Complexity Research; Why and How
Sorin Solomon Racah Institute of Physics HUJ Israel
Director, Complex Multi-Agent Systems Division, ISI TurinDirector, Lagrange Interdisciplinary Laboratory
for Excellence In Complexity
Complexity Sorin Solomon,
Racah Institute of Physics HUJ Israel Director, Complex Multi-Agent Systems Division, ISI Turin
MORE IS DIFFERENT (Anderson 72)(more is more than more)
Complex “Macroscopic” properties are often the collective effect of many simple “microscopic” components
(and independent on their details)
Director, Lagrange Interdisciplinary Lab for Excellence In Complexity
Phil Anderson Real world is controlled …– by the exceptional, not the mean;
– by the catastrophe, not the steady drip;
– by the very rich, not the ‘middle class’. we need to free ourselves from ‘average’ thinking.
SAME SYSTEM Reality Models
Complex ----------------------------------Trivial
Adaptive ----------------------------------Fixed dynamical law
Localized patches -----------------------Spatial Uniformity
Survival -----------------------------------Death
Discrete Individuals Continuum Density
Development -----------------------------Decay
Misfit was always assigned to the neglect of specific details.We show it was rather due to the neglect of the discreteness. Once taken in account => complex adaptive collective objects. emerge even in the worse conditions
Complexity Sorin Solomon,
Racah Institute of Physics HUJ Israel and
Director of the Complex Multi-Agent Systems Division, ISI Turin
MORE IS DIFFERENT (Anderson 72)(more is more than more)
Complex “Macroscopic” properties may be the collective effect of many simple “microscopic” components
(and independent on their details)
Lagrange Laboratory for Excellence In Complexity at ISI Torino support for students and researchersGeneral Integration Action in Complexity Science + 12 Specific Targeted Research Projects in Complexity (CO3)
The Multi-Agent Complex Systems Paradigm
MICRO - the relevant microscopic degrees of freedom
INTER - their fundamental interactions
MACRO - the macroscopic emerging collective objectsIntrinsically (3x) interdisciplinary:
-MICRO belongs to one science
-MACRO to another science
-Mechanisms: statistical mechanics (?) phase transitions, scale invariance,
The challenge : transcend traditional disciplinary research Complexity Research: More than a juxtaposition of expertises:
a new grammar with new interrogative forms allowing the formulation of new questions. Grow a new generation of bi- or multi-lingual scientists.
“MORE IS DIFFERENT” Complex Systems Paradigm
MICRO - the relevant elementary agents
INTER - their basic, simple interactions
MACRO - the emerging collective objects
Intrinsically (3x) interdisciplinary:
-MICRO belongs to one science
-MACRO to another science
-Mechanisms: a third science
traders
orders, transactions
herds,crashes,booms
Decision making, psychology
Financial economics
statistical mechanics, physicsmath, game theory, info
Complexity Sorin Solomon,
Racah Institute of Physics HUJ Israel Complex Multi-Agent Systems Division, ISI Turin
Lagrange Interdisciplinary Lab for Excellence In Complexity
MORE IS DIFFERENT (Anderson 72)(more is more than more)
Complex “Macroscopic” properties may be the collective effect of many simple “microscopic” components
(and independent on their details)
Phil Anderson “Real world is controlled …
– by the exceptional, not the mean;
– by the catastrophe, not the steady drip;
– by the very rich, not the ‘middle class’. we need to free ourselves from ‘average’ thinking.”
950C
1Kg
1cm
97
1cm
1Kg
99
1Kg
101
The breaking of macroscopic linear extrapolation
?Extrapolation?
BOILING PHASE TRANSITIONMore is different: a single molecule does not boil at 100C0
Simplest Example of a “More is Different” Transition
Water level vs. temperature
95 97 99 101
Example of “MORE IS DIFFERENT” transition in Finance:
Instead of Water Level: -economic index(Dow-Jones etc…)
Crash = result of collective behavior of individual traders
Chemicals almost free particles
DNA chains, proteins reproduction,evolution,synthesis
Cells chemotaxis, metabolism
Systems, Organisms health, perception self-non-self recognition
CONCEPTUAL AND DISCIPLINARY JUMPS
of the
MORE IS DIFFERENT
type
Statistical Mechanics
Phase Transition
Atoms,Molecules
Drops,Bubbles
Anderson abstractization
Complexity MICRO
MACRO More is different
BiologySocial Science
Brain ScienceEconomics and
Finance
BusinessAdministration ICT
Semiotics and Ontology
Chemicals
E-pages
Neurons
Words
people
Customers
Traders
Cells,lifeMeaning
Social groups
WWW
Cognition, perception
Markets
Herds, Crashes
Instead of temperature (energy / matter):
Exchange rate/interest rate
Value At Risk / liquid funds
Equity Price / Dividends
Equity Price / fundamental value
Taxation (without representation)/ Tea
Reality curves
DVD
VCR
CARS in USA 1895-1930
Product Propagation
Bass extrapolation formula vs
microscopic representation
Actual sales
Extrapolation
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
Microscopic view of a water drop: a network of linked water molecules
The water drop becomes vapors: the network splits in small clusters
The water drop becomes vapors: the network splits in small clusters
The water drop becomes vapors: the network splits in small clusters
The water drop becomes vapors: the network splits in small clusters
The water drop becomes vapors: the network splits in small clusters
Boiling is not a physical property of particular molecules
but a generic property of the cluster geometry
To understand, one does not need the details of the interactions.
Rather one can prove theorems on what is the density of links that
ensures the emergence or disintegration of clusters
Phase Transition
Product Propagation
BASS
VCR
SALES
Bass extrapolation formula vs
microscopic representation
VCR
Extrapolation
Actual sales
Product Propagation
BASS
VCR
SALES
Bass extrapolation formula vs
microscopic representation
VCR
Extrapolation
Actual sales
Also Belief Propagation
Reality curves
DVD
VCR
CARS in USA 1895-1930
Extrapolation
Product Propagation
Bass extrapolation formula vs
microscopic representation
Actual sales
Also Belief Propagation
- Microscopic Customers and Macroscopic Sales
MICRO – Customers, products / ideas / information
INTER – purchase, inform, learn, hear-say
MACRO – global trends, waves of sales (e.g. Tamaguchi), hits, flops, market fluctuations, anomalous diffusion
demarketing
Propagation effects:
- product propagation- spread of ideas
- epidemics - Internet viruses- Social ills: drugs, violence, terror- Credit networks and
bankruptcy avalanches
- production / trade practices
- real estate valuation
- tax paying habits
PotentialAdopters
RejectorsThe Square Lattice is
just for clarityThe effects demonstrated
are much more general
Density of potential adopters: 26/48>50% What Percent will actually adopt?
The Buyers are split in small clusters
The epidemics, bankruptcy avalanche, idea, product spread is limited to one cluster
Density of potential adopters: 26/48>50% What Percent will actually adopt? 7/48 < 15 %
Density of potential adopters: 26/48>50% What Percent will actually adopt? 7/48 < 15 %
Only 15 % will actually adopt! But what if add one more potential adopter?
If adds one more potential adopter 22 out of 27 potential adopters adopt 22/48~46%
Adopters Density 55%
This is not just a fortuitous case;
for larger systems the effect is even more dramatic
55%
55%
If lowering the price , or increasing quality, or decreasing taxes or subsidizing adopters
(or affecting credit rate) etc
one gains 5% more potential adopters Then
density of potential adopters = 60%
How much will this increase the actual adoption?
55%
60%55%
60%55%
60%55%
60% potential adopters
55% potential adopters
60% potential adopters
55% potential adopters 0%adoption55%
60%
59.3
Theorem
55%density
61%
Potential Adopters
Adopters fraction
0% salesPercolation transition
infin
itely
sha
rp
at in
finite
siz
e
Fractal Sales: Prediction Tool for product success (15/17)
fractal space distributionPrediction of campaign success (15/17) Goldenberg
Air-view of a sub-urban neighborhood;
crosses on the roofs indicate air-conditioner purchase
Stock market shock explainedPhysicists model recent trading frenzy.
Market 'spikes' are seen by traders as freak events.Physicists expect them
Small changes in product quality, price, external conditions can produce large effects(e.g. large market fluctuations)
Small deterioration in credit market can trigger large waves of bankruptcies
Stock market shock explainedPhysicists model recent trading frenzy.
Market 'spikes' are seen by traders as freak events.Physicists expect them
Lev Muchnik Phys. Scripta
“Levy, Solomon and Levy's
Microscopic Simulation of Financial Markets
points us towards the future of financial economics.
If we restrict ourselves to models which can be solved analytically, we will be modeling for our mutual entertainment, not to maximize explanatory or predictive power."
--HARRY M. MARKOWITZ, Nobel Laureate in Economics
Percolation transition
From non-sales at all to a lot of sales
Infinitely sharp at infinite size system
ALSO: effects of
Expectations Adaptation
Self tunning to criticality
Fractal fluctuations
and correlations
ALSO: effects of
Expectations Adaptation
Fractal space-time fluctuationsProduct Success prediction (15/ 17)
Resistance
Resistance
Resistance
Resistance
“ANTI-Percolation”
Antivirus
Figure 1 Comparing infection process evolution with (bottom) and without (top)immunization edges. On the top the network is being infected fully by the virus. Onthe bottom the virus cluster is reduced by more than half by introducing immunization edges. The blue (dark green) edges represent the original network (further immunization) edges. During the spread, an edge is coloured in red (turquoise) if it was used to infect (immunize) a node. In both cases we present foursnapshots of each network in different times. In addition, we present the time (t )varying graphs for the cluster development over time.
The blue, red and green lines are used to present the size of the susceptible, infected and immunized clusters,respectively. Note that in the bottom set, initially in snapshots 1 and 2, the virus cluster develops uninterruptedly until the immunization agent manages to escape the border of the virus cluster, in snapshot 3, and start immunizing the network; therefore the agent manages to immunize most of the network even though the virus had a head start.
Figure 2 Relative virus cluster size as a function of immunization link density(log–log scale). The dependence of the relative infected cluster size on the relative edge addition q, resulting from simulations over uncorrelated, scale-free networks with power exponent −3, mean degree 4 and network size 50,000–200,000 nodes. The ratio dependence shows a power-law form, with an exponent close to −4/3. The error bars present the 95% confidence interval.
Figure 5 Comparison of virus cluster sizes for the random-edge and thehoney-pot architectures. The different sets present different relative edge additions, q. The clusters in the random case are always larger than the clusters in the honey-pot case; as the network size grows, so does the gap between the two architectures. The reason behind this is that, whereas in the random case the cluster size remains fairly constant as we vary the network size, in the honey-pot case as the network grows, so does the effectiveness of the honey pots. This effect is mostly noted in the middle range of the density values where the immunization has an effect but the virus cluster is not extremely small. The error bars present the 95% confidence interval. show a power-law ratio dependence
Figure 6 The dependence of the virus cluster on the degree distribution powerexponent. We ran a sensitivity analysis where we varied the power exponent of thePareto degree distribution characterizing the underlying topology between 1.8 and 3,which includes all degree distributions found in real scale-free networks. As can be
seen, the effectiveness of the immunization process grows with the powerexponent, owing to the fact that lower exponents entail a higher density of edges,
which allows the virus to advance faster. However, this variation is still minorcompared with variations in the relative edge addition, q, and in the architecturetype, which are illustrated by the different data sets presented. The error bars
present the 95% confidence interval. HP = honey pot.
Parallel Networks
Consumption;
home
Expectations
work
- Microscopic Investors and Macroscopic Crashes /Power Laws
MICRO - Investors, individual capital ,shares
INTER - sell/buy orders, gain/loss
MACRO - social wealth distribution, market price fluctuations (cycles, crushes, booms, stabilization by noise)
-Microscopic Concepts and Macroscopic Ideas
MICRO - concepts, connections between concepts
INTER - creating/deleting/activating connections between concepts
- Microscopic Seers and Macroscopic Sight
MICRO - motion visual sensors for points and line elements.
INTER - time and space local data integration.
MACRO - Perception of 3 Dimensional global structure.
- Microscopic Picassos and Macroscopic Drawings
MICRO - local line / motion features, mental states, mental eventsINTER - line breaks and mind events(changes) vs line/mind inertia.MACRO - drawing shapes, emergence of representational meaning - Microscopic Doctors and Macroscopic Health
MICRO - Cells, Enzimes, Antigens, Antibodies INTER - producing, destroying, changing state of a cell/enzime, MACRO - immunity, health, infection, sickness, inflamation. -Microscopic Drivers / police and Macroscopic Jams
MICRO - carsINTER - go ahead/give way at intersections.MACRO - traffic flow, jamming; self-organization; useless police
Microscopic Grimm Brothers and Macroscopic Stories
MICRO – persons, relationsINTER – change in relations ; actingMACRO –plot, story, meaning
Internet study along the same lines1. physical,2. information flow and 3. emergent / cognitive.
LAYERS
Micro Macro
1. Cognitive / Social Layer Self-Organization
Content based servicerelationships, ringsPeer-To-Peer nets
Emergence of Collective
Complex Institutions with personality and
interests
2. WEB Information Layer
Sites, links, information storage and flow
Distributed Information
storage, processing retrieval, control, trust
3.INTERNETHardware Layer
Nodes, cables, data packets
Connectivity, robustness
• We exemplify this by studying two types of networks: Genetic networks and the World Wide Web. In the first case we formulate a model including only random unbiased gene duplications and mutations. In the second case, the basic moves are website generation and rapid surf-induced link creation (/ destruction). In both cases we do reproduce the experimental observations at all scales.
• For the genetic network our model predicts a slow convergence toward a directed structure composed roughly of a core directing toward a periphery. By contrast, the WWW presents rapidly changing non-stationary bi-directionally linked clusters.
• In the genetic case the rough picture outside the core is of a tree-like structure with arrows preferentially directed towards the branches. In the WWW case, the hierarchy is rather in the form of strongly linked clusters-within clusters and the connections between clusters are generically pointing in both directions.
Louzoun-LevA
B
C
DC
A
36546, Z=46
B
253270, Z=23
826, Z=2.83 442, Z=2.22
browse
link
Gene network
Internet network
Figure 1 - Mechanisms of individual node evolution: the 3 elementary processes defining our genetic model. Their effect is demonstrated on the configuration of the left upper corner (only links and nodes relevant for the explanation are explicitly shown). The effect of a Node copying elementary event is shown in A) . The blue node is “duplicated” by introducing the new brown node that has the same targets for its out-going links (and no incoming links at all). The Node removal is illustrated in B). The green node and all its links are deleted. The drawing C) illustrates the elementary operation of Link Mutation: the pink link is “copied”. i.e. a new link with the same origin but different target is created . These 3 elementary operations turn out to be sufficient for the formation of a steady state directional hierarchical scale free network, with the experimentally observed sub-motif distribution.
Microscopic Links, Macroscopic NetworksMICRO – Nodes, connections INTERACTIONS - Local changes (node / link (dis-) appearance)MACRO- Global connectivity, percolation, topologyNew generation of network studies:Instead of study generic properties that are not specific to any particular system, Study specific macroscopic collective properties implied by specific elementary interactions..
site links
Copied site /gene
links
site
Transitive link
browse
Gene activation
WWW linking
0 0.5 1 1.5 2 2.5 3 3.5 4 4.50
1
2
3
4
5
6
log10
(Number of Nodes)
log 10
(Ran
k)IncomingOutgoing
Figure 2 Node degree distribution in the genetic model- Incoming (empty circles) and outgoing (full squares) link distributions. The outgoing link distribution is normal (as experimentally observed). The incoming link distribution is scale free over more than three orders of magnitude (10-50,000). The straight line corresponds on this double logarithmic graph to a power law with exponent -2.2 (the value actually observed in most genetic networks). Increasing the network size has no effect on the exponent of the power distribution; it only increases its validity range.
Figure 3. Measures of the Small World character of the genetic network. The main graph represents the clustering coefficient (C) as a function of the degree k. The clustering coefficient of the node i, with a degree ki is defined as as Ci = 2ni/ki(ki - 1), where ni denotes the number of direct links connecting its ki nearest neighbors among themselves. Ci is equal to 1 if the neighbors of i are all connected one to the other. A random (Erdos-Renyi) graph would produce a flat very small clustering coefficient. One sees from the graph that the actual distribution, far from being a small constant is fitted by a straight line that represents on the double logarithmic scale a power law with exponent -1 as observed in genetic networks (and in contrast with the preferential attachment dynamics). The fit is excellent with no free parameters. The inset plots the distribution of distances between arbitrary pairs of nodes in the network. A regular one dimensional lattice would produce a distance proportional to the network size (50,000). The combination of a distance of the order of the log of the network size (inset) and the existence of high clustering coefficient are the hallmark of a "small world network".
1 1.5 2 2.5 3 3.5 40
0.5
1
1.5
2
2.5
3
3.5
log10
(Number of nodes)
log 10
(Ran
k)
Nodes rank distribution
IncomingOutgoing
3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 40.6
0.65
0.7
0.75
0.8
0.85
log10(Network Size)
log
10(<
d>)
IncomingPower law fitOutgoingPower law fit
Figure 6: Double logarithmic histogram of the nodes rank distribution in the WWW model. Results show a power law for both incoming (empty circles; exponent
2.1+/-0.1) and outgoing links (full squares; exponent 3.1 +/- 0.2). These exponents are in perfect agreement with the experimental data.
The inset shows the average node distance as predicted by the WWW model. The distance grows as the log of the network size, and thus the network is indeed a
small world network in agreement with the experiments.
Executive Abstract• Logistic dynamics has been recognized since 200 years to govern a
wide range of social, economic, biological and cognitive systems.• In the past the predictions of the logistic equation have been
invalidated almost systematically in many occasions.• In particular they predicted often falsely the decay of the dynamics in
adverse conditions.• We show that the correct accounting for the discrete character of the
elementary components of the system leads to dramatically different predictions:– In particular the emergence of adaptive collective objects that insure survival
and development in conditions in which the naïve continuous/ global treatment would predict complete and uniform decay.
– The emergence of stable Pareto-Zipf power laws even in very non-stationary conditions.
• We review a series of applications, predictions and their validation.
- map the interdisciplinary cooperation network(- people are nodes - cooperations andcommon papers, are links).
- give priority to people with high interdisciplinarity
rather then high rank / disciplinary authority
Discipline 2
Discipline 1
Subjects that need synthesis
Objective Algorithm to Evaluate Interdisciplinary researchers relevance
Discipline3
This was a Particular case of Logistics dynamics (with Corrections!!); Other:
technological change; innovations diffusion (Rogers)
new product diffusion / market penetration (Bass)
social change diffusion
X = number of people that have already adopted the change and
N -X = number of remaining customers
dX/dt ~ X(N – X )
Potential Adopters
Adopters
0% salesPercolation transition
infin
itely
sha
rp
at in
finite
siz
e
10
1
100Logarithmic scale
Naïve logistic
• Insert LSS book
POSTEXT
Shalit A. Erez T. Deters A. Hershberg U. Shir E. Solomon S.
My papers
I am here
Clusters automatically formed by elastic
connectionsand repelling forces
New (Dynamic, Distributed, Open, Free, Self-Org, Ontology
My papers
I am here
Clusters automatically formed by elastic
connectionsand repelling forces
New (Dynamic, Distributed, Open, Free, Self-Org, Ontology
EMPTY SHELLS
-emergence of High-Tech communities-start-ups connections to previous businesses-entrepreneurs emerging from old businesses-partners having previous common institutions
-emergence of High-Tech communities-start-ups connections to previous businesses-entrepreneurs emerging from old businesses-partners having previous common institutions
Realistic macroscopic simulations require a new causal framework:
discrete / delayed/ conditional / nested causality instead of the usual Markov infinitesimal one
TimeTime
New mathematical concept: Markov Webs