Role of Scaling in Developing an Understanding of How Systems Work OR the danger of Big Data
Download as PPT, PDF1 like529 views
The document discusses the role of scaling in understanding complex systems, emphasizing the importance of power laws and their implications across various scales. It critiques common misconceptions about complexity and power law distributions, pointing out the need for robust theories and data-driven solutions in adaptive business strategies. Additionally, it highlights the challenges brought by increasing complexity and rapid changes in a globalized and technologically advanced world.
![Urban Scaling: City Sizes - 1950 United Nations data for world cities
Compare the scale K and α coefficients of the power-law y(x) ≈ K x-α (and
Pareto β= α+1) with the q-exponential parameters for q slope and scale κ in
y(x) ~ [1 + (1–q) x/κ)]1/(1–q), fitted to entire size curves
Power laws and Zipf’s law might fit upper bin frequencies for city
sizes but not the whole curve
cumulative number
of cities at this log bin or higher
10000
1000
100
10
1
100000
1000000
city size
10000000
Dashed line = portion of distribution that is "power-law“
(but is exaggerated in the upper bins)
(White, Kejžar, Tsallis, and Rozenblat © 2005 working paper)](https://image.slidesharecdn.com/sfiscalingfinal-140131145452-phpapp01/85/Role-of-Scaling-in-Developing-an-Understanding-of-How-Systems-Work-OR-the-danger-of-Big-Data-42-320.jpg)
Role of Scaling in Developing an Understanding of How Systems Work OR the danger of Big Data
- 1. “Role of Scaling in Developing an Understanding of How Systems Work” Dr. Norman L. Johnson norman@santafe.edu http://CollectiveScience.com
- 2. How “SFI-ish” is your System? Local Distributed Diverse Global Centralized Uniform Synergistic Selective Noise Structure Robust Dynamic Fragile Static Scale-free Clustered Emergent Top-down Qualities to consider in the above measures • Dimensionality (spatially, functionally) • Connectivity – Access to information • Governing equations / Rules • Fitness function, quality/performance measures • Scalability
- 3. Power Laws and Complexity: Not without some controversy “Over the last decade or so, it has become reasonably common to see people (especially physicists) claiming that some system or other is complex, because it exhibits a power law distribution of event sizes. Despite its popularity, this is simply a fallacy. No one has demonstrated any relation between power laws and any kind of formal complexity measure. Nor is there any link tying power laws to our intuitive idea of complex systems as ones with strongly interdependent parts.” In “METHODS AND TECHNIQUES OFCOMPLEX SYSTEMS SCIENCE: AN OVERVIEW” by Cosma Shalizi
- 4. My Background Polymer physics Star Wars Novel fusion device Combustion modeling Hydrogen Fuel Program P&G multi-phase flows
- 6. My Background Polymer physics Star Wars Novel fusion device Combustion modeling Hydrogen Fuel Program P&G multi-phase flows Biological threat reduction & homeland security Bird Flu - Mitigation development
- 7. My Real Background Polymer physics Star Wars Novel fusion device Combustion modeling Hydrogen Fuel Program Future of the internet Self-organizing collectives Diversity and emergent problem solving • Finance applications Developmental view of evolution • Finance applications Failure of Experts P&G multi-phase flows Effects of rapid change on collectives • Finance applications Biological threat reduction & homeland security Bird Flu - Mitigation Identity formation and interaction • Psycho-social simulation models • Coexistence applications
- 8. Challenges Facing Business Increased ‘complexity’ (too vague) Change happening faster and faster Data-Poor to Data-Rich environment “Fall of the House of Experts” Globalization - connected markets Technology surprise / reset ?? Survival Guide 2006
- 9. Opportunities Facing Business Increased ‘complexity’ Success for adaptive businesses Faster Change Data-Rich environment “Creative destruction” Data-driven solutions “Fall of the House of Experts” Collective solutions Globalization - connected markets Expanding markets Technology surprise Large payoffs Above opportunities are not just additive All required understanding changing data, context, & knowledge Survival Guide 2006
- 10. Weather Prediction - Use of data and theory Theory of averages and outliers Theory provides “control” equations for the model Data generation Model building Theory provides the context of data - Is a datum an outlier? And if so, how much should it be “enforced” in the assimilation? Validation Rich spatial and temporal data provide both initial conditions of the model, validation and the “nudging” of the variables to agree with “theory” Prediction Bottom line: Bottom • • • • • • • • Data-rich environment required Data-rich environment required Robust theory required Robust theory required Ideal for chaotic or high-uncertainty systems Ideal for chaotic or high-uncertainty systems Rich interaction between data, analysis and theory Rich interaction between data, analysis and theory Data Assimilation quality compatibility density (time&space) prior history context
- 11. Origin of “The Theory” Data generation Theory of averages and outliers Discovery Increasing levels of discovery: • • • • • • • • • • Statistical characterization Statistical Dimensionless functionality (correlations) Dimensionless functionality (correlations) Scaling - self-similarity Scaling - self-similarity Descriptive “Laws” Descriptive Functional relationships Functional relationships • Static • Static • Dynamic (governing equations of change) • Dynamic (governing equations of change) • Higher moments (variation within) • Higher • Error generation • Error generation
- 12. Drucker’s Age of Discontinuity (1968) Method: Observe trends and predict changes, not details New Global information economy - Knowledge is new capital – New technologies create new technologies – High mobility of men, capital, and ideas New pluralism (diversity) – Will make obsolete old theories of economy, government and society New structure based on social responsibility and accountability Drucker’s prediction of globalization and faster change was 28 years ago, in the time of greatest IT change . The context in 1968 “In fact, IBM currently is selling 100s of computers a Survival Guide 2006
- 13. Power Laws and Complexity (con’t) “… it has been known for half a century that there are many, many ways of generating power laws, just as there are many mechanisms which can produce Poisson distributions, or Gaussians. “Perhaps the simplest one is that recently demonstrated by Reed and Hughes, namely exponential growth coupled with random observation times. “The observation of power laws alone thus says nothing about complexity (except in thermodynamic equilibrium), and certainly is not a reliable sign of some specific favored mechanism, such as self-organized criticality or highly-optimized tolerance.” In “METHODS AND TECHNIQUES OFCOMPLEX SYSTEMS SCIENCE: AN OVERVIEW” by Cosma Shalizi, Center for the Study of Complex Systems, University of Michigan, Ann Arbor
- 14. Importance of Power Laws - Farmer “Scale free behavior has important scientific implications because it strongly suggests that the same mechanism is at work across a range of different scales.”
- 15. Income distribution in US and Japan
- 16. US Firm Size by Employees
- 17. Why care about distributions? One approach to prediction is a dynamical theory, e.g., non-equilibrium statistical mechanics Analytical forms of dynamical theories require “nice” distributions for probability distribution functions (PDFs) PDFs map variables into observables Nice distributions Gaussian distribution p(θ) “Delta” or point distribution 0 1 θ Los Alamos
- 18. Probability Distributions (ref. Farmer) Normal (Gaussian) distribution Defined on ( −∞, ∞) Natural explanation: Central limit theorem: Sum of many random variables (second moment finite). Many applications: Maxwell: velocity distribution of particles in a gas Heights of individuals, IQ, ... Distribution is thin tailed: No one is 10 feet tall
- 19. Probability Distributions (ref. Farmer) Normal (Gaussian) distribution Defined on ( −∞, ∞) Natural explanation: Central limit theorem: Sum of many random variables (second moment finite). Many applications: Maxwell: velocity distribution of particles in a gas Heights of individuals, IQ, ... Distribution is thin tailed: No one is 10 feet tall
- 20. Probability Distributions (ref. Farmer) Exponential distribution Defined on [0, ∞) Natural explanation (1): Survival times for constant probability of decay Natural explanation (2): Equilibrium statistical mechanics (Maximum entropy subject to constraint on mean) Many applications: Radioactive decay, Energy distribution at equilibrium, ... Distribution is also thin tailed: characteristic scale x0.
- 21. Probability Distributions (ref. Farmer) Power law Defined on [a, ∞). α, a > 0. Natural explanation? Distribution is heavy tailed, no characteristic scale. Many applications – is there a common link? A power law is a linear relation between logarithms
- 22. Difference between Thin and Thick Tailed distributions Similar for frequent events 1/1000 event is twice as large for a power law 1/10,000 event is three and a half times as large “The probability of observing a fluctuation of 21% (the size of the famous negative S&P return on October 19, 1987) under the normal hypothesis is less than 10 −16, whereas the probability under the power law distribution is 0.08%. Under the normal distribution it is essentially impossible that this event could ever have occurred, whereas under a power law distribution such an event is to be expected.” - Farmer
- 23. Many Examples of Power-Law Distributions Frequency vs. measure earthquakes flood levels rainfall craters insurance claims income price changes firm size transaction volume price for order placement city size intensity of wars length of strikes DoD cost overruns frequency of word usage name frequency authored papers citations to papers patent profitability music sales movie sales book sales telephone calls number of email messages size of computer files hits on web page links to web sites
- 24. Many Examples of Power-Law Distributions Frequency vs. measure
- 25. What’s Special about Power Laws Power laws are scale invariant • Retains functional form under scaling Power law is an asymptotic relation • Only meaning full in a bounded region • Practically, real problems always have cut-offs
- 26. Unshaded region - Asymptotic Frequency vs. measure
- 27. What’s Special about Power Laws Power laws are scale invariant • Retains functional form under scaling Power law is an asymptotic relation • Only meaning full in a bounded region • Practically, real problems always have cut-offs Power laws have a threshold above which moments don’t exist • For this reason there is no such thing as an “average flood” - it is only possible to measure flood likelihoods in terms of quantiles, as in the statement “this is a 100 year flood”. (Farmer) Combinations (aggregations) of power laws remain a power law Any ”reasonable” function with moments that don’t exist, i.e. “truly heavy tails”, is a power law
- 28. Power Laws vs. Long-Memory Difficult to determine whether power laws or long-memory are present by empirical sampling Tails by definition only have a small number of events. May not have enough data to probe tails. Can have slow convergence to tail (slowly varying functions) Can have cutoffs of tail due to finite size effects (e.g. physical limits) Thin tailed distributions may mimic power law behavior over a wide range Statistical convergence for long-memory processes is very slow
- 29. Importance of Power Laws - Farmer “Scale free behavior has important scientific implications because it strongly suggests that the same mechanism is at work across a range of different scales.” “The real test is whether power laws can improve our predictive or explanatory power by leading to better models. Self-similarity is such a strong constraint that, even if only an approximation over a finite range, it is an important clue about mechanism.” “Ultimately, the best method to demonstrate that power laws are applicable is to construct theories that also have more detailed testable predictions.”
- 30. Diversity and Collective Prediction Prediction of collective behavior is generally easier at extremes of diversity or variation Locally and Globally Predictable Low Diversity Unpredictable Globally Predictable High Diversity Los Alamos
- 31. Diversity and Collective Prediction How does this translate to distribution functions? Locally and Globally Predictable Unpredictable Problem distributions: • Discrete distributions • Multi-modal distributions • Long-tailed distributions p(ø) 0 1 (e.g., power law, Globally Predictable p(θ) 0 1 θ instead of Gaussian statistics) Low Diversity High Diversity Los Alamos
- 32. Mechanisms for Generating Power Laws Critical points and deterministic dynamics Random processes Sampling from (terminated) exponential growth (Reed and Hughes) Mixtures Dimensional constraints Maximization principles Preferential attachment Non-equilibrium statistical mechanics
- 33. Causes of Anomalous Distributions Incomplete sampling Habitual behavior Non-equilibrium dynamics Co-evolving systems - markets “Structures” limit options Regulations, cultures, ... Boundary effects dominate Behavior at borders Coupling or interdependence of different levels Hierarchical organizations Clustering or localized regions of interaction Typical of social, information and power networks Los Alamos
- 34. Statistical Issues Relating to Power Laws “some common mistakes” Parameter Estimation Use linear regression to find the line of best fit to the points on the log-log plot. But the line minimizing the sum of squared errors is not a valid probability distribution, and so this is simply not a reliable way to estimate the distribution. Error Estimation Estimate of the standard error in the estimated slope and report this as the uncertainty in the power law. “This is an entirely unacceptable procedure.” On a log-log plot this violates the assumption that measured values have Gaussian fluctuations around their true means. Validation “The basic problem here is that any smooth curve looks like a straight line, if you confine your attention to a sufficiently small region and for some nonpower-law distributions, such sufficiently small ﾓ regions can extend over multiple orders of magnitude. In “METHODS AND TECHNIQUES OFCOMPLEX SYSTEMS SCIENCE: AN OVERVIEW” by Cosma Shalizi
- 35. THOUGHT LEADER FORUM Rat Studies of Maximum Carrying Capacity Cooperative social structure Control - no imposed social structure NIMH psychologist John B. Calhoun, 1971 Both systems loaded to 2 1/2 times the optimal capacity. Social order system can carry 8 times the optimal capacity. CSFB EMPOWERING CHANGE
- 36. THOUGHT LEADER FORUM Food Production Rate 1.20 Collective 1.00 Individual Condensed 0.80 0.60 Co-Operational 0.40 0.20 0.00 Production rate (food units/time units) 0 1 2 Formative 3 4 5 6 7 8 Rate of environmental change (tenth of degree/time unit) 0 2 2.5 4 CSFB EMPOWERING CHANGE
- 37. THOUGHT LEADER FORUM Collective Response to Rates of Change Stages in Development Unimpeded development Innovators are essential Collective actions lead to inefficiencies Potential system-wide failure Stable “no change” Change slower than collective response Change faster than collective response Change faster than individual response Condensed (optimization of collective) Co-Operational (synergism from individuals) Formative (creation of individual features) Featureless CSFB Rate of ChangeEMPOWERING CHANGE
- 38. On Constructing Theories (Models) and Testing 1. Replication is essential. 2. It is a good idea to share not just data but programs. 3. Always test the robustness of your model to changes in its parameters. (This is fairly common.) 4. Always test your model for robustness to small changes in qualitative assumptions. If your model calls for a given effect, there are usually several mechanisms which could accomplish it. If it does not matter which mechanism you actually use, the result is that much more robust. Conversely, if it does matter, the over-all adequacy of the model can be tested by checking whether that mechanism is actually present in the system. Altogether too few people perform such tests. In “METHODS AND TECHNIQUES OFCOMPLEX SYSTEMS SCIENCE: AN OVERVIEW” by Cosma Shalizi
- 39. European Conference on Complex Systems Paris, 14-18 November 2005 Civilizations as dynamic networks Cities, hinterlands, populations, industries, trade and conflict Douglas R. White © 2005 All rights reserved 50 slides - also viewable on drw conference paper website version 1.3 of 11/12/2005
- 40. Co-evolution time-series of Cities and City Networks City attributes and distributions begin Dynamics from Urban Hierarchy-Industries, _______Commerce, Finance Structural Cohesion Routes, Capacities Unit Formation (e.g. polities) Velocities and Magnitudes of trade City Sizes Hierarchy Demography/Resources Hinterland Productivity Conflicts periodize STATES from factions & coalitions to sovereignty - emergent Spatiopolitical units City Networks Interference and attempts at regulation Sources of boundary conflicts Organizational transformation of nodes MARKETS from structurally cohesive k-components - emergent Network units (overlap)
- 41. Superlinear ~ 1.67 Linear ~ 1 Sublinear ~ .85 ISCOM working paper
- 42. Urban Scaling: City Sizes - 1950 United Nations data for world cities Compare the scale K and α coefficients of the power-law y(x) ≈ K x-α (and Pareto β= α+1) with the q-exponential parameters for q slope and scale κ in y(x) ~ [1 + (1–q) x/κ)]1/(1–q), fitted to entire size curves Power laws and Zipf’s law might fit upper bin frequencies for city sizes but not the whole curve cumulative number of cities at this log bin or higher 10000 1000 100 10 1 100000 1000000 city size 10000000 Dashed line = portion of distribution that is "power-law“ (but is exaggerated in the upper bins) (White, Kejžar, Tsallis, and Rozenblat © 2005 working paper)
- 43. Origin of “The Theory” Data generation Theory of averages and outliers Discovery Increasing levels of discovery: Increasing levels of discovery: • Statistical characterization • Statistical • Dimensionless functionality (correlations) • Dimensionless functionality (correlations) • Scaling - self-similarity • Scaling - self-similarity • Descriptive “Laws” • Descriptive • Functional relationships • Functional relationships • Static • Static • Dynamic (governing equations of change) • Dynamic (governing equations of change) • Higher moments (variation within) • Higher • Error generation • Error generation
- 44. Summary Why true power law observations are exciting Essential to discovery of governing “laws” Often from a critical process in a complex system But power laws are only part of the process to developing a broader understanding. Not meaningful without a model or theory. Power laws behavior doesn’t mean complexity Power laws are (Idealized) asymptotic behavior - beware of the boundaries! Represent thick tails, rather than thin tailed distributions In general, anomalous distributions may dominate real systems Consider maturity of system evolution - “young” systems may be very different Or mixed maturity of sub-systems Consider optimization versus robustness Best resources on the subject (appear to be) from SFI
- 45. Challenges to Business • Increased ‘complexity’ • Change happening faster and faster • Data-Poor to Data-Rich environment • “Fall of the House of Experts” • Globalization - connected markets • Technology surprise Data generation Model building Theory of averages and outliers Validation Prediction Data Assimilation
- 46. References Shalizi, Cosma R., “METHODS AND TECHNIQUES OF COMPLEX SYSTEMS SCIENCE: AN OVERVIEW”, Chapter 1 (pp. 33-114) in Thomas S. Deisboeck and J. Yasha Kresh (eds.), Complex Systems Science in Biomedicine (New York:Springer, 2006) http://arxiv.org/abs/nlin/0307015 Farmer, J. Doyne & John Geanakoplos, “Power laws in economics and elsewhere”, DRAFT April 4, 2005 (chapter from a preliminary draft of a book called “Beyond equilibrium and efficiency”) Contact the authors for a copy. Farmer, J. Doyne, “Power laws”, Santa Fe Institute Summer School June 29, 2005. Contact the author for a copy. Holbrook, Morris B.. 2003. "Adventures in Complexity: An Essay on Dynamic Open Complex Adaptive Systems, Butterfly Effects, Self-Organizing Order, Co-evolution, the Ecological Perspective, Fitness Landscapes, Market Spaces, Emergent Beauty at the Edge of Chaos, and All That Jazz http://www.amsreview.org/articles/holbrook06-2003.pdf White, Douglas R., “Civilizations as dynamic networks: Cities, hinterlands, populations, industries, trade and conflict”, European Conference on Complex Systems Paris, 14-18 November 2005. http://eclectic.ss.uci.edu/~drwhite/ ppt/CivilizationsasDynamicNetworksParis.ppt For exceptional talks on Complexity in financial systems, see the Thought Leaders Forums: • • http://www.leggmason.com/thoughtleaderforum/2006/index.asp for 2003-2006 http://www.capatcolumbia.com/CSFB%20Thought%20Leader%20Forum.htm for 2000-2003
- 48. Topics for Discussion? What are the implications in your application area of the shift to being “data rich”? Are the moments around the norm important? Are you driven by the outliers or norms? How much can the results presented today be generalized? Stephanie’s table of correspondence? Worth more development / study? Are global properties always determined by the “supply chain” - as Geoff observed in organisms? How do physical/operational restrictions dominate the growth of a system? In complex composite systems, does the host always dominate the dynamics? Can human-designed systems be better planned by using general “laws”? Bio-inspired solutions - where/when valid? Are ant colonies more like bio organisms or human societies? What about material versus information differences? (If I give you an apple, I don’t have an apple; is information the same, why not?) How does basic scales affect the system: e.g., cell size is fixed, but computer size is not What can be concluded about network scaling and global system performance? Social systems and innovative reset (Singularity horizon?) Assumes infrastructure is largely unchanged Can innovation reset infrastructure? What about optimization of Efficiency versus Robustness? In times of faster change, robustness is maybe more important Stages of development and maturity How much what has been observed today is only for mature, well developed systems? How do early innovative systems behave? How do opportunities and approaches differ? Others?
Editor's Notes
- #2: Goal is to talk about Scaling under the rubric of survival in a fast changing, increasingly complex world. Introduction: This is not an area that I’m deeply familiar with, so I present the following much in the vain that you might do after you’ve spent some time reading about what’s been said and how it relates to your area of interest.
- #3: Different ways to think about your system of interest. There is no real judgment as to which side of the distribution a “SFI-ish” system might be located, but certainly the list captures many research and application areas that SFI is currently or in the past focused on. Note how these distributions are not independent. For example a robust system often has diverse and distributed components that are locally controlled, and have some degree of chaos or noise and often express emergent properties and are dynamic rather than static systems. Similarly a fragile system often has uniform and centralized components that are globally controlled and have a high degree of structure and top-down functions, and are often appear static until they break under stress or change.
- #4: No one at the SFI meeting was guilty of this. Nor any SFI researchers that I know. But it does illustrate that this isn’t a subject that is just academic and obvious. There is a lot of mystique about this topic that has been exploited in the popular literature, leading to this observation. Lesson: beware of selling complexity and justifying it by making associations with technical observations. (See the viewgraph near the end for the full reference.)
- #5: The next couple of viewgraphs represent some of my background - The main point is that I’ve been working closely with industry for the last 15 years at the interface between science and industry, particularly in the area of development of tools to reduce costs - mostly in the area of fluid dynamics modeling. The P&G project is a great example of how to close the gap between application and deep science: P&G came to the multiphase fluid dynamics group twice to ask for collaboration. They were turned away. On the third visit they made their pitch that they have rich data in need of theory (a major lesson on how data is becoming more available - more on this in a second). Secondly they suggested starting small ($10ks). Ultimately the project became a million dollars a year at LANL and ended up saving P&G about 20 times more than they spend - with a continued return long after the end of the project: They now market the resources under P&G’s name.
- #6: There was a time that I was very happy that the project with P&G was highly proprietary and I couldn’t talk about what I was working on. Until this came out.
- #7: Recent problems I’ve worked on have shown me that much of the complexity of problems facing society cannot be solved by technological solutions, but require an exceptional integration of science and policy awareness. One of the major challenges is not to solve new problems in the same way we tried to solve old problems. That is why we are here today.
- #8: These topics speak very much to the reason that the interface between science and industry in this new age is so important. Much is changing and requiring tools that weren’t available just a few years ago.
- #9: In my interactions with business this is the list that I hear are the biggest challenges - and is the reason that I think the first item is often the summary for all the items that follow. An why SFI is the interface to the future of doing business. I think this is the core problem - we see it in every facet of our personal and work lives. But Faster change refers not just to increased rates, but also the changes in the ordering of scales of change: for example, major changes occur in a worker’s life when the average lifespan of a company because much shorter than the average work lifespan of a worker - as has strongly happened over the last few decades (see the book Creative Destruction by Foster) The data rich environment has two consequences: 1) we may feel we have too much data, much irrelevant, to make decisions (information overload) and 2) we don’t have the tools to deal with the extra information. A relevant observation that connects item 1 and 3: computers are good processing large amounts of low complexity information, humans a good at processing small amounts of high complexity information. We have yet to develop the tool/resources/procedures for dealing with large amounts of highly complex information (See the papers on symbiotic intelligence at http://ishi.lanl.gov) “Fall of the house of experts” refers to the title of a talk I gave at an SFI public lecture - on how experts are failing us because of the complexity of the problems and how collective solutions a filling in the gap. Video available from SFI (a bit out of date) Globalization - well documented. An observation from the audience was that this is a complex issue that cuts many ways. Technology surprise: a major innovation, maybe not even technology but possibly a way of doing business (Amazon), can make rich companies poor quickly. Geof West observations about resetting the growth curve is quite relevant. I would add that some technologies can change the infrastructure, which may start a completely different growth curve. ?? Pick your own. - Much discussion on how it’s not just data rich, but knowing what data is important. I believe the issue also is the context of data - which gets lost in a fast changing world.
- #10: As with all challenges there are often opportunities. These are just a quick list, but this is an area for rich discussion among the business network. Note that these are not just additive - a small improvement in a couple of the above areas and have a multiplicative or exponential increase in opportunity.
- #11: One example of how prediction of complex-chaotic system (the weather) made great advances: By the combination of three advances: rich data sources world wide Theories or models that can give the data context Fast computers With these, we’ve been able to predict the worst of chaotic systems. I believe this is the paradigm for the future of business: as we develop better data and models, we’ll be able to make better predictions and decisions. Note that in order to deal with the chaotic nature of the system, data assimilation is required (using data real time to adjust the arrow to the future states)
- #12: This viewgraph illustrates the context and role of scaling or power laws in science (and business). Observations: Most businesses stop at correlations in dealing with large amounts of data. The challenge and big payoffs are from driving further down in the list. My view is that this is why we are all here today. The last two items are rarely touched even in well developed sciences, but are proving to be the real resources needed for decision makers in dealing with complex systems with potentially severe unintended consequences of decisions. Much of this can be captured under the rubric of UNCERTAINTY MANAGEMENT. Higher moments refer to the variation of the data around the mean Error generation refers to the tracking of uncertainty in systems or of the noise in a system. (search on infodynamics on the web for background)
- #13: Ducker was the master at prediction of the future, in the most challenging time of our history. I higlyh recommend reading the first chapter of this book on his approach to prediction. In many ways this is what scaling is all about: you focus on predicting trends, not details (technologies). Just to put Ducker’s achievement of predicting the information age in context: he includes the quote above as proof that computers are important. The personal computer didn’t even exist at this time.
- #14: Major point: observation of a power law (or more generally scaling) is of little use by itself. When supported by a model or theory, then it gets exciting. The cited paper by Reed and Hughes is useful because it addresses a simple process that causes powerlaw behavior that has not been appreciated. (See the viewgraph near the end for the full reference.)
- #15: The best technical summary for power law behavior is by Doyne Farmer (see references at end). A question was: how can you resolve the above statement with the second viewgraph on scaling and complexity? There are two different points: The statement that power laws or scale-free behavior implies complexity is not correct But generally power laws or scale-free behavior often, but not always (see the Reed reference in the previous viewgraph) imply a fundamental mechanism controls behavior over many scales.
- #16: An example
- #17: Another example that is particularly remarkable.
- #18: Just a quick reason why we care about distributions - from a science viewpoints. (We also care about them because they tell us about the behavior of the systems. ) The problem is that because we often try to develop analytical theories, we often force these distributions into unrealistic but “nice” ones. This has been a major source of controversy in many fields.
- #19: One common distribution.
- #20: Main point: thin tailed (more on this in a second)
- #21: Ditto on thin tailed.
- #22: Major note: Not thin tailed, but heavy tailed. These distributions can cause many problems in analytical treatment in science.
- #23: What’s the difference between thin and thick tails. An example.
- #24: Lots of examples.
- #25: Pictures. Note the range on the horizontal axis: For ones with only a few decades of range, these are questionable power laws (see later viewgraphs). Ones with many decades are pretty solid.
- #26: Previous speakers have addressed these: the first is important to applications. The second is a reminder that the real world has bookends that may not allow the power law region to be very broad.
- #27: The shaded portion illustrated the non-powerlaw regions.
- #28: These are more esoteric properties of paper laws - but very important in a general treatment and not generally appreciated. For a complete discussion see the paper (chapter) by Farmer in the references at at end.
- #29: Doyne Farmer has a section on long memory effects that can also lead to power law behavior. The point is that a system with no memory (or little) memory effects often are the systems that are identified to have power law behavior. But the long memory effect should also be a consideration when thick tail behavior is observed. This viewgraph also speaks to the difficulties around sampling systems with memory effects.
- #30: The over-all perspective: Yes, they are useful, but their full utility only comes out when supported with a theory or model. Note that this is not only to test the origin of the power aw behavior but also to help in the sampling strategy for data acquisition that is a challenge for these systems. Here the point is that incorrect sampling of distributions may lead to incorrect conclusions.
- #31: Back to distributions and prediction - with respect to diversity or heterogeneity of the system. Turns out that a little or a lot of diversity (that is well sampled) is good for prediction. The qualifier “well-sampled” diversity is required because some systems have lots of diversity that is poorly interconnected and therefore the diversity really doesn’t really get sampled, which has a major effect on the dynamics or robustness of the system - a prime example is a senescent ecosystem: lots of diversity but very restrained interactions. Same is true for old economies.
- #32: So what causes distributions to be not “nice”? One list is given above. You can read lots on this looking at the work by Tsallis (more on this in a bit).
- #33: From Farmer. This is an “academic” list of mechanisms for generation power laws. Books have been written on each of these topics.
- #34: This is a old viewgraph of mine (circa 2001) that describes the mechanism for generating anomalous distributions and mapping it over to real systems. This list has many similarities to the previous list for powerlaws and may be more general.
- #35: If you find yourself looking at data with apparent power law behavior, this advice is for you.
- #36: Because of a discussion during the break I added this viewgraph - to illustrate impact of habitual behavior on social systems. You can read more about this experiment in: http://www.capatcolumbia.com/CSFB%20TLF/2002/johnson_sidecolumn.pdf I’d generally recommend looking at the other talks at: http://www.leggmason.com/thoughtleaderforum/2006/index.asp for 2003-2006 http://www.capatcolumbia.com/CSFB%20Thought%20Leader%20Forum.htm for 2000-2003 Contact me for the reference.
- #37: The system is food foraging by ants with a moving food supply. It illustrate different stages of development. This illustrates the effect of rate of change on a self organizing system. Note how as the food moves faster, the collective contribution declines and becomes more erratic (the “error” bars for each stack). The take away is that rates of change do drive systems into different stages, each with their characteristics,properties, and different scaling behavior. Contact me for the movies and a paper on this topic. Also net-logo model for playing with it yourself.
- #38: Summary of the effect of change on the collective system; increased change forces the system to earlier stages in its development. The major lesson here is that most of the systems we have discussed today have had time to develop - a long time for biosystems. But many financial systems are relative young and potentially unstable. Are they as likely to show scaling or power law behavior? This is a major unanswered question.
- #39: Ok, so you have a model of the power law behavior, what should you do to investigate the robustness of the behavior and is accuracy as an explanation?
- #40: A complex model of human city and city network formation and dynamics. From an SFI regular. This is a excellent set of viewgraphs on the development of networks and dynamics on those networks
- #42: Doug’s examples in his area of scale free behavior. Note the distribution around the mean.
- #43: This plot is the main reason for showing excerpts from Doug’s set of viewgraphs. Doug makes the point that many of the curves he observes that might initially seem to be powerlaw are better fitted by an alternative “Q-exponential” fit that was developed by Tsallis. This in not a scale free distribution in general (although can reduce to one) and is argued to occur for many of the reasons that I listed in the previous viewgraph on the origins of anomalous distributions.
- #44: A summary: So here is where power law distribution fit into the bigger picture of discover of trends in the data. It is very important to not stop there, but to continue down the list if possible. And to develop theories if possible.
- #45: A summary of what we observed about power law behavior and what should be considered.
- #46: My speculation is that the future is all about data-rich, complex systems and developing tools that help us predict their behavior. The above process is the path that will be taken, and much of what has been discussed at this workshop address the discovery of the processes that will enable us ultimately to predict these systems.