Net-Gain Maximization (NGM) Overview

• NGM is an optimization pattern that subtracts explicit costs—such as promotion fees, switching expenses, or training energy—from gross benefits to determine net outcomes.
• The framework spans diverse domains, with applications in social-network promotion, wireless energy transfer, and integrated photonics, each employing distinct mathematical formulations.
• NGM methodologies leverage advanced algorithms like non-monotone submodular optimization, residual descent, and regret minimization to efficiently balance resource constraints against performance gains.

Net-Gain Maximization (NGM) denotes a recurring optimization pattern in which a decision maker seeks a net objective obtained after subtracting explicit expenditure, loss, penalty, or overhead from a gross benefit. In the cited literature, this pattern appears as expected influence gain minus promotion cost, payoff improvement minus switching cost, coalition profit after energy and QoS charges, and harvested energy minus channel-training expenditure (Gu et al., 2022, Zusai, 2018, Anglano et al., 2014, Zeng et al., 2015). The same language also appears in engineering settings where the central question is whether internal amplification exceeds propagation or coupling losses, and, in some fields, “net gain” denotes an observable rather than a control objective, which makes the term broad rather than uniform (Kao et al., 5 Feb 2026, Surkau et al., 30 Sep 2025).

1. Conceptual scope and terminology

Across the literature, NGM is not a single canonical formalism but a family of related constructions. Some papers use the label directly; others use closely related names such as profit maximization, net harvested energy, or net gains from switching. The common structure is the explicit balancing of a gain-side term against a cost-side term, or, more generally, against a term that measures the burden of obtaining or maintaining that gain.

A compact way to organize the main uses is the following.

Domain	Net quantity	Role
Social-network promotion	$f(m)=\sigma(m)-\delta(m)$	optimization objective
Evolutionary dynamics	$[\breve\pi_a(A')-q]_+$	switching incentive
Strategic network formation	$|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$	player utility
Wireless energy transfer	$$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$	optimization objective
Integrated photonics	$$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$	performance metric
QCD	$\Delta Q_q+1=1+T\partial_\mu\ln\ell$	thermodynamic observable

This distribution of meanings suggests that NGM is best understood as a structural motif rather than a field-specific doctrine. In some settings it is an optimization target; in others it is a residual, a utility, a scheduling index, or a measurable response quantity.

2. Mathematical archetypes

The most direct NGM form is the subtractive objective. In the bipartite social-network promotion model, the objective is explicitly

$f(m)=\sigma(m)-\delta(m),$

where $\sigma(m)$ is expected influence gain and $\delta(m)$ is linear promotion cost (Gu et al., 2022). In the 2-neighborhood network creation game, the same pattern appears as a player utility

$$u_u(\mathbf{s})=\left|N_2^{G(\mathbf{s})}(u)\right|-\alpha |S_u|,$$

with localized reachability as benefit and link purchases as cost (Haye et al., 10 Feb 2025). In green cellular cooperation, the coalition value is

$[\breve\pi_a(A')-q]_+$0

so revenue is offset by serving cost and coalition-formation cost (Anglano et al., 2014). In wireless energy transfer, the net quantity is

$[\breve\pi_a(A')-q]_+$1

which subtracts reverse-link training energy from average harvested energy (Zeng et al., 2015).

A second archetype is net gain from revision or deviation. In evolutionary dynamics, the relevant object is the maximal payoff improvement from switching after paying switching cost,

$[\breve\pi_a(A')-q]_+$2

with the aggregate maximized expected net gain later used as a Lyapunov function (Zusai, 2018). In game-theoretic equilibrium computation, the classical Nikaido-Isoda function measures exact unilateral deviation gain, while the Gradient-based Nikaido-Isoda function replaces global best-response improvement by one local steepest-descent step (Raghunathan et al., 2019).

A third archetype is gain under hard resource constraints. The dynamic influence-maximization game uses explicit budget constraints

$[\breve\pi_a(A')-q]_+$3

and may or may not include expenditure directly in stage utility. In the paper’s explicit two-player example, the stage utility becomes a true gain-minus-cost expression,

$[\breve\pi_a(A')-q]_+$4

but the broader framework is budget-constrained payoff maximization rather than universally subtractive NGM (Etesami, 2021). The same distinction is central in Flow-Based Network Creation Games, where agents maximize flow/connectivity benefit subject to a hard budget $[\breve\pi_a(A')-q]_+$5; the model is therefore a constrained gain-maximization framework, not a direct benefit-minus-cost formulation (Echzell et al., 2020).

This range of forms suggests a useful classification: subtractive objectives, constrained-gain objectives, deviation-gain residuals, and observational net quantities. The mathematics changes substantially across these classes even when the phrase “net gain” is shared.

3. Algorithmic and equilibrium methods

One major algorithmic line treats NGM objectives as non-monotone submodular optimization. In the social-network promotion problem, the expected influence term is DR-submodular on a bounded integer lattice, the linear cost is modular, and their difference remains DR-submodular. The resulting binary search double greedy algorithm is a randomized $[\breve\pi_a(A')-q]_+$6 time $[\breve\pi_a(A')-q]_+$7-approximation algorithm, improving over $[\breve\pi_a(A')-q]_+$8 lattice double greedy by shrinking coordinate intervals approximately by half (Gu et al., 2022).

A second line replaces direct optimization of net gain by minimization of a gain-based residual. The Gradient-based Nikaido-Isoda function is

$[\breve\pi_a(A')-q]_+$9

where $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$0 changes only player $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$1’s block by a steepest-descent step. This quantity is nonnegative and vanishes exactly at stationary Nash points for $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$2. Gradient descent on $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$3 converges sublinearly to a first-order stationary point, and in bilinear min-max games and multi-player quadratic games the GNI function is convex, yielding linear convergence to an equilibrium when one exists (Raghunathan et al., 2019).

A third line treats net gain as a dynamic disequilibrium certificate. In population games, the aggregate maximized expected net gain

$|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$4

is zero exactly at rest points for cost-benefit rationalizable dynamics. Under static stability and full cost-benefit rationalizability, its time derivative is nonpositive, so aggregate net gain decreases to zero and acts as a Lyapunov function (Zusai, 2018). This connects “negative incentives for deviation” to dynamic stability without requiring the dynamic to be written directly as an optimization routine.

A fourth line computes strategic NGM allocations by regret minimization. In the dynamic influence-maximization game, the state evolution under DeGroot mixing and campaign-time jumps yields a finite-horizon open-loop game. Under socially concave conditions, pure-strategy open-loop equilibria exist, and any no-regret algorithm in the repeated version converges to one. The paper uses projected gradient ascent as a concrete update rule for equilibrium computation (Etesami, 2021).

These algorithmic families show that NGM is not tied to a single solver. It appears as approximation on integer lattices, residual descent, Lyapunov analysis, and repeated-game learning.

4. Networked, social, and scheduling applications

In networked systems, NGM often appears as the operational core of a larger control or game-theoretic problem. In edge-server monitoring, the query-scheduling problem is modeled as a multi-action RMAB. The per-dispatcher net gain of querying server $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$5 is

$|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$6

namely the relative action value of querying versus not querying. The NGM policy schedules up to $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$7 dispatchers with the largest non-negative gains, and simulations report up to a $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$8 gain over the Round-Robin Policy and up to a $|N_2^{G(\mathbf{s})}(u)|-\alpha |S_u|$9 gain over the Never-Query Policy (Chamoun et al., 8 Sep 2025).

A closely related scheduling use appears in age-of-job minimization for task-specific machine networks. There, NGM is compared against max-weight and Whittle-index-based policies. The reported numerical pattern is not uniform across scales: two different max-weight policies can outperform the NGM policy in small systems, whereas the NGM policy improves as the system size scales and becomes asymptotically better than max-weight policies; for geometric service times, the WI policy gives the lowest age across all considered system sizes (Banerjee et al., 2 Feb 2026). This is a reminder that NGM-derived heuristics can have scale-dependent behavior even within the same architecture.

Strategic network formation gives a different reading of NGM. In the 2-neighborhood maximization game, agents buy edges to maximize localized reachability minus edge cost. The model yields sharp structural and welfare results: connected Nash equilibria have diameter at most $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$0 for $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$1, connected greedy equilibria have diameter at most $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$2 for $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$3 and at most $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$4 for $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$5, while the price of anarchy satisfies

$$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$6

and for $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$7 the greedy price of anarchy is $$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$8 (Haye et al., 10 Feb 2025). Localized net gain therefore produces compact equilibrium networks but not necessarily efficient ones.

In infrastructure sharing, NGM becomes coalition profit optimization. In green cellular networks, operators share base stations and users to maximize

$$\bar Q_{\text{net}}=\bar Q-E_1N_1-\sum_{n=1}^{N_2}E_{2,n}$$9

while coalition value is computed from optimized user-to-BS allocation and distributed through cooperative-game solution concepts. The resulting coalition-formation process converges to Nash-stable partitions and yields large profit improvements when energy costs are high and loads are heterogeneous (Anglano et al., 2014).

Taken together, these papers show that NGM in networked systems is typically state-dependent, resource-constrained, and often embedded in a larger scheduling or coalition mechanism rather than appearing as a single monolithic objective.

5. Physical and engineering realizations

In wireless energy transfer, NGM is a direct design principle. The two-phase training scheme for frequency-selective MISO WET uses Phase I to search $$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$0 sub-bands for strong channel norms and Phase II to refine the $$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$1 selected sub-bands for beamforming. The average harvested energy is

$$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$2

and NGM subtracts the two training costs (Zeng et al., 2015). The optimal Phase-II energy has a threshold form,

$$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$3

The asymptotics are especially informative: with many antennas, $$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$4 and only beamforming-oriented training survives; with many sub-bands, perfect-CSI gross harvested energy grows only logarithmically while net gain saturates because search cost grows too quickly.

Integrated photonics uses net gain in a more literal sense. In a thin-film lithium-niobate optical parametric amplifier, the on-chip phase-sensitive gain is

$$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$5

and the paper reports $$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$6 dB phase-sensitive gain at $$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$7 mW pump power, with total fibre-chip-fibre loss $$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$8 dB and thus external net gain up to $$G_{\mathrm{net,dB}}=G_{\mathrm{on\text{-}chip,dB}}-L_{\mathrm{f\!-\!c\!-\!f,dB}}$$9 dB. The mechanism is parametric down-conversion in thin-film lithium-niobate waveguides using adapted poling to preserve nonlinear coherence, and the $\Delta Q_q+1=1+T\partial_\mu\ln\ell$0 dB bandwidth is approximately $\Delta Q_q+1=1+T\partial_\mu\ln\ell$1 nm (Kao et al., 5 Feb 2026). In an erbium-doped lithium-niobate thin-film waveguide amplifier, the signal enhancement reaches $\Delta Q_q+1=1+T\partial_\mu\ln\ell$2 dB and the internal net gain reaches $\Delta Q_q+1=1+T\partial_\mu\ln\ell$3 dB in a $\Delta Q_q+1=1+T\partial_\mu\ln\ell$4-cm device at $\Delta Q_q+1=1+T\partial_\mu\ln\ell$5 nm, with $\Delta Q_q+1=1+T\partial_\mu\ln\ell$6 dB/cm, $\Delta Q_q+1=1+T\partial_\mu\ln\ell$7 dBm saturation power, and $\Delta Q_q+1=1+T\partial_\mu\ln\ell$8 dB noise figure (Cai et al., 2021). These two papers also make clear that photonic NGM is reference-plane dependent: “internal net gain” and “fibre-chip-fibre net gain” are distinct quantities.

Physical implementations also expose important measurement caveats. In Al$\Delta Q_q+1=1+T\partial_\mu\ln\ell$9Ga$f(m)=\sigma(m)-\delta(m),$0N/GaN multi-quantum wells, VSL measurements show a reliable maximum net gain of $f(m)=\sigma(m)-\delta(m),$1 at $f(m)=\sigma(m)-\delta(m),$2, while an apparently higher $f(m)=\sigma(m)-\delta(m),$3 value is attributed to anomalous amplification caused by crack-induced feedback and therefore overestimates intrinsic gain (Thai et al., 2021). In active cavities with net roundtrip gain,

$f(m)=\sigma(m)-\delta(m),$4

the condition $f(m)=\sigma(m)-\delta(m),$5 does not imply an unbounded driven steady-state solution; the exact Maxwell boundary-value solution remains finite, with side-tail-induced pre-excitation and interference providing the physical resolution (Mansuripur et al., 2013).

A more radical energy-balance use appears in kinematic fusion. There the proposed gain/loss ratio is

$f(m)=\sigma(m)-\delta(m),$6

and the central design principle is to recover the kinetic energy of elastically scattered ions before it thermalizes, because Coulomb scattering overwhelmingly dominates fusion events (Tsiper, 2020). Here NGM is not an information or scheduling problem but an energy-recapture problem.

These engineering cases show that NGM can denote either a formal optimization target or an experimentally measured net-amplification condition. In both roles, explicit accounting of overhead and loss is decisive.

6. Semantic heterogeneity, caveats, and research directions

The semantic range of “net gain” is widest in QCD, where the net quark number gain is not a maximization problem at all but a thermodynamic response observable. It is defined through the chemical-potential dependence of the Polyakov loop and anti-Polyakov loop,

$f(m)=\sigma(m)-\delta(m),$7

and it distinguishes meson-like, baryon-like, and deconfined screening regimes in heavy-quark QCD and, qualitatively, in low-temperature real QCD (Surkau et al., 30 Sep 2025). This is a substantive warning against treating every “net gain” phrase as an optimization problem.

The same caution applies across engineering and game-theoretic models. Flow-based network creation games maximize flow/connectivity benefit subject to a hard budget and therefore should be separated from explicit subtractive NGM; all agents use all of their $f(m)=\sigma(m)-\delta(m),$8 budget units in any Nash equilibrium, which is behavior characteristic of constrained gain maximization rather than benefit-minus-cost tradeoff (Echzell et al., 2020). In photonics, the distinction between internal and external net gain changes the meaning of the reported number. In measurement papers, apparent net gain may be inflated by cavity feedback, cracks, or reference-plane misalignment rather than by the intended mechanism.

This suggests a useful taxonomy for future work: objective NGM, residual NGM, constrained-gain analogues, and observational net-gain quantities. Several open directions already appear in the cited literature. In QCD, the proposed extensions include color-sensitive correlators such as $f(m)=\sigma(m)-\delta(m),$9, possible diquark signatures with value $\sigma(m)$0, and behavior near a critical point (Surkau et al., 30 Sep 2025). In social-network promotion, the current bipartite, linear-cost, unit-value activation model is a rigorous special case, and broader revenue heterogeneity or richer diffusion rules remain outside that formulation (Gu et al., 2022). In 2-neighborhood network creation, natural extensions include larger benefit radii $\sigma(m)$1, heterogeneous node values, and bilateral link consent (Haye et al., 10 Feb 2025). In dynamic influence games, the present equilibrium concept is open-loop; feedback or closed-loop equilibria are not part of the solved model (Etesami, 2021). In integrated OPAs, optimized edge couplers with insertion losses below $\sigma(m)$2 dB are expected to push net gain beyond $\sigma(m)$3 dB (Kao et al., 5 Feb 2026).

As a research category, NGM is therefore best viewed not as a single theory but as a recurring analytical question: what remains after all relevant acquisition, coordination, propagation, or switching costs are charged to the benefit term, and how should decisions be chosen once that net quantity is the object of interest?