probabilistic inference

The – Deep Learning Comprehensive Statistical Framework – Descriptive to Probabilistic Inference



Given the challenge of analyzing against the large data sets both structured (EHR data) and unstructured data; the emerging Healthcare analytics are around below discussed methods d (multivariate regression), e (neural-net) and f (multivariate probabilistic inference); Ingine is unique in the Hyperbolic Dirac Net proposition for probabilistic inference.

The basic premise in engineering The™ is in acknowledging the fact that in solving knowledge extraction from the large data sets (both structured and unstructured), one is confronted by very large data sets riddled with high-dimensionality and uncertainty.

Generally in solving insights from the large data sets the order in complexity is scaled as follows.

a)   Insights around :- “what”

For large data sets, descriptive statistics are adequate to extract a “what” perspective. Descriptive statistics generally delivers statistical summary of the ecosystem and the probabilistic distribution.

Descriptive statistics : Raw data often takes the form of a massive list, array, or database of labels and numbers. To make sense of the data, we can calculate summary statistics like the mean, median, and interquartile range. We can also visualize the data using graphical devices like histograms, scatterplots, and the empirical cdf. These methods are useful for both communicating and exploring the data to gain insight into its structure, such as whether it might follow a familiar probability distribution. 

b)   Univariate Problem :- “what”

Considering some simplicity in the variables relationships or is cumulative effects between the independent variables (causing) and the dependent variables (outcomes):-

i) Univariate regression (simple independent variables to dependent variables analysis)

c)    Bivariate Problem :- “what”

Correlation Cluster – shows impact of set of variables or segment analysis.

From above link :- In machine learningcorrelation clustering or cluster editing operates in a scenario where the relationships between the objects are known instead of the actual representations of the objects. For example, given a weighted graph G = (V,E), where the edge weight indicates whether two nodes are similar (positive edge weight) or different (negative edge weight), the task is to find a clustering that either maximizes agreements (sum of positive edge weights within a cluster plus the absolute value of the sum of negative edge weights between clusters) or minimizes disagreements (absolute value of the sum of negative edge weights within a cluster plus the sum of positive edge weights across clusters). Unlike other clustering algorithms this does not require choosing the number of clusters k in advance because the objective, to minimize the sum of weights of the cut edges, is independent of the number of clusters.

From above link. :- Correlation is a bivariate analysis that measures the strengths of association between two variables. In statistics, the value of the correlation coefficient varies between +1 and -1. When the value of the correlation coefficient lies around ± 1, then it is said to be a perfect degree of association between the two variables. As the correlation coefficient value goes towards 0, the relationship between the two variables will be weaker. Usually, in statistics, we measure three types of correlations: Pearson correlation, Kendall rank correlation and Spearman correlation

d)   Multivariate Analysis (Complexity increases) :- “what”

§ Multiple regression (considering multiple univariate to analyze the effect of the independent variables on the outcomes)

§ Multivariate regression – where multiple causes and multiple outcomes exists

 e)   Neural Net :- “what”

The above discussed challenges of analyzing multivariate pushes us into techniques such as Neural Net; which is the next level to Multivariate Regression Statistical Approach…. where multiple regression models are feeding into the next level of clusters, again an array of multiple regression models.

The above Neural Net method still remains inadequate in depicting “how” probably the human mind is operates. In discerning the health ecosystem for diagnostic purposes, for which “how”, “why” and “when” interrogatives becomes imperative to arrive at accurate diagnosis and target outcomes effectively. Its learning is “smudged out”. A little more precisely put: it is hard to interrogate a Neural Net because it is far from easy to see what are the weights mixed up in different pooled contributions, or where they come from.

“So we enter Probabilistic Computations which is as such Combinatorial Explosion Problem”.

f)    Hyperbolic Dirac Net (Inverse or Dual Bayesian technique): – “how”, “why”, “when” in addition to “what”.

All the above are still discussing the “what” aspect. When the complexity increases the notion of independent and dependent variables become non-deterministic, since it is difficult to establish given the interactions, potentially including cyclic paths of influence in a network of interactions, amongst the variables. A very simple example in just a simple case is that obesity causes diabetes, but the also converse is true, and we may also suspect that obesity causes type 2 diabetes cause obesity. In such situation what is best as “subject” and what is best as “object” becomes difficult to establish. Existing inference network methods typically assume that the world can be represented by a Directional Acyclic Graph, more like a tree, but the real world is more complex than that that: metabolism, neural pathways, road maps, subway maps, concept maps, are not unidirectional, and they are more interactive, with cyclic routes. Furthermore, discovering the “how” aspect becomes important in the diagnosis of the episodes and to establish correct pathways, while also extracting the severe cases (chronic cases which is a multivariate problem). Indeterminism also creates an ontology that can be probabilistic, not crisp.

Note: From Healthcare Analytics perspective most Accountable Care Organization (ACO) analytics addresses the above based on the PQRS clinical factors, which are all quantitative. Barely useful for advancing the ACO into solving performance driven or value driven outcomes most of which are qualitative.

To conduct HDN Inference, bear in mind that getting all the combinations of factors by data mining is “ combinatorial explosion ” problem, which lies behind the difficulty of Big Data as high dimensional data.

It applies in any kind of data mining, though it is most clearly apparent when mining structured data, a kind of spreadsheet with many columns, each of which are our different dimensions. In considering combinations of demographic and clinical factors, say A, B, C, D, E.., we ideally have to count the number of combinations (A), (A,B) (A, C) …(B, C, E)…and so on. Though sometimes assumptions can be made, you cannot always deduce a combination with many factors from those with fewer, nor vice versa. In the case of the number N of factors A,B,C,D,E,… etc. the answer is that there are 2N-1 possible combinations. So data with 100 columns as factors would imply about 


combinations, each of which we want to observe several times and so count them, to obtain probabilities. To find what we need without knowing what exactly it is in advance, distinguishes unsupervised data mining from statistics in which traditionally we test a hunch, a hypothesis. But worse still, in our spreadsheet the A, B, C, D, E are really to be seen as column headings with say about n possible different values in the columns below them, and so roughly we are speaking of potentially needing to count not just, say, males and females but each of nN different kinds of patient or thing. This results in truly astronomic number of different things, each to observe many time. If merely n=10, then nN is


There is a further implied difficulty, which in a strange way lifts much the above challenge from the shoulders of researchers and of their computers. In most cases of the above, must of the things we are counting contain many of the factors A,B,C,D, E..etc. Such concurrences of so many things is typically rare, so many of the things we would like to count will never be seen at all, and most of the rest will just be seen 1, 2, or 3 times. Indeed, any reasonably rich patient record with lots of data will probably be unique on this planet. However, most approaches are unable to make proper use of that sparse data, since it seems that it would need to be weighted and taken into account in the balance of evidence according to the information it contains, and it is not evident how. The zeta approach tells us how to do that. In short, the real curse of high dimensionality is in practice not that our computers lack sufficient memory to hold all the different probabilities, but that this is also true for the universe: even in principle we do not have all the data to work to determine probabilities by counting with even if we could count and use them. Note that probabilities of things that are never observed are, in the usual interpretation of zeta theory and of Q-UEL, assumed to have probability 1. In a purely multiplicative inference net, multiplying by probability 1 will have no effect. Information I = –log(P) for P = 1 means that information I = 0. Most statements of knowledge are, as philosopher Karl Popper argued, assertions awaiting refutation.

Nonetheless the general approach in the fields of semantics, knowledge representation, and reasoning from it is to gather all the knowledge that can be got into a kind of vast and ever growing encyclopedia. 

In The™ the native data sets have been transformed into Semantic Lake or Knowledge Representation Store (KRS) based on Q-UEL Notational Language such that they are now amenable to HDN based Inferences. Where possible, probabilities are assigned, if not, the default probabilities are again 1. 

Q-UEL Toolkit for Medical Decision Making :- Science of Uncertainty and Probabilities

Screen Shot 2016-08-24 at 11.07.49 AM

Quantum Universal Exchange Language

Emergent | Interoperability | Knowledge Mining | Blockchain


  1. It is a toolkit / framework
  2. Is an Algorithmic Language for constructing Complex System
  3. Results into a Inferential Statistical mechanism suitable for a highly complex system – “Hyperbolic Dirac Net”
  4. Involves an approach that is based on the premise that a Highly Complex System driven by the human social structures continuously strives to achieve a higher order in the entropic journey by continuos discerning the knowledge hidden in the system that is in continuum.
  5. A System in Continuum seeking Higher and Higher Order is a Generative System.
  6. A Generative System; Brings System itself as a Method to achieve Transformation. Similar is the case for National Learning Health System.
  7. A Generative System; as such is based on Distributed Autonomous Agents / Organization; achieving Syndication driven by Self Regulation or Swarming behavior.
  8. Essentially Q-UEL as a toolkit / framework algorithmically addresses interoperability, knowledge mining and blockchain; while driving the Healthcare Eco-system into Generative Transformation achieving higher nd higher orders in the National Learning Health System.
  9. It has capabilities to facilitate medical workflow, continuity of care, medical knowledge extraction and representation from vast large sets of structured and unstructured data, automating bio-statistical reasoning leading into large data driven evidence based medicine, that further leads into clinical decision support system including knowledge management and Artificial Intelligence; and public health and epidemiological analysis.


A Large Chaotic System driven by Human Social Structures has two contending ways.

a. Natural Selection – Adaptive – Darwinian – Natural Selection – Survival Of Fittest – Dominance

b. Self Regulation – Generative – Innovation – Diversity – Cambrian Explosion – Unique Peculiarities – Co Existence – Emergent

Accountable Care Organization (ACO) driven by Affordability Care Act transforms the present Healthcare System that is adaptive (competitive) into generative (collaborative / co-ordinated) to achieve inclusive success and partake in the savings achieved. This is a generative systemic response contrasting the functional and competitive response of an adaptive system.

Natural selection seems to have resulted in functional transformation, where adaptive is the mode; does not account for diversity.

Self Regulation – seems like is a systemic outcome due to integrative influence (ecosystem), responding to the system constraints. Accounts for rich diversity.

The observer learns generatively from the system constraints for the type of reflexive response required (Refer – Generative Grammar – Immune System –

From the above observation, should the theory in self regulation seem more correct and that adheres to laws of nature, in which generative learning occurs. Then, the assertion is “method” is offered by the system itself. System’s ontology has an implicate knowledge of the processes required for transformation (David Bohm – Implicate Order)

For very large complex system,

System itself is the method – impetus is the “constraint”.

In the video below, the ability for the cells to creatively create the script is discussed which makes the case for self regulated and generative complex system in addition to complex adaptive system.


Further Notes on Q-UEL / HDN :-

  1. That brings Quantum Mechanics (QM) machinery to Medical Science.
  2. Is derived from Dirac Notation that helped in defining the framework for describing the QM. The resulting framework or language is Q-UEL and it delivers a mechanism for inferential statistics – “Hyperbolic Dirac Net”
  3. Created from System Dynamics and Systems Thinking Perspective.
  4. It is Systemic in approach; where System is itself the Method.
  5. Engages probabilistic ontology and semantics.
  6. Creates a mathematical framework to advance Inferential Statistics to study highly chaotic complex system.
  7. Is an algorithmic approach that creates Semantic Architecture of the problem or phenomena under study.
  8. The algorithmic approach is a blend of linguistics semantics, artificial intelligence and systems theory.
  9. The algorithm creates the Semantic Architecture defined by Probabilistic Ontology :- representing the Ecosystem Knowledge distribution based on Graph Theory

To make a decision in any domain, first of all the knowledge compendium of the domain or the system knowledge is imperative.

System Riddled with Complexity is generally a Multivariate System, as such creating much uncertainty

A highly complex system being non-deterministic, requires probabilistic approaches to discern, study and model the system.

General Characteristics of Complex System Methods

  • Descriptive statistics are employed to study “WHAT” aspects of the System
  • Inferential Statistics are applied to study “HOW”, “WHEN”, “WHY” and “WHERE” probing both spatial and temporal aspects.
  • In a highly complex system; the causality becomes indeterminable; meaning the correlation or relationships between the independent and dependent variables are not obviously established. Also, they seem to interchange the position. This creates dilemma between :- subject vs object, causes vs outcomes.
  • Approaching a highly complex system, since the priori and posterior are not definable; inferential techniques where hypothesis are fixed before the beginning the study of the system become enviable technique.

Review of Inferential Techniques as the Complexity is Scaled

Step 1:- Simple System (turbulence level:-1)

Frequentist :- simplest classical or traditional statistics; employed treating data random with a steady state hypothesis – system is considered not uncertain (simple system). In Frequentist notions of statistics, probability is dealt as classical measures based only on the idea of counting and proportion. This technique is applied to probability to data, where the data sets are rather small.

Increase complexity: Larger data sets, multivariate, hypothesis model is not established, large variety of variables; each can combine (conditional and joint) in many different ways to produce the effect.

Step 2:- Complex System (turbulence level:-2)

Bayesian :- hypothesis is considered probabilistic, while data is held at steady state. In Bayesian notions of statistics, probability is of the hypothesis for a given sets of data that is fixed. That is, hypothesis is random and data is fixed. The knowledge extracted contains the more subjectivist notions of uncertainty, belief, reliability, or confidence often used in automated inference and decision support systems.

Additionally the hypothesis can be explored only in an acyclic fashion creating Directed Acyclic Graphs (DAG)

Increase the throttle on the complexity: Very large data sets, both structured and unstructured,  Hypothesis random, multiple Hypothesis possible, Anomalies can exist, There are hidden conditions, need arises to discover the “probabilistic ontology” as they represent the system and the behavior within.

Step 3: Highly Chaotic Complex System (turbulence level:-3)

Certainly DAG is now inadequate, since we need to check probabilities as correlations and also causations of the variables, and if they conform to a hypothesis producing pattern, meaning some ontology is discovered which describes the peculiar intrinsic behavior among a specific combinations of the variables to represent a hypothesis condition. And, there are many such possibilities within the system, hence  very chaotic and complex system.

Now the System itself seems probabilistic; regardless of the hypothesis and the data. This demands Multi-Lateral Cognitive approach

Telandic …. “Point – equilibrium – steady state – periodic (oscillatory) – quasiperiodic – Chaotic – and telandic (goal seeking behavior) are examples of behavior here placed in order of increasing complexity”

A Highly Complex System, demands a Dragon Slayer – Hyperbolic Dirac Net (HDN) driven Statistics (BI-directional Bayesian) for extracting the Knowledge from a Chaotic Uncertain System.

2nd Order Semantic Web and A.I driven Reasoning – 300 Years Plus of Crusade | Ingine Inc

Screenshot 2016-02-02 11.30.13

Chronology of Development of Hyperbolic Dirac Net (HDN) Inference.

From Above Link:-

1. 1763. Thomas Bayes was an English statistician, philosopher and Presbyterian minister who is known for having formulated a specific case of the theorem that bears his name: Bayes’ theorem.

 Bayes’s solution to a problem of inverse probability was presented in “An Essay towards solving a Problem in the Doctrine of Chances” which was read to the Royal Society in 1763 after Bayes’ death

From Above Link:-

In probability theory and statisticsBayes’ theorem (alternatively Bayes’ law or Bayes’ rule) describes the probability of an event, based on conditions that might be related to the event.

When applied, the probabilities involved in Bayes’ theorem may have different probability interpretations. In one of these interpretations, the theorem is used directly as part of a particular approach to statistical inference. With the Bayesian probability interpretation the theorem expresses how a subjective degree of belief should rationally change to account for evidence: this is Bayesian inference, which is fundamental to Bayesian statistics. However, Bayes’ theorem has applications in a wide range of calculations involving probabilities, not just in Bayesian inference.

From Above Link:-

Bayesian inference is a method of statistical inference in which Bayes’ theorem is used to update the probability for a hypothesis as more evidence or information becomes available. Bayesian inference is an important technique in statistics, and especially in mathematical statistics. Bayesian updating is particularly important in the dynamic analysis of a sequence of data. Bayesian inference has found application in a wide range of activities, including scienceengineeringphilosophymedicinesport, and law. In the philosophy of decision theory, Bayesian inference is closely related to subjective probability, often called “Bayesian probability“.

2. 1859, Georg Friedrich Bernhard Riemann proposed Riemann zeta function,function useful in number theory for investigating properties of prime numbers. Written as ζ(x), it was originally defined as the infinite series

ζ(x) = 1 + 2−x + 3−x + 4−x + ⋯.

The theory should perhaps be distinguished from an existing purely number-theoretic area sometimes also known as Zeta Theory, which focuses on the Riemann Zeta Function and the ways in which it governs the distribution of prime numbers

The Riemann zeta function is an extremely important special function of mathematics and physics that arises in definite integration and is intimately related with very deep results surrounding the prime number theorem. While many of the properties of this function have been investigated, there remain important fundamental conjectures (most notably the Riemann hypothesis) that remain unproved to this day. The Riemann zeta function is defined over the complex plane for one complex variable, which is conventionally denoted (instead of the usual ) in deference to the notation used by Riemann in his 1859 paper that founded the study of this function (Riemann 1859). It is implemented in the Wolfram Language as Zeta[s].

3. 1900. Ramanujan’s mathematical work was primarily in the areas of number theory and classical analysis. In particular, he worked extensively with infinite series, integrals, continued fractions, modular forms, q-series, theta functions, elliptic functions, the Riemann Zeta-Function, and other special functions.

Hardy wrote in Ramanujan’s obituary [14]:

There is always more i n one of Ramanujan’s formulae than meets the eye, as anyone who sets to work to verify those which look the easiest will soon discover. In some the interest lies very deep, in others comparatively near the surface; but there is not one, which is not curious and entertaining.

From above link :-

Now there is a famous account of the gifted Indian mathematician Ramanujan who when writing to Hardy at Cambridge regarding his early findings included the seemingly nonsensical result,

1 + 2 + 3 + 4 + ……(to infinity) = – 1/12.

Initially Hardy was inclined to think that he was dealing with a fraud, but on further reflection realized that Ramanujan was in fact describing the Riemann Zeta Function (for s = – 1). He could then appreciate his brilliance as one, who though considerably isolated and without any formal training, had independently covered much of the same ground as Riemann.

However it still begs the question as to what the actual meaning of such a result can be, for in the standard conventional manner of mathematical interpretation, the sum of the series of natural numbers clearly diverges.

The startling fact is that this result – though indirectly expressed in a quantitative manner – actually expresses a qualitative type relationship (pertaining to holistic mathematical interpretation).

Uncovering Ramanujan’s “Lost” Notebook: An Oral History


From above link :-

Whereas Ramanujan’s earlier work dealt largely with classical number-theoretic objects such as q-series, theta functions, partitions and prime numbers—exotic, startling, breathtaking identities built up from infinite series, integrals and continued fractions—in these newfound papers, Andrews found never-before-seen work on the mysterious “mock theta functions” hinted at in a letter written to Hardy in Ramanujan’s final months, pointing to realms at the very edge of the mathematical landscape. The content of Ramanujan’s lost notebook is too rich, too ornate, too strange to be developed within the scope of the present article. We provide a handful of stunning examples below, intended only to tantalize—perhaps mystify—the reader, who is encouraged to let his or her eyes wander across the page, picking patterns like spring flowers from the wild field of symbols.

The following are two fantastic q-series identities found in the lost notebook, published by Andrews soon after his discovery, in which is taken to be a complex number with |q| <1

Another surprising expression involves an example of a mock theta function provided by Ramanujan in the final letter he sent to Hardy

In the words of mathematician Ken Ono, a contemporary trailblazer in the field of mock theta functions, “Obviously Ramanujan knew much more than he revealed [14].” Indeed, Ramanujan then “miraculously claimed” that the coefficients of this mock theta function obey the asymptotic relation

The new realms pointed to by the work of Ramanujan’s final year are now understood to be ruled by bizarre mathematical structures known as harmonic Maass forms. This broader perspective was only achieved in the last ten years, and has led to cutting-edge science, ranging from cancer research to the physics of black holes to the completion of group theory. 

Yet details of George Andrews’s unearthing of Ramanujan’s notes are only sparsely sketched in the literature; one can detect but an outline of the tale surrounding one of the most fruitful mathematical discoveries of our era. In hopes of contributing to a more complete picture of this momentous event and its significance, here we weave together excerpts from interviews we conducted with Andrews documenting the memories of his trip to Trinity College, as well as from separate interviews with mathematicians Bruce Berndt and Ken Ono, who have both collaborated with Andrews in proving and extending the contents of Ramanujan’s famous lost notebook.

4. Elie Joseph Cartan, developed “Theory of Spinors

From above link:-

In geometry and physics, spinors are elements of a (complexvector space that can be associated with Euclidean space. Like geometric vectors and more general tensors, spinors transform linearly when the Euclidean space is subjected to a slight (infinitesimal) rotation. When a sequence of such small rotations is composed (integrated) to form an overall final rotation, however, the resulting spinor transformation depends on which sequence of small rotations was used, unlike for vectors and tensors. A spinor transforms to its negative when the space is rotated through a complete turn from 0° to 360° (see picture), and it is this property that characterizes spinors. It is also possible to associate a substantially similar notion of spinor to Minkowski space in which case the Lorentz transformations of special relativity play the role of rotations. Spinors were introduced in geometry by Élie Cartan in 1913. In the 1920s physicists discovered that spinors are essential to describe the intrinsic angular momentum, or “spin”, of the electron and other subatomic particles.

5. 1928, Paul A M Dirac derived the Dirac equation, which In particle physics, is a relativistic wave equation.

From above link:-

The quantum electrodynamical law which applies to spin-1/2 particles and is the relativistic generalization of the Schrödinger equation. In dimensions (three space dimensions and one time dimension), it is given by


6. 1930. Dirac publishes his book on his pivotal view of quantum mechanics, including his earliest mentions of an operator with the properties of the hyperbolic number such that hh = +1. It extends the theory of wave mechanics to particle mechanics. 
P. A. M. Dirac, The Principles of Quantum Mechanics, First Edition, Oxford University Press, Oxford (1930).

7. 1933. In his Nobel Prize Dinner speech, Dirac states that mechanical methods are applicable to all forms of human thought where numbers are involved.

8. 1939. DIRAC PUBLISHES HIS BRAKET NOTATION. It is incorporated into the third edition of his book.

P.A.M. Dirac (1939). A new notation for quantum mechanics, Mathematical Proceedings of the Cambridge Philosophical Society 35 (3): 416–418

9. 1974. Robson develops his Expected Information approach that preempts the Bayes Net method.

B. Robson, Analysis of the Code Relating Sequence to Conformation in Globular Proteins: Theory and Application of Expected Information, Biochem. J141, 853-867 (1974).

10. 1978. The Expected Information approach crystallizes as the GOR method widely used in bioinformatics.

Garnier, D. J. Osguthorpe, and B. Robson, Analysis of the Accuracy and Implications of Simple Methods for Predicting the Secondary Structure of Globular Proteins”, J. Mol. Biol. 120, 97-120 (1978). 

11. 1982 . Buchannan and Shortliffe describe the first medical Expert System. It is based on probabilistic statements, but sets a tradition of innovation and diverse controversial methods in automated medical inference.

Buchanan, E.H. Shortliffe, (1982) Rule Based Expert Systems. The Mycin Experiments of the Stanford Heuristic Programming Project, Addison-Wesley: Reading, Massachusetts.

12. 1985. Pearl Gives Full Accound the Bayes Net method.

Pearl, Probabilistic Reasoning in Intelligent Systems. San Francisco CA: Morgan Kaufmann (1985).

13. March 1989, Sir Tim Berners-less invented WWW: – Introduced non-linear linking of information across systems.

Tim laid out his vision for what would become the Web in a document called “Information Management: A Proposal”.Believe it or not, Tim’s initial proposal was not immediately accepted. In fact, his boss at the time, Mike Sendall, noted the words “Vague but exciting” on the cover. The Web was never an official CERN project, but Mike managed to give Tim time to work on it in September 1990. He began work using a NeXT computer, one of Steve Jobs’ early products.

14. 1997. Clifford Algebra using becomes more widely recognized as a tool for engineers as well as scientists and physicists.

Gürlebeck, W. Sprössig, Quaternionic and Clifford Calculus for Physicists and Engineers, Wiley, Chichester (1997)

15. 1999. Tim Berners-Lee described the Semantic Web vision in the following terms

I have a dream for the Web [in which computers] become capable of analyzing all the data on the Web, the content, links, and transactions between people and computers. A Semantic Web, which should make this possible, has yet to emerge, but when it does, the day-to-day mechanisms of trade, bureaucracy and our daily lives will be handled by machines talking to machines. The intelligent agents people have touted for ages will finally materialize. (1999)

16. 2000. Khrennikov gives description of a primarily h-complex quantum mechanics.

Khrenikov, Hyperbolic quantum mechanics, Cornell University Library, arXiv:quant-ph/0101002v1 (2000).

17. 2000. Bucholz and Sommer refine work showing that neural networks as inference systems modeled on the brain can usefully use the hypercomplex imaginary number h.

S. Buchholz, and G. Sommer, A hyperbolic multilayer perceptron International Joint Conference on Neural Networks, IJCNN 2000, Como,Italy, Vol. 2 of pp. 129-133. Amari, S-I and. Giles, C.L M. Gori. M. and Piuri, V. Eds. IEEE Computer Society Press, (2000).

18. 2003. Robson Points out that the Expected Information method in bioinformatics is really the use of the partially summated Riemann Zeta function, and a best choice for treatment of sparse data in data mining in general.

B Robson (2003) “Clinical and Pharmacogenomic Data Mining. 1. The generalized theory of expected information and application to the development of tools” J. Proteome Res. (Am. Chem. Soc.) 283-301, 2 

19. 2003. Nitta Shows that the power of the h-complex approach in neural nets is primarily due to its ability to solver the notorious exclusive-or logical problem in a single neuron.

Nitta, Solving the XOR problem and the detection of symmetry using a single complex-valued neuron, Neural Networks 16:8, 1101-1105, T. (2003).

20. 2003. Khrennikov consolidates the notion of an extensively h-complex quantum mechanics, but feels that i-complex, h-complex, and real world mechanics are three spate systems.

A.Khrennikov, A. Hyperbolic quantum mechanics, Adv. in Applied Clifford Algebras, Vol.13, 1 (2003). 

21.2004. Khrennikov notes possible relation between h-complex quantum mechanics and mental function.

Khrennikov, On Quantum-Like Probabilistic Structure of Mental Information, Open Systems Information Dynamics, Vol. 11, 3, 267-275 (2004).

22. 2004 Rochon shows that the full Riemann Zeta function is both i-complex and h-complex.

Rochon, A Bicomplex Riemann Zeta Function, Tokyo J. of Math.

23. 2004. Robson argues that zeta theory is a solution to high dimensionality problems in data mining.

Robson, The Dragon on the Gold: Myths and Realities for Data Mining in Biotechnology using Digital and Molecular Libraries, J. Proteome Res. (Am. Chem. Soc.) 3 (6), 1113 – 9 (2004).

24. 2005. Robson argues that all statements in zeta theory and in prime number theory are really statements relevant to data and data mining, and describes first link to Dirac’s quantum mechanics and Dirac’s braket notation.

Robson, Clinical and Pharmacogenomic Data Mining: 3. Zeta Theory As a General Tactic for Clinical Bioinformatics, J. Proteome Res. (Am. Chem. Soc.) 4(2); 445-455 (2005) 

25. 2005. Code CliniMiner/Fano based on Zeta Theory and prime number theory is used in first pioneering effort in data mining large number of patient records.

Mullins, I. M., M.S. Siadaty, J. Lyman, K. Scully, G.T. Garrett, G. Miller, R. Muller, B. Robson, C. Apte, C., S. Weiss, I. Rigoutsos, D. Platt, and S. Cohen, Data mining and clinical data repositories: Insights from a 667,000 patient data set, Computers in Biology and Medicine, 36(12) 1351 (2006). 

26. 2007. Robson recognizes that the imaginary number required to reconcile zeta theory with quantum mechanics and to allow Dirac notation to be used in inference is the hyperbolic imaginary number h, not the imaginary number i. Unaware of the work of Khrennikov, he makes no Khrennikov-like distinction between h-complex quantum mechanics and the everyday world.

Mullins, I. M., M.S. Siadaty, J. Lyman, K. Scully,G.T. Garrett, G.Miller, R. Muller, B.Robson, C. Apte, C., S. Weiss, I. Rigoutsos, D. Platt, and S. Cohen, Data mining and clinical data repositories: Insights from a 667,000 patient data set, Computers*in*Biology* and*Medicine, 36(12) 1351 (2006)

27. 2007. Robson recognizes that the imaginary number required to reconcile zeta theory with 
quantum mechanics and to allow Dirac notation to be used in inference is the hyperbolic imaginary number h, not the imaginary number i. Unaware of the work of Khrennikov, he makes no Khrennikov like distinction between h complex quantum mechanics and the every day world.

Robson, The New Physician as Unwitting Quantum Mechanic: Is Adapting Dirac’s Inference System Best Practice for Personalized Medicine, Genomics and Proteomics, J. Proteome Res. (A. Chem. Soc.), Vol. 6, No. 8: 3114 – 3126, (2007). 

Robson, B. (2007) “Data Mining and Inference Systems for Physician Decision Support in Personalized Medicine” Lecture and Circulated Report at the 1st Annual Total Cancer Care Summit, Bahamas 2007. 

28. 2008. Data Mining techniques using the full i-complex and h-complex zeta function are developed.

Robson, Clinical and Pharmacogenomic Data Mining: 4. The FANO Program and Command Set as an Example of Tools for Biomedical Discovery and Evidence Based Medicine” J. Proteome Res., 7 (9), pp 3922–3947 (2008). 

29. 2008. Nitta and Bucholtz explore decision process boundaries of h-complex neural nets.

Nitta, and S. Bucholtz, On the Decision Boundaries of Hyperbolic Neurons. In 2008 International Joint Conference on Neural Networks (IJCNN). 

30. 2009. Semantic Web starts to emerge but runs into bottleneck regarding the best approach for probabilistic treatment.

Prediou and H. Stuckenschmidt, H. Probabilistic Models for the SW – A Survey. publication/ Predoiu08Survey.pdf (last accessed 4/29/2010) 

31. 2009. Baek and Robson propose that, for reasons of bandwidth limitations and security, the Internet should consist of data-centric computing by smart software robots. Robson indicates that they could be based on h-complex inference systems and link to semantic theory.

Robson B.. and Baek OK. The Engines of Hippocrates. From the Dawn of Medicine to Medical and Phrmaceuteutical Infomatics, Wiley, 2009. 

Robson B. (2009) “Towards Intelligent Internet-Roaming Agents for Mining and Inference from Medical Data”, Future of Health Technology Congress, Technology and Informatics, Vol. 149, 157-177 IOS Press 

Robson, B. (2009) “Links Between Quantum Physics and Thought” (A. I. Applications in Medicine) , Future of Health Technology Congress, Technology and Informatics, Vol. 149, 157-177 IOS Press. 

32. 2009. Nivitha et al. develop new learning algorithms for complex-valued networks.

S. Savitha, S. Suresh, S. Sundararajan, and P, Saratchandran, A new learning algorithm with logarithmic performance index for complex-valued neural networks, Neurocomputing 72 (16-18), 3771-3781 (2009).

33. 2009. Khrennikov argues for the h-complex Hilbert space as providing the “contextual” (underlying rationale, hidden variables etc.) for all quantum mechanics.

Khrennikov, Contextual Approach to Quantum Formalism, Springer (2009) 

34. 2010. Robson and Vaithiligam describe how zeta theory and h-complex probabilistic algebra can resolves challenges in data mining by the pharmaceutical industry.

Robson and A. Vaithiligam, Drug Gold and Data Dragons: Myths and Realities of Data Mining in the Pharmaceutical Industry pp25-85 in Pharmaceutical Data Mining, Ed Balakin, K. V. , John Wiley Sons (2010).

35. 2010. PCAST. December Report by the US President’s Council of Advisors on science and Technology calls for an XML-like Universal Exchange Langue for medicine including disaggregation for the patient record on the Internet for patient access, security, and privacy. report.pdf 

36. 2011. First description of Q-UEL in response to PCAST 2010.

Robson, B., Balis, U. G. J. and Caruso, T. P. (2011)“Considerations for a Universal Exchange Language for Healthcare.” In Proceedings of 2011 IEEE 13th International Conference on e-Health Networking, Applications and Services (Healthcom 2011), 173– 176. Columbus, MO: IEEE, 2011. 

 37. 2011. Robson and Colleagues develop the method of match-and-edit instructions for extracting

Robson, B., Li, J., Dettinger, R., Peters, A., and Boyer, S.K. (2011), Drug discovery using very large numbers of patents. General strategy with extensive use of match and edit operations. Journal of Computer-Aided Molecular Design 25(5): 427-441 

38. 2011. Kuroe et al. consolidate the theory of h– complex neural nets.

Kuroe, T. Shinpei, and H. Iima, Models of Hopfield-Type Clifford Neural Networks and Their Energy Functions – Hyperbolic and Dual Valued Networks, Lecture Notes in Computer Science, 7062, 560 (2011).

39. 2012. Robson argues that h-complex algebra is an appropriate basis for Artificial Intelligence in the Pharmaceutical Industry.

Robson, B. (2012) “Towards Automated Reasoning for Drug Discovery and Pharmaceutical Business Intelligence”, Pharmaceutical Technology and Drug Research, 2012 1: 3 ( 27 March 2012 ) 

40. 2013. Goodman and Lassiter attempt to reconcile and restore interest in probabilistic semantics after a long period of domination by classical logic. 
N. D. Goodman and D. Lassiter, Probabilistic Semantics and Pragmatics: Uncertainty in Language and Thought,

41. 2013. Robson argues for importance of h-complex approach for measures in epidemiology. Robson, B. (2013)

“Towards New Tools for Pharmacoepidemiology”, Advances in Pharmacoepidemiology and Drug Safety, 1:6,

42. 2013 Robson promotes Q-UEL from a public health perspective.
B. Robson, Rethinking Global Interoperability in Healthcare. Reflections and Experiments of an e-Epidemiologist from Clinical Record to Smart Medical Semantic Web Johns Hopkins Grand Rounds Lectures (last accessed 3/14/2013).

Screenshot 2016-02-02 11.30.13

43. 2013 Robson and Caruso describe first version of Q-UEL in greater Detail.

Robson, B, and TP Caruso (2013) “A Universal Exchange Language for Healthcare” MedInfo ’13: Proceedings of the 14th World Congress on Medical and Health Informatics, Copenhagen, Denmark, Edited by CU Lehmann, E Ammenwerth, and C Nohr. IOS Press, Washington, DC, USA.;

44. 2014. Robson et al. release formal description of consolidated second version of Q-UEL.

Robson, T. P. Caruso and U. G. J. Balis, Suggestions for a Web Based Universal Exchange and Inference Language for Medicine, Computers in Biology and Medicine, 
43(12) 2297 (2013).

45. 2013. Moldoveneua expresses view that hyperbolic quantum mechanics can’t also include wave mechanics. Possible attack on Khrennikov’s idea that hyperbolic quantum mechanics can show 
interference as for waves. Signs of growing sense that hyperbolic quantum mechanics is simply the everyday world described in terms of the machinery of traditional quantum mechanics.

Moldoveanu, Non viability of hyperbolic quantum mechanics as a theory of Nature, Cornell University Library, arXiv:1311.6461v2 [quant-ph] (2013).

46. 2013. First full description of the Hyperbolic Dirac Net and its relation to Q-UEL and to Bayes Nets.

Robson, Hyperbolic Dirac Nets for Medical Decision Support. Theory, Methods, and Comparison with Bayes Nets, Computers in Biology and Medicine, 51, 183 (2013).

47. 2014. Kunegis et al.c develop h-complex algorithms for dating recommender systems.

Kunegis, G. Gröner, and T, Gottrron, On-Line Dating Recommender Systems, the Split Complex Number Approach, (Like/Dislike, Similar/Disimilar) http://userpages.uni- complex-number-approach.pdf (last accessed 6/1/2014).

48. 2015. Robson describes extension of Hyperbolic Dirac Net to semantic reasoning and probabilistic lingusitics. 

Robson, B. “POPPER, a Simple Programming Language for Probabilistic Semantic Inference in Medicine. Computers in Biology and Medicine ” Computers in biology and Medicine”, (in press), DOI: 10.1016/j.compbiomed.2014.10.011 (2015).

49. 2014. Yosemite Manifesto – a response to PCAST 2010 that the Semantic Web should provide healthcare IT, al though preempted by Q-UEL (last accessed 7/5/2014). 

50. 2015. Robson et al. describe medical records in Q-UEL format and PCAST disaggregation for patient security and privacy.

Robson, B., Caruso, T, and Balis, U. G. J. (2015) “Suggestions for a Web Based Universal Exchange and Inference Language for Medicine. Continuity of Patient Care with PCAST Disaggregation.” Computers in Biology and Medicine (in press) 01/2015; 56:51. DOI: 10.1016/j.compbiomed.2014.10.022 

51. 2015. Mathematician Steve Deckelman of U. Wisconsin-Stout and Berkeley validates the theoretical principles Hyperbolic Dirac Net.

Deckelman and Robson, B. (2015)“Split-Complex Numbers and Dirac Bra-Kets” Communications in Information andSystems (CIS), in press.

From Above Link:-

The inference net on which this dualization is performed is defined as an estimate of a probability as an expression comprising simpler probabilities and or association measures, i.e. each with fewer attributes (i.e. arguments, events, states, observations or measurements) that the joint probability estimated, where each attribute corresponds to nodes of a general graph and the probabilities or association measures represent their interdependencies as edges. It is not required that the inference net be an acyclic directed graph, but the widely used BN that satisfies that description by definition is a useful starting point for making use of the given probabilities to address the same or similar problems. Specifically for the estimation of a joint probability, and HDN properly constructed with prior probabilities, and whether or not it contains cyclic paths, is purely real valued and its construction principles represent a generalization of Bayes Theorem. Any imaginary part indicates the degree of departure from Bayes Theorem over the net as a whole, and the direction of conditionality in which the degree of departure occurs, and thus the HDN provides an excellent book-keeping tool that Bayes Theorem is satisfied overall. Specially for the estimation of a conditional probability, it follows conversely from the above that any expression for a joint probability validated by the above means can serve as the generator of an HDN for the estimation of a conditional probability simply by dividing it by the HDN counterparts of prior probabilities, whence the resulting net is not purely real save by coincidence of probability values.

52. 2015. Implementation of a web based universal exchange and inference language for medicine: Sparse data, probabilities and inference in data mining of clinical data repositories

Barry Robson and Srinidhi Boray

52. 2015. Robson, B., and S. Boray, The Structure of Reasoning in Answering Multiple Choice Medical Licensing Examination Questions. Computer Studies   towards Formal Theories of Clinical Decision Support and Setting and Answering Medical Licensing Examinations, Workshop Lecture presentation, Proceedings of the IEEE International conference of Bioinformatics and Biomedicine, 9th-11th November, Washington DC (2015)










Applying Quantum Theory for Deep Healthcare Analytics (Semantic Algebra) – Dr. Barry Robson (RQSA)

Quantum driven Cognitive Computing

RQSA Theory – Develops algorithm for The QEXL Approach; overcoming limitations in gold standard Bayesian Network; while allowing for creation of generative models. Bayesian as such is an adaptive technique.

(March 8 2013, Version March 10 2013)


1.1. Purpose and Background. The following document describes principle features of a mathematical system of practical importance in probabilistic ontology and semantics, and their applications inference and automated reasoning. Whilst applications are wide, healthcare may be the most pressing need [1].  The focus here is on aspects of practical importance in (a) decision support systems as Expert Systems (ES) derived from probabilistic rules formulated by experts, (b) decision support systems based on automated unsupervised data mining of structured data (DM), and (c) the Semantic Web (SW) and mining of unstructured data, here, essentially text analytics (TA).  The primary areas of interest to the author is in Evidence Based Medicine (EBM), Comparative Effectiveness Research (CER), epidemiology, and Clinical Decision support Systems (CDSS), and bioinformatics, which provide useful reference points and test beds for more general application.

1.2. The Need for a New Approach.  In general we are beset by a plethora of approaches for reasoning probabilistically when many probabilistic rules are available, and the same uncertainty about best practice poses problems for the SW to go probabilistic [2].  The easiest relations between things to handle probabilistically, i.e. to extract, quantify, and reason with, are “are associated with”, as in “A is associated with B and C and D” [2-6], and it has found many biomedical applications [7-14]. From many such and self probabilities or information, we can express “A if B and C and D” are readily derived. These are of course not the only relationships between things that we use when communicating information by natural language, but the latter are noteworthy in forming the basis of the Bayesian network or Bayes Net (BN) [15], which is a gold standard for probabilistic inference. It is theoretically well founded. However, the reason that we see a plethora of further approaches is that BN’s adherence to a strict set of axioms to avoid apparent difficulties makes it very restricted in application. A BN (one that is truly a BN by original definition confines) confines itself to use of conditional probabilities that reflect the above “if” relationship, to multiplicative operations implying just logical AND, it considers only one direction of conditionality when in general we are interested in inference about etiology or causes as well as outcomes, and not least it confines itself to acyclic directed graphs when networks of knowledge representation are in practice rich in cyclic paths. That last is because the ideal full treatment is a fully connected graph, not neglecting any relationship, which is necessarily dominated by cycles.  Note that neglect of relationships by a BN is equivalent to saying that they are they are there with probability 1, and whilst to be fair to BNs it is true that this implies absence of information I = −log(P), it remains an extreme assumption that  is evidently not justified when we consider the other conditional probability terms and in the context of all available data, often called the problem of coherence.

1.3. Advantages of the Present Approach. The advantages of the present approach can be referenced with respect to BN as the gold standard that does not  traditionally provide the following.

(1) Bidirectional inference, i.e. etiologies as well as outcomes.

(2) Intrinsic treatment of coherence as Bayes Rule.

(3) Cyclic paths are allowed, fall out naturally as part of the theory, and do not require iteration.

(4) Not confined to AND logic

(5) Not confined to conditional probabilities. Relators and operators may be used symbolically albeit with probabilities, or as matrices or algorithms.

(6) Probability distributions represented by vectors.

(7) Metarules with binding variables such to generate new rules and evolve the old, evolving the old. Metarules are also used to define words from simpler vocabularies.

(8) Handling of negation and, when there are double negation etc., conversion to canonical forms.

(9) Reconciliation into one rule of rules that overlap in information content or are semantically equivalent, including reconciliation of their probabilities.

1.4. The Pursuit of Best Practice for the Theoretical Basis. Quantum mechanics (QM) claims to be a system of best practice for representing observations and for inference from them on all scales, although the notorious predictions by QM in the narrow band of scale of everyday human experience  has  discouraged investigation of applicability. This is due to the perception of QM as wave mechanics, even though we are not for everyday purposes usually interested in inference about waves. For that reason, the method is based more specifically on the larger QM of Paul A. M. Dirac [16-17] who established the theoretical basis of particle physics.  Penrose [18] provides an excellent primer.  That there is sufficient breadth to Dirac’s perception of QM to encompass semantics is indicated by Dirac’s Nobel Prize Banquet Speech in 1933, “The methods of theoretical physics should be applicable to all those branches of thought in which the essential features are expressible with numbers”. Dirac was  certainly modestly referring to his extensions to physics as further extensible to  human language and thought, because  in interviews he explicitly excluded poetry and  (more controversially) economics as subjective.

The author has published extensively in areas of some relevance, but the bibliography [19-31] refers to publications since 2005 which have some relevance to an idea first broached in Ref. [31]. The present report collates the observations that have survived as useful, adjusts some nomenclature as well as perceptions, and adds integrating material.  “Best Practice” would be presumptive for these or indeed anyone’s publications  (though “pursuit of Best Practice” validly reflects the intent). Indeed, these publications show various degrees of development from some naïve initial observations, and represent a learning curve, and so are presented in reverse chronological order. This is because whilst the author has some formal training as doctorates in, essentially, biophysics and theoretical and computational chemistry, the quantum mechanics of molecules then touched (and still does) only upon aspects related to pre-Dirac quantum mechanics. That pre-Dirac era is essentially that of Schrödinger that  describes quantum mechanics as wave mechanics. It is based on the imaginary number i, the number such that ii = −1. The problematic wave nature arises because exponentials of i-complex values are periodic or wave functions: eiq =  cos(q) + i sin(q). However, Dirac rediscovered another imaginary number, although it was first noted by Cockle in 1848 and relates to a broader Clifford calculus originating roughly around that time.  It has very different consequences, as follows.



2.1. The Hyperbolic Imaginary Number. The “new” imaginary number is here represented as h such that hh = +1.  Dirac developed what is now known as the Clifford-Dirac algebra. It, and h, arose in consideration of the origins of mass, introduced general relativity into quantum mechanical systems, and founded the current “standard model” of particle physics. Actually, the Clifford Dirac algebra and particle physics has several distinct imaginary numbers (that can all be represented as matrices). However they fall into the two general classes, of character i in that their squares are −1, and of character h in that their squares are +1. Otherwise, two different flavors of imaginary number  are anticommutative, meaning that if the order is changed in which they form a product, the sign of the product changes. This includes hi = −ih.  In the core theory presented here, different flavors of imaginary number do not meet up in the equations, and the focus is primarily upon h in isolation, meaning that focus is on purely hyperbolic complex or h-complex algebra (although real numbers are of course present, and that algebra sometimes delivers purely real-valued results, because hh = +1, and h−h = 0). In physics, those of h character include Dirac’s linear operator s, g0 (or gtime), and g5. Significantly for what follows,  they particularly appear  in physics as the equivalent to ½ (1+h) and ½ (1−h) multiplied by real or usually i-complex expressions, which are called spinors. As dual spinors in which these two spinors are involved, they relate to key expressions in quantum field theory, as well as the in theory discussed here.

2.2. The h-complex Hyperbolic Function. Unlike in i-complex algebra where eiq =  cos(q) + isin(q), we are not usually concerned with making inference about waves, which may be important in physics and chemistry but not much in everyday life. As it happens, h is frequently called the hyperbolic number, and that is so because ehq  is a hyperbolic, not a trigonometric and so periodic function (see next the Section after next for an interesting example and its physical consequences). We can express the h-complex hyperbolic function in the author’s notation (which will be valuable later below).

ehq  =  i*eq + ie+q   =  cosh(q) + h sinh(q)                                                                                (1)

where the notation means

i =  ½ (1+h)

i*   =  ½ (1−h)                                                                                                                                   (2)

For completeness and to demonstrate consistency with quantum mechanics, as well as for a few practical applications for present purposes (use of waves and wavelets), note that a somewhat more general description of quantum mechanics is given by

e+hiq  =  ehiq   =   i*eiq + ie+iq  =   cos(q) + hi sin(q)                                                                (3)

The full treatment following Dirac would be to resolve i into three kinds, one for each dimension of space, implying a Clifford-Dirac product of four imaginary numbers that leads to i =  ½ (1+g5) and i* =  ½ (1-g5), to same effect as g5 is a flavor of h:  i.e. g5 g5  = +1. Although algebraically it does not of course matter, the author writes the i* term as the “lead term” i.e. as the focus of attention, for several reasons, one of which will become apparent later.

2.3. Complex Conjugation. The asterisk * used above and in other contexts below means forming the complex conjugate, i.e. changing the sign of the imaginary part, equivalent to replacing all +h by –h and vice versa. It is applied much more generally in the theory than just to rationalize i* and i*.  Usually defined for i, we need to assume that it is extensible to h. However, in cases less relevant here where we need to think in terms of both imaginary numbers, note that we need to think of applying complex conjugation as  (hi)* = ih = –hi, not h*i* = hi.

2.4. Physical and Statistical Interpretations.  As indicated above, the relation to classical as opposed to wave behavior arises because ehq  can be expressed in terms of hyperbolic functions. That means it can be expressed in terms of Gaussian functions (“normal distributions”) and their reciprocals by choice of variables in q. An example from physics is the choice of distance x = xt – x0 from starting point x0 in quantum mechanical expressions where a particle moves with momentum p=mx/t, giving q = -2pmx2/ht, with mass m, time t and Planck’s constant. Incidentally, Planck’s constant h is not of course be confused with the hyperbolic number h, and so when appearing elsewhere in discussions of physics the latter is usually written by the author in italic bold, as h, to make clear that distinction. The normalization procedure (Dirac recipe) is described later below, but for present purposes it suffices that the exponential of -2pmx2/ht is proportional to the probability of the absolute value |x|, and more generally, and as a practical application for the present context, it represents a normal distribution q = -½x2/s2 where starting point x0 becomes a mean value and √(ht/4pm) becomes the standard deviation s. The use of h does to some extent dictate a model of physical observation. To show consistency with quantum mechanics and interestingly consistency with the collapse of the wave function (e.g. Penrose reduction) interpretation, the above essentially suggests a kind of diffusion model, in which in isolation an i-complex description as a wave is the lowest energy state, but an observation of the particle as a particle and not a wave, or analogous physical perturbations, swings the description to a now lower energy h-complex one by “forcing the particle description”. This establishes a new x0 around which the particle collapses, although there is a accuracy of measurement conveyed by a standard deviation s  in general, and at very least, for the most accurate measurements that are in principle possible, we have s = √(ht/4pm). Following the perturbation, the i-complex description is restored as the lowest energy state[1]. This relates to the physicists’ spread of the wave function, but it would be billions of years to see the phenomena for an object on the scale of a household object (by which time other entropic considerations would have had their effect). It is interesting to note that the collapse of the wave function seen this way does not necessarily imply a discontinuous ih jump (or hii jump in the broader description of Eqn. 3) but a progressive rotation in time t or to an extent governed by the energy of the perturbation that may be interpretable as a field, of which demanding to measure it as a particle rather than wave is just a particularly strong case. The practical application here is that one can consider wave packets or wavelets that are progressively localized wave descriptions, and a Gaussian in the limit of being maximally localized. The applications outside of physics appear, however, to be in specialist areas such as probabilistic treatment of image analysis, and distributions generally are better described in terms of h-complex vectors, described later.

In consequence, we are considering the primary applications below as relating to the case when the parameters in q, and notably time, are fixed so that the exponentials merely relate to the single probability value such as P(A) of a state, event, observation, measurement or description A. P(A) that can empirically replace the concept of the exponential (as the statistical weight) and any normalizing factor for it.   That said, the exponential form will make appearance in which the physicists’ q  identifies with Fano’s  mutual information  I(A; B) between A and B, as described below, though we will also think of eI(A;B) as association constant K(A; B). The logarithm of the wave function y proportional to eq is seen as information that is somehow encoded. Note that eiI for any information I is a periodic function that bounds the different information and resultant probability values that one may have to the interval 0..2p. In contrast, ehI does not. h may be interpreted as adding the capacity for additional information that localizes the wave function as a Gaussian due to observations made, with a precision due to variance(i.e. square of standard deviation s2)  as a counterpart of the physicists’ ht/4pm, which relates to the physicists’ notion of  q as the action written in units of Plank’s constant h. For consistency with the physicist’s interpretation as wave function collapse due to loss and movement of information from the system, the treatment below should therefore be seen, reasonably enough,  as the gain of information to the observer. In practice the information comes from sampling and counting of an everyday system as a population, or from our belief in the result that we would obtain by doing so.

2.5. Eigenvalues of h. Unlike the case of i that has imaginary eigenvalues +i and –i, we can also (as Dirac noted) replace h by its eigenvalues that are real, either +1, or −1. This is equivalent to treating i  and  i* as linear operators with eigenvalues 0 and 1, but specifically meaning that we can set i =  1 and i*  =  0  to get one solution, and i =  0 and i* =  1  to get the other, giving two plausible physically real eigensolutions, or two sets of i-complex ones when expressions are also i-complex. Basic quantum mechanical texts gloss over this, jumping straight to eiq  and e+iq as the intuitive solutions and solving each separately,  but Dirac made it clear by stating that a wave function is always decomposable into two parts, one a product with  i and one a product with  i* (although he didn’t use that notation). In physics, they typically relate to solutions in matters of direction in time or chirality (handedness) and more generally to direction in conditionality. That is meant in the same sense that conditional probability P(A|B) = PA, B)/P(B) is of  reverse conditionality to P(B|A) = P(A, B)/P(A). In other words, the two eigensolutions do not imply indeterminacy in the sense that multiple eigenvalues would be possible interpretations.  Rather, they simply relate to two directions of inference in the network and two directions corresponding of effect of the terms in it. However, we cannot compute P(A|B) from P(B|A) or vice versa by taking the adjoint  † as the transpose and/or complex conjugate of either one of them because a classical probability is a scalar value, and has no imaginary part. In other words, P(A|B) is purely a symbolic adjoint of P(B|A). However, the effect of our i and i* operators is to render real values h-complex, and it should be held in mind that since (i)* = i* and (i*)* = i*, then (i*P(A|B) + iP(B|A))* = (i*P(B|A) + iP(A|B))*.  Then given the (0,1) eigenvalues of these operators, we have P(A|B) and P(B|A) as the two eigensolutions.

2.6. Iota Algebra. Alternatively to thinking in terms of of Eqn. 3, one can for present purposes think of pre-Dirac quantum mechanics in which h replaces i, which in physics is called the Lorentz rotation, and it is arguably a generalization of the Wick rotation in which time t is replaced by imaginary time it to render quantum mechanical expressions classical. The resulting purely h-complex algebra takes some practice and excessive familiarity with i-complex algebra can sometime be more a hindrance than a help, because one can jump to conclusions that do not hold when h replaces i, and conversely miss important algebraic opportunities that h provides. Fortunately, h-complex algebra can be rendered in a form making manipulation much easier than in i-complex algebra. The above spinor forms, more generally quantities of form i*x and iywhere  x and y are not h-complex,  are so-called by analogy with Dirac’s treatment, and can be considered as comprising spinor operators i*and i with a very convenient algebra of their own. We can usually avoid discussion any of h by using them, i.e. by using i algebra or iota algebra alone. Its simple properties include the idempotent property ii = i and i*i* = i* which for example means that ei = i  and  ei* = i*, and similarly all powers and logarithms leave i or i* changed, but not of course the non-h-complex terms they multiply.  They also include the annihilation property  ii* =  i*i = 0 that by annihilating terms in greatly facilitates multiplying h-complex expressions, and the normalization property  i+i*  =  i*+i  = 1, with the important effect that if a dual spinor form ix  +i*y where x and y are real values, then ix +i*y = x + y.  Note also that (ix +i*y)* = (i*y +i*x), important as the general statement that was implied in stating (i*P(A|B) + iP(B|A))* = (i*P(B|A) + iP(A|B))*.   Eqns. 1 and 3 cover trigonometry and hyperbolic functions. That covers almost all the new algebra that we need here, but for completeness, because the Riemann zeta function z(s=1, n)  = 1 + 2-s + 3-s +…+ n-s,  is used in data mining and treatment of finite data to estimate information and probability values, it should be noted that z(x+hy, n) = i* z(x−y, n)  + i z(x+y, n).

2.7. Dirac Notation. Let  q be a function of A and B where are A and B are observed states or events that can take on particular values, being prepared that way, or by observation. More precisely, using the author’s notation, we should write A:=a and B:=b for these states or events as measurements. Here A and B are metadata or data descriptors such as momentum in physics or systolic blood pressure in medical life, and the specific values they have are a and b respectively, constituting the orthodata or specific manifestation of A and B. However, we can take A and B etc. as implying A:=a, B:=b etc. for brevity. We have, for example, in Dirac’s bracket or braket notation,  where we can speak of a bra part <A| and a ket part |B>,

<A|B>  =  ke+hiq  =  kehiq   =   i*keiq + ike+iq  =   k(cos(q) + hi sin(q))                                  (4)

However, we shall assume that <A|B> is purely hyperbolic, i.e. there is a Lorentz rotation which here implies i → 1, and although this is not usually the case in quantum mechanical texts, a purely hyperbolic <A|B>  is a valid solution. Above, k is a real valued constant dependent on the nature and scale of the system, and importantly it must be such that k relates to P(A) and P(B) meaning P(A:=a) and P(B:=b) so that the probability  of a measurement value is purely one of chance without a prior observation of one of A and B, or in quantum mechanical language, without preparing a value of A or B.  For example, in quantum mechanics texts considering a particle on a circular orbit of length L, we see k=1/L2.


2.8. Dirac Recipe for Observable Probabilities. The fact that we set a set a prior value of A and B, and then measure B or A as conditional upon it, means that that we think in terms of  calculating  P(A|B) and P(B|A), in quantum mechanics a process that algebraically implies first a ket |B> or bra |A> normalization as the preparation of B and A.  We can write the bra normalized <A|B> as `<A|B> and the ket normalized<A|B>, whatever that might mean algebraically at this stage. In fact, in the earlier example q = 2pmx2/ht, we are conceptually obliged to apply ket normalization since q = -2pmx2t/ht becomes q = +2pmx2/ht in the term ike+iq  which can exceed an upper valid probability of 1 if  t  and m are, as in our everyday world,  positive.  Following Dirac’s recipe for observable probabilities we apply after ket normalization  P(A|B) =  <A|B> (<A|B>)*, sometimes written as the square of the absolute magnitude |<A|B>|2 according to the Born rule but implying that <A|B> is ket normalized according to the Dirac recipe. In fact, the interpretation as  |<A|B>|2  is peculiar to i-complex algebra and is not presumed here, and the more fundamental interpretation is that observation implies a projection operator P = |A><B|  acting on vectors <A| or |B> (a vector interpretation is preserved in the current theory – see later below) such that

<A| P |B> = <A|  |B><A| | B> = <A|B><B|A> = <A|B><A|B>*                                          (5)

We cannot yet express <A|B> as a function of P(A|B) and P(B|A), but the solution must satisfy

<A|B> = i *<A|B> +  i<A|B>  = i *<A|B> +  i<A|B>  = i *<A|B> +  i (<A|B>)*

= <A|B> <A|B>                                                                  (6)

P(A|B)  =<A|B> (<A|B>)*                                                                                                   (7)

P(B|A)  = <A|B> (<A|B>)*  = (<A|B>*) ((<A|B>*))*                                                      (8)

We can see the requirements for <A|B> emerging from the above by inspection, but when we apply it to Eqns. 1,3, and 4, we find that P(A|B) = P(B|A), which is the special case P(A) = P(B).

2.9. Conjugate and Non-Conjugate Variables. To move towards the required theory, we note that  Eqns. 1,3, and 4 have a conjugate symmetry not suitable for our general and more classical purposes. That is, they are composed as i*x + iy.  such that xy = 1 for all values of x and y. It arises because A:=a is a simple  function f of B:=b and vice versa, so that the probabilities are predetermined as P(A:=a) = P(f(B:=b)) = P(f(A:=a)) = P(B). That is not generally true in quantum mechanics either, but relates to the important special case of conjugate variables such as momentum and position, or energy and time, and generally where the action A = (A:=a)(B:=b) such that  q = 2pA/h. Classically, we can also be measuring values which are also such conjugate variables, like pressure P and volume V in the gas law PV = RT where R is a constant and when the absolute temperature T is constant, or current I and resistance R in the electrical engineering equation V= IR for constant voltage V. However, such cases in inference are rare, and in practice the extent to which P(A:=a) ≈ P(f(B:=b)) is more interesting as deducible from inference than as input, and moreover as part of a more general description of a relationship between A and B, i.eas the association constant

K(A; B) = P(A, B) / P(A)P(B) =  eI(A; B)                                                                                (9)

where I(A; B) is Fano’s mutual information, between A and B, noting K(A; B) = K(B; A) and I(A; B) = I(B; A)..  We will generally only require for our applications that measurements and observations are such that 0 ≤ xy ≤ 1, and more specifically that 0 ≤ x ≤ 1 and 0 ≤ y ≤ 1, because we will relate them directly to empirical probabilities from data mining or human assignment.  We can think of conjugate symmetry xy < 1 as the mother or prototype form, and we might say in physics that the dual spinor is a system in an asymmetric field that breaks the conjugate symmetry. Whilst this is most generally is of no relevance here, there is arguably one exception:  an observation or measurement that implies the above described  normalization as part of the Dirac recipe is an interaction analogous to an asymmetric field,  and  it sets one of x or y to 1.

2.10. The Braket as a Simple Linear Function of Empirical Probabilities.  To all the above consideration there is only one and simple interpretation:

<A|B>  = i*P(A|B) +  i P(B|A) = [i *P(A) +  iP(B)] K(A; B) = [i *P(A) +  iP(B)] eI(A; B)

½ h[P(A|B) + P(B|A) + ½ h [P(B|A) – P(A|B)]                                     (10)

which can be shown to satisfy Eqns. 6-9.  Several things may be noted here and in Sections immediately following. First, we can see form the last line, analogous to the quantum mechanical Hermitian commutator form, that it is not entirely true that it is the only solution; we can replace +h by –h and satisfy most of the above discussion, with one exception, that in the case of conjugate variables it was argued that we must normalize the ket, not the bra, and that can be shown to imply Eqn. 10 as written. Second, specifically in the h-complex algebra, following the Dirac recipe does not suggest that we consider the square roots of probabilities in such expressions. In particular, our bra and ket normalizations become

<A|B>  = i*P(A|B) +  i                                                                              (11)

<B|A>  = i* +  i P(B|A)                                                                              (12)

Multiplying these by their own complex conjugates delivers P(A|B) and P(B|A) respectively. Computationally, is equivalent to using the following where Re and Im are the real and imaginary parts.

P(A|B) = Re<A|B> – Im<A|B>                                                                                              (13)

P(B|A) = Re<A|B> + Im<A|B>                                                                                             (14)

2.11. Classical Probabilistic Behavior. Above it was noted that h-complex functions exhibits classical local distribution functions.. The further point about Eqn. 10 is that it also yields classical probabilistic behavior. Notably for the chain rule P(A|C) ≈  P(A|B)P(B|C) and so P(C|A) ≈ P(C|A)  that assume A and C independence, which is often physically the case, we obtain

<A|C> ≈ <A|B><B|C> = i*P(A|B)P(B|C) +  iP(C|B)P(B|A)

= [i*P(A) +  iP(B)] K(A; B) K(B; C)                                          (15)

and similarly for <A|B><B|C><C|D> and so on.

2.12. Observation Brakets. There is an important special case of Eqns. 11 and 12 when A and B are statically independent, i.e. K(A; B) = 1, and when we are at least 100% sure that we are performing a preparation or measurement, which is here represented by ? where P(?) = 1. It gives us two terms which are the counterparts of prior self-probabilities in an inference network.

<B|?>  = i * P(B|?)  +  i   =   i* P(B) +  i                                                         (16)

<?|A>  = i* +  i P(A|?)  = i* +  i P(B)                                                             (17)

Note that the following t can be readily shown with a little standard algebra and recalling i *+ i = 1 so  that  i*x + ix = x ,

<?|A><A|B><B|C><C|D<D|E><E|?>   = P(A, B)P(B, C)P(C,D)P(D|E) / P(B)P(C)P(D)   (18)

which is an estimate of  <A|E> = P(A|E), a joint probability that estimates the joint probability P(A, B, C, D, E).

2.13. Dirac Nets and Coherence.  A network built from brakets (or bra-operator-kets – see later below) may be dubbed a Dirac Net. It is common as in Bayes Nets to consider a joint probability, although we can always make it conditional on say X or (X,Y) by dividing by P(X) and P(X, Y), which makes most sense of course when X and Y are states and events in the network. We can make a joint probability first by providing a full provision of observational brakets as all terminal nodes of the network, with a caveat as shortly to be mentioned.  An advantage of  Dirac Nets is that they can be used to ensure that the network is coherent with itself and all available data , meaning that all P(A|B)P(B) = P(B|A)P(A) and so on, which is Bayes Theorem. Ironically, Bayes Nets as usually defined as acyclic directed graphs do not consider this, since they only see one direction, P(A|B)P(B). If there is not such balance overall, or if it is conditional, and in a   non-trivial way (e.g. missing probabilities are not 1), a Dirac Net value will have an imaginary part, positive or negative. In conversion of a Bayes Net to a Dirac Net, we will encounter branch points such as for example P(A| B, C) P(B|D)P(C|E). They allow that B and C are not independent in P(A| B, C)  in one direction, but that they are independent in  P(B|D)P(C|E)  in the other direction. This will typically show up as a complex and not purely real value for a network. Consideration show that we need to correct this (arguable) mistreatment by a Bayes Net by multiplying by

<? ; B, C> =  i* + iK(B; C) = <? | B, C> / <? | B > <? | C>                                                             (19)

Note that the construction <A|B><B|C>  <B|D>  etc., where we have a branch with two As, or indeed  <C|A>  <A|D> as a branch in the other direction, is valid. We need to correct accordingly in the first case by i* + i K(A; A) and in the second case by i*K(A; A) + i, where K(A; A) = P(A, A) / P(A)P(A). That is, by 1/P(A) if they are absolutely indistinguishable so that P(A, A) = P(A). But that is not necessarily the case if A is Bernoulli countable state, when P(A, A) = P(A)P(A).in that case no such correction is required. The general case of degrees of distinguishability  is considered immediately after the following Section.

2.14. Cyclic Paths.  Oddly, traditional Bayes Nets deny the possibility of cyclic paths, yet in a Dirac Net using observational brakets creates cyclic paths in order to get the answer we require.  A full and proper use of observational brakets giving a purely real value indicates a joint probability, not a conditional one, but it applies whatever additional states, say F  replace ?. Indeed it applies in replacement by two different states, say F and G, providing P(F) = P(G). It is an important feature of cyclic paths in an inference network formed by these methods that they are purely real, and appear to pose no special problems when a system is described where probabilities are in steady state, or sampled on a timescale much shorter than that over which it evolves. Such considerations show that <A|B><A|B>* = <A|B><B|A>* that yields observable probabilities  is a simple case of such as cyclic path.

2.15. Distinguishability of States and Events. Actually, the simplest cyclic path is <A|A>. In quantum mechanics , texts often state that <A|A> = 1, but that is not in general true, and is just one possibility for the distinguishability of A and B on a continuum for <A|B> as a real line defined by P(A|B) = P(B|A), i.e.  P(A) = P(B).  <A|B> = 1 is the case when A and B are absolutely indistinguishable, so that P(A) = P(B) = 1, because P(A, A) =A. If they are distinguishable by recurrence so that the As can be counted, as when counting males in a population, then recurrences are independent (Bernoulli sampling) and <A|A> = P(A), because P(A,A) = P(A)P(A) Note here that <A, A, A | A> =  P(A)3  and so on have meaning as concurrence of As.  If A and B are totally indistinguishable, they are mutually exclusive, giving the orthogonal case, because P(A,B) = 0.  Of course, many interesting cases do not satisfy P(A|B) = P(B|A), and we need to move from a real line to a plane, which can be described by a complex value. The valid region for probabilities in such a plane is in our case  the  h-complex iota space. It is contained by the vectors connecting the values 0  →  i  →  1 → i* 0. With that, we are now ready to proceed to the ontological interpretation implying the verb to be, and on to other verbs and relationships in general, which will require a h-complex vector and matrix interpretation analogous to that in quantum mechanics.



3.1 Introduction and Overview of Agenda. Conditionality discussed in Section 1 is an example case of a relationship, but has several interpretations.  Quantum mechanics does not well differentiate <B|A> from , since conditionality is typically seen as inevitably due to information relayed through cause and effect. The author uses “ontology” as closer to its older meaning of use of categorical relationships, i.e. the use of, in English, the verb “to be” and certain other related verbs as the relational operator or relator. More precisely, the term ontology is used here for the different interpretations of the relationship between nouns or noun phrases A and B in P(A|B) and <A|B>, that would reflect the way we sample to get counts that lead to the conditional probabilities. They can be distinguished by convenient readable choice of relator word or phrase in the form <A| relator |B> such as ‘if’. Whereas it is tempting, and indeed desirable to use symbols for ontological relators in particular, in applications natural language is used for readability.

Important examples follow. Here we shall not be too fussy about plurality of nouns and corresponding verb persons at present, although it can have effect in assigning probabilities. Rather, the more general existential notions of “some” and of the universal notions of “all” are the focus here(see later below).  The asterisk that implies complex conjugation can for the moment be considered as used symbolically, to indicate and active-passive inversion of the relator on which it acts.

if* =  implies (general conditionality)

when* = then (coincidence in time)

are* = include (categorical, set theoretic)

causes* = is caused by (causality)

“Semantics” is larger: it also includes other verbs such as verbs of action, where we have no choice but to write <A| relator |B> because there is way to construe that meaning by <A|B>.  Other verbs primarily differ in the probabilities they convey, so the verb “to be” is regarded as the mother form from which they are derived. Meaning, in contrast, comes (a) from the h-complex probabilistic knowledge network of h-complex terms that specifies meaning as the network context, and (b) from definitions that are not fundamentally different in action from specifications of syllogisms and other logical laws, that evolve the network to generate new probabilistic statements.

In both ontological and full semantic treatment, we can think of the nodes such as A as states having self, marginal, or  prior probabilities P(A), but they can also be viewed as parameters that set P(A|B) and P(B|A), as can be seen in Eqn. 10.  Often that is the same thing, but we may not know these probabilities associated with the states as nodes, and subjectively at least, it is easier to assign conditional probabilities P(A|B) and P(B|A), from which P(A) and P(B) follow, given KA; B) = eI(A; B), the value of which can be changed to describe the probabilistic properties of  different and diverse relationships.  More correctly node probabilities are to be seen as h-complex values vectors of state  |A> where we will equally well need <A|, but we can easily get <A| from |A>* as discussed below. The minimal perspective is that A and B are vectors of one lement, and more precisely the observational brakets discussed above, < ?|A>  =  i* + iP(A) = <A|?>* = (i*P(A) + iP(A))*. However they can be full vectors. Relators are operators that act on <A| or |B> first, to same effect, establishing the probability of the specified relationship. The development below is an agenda essentially follows symbolic manipulation  → semiquantitative manipulation → symbolic projection with quantification → sufficient vectors→ distribution vectors. We can work at any stage on this continuum between symbolic and a full vector-matrix treatment.

3.2 Hermitian Operators as Relationship Operators.  In general, any statements <A| relator |B> represent relationships in the network called probabilistic rules, or simply rules, that the applications import as XML-like tags. We can see a knowledge network as consisting of nodes or vertices A, B, C,  that represent nouns and noun phrases as analogous to the physicists’’ states, or measurements concerning the actual values of states, and vertices between them as the relationships. The two pieces of relationship information in each direction of each edge are

<A| relator |B> =                                                                                         (20)


<B| relator |A> = <A| relator* |B>                                                                                          (21)

These two equations each individually represent the active-passive inverse form of the statement with no change in meaning, as in and <‘type 2 diabetes’ | is caused by | obesity>, where “is cause by” ≡ causes*. The two equations are connected by the complex conjugate of the whole rule:

*  =                                                                                        (22)

So we can thing of a directed edge as associated with as a single complex value, encoding both directions of the relationship. The above define ontological and semantic relationships as Hermitian operators, an important class of QM operators related to data from observations and measurements. Were they not Hermitian, then * = *, as in *  =  <‘type 2 diabetes’ | is caused by | obesity>, which is true as active-passive inversion, an example of semantic equivalence, but it now loses the ability to represent two different directions of distinct effect, and in terms of the meaning, the graph is no longer a directed graph.

In the following few Sections, we would appear to be making use of the above symmetry rules in a way that is, if not exactly simply symbolic, nonetheless nominally quantitative. That this is not necessarily the case is discussed much later below. We shall focus first largely on the general semantic implications of the categorical case that is, after all extensible to verbs of action: = The two reasons why this not a good idea ouside of formulating the case with non-categorical verbs is that (a)  it multiplies the number of nodes represented in states by having a variety of properties associated with each same noun, and (b) we will typically want to inference by referring to the object noun as a state, here cat, rather than to a quality of the subject noun, here dogs as cat-chasers.

3.3. Existential and Universal Quantification. In the ontological interpretation that is specifically categorical, <A|B> is translated as follows.

<A|B> =  = =   = <A| are B>*                    (23)

One consequence is that we can think of  <A|B> in a very simple way.

  =  i P(“A are B”) +  i*P(“B are A”)

=   ½ [(“A are B”) +  P(“B are A”)] + ½ h[[(“A are B”) −  P(“B are A”)]           (24)

Note that we have switched i and  i * for  a more readable styles, as it is nice that the first term iP(“A are B”)  images .  It is important to note that the existential notions of  “some” and the universal notions of “all” are not required in these expressions, although they could be. Rather, they would relate to specific choices  the values of P(“A are B”) and P(“B are A”) dictate that.  When they are used,  they follow a QM rule that is here applied to quantifiers such as ‘the’, ‘a’, ‘one’, ‘two’ ‘many’: when such an entity is moved outside the bra, its complex conjugate is taken (however, when moved outside the ket, it is unchanged.

< quantifier A | relator |B>  = < A | quantifier*  relator |B>                                                   (25)

< A | relator | relator |B>  = < A |  relator quantifier*  relator |B>                                          (26)

So,  for example, = . So armed, we can express the extent of existential quantification as

= = Re<A| are |B> = ½ [(“A are B”) +  P(“B are A”)]                                                                                                                                                       (27)

= Im <A| are |B> = ½  h[(“A are B”) P(“B are A”)],

Im <A| are |B>  >   0                                                               (28)

= Im <A| are |B> = ½  h[(“A are B”) P(“B are A”)],

Im <A| are |B>  <   0                                                               (29)

The above “greater than” or “less than” is for ease of interpretation, but it really represents a continuum, not a discontinuity. To put it another way,  Im <A| are |B> reflects the universal case on a scale −1 to +1  mapping from  “all A are B” to “all B are A”, and if is approximately zero , we can simply say that “some A are B” and no more. It is clear logical why the existential case subtends the universal: if all A are B, or all B are A, it necessarily follows that some A are B (and some B are A).

3.4. Trivially and Non-trivially Hermitian Relationships. More generally, relators are non-trivially Hermitian,  because   is not the same as in meaning or in probability. Or rather, they have that capacity, because a trivially Hermitian operator obeys the rules in Section 2.1. In addition, however, the latter obeys = as in < Jack | marries | Jill>, and note that whilst we can write “gets married by” = married*, here married = married*. Note that English often carries different meanings in verbs as if they were overloaded operators.  Saying “the priest marries Jack and Jill” and even “the priest marries Jill” are not taken by the human reader as the same kind of meaning of “marries”, but we would need rules from the context to say that in the second case, considering whether in that religion priests can get married. The priest performs the ceremony of marriage, or “causes to be married”. Text analytics would need to perform context dependent task to distinguish the meaning, and we might say that a second operator holds with marries(2)* = ‘got married by’(2), but marries ≠ married*, so part of the distinct definition is that the verb is in the second meaning is  not trivially Hermitian.  We could write < The priest | causes | marriage> being careful that there is not another marriage performed by the priest in the same relevancy set of rules, and we note that causes and of  are non-trivially Hermitian. For ‘of’, however, a categorical interpretation requires caution. In principle, we could say that aces of spades are the the same thing as the spades of aces, since the set of spades includes an ace, and vice versa. However ‘of’ carries linguistically an non-trivial Hermitian sense, of “owned by” or that in that B is a larger set of things than A. These difficulties vanish, to some extent, in the present system. Rather, <A| relator |B> quantifies the trivial or non-trivial nature of the spefiic relation in the context of a statement. It carries asymmetry as potentially different probabilities. P(Jack marries Jill”) = P(“Jill marries Jack”) but so <A| relator |B> is then purely real, but P(“The priest marries Jill”) > P(“Jill marries the priest”) even if neither can, a prior, be said to be said to probability 0 or 1.

3.5. Para-ontological Relationships. Note that

<A| are equivalent to |B> = <A|B> =<B|A> = = 1, if P(A|B) = P(B|A)             (30)

In some sense, there is even a mother form of <A|B>, which is when A and B are seen together  at random. Following Section 2.15, if there is no such association, such that A and B are independent, then <A|B> = P(A) P(B), and mutual information I(A; B) = 0 (i.e. K(A; B) = 1). Eqn. 10 then becomes, say,

= <A|B>  = <B|A> = i*P(A)  +  i P(B)

=  ( i*P(A)  +  i)  ( i*  +  i P(B))    =   <A|?><?|B>                              (31)

This lies in the scope of QM, but just means that we have the special case P(A) =P(A|B) and P(B|A) = P(A). By the Dirac recipe

<A|B> (<A|B>)*  =   ( i*P(A)  +  i)  ( i*P(A)  +  i) * = <A|?><?|A>

= ( i*P(A)  +  iP(A)) = P(A)                                                                         (32)

But we are not confined to conjugate variables, and can bra-normalize.

<A|B> (<A|B>)*  =   ( i* +  iP(B)) ( i*+  iP(B)) *  <B|?><?|B> = P(B)                                     (33)

In general <A|?><?|A>  is not a relation but stands for the self probability P(A) of a node A in the network. There in this way no information describing the relationship, so this can be computed de novo as required from the self-probabilities P(A) and P(B) of nodes A and B. We do not need a rule (however, recall that its omission in a purely multiplicative network of which we wish to express the join probability implies that it is there with probability 1, not P(A)P(B)).

3.6. Negation. It must always be the case that <A|?><?|A>  > 0 although to be pedantic we might say “providing A exists”. But it does not follow that <A|?><?|B>  > 0. Note, then, the case when A and B are so distinguishable that they are mutually exclusive.

=  <A|B> = 0     =>    = 1                                                        (34)

It is the case of orthogonal vectors <A| and |B> discussed later below. This is actually an important case, because it is quite plausible to build a network in which nodes A, B, C, have distinct and non-overlapping meanings, or that some nodes in a network do. Cats and dogs can be nodes, and in the ontological interpretation = , evidently the value is zero. But it does not mean that, for example,   is zero.

At first glance the rule, that when we move a quantifier outside of the bra we take the complex conjugate, at least symbolically, need not be applied here. That is because at first glance  it seems that the words ‘no’, ‘not’, ‘none’ and ‘non-‘ do not obviously change the meaning when so moved, as if trivially Hermitian, e.g. not = not*. The situation is more subtle.

< no A | are |B> = < A | none are |B> = <A| are not |B> = <A| are |not B> =

= = = 1              (35)

The last equality   = , which is not completely obvious to initial contemplation, is the logical law of the contrapositive. It includes, for example, “mammals are cats” ≡ “non-cats are non-mammals”, which takes a momen’s thought.   But even more subtly, it holds quantitatively  only under certainty, otherwise, using our original example, and can have different associated probabilities. Conversely we can see that does not have the same value as in any event, because while the latter is absolutely true as read, the former is only very occasionally true, in the sense that non-birds can be reptiles or fish, or trees etc. For these reasons we apply the out-of-bra and out-of-ket rules, and distinguish between

= <A| are |not B> =

=  = <A| are not* |B>                                                      (36)

It may well be argued that these are not the same thing anyway, even to causal inspection, but then consistent with that, we are saying that ‘not*’ is different to ‘not’. The real point is it conveniently places the two distinct forms of negation within the relator phrase, i.e. as a property of the relationship.

3.7. Subjective Quantification as Semiquantitative Quantification.  In several Sections following, how statements in a network interact depends on assigning probabilities that are valid combinations in that context, even if not necessarily true as reasonable estimates. If data is structured, even containing relationships, we can establish detailed probability values at least for that set of data. But reasonable estimates are of course desirable in every case, including  textual and anecdotal statements about relationships, and need not be confined to the idea of an authoritative statement having probability 1 merely because the author stated it.  This is harder, but probabilities are nonetheless constrained or guided in many examples considered so far. We could frequently at least say that forward and reverse values are equal, or one specified one is a lot larger than the other, but the other is definitely not zero. For example P(“cats are mammals”) = 1, while the reverse probability P(“mammals are cats”) is problematic. Consistent with the forward direction, it is reasonable that such probabilities relate to the size of sets they describe. For example, given that what you are sampling is a mammal, what is the chance of it being a cat, or what proportion of individual mammalian animals are also cats?  Without any further conditions, this would require knowledge of the number of mammals and number of cats (number of individuals, not species). It is the kind of calculation that some people, and especially demographers, like to do as an exercise in indirect estimation. For example, in a detailed analysis Legay estimated 400 million cats in the world (although others estimate more than 500 million).  Estimates of the number of mammalian species is relevant to estimating the number of individual mammals, and is  around 6000 From that one guesses that the number of mammals, if cats were representative, is 400 x 6000 x 1000000 = 2.4 x 1012.  Importantly also, if the 6000 mammalian species were equally populated, we could then say that P(cats | mammals) = P(“mammals are cats”) ≈ 1/6000 ≈  0. 00017. That gives some idea of the magnitudes if figures encountered,  but assuming equal sized populations a bad assumption. Typically such distributions follow Zipf’s law that predicts that out of a population M of N elements, the probability P(e(k) | M)  of elements of rank k is

P(e(k) | M)  =  (z(s, k)  ̶   z(s, k ̶ 1)) / z(s, N)                                                                         (37)

where we have written it maximally in terms of Riemann’s zeta function [3] defined as  z(s, n)  =  1 + 2 ̶  s + 3 ̶  s + …+ n ̶  s ).  The zeta function has more general significance as the amount of information obtained by counting things. Although it does not immediately solve the P(cats | mammals)  estimation problem, we note that P(cats | mammals) also estimated in terms of zeta functions is in our example case

P(cats | mammals)  =  ez(s=1, n[cats, mammals])  ̶  z(s=1, n[mammals]) =  ez(s=1, n[cats])  ̶  z(s=1, n[mammals])         (38)

Here n[A] in general means “number of A”, i.e. the counted number, or observed frequency,  of A. Note n[cats, mammals] = n[cats], since all cats are mammals. Putting the above together (along with, strictly speaking, considerations of the next Section) we have for 6000 mammalian species, and with cats ranked k[cats] as the k[cats] most populous species,

P(cats | mammals) = P(“mammals are cats”)

= ez(s=1, n[cats, mammals])  ̶  z(s=1, n[mammals])

= ez(s=1, n[cats])  ̶  z(s=1, n[mammals])

= (z(s, k[cats])  ̶   z(s, k[cats] ̶ 1) / z(s, N=6000)

=  k[cats] ̶ s / z(s, N=6000)

logeP(cats | mammals) =  z(s=1, n[mammals])  ̶   z(s=1, n[cats])

=  s logek  +  loge z(s, N=6000)

For the simplest case of s = 1             and large amounts of data,

P(cats | mammals) = P(“mammals are cats”)

= ez(s=1, n[cats])  ̶  z(s=1, n[mammals])

≈   n[cats]) / n[mammals]

= (z(s, k[cats])  ̶   z(s, k[cats] ̶ 1) / z(s, N=6000)

=  1 /   ( k[cats] ( loge(6000) + 0.5772))

z(s=1, n[mammals])  ̶   z(s=1, n[cats])

=  logek  +  loge (loge z(s, N=6000) + 0.5772))

≈    logek  +  2

where 0.5772… = g,  the Euler-Mascheroni constant relating logarithms and zeta functions That loge ( loge z(s=1, N=6000) + 0.5772)) ≈ 2 is a reasonable approximation for a broad number of  estimates of any number of taxonomic groups considered of which the group of specific interest, here domestic cats, is one: for N=100,  it is more precisely 1.5, for N=1000,000, it is more precisely 2.6. We can note now that 1 ≤ k ≤ N, since it is the kth of the N groups. Hence logek cannot itself exceed the value of  approximately 2 and so, again approximately, it  lies in the range 0…2.  For the 10th in the ranked list it would be about 0.8, for the  100th about 1.5, for the 1000th about 1.9, and for  the 3000th (the median of the ranked list), it would be about 2.1. As a ball park estimate of what we expect for  z(s=1, n[mammals])  ̶   z(s=1, n[cats]), we have z(s=1, n[mammals])  ̶   z(s=1, n[cats]) ≈ 4. In consequence,

P(cats | mammals) = P(“mammals are cats”) = ez(s=1, n[cats])  ̶  z(s=1, n[mammals])  ≈  0.018.

However, we assumed s=1, and s can be critical in the Zip’s distribution. For large s and large N, z(s, N) → 1

P(cats | mammals) = P(“mammals are cats”)

= ez(s=1, n[cats])  ̶  z(s=1, n[mammals])

= (z(s, k[cats])  ̶   z(s, k[cats] ̶ 1) / z(s, N=6000)

=  k[cats] ̶ s + 1


z(s=1, n[mammals])  ̶   z(s=1, n[cats])

=  logek  +  loge (loge z(s, N=6000) + 0.5772))

≈   2s, s >> 1

and so

P(cats | mammals) = P(“mammals are cats”)  ≈  e ̶ 2s, s >> 1, and as reasoned earlier, P(cats | mammals) = P(“mammals are cats”)  →  ~ 0.018 when s → 1. Note that if we simply set s=1 in e ̶ 2s, then e ̶ 2  ≈ 0.14. The above 0.018 is two orders of magnitude higher than the value of 0.00017 reasoned  even earlier above if all species in the Mammalia are equally densely populated, but that kind of qualitative disagreement is what may be expected by the more realistic distribution of mammalian species that we see, and by Zipf’s law. It suggests that s must be significantly greater than 1. More precisely, we have   ̶ loge0. 00017 = 8.68 ≈ 2s, and so s ≈ 4.

3.9. Objective Treatment in Terms of Zeta Functions. The bases of the follow considerations are found in Refs. [3-6] and references therein. It  is useful for calculations to note that we can rewrite Eqn. 10 in zeta function terms, with N as total data. Retaining expected frequencies, this is as follows.

<A|B>   = [ i*P(A)  +  iP(B)]eI(A; B)  

               =  [ i* ez(s=1, n[A])  ̶  z(s=1, N)  +  i ez(s=1,n[B])  ̶  z(s=1, N)]  ez(s=1,n[A, B])  ̶  e[A, B])                

               =  [ i* ez(s=1, n[A])i ez(s=1,n[B])] ez(s=1,n[A, B])]   ̶  z(s=1, e[A, B])  ̶  z(s=1, N)                                   (39)

Here e[ ] is an expected frequency, calculated on the classical, e.g.chi-square test basis: e[A, B] = n[A]n[B]/N = n[A]P(B) = P(A)n[B]. In the cat and mammal example, the above becomes

i* ez(s=1, n[mammals])   i ez(s=1,n[cats])]ez(s=1, n[cats, mammals]])   ̶  z(s=1,e[ cats, mammals]) e  ̶  z(s=1, n[animals])      

=  i* ez(s=1, n[mammals])   iez(s=1,n[cats])]e  ̶ z(s=1,n[mammals])  ez(s=1, n[animals])  e  ̶  z(s=1, n[animals])              

=  i* ez(s=1, n[mammals])   iez(s=1,n[cats])]e  ̶ z(s=1,n[mammals])

=   i*   i ez(s=1,n[mammals])  ̶  z(s=1, n[cats])                                                                                   (40)

The above estimate of P(cats | mammals) reflects the choice of  z(s=1, n[cats, mammals])  ̶  z(s=1, n[mammals])  as the expected information E( I(mammals | cats) | D[cats, mammals]) given data D.  Incidentally, note that n[cats, mammals] means n[cats & mammals], and the converse probability P(mammals | cats) =1 reflects the fact that we know that all cats are mammals, so we could replace n[cats, mammals]) by n[cats].

For what follows,  note that we can get rid of the sometimes problematic total amount of data N by avoiding expected frequencies e[   ].

<A|B>   = [ i*P(A|B)  +  iP(B|A)]

               =  [ i* ez(s=1, n[A, B])  ̶  z(s=1, n[B])  +  i ez(s=1,n[A,B])  ̶  z(s=1, n[A]            )

               =  [ i* ez(s=1, n[B])i ez(s=1,n[A])] ez(s=1,n[A, B])]                                                                            (41)

3.10. Subjective Quantification. The two preceding Sections essentially addressed subjective and objective data, in that order. We can represent either, or combine both, as follows

Importantly, we can include expected frequencies of another kind [1,5,6], namely b[  ] that are based on subjective prior belief about the values.

<A|B>   = [ i*P(A|B)  +  iP(B|A)]

                  =  [ i* ez(s=1, n[B])+b[B])i ez(s=1,n[A])+b[A])] ez(s=1,n[A, B]+b[A, B])

≈  [ i* Be(A|B)ez(s=1, n[A])iBe(B|A) ez(s=1,n[B])] ez(s=1,n[A, B])]                                   (42)

Here Be( ) are the pior, probability-like, degrees of belief, in the values, Be(A|B) = ez(s=1, b[A|B]-b[A] and Be(B|A) = ez(s=1, b[A|B]-b[A]. If we have zero objective (frequentist) information from counting, of course <A|B>   = i* Be(A|B)  +  iBe(B|A). The use of Be(  ) rapidly becomes a good approximation as the b[ ] increase; if any of their their values approach  then the full zeta  function representation should be used.

3.11. Chain of Effect.  Further development benefits from the above quantitative and semi-quantitative ideas and the constraints imposed on the nature of the probabilities when used collectively in inference. Notably, when consider AB |B> CD |D> have to either assume that some kind of linker such as <B|C> orBC |C> has a value of approximately 1, or provide a linker. A note should be made here on why we care. After all, the example   addresses little more than (a) establishing a web of relationships that help define nouns and noun-phrases, (b) can occasionally (albeit rarely) be important  inference in deducing that dogs chase some mice-chasers, and (c) for <mice|?> establishing a collective truth of the statements. But there can be an important meaning when there is an implicit or explicit chain of effect.   Categorical and causal relationships are a clear case of such. However, catching, transporting, selling, and eating fish can have important meaning as a chain of effect for an epidemiologist when one of several possible lakes are contaminated and the outcome is an incidence of food poisoning, and when they wish to establish most probable origins as well as predict outcomes. Indeed, every action exerted or something in the vicinity leaves some sort of trace on where and what it acts upon (e.g. culprit DNA), this being a principle of forensic science. Some are, however, more interesting than others. A large knowledge network may contain interesting and uninteresting cases from a probabilistic inference perspective, but even the uninteresting cases can affect the probabilities finally deduced from a joint probability that addresses a topic of interest. Those rules that are directly relevant constitute the relevancy set.

3.12. Symbolic Projection. There is an intermediate possibility between the scalar treatment so far, and the vector-matrix approach to follow. It is useful when probabilities are to be assigned by a human expert. For a more general relationship such as a verb of action, it is convenient to think of a symbolic projection of value from the operator into the bra and ket. It is an intermediate step towards a less extensively symbolic treatment.

<A| relator |B> =  i* P(A:=relator,  B:=relator*)   +  iP(B:=relator, A:=relator*)

=   i* <A:=relator | B:=relator*> +  i<B:=relator | A:=relator*>                                (41)

By thinking of the relator as “causes”, which is a kind of  prototype action verb, we have a reference point allowing us to say the following under the causal interpretation of conditional probabilities as a joint probability.

<A| causes |B> =  i* P(A:=causes,  B:=causes*)   +  iP(B:=causes, A:=causes*)

= <B|A> =  i* P(B|A)   +  iP(A|B)


That is, P(A|B) = P(B:=causes, A:=causes*). Consistent with this, <?|A><A| relatorAB |B><B| relatorBC |C><C|?> requires linkers such as i*P(B:=relator*) and iP(A:=relator*)  to fill in gaps to complete the chain rule such as P(A|C)  ≈P(A|B)P(B|C). Including links and associating then with appropriate brakets in a network intended to calculate a joint probability, it can readily be shown to be equivalent to using the following for each braket. It gives us Eqn. 41. <?|A><A| relatorAB |B><B| relatorBC |C><C|?>  does not, as a consequence,  contain redundant terms, and although Eqn. 35 is not conditional in its probabilities, it remains asymmetric and typically has an h-complex value. So in asking a human about the probability that A causes B and the probability for the converse, the question in this case is as to with what probability doogs  doing chasing occurs with cats being chased, and the probability that cats doing chasing occurs with dogs being chased.

There is an algebra that links the metadata operator ‘:=’ to conditionality, e.g. P(mammals:=cats) = P(mammals | cats). It suggests P(A|B) = P(A:=causes,  B:=causes*)  = P(A , B  | causes, causes*), but that is not correct because it loses the information as to which of A and B are the cause. It exemplifies the limitations of an approach that is still essentially in terms of conditional probabilities and <A|B>, when it is QM’s <A| operator |ket> that is needed.

3.13. Sufficient Vectors. The following has the limitation that it is only really suitable for (a) ontological relations (i.e. that are some kind of interpretation of <A|B>, not necessarily categorical), and (b) trivially Hermitian relators in general. Actually, the overall non-trivial Hermitian effect can be represented, but it is dependent on choosing particular different values for the probabilities P(A), P(B) etc. implied by the nodes. That may conflict with values that we want the nodes to have when engaged in other interactions. That <A| and |B> are vectors of at least two elements is not strictly true in the Dirac notation. The significance of the notation <A|  and |A> is that  one is the transpose of the other, but we add the requirement that one is also the complex conjugate of the other, and so the general requirement is,

<A| = |A>*                                                                                                                              (24)

The above double * and T consideration follows from the use of complex algebra (i-complex and h-complex) to represent directions of effect, as in <A|B> = <B|A>* and = *. But if x is a scalar real number, it is unchanged by the transpose when seen as a vector or matrix of one element, and unchanged by complex conjugation because its imaginary part is zero, and the Dirac notation as implying the above transformations means that

<x| =  x* = x,    |x> = x,                                                                                                            (25)

On the other hand, it also follows that if we address a complex scalar value say x+hy where y is also a scalar real value,

< x+hy | =  (x+hy)* = xhy,    | x+hy>  = x+hy,                                                                       (26)

Consistent with that, but not requiring that the vectors are of one element, we can think of the observation brakets as vectors (not necessarily of one element), and a projection matrix over the whole space as an identity matrix|?><!–?| = I, then times eI(A; B) To see the relation with QM for operator, we can write

<a|  relator=”” |b=””>  =  <a|  <b=””>|?><!–?| eI(A; B)   |B>  =  <a|  <b=””>I eI(A; B)   |B>                                         (27)

It is sufficient to think of vectors of one element  <a| ==”” <a|<b=””>?> = i *P(A) + i, and  |B> = <?|B> = i * + iP(B), because the above equation holds true for I = 1.  Recall however that we can write <a|b> = [i *P(A) + iP(B)] eI(A; B)   = <a|?> eI(A; B)  <?|B>, so we can write

<a|  relator=”” |b=””>  =  <a|  e<sup=””>I(A; B)   |B>  =  <a|?>  eI(A; B)   <?|B>  = <a|b>                             (28)

What this curious equation means is that <a|?>  eI(A; B)   <?|B>  is the case <a|?><?|B>  for I(A; B) =0, and A and B are independent. In this case, the relator here inserts mutual information on A and B as independent. There is just a single eigenvalue, though we could consider that an relator implies a distinct eigenvalue from the space of all possible relators. In that sense we can represent the probabilistic effect of any relator I eI(A; B)   just by inserting into <a|b> an exponential of a new I(A; B). We can accept any I(A; B) because it has the property of being independent of P(A) and P(B) respectively, since it abstracts them as in I(A; B) = lnP(A, B) – ln(A) – lnP(B). I(A; B) varies much less than P(A) and P(B) when used as a metric that is measured from  on population and applied to another, including the population of a single patient for which we wish to perform inference as to best action. However, the limitations mentioned, at the beginning of this Section, hold.

3.14. Operators Acting on Orthogonal Sufficient Vectors. The following works for orthogonal vectors, i.e. when <a|b> = 0. The extent to which that is not actually a restriction is discussed below. Recall that [a, b; c, d] [p, q]T = [ap+bq, cp+dq]T is the product of a matrix with a column vector, and [p, q] [a, b; c, d] = [pa+qc, pb+qd] is the product of a row vector with a matrix. Let us first consider the matrix

Q(A, B) = [0,  i eI(A;B) ; i*eI(A;B),0]  =  [0,  ii*,0]  eI(A;B)                                                    (29)

withectors <a| and=”” |b=””> are of consistent form satisfying <a| ==”” |a=””>*.

|B>  = [i*P(B),  i)]T

<b|  = =”” [<b=””>iP(B),  i*]

<a|  ==”” [<b=””>iP(A),  i*]                                                                                                                    (30)

Q(A,B) |B>  =  [0,  i eI(A;B) ; i*eI(A;B),0]  [i*P(B),  i)] =   [ie I(A;B), i*P(B) e I(A;B)]T                        (31)

<a| Q(A,B) |B>  =    [iP(A), i*] [iP(e I(A;B), i*P(B)e I(A;B)] T = iP(A)e I(A;B) + i*P(B)e I(A;B)

=i*P(A|B) +iP(B|A)                                                                                                 (32)

This is not a restriction if we ensure that all relationships are of character. In other words, even for the basic ontological cases we use [0,  ii*,0] eI(A;B)  as a joining operator, and one that stands for  (using our English type notation of earlier) ‘if’, ‘whe’n,’ is caused by’, and ‘include’. If we want to have another relation we use a new I(A; B) to represent its different probabilities. This attempt has a serious problem, however, in the complex conjugates of the matrices the idempotent multiplications ii= i  and i*i* = i*are replaced by the annihilations are  ii* = i*i = 0, and so the net result is <a| Q(A,B) |B>  =   0. We would be confined to replacing, for example   by <b| include=”” |a=””>, because = <a| are=”” |b=””>  = 0.   Constructions like   by <b| include=”” |a=””> are semantically equivalent in the categorical interpretation, but irksome and restricting. However, it has its uses, notably as the effective matrix with time dependent modifiers. For example,

<a|  now=”” destroys=”” |b=””>  = 1

= 0

This is beyond scope of the present introduction. For present purposes, it is good that we can prevent annihilation by using the matrix [0, i*+i eI(A;B) ; i*eI(A;B)+i,0]  that leaves an idempotent multiplication. Using vectors

|B> = [i*,  i P(B)]T

<a| ==”” [<b=””>i, i* P(A)]                                                                                                                      (33)

we have

R(A,B) |B> = [0, i*+i eI(A;B) ; i*eI(A;B)+i,0]  [i*,  i P(B)]T =[iP(B)e I(A;B), i*e I(A;B)]T                         (34)

<a| R(A,B) |B>  =  [i, i* P(A)] [iP(B)e I(A;B), i*e I(A;B)] T = iP(B)e I(A;B) + i*P(A)e I(A;B)

=i*P(A|B) +iP(B|A)                                                                                                 (36)

We can again change I(A; B) to represent the probabilities involved in a new relationship.

3.15. Distribution Vectors. In general, the appearance of an operator implies in QM that it is a matrix and that we see <a| and=”” |b=””> as vectors between which the operator sits, and it can act on either <a| or=”” |b=””> first to same effect.  Following QM exactly, we have

<a| = =”” [<a|y<sub=””>0>, <a|y1>, <a|y2>, ….]                                                                                  (32)

|B> =  [<y0|B>, <y1|B>, <y2|BA>, ….]T                                                                               (33)

which obey the general rules described in the next section with T here indication transposition to a column vector, and for which vector multiplication satisfies the QM law of composition of probability amplitudes (and exemplifies a kind of inference network).

<a|b> = Si = 0,1,2,3,…n<a|yi><yi||B>                                                                                       (34)

The problem is that  in QM,  y is the universal wave function (universal quantum state) to which the probabilities of all other states A, B, C, etc can be referred with high precision. To do that it represents a information repository with a trans-astronomic, indeed by definition, cosmological, number of bits 0,1,0,0,1,1,1,…In practice, QM practitioners  focus not on the universe but a specific subsystem of interest.  In our case, we need to choose an ubiquitous if not universal state that is of interest. The obvious one for A as, say, obesity, or diabetes type 2, or systolic blood pressure  140 mmHg, or systolic blood pressure greater than  140 mmHg, is that y  is age, and the indices of  y 0, y1, y2 are its value (how many years old).  Whatever the base chosen (here age), vectors represent probability distributions when expressed as above, and so represent distribution vectors.

Use of matrices to act on such distributions is rare in our current applications and beyond scope of tis introduction. Note, however, that we can construct projection operators generally as Pi = 0,1,2,3,…n|yi><yi||. We can sum over many P to obtain a new P  if their component ket and bra are not orthogonal, and the result is the identity operator I if we sum over the  the whole larger space, not just n-dimensional, meaning we consider  more elements than just the n elements in the above vectors. We can multiply a projection operator by the analogue of the exponential of mutual information to obtain an operator (which is not a projection operator, since the effect of including this exponential is that result squared does not return itself, but itself times the square of  this exponential).


  • Robson, B., and Baek, O.K. (2009) “The Engines of Hippocrates: From the Dawn of Medicine to Medical and Pharmaceutical Informatics” B. Robson and Ok Beak (2009), John Wiley & Sons
  2. Robson , B. (2003)  “clinical and Pharmacogenomic Data Mining. 1. the generalized theory of expected information and application to the development of tools” J. Proteome Res. (Am. Chem. Soc.) 283-301, 2
  3. Robson, B.,  and Mushlin, R.  (2004) “clinical and Pharmacogenomic Data Mining.. 2. A Simple Method for the Combination of Information from Associations and Multivariances to Facilitate Analysis, Decision and Design in Clinical Research and Practice. J. Proteome Res. (Am. Chem. Soc.) 3(4); 697-711
  4. Robson, B (2005) . Clinical and Pharmacogenomic Data Mining: 3. Zeta Theory As a General Tactic for Clinical Bioinformatics. J. Proteome Res. (Am. Chem. Soc.) 4(2); 445-455
  5. Robson, B. (2008) Clinical and Pharmacogenomic Data Mining: 4. The FANO Program and Command Set as an Example of Tools for Biomedical Discovery and Evidence Based Medicine” J. Proteome Res., 7 (9), pp 3922–3947
  6. Mullins, I. M., Siadaty, M. S., Lyman, J., Scully, K., Garrett, C. T., Miller, W. G., Robson, B., Apte, C., Weiss, S., Rigoutsos, Platt, D., Cohen, S., Knaus, W. A. (2006) “Data mining and clinical data repositories: Insights from a 667,000 patient data set” Computers in Biology and Medicine, Dec;36(12):1351-77
  7. Robson, B., Li, J., Dettinger, R., Peters, A., and Boyer, S.K. (2011), Drug discovery using very large numbers of patents. General strategy with extensive use of match and edit operations. Journal of Computer-Aided Molecular Design 25(5): 427-441 (2011)
  8. Svinte, M., Robson, B., and Hehenberger, H.(2007)  “Biomarkers in Drug Dvelopment and Patient Care” Burrill 2007 Person. Med. Report. Vol. 6,  3114 – 3126. 8
  9. Robson, B. and McBurney, R. (2013) “The Role of Information, Bioinformatics, and Genomics” pp 77-94 in  Drug Discovery and Development. Technology in Transition, Churcill Livingstone, Elsevier
  10. Robson, B (2013) “Rethinking Global Interoperability in Healthcare. Reflections and Experiments of an e-Epidemiologist from Clinical Record to Smart Medical Semantic Web” Johns Hopkins Grand Rounds  Lectures 80245ac77f9d4fe0a2a2 bbf300caa8be1d
  11. Robson, B. (2013)“Towards New Tools for Pharmacoepidemiology”, Advances in Pharmacoepidemiology and Drug Safety, 1:6,, in press.
  12. Robson, B. and McBurney, R.  (2012) “The Role of Information, Bioinformatics and Genomics”, pp77-94 In Drug Discovery and Development: Technology In Transition, Second Edition, Ed.  Hill, R.G., Rang, P. Eds. Elsevier Press.
  13. Robson, B., Li, J., Dettinger, R., Peters, A., and Boyer, S.K. (2011), “Drug Discovery Using Very Large Numbers Of Patents. General Strategy With Extensive Use Of Match And Edit Operations”, Journal of Computer-Aided Molecular Design 25(5): 427-441 (2011)
  16. Dirac,. P. M. (1930) “The Principles of Quantum Mechanics”, Oxford University Press.
  17. Penrose, R.  (2004) “The Road to Reality: A Complete Guide to the Laws of the Universe”, Vintage Press
  18. Robson, B. “Schrödinger’s Better Patients”, Lecture and Synopsis  University of North Carolina, (4/16/2012)
  19. Robson, B. (2012) “Towards Automated Reasoning for Drug Discovery and Pharmaceutical Business Intelligence”,  Pharmaceutical Technology and Drug Research, 2012 1: 3 ( 27 March 2012 )
  20. Robson, B.,  Balis, UGC , and Caruso, T.P. (2012), “Considerations for a Universal Exchange Language for HealthcareIEEE Healthcom ’11 Conference Proceedings, June 13-15, 2011, Columbia, MO pp 173-176
  21. Robson, B (2009),  “Towards Intelligent Internet-Roaming Agents for Mining and Inference from Medical Data”, Studies in Health Technology and Informatics,
    Vol. 149  pp 157-177

    • Robson,B.  (2009),  “Links Between Quantum Physics and Thought  (for Medical A.I. Decision Support Systems) ”, Studies in Health Technology and Informatics, Vol. 149, pp 236-248
    • Robson. B (2009) “Artificial Intelligence for Medical Decisions”  14th Future of Health Technology Congress,  MIT , September 29-30 2010 and Robson. B (2009) “Using Deep Models of Medicine and Common Sense to Answer ad hoc Clinical Queries” 14th Future of Health Technology Congress,  MIT , September 28-29, 2009.
    • Robson B., and Baek OK (2009), “The Engines of Hippocrates: From the Dawn of Medicine to Medical and Pharmaceutical Infromatics.” BOOK,  600 pages, pub. Wiley.
    • Robson B., and  Vaithilingam A.   (2009) “Drug Gold and Data Dragons. Myths and Realities of Data Mining in the Pharmaceutical Industry” in Pharmaceutical Data Mining: Approaches and Applications for Drug Discovery,  Ed. Konstantin V. Balakin, Pub. Wiley
    • Robson B.  (2008)  “Clinical and Pharmacogenomic Data Mining: 4. The FANO Program and Command Set as an Example of Tools for Biomedical Discovery and Evidence Based Medicine” J. Proteome Res. (Am. Chem. Soc.). ” J. Proteome Res., 7 (9), pp 3922–3947
    • Robson B., and Vaithilingam, A. (2008)  “Protein Folding Revisited” in Molecular Biology and Translational Science, Vol. 84 Ed. Kristi A.S. Gomez., Elsevier Inc.
    • Robson B. and Vaithilingam, A. (2007)  A. “New Tools for Epidemiology, Data Mining, and Evidence Based Medicine”. Poster,  10th World Congress in Obstetrics & Gynecology, Grand Cayman, 2007.
    • Robson, B. (2007)  “Data Mining and Inference Systems for Physician Decision Support in Personalized Medicine”. Lecture and Circulated Report at the 1st Annual Total Cancer Care Summit, Bahamas 2007
    • Robson B.  (2007) “The New Physician as Unwitting Quantum Mechanic: Is Adapting Dirac’s Inference System Best Practice for Personalized Medicine, Genomics and Proteomics?”  J Proteome Res. (Am. Chem. Soc.), Vol. 6, No. 8: pp 3114 – 3126

    [1] It should be declared that a Gaussian function can also be reached within an i-complex description provided that we see a particle as a harmonic oscillator in the ground state, for which the wave function solution is well known to be a Gaussian function. The particle is then seen as oscillating around the fixed point x0. In that sense, h-complex quantum mechanical descriptions relate to ground state harmonic oscillations in the i-complex description,  but do not explain why a wave and not particle is a lower energy state in the absence of perturbation (unless energy is withdrawn from the system by perturbation). However, the rest mass as a kind of ground state of motion in the Dirac equation for mass predicts the mass as due to an oscillation. Here the rest mass is computed via igtime ∂y/∂t where gtime is a hyperbolic number, suggesting the more general hi description as starting point.

Quantum Theory driven (QEXL Approach) Cognitive Computing Architecture resolving Healthcare Interoperability (BigData – HIE/ ACO )

[healthcare cognitive computing platform]

Conquering Uncertainties Creating Infinite Possibilities

(Possible application :- Achieving Algorithm Driven ACO)



The QEXL Approach is a Systems Thinking driven technique that has been designed with the intension of developing “Go To Market” solutions for Healthcare Big Data applications requiring integration between Payor, Provider, Health Management (Hospitals), Pharma etc; where the systemic complexities tethering on the “edge of chaos” pose enormous challenges in achieving interoperability owing to existence of plethora of healthcare system integration standards and management of the unstructured data in addition to structured data ingested from diverse sources. Additionally, The QEXL Approach targets for the creation of Tacit  Knowledge Sets by inductive techniques and probabilistic inference from the diverse sets of data characterized by volume, velocity and variability. In fact, The QEXL Approach facilitates algorithmic driven Proactive Public Health Management, while rendering business models achieving Accountable Care Organization most effective.

The QEXL Approach is an integrative multivariate declarative cognitive architecture proposition to develop Probabilistic Ontology driven Big Data applications creating interoperability among Healthcare systems. Where, it is imperative to develop architecture that enable systemic capabilities such as Evidence Based Medicine, Pharmacognomics, biologics etc; while also creating  opportunities for studies such as Complex Adaptive System (CAS). Such approach is vital to develop ecosystem as an response to mitigate the Healthcare systemic complexities. Especially CAS studies makes it possible to integrate both macro aspects (such as epidemiology) related to Efficient Heathcare Management Outcomes ; and micro aspects (such as  Evidence Based Medicine and Pharmacogenomics that helps achieve medicine personalization) achieving Efficacy in the Healthcare delivery, to help achieve systemic integrity. In The QEXL Approach QEXL stands for “Quantum Exchange Language”, and Q-UEL is the initial proposed language. The QEXL Consortium embraces Quantal Semantics, Inc; (NC) and Ingine, Inc; (VA), and collaborates with The Dirac Foundation (UK), which has access to Professor Paul Dirac’s unpublished papers. The original consortium grew as a convergence of responses to four stimuli:

  1. The “re-emerging” interest in Artificial Intelligence (AI) as “computational thinking”, e.g. under the American Recovery Act;
  2. The President’s Council of Advisors on Science and Technology December 2010 call for an “XML-like” “Universal Exchange Language” (UEL) for healthcare;
  3. A desire to respond to the emerging Third World Wide Web (Semantic Web) by an initiative based on generalized probability theory  – the Thinking Web; and
  4. In the early courses of these  efforts, a greater understanding  of what Paul Dirac meant in his  Nobel Prize dinner speech where he stated that quantum mechanics should be applicable to all aspects of human thought.

The QEXL Approach

The QEXL Approach is developed based on considerable experiences in Expert Systems, linguistic theory, neurocognitive science, quantum mechanics, mathematical and physics-based approaches in Enterprise Architecture, Internet Topology, Filtering Theory, Semantic Web, Knowledge Lifecycle Management, and principles of Cloud Organization and Integration. The idea for well-formed probabilistic programming reasoning language is simple.  Importantly, also, the more essential features of it for reasoning and prediction are correspondingly simple such that the programmers are not necessarily humans, but structured and unstructured (text-analytic) “data mining” software robots. We have constructed a research prototype Inference Engine (IE) network (and more generally a program) that “simply” represents a basic Dirac notation and algebra compiler, with the caveat that it extends to Clifford-Dirac algebra; notably a Lorentz rotation of the imaginary number i (such that ii = -1) to the hyperbolic imaginary number h (such that hh = +1) corresponding to Dirac’s s, and gtime or g5) is applied.

[Outside the work of Dr. Barry Robson, this approach has not been tried in the inference and AI fields, with one highly suggestive exception: since the late 1990s it has occasionally been used in the neural network field by T. Nitta and others to solve the XOR problem in a single “neuron” and to reduce the number of “neurons” generally. Also suggestively, in particle physics it may be seen as a generalization of the Wick rotation time i x time used by Richard Feynman and others to render wave mechanics classical.  It retains the mathematical machinery and philosophy of Schrödinger’s wave mechanics but, instead of probability amplitudes as wave amplitudes, it yields classical but complex probability amplitudes encoding two directions of effect: “A acts on B, and B differently on A”. It maps to natural language where words relate to various types of real and imaginary scalar, vector, and matrix quantities. Dirac’s becomes the XML-like semantic triple . ]  

The QEXL Approach involves following  interdependent components.

  • Q-UEL (Probabilistic Inference + Phenomenon Of Interest): Addresses global issues that potentially pervade all human endeavors, and hence universal interoperability is of key importance
  •  (Inference Engine + Semantic Inferencing): Project addressing universal meaning underlying diverse natural languages on the Internet, and the use of that in knowledge representation
  • Inference Engine + Decentralized Infra: A link infrastructure for intra- and inter-cloud interoperability and integration in a coherent high level “metaware” environment. This component can also be explored to be replaced with simpler industry ready solutions such as MarkLogic® Enterprise NoSQL Database on Hadoop Distributed File System.

In an endeavor of this kind the partitions-of-work are inevitably artificial; it is important that this does not impede the integrity of optimal solutions.  The most important aspect in The QEXL Approach is, in essence where architecturally Probabilistic Inference (PI) and Data Architecture for the Inference Engine (IE)  is designed to be cooperative; software robots are created while PI and IE interact; and the inference knowledge gained by the PI and IE provide rules for solvers (robots) to self compile and conduct queries etc. This is therefore the grandeur of the scheme: This approach will have facilitated programming by nice compilers so that writing the inference network is easy, but it is not required to write the inference net as input code to compile, with the exception of reusable metarules as Dirac expressions with variables to process other rules by categorical and higher order logic. The robots are designed and programmed to do the remaining coding required to perform as solvers. So the notion of a compiler disappears under the hood. The robots are provided with well-formed instructions as well formed queries. Once inferences are formed, different “what – if” questions can be asked. Given that probability or that being the case, what is the chance of… and so on. It is as if having acquired knowledge, Phenomenon Of Interest (POI) is in a better state to explore what it means. Hyperbolic Dirac Networks (HDNs) are inference networks capable of overcoming the limitations imposed by Bayesian Nets (and statistics) and creating generative models richly expressing the “Phenomenon Of Interest” (POI) by the action of expressions containing binding variables. This may be thought of as an Expert System but analogous to Prolog data and Prolog programs that act upon the data, albeit here a “probabilistic Prolog”. Upfront should be stated the advantages over Bayes Nets as a commonly used inference method, but rather than compete with such methods the approach may be regarded as extending them. Indeed a Bayes Net as a static directed acyclic conditional probability graph is a subset of the Dirac Net as a static or dynamic general bidirectional graph with generalized logic and relationship operators, i.e. empowered by the mathematical machinery of Dirac’s quantum mechanics.

 The QEXL Approach Theory :- Robson Quantitative Semantics Algebra (RQSA)

Developed by Dr. Barry Robson

Theory :- The QEXL Approach based on Robson Quantitative Semantics Algebra – RQSA (Link to development of algorithm – overcoming limitations of Gold Stand Bayesian Network – to solve uncertainty while developing probabilistic ontology)

Impact Of The QEXL Approach

Impact of The QEXL Approach creating Probabilistic Ontology based on Clifford-Dirac algebra has immense opportunity in advancing the architecture to tackle large looming problems involving System of Systems; in which vast uncertain information emerge. Generally, as such systems are designed and developed employing Cartesian methods; such systems do not offer viable opportunity to deal with vast uncertain information when ridden with complexity. Especially when the context complexity poses multiple need for ontologies, and such a system inherently defies Cartesian methods. The QEXL Approach develops into an ecosystem response while it overcomes the Cartesian dilemma (link to another example for Cartesian Dilemma) and allows for generative models to emerge richly expressing the POI. The models generatively develops such that the POI behavior abstracted sufficiently lend for the IE and the Solvers to varieties of studies based on evidence and also allows for developing systemic studies pertaining to Complex Adaptive System and Complex Generative Systems afflicted by multiple cognitive challenges. Particularly, The QEXL Approach has potential to address complex challenges such as evidence based medicine (EBM); a mission that DoD’s Military Health System envisions while it modernizes its Electronics Health Record System – Veterans Health Information Systems and Technology Architecture (VistA). Vast potential also exists in addressing Veteran Administration’s (VA) Million Veteran Program (MVP); an effort by VA to consolidate genetic, military exposure, health, and lifestyle information together in one single database. By identifying gene-health connections, the program could consequentially advance disease screening, diagnosis, and prognosis and point the way toward more effective, personalized therapies.

Although The QEXL Approach is currently targeted to the healthcare and pharmaceutical domains where recognition of uncertainty is vital in observations, measurements and predictions, and probabilities underlying a variety of medical metrics, the scope of application is much more general. The QEXL Approach is to create a generic multivariate architecture for complex system characterized by Probabilistic Ontology that employing generative order will model “POI” facilitating creation of “communities of interest” by self-regulation in diverse domains of interest, requiring integrative of disciplines to create complex studies. The metaphor of “Cambrian Explosion” may aptly represent the enormity of the immense possibilities in advancing studies that tackle large systemic concerns riddled with uncertain information and random events that The QEXL Approach can stimulate.


The inference engine can be conceptualized into solutions such as MarkLogic NoSQL + Hadoop (HDFS).

It is interesting to note that in the genesis of evolving various NoSQL solutions based on Hadoop few insights have emerged related to need for designing the components recognizing their cooperative existence.

The Goal of The QEXL Approach: Is all about Contextualization 

The goal employing The QEXL Approach is to enable the realization of cognitive multivariate architecture for Probabilistic Ontology, advancing the Probabilistic Ontology based architecture for context specific application; such as Healthcare. Specifically, The QEXL Approach will develop PI  that helps in the creation of generative models that depicts the systemic behavior of the POI riddled with vast uncertain information. Generally, uncertainty in the vast information is introduced by the System of Systems complexity that is required to resolve multiples of ontologies, standards etc., these further introduce cognitive challenges. The further goal of The QEXL Approach is to overcome such challenges, by addressing interoperability at all levels, including the ability to communicate data and knowledge in a way that recognizes uncertainty in the world, so that automated PI and decision-making is possible. The aim is semiotic portability, i.e. the management of signs and symbols that deals especially with their function and interactions in both artificially constructed and natural languages. Existing systems for managing semantics and language are mostly systems of symbolic, not quantitative manipulation, with the primary exception of BayesOWL.  RQSA, or Robson Quantitative Semantic Algebra by its author Dr. Barry Robson, to distinguish it from other analogous systems, underlies Q-UEL. It is the development of (a) details of particular aspect of Dirac’s notation and algebra that is found to be of practical importance in generalizing and correctly normalizing Bayes Nets according to Bayes Theorem (i.e. controlling coherence, which ironically Bayes Nets usually neglect, as they are unidirectional), (b) merged with the treatment of probabilities and information based on finite data using the Riemann Zeta function that he has employed for many years in bioinformatics and data mining (, and (c) the extension to more flavors of hyperbolic imaginary number to encode intrinsic “dimensions of meaning” under a revised Rojet’s thesaurus system.

The Layers of the Architecture Created by The QEXL Approach

The QEXL Layered View

The QEXL Layered View

Layer 1- Contextualization: Planning, Designing driven by Theories 

A.    Probabilistic Ontology creating Inferencing leading into Evidence Based Medicine

i.     Aspects addressed by Q-UEL Tags and Kodaxil Inferencing

  1. Autonomy / Solidarity
  2. Inferencing (Kodaxil and Q – UEL)
  3. MetaData
  4. Security / Privacy
  5. Consented vs Un-consented Data
  6. Creating Incidence Rule (predicated – Q-UEL and Kodaxil)

ii.     Kodaxil:-  Enforcing Semantics across data sources (global text and data interoperability) – universal meaning underlying diverse natural languages on the Internet

iii.     Fluxology:- Logical Meta Data Cloud (A link infrastructure for intra- and inter-cloud interoperability and integration in a international setting)

  1. Adaptive
  2. Emergent Data Usage Patterns (networks of networks – enables by Probabilistic Ontology rules)
  3. Modeless Emergent Hierarchies
  4. Federation and Democratization Rule for Data (contract, trust, certificates, quality)

B.    Development of Probabilistic Model Representing Universal Abstraction of Phenomenon Of Interest

C.   Targeting Architecture to Application

  • Evidence Based Medicine
  • Genomics
  • Systemic Healthcare Studies
  • etc

Layer 2 – A: Operational Architecture (Logical )

A.    Reference Architecture

  1. Business Con Ops (Use cases)
  2. Conceptual Target Solution Architecture

Layer 2 – B: Data Management – Data Ingestion and Processing 

  1.  The processing of entries in the source data into form suitable for data mining
  2. The data mining of that processed data to obtain summary rules
  3. The capture of the appropriate released summary rules for inference

B.    Data Storage and Retrieval, Transactions

  1. Secure Storage and Retrieval
  2. Enable Secure Transactions
  3. Secure Data Exchange among several stake-holders and data owners

C.    Data Lifecycle, Data Organization Rules, Data Traceability to the Events, 

  1. Security and privacy by encryption and disaggregation of the EHR in a manner that is balanced against authorized access for extraction of global clinical and biomedical knowledge.
  2. Mechanisms for fine-grained consent permitting sharing and data mining.
  3. Mechanisms for secure alerting of patient or physician by backtrack when an authorized researcher or specialist notes that a patient is at risk.
  4. Structure and format that allows all meaningful use cases to be applied in reasonable time, including large-scale data mining.
  5. Assemblies across sources and data users forming contextual work patterns
  6. Hardened Security Framework

D.    Large EHR repository scaling

E.    Data Mining Rules

F.     Extracting and creating Incidence Rules

G.    Experimenting, observing and creating Semantic Inferences

H.    Visualization 

 The below two layers can be implemented in varieties of BigData platforms such as Hortonworks, Pivotal, Altiscale

Layer 3 – Application Layer (Schema-less for structured and unstructured Knowledge Repository – KRS)

Layer 4 – Infrastructure Architecture (Physical) (Hadoop and MapReduce for Large Data File-management and Processing; and Distributed / Concurrent Computations)

Big Data still in naivety proves Elusive to Big Picture and Randomness

Big Data as it exists today is not necessarily a tool that creates the Big Picture. Most systems that exists today are designed based on predicated methods. These rely on known variables, the only constraint unknown are the values of the variables. Predicated system are constrained to create newer insight as they are bounded by system boundaries defined by known variables. The new emergent variables can be considered as the non-causal agents until the newer order of the complex system begins to emerge; in which the newer variables manifests.

To develop wider and deeper insight about the behavior of the complex system as the system driven by entropy moves from one order to next; it does require to bring into its realm the possible newer variables. Lack of adequate techniques (Bayesian network included) in discerning all the variables that affect the system behavior creates uncertainty. Predicated system as such are blind sighted to such uncertainties. This creates inadequacies in methods when dealing with the randomness in the system. Emergence of newer variables that begin to affect the system behavior become the newer causes. Predicated system cannot hone on the variables that  lie dormant as the non-causal agents.

The hidden non-causal agents can be discovered by creating 2nd order generative semantic models. Predicative systems as such are not designed to dramatically alter its behavior; but only reflect the behavior constrained by those variables considered in its initial design