Mean Binary Freshness (MBF)

• Mean Binary Freshness (MBF) is a metric that defines the time-averaged fraction of objects that are up-to-date in systems like caching, recommendations, and monitoring.
• It leverages closed-form and renewal–reward formulations to enable efficient performance analysis and optimization under diverse system models.
• MBF’s applications span cache updating, network information dissemination, and recommendation systems, guiding actionable system design decisions.

Mean Binary Freshness (MBF) quantifies, for a given system and over time, the fraction of objects (files, recommendations, or node states) that are "fresh"—i.e., identical to their most up-to-date version at a reference source, or, analogously, novelties not recently presented in a recommendation context. MBF provides a scalar, actionable performance metric in fields including cache systems, recommender systems, pull-based monitoring, and networked information dissemination. It formalizes binary freshness as a long-term mean of per-item, per-session, or per-node binary indicators, and admits closed-form expressions and efficient optimization under a variety of system models.

1. Formal Definition Across Contexts

MBF is fundamentally a time- or session-averaged indicator of exact freshness. The precise operationalization depends on the system:

• Caching and Information Dissemination Systems: For each object $i$, define $f(i, t) = 1$ if the version at the consumer (cache, user, node) at time $t$ matches the most recent source version, $0$ otherwise. MBF for object $i$ is

$$\overline{F}_i = \lim_{T \to \infty}\frac{1}{T}\int_0^T f(i, t)\,dt,$$

and the aggregate MBF is the mean over all objects or weighted by request probabilities (Bastopcu et al., 2020, Bastopcu et al., 2021, Hasan et al., 4 Aug 2025, Poojary et al., 2017).

• Recommendation Systems: For session $n$ with recommendation set $R_n$ ($t$ items), and the union $A^{(n)}_k$ of the previous $f(i, t) = 1$0 recommended sets, MBF is the fraction of current recommendations not appearing in the last $f(i, t) = 1$1 sessions, averaged over $f(i, t) = 1$2 sessions:

$f(i, t) = 1$3

(Malladi et al., 2016).

• Markov Sources and Pull-Based Monitoring: For a Markov chain $f(i, t) = 1$4 and its estimate $f(i, t) = 1$5 at a monitor, MBF is

$f(i, t) = 1$6

(Akar et al., 2023, Liyanaarachchi et al., 20 Jan 2026).

In networked and cache updating systems, alternative derivations via renewal–reward theory or stochastic hybrid systems (SHS) confirm that MBF equals the probability that, in each renewal cycle, an item is successfully refreshed before the next obsolescence event (Hasan et al., 4 Aug 2025, Bastopcu et al., 2021, Bastopcu et al., 2020).

2. Model Specializations and Closed-Form MBF Expressions

Mean Binary Freshness admits closed-form or recursive characterizations under various architectures:

• Caching Under Freshness Constraints: For request probability $f(i, t) = 1$7 and required freshness window $f(i, t) = 1$8 (in slots), MBF for LRU and LP policies is directly tied to hit rate:

$f(i, t) = 1$9

(Poojary et al., 2017).

• Cache Updating Systems: For update rates $t$0 (source, cache, user), MBF at user is a product of "win probabilities" in exponential races:

$t$1

(Bastopcu et al., 2020).

• Gossip and Clustered Networks: For a node $t$2 in a network with source update rate $t$3 and per-node refresh probability $t$4, MBF is

$t$5

where $t$6 depends on network structure (disjoint, ring, fully connected, clustered) and dissemination rates. E.g.,

$t$7

and for full RC-gossip, see explicit product forms in (Hasan et al., 4 Aug 2025, Bastopcu et al., 2021).

• Remote Monitoring of Markov Sources: For a $t$8-state CTMC with stationary distribution $t$9, generator $0$0, and Poisson query rate $0$1, FWE model MBF is:

$0$2

where $0$3 are the positive eigenvalues of $0$4 and $0$5 are derived from the spectral decomposition (Akar et al., 2023).

• Recommendation Systems: Session-wise MBF is computed by counting the novel items (not served in the past $0$6 sessions) per session and averaging (Malladi et al., 2016). See Section 3 for concrete algorithms.

3. Algorithmic Computation and Optimization

MBF is amenable to efficient computation and, in many contexts, direct optimization:

• Session Aggregation (Recommendation Systems): Maintain a sliding window of recent $0$7 recommendation sets; in each session, count the number of novel items in $0$8 (w.r.t. the last $0$9 sessions), and average over $i$0 sessions for MBF. Pseudocode is provided in (Malladi et al., 2016).
• Gossip and Clustered Network Optimization: In RC-Gossip, adjust cluster size $i$1 to maximize MBF using closed-form solutions to transcendental equations derived from the renewal–reward structure. MBF as a function of $i$2 is strictly concave under standard assumptions, enabling efficient root-finding (Hasan et al., 4 Aug 2025).
• Cache Update and User Polling Policies: Under joint budget constraints (total cache and user rates), optimal policies follow a thresholding structure, with closed-form water-filling–like solutions for $i$3 and $i$4:

$i$5

with similar expressions for $i$6. Files with marginal gain below the threshold receive zero allocation (Bastopcu et al., 2020).

• Markov Sampling and Query Scheduling: For CTMC monitoring under generic delays, optimal waiting-based querying maximizes MBF by choosing a delay-dependent threshold, computable via Dinkelbach’s method and renewal theory (Liyanaarachchi et al., 20 Jan 2026). Explicit threshold expressions are given for the policy:

$i$7

where $i$8 solves a first-order condition.

• Multi-Source and Multi-Object Systems: For heterogeneous sources or objects, MBF-maximizing polling is solved via water-filling, yielding quadratic complexity in $i$9, linear for two-state cases (Akar et al., 2023).

4. Connections to Renewal-Reward, SHS, and Temporal Diversity

The renewal–reward approach is foundational for MBF analysis in Poisson-driven and Markovian environments. Under this viewpoint:

• Each obsolescence (or source version jump) initiates a renewal cycle; MBF is the expected fraction of time a given item or node is fresh in the cycle, i.e.,

$$\overline{F}_i = \lim_{T \to \infty}\frac{1}{T}\int_0^T f(i, t)\,dt,$$0

(Hasan et al., 4 Aug 2025, Bastopcu et al., 2021, Liyanaarachchi et al., 20 Jan 2026).

• For gossip, cache, and Markov monitoring systems, the renewal–reward result agrees with steady-state solutions from the SHS (stochastic hybrid system) formalism but is algebraically simpler and illuminates optimization structure (Hasan et al., 4 Aug 2025, Bastopcu et al., 2021).
• In recommendation systems, MBF reduces to a temporal diversity metric, capturing the mean dissimilarity in time between current and recent item sets (Malladi et al., 2016).

5. Design Implications and Applications

MBF is widely applicable and supports actionable system design:

• Recommendation Systems: Operators can set production-level "freshness thresholds," ensuring a minimum MBF per user, and dynamically tune the exploration–exploitation trade-off via feedback or metric-based loops (Malladi et al., 2016).
• Caching Systems: MBF prescribes optimal cache sizing and fresh-aware replacement strategies. The universal upper bound,

$$\overline{F}_i = \lim_{T \to \infty}\frac{1}{T}\int_0^T f(i, t)\,dt,$$1

is attained or approached by LP (if most-popular) or properly parameterized LRU/M-LRU, and modulated further by fresh-aware policies like LEH (Poojary et al., 2017).

• Gossip and Networks: RC-Gossip, which adaptively targets updates to only stale nodes, substantially raises MBF over static gossip by maximizing the per-stale update rate at each instant. Optimal cluster sizes in networked architectures are computable, balancing between dissemination efficiency and load (Hasan et al., 4 Aug 2025, Bastopcu et al., 2021).
• Remote Monitoring: In query-based sampling, MBF directly motivates threshold-based or delay-adaptive sampling, providing explicit performance gains and principled water-filling allocations among heterogeneous sources (Akar et al., 2023, Liyanaarachchi et al., 20 Jan 2026).
• Cache Updating: Optimal update-rate allocation follows a threshold policy in both the cache and user domains, with inactive files dropped automatically if their potential gain falls below the threshold, driving efficient resource use (Bastopcu et al., 2020).

A key insight is that MBF frequently admits strict concavity in system parameters (e.g., query rate, cluster size), greatly facilitating optimization.

6. Limitations, Asymptotics, and Performance Insights

MBF captures freshness under the precise definition of "exact match" with the source or with recent history. Its value typically decays in large systems unless appropriately compensated by design choices:

• In large gossip networks without clustering or with slow dissemination relative to obsolescence ($$\overline{F}_i = \lim_{T \to \infty}\frac{1}{T}\int_0^T f(i, t)\,dt,$$2), MBF approaches zero; clustering and fully connected architectures slow this decay (Bastopcu et al., 2021, Hasan et al., 4 Aug 2025).
• MBF's strict concavity in many cases simplifies optimizing sampling, polling, or forwarding rates, enabling practical system design.
• Gains of 20%–150% in MBF are reported for waiting-based or thresholded querying vs. zero-wait or memoryless strategies, especially under non-instantaneous feedback and asymmetric delay distributions (Liyanaarachchi et al., 20 Jan 2026).
• MBF-based policies are robust to statistical heterogeneity and scale, with water-filling and thresholding structures maintaining near-optimality under complex constraints (Akar et al., 2023, Bastopcu et al., 2020).

7. Empirical Studies and Practical Prescriptions

Simulations and empirical studies confirm the efficacy and tractability of MBF-based optimization:

• Fresh-aware cache policies (LEH, M-LRU) achieve 5–10% additional MBF over baseline LRU, especially when item popularity and freshness requirements are heterogeneous (Poojary et al., 2017).
• RC-Gossip achieves up to 50% higher MBF in typical settings, with optimal cluster sizes depending on dissemination and obsolescence rates (Hasan et al., 4 Aug 2025).
• In recommendation systems, MBF is an actionable KPI; active policies enforcing a minimum MBF ensure explicit control of repetition and user exposure to novelty (Malladi et al., 2016).
• For Markov monitoring with generic delays, perceived-age-thresholding is simple to implement and achieves significant MBF gains; optimal thresholds can be computed via Dinkelbach’s method with negligible overhead (Liyanaarachchi et al., 20 Jan 2026).

MBF thus provides a theoretically grounded, operationally meaningful, and widely applicable metric for freshness in communication, caching, recommendation, and monitoring systems. Its flexibility in modeling, analytic tractability, and actionable design implications have led to its adoption as a primary freshness metric across multiple domains.