Global Utility in Systems

• Global utility is a cross-disciplinary concept defined as system-wide objectives and latent preference indicators rather than isolated component measures.
• It integrates diverse methodologies such as MC-EM, Gibbs sampling, and convex optimization to handle competing criteria like fairness, energy-efficiency, and privacy constraints.
• Practical applications span social choice, network optimization, multi-agent reinforcement learning, and infrastructure diagnostics where local signals are aggregated to form a global perspective.

Global utility is a cross-disciplinary term used for objective functions, latent preference states, monitoring capabilities, and deployment diagnostics defined at the scale of an entire system rather than an isolated component. In social choice, it is the latent vector $\vec\theta=(\theta_1,\dots,\theta_m)$ from which observed rankings arise as noisy realizations (Soufiani et al., 2012). In wireless and network optimization, it is a network-wide objective that jointly captures competing criteria such as the EE–SE tradeoff and fairness (Deng et al., 2014). In multi-agent multi-objective reinforcement learning, it is the aggregate induced by agents’ preference vectors and is tied to Bayesian Nash equilibrium under utility heterogeneity (Li et al., 12 Nov 2025). In AI deployment for next-billion-user contexts, it is explicitly not a universal score of social value, but a diagnostic of whether systems possess the technical, economic, and governance properties likely to support useful deployment (Rawat et al., 29 May 2026).

1. Field-dependent meanings of global utility

Across the cited literature, “global utility” does not denote a single formal object. It may refer to a latent common signal underlying noisy rankings, a scalar system objective over shared resources, a privacy-aware measure of downstream usefulness under a fixed budget, a global monitoring capability, or a diagnostic of real-world adoptability. This suggests that the term is best understood relationally: “global” specifies the scale at which utility is defined, whether that scale is collective preference, network-wide allocation, full-outcome valuation, or deployment across globally diverse environments.

Domain	Global utility denotes	Source
Social choice	Latent utility vector $\vec\theta$ inducing a representative ranking	(Soufiani et al., 2012)
Uplink MU-MIMO	Network-wide proportional-fair objective combining EE–SE tradeoff and fairness	(Deng et al., 2014)
MAMORL	Average across agents of individual utility under preference vectors $\bm w_i$	(Li et al., 12 Nov 2025)
Global DP	Utility of released summary statistics under one global budget $\epsilon$	(Nunes et al., 2024)
Next-billion AI	Diagnostic of AI usefulness under constraint, not a universal score	(Rawat et al., 29 May 2026)

A recurring distinction is between globally defined value and locally observed or optimized surrogates. Several papers argue that local signals are insufficient: instantaneous heuristic token scores are not globally meaningful across attention heads (Tang et al., 9 Feb 2026); safety-utility conflicts are not uniformly distributed across Transformer parameters (Cai et al., 7 Jan 2026); and in decentralized MARL, hidden or unstructured preferences make equilibrium unattainable unless global preferences are observable or structurally modeled (Li et al., 12 Nov 2025).

2. Latent preference, full-outcome valuation, and elicitation

In random utility theory for social choice, global utility is the unobserved vector of alternative-level scores $\vec\theta$, interpreted as a “global” state of the world representing the latent quality or desirability of each alternative. Each agent independently draws a random score $X_j$ from a parameterized distribution $\mu_j(\cdot\mid \theta_j)$, and a ranking $\pi$ is generated by ordering the realized scores: $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$ In the location-family formulation,

$X_j=\theta_j+\zeta_j,$

so individual rankings are noisy samples generated by perturbing a common latent signal. The inference problem is to maximize

$\vec\theta$0

Under log-concave noise densities in the location-family subclass, the log-likelihood is concave in $\vec\theta$1, and after fixing one parameter, the set of global maximizers is bounded if and only if the data satisfy the paper’s strong connectivity-type condition. The paper then uses MC-EM, Gibbs sampling from truncated conditional distributions, and Rao–Blackwellization to make inference scalable beyond Plackett–Luce (Soufiani et al., 2012).

The GAI elicitation literature uses “global utility” differently: global elicitation asks about full outcomes $\vec\theta$2, whereas local elicitation asks about small subsets of attributes while exploiting GAI structure. Under generalized additive independence,

$\vec\theta$3

but subutility factors are not semantically independent unless they are anchored and calibrated carefully. The paper’s contribution is to replace difficult global outcome queries with local utility queries using top and bottom anchors, conditioning sets $\vec\theta$4, and local value functions $\vec\theta$5, while still preserving the semantics of expected utility theory. Only $\vec\theta$6 full-outcome anchor queries are needed for the $\vec\theta$7 factors, with the remainder of elicitation handled locally (Braziunas et al., 2012).

A common misconception is that global utility in preference modeling must be directly observable or directly queryable. The cited work instead treats it as a latent full-outcome object that can often be inferred or reconstructed from structured local assessments rather than measured directly (Soufiani et al., 2012).

3. Global utility as a system-wide optimization objective

In uplink MU-MIMO, global utility is a single scalar network objective that jointly captures the EE–SE tradeoff for each user and fairness across users via a proportional-fair aggregation. With

$\vec\theta$8

the paper defines a Cobb-Douglas-style per-user utility

$\vec\theta$9

and then

$\bm w_i$0

The network-wide problem is $\bm w_i$1 subject to per-user and sum-power constraints. Although the original problem is non-convex, the paper proves each $\bm w_i$2 is unimodal, shrinks the feasible set to $\bm w_i$3 without changing the optimum, and converts the problem into an equivalent convex problem. It then gives both a centralized algorithm requiring global system information and a primal-dual distributed algorithm that converges to the global optimum (Deng et al., 2014).

A broader distributed-power-control formulation treats global utility as any nonnegative system utility $\bm w_i$4, with no requirement that $\bm w_i$5 be concave, continuous, differentiable, monotone, or additive across links. GLAD applies Gibbs sampling to asynchronous power updates and yields a stationary distribution

$\bm w_i$6

As $\bm w_i$7, the mass concentrates on the set of global optimal power allocations. I-GLAD reduces signaling overhead through infrequent message passing, and NI-GLAD further reduces processing complexity by processing only neighboring control messages, with a neighborhood-dependent trade-off in optimality (Qian et al., 2010).

In complete bipartite matching, the global utility function is an arbitrary map $\bm w_i$8 defined on the set of perfect matchings rather than an additive edge-weight sum. The objective

$\bm w_i$9

is NP-hard by reduction from 3-CNF-SAT. The paper samples from the Gibbs distribution

$\epsilon$0

using a reversible, irreducible, aperiodic DTMC on the space of perfect matchings, proves a conductance lower bound, and derives a polynomial mixing-time bound when $\epsilon$1 is constant or grows only logarithmically with input size (Moothedath et al., 2017).

In UAV-assisted small cells, the system utility is not throughput, fairness, or energy efficiency alone, but the maximum possible number of heterogeneous users that can be served under limited energy resources while satisfying user-specific rate and coverage demands. The main problem is

$\epsilon$2

subject to per-user rate-coverage probability, total power, total normalized time, and feasibility constraints. The paper proves joint pseudo-concavity of the rate-coverage probability in $\epsilon$3, performs iterative feasibility checking over $\epsilon$4, and gives an equivalent reformulation aimed at minimizing per-user energy consumption. The reported outcome is that almost $\epsilon$5 more users can be served using the proposed joint-optimal resource allocation than with uniform allocation (Lohan et al., 2019).

4. Global utility in modern learning systems

For task-agnostic KV-cache eviction, global utility is defined as future, long-horizon model output preserved under a fixed total memory budget rather than whatever tokens look most important at the current step. The paper introduces an oracle importance

$\epsilon$6

and a head-level global objective

$\epsilon$7

After isotonic regression via PAVA, the relaxed per-head loss curves become convex and non-increasing, with effective marginal gains

$\epsilon$8

The resulting greedy solver selects the largest feasible marginal gains globally across heads. An offline profiling protocol then produces a lookup table $\epsilon$9 for near-zero-overhead runtime use. On LongBench and RULER, the reported result is an $\vec\theta$0 reduction in KV cache size with minimal performance degradation (Tang et al., 9 Feb 2026).

In safety alignment, the paper explicitly rejects the view that the safety-utility conflict is a homogeneous multi-objective optimization problem over all parameters. CAST computes head-level optimization conflict

$\vec\theta$1

functional sensitivity

$\vec\theta$2

and the combined conflict score

$\vec\theta$3

Heads are ranked into a Risky Zone and a Safe Zone, and alignment updates only selected heads via LoRA on the attention query projection $\vec\theta$4. The reported empirical claim is that the drop in general capabilities mainly comes from updating a small group of “high-conflict” heads, and that Safe Zone tuning dominates Full-SFT and Random-SFT on the safety-versus-utility Pareto frontier (Cai et al., 7 Jan 2026).

In heterogeneous multi-agent multi-objective RL, the paper’s main claim is that direct access to, or structured modeling of, global utility functions is necessary for the Bayesian Nash Equilibrium under decentralized execution constraints. Agent $\vec\theta$5’s utility is

$\vec\theta$6

and the multi-agent multi-objective return is

$\vec\theta$7

When preferences are unobservable and unstructured, the paper states that the BNE concept is inapplicable; when preferences are observable or of the form $\vec\theta$8, BNE exists. AA-MAMORL addresses this using a centralized attention-based critic that learns a joint belief over other agents’ utility functions and policies during training, while execution remains decentralized (Li et al., 12 Nov 2025).

A plausible implication is that globally defined utility in contemporary machine learning is increasingly treated as a resource-allocation problem over heterogeneous modules rather than a uniform scalar objective applied identically to all parameters or heads.

5. Privacy, fairness, and budgeted disclosure

In non-interactive Global Differential Privacy for summary statistics, global utility is tied to how a single privacy budget $\vec\theta$9 is distributed across released statistics so as to improve the utility of the equations an analyst is predicted to compute later. The paper defines

$X_j$0

with

$X_j$1

and the utility-loss metric

$X_j$2

Lower metric means better utility. The intended workflow is to predict future equations, identify the statistics those equations depend on, vary $X_j$3 subject to the global privacy budget constraint, compute the metric, and select the allocation with minimum metric. The paper is explicit that this preserves privacy because only budget allocation changes, not the total budget itself, and equally explicit that the benefit depends on correctly predicting analyst equations (Nunes et al., 2024).

The privacy-utility trade-off literature formalizes global utility differently. In the global privacy framework, a curator observes a full database $X_j$4, releases $X_j$5 through a mechanism $X_j$6, and utility is measured by the average Hamming distortion

$X_j$7

The privacy-distortion function is

$X_j$8

and the paper gives upper and lower bounds for differential privacy, approximate differential privacy, maximal information, maximal leakage, Rényi differential privacy, Sibson mutual information, and mutual information. In this formulation, global utility is the permitted distortion budget $X_j$9, while privacy is minimized subject to that utility constraint (Zhong et al., 2022).

In differentially private federated learning, utility is defined at the device level as learning gain minus the cost of participation, but the paper treats the global model as a “precious commodity” whose quality should not be uniformly disclosed to all devices. It defines a deviation factor

$\mu_j(\cdot\mid \theta_j)$0

and shares a device-specific noisy global model

$\mu_j(\cdot\mid \theta_j)$1

with larger noise for higher deviation factors. The reported result is a $\mu_j(\cdot\mid \theta_j)$2 reduction in the standard deviation of devices’ energy cost relative to the benchmark, while the standard deviation of training loss varies around $\mu_j(\cdot\mid \theta_j)$3 (Alvi et al., 2021).

A common point across these privacy papers is that global utility is not treated as privacy’s opposite in the abstract. It is operationalized through downstream equations, distortion budgets, or device-level learning gain under a shared model, and is therefore sensitive to allocation and disclosure design.

6. Global monitoring, people in the loop, and infrastructure-scale utility

For bolide detection and planetary defense, the cited work argues that infrasound is especially valuable because it provides continuous, passive, global coverage. Unlike ground cameras, which are geographically limited, or space-based systems that cover only parts of the globe or lack a dedicated bolide-monitoring payload, infrasound can detect large atmospheric explosions and bolides anywhere on Earth, day or night, and through clouds and darkness. The IMS network includes $\mu_j(\cdot\mid \theta_j)$4 fully installed and certified stations out of $\mu_j(\cdot\mid \theta_j)$5 planned. In the 23 July 2008 Tajikistan bolide case, the event was detected by two IMS infrasound stations at $\mu_j(\cdot\mid \theta_j)$6 km and $\mu_j(\cdot\mid \theta_j)$7 km, and the paper attributes the unexpected detection path to acoustic energy trapped in a weak but leaky stratospheric AtmoSOFAR channel. The analysis infers a main breakup altitude of $\mu_j(\cdot\mid \theta_j)$8 km, a spherical blast from gross fragmentation, and an energy estimate of $\mu_j(\cdot\mid \theta_j)$9–$\pi$0 kt TNT equivalent (Silber, 2024).

In people-centric IoT, the phrase “new global utility” denotes a pervasive, multi-tenant, utility-like layer of sensing, communication, and actuation embedded into everyday life. The paper distinguishes Human-in-the-loop, where humans provide context and feedback, from User-in-the-loop, where the system actively influences user behavior through suggestions, incentives, or penalties. In its representative urban Manhattan-grid scenario of about $\pi$1, the deployment includes $\pi$2 city blocks, $\pi$3 stationary machines, $\pi$4 mobile wearable machines, $\pi$5 vehicles, and $\pi$6 pedestrians. The strongest reported performance result is that NB-IoT with Type 2 user involvement can achieve up to $\pi$7 higher energy efficiency. The paper treats incentivization and billing mechanisms as necessary conditions for commoditization of this people-centric global utility (Petrov et al., 2017).

These two uses of the term differ materially. In the infrasound study, global utility is a monitoring capability defined by continuous, passive, global reach. In the IoT study, it is a socio-technical infrastructure in which human context and human actions become part of the control loop. What they share is the claim that global-scale utility depends on mechanisms—atmospheric propagation in one case, incentivized user participation in the other—that are not visible if one looks only at local instrumentation.

7. Contextual usefulness, utility-sector intelligence, and adoption diagnostics

The Next-Billion AI Index reframes global utility as a diagnostic of AI usefulness under the constraints of the global majority rather than a frontier capability score. It organizes $\pi$8 dimensions under three themes—Effective Efficiency, Operational Practicality, and Societal Integrity—and explicitly states that it is not a universal measure of usefulness for the global majority. The formative expert evaluation uses eleven one-hour semi-structured interviews with founders, developers, product leaders, engineering leads, heads of AI, a CTO, a product manager, and a data scientist. Participants identified cost, usability, reliability, and trust as especially important, and $\pi$9 participants chose cost-effectiveness as a top consideration. The paper warns that nexbax should be used with explanatory evidence and deployment context, not as a standalone scalar ranking (Rawat et al., 29 May 2026).

In smart energy infrastructure, “utility” refers both to objective functions and to the enterprise-scale public-utility sector. The unified framework described in the paper integrates a diffusion-based synthetic data generator, a transformer-based demand forecaster, a graph-autoencoder leak-risk estimator, a Simulated Bifurcation solver for compressor scheduling and demand-response activation, a retrieval-augmented billing agent, and a deterministic carbon-attribution module. The architecture is organized into Data, Analytics, Optimisation, and Engagement layers. Reported results include gas forecasting MAPE of $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$0, electricity forecasting MAPE of $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$1, leak detection Precision $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$2, Recall $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$3, F1 $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$4, and SB convergence in $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$5–$$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$6 iterations, with emissions reduced by $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$7 versus greedy (Manjunath et al., 15 May 2026).

For unsupervised pattern mining from utility usage data, Online Gaussian LDA uses multi-utility streams—electricity, water, gas, and HEMS sensors—to extract “global components” that summarize recurring consumption patterns. The model treats windows of utility usage data as documents and latent continuous-valued patterns as Gaussian topics. On the reported dataset of around $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$8 TB from one house over one month, the method is evaluated by perplexity, pattern regularity, and regression-based energy coherence. The best trade-off in the paper’s batch-size study is batch size $$\Pr(\pi\mid \vec\theta)=\Pr(X_{\pi(1)} > X_{\pi(2)} > \cdots > X_{\pi(m)}).$$9, with perplexity $X_j=\theta_j+\zeta_j,$0, and the learned patterns recur across corresponding weekdays in different weeks (Mohamad et al., 2019).

A recurring controversy in this literature concerns whether global utility can be reduced to a single number. The next-billion AI framework explicitly rejects that reduction (Rawat et al., 29 May 2026); the smart-infrastructure framework instead composes forecasting, leak analysis, scheduling, billing, and carbon accounting as interconnected layers (Manjunath et al., 15 May 2026); and the multi-utility pattern-mining work uses global components as a latent representation rather than an explicit scalar welfare function (Mohamad et al., 2019). This suggests that, in infrastructure and deployment settings, global utility is often a structured diagnostic or operational stack rather than a one-dimensional maximand.