approaches to reconstruction

The section surveys reconstruction approaches that extremize functionals beyond Shannon entropy or KL divergence, including spectral entropy, Cressie-Read divergences, and other entropic families.

3.4.1. Spectral entropy

  • Spectral entropy applies Shannon entropy to eigenvalue-induced probabilities, and reconstruction compares spectral densities of empirical-graph and model operators.The network density matrix is constructed to be positive semi-definite with unit trace; the cited alternative does not satisfy subadditivity.Source: §3.4.1 · pp. 32–33

3.4.2. The Cressie-Read family of power divergences

  • Cressie-Read power divergences generalize KL divergence; as γ varies, they induce different reconstruction forms, with γ → 0 recovering Shannon-entropy maximization.The procedure first derives a distribution from constraints and then optimizes the divergence over unknown parameters.Source: §3.4.2 · pp. 33–35
  • Cressie-Read reconstruction recasts inference of table entries from row and column marginals as a constrained optimization over probability coefficients.The entries are normalized by marginals and total n; the table representation illustrates the inverse problem.Source: §3.4.2 · p. 34, Table 1 · p. 35
  • For arbitrary γ, the induced entry forms differ from the usual exponential form, so reconstruction performance can differ from the KL-based approach.The divergence generalizes KL and is not a proper metric for every γ; γ = −1/2 gives a proper metric distance.Source: §3.4.2 · pp. 33, 35

3.4.3. Other entropic families

  • Rényi and Tsallis entropies generalize Shannon entropy through a free parameter and include Shannon entropy as a limiting case.A thorough performance comparison with extensive entropies is still missing because nontrivial constraints for the non-extensive formalism have not been attempted.Source: §3.4.3 · pp. 35–36

3.5. Beyond Shannon entropy: alternative approaches to reconstruction

This section surveys reconstruction approaches beyond Shannon-based optimization, including copula and fitness models, heuristic link probabilities, and the Minimum Density algorithm. It also notes that Minimum Density is presented mainly for bounding systemic risk rather than producing realistic networks, with its parameter allowing exploration of different link densities.

3.5.1. The “copula” approach

  • The copula approach generalizes MaxEnt by coupling observed marginals through a chosen dependence function, then uses IPF to restore marginals, but cannot reproduce sparse-network topology.The independent copula recovers the MaxEnt estimate.Source: §3.5.1 · p. 37

3.5.2. A Bayesian approach to network reconstruction

  • Fitness models assign latent node fitnesses and connection probabilities; suitable choices yield P(k) ∝ k^-2, showing static models can reproduce power laws without preferential attachment.A weighted extension models link presence as Bernoulli and weights exponentially; Gibbs sampling explores configurations while preserving constraints exactly.Source: §3.5.2 · pp. 37–39

3.5.3. A comment on the Bayesian approaches to reconstruction

  • Although labeled Bayesian, the three algorithms use one or more free parameters estimated via likelihood conditions, not drawn from prior distributions; as reconstruction methods, Bayesian models need at least partial tuning.Source: §3.5.3 · pp. 39–40

3.5.4. The Montagna & Lux approach

  • Montagna and Lux define link probabilities ad hoc using node strengths and density-adjusting parameters, then assign realized link weights in proportion to node size.Their original approach generates an interbank network using power-law-distributed bank interbank assets, rather than reconstructing a network.Source: §3.5.4 · p. 40

3.5.5. The probability map of Ha laj & Kok

  • Ha laj and Kok use geographic group membership and aggregated out-strengths to set link probabilities, but treat within-area links as random and require substantial extra information.The algorithm may satisfy constraints only approximately because its final allocation step is numerical.Source: §3.5.5 · p. 42

3.5.6. The Minimum Density algorithm

  • The Minimum Density algorithm [121] allocates observed marginals across as few links as possible, addressing dense MaxEnt reconstructions that may underestimate systemic risk.It chooses node pairs with mismatched strength deficits and assigns the largest feasible weight, an update that induces disassortative topology.Source: §3.5.6 · pp. 42–43
  • The Minimum Density algorithm can provide an upper bound on systemic risk, which MaxEnt’s lower bound complements to bracket the true value.The method’s stated rationale is risk bounding rather than network reconstruction.Source: §3.5.6 · p. 44
  • The algorithm generates ensembles with link density close to the minimum possible, but real networks with such low density are rarely observed.Different topological structures arise from variability in the final step.Source: §3.5.6 · p. 44
  • Tuning θ from 0 to 1 relaxes the minimum-density constraint by allocating percentages of marginals, allowing exploration across the full range of link densities.The procedure repeats the preceding steps until all marginals have been allocated.Source: §3.5.6 · pp. 43–44

4. Testing the network reconstruction

The section evaluates network reconstruction through binary link classification, weight similarity, and the reproduction of broader network structure. It cautions that individual metrics can reward dense predictions or be difficult to compare across candidates.

4.1. Statistical indicators

  • Binary reconstruction is assessed as link classification using true positives, false negatives, true negatives, and false positives.Sensitivity or TPR measures correctly recovered observed links; specificity measures correctly recovered missing links, and FPR is its complement.Source: §4.1 · pp. 44, 46
  • AUC summarizes TPR against FPR: a perfect classifier scores 1, a random classifier 1/2, and a non-random classifier lies above 1/2.High TPR alone is insufficient because predicting too many links can raise it while missing absent connections.Source: §4.1 · p. 46
  • PPV complements TPR by measuring the share of predicted links that are correct, while accuracy counts correctly classified present and missing links.Unlike TPR, a dense reconstruction method cannot trivially obtain a large PPV.Source: §4.1 · p. 47
  • MaxEnt’s expected positive predictive value (PPV) equals network density, matching the Directed Random Graph Model lower bound for reconstruction PPV.The passage describes this as weak performance unless configurations are very dense; MaxEnt’s expected accuracy also equals density.Source: §4.1 · p. 47
  • For weighted networks, cosine similarity compares observed and reconstructed weight matrices, but norm-based distances are unbounded and hinder comparisons across candidates.The section also lists L1 and L2 distances and an error measure; ensemble metrics can be evaluated using the expected weight matrix.Source: §4.1 · p. 48

4.2. Topological indicators

Topological evaluation compares reconstructed weights and higher-order network patterns with their observed counterparts, accounting for uncertainty when methods generate ensembles. The section also warns that inferred mesoscale organization can depend on the benchmark model.

4.2.1. Testing weights reconstruction

  • For weight reconstruction, scatter observed realized weights against their conditional estimates to compare only realized connections, including methods with different predicted topologies.This is useful when algorithms predict the same expected weights but different network structures.Source: §4.2.1 · p. 49

4.2.2. Testing higher-order patterns reconstruction

  • Higher-order patterns are assessed by comparing observed network quantities with ensemble means and standard deviations, using z-scores to judge whether discrepancies are significant.A discrepancy beyond the chosen threshold indicates that the imposed constraints do not fully explain that observed structure.Source: §4.2.2 · pp. 49, 51
  • When whole-ensemble calculations are infeasible, the delta method estimates pattern means and deviations; its estimates are exact for linear degree and strength constraints.For other quantities, the section describes numerical estimation by sampling a properly selected subset of network configurations.Source: §4.2.2 · pp. 49–50

4.2.3. What may and what may not be reconstructed

  • SBM and dcSBM benchmarks can interpolate between core-periphery and bipartite partitions, but the inferred organization of empirical interbank networks varies with the chosen benchmark.In the described generative example, increasing degree heterogeneity shifts the structure from bipartite toward core-periphery.Source: §4.2.3 · p. 53
  • Empirical interbank networks show benchmark-dependent mesoscale structure: SBM indicates bipartite organization at daily scales and core-periphery at longer scales, whereas dcSBM indicates bipartite structure at both.For strong degree heterogeneity, the section says dcSBM should be preferred.Source: §4.2.3 · p. 53

4.3. Dynamical indicators

Dynamical indicators characterize how distress moves through financial networks and quantify banks’ systemic impact and vulnerability. Reconstruction methods provide exposure scenarios when individual exposures are unavailable, allowing resilience tests.

4.3. Dynamical indicators

  • Reconstruction methods generate exposure scenarios compatible with available information, enabling tests of individual institutions’ and system-wide resilience.Systemicness measures a bank’s total loss induced on the system; vulnerability measures its loss when the system is under distress.Source: §4.3 · p. 54

4.3.1. Balance sheets and financial networks

  • Interbank loans form monopartite networks, while bank holdings of securities form bipartite bank-security networks.These networks emerge from interconnections represented on bank balance sheets.Source: §4.3.1 · pp. 54–55

4.3.2. Counterparty risk and credit shocks

  • DebtRank models credit shocks from defaults and mark-to-market losses, allowing distress to propagate even without defaults.It iteratively spreads distress; impact and vulnerability measure resulting equity losses while excluding first-round exogenous shocks.Source: §4.3.2 · pp. 55–56

4.3.3. Rollover risk and liquidity shocks

  • Rollover stress can prompt liquidity hoarding and illiquid-asset fire sales, producing losses that propagate as liquidity shocks.In an annealed interbank network, liquidity shocks can be incorporated into DebtRank; the impact term is ψχw_ij/e_j.Source: §4.3.3 · pp. 56–57

4.3.4. Overlapping portfolios and fire-sales spillovers

  • Overlapping securities holdings transmit fire-sale losses across banks, and leverage targeting can amplify price declines through asset sales.The model represents asset reallocation and security price impacts, which depend on securities’ illiquidity.Source: §4.3.4 · pp. 57–58
  • Portfolio-overlap indicators define systemicness as a bank’s contribution to system equity wiped out and vulnerability as the impact on its equity through other banks’ deleveraging.Their expressions depend on difficult-to-estimate illiquidity parameters; a diagonal price-impact matrix makes all illiquidity parameters equal.Source: §4.3.4 · p. 58

5. Model selection criteria

Model-selection criteria compare reconstruction models by balancing fit against complexity to limit overfitting. The section covers nested-model tests, information criteria, and model averaging.

5.1. The Likelihood Ratio Test

  • The Likelihood Ratio Test compares two nested models to test whether enlarging the parameter space is warranted.Its threshold is based on a chi-squared distribution with degrees of freedom equal to the difference in parameter counts.Source: §5.1 · p. 59

5.2. The Akaike Information Criterion

  • AIC balances maximized log-likelihood against parameter count, while AICc penalizes complexity more when empirical observations are few.The passage gives n/M_r < 40 as a rule of thumb for using AICc.Source: §5.2 · p. 60

5.3. The Bayesian Information Criterion

  • BIC discounts both parameter count and observation count, tends to select less complex models than AIC, and has no settled performance advantage.Source: §5.3 · p. 60

5.4. A quick look at multimodel averaging

  • Akaike or Bayesian weights quantify relative model support and can weight-average parameter estimates when no model wins.Models with Δ ≤ 2 have substantial statistical support, while those with Δ > 10 have essentially none.Source: §5.4 · pp. 61–62
  • Except for the first two nearly fully connected networks, the ECM outperforms the WCM, achieving unit probability within machine precision.The comparison indicates that degree information is non-redundant for reconstructing networks with non-trivial topology.Source: §5.4 · p. 61, Table 2 · p. 61

6. Conclusions and perspectives

The review compares reconstruction methods for economic and financial networks, where performance depends on the information available and the indicators used. It highlights the value and limits of link-density information and discusses adapting methods to new disclosures, including OTC-market reporting.

6.1. Comparing different reconstruction algorithms on real-world networks

  • For sparse financial networks, two-step methods that estimate topology separately from weights can outperform alternatives; the fitness-induced ERG model reproduces degree sequences and disassortative trends.The density-corrected DWCM matches marginals and density but may miss higher-order topology; the fitness-induced model is also reported to reproduce structural details better than methods using the same information.Source: §6.1 · pp. 62–63
  • Indicator choice matters: structural measures favor sparser networks, while exposure-magnitude measures favor more even allocations; the fitness-induced ERG model leads ensemble methods across reported measures.The review says performance also depends on the specific reconstruction problem and that empirical comparisons have considered many networks and indicators.Source: §6 · p. 62, §6.1 · p. 63
  • The fitness-induced ERG model performs strongly on binary topology, while Bayesian methods are versatile for generating scenarios for stress tests rather than reconstructing one particular network.For weights, the fitness-induced model leads under L1 and L2 norms, whereas the Bayesian approach leads under PTS; the Weighted Random Graph Model also scores highly under PTS.Source: §6.1 · p. 63
  • In bipartite bank-firm credit networks, ERG variants perform consistently well across aggregation levels, but no method is best on every indicator or level.For systemic-risk indicators, the model inspired by is closest overall to the actual network; Minimum-Density performance varies across aggregation levels.Source: §6.1 · pp. 63, 65

6.2. Is link density really needed?

  • Methods calibrated to observed link density outperform methods that cannot use it, making density important information for accurate reconstruction.The review also notes that density cannot be inferred from marginals alone and may need to be estimated from a carefully selected subgraph.Source: §6.2 · p. 65
  • Link density cannot be deduced from marginals alone, and subgraph-based estimates require careful entry selection to avoid sampling bias.The cited work recommends random selection; selecting nodes by total strength can make the estimated density depend strongly on the subset.Source: §6.2 · p. 65
  • When MaxEnt estimates observed weights satisfactorily, adding topological information can reduce weight-estimate error by shrinking the ensemble; strengths are better used first to estimate degrees.The reported ratio is below 1 when p_iα > 1/2.Source: §6.2 · p. 66

6.3. Policy-making implications: the case of the OTC market

  • OTC reporting rules disclose selected adjacency-matrix entries alongside marginals, and adapting reconstruction methods to such information can help evaluate the rules and inform data-sharing agreements.Known entries can be incorporated through ERG constraints or by modifying IPF to subtract known values from marginals before reallocating the remainder.Source: §6.3 · p. 66

Appendix A. A combinatorial derivation of Shannon entropy

The appendix derives Shannon entropy combinatorially by counting typical long messages from a probabilistic source. It interprets entropy as the average information needed to transmit such a message.

Appendix A. A combinatorial derivation of Shannon entropy

  • For long sequences from an independent source, symbol frequencies approach their probabilities, making sequences with those typical frequencies overwhelmingly more likely.The derivation therefore restricts attention to messages whose symbol counts satisfy n_i/n ≃ P_i.Source: Appendix A · p. 67
  • Shannon entropy quantifies the average number of bits needed to transmit a typical sequence, obtained from the logarithm of the number of typical messages per symbol.The derivation uses the law of large numbers and Stirling approximation.Source: Appendix A · pp. 67–68
  • With high probability, observed messages belong to a typical subset, yielding the asymptotic equipartition property.The appendix notes that it uses natural logarithms, so information is measured in nats; one nat equals log2 e bits.Source: Appendix A · p. 68

Appendix B. Sketching a principled derivation of Shannon entropy

This appendix sketches a derivation of Shannon entropy from the Shannon-Khinchin axioms. Additivity for independent subsystems leads to a logarithmic entropy function.

Appendix B. Sketching a principled derivation of Shannon entropy

  • The fourth Shannon-Khinchin axiom requires entropy to add for two independent subsystems.Differentiating the additivity relation with respect to each subsystem produces a differential equation.Source: Appendix B · p. 68
  • The resulting differential equation has a logarithmic solution, S(W) = k ln W.The derivation identifies the derived function as constant before solving.Source: Appendix B · p. 68
  • For dependent subsets, conditional entropy is defined as a weighted average over the configurations of subsystem B induced by subsystem A.The weights are the fractions of subsystem-A configurations inducing the corresponding B configurations.Source: Appendix B · p. 68

Appendix C. Two relevant properties of Shannon entropy

This appendix establishes two properties of Shannon entropy relevant to the review: constrained exponential distributions maximize entropy, while the unconstrained maximum occurs at the uniform distribution.

Appendix C. Two relevant properties of Shannon entropy

  • Among distributions satisfying the same constraints, an exponential distribution has greater Shannon entropy than any other distribution.Equality holds only when the probability coefficients of the two distributions coincide.Source: Appendix C · p. 69
  • Without constraints, Shannon entropy is maximized by the uniform distribution over configurations.The appendix verifies this by evaluating the stationary point and showing that the Hessian is negative-definite for distributions with positive probabilities.Source: Appendix C · pp. 69–70

Appendix D. A notable, continuous case

For a continuous real variable with constrained mean and variance, maximizing continuous Shannon entropy yields a Gaussian density; a negative second functional derivative verifies it is a maximum.

Appendix D. A notable, continuous case

  • The appendix applies continuous Shannon-entropy maximization to find the least-biased density for a real variable with constrained mean and variance.It extends the preceding case, which constrained only first moments.Source: Appendix D · p. 70
  • The optimization uses Lagrange multipliers for the three constraints and seeks a density that makes the functional derivative vanish.Source: Appendix D · p. 70
  • The maximizing density is Gaussian, with variance identified as σ^2 = µ2 − µ^2.Source: Appendix D · p. 71
  • A negative second functional derivative establishes that the Gaussian is a maximum of the constrained entropy.The appendix gives the second derivative as −δ(x−x′).Source: Appendix D · p. 71
Loading 1806.06941v1…