Information Parity in Complex Networks

• Information Parity is an information-theoretic framework that assesses balanced influence, fairness, and cost parity in network structures.
• It employs geodesic distance statistics and mutual information-inspired measures to reveal block structures and fairness tradeoffs in networks.
• Additionally, the concept extends to IP-over-ICN architectures, interconnection economics, and registry geo-consistency, offering actionable insights for network design and optimization.

Information Parity (IP) denotes several closely related but nonidentical concepts in contemporary research. In its most explicit sense, it is an information-theoretic network measure that quantifies the consonance of influence among nodes with respect to the whole network architecture. In adjacent literatures, the same phrase or a directly related parity principle is used for bounded statistical parity in fair representation design, for decomposing group-fairness tradeoffs, for equalizing backbone-cost burden in peering, for preserving IP-visible service behavior over an ICN substrate, and for testing whether IP registration records remain consistent with measured operational reality (Viol et al., 2019, Zamani et al., 18 Aug 2025, Hamman et al., 2024, Nikkhah et al., 2023, Xylomenos et al., 2018, Beverly et al., 11 Feb 2026).

1. Scope and terminological boundaries

The strictest named definition of Information Parity appears in complex-networks research, where it is introduced as a novel measure based on information theory. Later work extends the parity vocabulary into fairness and privacy, where bounded statistical parity is formalized as a mutual-information constraint, and into group-fairness analysis, where statistical parity, equalized odds, and predictive parity are expressed as overlapping information quantities. Other papers are connected to Information Parity more by explicit analogy than by terminology: peering fairness is defined as equalizing net backbone transportation costs, IP-over-ICN aims to preserve IP-visible service semantics while changing the transport core, and registry geo-consistency asks whether recorded information matches actual use (Viol et al., 2019, Zamani et al., 18 Aug 2025, Hamman et al., 2024, Nikkhah et al., 2023, Trossen et al., 2015, Beverly et al., 11 Feb 2026).

A recurring source of ambiguity is that the acronym “IP” also denotes Internet Protocol or intellectual property in neighboring work. “IP-over-ICN” uses IP in the Internet Protocol sense, whereas IPProtect and IP-CLIP concern protection of the intellectual property of datasets and vision-LLMs (Trossen et al., 2015, Singh et al., 2022, Wang et al., 4 Mar 2025). This suggests that “Information Parity” functions both as a named construct and as a broader editorial shorthand for balanced, equivalent, or reality-consistent informational relations.

2. Geodesic-distance formulation in complex networks

For an unweighted, undirected network $G(V,E)$ with $N$ nodes and geodesic distance matrix $D_{ij}$, Information Parity is defined from the statistics of shortest-path distances. For node $i$,

$$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$

which is the probability that a randomly chosen node lies at geodesic distance $r$ from $i$. For a pair $(i,j)$,

$$p_{ij}(r)= \frac{1}{N-2}\sum_{\substack{k \in V \ k\neq i,j}} \delta_{D_{ik},r}\,\delta_{D_{jk},r},$$

which is the probability that a node $k$ is at the same distance $N$0 from both $N$1 and $N$2. The Information Parity score is

$N$3

The quantity is inspired by mutual information but is not standard mutual information between two random variables, because it is built from equidistance statistics rather than a conventional joint distribution (Viol et al., 2019).

The interpretation is architectural rather than merely local. High $N$4 means that two nodes have similar distance distributions to the rest of the network and similar equidistant structure, so they are influenced by, and can influence, the network in similar ways. Low or negative values indicate substantial divergence in these patterns; the paper refers to negative values as information disparity. The supplementary formulation decomposes the score by distance through

$N$5

and also gives a generalized form using separate radii $N$6 and $N$7.

Empirically, the measure reveals block structure in Zachary’s Karate Club network: within-group IP is higher than between-group IP, and members who ended up in the same faction tend to have more similar influence profiles. In human brain networks, anatomical inter-hemispheric homologous regions near the mid-sagittal plane have higher IP than other homologous pairs, while functional networks show, on average, greater information parity for inter-hemispheric correspondent regions in comparison to the whole network. The reported positive relationship between IP and Pearson correlation in functional networks suggests that functional correlations can be partially explained by the symmetry of overall network position rather than by first-neighbor overlap alone (Viol et al., 2019).

3. Information-theoretic fairness and bounded statistical parity

In fair representation design, Information Parity is formalized through the constrained design of a representation $N$8 from useful data $N$9, task variable $D_{ij}$0, and sensitive attribute $D_{ij}$1. The central bounded statistical parity constraint is

$D_{ij}$2

where $D_{ij}$3 measures how much information the representation reveals about the sensitive attribute. Perfect demographic parity is recovered at $D_{ij}$4, in which case $D_{ij}$5 and $D_{ij}$6 are independent. Utility and compression are incorporated through

$D_{ij}$7

which maximizes task information subject to bounded leakage and bounded rate. The paper derives the upper bound

$D_{ij}$8

and constructive lower bounds based on the Functional Representation Lemma and a tighter Strong Functional Representation Lemma satisfying

$D_{ij}$9

A central conclusion is that allowing non-zero leakage can improve utility, so bounded leakage is not merely a relaxation but can enlarge achievable utility regions (Zamani et al., 18 Aug 2025).

A complementary treatment analyzes group fairness by representing the three classical gaps as information quantities about the sensitive attribute $i$0: $i$1 Using partial information decomposition (PID), the relations become

$i$2

$i$3

$i$4

This makes the overlap and disagreement among fairness notions exact rather than heuristic. Statistical parity and equalized odds overlap through $i$5, equalized odds and predictive parity overlap through $i$6, statistical parity has a redundant-only component, and predictive parity has a component unique to $i$7. The paper also states that if $i$8, at least one fairness gap must be nonzero, and all three gaps are zero if and only if $i$9. Under equalized odds, statistical parity and predictive parity trade off linearly: $$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$0 Together, these results locate Information Parity within a precise fairness–utility–compression geometry rather than a single scalar criterion (Hamman et al., 2024).

4. Cost-burden parity in interconnection economics

A different use of parity arises in peering and interconnection, where fairness is defined as parity in backbone transportation burden. The traffic-sensitive backbone cost model is

$$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$1

with access and middle-mile costs treated as borne by subscribers rather than peering partners. For two Tier-1 ISPs that interconnect at all $$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$2 major IXPs under hot potato routing, the fair fee paid by ISP$$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$3 to ISP$$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$4 is defined as the amount that equalizes net costs: $$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$5 If $$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$6, the fee is zero and settlement-free peering is fair. If $$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$7, the downstream-heavy ISP pays the other; if $$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$8, the payment reverses (Nikkhah et al., 2023).

The same principle extends to ISP–transit-provider and ISP–content-provider relations, but localization becomes decisive. For an ISP and transit provider,

$$p_i(r)= \frac{1}{N-1}\sum_{\substack{k \in V \ k\neq i}} \delta_{D_{ik},r},$$9

so the fee increases with downstream-heavy non-video traffic, decreases with localization $r$0, and is more sensitive when the video ratio $r$1 is larger. The settlement-free condition is

$r$2

For direct ISP–content-provider peering, the fee depends on the number of interconnection points $r$3, the localization fraction $r$4, and the hot- versus cold-potato distances. The paper’s special conclusion is that if the content provider peers at all $r$5 IXPs, then 50% localization is sufficient for settlement-free peering.

The principal controversy addressed in this literature is the claim that ISPs should be compensated regardless of localization. The paper explicitly rejects that claim and dispenses with the ISP argument that compensation should be paid regardless of localization. The resulting parity principle is conditional: settlement-free peering is appropriate when costs are balanced, whereas paid peering is appropriate when one side imposes more backbone cost on the other (Nikkhah et al., 2023).

5. Service-level parity in IP-over-ICN architectures

In networking systems, parity is operational rather than statistical. The IP-over-ICN proposal explicitly rejects a wholesale replacement of IP and instead uses ICN as an enabling substrate to improve existing IP-based services for a single operator. The architecture is gateway-based: user devices continue speaking standard IP-based protocols such as HTTP, CoAP, TCP, or raw IP datagrams; a Network Attachment Point (NAP) bridges the access edge into the ICN domain; and an ICN border gateway restores IP-level reachability beyond the operator domain. Inside the operator network, the Blackadder/PURSUIT model supplies the three core functions Rendezvous (RV), Topology Management (TM), and Forwarding (FN), with interfaces ICNPR, ICNSR, ICNRT, ICNTp, and ICNF. IP communication is reinterpreted as information exchange over a name: sending from address $r$6 to $r$7 becomes publishing to the ICN name $r$8, while the receiver subscribes to that name. The namespace is split into scopes $r$9 and $i$0 for internal and external addresses, and subnets are represented as scopes so that an entire subnet can be subscribed to. The design requires no changes to end devices and no changes to applications, supports multiple IP abstractions, and is constrained to a single-operator scenario (Trossen et al., 2015).

The operator-facing rationale for “better” IP is fourfold. HTTP unicast streaming can preserve the same unicast service abstraction while using multipoint delivery internally; constrained and IoT settings can attach security and privacy rules to namespaces rather than to constrained devices; VLAN-like services can use ICN’s resource awareness together with short-term network metrics and content or service popularity; and cache-aware resource management can place popular content more fairly and automatically, with predictive cache population mentioned as a possible benefit. The resulting characterization is concise: IP semantics at the service boundary, ICN semantics in the transport core (Trossen et al., 2015).

The POINT project demonstrates the same parity principle in a production setting. It preserves the familiar IP service model at the edges, inserts an ICN core behind NAP gateways, and uses RV, TM, and Forwarding Nodes with Forwarding Identifiers (FIDs) that encode the set of links a packet must traverse. Forwarding nodes apply simple bitwise operations rather than maintaining per-flow routing tables, yielding stateless forwarding with multicast and native anycast. Standard SDN switches, Open vSwitch, and OpenDaylight form the forwarding fabric, with TM interacting bidirectionally with the SDN controller. The production-network trial ran in PrimeTel’s operational network in Cyprus with more than 30 volunteers, three servers, ADSL access, two SDN switches, and dual trunk links. For HLS, POINT exploited coincidental multicast by grouping quasi-synchronized HTTP requests and multicasting a single response through the ICN core. For IPTV, multicast trees were formed at source-side NAPs and could be recomputed rapidly because multicast trees are built by bitwise combination of unicast paths. Under congestion, HLS server failure, and IPTV link failure, the paper reports that failover and rerouting were transparent or immediate, users did not notice failures under POINT, and EEG measurements aligned with questionnaires and interviews in showing lower frustration during fault conditions (Xylomenos et al., 2018).

6. Registry geo-consistency and parity between records and reality

A further operational meaning of Information Parity concerns whether the information recorded in RIR and WHOIS systems matches the actual geographic use of IP resources. The WHEREIS framework defines geo-consistency through three locations: $i$1, the RIR that registered or delegated the prefix; $i$2, the RIR responsible for the registered organization’s country; and $i$3, the RIR region inferred from active geolocation measurements. The inference procedure uses RIPE Atlas, selects responsive addresses in each prefix, probes them from geographically distributed vantages, takes the minimum RTT as a propagation estimate, converts RTT to distance using a $i$4 speed-of-light assumption in fiber, and then rules out impossible regions. The taxonomy has five classes: Fully Geo-consistent (FC), Organization Geo-consistent (OC), Organization Geo-inconsistent (OI), Registry Geo-inconsistent (RI), and Fully Geo-inconsistent (FI) (Beverly et al., 11 Feb 2026).

After filtering to prefixes aligned with BGP advertisements or their subnets, the audit covers 62,084 IPv4 prefixes and 43,000 IPv6 prefixes. For IPv4, the reported counts are FC 60,972 (98.2%), OC 653 (1.1%), OI 255 (0.4%), RI 166 (0.6%), and FI 38 (0.1%). For IPv6, they are FC 42,019 (97.7%), OC 739 (1.7%), OI 128 (0.3%), RI 106 (0.2%), and FI 8 (0.0%). In aggregate, more than 98% of the examined prefixes are consistent with the geolocation inferences, and 98.9% of IPv6 prefixes are either FC or OC. At the same time, the work emphasizes strong regional variation: RIPE exceeds 99.6% consistency in IPv4, ARIN is about 98.8%, and AFRINIC is the major IPv4 outlier at about 92.2% fully consistent, with nearly 1% fully geo-inconsistent. IPv6 registrations are no more consistent than IPv4, which the paper interprets as evidence that the issue is structural rather than technical (Beverly et al., 11 Feb 2026).

Validation is multi-layered. On 91 responsive Speedtest servers with published geolocations, the method reports 100% accuracy for OC and RI inferences after manual checking. ARIN, RIPE, and AFRINIC participated in consultation and validation; AFRINIC staff categorized a shared sample of 106 prefixes as 69 FI and 37 RI, with 19 known and acceptable uses, 77 known and unaccepted uses under policy, and 10 unknown or mixed cases. The inconsistencies also propagate into commercial geolocation databases: MaxMind, IPinfo, and DB-IP detect out-of-region prefixes at rates ranging from 7% to 92%, with IPinfo generally detecting the most and DB-IP often the fewest. In this setting, Information Parity becomes a question of transparency and truthfulness: whether the registry’s information surface remains parity-preserving with operational reality, or whether stale or inaccurate metadata affects incident response, troubleshooting, attribution, route optimization, CDN placement, geolocation-based access control, and policy enforcement (Beverly et al., 11 Feb 2026).