Space (mathematics)

From Wikipedia, the free encyclopedia

A hierarchy of mathematical spaces: The inner product induces a norm. The norm induces a metric. The metric induces a topology.

In mathematics, a space is a set (sometimes called a universe) with some added structure.

Mathematical spaces often form a hierarchy, i.e., one space may inherit all the characteristics of a parent space. For instance, all inner product spaces are also normed vector spaces, because the inner product induces a norm on the inner product space such that:

${\displaystyle \left\|x\right\|={\sqrt {\langle x,x\rangle }},}$

where the norm is indicated by enclosing in double vertical lines, and the inner product is indicated enclosing in by angle brackets.

Modern mathematics treats "space" quite differently compared to classical mathematics.

History

Before the golden age of geometry

In the ancient mathematics, "space" was a geometric abstraction of the three-dimensional space observed in everyday life. About 300 BC, Euclid gave axioms for the properties of space. Euclid built all of mathematics on these geometric foundations, going so far as to define numbers by comparing the lengths of line segments to the length of a chosen reference segment.

The method of coordinates (analytic geometry) was adopted by René Descartes in 1637.^[1] At that time, geometric theorems were treated as an absolute objective truth knowable through intuition and reason, similar to objects of natural science;^[2]:11 and axioms were treated as obvious implications of definitions.^[2]:15

Two equivalence relations between geometric figures were used: congruence and similarity. Translations, rotations and reflections transform a figure into congruent figures; homotheties — into similar figures. For example, all circles are mutually similar, but ellipses are not similar to circles. A third equivalence relation, introduced by Gaspard Monge in 1795, occurs in projective geometry: not only ellipses, but also parabolas and hyperbolas, turn into circles under appropriate projective transformations; they all are projectively equivalent figures.

The relation between the two geometries, Euclidean and projective,^[2]:133 shows that mathematical objects are not given to us with their structure.^[2]:21 Rather, each mathematical theory describes its objects by some of their properties, precisely those that are put as axioms at the foundations of the theory.^[2]:20

Distances and angles are never mentioned in the axioms of the projective geometry and therefore cannot appear in its theorems. The question "what is the sum of the three angles of a triangle" is meaningful in the Euclidean geometry but meaningless in projective geometry.

A different situation appeared in the 19th century: in some geometries the sum of the three angles of a triangle is well-defined but different from the classical value (180 degrees). The non-Euclidean hyperbolic geometry, introduced by Nikolai Lobachevsky in 1829 and János Bolyai in 1832 (and Carl Gauss in 1816, unpublished)^[2]:133 stated that the sum depends on the triangle and is always less than 180 degrees. Eugenio Beltrami in 1868 and Felix Klein in 1871 obtained Euclidean "models" of the non-Euclidean hyperbolic geometry, and thereby completely justified this theory.^[2]:24[3]

This discovery forced the abandonment of the pretensions to the absolute truth of Euclidean geometry. It showed that axioms are not "obvious", nor "implications of definitions". Rather, they are hypotheses. To what extent do they correspond to an experimental reality? This important physical problem no longer has anything to do with mathematics. Even if a "geometry" does not correspond to an experimental reality, its theorems remain no less "mathematical truths".^[2]:15

A Euclidean model of a non-Euclidean geometry is a clever choice of some objects existing in Euclidean space and some relations between these objects that satisfy all axioms (therefore, all theorems) of the non-Euclidean geometry. These Euclidean objects and relations "play" the non-Euclidean geometry like contemporary actors playing an ancient performance. Relations between the actors only mimic relations between the characters in the play. Likewise, the chosen relations between the chosen objects of the Euclidean model only mimic the non-Euclidean relations. It shows that relations between objects are essential in mathematics, while the nature of the objects is not.

The golden age and afterwards: dramatic change

According to Nicolas Bourbaki,^[2]:131 the period between 1795 (Geometrie descriptive of Monge) and 1872 (the "Erlangen programme" of Klein) can be called the golden age of geometry. The original space investigated by Euclid is now called three-dimensional Euclidean space. Its axiomatization, started by Euclid 23 centuries ago, was reformed with Hilbert's axioms, Tarski's axioms and Birkhoff's axioms. These axiom systems describe space via primitive notions (such as "point", "between", "congruent") constrained by a number of axioms.

Analytic geometry made great progress, and it succeeded in replacing theorems of classical geometry with computations via invariants of transformation groups.^[2]:134,5 This technique applies simultaneously to many geometries. A definition of space "from scratch", as in Euclid, is now not often used since it does not reveal the relation of this space to other spaces. Since this time, new theorems of classical geometry are of more interest to amateurs rather than to professional mathematicians.^[2]:136 However, the heritage of the classical geometry was not lost. According to Bourbaki,^[2]:138 "passed over in its role as an autonomous and living science, classical geometry is thus transfigured into a universal language of contemporary mathematics".

Simultaneously, numbers began to displace geometry as the foundation of mathematics. For instance, in Richard Dedekind's 1872 essay Stetigkeit und irrationale Zahlen (Continuity and irrational numbers), he asserts that points on a line ought to have the properties of Dedekind cuts, and that therefore a line was the same thing as the set of real numbers. Dedekind is careful to note that this is an assumption that is incapable of being proven. In modern treatments, Dedekind's assertion is often taken to be the definition of a line, thereby reducing geometry to arithmetic. Three-dimensional Euclidean space is defined to be an affine space whose associated vector space of differences of its elements is equipped with an inner product.^[4] Also, a three-dimensional projective space is now defined non-classically, as the space of all one-dimensional subspaces (that is, straight lines through the origin) of a four-dimensional vector space. This shift in foundations requires a new set of axioms, and if these axioms are adopted, the classical axioms for geometry become theorems.

According to the famous inaugural lecture given by Bernhard Riemann in 1854, every mathematical object parametrized by ${\displaystyle n}$ real numbers may be treated as a point of the ${\displaystyle n}$-dimensional space of all such objects.^[2]:140 Contemporary mathematicians follow this idea routinely and find it extremely suggestive to use the terminology of classical geometry nearly everywhere.^[2]:138

An object parametrized by n complex numbers may be treated as a point of a complex n-dimensional space. However, the same object is also parametrized by 2n real numbers (if c is a complex number, then c = a + bi, where a and b are real), thus, a point of a real 2n-dimensional space. The complex dimension differs from the real dimension. This is only the tip of the iceberg. The "algebraic" concept of dimension applies to vector spaces. For topological spaces there are several dimension concepts including inductive dimension and Hausdorff dimension, which can be non-integer (especially for fractals). Some kinds of spaces (for instance, measure spaces) admit no concept of dimension at all.

Functions are important mathematical objects. Usually they form infinite-dimensional function spaces, as noted already by Riemann^[2]:141 and elaborated in the 20th century by functional analysis.

In order to fully appreciate the generality of this approach one should note that mathematics is "a pure theory of forms, which has as its purpose, not the combination of quantities, or of their images, the numbers, but objects of thought" (Hermann Hankel, 1867).^[2]:21 This is a controversial characterization of the purpose of mathematics, which is not necessarily committed to the existence of "objects of thought".

A space consists now of selected mathematical objects (for instance, functions on another space, or subspaces of another space, or just elements of a set) treated as points, and selected relationships between these points. It shows that spaces are just mathematical structures of convenience. One may expect that the structures called "spaces" are more geometric than others, but this is not always true. For example, a differentiable manifold (called also smooth manifold) is much more geometric than a measurable space, but no one calls it "differentiable space" (nor "smooth space").

Taxonomy of spaces

Three taxonomic ranks

Spaces are classified on three levels. Given that each mathematical theory describes its objects by some of their properties, the first question to ask is: which properties?

For example, the upper-level classification distinguishes between Euclidean and projective spaces, since the distance between two points is defined in Euclidean spaces but undefined in projective spaces. These are spaces of different types.

Another example. The question "what is the sum of the three angles of a triangle" makes sense in a Euclidean space but not in a projective space; these are spaces of different types. In a non-Euclidean space the question makes sense but is answered differently, which is not an upper-level distinction.

Also, the distinction between a Euclidean plane and a Euclidean 3-dimensional space is not an upper-level distinction; the question "what is the dimension" makes sense in both cases.

In terms of Bourbaki^[5] the upper-level classification is related to "typical characterization" (or "typification"). However, it is not the same (since two equivalent structures may differ in typification).

On the second level of classification one takes into account answers to especially important questions (among the questions that make sense according to the first level). For example, this level distinguishes between Euclidean and non-Euclidean spaces; between finite-dimensional and infinite-dimensional spaces; between compact and non-compact spaces, etc.

In terms of Bourbaki^[5] the second-level classification is the classification by "species". Unlike biological taxonomy, a space may belong to several species.

On the third level of classification, roughly speaking, one takes into account answers to all possible questions (that make sense according to the first level). For example, this level distinguishes between spaces of different dimension, but does not distinguish between a plane of a three-dimensional Euclidean space, treated as a two-dimensional Euclidean space, and the set of all pairs of real numbers, also treated as a two-dimensional Euclidean space. Likewise it does not distinguish between different Euclidean models of the same non-Euclidean space.

More formally, the third level classifies spaces up to isomorphism. An isomorphism between two spaces is defined as a one-to-one correspondence between the points of the first space and the points of the second space, that preserves all relations between the points, stipulated by the given "typification". Mutually isomorphic spaces are thought of as copies of a single space. If one of them belongs to a given species then they all do.

The notion of isomorphism sheds light on the upper-level classification. Given a one-to-one correspondence between two spaces of the same type, one may ask whether it is an isomorphism or not. This question makes no sense for two spaces of different type.

Isomorphisms to itself are called automorphisms. Automorphisms of a Euclidean space are motions and reflections. Euclidean space is homogeneous in the sense that every point can be transformed into every other point by some automorphism.

Two relations between spaces, and a property of spaces

Topological notions (continuity, convergence, open sets, closed sets etc.) are defined naturally in every Euclidean space. In other words, every Euclidean space is also a topological space. Every isomorphism between two Euclidean spaces is also an isomorphism between the corresponding topological spaces (called "homeomorphism"), but the converse is wrong: a homeomorphism may distort distances. In terms of Bourbaki,^[5] "topological space" is an underlying structure of the "Euclidean space" structure. Similar ideas occur in category theory: the category of Euclidean spaces is a concrete category over the category of topological spaces; the forgetful (or "stripping") functor maps the former category to the latter category.

A three-dimensional Euclidean space is a special case of a Euclidean space. In terms of Bourbaki,^[5] the species of three-dimensional Euclidean space is richer than the species of Euclidean space. Likewise, the species of compact topological space is richer than the species of topological space.

Euclidean axioms leave no freedom, they determine uniquely all geometric properties of the space. More exactly: all three-dimensional Euclidean spaces are isomorphic. In this sense we have "the" three-dimensional Euclidean space. In terms of Bourbaki, the corresponding theory is univalent. In contrast, topological spaces are generally non-isomorphic, their theory is multivalent. A similar idea occurs in mathematical logic: a theory is called categorical if all its models of the same cardinality are isomorphic. According to Bourbaki,^[6] the study of multivalent theories is the most striking feature which distinguishes modern mathematics from classical mathematics.

Types of spaces

Overview of types of abstract spaces. An arrow from space A to space B implies that space A is also a kind of space B. That means, for instance, that a normed vector space is also a metric space.

Linear and topological spaces

Two basic spaces are linear spaces (also called vector spaces) and topological spaces.

Linear spaces are of algebraic nature; there are real linear spaces (over the field of real numbers), complex linear spaces (over the field of complex numbers), and more generally, linear spaces over any field. Every complex linear space is also a real linear space (the latter underlies the former), since each real number is also a complex number.^{[details 1]} Linear operations, given in a linear space by definition, lead to such notions as straight lines (and planes, and other linear subspaces); parallel lines; ellipses (and ellipsoids). However, orthogonal (perpendicular) lines cannot be defined, and circles cannot be singled out among ellipses. The dimension of a linear space is defined as the maximal number of linearly independent vectors or, equivalently, as the minimal number of vectors that span the space; it may be finite or infinite. Two linear spaces over the same field are isomorphic if and only if they are of the same dimension.

Topological spaces are of analytic nature. Open sets, given in a topological space by definition, lead to such notions as continuous functions, paths, maps; convergent sequences, limits; interior, boundary, exterior. However, uniform continuity, bounded sets, Cauchy sequences, differentiable functions (paths, maps) remain undefined. Isomorphisms between topological spaces are traditionally called homeomorphisms; these are one-to-one correspondences continuous in both directions. The open interval ${\displaystyle (0,1)}$ is homeomorphic to the whole real line ${\displaystyle (-\infty ,\infty )}$ but not homeomorphic to the closed interval ${\displaystyle [0,1]}$, nor to a circle. The surface of a cube is homeomorphic to a sphere (the surface of a ball) but not homeomorphic to a torus. Euclidean spaces of different dimensions are not homeomorphic, which seems evident, but is not easy to prove. Dimension of a topological space is difficult to define; "inductive dimension" and "Lebesgue covering dimension" are used. Every subset of a topological space is itself a topological space (in contrast, only linear subsets of a linear space are linear spaces). Arbitrary topological spaces, investigated by general topology (called also point-set topology) are too diverse for a complete classification (up to homeomorphism). They are inhomogeneous (in general). Compact topological spaces are an important class of topological spaces ("species" of this "type"). Every continuous function is bounded on such space. The closed interval ${\displaystyle [0,1]}$ and the extended real line ${\displaystyle [-\infty ,\infty ]}$ are compact; the open interval ${\displaystyle (0,1)}$ and the line ${\displaystyle (-\infty ,\infty )}$ are not. Geometric topology investigates manifolds (another "species" of this "type"); these are topological spaces locally homeomorphic to Euclidean spaces. Low-dimensional manifolds are completely classified (up to homeomorphism).

The two structures discussed above (linear and topological) are both underlying structures of the "linear topological space" structure. That is, a linear topological space is both a linear (real or complex) space and a (homogeneous, in fact) topological space. However, an arbitrary combination of these two structures is generally not a linear topological space; the two structures must conform, namely, the linear operations must be continuous.

Every finite-dimensional (real or complex) linear space is a linear topological space in the sense that it carries one and only one topology that makes it a linear topological space. The two structures, "finite-dimensional (real or complex) linear space" and "finite-dimensional linear topological space", are thus equivalent, that is, mutually underlying. Accordingly, every invertible linear transformation of a finite-dimensional linear topological space is a homeomorphism. In the infinite dimension, however, different topologies conform to a given linear structure, and invertible linear transformations are generally not homeomorphisms.

Affine and projective spaces

It is convenient to introduce affine and projective spaces by means of linear spaces, as follows. An ${\displaystyle n}$-dimensional linear subspace of an ${\displaystyle (n+1)}$-dimensional linear space, being itself an ${\displaystyle n}$-dimensional linear space, is not homogeneous; it contains a special point, the origin. Shifting it by a vector external to it, one obtains an ${\displaystyle n}$-dimensional affine space. It is homogeneous. In the words of John Baez, "an affine space is a vector space that's forgotten its origin". A straight line in the affine space is, by definition, its intersection with a two-dimensional linear subspace (plane through the origin) of the ${\displaystyle (n+1)}$-dimensional linear space. Every linear space is also an affine space.

Every point of the affine space is its intersection with a one-dimensional linear subspace (line through the origin) of the ${\displaystyle (n+1)}$-dimensional linear space. However, some one-dimensional subspaces are parallel to the affine space; in some sense, they intersect it at infinity. The set of all one-dimensional linear subspaces of an ${\displaystyle (n+1)}$-dimensional linear space is, by definition, an ${\displaystyle n}$-dimensional projective space. Choosing an ${\displaystyle n}$-dimensional affine space as before one observes that the affine space is embedded as a proper subset into the projective space. However, the projective space itself is homogeneous. A straight line in the projective space, by definition, corresponds to a two-dimensional linear subspace of the ${\displaystyle (n+1)}$-dimensional linear space.

Defined this way, affine and projective spaces are of algebraic nature; they can be real, complex, and more generally, over any field.

Every real (or complex) affine or projective space is also a topological space. An affine space is a non-compact manifold; a projective space is a compact manifold.

Metric and uniform spaces

Distances between points are defined in a metric space. Every metric space is also a topological space. Bounded sets and Cauchy sequences are defined in a metric space (but not just in a topological space). Isomorphisms between metric spaces are called isometries. A metric space is called complete if all Cauchy sequences converge. Every incomplete space is isometrically embedded into its completion. Every compact metric space is complete; the real line is non-compact but complete; the open interval ${\displaystyle (0,1)}$ is incomplete.

A topological space is called metrizable, if it underlies a metric space. All manifolds are metrizable.

Every Euclidean space is also a complete metric space. Moreover, all geometric notions immanent to a Euclidean space can be characterized in terms of its metric. For example, the straight segment connecting two given points ${\displaystyle A}$ and ${\displaystyle C}$ consists of all points ${\displaystyle B}$ such that the distance between ${\displaystyle A}$ and ${\displaystyle C}$ is equal to the sum of two distances, between ${\displaystyle A}$ and ${\displaystyle B}$ and between ${\displaystyle B}$ and ${\displaystyle C}$.

Uniform spaces do not introduce distances, but still allow one to use uniform continuity, Cauchy sequences, completeness and completion. Every uniform space is also a topological space. Every linear topological space (metrizable or not) is also a uniform space. More generally, every commutative topological group is also a uniform space. A non-commutative topological group, however, carries two uniform structures, one left-invariant, the other right-invariant. Linear topological spaces are complete in finite dimension but generally incomplete in infinite dimension.

Normed, Banach, inner product, and Hilbert spaces

Vectors in a Euclidean space are a linear space, but each vector ${\displaystyle x}$ has also a length, in other words, norm, ${\displaystyle \|x\|}$. A (real or complex) linear space endowed with a norm is a normed space. Every normed space is both a linear topological space and a metric space. A Banach space is a complete normed space. Many spaces of sequences or functions are infinite-dimensional Banach spaces.

The set of all vectors of norm less than one is called the unit ball of a normed space. It is a convex, centrally symmetric set, generally not an ellipsoid; for example, it may be a polygon (on the plane). The parallelogram law (called also parallelogram identity) ${\displaystyle \|x-y\|^{2}+\|x+y\|^{2}=2\|x\|^{2}+2\|y\|^{2}}$ generally fails in normed spaces, but holds for vectors in Euclidean spaces, which follows from the fact that the squared Euclidean norm of a vector is its inner product to itself.

An inner product space is a (real or complex) linear space endowed with a bilinear (or sesquilinear) form satisfying some conditions and called inner product. Every inner product space is also a normed space. A normed space underlies an inner product space if and only if it satisfies the parallelogram law, or equivalently, if its unit ball is an ellipsoid. Angles between vectors are defined in inner product spaces. A Hilbert space is defined as a complete inner product space. (Some authors insist that it must be complex, others admit also real Hilbert spaces.) Many spaces of sequences or functions are infinite-dimensional Hilbert spaces. Hilbert spaces are very important for quantum theory.^[7]

All ${\displaystyle n}$-dimensional real inner product spaces are mutually isomorphic. One may say that the ${\displaystyle n}$-dimensional Euclidean space is the ${\displaystyle n}$-dimensional real inner product space that's forgotten its origin.

Smooth and Riemannian manifolds (spaces)

Smooth manifolds are not called "spaces", but could be. Smooth (differentiable) functions, paths, maps, given in a smooth manifold by definition, lead to tangent spaces. Every smooth manifold is a (topological) manifold. Smooth surfaces in a finite-dimensional linear space (like the surface of an ellipsoid, not a polytope) are smooth manifolds. Every smooth manifold can be embedded into a finite-dimensional linear space. A smooth path in a smooth manifold has (at every point) the tangent vector, belonging to the tangent space (attached to this point). Tangent spaces to an ${\displaystyle n}$-dimensional smooth manifold are ${\displaystyle n}$-dimensional linear spaces. A smooth function has (at every point) the differential, – a linear functional on the tangent space. Real (or complex) finite-dimensional linear, affine and projective spaces are also smooth manifolds.

A Riemannian manifold, or Riemann space, is a smooth manifold whose tangent spaces are endowed with inner product (satisfying some conditions). Euclidean spaces are also Riemann spaces. Smooth surfaces in Euclidean spaces are Riemann spaces. A hyperbolic non-Euclidean space is also a Riemann space. A curve in a Riemann space has the length. A Riemann space is both a smooth manifold and a metric space; the length of the shortest curve is the distance. The angle between two curves intersecting at a point is the angle between their tangent lines.

Waiving positivity of inner product on tangent spaces one gets pseudo-Riemann (especially, Lorentzian) spaces very important for general relativity.

Measurable, measure, and probability spaces

Waiving distances and angles while retaining volumes (of geometric bodies) one moves toward measure theory. Besides the volume, a measure generalizes area, length, mass (or charge) distribution, and also probability distribution, according to Andrey Kolmogorov's approach to probability theory.

A "geometric body" of classical mathematics is much more regular than just a set of points. The boundary of the body is of zero volume. Thus, the volume of the body is the volume of its interior, and the interior can be exhausted by an infinite sequence of cubes. In contrast, the boundary of an arbitrary set of points can be of non-zero volume (an example: the set of all rational points inside a given cube). Measure theory succeeded in extending the notion of volume (or another measure) to a vast class of sets, so-called measurable sets. Indeed, non-measurable sets almost never occur in applications, but anyway, the theory must restrict itself to measurable sets (and functions).

Measurable sets, given in a measurable space by definition, lead to measurable functions and maps. In order to turn a topological space into a measurable space one endows it with a σ-algebra. The σ-algebra of Borel sets is most popular, but not the only choice (Baire sets, universally measurable sets etc. are used sometimes). Alternatively, a σ-algebra can be generated by a given collection of sets (or functions) irrespective of any topology. Quite often, different topologies lead to the same σ-algebra (for example, the norm topology and the weak topology on a separable Hilbert space). Every subset of a measurable space is itself a measurable space.

Standard measurable spaces (called also standard Borel spaces) are especially useful. Every Borel set (in particular, every closed set and every open set) in a Euclidean space (and more generally, in a complete separable metric space) is a standard measurable space. All uncountable standard measurable spaces are mutually isomorphic.

A measure space is a measurable space endowed with a measure. A Euclidean space with Lebesgue measure is a measure space. Integration theory defines integrability and integrals of measurable functions on a measure space.

Sets of measure 0, called null sets, are negligible. Accordingly, a ${\displaystyle {\bmod {0}}}$ isomorphism is defined as isomorphism between subsets of full measure (that is, with negligible complement).

A probability space is a measure space such that the measure of the whole space is equal to 1. The product of any family (finite or not) of probability spaces is a probability space. In contrast, for measure spaces in general, only the product of finitely many spaces is defined. Accordingly, there are many infinite-dimensional probability measures (especially, Gaussian measures), but no infinite-dimensional Lebesgue measure.

Standard probability spaces are especially useful. Every probability measure on a standard measurable space leads to a standard probability space. The product of a sequence (finite or not) of standard probability spaces is a standard probability space. All non-atomic standard probability spaces are mutually isomorphic ${\displaystyle {\bmod {0}};}$ one of them is the interval ${\displaystyle (0,1)}$ with Lebesgue measure.

These spaces are less geometric. In particular, the idea of dimension, applicable (in one form or another) to all other spaces, does not apply to measurable, measure and probability spaces.

A topological space becomes also a measurable space when endowed with the Borel σ-algebra.^{[details 2]} However, the topology is not uniquely determined by its Borel σ-algebra; and not every σ-algebra is the Borel σ-algebra of some topology.^{[details 3]}

Non-commutative geometry

The theoretical study of calculus, known as mathematical analysis, led in the early 20th century to the consideration of linear spaces of real-valued or complex-valued functions. The earliest examples of these were function spaces, each one adapted to its own class of problems. These examples shared many common features, and these features were soon abstracted into Hilbert spaces, Banach spaces, and more general topological vector spaces. These were a powerful toolkit for the solution of a wide range of mathematical problems.

The most detailed information was carried by a class of spaces called Banach algebras. These are Banach spaces together with a continuous multiplication operation. An important early example was the Banach algebra of essentially bounded measurable functions on a measure space X. This set of functions is a Banach space under pointwise addition and scalar multiplication. With the operation of pointwise multiplication, it becomes a special type of Banach space, one now called a commutative von Neumann algebra. Pointwise multiplication determines a representation of this algebra on the Hilbert space of square integrable functions on X. An early observation of von Neumann was that this correspondence also worked in reverse: Given some mild technical hypotheses, a commutative von Neumann algebra together with a representation on a Hilbert space determines a measure space, and these two constructions (of a von Neumann algebra plus a representation and of a measure space) were mutually inverse.

von Neumann then proposed that non-commutative von Neumann algebras should have geometric meaning, just as commutative von Neumann algebras do. Together with Francis Murray, he produced a classification of von Neumann algebras. The direct integral construction shows how to break any von Neumann algebra into a collection of simpler algebras called factors. von Neumann and Murray classified factors into three types. Type I was nearly identical to the commutative case. Types II and III exhibited new phenomena. A type II von Neumann algebra determined a geometry with the peculiar feature that the dimension could be any non-negative real number, not just an integer. Type III algebras were those that were neither types I nor II, and after several decades of effort, these were proven to be closely related to type II factors.

A slightly different approach to the geometry of function spaces developed at the same time as von Neumann and Murray's work on the classification of factors. This approach is the theory of C*-algebras. Here, the motivating example is the C*-algebra ${\displaystyle C_{0}(X)}$, where X is a locally compact Hausdorff topological space. By definition, this is the algebra of continuous complex-valued functions on X that vanish at infinity (which loosely means that the farther you go from a chosen point, the closer the function gets to zero) with the operations of pointwise addition and multiplication. The Gelfand–Naimark theorem implied that there is a correspondence between commutative C*-algebras and geometric objects: Every commutative C*-algebra is of the form ${\displaystyle C_{0}(X)}$ for some locally compact Hausdorff X. The non-commutative C*-algebras, therefore, can be interpreted as non-commutative spaces, much like non-commutative von Neumann algebras.

Both of these examples are now cases of a field called non-commutative geometry. The specific examples of von Neumann algebras and C*-algebras are known as non-commutative measure theory and non-commutative topology, respectively. Non-commutative geometry is not merely a pursuit of generality for its own sake and is not just a curiosity. Non-commutative spaces arise naturally, even inevitably, from some constructions. For example, consider the non-periodic Penrose tilings of the plane by kites and darts. It is a theorem that, in such a tiling, every finite patch of kites and darts appears infinitely often. As a consequence, there is no way to distinguish two Penrose tilings by looking at a finite portion. This makes it impossible to assign the set of all tilings a topology in the traditional sense. Despite this, the Penrose tilings determine a non-commutative C*-algebra, and consequently they can be studied by the techniques of non-commutative geometry. Another example, and one of great interest within differential geometry, comes from foliations of manifolds. These are ways of splitting the manifold up into smaller-dimensional submanifolds called leaves, each of which is locally parallel to others nearby. The set of all leaves can be made into a topological space. However, the example of an irrational rotation shows that this topological space can be bizarre and the techniques of classical measure theory may be useless on it. However, there is a non-commutative von Neumann algebra associated to the leaf space of a foliation, and once again, this gives an otherwise unintelligible space a good geometric structure.

Schemes

Algebraic geometry studies the geometric properties of polynomial equations. Polynomials are a type of function defined by the basic arithmetic operations of addition and multiplication. Because of this, they are closely tied to algebra. Algebraic geometry offers a way to apply geometric techniques to questions of pure algebra, and vice versa.

The type of space that underlies most modern algebraic geometry was introduced by Alexander Grothendieck and is called a scheme. One of the building blocks of a scheme is a topological space. Topological spaces have continuous functions, but continuous functions are too general to reflect the underlying algebraic structure of interest. The other ingredient in a scheme, therefore, is a sheaf on the topological space, called the "structure sheaf". On each open subset of the topological space, the sheaf specifies a collection of functions, called "regular functions". The topological space and the structure sheaf together are required to satisfy conditions that mean the functions come from algebraic operations.

Like manifolds, schemes are defined as spaces which are locally modeled on a familiar space. In the case of manifolds, the familiar space is Euclidean space. For a scheme, the local models are called affine schemes. Affine schemes provide a direct link between algebraic geometry and commutative algebra. The fundamental objects of study in commutative algebra are commutative rings. If R is a commutative ring, then there is a corresponding affine scheme ${\displaystyle \operatorname {Spec} R}$ which translates the algebraic structure of R into geometry. Conversely, every affine scheme determines a commutative ring, the global sections of its structure sheaf. These two operations are mutually inverse, so affine schemes provide a new language with which to study questions in commutative algebra. By definition, every point in a scheme has an open neighborhood which is an affine scheme.

There are many schemes which are not affine. Often, this is an unavoidable consequence of the geometry of the scheme. The most important example of this is projective space. Projective space satisfies a condition called properness which is analogous to compactness. Affine schemes cannot be proper (except in trivial situations like when the scheme has only a single point), and hence no projective space is an affine scheme (except for zero-dimensional projective spaces). Projective space is closely related to the theory of perspective and to homogeneous polynomials. Projective schemes, meaning those that arise as closed subschemes of a projective space, are the single most important family of schemes.^[8]

Several generalizations of schemes have been introduced. Michael Artin defined an algebraic space to be an object which is the quotient of a scheme by certain types of equivalence relations, specifically, equivalence relations which define étale morphisms. Algebraic spaces retain many of the useful properties of schemes while simultaneously being more flexible. For instance, the Keel–Mori theorem can be used to show that many moduli spaces are algebraic spaces.

More general than an algebraic space is a Deligne–Mumford stack. DM stacks are similar to schemes, but they permit singularities that cannot be described solely in terms of polynomials. They play the same role for schemes that orbifolds do for manifolds. For example, the quotient of the affine plane by a finite group of rotations around the origin yields a Deligne–Mumford stack that is not a scheme or an algebraic space. Away from the origin, the quotient by the group action identifies finite sets of equally spaced points on a circle. But at the origin, the circle consists of only a single point, the origin itself, and the group action fixes this point. In the quotient DM stack, however, this point comes with the extra data of being a quotient. This kind of refined structure is useful in the theory of moduli spaces, and in fact, it was originally introduced to describe moduli of algebraic curves.

A yet further generalization are the algebraic stacks, also called Artin stacks. DM stacks are limited to quotients by finite group actions. While this suffices for many problems in moduli theory, it is too restrictive for others. Artin stacks permit more general quotients, and hence more moduli problems can be treated using Artin stacks than DM stacks.

Topoi

In Grothendieck's work on the Weil conjectures, he introduced a new type of topology now called a Grothendieck topology. A topological space (in the ordinary sense) axiomatizes the notion of "nearness," making two points be nearby if and only if they lie in many of the same open sets. By contrast, a Grothendieck topology axiomatizes the notion of "covering." A covering of a space is a collection of subspaces that jointly contain all the information of the ambient space. Since sheaves are defined in terms of coverings, a Grothendieck topology can also be seen as an axiomatization of the theory of sheaves.

Grothendieck's work on his topologies led him to the theory of topoi, which he believed to be one of his greatest mathematical ideas. A sheaf (either on a topological space or with respect to a Grothendieck topology) is used to express local data. The category of all sheaves carries all possible ways of expressing local data. Since topological spaces are constructed from points, which are themselves a kind of local data, the category of sheaves can therefore be used as a replacement for the original space. Grothendieck consequently defined a topos to be a category of sheaves and studied topoi as objects of interest in their own right. These are now called Grothendieck topoi.

Every topological space determines a topos, and vice versa. There are topological spaces where taking the associated topos loses information, but these are generally considered pathological. (A necessary and sufficient condition is that the topological space be a sober space.) Conversely, there are topoi whose associated topological spaces do not capture the original topos. But, far from being pathological, these topoi can be of great mathematical interest. For instance, Grothendieck's theory of étale cohomology (which eventually led to the proof of the Weil conjectures) can be phrased as cohomology in the étale topos of a scheme, and this topos does not come from a topological space.

Topological spaces in fact lead to very special topoi called locales. The set of open subsets of a topological space determines a lattice. The axioms for a topological space cause these lattices to be complete Heyting algebras. The theory of locales takes this as its starting point. A locale is defined to be a complete Heyting algebra, and the elementary properties of topological spaces are re-expressed and reproved in these terms. The concept of a locale turns out to be more general than a topological space, in that every sober topological space determines a unique locale, but many interesting locales do not come from topological spaces. Because locales need not have points, the study of locales is somewhat jokingly called pointless topology.

Topoi also display deep connections to mathematical logic. Every Grothendieck topos has a special sheaf called a subobject classifier. This subobject classifier functions like the set of all possible truth values. In the topos of sets, the subobject classifier is the set ${\displaystyle \{0,1\}}$, corresponding to "False" and "True". But in other topoi, the subobject classifier can be much more complicated. Lawvere and Tierney recognized that axiomatizing the subobject classifier yielded a more general kind of topos, now known as an elementary topos, and that elementary topoi were models of intuitionistic logic. In addition to providing a powerful way to apply tools from logic to geometry, this made possible the use of geometric methods in logic.

Publication bias

From Wikipedia, the free encyclopedia

Publication bias is a type of bias that occurs in published academic research. It occurs when the outcome of an experiment or research study influences the decision whether to publish or otherwise distribute it. Publication bias matters because literature reviews regarding support for a hypothesis can be biased if the original literature is contaminated by publication bias.^[1] Publishing only results that show a significant finding disturbs the balance of findings.^[2]

Studies with significant results can be of the same standard as studies with a null result with respect to quality of execution and design.^[3] However, statistically significant results are three times more likely to be published than papers with null results.^[4]

Multiple factors contribute to publication bias.^[1] For instance, once a scientific finding is well established, it may become newsworthy to publish reliable papers that fail to reject the null hypothesis.^[5] It has been found that the most common reason for non-publication is simply that investigators decline to submit results, leading to non-response bias. Factors cited as underlying this effect include investigators assuming they must have made a mistake, failure to support a known finding, loss of interest in the topic, or anticipation that others will be uninterested in the null results.^[3] The nature of these issues and the problems that have been triggered, have been referred to as the 5 diseases that threaten science, which include: "significosis, an inordinate focus on statistically significant results; neophilia, an excessive appreciation for novelty; theorrhea, a mania for new theory; arigorium, a deficiency of rigor in theoretical and empirical work; and finally, disjunctivitis, a proclivity to produce large quantities of redundant, trivial, and incoherent works." ^[6]

Attempts to identify unpublished studies often prove difficult or are unsatisfactory.^[1] In an effort to combat this problem, some journals require that studies submitted for publication are pre-registered (registering a study prior to collection of data and analysis) with organizations like the Center for Open Science.

Other proposed strategies to detect and control for publication bias^[1] include p-curve analysis ^[7] and disfavoring small and non-randomised studies because of their demonstrated high susceptibility to error and bias.^[3]

Definition

Publication bias occurs when the publication of research results depends not just on the quality of the research but also on the hypothesis tested, and the significance and direction of effects detected.^[8] The term was first used in 1959 by statistician Theodore Sterling to refer to fields in which "successful" research is more likely to be published. As a result, "the literature of such a field consists in substantial part of false conclusions resulting from type-I errors".^[9]

Publication bias is sometimes called the "file drawer effect," or "file drawer problem." This term suggests that results not supporting the hypotheses of researchers often go no further than the researchers' file drawers, leading to a bias in published research.^[10] The term "file drawer problem" was coined by Rosenthal in 1979.^[11]

Positive-results bias, a type of publication bias, occurs when authors are more likely to submit, or editors are more likely to accept, positive results than negative or inconclusive results.^[12] Outcome reporting bias occurs when multiple outcomes are measured and analyzed, but the reporting of these outcomes is dependent on the strength and direction of its results. A generic term coined to describe these post-hoc choices is HARKing ("Hypothesizing After the Results are Known").^[13]

Evidence

Meta-analysis of stereotype threat on girls' math scores showing asymmetry typical of publication bias. From Flore, P. C., & Wicherts, J. M. (2015)^[14]

The presence of publication bias in the literature has been most extensively studied in biomedical research. Investigators following clinical trials from the submission of their protocols to ethics committees (or regulatory authorities) until the publication of their results observed that those with positive results are more likely to be published.^[15][16][17] In addition, studies often fail to report negative results when published, as demonstrated by research comparing study protocols with published articles.^[18][19]

The presence of publication bias was investigated in meta-analyses. The largest such analysis investigated the presence of publication bias in systematic reviews of medical treatments from the Cochrane Library.^[20] The study showed that statistically positive significant findings are 27% more likely to be included in meta-analyses of efficacy than other findings. Results showing no evidence of adverse effects have a 78% greater probability of inclusion in safety studies than statistically significant results showing adverse effects. Evidence of publication bias was found in meta-analyses published in prominent medical journals.^[21]

Impact on meta-analysis

Where publication bias is present, published studies are no longer a representative sample of the available evidence. This bias distorts the results of meta-analyses and systematic reviews. For example, evidence-based medicine is increasingly reliant on meta-analysis to assess evidence.

Meta-analyses and systematic reviews can account for publication bias by including evidence from unpublished studies and the grey literature. The presence of publication bias can also be explored by constructing a funnel plot in which the estimate of the reported effect size is plotted against a measure of precision or sample size. The premise is that the scatter of points should reflect a funnel shape, indicating that the reporting of effect sizes is not related to their statistical significance.^[22] However, when small studies are predominately in one direction (usually the direction of larger effect sizes), asymmetry will ensue and this may be indicative of publication bias.^[23]

Because an inevitable degree of subjectivity exists in the interpretation of funnel plots, several tests have been proposed for detecting funnel plot asymmetry.^[22][24][25] These are often based on linear regression, and may adopt a multiplicative or additive dispersion parameter to adjust for the presence of between-study heterogeneity. Some approaches may even attempt to compensate for the (potential) presence of publication bias,^[20][26][27] which is particularly useful to explore the potential impact on meta-analysis results.^[28][29][30]

Compensation examples

Two meta-analyses of the efficacy of reboxetine as an antidepressant demonstrated attempts to detect publication bias in clinical trials. Based on positive trial data, reboxetine was originally passed as a treatment for depression in many countries in Europe and the UK in 2001 (though in practice it is rarely used for this indication). A 2010 meta-analysis concluded that reboxetine was ineffective and that the preponderance of positive-outcome trials reflected publication bias, mostly due to trials published by the drug manufacturer Pfizer. A subsequent meta-analysis published in 2011, based on the original data, found flaws in the 2010 analyses and suggested that the data indicated reboxetine was effective in severe depression. Examples of publication bias are given by Goldacre^[31] and Wilmshurst.^[32]

In the social sciences, a study of published papers exploring the relationship between corporate social and financial performance found that "in economics, finance, and accounting journals, the average correlations were only about half the magnitude of the findings published in Social Issues Management, Business Ethics, or Business and Society journals".^[33]

One example cited as an instance of publication bias is the refusal to publish attempted replications of Bem's work that claimed evidence for precognition by The Journal of Personality and Social Psychology (the original publisher of Bem's article).^[34]

An analysis^[35] comparing studies of gene-disease associations originating in China to those originating outside China found that those conducted within the country reported a stronger association and a more statistically significant result.^[36]

Risks

Ioannidis argued that "claimed research findings may often be simply accurate measures of the prevailing bias."^[37] He lists the following factors as those that make a paper with a positive result more likely to enter the literature and suppress negative-result papers:

• The studies conducted in a field are smaller.
• The effect sizes are smaller.
• There is both a greater number and lesser preselection of tested relationships.
• There is greater flexibility in designs, definitions, outcomes, and analytical modes.
• There is prejudice (financial interest or otherwise).
• More teams are involved in a particular scientific field and chasing statistical significance.

Other factors include experimenter bias and white hat bias.

Remedies

Publication bias can be contained through better-powered studies, enhanced research standards, and careful consideration of true and non-true relationships.^[37] Better-powered studies refer to large studies that deliver definitive results or test major concepts and lead to low-bias meta-analysis. Enhanced research standards such as the pre-registration of protocols, the registration of data collections and adherence to established protocols are other techniques. To avoid false-positive results, the experimenter must consider the chances that they are testing a true or non-true relationship. This can be untertaken by properly assessing the false positive report probability based on the statistical power of the test^[38] and reconfirming (whenever ethically acceptable) established findings of prior studies known to have minimal bias.

Study registration

In September 2004, editors of prominent medical journals (including the New England Journal of Medicine, The Lancet, Annals of Internal Medicine, and JAMA) announced that they would no longer publish results of drug research sponsored by pharmaceutical companies, unless that research was registered in a public clinical trials registry database from the start.^[39] Furthermore, some journals (e.g. Trials), encourage publication of study protocols in their journals.^[40]

The World Health Organization (WHO) agreed that basic information about all clinical trials should be registered at the study's inception, and that this information should be publicly accessible through the WHO International Clinical Trials Registry Platform. Additionally, public availability of complete study protocols, alongside reports of trials, is becoming more common for studies .^[41]