Social physics

Definition

• Social physics is a broad, operational collection of active research topics in which physicists substantially contribute to resolving societal problems.The authors favor an inclusive, practical definition based on what physicists actually do.Source: §1 · pp. 5–6

Intellectual roots

• Probability and statistics moved from social science into statistical physics, while physicists now apply physical methods to quantify social phenomena.This trajectory is presented as a completed circle between social inquiry and physics.Source: §1 · p. 5

Multidisciplinarity

• Physics’ quantitative successes have encouraged other disciplines to adopt physical and mathematical modelling, placing physics at the roots of modern multidisciplinarity.Examples span ecology, epidemiology, virology, neuroscience, medicine, psychology, and sociology.Source: §1 · pp. 5–6

Scope of the review

• The review examines topics enabling modern life, including urban dynamics and traffic, financial markets, cooperation, social networks, and human-machine networks.These topics are linked to cities, markets, cooperative civil life, network organization, and increasingly intelligent computers.Source: §1 · p. 7
• It also addresses threats to modern life, including crime, migration, contagion, chemical pollution, and climate change, highlighting physics’ versatility across societal challenges.The review connects these themes to humanitarian crises, xenophobia, disease transmission, environmental risks, and Earth’s climate system.Source: §1 · pp. 7–8

2.1. The definition of a city

• City definitions remain nontrivial: research commonly uses administrative boundaries, while density-threshold and commuting-based algorithms define cities from population distributions and flows.The density method joins wards above a threshold ρ0, while the commuting method recursively groups wards around populous seed regions according to commuting fractions.Source: §2.1 · pp. 8–9

2.2. Size of cities

• City-size distributions often follow power laws, and models reproduce them through multiplication-diffusion or rich-gets-richer mechanisms, with direct empirical support for cumulative growth advantages.The multiplication-diffusion model produces emergent power laws across a broad parameter range, while larger cities attracting more people provides a common mechanism and has empirical support.Source: §2.2 · pp. 10–11

2.3. City growth

• Models of spatial city growth reproduce many real features but can misrepresent development history, notably by embedding old regions entirely within newer built environments.Alternative approaches include reaction-diffusion and diffusion-limited-aggregation-inspired models, some incorporating urban-planning rules or geographic data.Source: §2.3 · pp. 11–12

2.4. Networks within and of cities

• Urban-network models span idealised city systems and internal street growth: Christaller’s theory predicts hierarchical hexagonal settlements, while link-cost models interpolate between infrastructure-like and airline-like networks [70, 71, 77].Street-growth models connect centres by minimising cumulative distances, with tree-like structures developing loops as road density increases.Source: §2.4 · pp. 12, 14, 15, Figure 4 · p. 14

2.5. Segregation

• Schelling’s model shows that spatial constraints can generate strong segregation even when residents tolerate substantial proximity to another ethnicity.The final same-ethnicity neighbour fraction increases non-linearly with both the tolerance threshold b and occupancy f, when the dynamics converge.Source: §2.5 · pp. 15–16

2.6. Scaling theory of cities

• City functions scale differently with size: collaboration-intensive sectors such as research and development scale superlinearly, whereas spatially distributed facilities such as gas stations scale sublinearly.Related work reformulates scaling using population density and models the parameter regions in which a city can exist.Source: §2.6 · pp. 16–17

2.7. Human mobility / 2.8. Trajectory analysis

• Human movement is highly structured: gravity and mechanistic radiation models describe flows, while trajectory data show that the next location can be predicted with around 90 % probability and that detour indices approach 1.3 after 20-30 km.The detour index can exceed two for very short trips, and city form influences routes through measures such as inness; highway ring-road cities tend to have negative inness.Source: §2.7 · p. 17, §2.8 · pp. 18–19

2.9. Navigability / 2.10. Future outlook

• Navigability can be quantified with greedy navigators, whose average route length relative to the shortest path defines a navigability index and exposes failures in difficult layouts.The section presents urban physics as an observation-driven field whose future goals include sustainable, energy-efficient cities and anticipating challenges and opportunities from urbanisation.Source: §2.9 · pp. 19–20, §2.10 · p. 20

3. Traffic flows

• Traffic-flow research treats vehicles as a complex system with emergent behavior across spatial and temporal scales, motivated more by qualitative explanation than forecasting.Models range from continuous density descriptions and car-following theories to cellular automata, linking traffic research with statistical physics.Source: §3 · pp. 20–21, §3.1 · p. 21

3.1. Observed phenomena

• Traffic models capture self-organized phenomena including free-flow and congestion phases, phantom jams moving opposite traffic, hysteresis, and the pinch effect.Flow–occupancy relations are tent-shaped, while small jams can merge into larger wide moving jams.Source: §3.1 · pp. 21, 23

3.2. Traffic-flow models

• Coupled-map lattices reproduce empirical flow-density relations, while the Nagel-Schreckenberg model reproduces flow-density curves and phantom traffic jams despite its simplicity.More stylized reductions connect traffic models to non-equilibrium statistical mechanics but cannot reproduce all statistical features of highway traffic.Source: §3.2 · pp. 27–29

3.3. Pedestrian traffic flows

• Pedestrian traffic differs from vehicular flow and is modeled as semi-intelligent particles whose interactions explain trails, lanes, and escape panic.Models typically include social repulsion and physical contact forces; escape panic additionally requires tangential and radial force components.Source: §3.3 · p. 29

3.4. Future outlook

• New datasets from autonomous vehicles, drones, and tower-mounted cameras may reveal statistical laws that test whether mechanistic models explain phantom traffic jams.The future direction is to use high-quality observations of congestion onset to measure jam frequency and refine minimal explanatory models.Source: §3.4 · pp. 29–30

4. Econophysics

• Econophysics replaces fixed rational-choice preferences with interacting-agent models and concentrates empirical research on data-rich markets, firms, trade, and money flows.Scarce, short, noisy macroeconomic datasets have received less attention, although networks may expand econophysics’ future role in macroeconomics.Source: §4 · p. 31

4.1. The advent of econophysics

• Econophysics emerged through financial time-series findings: Mantegna reported power-law price changes [151], and Mantegna and Stanley found rescaled symmetric power-law tails across time scales.The field subsequently gained institutional recognition through meetings, journal publication, monographs, and expanding research beyond financial markets.Source: §4.1 · pp. 31–32

4.2. Agent-based modelling: The dealer model

• The dealer model shows that deterministic behavior generates noisy dynamics through nonlinear interactions, while trend following produces scale-invariant fluctuations resembling stochastic random walks.With N ≥3 dealers, the price evolution becomes highly nonlinear and chaotic; the model identifies chaos-inducing interactions and trend following as distinct mechanisms.Source: §4.2 · pp. 34–35
• Trend-following dealer dynamics reproduce power-law price changes, volatility clustering through an ARCH mechanism, abnormal diffusion, and empirical dealing-time statistics [151, 158, 159, 165, 166, 167, 168, 169].A time-dependent Poisson process with an estimated T = 150 s explains clustered deal intervals and produces a 1/f power spectrum.Source: §4.2 · p. 35, §4.3 · p. 36
• The stochastic dealer model extends market realism by generating bubbles, intervention-related risk responses, and cross-currency correlations driven by a small number of triangular-arbitrage dealers.Its extensions support assimilating financial time-series data, planning intervention strategies, and predicting subsequent market responses.Source: §4.1 · p. 31, §4.2 · pp. 35–36

4.3. Time-series modelling: The PUCK model

• The PUCK coefficient b(t) distinguishes stable attraction to the moving average from unstable repulsion, with b(t) < −2 producing exponentially growing or declining price fluctuations.This threshold characterizes bubbles and crashes, while cubic potentials generally indicate asymmetric price movements.Source: §4.3 · pp. 37–38
• The PUCK model reproduces power-law price changes, short-time abnormal diffusion, and volatility clustering across various financial markets.It generalizes moving-average dynamics through a time-dependent market-potential function.Source: §4.3 · p. 37
• PUCK potential estimates connect market dynamics to dealer behavior: contrarian dealers yield stable potentials, trend followers unstable potentials, and asymmetric trend following higher-order potentials.The market-potential coefficient can be derived theoretically from the dealer model, linking the potential to trend-following behavior.Source: §4.3 · p. 37, Figure 13 · p. 38

Financial Brownian motion

• Order-book models inject bids and asks onto a price axis, where crossing orders annihilate into the latest market price while non-crossing orders accumulate.Ref. documented empirical laws for order injections and cancellations and proposed a Zero Intelligence model requiring no intelligent dealer strategy.Source: §4.4 · p. 39
• The Financial Brownian Particle model maps accumulated bids and asks to surrounding molecules and the bid–ask gap to a colloidal particle that undergoes random-walk-like price movement.As the particle moves, orders behind it refill the opened space while orders ahead are displaced, consistent with dealer models and trend following.Source: §4.4 · p. 39, Figure 14 · p. 40
• Inner-layer orders drive market-price movements, whereas outer-layer orders provide drag resistance, producing a fluctuation-dissipation relation modeled by a Langevin equation.The drag coefficient normalized by colloidal-particle mass can be estimated from market-price data.Source: §4.4 · p. 40, Figure 15 · p. 41
• 0.05 is the typical Knudsen number for dollar-yen and dollar-Euro markets, but values exceed 0.1 during turmoil, making the continuous-price assumption questionable.Continuous-price modelling is considered valid below 0.01, whereas discrete-time descriptions are needed above 0.1; intermediate values imply a transitional regime.Source: §4.4 · p. 41

Microscopic data: trading logs of individual traders

• Microscopic data from 1,015 traders, including 134 HFTs, directly validates the dealer model’s assumed trend-following behavior and buy-sell spread distributions.The data comprise anonymized order submissions, cancellations, and executions in the USD/JPY market from 5–10 June 2016, with millisecond time precision and $1M minimum volume.Source: §4.5 · pp. 42–43

HFTs as liquidity providers

• HFTs typically maintain two-sided quotes, indicating liquidity provision or market making under the Electronic Broking Services market rules.Their bid-ask spreads can be interpreted as round-trip profit estimates or risk evaluations against adverse selection.Source: §4.5 · pp. 43–44

Trend-following behaviour

• A rescaled master curve holds for the top 20 HFTs, demonstrating a common statistical trend-following relation between future mid-price changes and prior market-price changes.The conditional mean is approximately linear for small prior price changes and saturates for large changes.Source: §4.5 · pp. 44–45, Figure 17 · p. 45
• The conditional variance of HFT mid-price changes is independent of prior market-price changes, so adjustment magnitude is an intrinsic property of each HFT.Thus, historical price changes affect the mean response but not its variance.Source: Figure 17 · p. 45, §4.5 · pp. 45–46

Modelling based on microscopic evidence

• HFTs keep fewer than 10 live orders with typically one volume unit, whereas LFTs use power-law-distributed volumes, supporting distinct trader classes and motivating the model’s assumptions.The dataset’s figure reports 135 HFTs responsible for almost 90% of submissions, while 81.5% of transactions are one-to-one.Source: §4.5 · p. 46, Figure 18 · p. 47
• Trend-following random walks, binary price-matching transactions, and post-trade resubmissions produce collective order-book motion and the layered structure observed in the market.The centre of mass is a slow macroscopic variable, with diffusion proportional to N^-1 without trend following, supporting the micro-macro reduction.Source: §4.5 · pp. 48–49, Figure 19 · p. 49

Financial Liouville equation

• The financial Liouville equation is an exact time-evolution equation equivalent to the original stochastic HFT model, providing the starting point for statistical-physics analysis.It describes the N-body probability density through a linear Liouville operator without approximation.Source: §4.6 · p. 50

Financial BBGKY hieararchy

• BBGKY reduction integrates out most variables to obtain a lowest-order hierarchy in one- and two-body relative-price distributions, with transactions represented by a collision integral.The three-body collision integral is dropped because it becomes irrelevant for large N.Source: §4.6 · pp. 50–52

Financial Boltzmann equation

• The closed financial Boltzmann equation has an analytically solvable leading-order steady state given by a tent function, while its next-to-leading order is also accessible.Closure uses the molecular-chaos mean-field approximation under an appropriate boundary condition for large N.Source: §4.6 · p. 52
• The average order-book profile is analytically derivable for any spread distribution ρ(L), demonstrating that the microscopic HFT model is analytically well tractable.The profile follows from the financial Boltzmann equation after approximating the spread summation by an integral.Source: §4.6 · p. 53

Financial Langevin equation

• By additional coarse-graining, the financial Langevin equation is obtained; when ˜c ⪆1 and ∆˜p*⪅1, the marginal-to-strong trend-following regime, price-change statistics asymptotically obey an exponential law.Source: §4.6 · p. 53

Numerical confirmation

• Numerical simulations closely confirm the analytical order-book profile and exponential price-change distribution, because the model’s collision rule rapidly suppresses repeated pairwise interactions and two-body correlations.This mechanism supports molecular chaos even though the price space is one-dimensional, where mean-field approximations are generally expected to be weaker.Source: §4.6 · pp. 53–54

• The normalised average order-book profile shows excellent agreement with Eq. (37) without additional parameter fitting.The profile is composed of the best prices of individual HFTs.Source: §4.7 · p. 54, Figure 21 · p. 55
• Rescaling price changes by the time-dependent decay length κ collapses two-hourly data onto the exponential master curve predicted by Eq. (35).At one-tick time scales, the distribution follows an exponential law whose κ depends on the selected period.Source: §4.7 · p. 55, Figure 21 · p. 55
• α = 3.6 ± 0.13 characterises the weekly fat-tailed price-change distribution, emerging at longer time scales from superposed two-hourly exponential distributions.The transition from short-scale exponential behavior to long-scale power-law behavior agrees with previous findings.Source: §4.7 · pp. 55–56, Figure 21 · p. 55
• m = 3.5 ± 0.13 describes the decay-length distribution, and α ≈ m confirms consistency between the resulting fat-tail exponents.The decay-length probability density follows Q(κ) ∝κ−(m+1).Source: Figure 21 · p. 55, §4.7 · p. 56

Future outlook: towards market ecology

• Ref. extends a microscopic trend-following model by testing whether HFTs apply exponential moving averages in practice.The analysis uses a regression relation with delayed trend signals, coarse-grained price series, and iteratively estimated nonlinear regression coefficients.Source: §4.8 · p. 56
• 85 % of examined HFTs conform to an exponential moving average, while 15 % use other trading strategies.HFTs divide into short-term trend followers of about 4 ticks (30 s), intermediate-term followers of about 20 ticks (3 min), and long-term followers of about 40 ticks (6 min).Source: §4.8 · p. 57
• Identifying trading strategies is a first step toward understanding market ecology, including how strategies affect liquidity and price formation.Precise understanding of these interactions could make regulatory market simulators plausible and help regulators enhance market liquidity and stability.Source: §4.8 · p. 57
• Scarcity of microscopic data has so far prevented in-depth market-ecology analysis, though the authors argue that the field is becoming attainable.Source: §4.8 · p. 57

5.1. Social dilemmas

• Evolutionary game theory explains social dilemmas through selection between cooperation and defection, with prisoner’s dilemma defection as an ESS but coexistence of cooperation and defection in the snowdrift dilemma.Direct reciprocity can transform a prisoner’s dilemma into a stag-hunt dilemma when continuation conditions are sufficient, making cooperation an ESS.Source: §5.1 · pp. 58, 59, 61
• Dilemma-strength parameters reduce dyadic games to two dimensions while preserving evolutionary outcomes, and cooperation-promoting mechanisms operate through positive assortment.Games with identical dilemma strength are evolutionarily equivalent even when their four payoffs differ; an efficiency-deficit measure extends the framework to more complex dilemmas.Source: §5.1 · pp. 60–62

5.2. Cooperation in networks with pairwise interactions

• Network structure can promote cooperation in the prisoner’s dilemma, but it decreases cooperators in the snowdrift dilemma relative to well-mixed populations.Under weak selection, cooperation is favored when b/c > k, while link-based selection can sustain high cooperation across broad game setups and produce dynamic strategy interchange [240, 252, 255, 256].Source: §5.2 · pp. 63–66

5.3. Cooperation in multilayer networks

• In multilayer public-goods games with coupled utilities, cooperation is strongly promoted in one layer but not the other because coupling dampens exploitation of cooperators in the former layer [277].Multilayer networks represent interdependent layers with intralayer and interlayer links, allowing social connectivity and evolutionary interactions to differ across layers.Source: §5.3 · pp. 66–68, Figure 24 · p. 68
• Only intermediate threshold values E produce full synergy between multilayer networks, as cooperative interdependencies seed clusters that eventually take over both layers.Large E disconnects the layers, whereas small E permits defectors to sustain interdependent defectors and produces suboptimal synergy.Source: §5.3 · p. 69, Figure 25 · p. 70

5.4. Cooperation in temporal networks

• Temporal networks can outperform static networks when evolutionary dynamics are fast, but cooperation is most resistant to defection at intermediate aggregation periods and can be undermined by bursty interactions.Increasing g improves cooperation beyond static networks, whereas burstiness and unfavorable values of g and ∆t can nullify these benefits.Source: §5.4 · pp. 71, 73, 74, Figure 27 · p. 73

5.5. Cooperation in networks with higher-order interactions

• Hypergraphs permit nonzero cooperation below the well-mixed threshold r > 1, but full cooperation still requires r > 1 and increasing connectivity makes them approach well-mixed behavior.Scale-free hypergraphs can make cooperation harder, while moderate hyperdegree heterogeneity has no effect; empirical APS data indicate optimal collaborations usually contain two to five scientists.Source: §5.5 · pp. 75, 76, 78, Figure 28 · p. 76

5.6. Empirical facts about human cooperation

• Human cooperation experiments produce ambiguous support for payoff-difference imitation and reveal phenotypes, reciprocal responses, history dependence, and greater cooperation among acquaintances without incentives.Peer punishment and network reciprocity also show mixed or constrained experimental support, while some evidence favors imitation of more successful others.Source: §5.6 · pp. 79–81
• Reward opportunities can double cooperation even when volunteers almost never reward others, because the decoy effect makes ordinary cooperation appear more attractive than defection.Cooperation increases from the start, before rewards occur, illustrating how cognitive biases can override apparent incentives and complicate evolutionary models.Source: §5.6 · pp. 81, 83
• Frequent, free network updating sustains high cooperation by enabling cooperators to avoid defectors, promoting positive assortment without punishment costs or retaliation risks.Static, randomly shuffled, or infrequently rewired networks do not reliably maintain cooperation; dynamic networks also avoid second-order social dilemmas.Source: §5.6 · pp. 82–83
• Psychological factors can override rational incentives: individuals differ in cooperativeness, intuitions favor acquaintances over strangers, and punishment may serve intimidation or retaliation.Relative payoff performance guides decisions when context obscures others’ intentions, but direct responses to others’ actions can become more important when intentions are clear.Source: §5.6 · pp. 82–83

5.7. Future outlook

• Connecting cooperation theory with experiments is hindered by psychology’s and behavioral economics’ replication crisis, demanding stricter significance criteria, larger samples, preregistration, and broader collaboration.These standards improve rigor but increase the logistical, time, and effort costs of behavioral research.Source: §5.7 · pp. 83–84
• Strong theoretical expectations are needed to guide hypotheses and scrutinize surprising findings; without them, experiments may produce disconnected knowledge despite impeccable methodology.The cooperation field is not in that dire state, but evolutionary game theory can still generate many tweaks and scenarios without approaching general principles such as positive assortment.Source: §5.7 · p. 84
• Future models should separately target societal incentives, such as fighting corruption or encouraging vaccination, and general evolutionary principles such as reciprocal altruism or social norms.This division should narrow experimental hypotheses, strengthen empirical support for cooperation models, and yield more definitive answers about when and why humans cooperate.Source: §5.7 · pp. 84–85

6. Networks and communities

• Network science encompasses applications from social and animal networks to finance, infrastructure, climate, medicine, and epidemiology, with community structure forming one focused branch.The field also studies random walks, synchronization, cooperation, social balance, innovation, and other network dynamics.Source: §6 · p. 85

6.1. Community detection: contexts and methods

• Community detection simplifies complex networks, informs understanding of their dynamics and measurement errors, and requires methods chosen according to context.Cut-based, clustering, stochastic-equivalence, and dynamic methods address different goals; generative models provide a statistical blueprint for assessing missing or unexpected links.Source: §6.1 · pp. 86–87, Figure 30 · p. 88

6.2. Introducing stochastic block models

• SBMs model community structure through statistical inference and extend naturally to overlapping, weighted, multilayer, temporal, attributed, and annotated networks.Their parameters describe network structure and node relationships, supporting analysis and prediction across diverse topological and dynamical settings.Source: §6.2 · pp. 88–89

6.3. Defining stochastic block models

• Canonical SBMs impose degree and link-count constraints on average, whereas microcanonical SBMs impose them exactly; degree-corrected SBMs additionally model node-degree heterogeneity.Canonical and microcanonical forms become asymptotically equivalent when node degrees and intercommunity link counts are sufficiently large, but can differ substantially in small or sparse networks.Source: §6.3 · pp. 91–94

6.4. Statistical inference of communities

• SBM inference seeks the node partition that maximizes likelihood, but naive optimization can produce singleton communities, making estimation of the number of communities essential.K may be selected parametrically using penalties such as BIC or MDL, or determined internally through non-parametric approaches, including Peixoto’s microcanonical formulation.Source: §6.4 · pp. 95–96
• Bayesian marginalisation shows canonical and microcanonical formulations are equivalent in expectation, although their total number of links may fluctuate in the canonical form.The microcanonical likelihood and uniform prior are recovered within the Bayesian formulation, while the remaining prior differs only through link-count fluctuations.Source: §6.4 · p. 99
• The MDL interpretation makes the non-parametric microcanonical SBM an implementation of Occam’s razor: model complexity penalizes additional communities and prevents each node becoming its own community.The description length separates the bits needed to describe the network given parameters from those needed to describe the model, with the latter penalizing parameter growth.Source: §6.4 · pp. 99–100
• Nested stochastic block models address the resolution limit by recursively using higher-level multigraphs as priors, enabling community detection in large assortative or disassortative networks.A flat microcanonical SBM merges 64 disconnected 10-node cliques into 32 communities, whereas the nested construction repeatedly models community structure at higher levels.Source: §6.4 · p. 100
• Community-affiliation inference commonly uses MCMC methods, including Metropolis and Metropolis-Hastings, alongside variational and greedy approaches for complicated posterior distributions.MCMC avoids requiring the target distribution’s normalisation constant; Metropolis-Hastings additionally relaxes the symmetric-proposal requirement.Source: §6.4 · pp. 101–102

6.5. Future outlook

• Future community detection increasingly uses topic models, matrix factorisation, and deep neural networks, but generative statistical methods are expected to remain strong because of their rigour.Machine-learning methods are likely to prosper for very large, multilayer, dynamic, and incomplete networks, while deep methods weaken the connection to generative processes and statistical inference.Source: §6.5 · pp. 102–104

7. Human-machine networks

• Data-driven modelling can support socioeconomic monitoring, public health and wealth, safety, crime control, epidemic mapping, disaster management, social inclusion, and social-good algorithms.The approach seeks expressive-yet-simple models calibrated and validated against real-world data rather than black-box models that neglect social-system principles.Source: §7 · p. 107
• Human-machine networks create interconnected socio-technical systems whose scalability must be balanced with robustness, security, transparency, privacy, accountability, fairness, and human oversight.The black-box society reflects limited public awareness of how centralised platforms and opaque AI systems use data and affect lives; Trustworthy AI frames responses as lawful, ethical, and robust.Source: Figure 34 · p. 106, §7 · pp. 107–108
• Research on human and machine behaviour remains fragmented across Human-Imitative AI, Intelligence Augmentation, and Intelligent Infrastructure, which should be integrated to study large-scale human behaviour.The chapter surveys human and machine behaviour, explains AI methodology, reviews social-good applications, and uses AI agents in game theory to examine cooperation.Source: §7 · p. 108

7.1. Literature walkthrough

• The literature walkthrough organizes research into general AI techniques, AI for societal challenges, AI in social physics, and collective human-machine behaviour, while machine-behaviour studies span individual machines, machine networks, and human-machine networks.Human-machine networks include public-resource computing, crowdsourcing, web-search engines, crowdsensing, online markets, social media, multiplayer games and virtual worlds, and mass collaborations, with hybrid types increasingly emerging.Source: §7.1 · pp. 109, 111, 112
• Deep reinforcement learning is extending from single-agent to multi-agent settings focused on emergent behaviour, communication, cooperation, and agents modelling agents, with cooperative tasks especially relevant to social physics.The walkthrough also covers supervised, unsupervised, semi-supervised, generative, explainable, and physics-informed machine-learning research, including statistical-mechanics insights into deep-learning models.Source: §7.1 · pp. 109–111

7.2. Fundamentals of artificial intelligence

• Current AI research realizes narrow, task-specific intelligence rather than artificial general intelligence, while machine learning learns predictions and decisions from data by generalizing beyond training samples.The fundamentals cover supervised, unsupervised, reinforcement, and semi-supervised learning, with semi-supervised methods addressing costly or incomplete labels and enabling graph applications such as community detection and link prediction.Source: §7.2 · pp. 112, 113, 115, 116
• Machine-learning performance depends on the hypothesis set, error function, optimization procedure, dataset size, and complexity trade-offs that determine bias, variance, generalization, scalability, and robustness.Overly simple models cannot represent relevant knowledge, whereas overly complex models can fit noise; optimization may also fail because of local optima and limited computational search.Source: §7.2 · pp. 113–115, Figure 36 · p. 114
• Reinforcement learning agents learn policies that maximise long-term cumulative rewards through feedback while balancing exploitation of known actions against exploration of potentially better alternatives.The framework models states, actions, transition probabilities, and rewards in complex or uncertain environments where the optimal policy is not known beforehand.Source: §7.2 · p. 116, Figure 37 · p. 117

7.3. Deep learning

• Deep learning automates feature extraction through multilayer neural networks, reducing reliance on handcrafted domain expertise and achieving human-level or better performance across several tasks.Its learned sparse representations reduce redundant features and make downstream processing less intensive, but popular methods remain highly data-inefficient.Source: §7.3 · pp. 117–119

7.4. Learning to learn

• Meta-learning exploits previous learning experiences to generalise rapidly to novel tasks, with methods achieving human and superhuman performance on simple one-shot classification.Few-shot learning trains across varying classes, while optimisation-based approaches learn shared initial parameters that enable quick task specialisation with fewer examples.Source: §7.4 · pp. 121–122

7.5. AI agents for promoting cooperation

• Learning with opponent-learning awareness induces tit-for-tat and cooperation among reinforcement learners, while reinforcement learning and evolutionary algorithms construct strong strategies across iterated social dilemmas.In a 5 % noise tournament comparing 176 strategies, LookerUp performed best in the standard tournament, whereas tit-for-tat won no matches in that particular tournament.Source: §7.5 · pp. 123–124
• Human-machine cooperation depends on agents that are general, flexible, and fast-learning, but revealing that an opponent is non-human can reduce human cooperativeness, creating a transparency-efficiency tradeoff.Interpretable algorithmic decision-making is proposed as a way to support stakeholder agency and trust.Source: §7.5 · pp. 125–126
• Low-noise non-human agents placed centrally in a constrained colour-coordination network improved conflict resolution, success rates, and solution speed by nudging humans toward productive deviations.The section also warns that technology-mediated human interactions can enable deception or manipulation, motivating human-centric solutions that foster beneficial interactions and avoid harmful ones.Source: §7.5 · pp. 126–127

7.6. Future outlook

• Successful integration of human-machine networks requires mutual communication, trust, clarity, understanding, and human-centric explainability, alongside methods to assess AI’s sustained social, cultural, and political effects.These safeguards are presented as prerequisites for using AI on societal issues including sustainable development, pandemic response, environmental sustainability, social-media abuse, public safety, and crime prevention.Source: §7.6 · p. 127

8.1. The broken windows theory

• Crime remains recurrent and difficult to suppress: target heterogeneities, learning, and feedback can generate spatially localised hotspots and repeat and near-repeat victimisation, with enhanced odds of subsequent victimisation.Crime dynamics are complex, making mitigation and displacement highly non-trivial; stronger punishment does not necessarily reduce crime.Source: §8 · p. 128, §8.1 · p. 130

8.2. Crime hotspots

• Crime hotspots arise from nonlinear feedback between offenders’ biased movement and dynamically increasing attractiveness after victimisation, producing multiple burglary regimes.The model represents repeat and near-repeat victimisation through a decaying dynamic attractiveness field and generates complex aggregation patterns resembling observed hotspots.Source: §8.2 · pp. 131, 133
• Only subcritical crime hotspots can be permanently eradicated by suitable suppression, whereas supercritical hotspots are displaced and re-emerge nearby.This result follows from coupled reaction-diffusion equations for attractiveness and offender density, with suppression applied at selected locations and times.Source: §8.2 · p. 134, Figure 42 · p. 134

8.3. Crime as a self-exciting point process

• Self-exciting point processes outperform fixed-kernel hotspot maps by balancing exogenous and endogenous crime contributions and inferring patterns directly from data.The method models parent crimes, distance-decaying offspring events, and background variation from weather, seasonality, and time of day.Source: §8.3 · pp. 135–136

8.4. Social dilemmas of crime

• Informants can drive a crime-dominated society to a crime-free state when at least some are initially present, although optimal recruitment depends sharply on costs, parameters, and resource constraints.Simulations and deterministic analysis find utopia for I0 > 0 regardless of δ, θ, and ϵ, while optimal recruitment strategies vary with conditions and added information yields only marginally lower costs.Source: §8.4 · p. 139
• The three-strategy spatial inspection game exhibited continuous and discontinuous phase transitions, including a robust cyclic-dominance phase where criminals, inspectors, and citizens coexist.The model used Monte Carlo simulations on a periodic square lattice, with strategy adoption incorporating uncertainty through a finite K.Source: §8.4 · pp. 141–142

8.5. Criminal networks

• Criminal-network interventions can backfire: targeted disruption attempts in Dutch cannabis networks were unsuccessful or increased network efficiency through reorganisation, implying that intervention is most effective early in network growth.Criminal connectivity is also associated with higher crime rates, while biased Lévy-walk modelling reproduced actual gang networks and helped infer rivalry structures.Source: §8.5 · pp. 143–144, Figure 45 · p. 144

8.6. Rehabilitation and recidivism

• Balanced punishment and rehabilitation maximized reform: for N = 400, h = 0.3, τ = 1.5, and θ = 0.35 produced the highest P/U ratio under finite resources.Excessively harsh or lenient punishment was less effective, while impactful intervention was especially important during offenders’ first stages of return to society.Source: §8.6 · p. 151, Figure 48 · p. 152

8.7. Rosy outlooks for less crime

• Physics-inspired criminology supports future models with heterogeneous cities, evolving social networks, richer strategies, alternative decision rules, and self-organising criminal-network growth to improve crime mitigation.The reviewed work argues that crime policies should account for interdependence and emergent collective dynamics, while rehabilitation and recidivism remain comparatively sparse research areas.Source: §8.7 · pp. 151–153, §8.6 · p. 151

9. Migrations

• 58% of EU citizens expressed immigration concerns in the second half of 2015, compared with 24% one year earlier and 16% two years earlier, indicating a possible collective tipping process.The concentration of public concern around one dimension parallels long-range order and correlated behaviour in physical systems.Source: §9 · p. 156

9.1. Tolerance

• Mean-field and numerical results agree for predicting radicalisation and critical thresholds in an artificial society where newcomer fitness and local tolerance govern population dynamics.The model represents locals and newcomers as network-linked populations whose fitness ratio controls newcomer growth, while tolerance saturation can trigger radicalisation or integration.Source: §9.1 · pp. 159–160
• Immigration outcomes range from mutualism to newcomer dominance or antagonism, depending on local tolerance and newcomers’ integration rate.Sufficient tolerance with reasonable integration supports mutually beneficial interaction; slow integration can produce newcomer majority, while slow integration combined with low tolerance can produce turmoil and violence.Source: §9.1 · p. 160, Figure 52 · p. 162

9.2. Integration and culture

• Immigrant earnings initially lag natives by 17%, the gap disappears after 15 years, and earnings reach 11% above natives’ after 30 years in early cross-sectional analysis, whereas Borjas judged parity extremely unlikely.The contrasting conclusions show that measured economic integration depends on how immigrant cohorts and changing skills are accounted for.Source: §9.2 · p. 161
• Axelrod’s model shows that local cultural convergence can coexist with global polarisation, making immigrant integration vulnerable to exclusion without concerted government effort.The model uses multidimensional cultural traits, interaction probability proportional to cultural similarity, and adaptive bottom-up influence; crises may intensify polarisation by narrowing salient cultural dimensions.Source: §9.2 · p. 164

9.3. Populism and polarisation

• When annual immigrant inflow exceeded 0.4% of a country’s population, right-wing populist voters increased annually by 1%–5% during 2014–2016.The relationship depended on both immigrant prevalence within a country and total inflow into the EU, and the proposed network model explains it through local outsider exposure and tipping points.Source: §9.3 · pp. 165–166
• A 0.5% annual outsider inflow produces a discontinuous jump in right-wing populist supporters after about 37 years, with identical 5% shocks becoming disproportionately stronger near the tipping point.Global influences can seed populist ideas, but local influences drive the initial increase; near the tipping point, opinion contagion overtakes them and populist supporters dominate society.Source: §9.3 · pp. 167–168
• In opinion dynamics, consensus requires sufficiently large tolerance threshold d, whereas smaller d values produce two or more persistent competing opinions.In lattice networks, lower thresholds yield sizeable clusters of distinct opinions, while opinions within each non-percolating cluster remain similar but not identical.Source: §9.3 · p. 168, Figure 54 · p. 169
• Limited communication across migration-related rifts can polarise society, with partisanship, religion, geography, and echo chambers also reinforcing divisions, though online effects remain debated.Polarisation can increase political participation as sub-populations reinforce their own positions.Source: §9.3 · p. 169

9.4. Future outlook

• An interdisciplinary, data-driven agenda should combine network science, statistical physics, evolutionary game theory, Bayesian statistics, econometrics, and machine learning to address migration’s globally interdependent dilemmas.Countries face complementary choices over accepting immigrants and integration capacity, or replenishing populations amid emigration and future socioeconomic instability.Source: §9.4 · pp. 170–171
• Future migration models should represent how inflow speed interacts with native acceptance and migrant integration, because moderate inflows may permit adaptation while extreme inflows can trigger antagonistic responses.Tolerance is therefore relative to inflow magnitude, and failing to model this dynamic may allow right-wing populism to threaten liberal and democratic values.Source: §9.4 · pp. 170–171

10.1. Data-driven metapopulation models

• Metapopulation models represent local transmission within populations and transmission between populations caused by movements of infectious individuals, often modeled through transportation networks.They can incorporate stochastic movement, commuting memory, socioeconomic stratification, age-related factors, and digital data describing population flows, including mobility reports and travel by aeroplanes, railways, or highways.Source: §10.1 · pp. 173–175

10.2. Digital epidemiology

• Search and social-media data support near-real-time disease surveillance and prediction, while ensemble models improve forecasts by combining complementary digital, health, and epidemiological data sources.Google query frequencies closely correlated with influenza cases reported by US public-health agencies. Twitter achieved 85% accuracy for weekly influenza-prevalence changes, and Brazilian dengue prediction reached a national correlation coefficient of 0.98.Source: §10.2 · pp. 176, 177, 179
• Digital epidemiology also captures vaccine sentiment, adverse events, human movements, social-network proximity, and disaster displacement, but individual-source methods face changing search behavior and other limitations.Negative vaccine sentiment is more contagious than positive sentiment; analysis of 2 billion tweets identified 27 drug-related adverse events; Facebook data linked Covid-19 spread to regional social connections.Source: §10.2 · pp. 176, 178, 179

10.3. Analytical results from metapopulation models

• Network heterogeneity lowers epidemic thresholds, motivating hub-focused vaccination strategies, while metapopulation analyses extend threshold results to recurrent mobility patterns.The epidemic arrival time measures the timing of the first imported case and supports early inference of epidemiological parameters.Source: §10.3 · p. 181
• A closed-form epidemic-arrival-time distribution was derived for increasingly complex metapopulation networks, under Poisson exportation and early exponential growth assumptions [890].The analytical expression was compared with simulations across epidemic scenarios whose doubling and generation times ranged from 3 to 30 days, covering Covid-19 and Ebola timescales.Source: §10.3 · pp. 181–182
• The metapopulation model derives non-homogeneous Poisson exportation, EAT distributions and expectations, and recursive relations for later arrivals.The first-n infection exportation intensity is αij exp(λit), while conditional later-arrival distributions can be reinterpreted as EAT distributions with an updated seed size.Source: §10.3 · pp. 183–184
• Shortest-path-tree EAT analytics require adjustments for hub effects, continuous seeding, and longer-path reduction, with numerical validation of the resulting framework.Hub adjustment changes the effective local growth rate while preserving the exportation process, and continuous seeding accounts for infections imported after the first arrival.Source: §10.3 · pp. 186, 187, 189, Figure 59 · p. 187
• The combined analytical framework accurately estimates epidemic arrival times for almost all populations in the worldwide air-transportation network.It combines two-population analytics with hub-effect and continuous-seeding adjustments, path reduction, and path superposition.Source: Figure 60 · p. 190, §10.3 · p. 190

10.4. Bayesian inference

• Bayesian inference combines prior distributions with likelihood functions, but flat or diffuse priors can bias estimates or cause convergence failures in high-dimensional problems.Weakly informative priors may be preferable when many parameters must be inferred.Source: §10.4 · pp. 191–192
• R0 estimates for the 2009 influenza A(H1N1) epidemic were 1.65 with 95 % CI [1.54, 1.77], 1.75 with 95 % CI [1.64, 1.88], or 1.89 with 95 % CI [1.77, 2.01], depending on outbreak start date.Bayesian inference produced the same estimates as the GLEAM simulator while avoiding high-performance computing requirements.Source: §10.4 · pp. 192–193, Figure 61 · p. 194
• A second Bayesian case study models Covid-19 spread from Wuhan using exponentially growing infections inferred from the first 425 confirmed cases and a mean detection delay of D = 10 days.The delay includes a 5–6-day incubation period and a 4–5-day delay from symptom onset to detection.Source: §10.4 · pp. 193–194
• A Bayesian mobility model estimated Wuhan’s Covid-19 doubling time at 7.31 days (95 % CrI [6.26, 9.66]) and R0 estimates mostly near 2.38–2.7.The model used international exportation data, Markov Chain Monte Carlo sampling, and mobility-based risk calculations before Wuhan’s 23 January 2020 lockdown; 131 Chinese cities had over 50% importation risk by 22 January.Source: §10.4 · pp. 195–196, Figure 62 · p. 197

10.5. Challenges and future work

• Intracity mobility can reveal epidemiological disparities hidden by homogeneous intercity metapopulation models, including higher infection rates among disadvantaged racial and socioeconomic groups.These groups have relatively little control over reducing mobility and consequent exposure to infectious diseases.Source: §10.5 · pp. 196, 198
• Age-assortative contact patterns leave younger populations potentially more vulnerable to infectious diseases, although biological factors can attenuate this vulnerability.School children and young adults tend to mix with similarly aged people, while early SARS-CoV-2 variants provided an example of an attenuating biological circumstance.Source: §10.5 · p. 198
• About 19 % of infectious individuals seed 80 % of Covid-19 local transmissions, so models incorporating infectiousness heterogeneity predict rarer but more explosive outbreaks.SARS-CoV and SARS-CoV-2 transmission follows fat-tailed secondary-infection distributions, requiring models to represent heterogeneity in contacts, infectiousness, and other parameters.Source: §10.5 · p. 198
• Social distancing and rapid isolation can control Covid-19 and shorten serial intervals, making optimization of intervention combinations a major epidemiological-modeling priority.Social distancing alone controlled Covid-19 in China, while fast isolation reduced the time between successive disease onsets in transmission chains.Source: §10.5 · p. 198
• Digital epidemiology needs standardized, purpose-built data systems and visual analytics because general platforms such as Google Trends, Twitter, and Facebook were not designed for epidemiological use.Interactive visual-analytics environments can couple data mining with agent-based epidemic modeling to support public-health decision making.Source: §10.5 · pp. 198–199

11. Environment

• Environmental protection often requires costly investments or land-use restrictions, making social consensus difficult even when degradation costs outweigh economic benefits.Scientific evidence and quantified degradation costs can nevertheless enable action, as shown by international agreements and recovery following evidence on ozone-depleting substances.Source: §11 · pp. 199–200

11.1. Pollution

• Conventional toxicity testing inadequately captures toxic effects because species and individuals differ in sensitivity, laboratory exposures are brief and standardized, environmental conditions vary, ecological feedbacks alter exposure, and mixtures can exacerbate toxicity.Source: §11.1 · p. 201
• Physics-based frameworks can link biological scales and support falsifiable, modular predictions, but current ecotoxicology still falls short of modelling multigenerational, multispecies, heterogeneous environments with interacting toxicants and stressors.AOPs remain unsuitable for environmental ecotoxicology, while DEB models do not yet capture the ultimate multi-population, multi-toxicant goal.Source: §11.1 · pp. 203, 206, 210
• Mechanistic frameworks, especially DEB models, causally link environmental conditions to organismal growth, reproduction, hazard, and toxicant effects, enabling predictions beyond empirical dose-response curves.DEB models predicted bacterial population dynamics at exposures up to seven times those used for fitting and supported toxicokinetic, toxicodynamic, and nano-toxicity analyses.Source: §11.1 · pp. 205, 210

11.2. Physical habitat loss

• Habitat preservation requires balancing human activities with biodiversity conservation, but management often approaches the maximum damage ecosystems can sustain and can reduce resilience to pressures such as climate change.Protected areas and environmental impact assessments remain the most effective tools available, yet tourism and fisheries illustrate conflicts between economic stability and conservation goals.Source: §11.2 · pp. 210–211
• Environmental impact studies incentivize firms to meet minimum standards, effectively managing industrial activity toward the maximum acceptable environmental degradation.Acceptability reflects environmental, economic, social, health, and political considerations, with red lines tied to ecological and health tipping points.Source: §11.2 · pp. 211–212

11.3. Tipping points

• Tipping points occur when small perturbations induce major state changes, and molecular loss of homeostatic control can initiate runaway damage leading to cell or organism death.Ecological tipping points can also occur at population and ecosystem levels, so apparently healthy individuals may coexist with impending population extinction.Source: §11.3 · pp. 212–213
• Above a persistence threshold, increasing predator mortality causes predator extinction, rare adult prey and zooplankton, abundant juvenile prey, and an ecosystem tipping point.Near the threshold, an adverse event affecting adult prey can trigger the same collapse even when predator mortality remains below the threshold.Source: §11.3 · pp. 218–219

11.4. Future outlook

• Environmental science should identify tipping points, quantify externalities, monitor ecosystems, assess risks, and use mechanistic models integrated with data to support precautionary adaptive management.Long-term solutions require interaction among environmental sciences, social sciences, economics, and politics, while increasingly complex models must remain grounded in reality.Source: §11.4 · pp. 219, 221, 222
• The authors recommend replacing the question of how much environmental change is acceptable with constructive goals such as improving environments through targeted, interdisciplinary management.In Croatian national parks, concentrating visitors, using incentives, relocating attractions, and limiting simultaneous visitors reduced degradation while increasing visitors, happiness, and entrance-fee income.Source: §11.4 · pp. 220–221

12. Global climate change

• Climate variables causally affect mortality, agriculture, economic productivity, violence, conflict, and migration across regional and global settings.Examples include a 7.3% mortality increase among rural Indians after a one-standard-deviation increase in high-temperature days and greater permanent outward migration in Indonesia from multi-year temperature changes than from other natural disasters.Source: §12 · p. 223, Figure 72 · p. 224
• Human activities caused more than half of the observed increase in global average surface temperature from 1951 to 2010, according to the IPCC’s fifth assessment report.Fossil-fuel combustion emits greenhouse gases that block convective heat from escaping into space, producing warming.Source: §12 · p. 225
• Mitigating climate risks requires integrating climate science with social sciences and humanities to mobilise human action and build a unified understanding of Earth and its population.Social tipping interventions such as climate education, engagement, and emissions disclosure could foster mitigation.Source: §12 · pp. 225, 227
• Climate extremes and tipping points threaten societies, while early-warning signals may help anticipate abrupt transitions and strengthen resilience.Critical slowing down is indicated by increasing autocorrelation and variance; related signals include detrended fluctuation analysis, power spectra, flickering, skewness, and kurtosis.Source: §12 · pp. 225–227

12.1. Modelling the climate system

• Climate models range from energy-balance to general circulation models, but no single model captures all nonlinear subsystem interactions, limiting longer-term predictive skill.Integrated assessment models address this gap by linking socioeconomic systems, the biosphere, and the atmosphere to simulate climate-stabilisation policies, although complex models may miss nuanced social behaviours.Source: §12.1 · pp. 229, 232, 233

12.2. Climate networks

• Climate networks represent geographical locations as nodes and statistical similarity between climate records as links, enabling analysis and prediction of diverse climate phenomena.Applications include ENSO, extreme rainfall, the Indian summer monsoon, Atlantic circulation and oscillations, teleconnection paths, and CO2 impacts [1103, 1104, 1105, 1106, 1107, 1108, 1109, 1110, 1111, 1112, 1113, 1114, 1115, 1116, 1117, 1118, 1119, 1120].Source: §12.2 · pp. 233–234
• Climate networks quantify links from correlated or synchronised climate time series, threshold weak connections, and analyse the resulting topology to reveal climatic events and teleconnections.Pearson cross-correlation provides link weights, while event synchronisation can determine link strength and time-delay direction.Source: §12.2 · pp. 234, 236, 237
• One year ahead, climate-network methods reliably predicted El Niño events and outperformed state-of-the-art climate models, including the 2014–2016 event.A network alarm required W^y to cross Θ = 2.82 from below while NINO3.4 remained below 0.5 °C; it also predicted the 2018–2019 event one year ahead.Source: §12.2 · pp. 237–238, Figure 74 · p. 239
• More than 60% of extreme-rainfall events in the Central Andes were predicted, rising above 90% during El Niño conditions, by event-synchronisation climate networks.Network divergence identified source and sink nodes and propagation pathways along which extreme events were predictable.Source: §12.2 · pp. 240–241

12.3. Impact of Rossby waves on air pollution

• Rossby waves modulate surface air pollution through alternating cyclone and anticyclone regimes that change vertical motion, winds, air stability, and pollutant accumulation.Strong winds during downwelling and upwelling keep pollution low, whereas the reversal of these regimes weakens winds and permits accumulation.Source: §12.3 · pp. 243–244

12.4. Critical phenomena in the climate system

• The critical exponent for precipitation was universal across climatic regions, with a value of 0.215 ± 0.02, while variance scaling yielded γ/ν = 1.58.The same power-law fit described precipitation above the critical water-vapour value across ocean basins, and finite-size scaling supported the critical-phenomena framework.Source: §12.4 · pp. 245–247
• Atmospheric precipitation undergoes a continuous, apparently self-organised phase transition when water vapour crosses a critical point, as rapid convective buoyancy release overtakes slow large-scale forcing.Below the critical point precipitation is minimal; after crossing it, precipitation increases rapidly, with σ^2_P(w_c; L) ∝ L^-0.42 near criticality.Source: §12.4 · pp. 245–247
• Observations, theory, and climate models indicate that global warming expands the Hadley cell poleward and weakens it, including an observed expansion of approximately 2° from 1979 to 2005.The conclusion draws on satellite observations, a proposed physical mechanism, and 30 CMIP5 coupled-model simulations.Source: §12.4 · p. 247
• Percolation analysis finds a discontinuous first-order transition whose largest connected climate-network component at threshold is located in the tropics, analogous to the Hadley cell.The threshold is identified from the largest jump in the order parameter G1 as links are added, and the result remains discontinuous across grid resolutions.Source: §12.4 · pp. 248–249
• Most CMIP5 models and reanalysis data show poleward expansion of the Hadley-cell analogue and weakening critical link weight across climate scenarios, with strong correlation to conventional estimates.The analogue is the largest connected network component at the percolation threshold; the same qualitative trends appear for Hadley-cell size and intensity, though conventional results are more ambiguous in some cases.Source: §12.4 · p. 249, Figure 78 · p. 250
• Hadley-cell poleward expansion may produce drier tropical or subtropical regions and shift maximum tropical-cyclone intensity poleward, including risks for northern India, southern Africa, and western Australia.Climate-network analysis may help identify regions more likely to experience precipitation decline or hurricane intensification, supporting timely risk mitigation.Source: §12.4 · pp. 249–250

12.5. Future outlook

• Future climate research must integrate climate predictions into social dynamics and model cascading failures from extreme events across critical interdependent infrastructures.Addressing this gap is presented as crucial for building a climate-resilient society.Source: §12.5 · p. 251
• Climate-network similarity measures can obscure physical mechanisms, motivating machine learning, AI, and visual analytics to improve climate understanding and numerical prediction.The black-box structure of climate systems makes mechanism discovery a central challenge for future research.Source: §12.5 · p. 251

• The review’s account is necessarily selective, omitting topics including the physics of art, agriculture, gastronomy, ethnology, and civil unrest, while its definition of social physics may be considered overly broad.The authors suggest that “Human physics” could be a more fitting title for topics affecting humans, especially near engineering, artificial intelligence, and climate modelling.Source: §13 · pp. 251–252
• Open dialogue is essential because physicists’ imperious attitudes can undermine mutual respect, understanding, and the success of multidisciplinary collaboration.Patience and care in resolving cross-disciplinary differences can produce insightful and impactful research.Source: §13 · p. 252
• Before applying quantitative methods, physicists should seek expert input because intuition and common sense can produce simplistic or naive hypotheses and assumptions.Their strengths include rigorous data analysis and complex numerical simulation, but these methods depend on problem-relevant hypotheses and model assumptions.Source: §13 · p. 252
• When quantitative methods are employed, physicists can contribute by identifying broad patterns and simplifying problems while retaining their main processes.Such approximations may conflict with disciplines that prioritize exceptions, so physicists also need to communicate their perspective to collaborators.Source: §13 · p. 252