You are currently browsing the category archive for the ‘math.RA’ category.
The fundamental notions of calculus, namely differentiation and integration, are often viewed as being the quintessential concepts in mathematical analysis, as their standard definitions involve the concept of a limit. However, it is possible to capture most of the essence of these notions by purely algebraic means (almost completely avoiding the use of limits, Riemann sums, and similar devices), which turns out to be useful when trying to generalise these concepts to more abstract situations in which it becomes convenient to permit the underlying number systems involved to be something other than the real or complex numbers, even if this makes many standard analysis constructions unavailable. For instance, the algebraic notion of a derivation often serves as a substitute for the analytic notion of a derivative in such cases, by abstracting out the key algebraic properties of differentiation, namely linearity and the Leibniz rule (also known as the product rule).
Abstract algebraic analogues of integration are less well known, but can still be developed. To motivate such an abstraction, consider the integration functional from the space of complex-valued Schwarz functions to the complex numbers, defined by
where the integration on the right is the usual Lebesgue integral (or improper Riemann integral) from analysis. This functional obeys two obvious algebraic properties. Firstly, it is linear over , thus
for all , where is the translation of by . Motivated by the uniqueness theory of Haar measure, one might expect that these two axioms already uniquely determine after one sets a normalisation, for instance by requiring that
This is not quite true as stated (one can modify the proof of the Hahn-Banach theorem, after first applying a Fourier transform, to create pathological translation-invariant linear functionals on that are not multiples of the standard Fourier transform), but if one adds a mild analytical axiom, such as continuity of (using the usual Schwartz topology on ), then the above axioms are enough to uniquely pin down the notion of integration. Indeed, if is a continuous linear functional that is translation invariant, then from the linearity and translation invariance axioms one has
for all and non-zero reals . If is Schwartz, then as , one can verify that the Newton quotients converge in the Schwartz topology to the derivative of , so by the continuity axiom one has
Next, note that any Schwartz function of integral zero has an antiderivative which is also Schwartz, and so annihilates all zero-integral Schwartz functions, and thus must be a scalar multiple of the usual integration functional. Using the normalisation (4), we see that must therefore be the usual integration functional, giving the claimed uniqueness.
Motivated by the above discussion, we can define the notion of an abstract integration functional taking values in some vector space , and applied to inputs in some other vector space that enjoys a linear action (the “translation action”) of some group , as being a functional which is both linear and translation invariant, thus one has the axioms (1), (2), (3) for all , scalars , and . The previous discussion then considered the special case when , , , and was the usual translation action.
Once we have performed this abstraction, we can now present analogues of classical integration which bear very little analytic resemblance to the classical concept, but which still have much of the algebraic structure of integration. Consider for instance the situation in which we keep the complex range , the translation group , and the usual translation action , but we replace the space of Schwartz functions by the space of polynomials of degree at most with complex coefficients, where is a fixed natural number; note that this space is translation invariant, so it makes sense to talk about an abstract integration functional . Of course, one cannot apply traditional integration concepts to non-zero polynomials, as they are not absolutely integrable. But one can repeat the previous arguments to show that any abstract integration functional must annihilate derivatives of polynomials of degree at most :
Clearly, every polynomial of degree at most is thus annihilated by , which makes a scalar multiple of the functional that extracts the top coefficient of a polynomial, thus if one sets a normalisation
for any polynomial . So we see that up to a normalising constant, the operation of extracting the top order coefficient of a polynomial of fixed degree serves as the analogue of integration. In particular, despite the fact that integration is supposed to be the “opposite” of differentiation (as indicated for instance by (5)), we see in this case that integration is basically (-fold) differentiation; indeed, compare (6) with the identity
In particular, we see, in contrast to the usual Lebesgue integral, the integration functional (6) can be localised to an arbitrary location: one only needs to know the germ of the polynomial at a single point in order to determine the value of the functional (6). This localisation property may initially seem at odds with the translation invariance, but the two can be reconciled thanks to the extremely rigid nature of the class , in contrast to the Schwartz class which admits bump functions and so can generate local phenomena that can only be detected in small regions of the underlying spatial domain, and which therefore forces any translation-invariant integration functional on such function classes to measure the function at every single point in space.
The reversal of the relationship between integration and differentiation is also reflected in the fact that the abstract integration operation on polynomials interacts with the scaling operation in essentially the opposite way from the classical integration operation. Indeed, for classical integration on , one has
for Schwartz functions , and so in this case the integration functional obeys the scaling law
In contrast, the abstract integration operation defined in (6) obeys the opposite scaling law
Remark 1 One way to interpret what is going on is to view the integration operation (6) as a renormalised version of integration. A polynomial is, in general, not absolutely integrable, and the partial integrals
diverge as . But if one renormalises these integrals by the factor , then one recovers convergence,
thus giving an interpretation of (6) as a renormalised classical integral, with the renormalisation being responsible for the unusual scaling relationship in (7). However, this interpretation is a little artificial, and it seems that it is best to view functionals such as (6) from an abstract algebraic perspective, rather than to try to force an analytic interpretation on them.
As noted earlier, this integration functional has a translation invariance associated to translations along the real line , as well as a dilation invariance by real dilation parameters . However, if we refine the class of functions somewhat, we can obtain a stronger family of invariances, in which we allow complex translations and dilations. More precisely, let denote the space of all functions which are entire (or equivalently, are given by a Taylor series with an infinite radius of convergence around the origin) and also admit rapid decay in a sectorial neighbourhood of the real line, or more precisely there exists an such that for every there exists such that one has the bound
whenever . For want of a better name, we shall call elements of this space Schwartz entire functions. This is clearly a complex vector space. A typical example of a Schwartz entire function are the complex gaussians
where are complex numbers with . From the Cauchy integral formula (and its derivatives) we see that if lies in , then the restriction of to the real line lies in ; conversely, from analytic continuation we see that every function in has at most one extension in . Thus one can identify with a subspace of , and in particular the integration functional (8) is inherited by , and by abuse of notation we denote the resulting functional as also. Note, in analogy with the situation with polynomials, that this abstract integration functional is somewhat localised; one only needs to evaluate the function on the real line, rather than the entire complex plane, in order to compute . This is consistent with the rigid nature of Schwartz entire functions, as one can uniquely recover the entire function from its values on the real line by analytic continuation.
Of course, the functional remains translation invariant with respect to real translation:
However, thanks to contour shifting, we now also have translation invariance with respect to complex translation:
where of course we continue to define the translation operator for complex by the usual formula . In a similar vein, we also have the scaling law
for any , if is a complex number sufficiently close to (where “sufficiently close” depends on , and more precisely depends on the sectoral aperture parameter associated to ); again, one can verify that lies in for sufficiently close to . These invariances (which relocalise the integration functional onto other contours than the real line ) are very useful for computing integrals, and in particular for computing gaussian integrals. For instance, the complex translation invariance tells us (after shifting by ) that
when with , and then an application of the complex scaling law (and a continuity argument, observing that there is a compact path connecting to in the right half plane) gives
using the branch of on the right half-plane for which . Using the normalisation (4) we thus have
This is a basic illustration of the power that a large symmetry group (in this case, the complex homothety group) can bring to bear on the task of computing integrals.
One can extend this sort of analysis to higher dimensions. For any natural number , let denote the space of all functions which is jointly entire in the sense that can be expressed as a Taylor series in which is absolutely convergent for all choices of , and such that there exists an such that for any there is for which one has the bound
whenever for all , where and . Again, we call such functions Schwartz entire functions; a typical example is the function
where is an complex symmetric matrix with positive definite real part, is a vector in , and is a complex number. We can then define an abstract integration functional by integration on the real slice :
where is the usual Lebesgue measure on . By contour shifting in each of the variables separately, we see that is invariant with respect to complex translations of each of the variables, and is thus invariant under translating the joint variable by . One can also verify the scaling law
for complex matrices sufficiently close to the origin, where . This can be seen for shear transformations by Fubini’s theorem and the aforementioned translation invariance, while for diagonal transformations near the origin this can be seen from applications of one-dimensional scaling law, and the general case then follows by composition. Among other things, these laws then easily lead to the higher-dimensional generalisation
whenever is a complex symmetric matrix with positive definite real part, is a vector in , and is a complex number, basically by repeating the one-dimensional argument sketched earlier. Here, we choose the branch of for all matrices in the indicated class for which .
where is now a complex variable
is the adjoint
is a complex matrix with positive definite Hermitian part, are column vectors in , is a complex number, and is times Lebesgue measure on . (The factors of two here turn out to be a natural normalisation, but they can be ignored on a first reading.) As we shall see later, such integrals are relevant when performing computations on the Gaussian Unitary Ensemble (GUE) in random matrix theory. Note that the integrand here is not complex analytic due to the presence of the complex conjugates. However, this can be dealt with by the trick of replacing the complex conjugate by a variable which is formally conjugate to , but which is allowed to vary independently of . More precisely, let be the space of all functions of two independent -tuples
of complex variables, which is jointly entire in all variables (in the sense defined previously, i.e. there is a joint Taylor series that is absolutely convergent for all independent choices of ), and such that there is an such that for every there is such that one has the bound
whenever . We will call such functions Schwartz analytic. Note that the integrand in (11) is Schwartz analytic when has positive definite Hermitian part, if we reinterpret as the transpose of rather than as the adjoint of in order to make the integrand entire in and . We can then define an abstract integration functional by the formula
thus can be localised to the slice of (though, as with previous functionals, one can use contour shifting to relocalise to other slices also.) One can also write this integral as
and note that the integrand here is a Schwartz entire function on , thus linking the Schwartz analytic integral with the Schwartz entire integral. Using this connection, one can verify that this functional is invariant with respect to translating and by independent shifts in (thus giving a translation symmetry), and one also has the independent dilation symmetry
for complex matrices that are sufficiently close to the identity, where . Arguing as before, we can then compute (11) as
In particular, this gives an integral representation for the determinant-reciprocal of a complex matrix with positive definite Hermitian part, in terms of gaussian expressions in which only appears linearly in the exponential:
This formula is then convenient for computing statistics such as
for random matrices drawn from the Gaussian Unitary Ensemble (GUE), and some choice of spectral parameter with ; we review this computation later in this post. By the trick of matrix differentiation of the determinant (as reviewed in this recent blog post), one can also use this method to compute matrix-valued statistics such as
However, if one restricts attention to classical integrals over real or complex (and in particular, commuting or bosonic) variables, it does not seem possible to easily eradicate the negative determinant factors in such calculations, which is unfortunate because many statistics of interest in random matrix theory, such as the expected Stieltjes transform
which is the Stieltjes transform of the density of states. However, it turns out (as I learned recently from Peter Sarnak and Tom Spencer) that it is possible to cancel out these negative determinant factors by balancing the bosonic gaussian integrals with an equal number of fermionic gaussian integrals, in which one integrates over a family of anticommuting variables. These fermionic integrals are closer in spirit to the polynomial integral (6) than to Lebesgue type integrals, and in particular obey a scaling law which is inverse to the Lebesgue scaling (in particular, a linear change of fermionic variables ends up transforming a fermionic integral by rather than ), which conveniently cancels out the reciprocal determinants in the previous calculations. Furthermore, one can combine the bosonic and fermionic integrals into a unified integration concept, known as the Berezin integral (or Grassmann integral), in which one integrates functions of supervectors (vectors with both bosonic and fermionic components), and is of particular importance in the theory of supersymmetry in physics. (The prefix “super” in physics means, roughly speaking, that the object or concept that the prefix is attached to contains both bosonic and fermionic aspects.) When one applies this unified integration concept to gaussians, this can lead to quite compact and efficient calculations (provided that one is willing to work with “super”-analogues of various concepts in classical linear algebra, such as the supertrace or superdeterminant).
Abstract integrals of the flavour of (6) arose in quantum field theory, when physicists sought to formally compute integrals of the form
where are familiar commuting (or bosonic) variables (which, in particular, can often be localised to be scalar variables taking values in or ), while were more exotic anticommuting (or fermionic) variables, taking values in some vector space of fermions. (As we shall see shortly, one can formalise these concepts by working in a supercommutative algebra.) The integrand was a formally analytic function of , in that it could be expanded as a (formal, noncommutative) power series in the variables . For functions that depend only on bosonic variables, it is certainly possible for such analytic functions to be in the Schwartz class and thus fall under the scope of the classical integral, as discussed previously. However, functions that depend on fermionic variables behave rather differently. Indeed, a fermonic variable must anticommute with itself, so that . In particular, any power series in terminates after the linear term in , so that a function can only be analytic in if it is a polynomial of degree at most in ; more generally, an analytic function of fermionic variables must be a polynomial of degree at most , and an analytic function of bosonic and fermionic variables can be Schwartz in the bosonic variables but will be polynomial in the fermonic variables. As such, to interpret the integral (14), one can use classical (Lebesgue) integration (or the variants discussed above for integrating Schwartz entire or Schwartz analytic functions) for the bosonic variables, but must use abstract integrals such as (6) for the fermonic variables, leading to the concept of Berezin integration mentioned earlier.
In this post I would like to set out some of the basic algebraic formalism of Berezin integration, particularly with regards to integration of gaussian-type expressions, and then show how this formalism can be used to perform computations involving GUE (for instance, one can compute the density of states of GUE by this machinery without recourse to the theory of orthogonal polynomials). The use of supersymmetric gaussian integrals to analyse ensembles such as GUE appears in the work of Efetov (and was also proposed in the slightly earlier works of Parisi-Sourlas and McKane, with a related approach also appearing in the work of Wegner); the material here is adapted from this survey of Mirlin, as well as the later papers of Disertori-Pinson-Spencer and of Disertori.
The determinant of a square matrix obeys a large number of important identities, the most basic of which is the multiplicativity property
whenever are square matrices of the same dimension. This identity then generates many other important identities. For instance, if is an matrix and is an matrix, then by applying the previous identity to equate the determinants of and (where we will adopt the convention that denotes an identity matrix of whatever dimension is needed to make sense of the expressions being computed, and similarly for ) we obtain the Sylvester determinant identity
This identity, which relates an determinant with an determinant, is very useful in random matrix theory (a point emphasised in particular by Deift), particularly in regimes in which is much smaller than .
Another identity generated from (1) arises when trying to compute the determinant of a block matrix
where is an matrix, is an matrix, is an matrix, and is an matrix. If is invertible, then we can manipulate this matrix via block Gaussian elimination as
and on taking determinants using (1) we obtain the Schur determinant identity
relating the determinant of a block-diagonal matrix with the determinant of the Schur complement of the upper left block . This identity can be viewed as the correct way to generalise the determinant formula
It is also possible to use determinant identities to deduce other matrix identities that do not involve the determinant, by the technique of matrix differentiation (or equivalently, matrix linearisation). The key observation is that near the identity, the determinant behaves like the trace, or more precisely one has
for any bounded square matrix and infinitesimal . (If one is uncomfortable with infinitesimals, one can interpret this sort of identity as an asymptotic as .) Combining this with (1) we see that for square matrices of the same dimension with invertible and invertible, one has
for infinitesimal . To put it another way, if is a square matrix that depends in a differentiable fashion on a real parameter , then
Let us see some examples of this differentiation method. If we take the Sylvester identity (2) and multiply one of the rectangular matrices by an infinitesimal , we obtain
applying (4) and extracting the linear term in (or equivalently, differentiating at and then setting ) we conclude the cyclic property of trace:
To manipulate derivatives and inverses, we begin with the Neumann series approximation
for bounded square and infinitesimal , which then leads to the more general approximation
for square matrices of the same dimension with bounded. To put it another way, we have
whenever depends in a differentiable manner on and is invertible.
We can then differentiate (or linearise) the Schur identity (3) in a number of ways. For instance, if we replace the lower block by for some test matrix , then by (4), the left-hand side of (3) becomes (assuming the invertibility of the block matrix)
while the right-hand side becomes
extracting the linear term in (after dividing through by (3)), we conclude that
As was an arbitrary matrix, we conclude from duality that the lower right block of is given by the inverse of the Schur complement:
One can also compute the other components of this inverse in terms of the Schur complement by a similar method (although the formulae become more complicated). As a variant of this method, we can perturb the block matrix in (3) by an infinitesimal multiple of the identity matrix giving
By (4), the left-hand side is
From (5), we have
which relates the trace of the inverse of a block matrix, with the trace of the inverse of one of its blocks. This particular identity turns out to be useful in random matrix theory; I hope to elaborate on this in a later post.
As a final example of this method, we can analyse low rank perturbations of a large () matrix , where is an matrix and is an matrix for some . (This type of situation is also common in random matrix theory, for instance it arose in this previous paper of mine on outliers to the circular law.) If is invertible, then from (1) and (2) one has the matrix determinant lemma
if one then perturbs by an infinitesimal matrix , we have
Extracting the linear component in as before, one soon arrives at
assuming that and are both invertible; as is arbitrary, we conclude (after using the cyclic property of trace) the Sherman-Morrison formula
for the inverse of a low rank perturbation of a matrix . While this identity can be easily verified by direct algebraic computation, it is somewhat difficult to discover this identity by such algebraic manipulation; thus we see that the “determinant first” approach to matrix identities can make it easier to find appropriate matrix identities (particularly those involving traces and/or inverses), even if the identities one is ultimately interested in do not involve determinants. (As differentiation typically makes an identity lengthier, but also more “linear” or “additive”, the determinant identity tends to be shorter (albeit more nonlinear and more multiplicative) than the differentiated identity, and can thus be slightly easier to derive.)
where is an matrix, is an matrix, is an matrix, and is an matrix, assuming that , and are all invertible.
Exercise 2 Let be invertible matrices. Establish the identity
and differentiate this in to deduce the identity
(assuming that all inverses exist) and hence
Rotating by then gives
which is useful for inverting a matrix that has been split into a self-adjoint component and a skew-adjoint component .
Mathematicians study a variety of different mathematical structures, but perhaps the structures that are most commonly associated with mathematics are the number systems, such as the integers or the real numbers . Indeed, the use of number systems is so closely identified with the practice of mathematics that one sometimes forgets that it is possible to do mathematics without explicit reference to any concept of number. For instance, the ancient Greeks were able to prove many theorems in Euclidean geometry, well before the development of Cartesian coordinates and analytic geometry in the seventeenth century, or the formal constructions or axiomatisations of the real number system that emerged in the nineteenth century (not to mention precursor concepts such as zero or negative numbers, whose very existence was highly controversial, if entertained at all, to the ancient Greeks). To do this, the Greeks used geometric operations as substitutes for the arithmetic operations that would be more familiar to modern mathematicians. For instance, concatenation of line segments or planar regions serves as a substitute for addition; the operation of forming a rectangle out of two line segments would serve as a substitute for multiplication; the concept of similarity can be used as a substitute for ratios or division; and so forth.
A similar situation exists in modern physics. Physical quantities such as length, mass, momentum, charge, and so forth are routinely measured and manipulated using the real number system (or related systems, such as if one wishes to measure a vector-valued physical quantity such as velocity). Much as analytic geometry allows one to use the laws of algebra and trigonometry to calculate and prove theorems in geometry, the identification of physical quantities with numbers allows one to express physical laws and relationships (such as Einstein’s famous mass-energy equivalence ) as algebraic (or differential) equations, which can then be solved and otherwise manipulated through the extensive mathematical toolbox that has been developed over the centuries to deal with such equations.
However, as any student of physics is aware, most physical quantities are not represented purely by one or more numbers, but instead by a combination of a number and some sort of unit. For instance, it would be a category error to assert that the length of some object was a number such as ; instead, one has to say something like “the length of this object is yards”, combining both a number and a unit (in this case, the yard). Changing the unit leads to a change in the numerical value assigned to this physical quantity, even though no physical change to the object being measured has occurred. For instance, if one decides to use feet as the unit of length instead of yards, then the length of the object is now feet; if one instead uses metres, the length is now metres; and so forth. But nothing physical has changed when performing this change of units, and these lengths are considered all equal to each other:
It is then common to declare that while physical quantities and units are not, strictly speaking, numbers, they should be manipulated using the laws of algebra as if they were numerical quantities. For instance, if an object travels metres in seconds, then its speed should be
where we use the usual abbreviations of and for metres and seconds respectively. Similarly, if the speed of light is and an object has mass , then Einstein’s mass-energy equivalence then tells us that the energy-content of this object is
Note that the symbols are being manipulated algebraically as if they were mathematical variables such as and . By collecting all these units together, we see that every physical quantity gets assigned a unit of a certain dimension: for instance, we see here that the energy of an object can be given the unit of (more commonly known as a Joule), which has the dimension of where are the dimensions of mass, length, and time respectively.
There is however one important limitation to the ability to manipulate “dimensionful” quantities as if they were numbers: one is not supposed to add, subtract, or compare two physical quantities if they have different dimensions, although it is acceptable to multiply or divide two such quantities. For instance, if is a mass (having the units ) and is a speed (having the units ), then it is physically “legitimate” to form an expression such as , but not an expression such as or ; in a similar spirit, statements such as or are physically meaningless. This combines well with the mathematical distinction between vector, scalar, and matrix quantities, which among other things prohibits one from adding together two such quantities if their vector or matrix type are different (e.g. one cannot add a scalar to a vector, or a vector to a matrix), and also places limitations on when two such quantities can be multiplied together. A related limitation, which is not always made explicit in physics texts, is that transcendental mathematical functions such as or should only be applied to arguments that are dimensionless; thus, for instance, if is a speed, then is not physically meaningful, but is (this particular quantity is known as the rapidity associated to this speed).
These limitations may seem like a weakness in the mathematical modeling of physical quantities; one may think that one could get a more “powerful” mathematical framework if one were allowed to perform dimensionally inconsistent operations, such as add together a mass and a velocity, add together a vector and a scalar, exponentiate a length, etc. Certainly there is some precedent for this in mathematics; for instance, the formalism of Clifford algebras does in fact allow one to (among other things) add vectors with scalars, and in differential geometry it is quite common to formally apply transcendental functions (such as the exponential function) to a differential form (for instance, the Liouville measure of a symplectic manifold can be usefully thought of as a component of the exponential of the symplectic form ).
However, there are several reasons why it is advantageous to retain the limitation to only perform dimensionally consistent operations. One is that of error correction: one can often catch (and correct for) errors in one’s calculations by discovering a dimensional inconsistency, and tracing it back to the first step where it occurs. Also, by performing dimensional analysis, one can often identify the form of a physical law before one has fully derived it. For instance, if one postulates the existence of a mass-energy relationship involving only the mass of an object , the energy content , and the speed of light , dimensional analysis is already sufficient to deduce that the relationship must be of the form for some dimensionless absolute constant ; the only remaining task is then to work out the constant of proportionality , which requires physical arguments beyond that provided by dimensional analysis. (This is a simple instance of a more general application of dimensional analysis known as the Buckingham theorem.)
The use of units and dimensional analysis has certainly been proven to be very effective tools in physics. But one can pose the question of whether it has a properly grounded mathematical foundation, in order to settle any lingering unease about using such tools in physics, and also in order to rigorously develop such tools for purely mathematical purposes (such as analysing identities and inequalities in such fields of mathematics as harmonic analysis or partial differential equations).
The example of Euclidean geometry mentioned previously offers one possible approach to formalising the use of dimensions. For instance, one could model the length of a line segment not by a number, but rather by the equivalence class of all line segments congruent to the original line segment (cf. the Frege-Russell definition of a number). Similarly, the area of a planar region can be modeled not by a number, but by the equivalence class of all regions that are equidecomposable with the original region (one can, if one wishes, restrict attention here to measurable sets in order to avoid Banach-Tarski-type paradoxes, though that particular paradox actually only arises in three and higher dimensions). As mentioned before, it is then geometrically natural to multiply two lengths to form an area, by taking a rectangle whose line segments have the stated lengths, and using the area of that rectangle as a product. This geometric picture works well for units such as length and volume that have a spatial geometric interpretation, but it is less clear how to apply it for more general units. For instance, it does not seem geometrically natural (or, for that matter, conceptually helpful) to envision the equation as the assertion that the energy is the volume of a rectangular box whose height is the mass and whose length and width is given by the speed of light .
But there are at least two other ways to formalise dimensionful quantities in mathematics, which I will discuss below the fold. The first is a “parametric” model in which dimensionful objects are modeled as numbers (or vectors, matrices, etc.) depending on some base dimensional parameters (such as units of length, mass, and time, or perhaps a coordinate system for space or spacetime), and transforming according to some representation of a structure group that encodes the range of these parameters; this type of “coordinate-heavy” model is often used (either implicitly or explicitly) by physicists in order to efficiently perform calculations, particularly when manipulating vector or tensor-valued quantities. The second is an “abstract” model in which dimensionful objects now live in an abstract mathematical space (e.g. an abstract vector space), in which only a subset of the operations available to general-purpose number systems such as or are available, namely those operations which are “dimensionally consistent” or invariant (or more precisely, equivariant) with respect to the action of the underlying structure group. This sort of “coordinate-free” approach tends to be the one which is preferred by pure mathematicians, particularly in the various branches of modern geometry, in part because it can lead to greater conceptual clarity, as well as results of great generality; it is also close to the more informal practice of treating mathematical manipulations that do not preserve dimensional consistency as being physically meaningless.
Given a function between two sets , we can form the graph
which is a subset of the Cartesian product .
There are a number of “closed graph theorems” in mathematics which relate the regularity properties of the function with the closure properties of the graph , assuming some “completeness” properties of the domain and range . The most famous of these is the closed graph theorem from functional analysis, which I phrase as follows:
Theorem 1 (Closed graph theorem (functional analysis)) Let be complete normed vector spaces over the reals (i.e. Banach spaces). Then a function is a continuous linear transformation if and only if the graph is both linearly closed (i.e. it is a linear subspace of ) and topologically closed (i.e. closed in the product topology of ).
I like to think of this theorem as linking together qualitative and quantitative notions of regularity preservation properties of an operator ; see this blog post for further discussion.
The theorem is equivalent to the assertion that any continuous linear bijection from one Banach space to another is necessarily an isomorphism in the sense that the inverse map is also continuous and linear. Indeed, to see that this claim implies the closed graph theorem, one applies it to the projection from to , which is a continuous linear bijection; conversely, to deduce this claim from the closed graph theorem, observe that the graph of the inverse is the reflection of the graph of . As such, the closed graph theorem is a corollary of the open mapping theorem, which asserts that any continuous linear surjection from one Banach space to another is open. (Conversely, one can deduce the open mapping theorem from the closed graph theorem by quotienting out the kernel of the continuous surjection to get a bijection.)
It turns out that there is a closed graph theorem (or equivalent reformulations of that theorem, such as an assertion that bijective morphisms between sufficiently “complete” objects are necessarily isomorphisms, or as an open mapping theorem) in many other categories in mathematics as well. Here are some easy ones:
Theorem 2 (Closed graph theorem (linear algebra)) Let be vector spaces over a field . Then a function is a linear transformation if and only if the graph is linearly closed.
Theorem 3 (Closed graph theorem (group theory)) Let be groups. Then a function is a group homomorphism if and only if the graph is closed under the group operations (i.e. it is a subgroup of ).
Theorem 4 (Closed graph theorem (order theory)) Let be totally ordered sets. Then a function is monotone increasing if and only if the graph is totally ordered (using the product order on ).
Remark 1 Similar results to the above three theorems (with similarly easy proofs) hold for other algebraic structures, such as rings (using the usual product of rings), modules, algebras, or Lie algebras, groupoids, or even categories (a map between categories is a functor iff its graph is again a category). (ADDED IN VIEW OF COMMENTS: further examples include affine spaces and -sets (sets with an action of a given group ).) There are also various approximate versions of this theorem that are useful in arithmetic combinatorics, that relate the property of a map being an “approximate homomorphism” in some sense with its graph being an “approximate group” in some sense. This is particularly useful for this subfield of mathematics because there are currently more theorems about approximate groups than about approximate homomorphisms, so that one can profitably use closed graph theorems to transfer results about the former to results about the latter.
A slightly more sophisticated result in the same vein:
Theorem 5 (Closed graph theorem (point set topology)) Let be compact Hausdorff spaces. Then a function is continuous if and only if the graph is topologically closed.
Indeed, the “only if” direction is easy, while for the “if” direction, note that if is a closed subset of , then it is compact Hausdorff, and the projection map from to is then a bijective continuous map between compact Hausdorff spaces, which is then closed, thus open, and hence a homeomorphism, giving the claim.
Note that the compactness hypothesis is necessary: for instance, the function defined by for and for is a function which has a closed graph, but is discontinuous.
A similar result (but relying on a much deeper theorem) is available in algebraic geometry, as I learned after asking this MathOverflow question:
Theorem 6 (Closed graph theorem (algebraic geometry)) Let be normal projective varieties over an algebraically closed field of characteristic zero. Then a function is a regular map if and only if the graph is Zariski-closed.
Proof: (Sketch) For the only if direction, note that the map is a regular map from the projective variety to the projective variety and is thus a projective morphism, hence is proper. In particular, the image of under this map is Zariski-closed.
Conversely, if is Zariski-closed, then it is also a projective variety, and the projection is a projective morphism from to , which is clearly quasi-finite; by the characteristic zero hypothesis, it is also separated. Applying (Grothendieck’s form of) Zariski’s main theorem, this projection is the composition of an open immersion and a finite map. As projective varieties are complete, the open immersion is an isomorphism, and so the projection from to is finite. Being injective and separable, the degree of this finite map must be one, and hence and are isomorphic, hence (by normality of ) is contained in (the image of) , which makes the map from to regular, which makes regular.
The counterexample of the map given by for and demonstrates why the projective hypothesis is necessary. The necessity of the normality condition (or more precisely, a weak normality condition) is demonstrated by (the projective version of) the map from the cusipdal curve to . (If one restricts attention to smooth varieties, though, normality becomes automatic.) The necessity of characteristic zero is demonstrated by (the projective version of) the inverse of the Frobenius map on a field of characteristic .
There are also a number of closed graph theorems for topological groups, of which the following is typical (see Exercise 3 of these previous blog notes):
Theorem 7 (Closed graph theorem (topological group theory)) Let be -compact, locally compact Hausdorff groups. Then a function is a continuous homomorphism if and only if the graph is both group-theoretically closed and topologically closed.
The hypotheses of being -compact, locally compact, and Hausdorff can be relaxed somewhat, but I doubt that they can be eliminated entirely (though I do not have a ready counterexample for this).
In several complex variables, it is a classical theorem (see e.g. Lemma 4 of this blog post) that a holomorphic function from a domain in to is locally injective if and only if it is a local diffeomorphism (i.e. its derivative is everywhere non-singular). This leads to a closed graph theorem for complex manifolds:
Theorem 8 (Closed graph theorem (complex manifolds)) Let be complex manifolds. Then a function is holomorphic if and only if the graph is a complex manifold (using the complex structure inherited from ) of the same dimension as .
Indeed, one applies the previous observation to the projection from to . The dimension requirement is needed, as can be seen from the example of the map defined by for and .
(ADDED LATER:) There is a real analogue to the above theorem:
Theorem 9 (Closed graph theorem (real manifolds)) Let be real manifolds. Then a function is continuous if and only if the graph is a real manifold of the same dimension as .
Note though that the analogous claim for smooth real manifolds fails: the function defined by has a smooth graph, but is not itself smooth.
(ADDED YET LATER:) Here is an easy closed graph theorem in the symplectic category:
Theorem 10 (Closed graph theorem (symplectic geometry)) Let and be smooth symplectic manifolds of the same dimension. Then a smooth map is a symplectic morphism (i.e. ) if and only if the graph is a Lagrangian submanifold of with the symplectic form .
In view of the symplectic rigidity phenomenon, it is likely that the smoothness hypotheses on can be relaxed substantially, but I will not try to formulate such a result here.
There are presumably many further examples of closed graph theorems (or closely related theorems, such as criteria for inverting a morphism, or open mapping type theorems) throughout mathematics; I would be interested to know of further examples.
where . One easily verifies that these expressions are well-defined (and depend smoothly on and ) when and are sufficiently small.
We have the famous Baker-Campbell-Hausdoff-Dynkin formula:
Theorem 1 (BCH formula) Let be a finite-dimensional Lie group over the reals with Lie algebra . Let be a local inverse of the exponential map , defined in a neighbourhood of the identity. Then for sufficiently small , one has
See for instance these notes of mine for a proof of this formula (it is for , but one easily obtains a similar proof for ).
for sufficiently small , and the inverse operation by (one easily verifies that for all small ).
It is tempting to reverse the BCH formula and conclude (the local form of) Lie’s third theorem, that every finite-dimensional Lie algebra is isomorphic to the Lie algebra of some local Lie group, by using (3) to define a smooth local group structure on a neighbourhood of the identity. (See this previous post for a definition of a local Lie group.) The main difficulty in doing so is in verifying that the definition (3) is well-defined (i.e. that is always equal to ) and locally associative. The well-definedness issue can be trivially disposed of by using just one of the expressions or as the definition of (though, as we shall see, it will be very convenient to use both of them simultaneously). However, the associativity is not obvious at all.
With the assistance of Ado’s theorem, which places inside the general linear Lie algebra for some , one can deduce both the well-definedness and associativity of (3) from the Baker-Campbell-Hausdorff formula for . However, Ado’s theorem is rather difficult to prove (see for instance this previous blog post for a proof), and it is natural to ask whether there is a way to establish these facts without Ado’s theorem.
After playing around with this for some time, I managed to extract a direct proof of well-definedness and local associativity of (3), giving a proof of Lie’s third theorem independent of Ado’s theorem. This is not a new result by any means, (indeed, the original proofs of Lie and Cartan of Lie’s third theorem did not use Ado’s theorem), but I found it an instructive exercise to work out the details, and so I am putting it up on this blog in case anyone else is interested (and also because I want to be able to find the argument again if I ever need it in the future).
Jordan’s theorem is a basic theorem in the theory of finite linear groups, and can be formulated as follows:
Informally, Jordan’s theorem asserts that finite linear groups over the complex numbers are almost abelian. The theorem can be extended to other fields of characteristic zero, and also to fields of positive characteristic so long as the characteristic does not divide the order of , but we will not consider these generalisations here. A proof of this theorem can be found for instance in these lecture notes of mine.
I recently learned (from this comment of Kevin Ventullo) that the finiteness hypothesis on the group in this theorem can be relaxed to the significantly weaker condition of periodicity. Recall that a group is periodic if all elements are of finite order. Jordan’s theorem with “finite” replaced by “periodic” is known as the Jordan-Schur theorem.
The Jordan-Schur theorem can be quickly deduced from Jordan’s theorem, and the following result of Schur:
Theorem 2 (Schur’s theorem) Every finitely generated periodic subgroup of a general linear group is finite. (Equivalently, every periodic linear group is locally finite.)
Remark 1 The question of whether all finitely generated periodic subgroups (not necessarily linear in nature) were finite was known as the Burnside problem; the answer was shown to be negative by Golod and Shafarevich in 1964.
Let us see how Jordan’s theorem and Schur’s theorem combine via a compactness argument to form the Jordan-Schur theorem. Let be a periodic subgroup of . Then for every finite subset of , the group generated by is finite by Theorem 2. Applying Jordan’s theorem, contains an abelian subgroup of index at most .
In particular, given any finite number of finite subsets of , we can find abelian subgroups of respectively such that each has index at most in . We claim that we may furthermore impose the compatibility condition whenever . To see this, we set , locate an abelian subgroup of of index at most , and then set . As is covered by at most cosets of , we see that is covered by at most cosets of , and the claim follows.
Note that for each , the set of possible is finite, and so the product space of all configurations , as ranges over finite subsets of , is compact by Tychonoff’s theorem. Using the finite intersection property, we may thus locate a subgroup of of index at most for all finite subsets of , obeying the compatibility condition whenever . If we then set , where ranges over all finite subsets of , we then easily verify that is abelian and has index at most in , as required.
Below I record a proof of Schur’s theorem, which I extracted from this book of Wehrfritz. This was primarily an exercise for my own benefit, but perhaps it may be of interest to some other readers.
Let be a Lie group with Lie algebra . As is well known, the exponential map is a local homeomorphism near the identity. As such, the group law on can be locally pulled back to an operation defined on a neighbourhood of the identity in , defined as
where is the local inverse of the exponential map. One can view as the group law expressed in local exponential coordinates around the origin.
An asymptotic expansion for is provided by the Baker-Campbell-Hausdorff (BCH) formula
for all sufficiently small , where is the Lie bracket. More explicitly, one has the Baker-Campbell-Hausdorff-Dynkin formula
for all sufficiently small , where , is the adjoint representation , and is the function
which is real analytic near and can thus be applied to linear operators sufficiently close to the identity. One corollary of this is that the multiplication operation is real analytic in local coordinates, and so every smooth Lie group is in fact a real analytic Lie group.
It turns out that one does not need the full force of the smoothness hypothesis to obtain these conclusions. It is, for instance, a classical result that regularity of the group operations is already enough to obtain the Baker-Campbell-Hausdorff formula. Actually, it turns out that we can weaken this a bit, and show that even regularity (i.e. that the group operations are continuously differentiable, and the derivatives are locally Lipschitz) is enough to make the classical derivation of the Baker-Campbell-Hausdorff formula work. More precisely, we have
Theorem 1 ( Baker-Campbell-Hausdorff formula) Let be a finite-dimensional vector space, and suppose one has a continuous operation defined on a neighbourhood around the origin, which obeys the following three axioms:
- (Approximate additivity) For sufficiently close to the origin, one has
(In particular, for sufficiently close to the origin.)
- (Associativity) For sufficiently close to the origin, .
- (Radial homogeneity) For sufficiently close to the origin, one has
for all . (In particular, for all sufficiently close to the origin.)
Then is real analytic (and in particular, smooth) near the origin. (In particular, gives a neighbourhood of the origin the structure of a local Lie group.)
Indeed, we will recover the Baker-Campbell-Hausdorff-Dynkin formula (after defining appropriately) in this setting; see below the fold.
The reason that we call this a Baker-Campbell-Hausdorff formula is that if the group operation has regularity, and has as an identity element, then Taylor expansion already gives (2), and in exponential coordinates (which, as it turns out, can be defined without much difficulty in the category) one automatically has (3).
We will record the proof of Theorem 1 below the fold; it largely follows the classical derivation of the BCH formula, but due to the low regularity one will rely on tools such as telescoping series and Riemann sums rather than on the fundamental theorem of calculus. As an application of this theorem, we can give an alternate derivation of one of the components of the solution to Hilbert’s fifth problem, namely the construction of a Lie group structure from a Gleason metric, which was covered in the previous post; we discuss this at the end of this article. With this approach, one can avoid any appeal to von Neumann’s theorem and Cartan’s theorem (discussed in this post), or the Kuranishi-Gleason extension theorem (discussed in this post).
(One can of course define Lie algebras over other fields than the complex numbers , but in order to avoid some technical issues we shall work solely with the complex case in this post.)
An important special case of the abstract Lie algebras are the concrete Lie algebras, in which is a vector space of linear transformations on a vector space (which again can be either finite or infinite dimensional), and the bilinear form is given by the usual Lie bracket
It is easy to verify that every concrete Lie algebra is an abstract Lie algebra. In the converse direction, we have
To prove this theorem, we introduce the useful algebraic tool of the universal enveloping algebra of the abstract Lie algebra . This is the free (associative, complex) algebra generated by (viewed as a complex vector space), subject to the constraints
This algebra is described by the Poincaré-Birkhoff-Witt theorem, which asserts that given an ordered basis of as a vector space, that a basis of is given by “monomials” of the form
where is a natural number, the are an increasing sequence of indices in , and the are positive integers. Indeed, given two such monomials, one can express their product as a finite linear combination of further monomials of the form (3) after repeatedly applying (2) (which we rewrite as ) to reorder the terms in this product modulo lower order terms until one all monomials have their indices in the required increasing order. It is then a routine exercise in basic abstract algebra (using all the axioms of an abstract Lie algebra) to verify that this is multiplication rule on monomials does indeed define a complex associative algebra which has the universal properties required of the universal enveloping algebra.
The abstract Lie algebra acts on its universal enveloping algebra by left-multiplication: , thus giving a map from to . It is easy to verify that this map is a Lie algebra homomorphism (so this is indeed an action (or representation) of the Lie algebra), and this action is clearly faithful (i.e. the map from to is injective), since each element of maps the identity element of to a different element of , namely . Thus is isomorphic to its image in , proving Theorem 1.
In the converse direction, every representation of a Lie algebra “factors through” the universal enveloping algebra, in that it extends to an algebra homomorphism from to , which by abuse of notation we shall also call .
One drawback of Theorem 1 is that the space that the concrete Lie algebra acts on will almost always be infinite-dimensional, even when the original Lie algebra is finite-dimensional. However, there is a useful theorem of Ado that rectifies this:
Theorem 2 (Ado’s theorem) Every finite-dimensional abstract Lie algebra is isomorphic to a concrete Lie algebra over a finite-dimensional vector space .
Among other things, this theorem can be used (in conjunction with the Baker-Campbell-Hausdorff formula) to show that every abstract (finite-dimensional) Lie group (or abstract local Lie group) is locally isomorphic to a linear group. (It is well-known, though, that abstract Lie groups are not necessarily globally isomorphic to a linear group, but we will not discuss these global obstructions here.)
Ado’s theorem is surprisingly tricky to prove in general, but some special cases are easy. For instance, one can try using the adjoint representation of on itself, defined by the action ; the Jacobi identity (1) ensures that this indeed a representation of . The kernel of this representation is the centre . This already gives Ado’s theorem in the case when is semisimple, in which case the center is trivial.
The adjoint representation does not suffice, by itself, to prove Ado’s theorem in the non-semisimple case. However, it does provide an important reduction in the proof, namely it reduces matters to showing that every finite-dimensional Lie algebra has a finite-dimensional representation which is faithful on the centre . Indeed, if one has such a representation, one can then take the direct sum of that representation with the adjoint representation to obtain a new finite-dimensional representation which is now faithful on all of , which then gives Ado’s theorem for .
It remains to find a finite-dimensional representation of which is faithful on the centre . In the case when is abelian, so that the centre is all of , this is again easy, because then acts faithfully on by the infinitesimal shear maps . In matrix form, this representation identifies each in this abelian Lie algebra with an “upper-triangular” matrix:
This construction gives a faithful finite-dimensional representation of the centre of any finite-dimensional Lie algebra. The standard proof of Ado’s theorem (which I believe dates back to work of Harish-Chandra) then proceeds by gradually “extending” this representation of the centre to larger and larger sub-algebras of , while preserving the finite-dimensionality of the representation and the faithfulness on , until one obtains a representation on the entire Lie algebra with the required properties. (For technical inductive reasons, one also needs to carry along an additional property of the representation, namely that it maps the nilradical to nilpotent elements, but we will discuss this technicality later.)
This procedure is a little tricky to execute in general, but becomes simpler in the nilpotent case, in which the lower central series becomes trivial for sufficiently large :
Theorem 3 (Ado’s theorem for nilpotent Lie algebras) Let be a finite-dimensional nilpotent Lie algebra. Then there exists a finite-dimensional faithful representation of . Furthermore, there exists a natural number such that , i.e. one has for all .
The second conclusion of Ado’s theorem here is useful for induction purposes. (By Engel’s theorem, this conclusion is also equivalent to the assertion that every element of is nilpotent, but we can prove Theorem 3 without explicitly invoking Engel’s theorem.)
Below the fold, I give a proof of Theorem 3, and then extend the argument to cover the full strength of Ado’s theorem. This is not a new argument – indeed, I am basing this particular presentation from the one in Fulton and Harris – but it was an instructive exercise for me to try to extract the proof of Ado’s theorem from the more general structural theory of Lie algebras (e.g. Engel’s theorem, Lie’s theorem, Levi decomposition, etc.) in which the result is usually placed. (However, the proof I know of still needs Engel’s theorem to establish the solvable case, and the Levi decomposition to then establish the general case.)
I’ve just uploaded to the arXiv my paper “Outliers in the spectrum of iid matrices with bounded rank perturbations“, submitted to Probability Theory and Related Fields. This paper is concerned with outliers to the circular law for iid random matrices. Recall that if is an matrix whose entries are iid complex random variables with mean zero and variance one, then the complex eigenvalues of the normalised matrix will almost surely be distributed according to the circular law distribution in the limit . (See these lecture notes for further discussion of this law.)
The circular law is also stable under bounded rank perturbations: if is a deterministic rank matrix of polynomial size (i.e. of operator norm ), then the circular law also holds for (this is proven in a paper of myself, Van Vu, and Manjunath Krisnhapur). In particular, the bulk of the eigenvalues (i.e. of the eigenvalues) will lie inside the unit disk .
However, this leaves open the possibility for one or more outlier eigenvalues that lie significantly outside the unit disk; the arguments in the paper cited above give some upper bound on the number of such eigenvalues (of the form for some absolute constant ) but does not exclude them entirely. And indeed, numerical data shows that such outliers can exist for certain bounded rank perturbations.
In this paper, some results are given as to when outliers exist, and how they are distributed. The easiest case is of course when there is no bounded rank perturbation: . In that case, an old result of Bai and Yin and of Geman shows that the spectral radius of is almost surely , thus all eigenvalues will be contained in a neighbourhood of the unit disk, and so there are no significant outliers. The proof is based on the moment method.
Now we consider a bounded rank perturbation which is nonzero, but which has a bounded operator norm: . In this case, it turns out that the matrix will have outliers if the deterministic component has outliers. More specifically (and under the technical hypothesis that the entries of have bounded fourth moment), if is an eigenvalue of with , then (for large enough), will almost surely have an eigenvalue at , and furthermore these will be the only outlier eigenvalues of .
Thus, for instance, adding a bounded nilpotent low rank matrix to will not create any outliers, because the nilpotent matrix only has eigenvalues at zero. On the other hand, adding a bounded Hermitian low rank matrix will create outliers as soon as this matrix has an operator norm greater than .
When I first thought about this problem (which was communicated to me by Larry Abbott), I believed that it was quite difficult, because I knew that the eigenvalues of non-Hermitian matrices were quite unstable with respect to general perturbations (as discussed in this previous blog post), and that there were no interlacing inequalities in this case to control bounded rank perturbations (as discussed in this post). However, as it turns out I had arrived at the wrong conclusion, especially in the exterior of the unit disk in which the resolvent is actually well controlled and so there is no pseudospectrum present to cause instability. This was pointed out to me by Alice Guionnet at an AIM workshop last week, after I had posed the above question during an open problems session. Furthermore, at the same workshop, Percy Deift emphasised the point that the basic determinantal identity
for matrices and matrices was a particularly useful identity in random matrix theory, as it converted problems about large () matrices into problems about small () matrices, which was particularly convenient in the regime when and was fixed. (Percy was speaking in the context of invariant ensembles, but the point is in fact more general than this.)
From this, it turned out to be a relatively simple manner to transform what appeared to be an intractable matrix problem into quite a well-behaved matrix problem for bounded . Specifically, suppose that had rank , so that one can factor for some (deterministic) matrix and matrix . To find an eigenvalue of , one has to solve the characteristic polynomial equation
This is an determinantal equation, which looks difficult to control analytically. But we can manipulate it using (1). If we make the assumption that is outside the spectrum of (which we can do as long as is well away from the unit disk, as the unperturbed matrix has no outliers), we can divide by to arrive at
Now we apply the crucial identity (1) to rearrange this as
The crucial point is that this is now an equation involving only a determinant, rather than an one, and is thus much easier to solve. The situation is particularly simple for rank one perturbations
in which case the eigenvalue equation is now just a scalar equation
that involves what is basically a single coefficient of the resolvent . (It is also an instructive exercise to derive this eigenvalue equation directly, rather than through (1).) There is by now a very well-developed theory for how to control such coefficients (particularly for in the exterior of the unit disk, in which case such basic tools as Neumann series work just fine); in particular, one has precise enough control on these coefficients to obtain the result on outliers mentioned above.
The same method can handle some other bounded rank perturbations. One basic example comes from looking at iid matrices with a non-zero mean and variance ; this can be modeled by where is the unit vector . Here, the bounded rank perturbation has a large operator norm (equal to ), so the previous result does not directly apply. Nevertheless, the self-adjoint nature of the perturbation has a stabilising effect, and I was able to show that there is still only one outlier, and that it is at the expected location of .
If one moves away from the case of self-adjoint perturbations, though, the situation changes. Let us now consider a matrix of the form , where is a randomised version of , e.g. , where the are iid Bernoulli signs; such models were proposed recently by Rajan and Abbott as a model for neural networks in which some nodes are excitatory (and give columns with positive mean) and some are inhibitory (leading to columns with negative mean). Despite the superficial similarity with the previous example, the outlier behaviour is now quite different. Instead of having one extremely large outlier (of size ) at an essentially deterministic location, we now have a number of eigenvalues of size , scattered according to a random process. Indeed, (in the case when the entries of were real and bounded) I was able to show that the outlier point process converged (in the sense of converging -point correlation functions) to the zeroes of a random Laurent series
where are iid real Gaussians. This is basically because the coefficients of the resolvent have a Neumann series whose coefficients enjoy a central limit theorem.
On the other hand, as already observed numerically (and rigorously, in the gaussian case) by Rajan and Abbott, if one projects such matrices to have row sum zero, then the outliers all disappear. This can be explained by another appeal to (1); this projection amounts to right-multiplying by the projection matrix to the zero-sum vectors. But by (1), the non-zero eigenvalues of the resulting matrix are the same as those for . Since annihilates , we thus see that in this case the bounded rank perturbation plays no role, and the question reduces to obtaining a circular law with no outliers for . As it turns out, this can be done by invoking the machinery of Van Vu and myself that we used to prove the circular law for various random matrix models.
In the previous lectures, we have focused mostly on the equidistribution or linear patterns on a subset of the integers , and in particular on intervals . The integers are of course a very important domain to study in additive combinatorics; but there are also other fundamental model examples of domains to study. One of these is that of a vector space over a finite field of prime order. Such domains are of interest in computer science (particularly when ) and also in number theory; but they also serve as an important simplified “dyadic model” for the integers. See this survey article of Green for further discussion of this point.
The additive combinatorics of the integers , and of vector spaces over finite fields, are analogous, but not quite identical. For instance, the analogue of an arithmetic progression in is a subspace of . In many cases, the finite field theory is a little bit simpler than the integer theory; for instance, subspaces are closed under addition, whereas arithmetic progressions are only “almost” closed under addition in various senses. (For instance, is closed under addition approximately half of the time.) However, there are some ways in which the integers are better behaved. For instance, because the integers can be generated by a single generator, a homomorphism from to some other group can be described by a single group element : . However, to specify a homomorphism from a vector space to one would need to specify one group element for each dimension of . Thus we see that there is a tradeoff when passing from (or ) to a vector space model; one gains a bounded torsion property, at the expense of conceding the bounded generation property. (Of course, if one wants to deal with arbitrarily large domains, one has to concede one or the other; the only additive groups that have both bounded torsion and boundedly many generators, are bounded.)
The starting point for this course (Notes 1) was the study of equidistribution of polynomials from the integers to the unit circle. We now turn to the parallel theory of equidistribution of polynomials from vector spaces over finite fields to the unit circle. Actually, for simplicity we will mostly focus on the classical case, when the polynomials in fact take values in the roots of unity (where is the characteristic of the field ). As it turns out, the non-classical case is also of importance (particularly in low characteristic), but the theory is more difficult; see these notes for some further discussion.