Welcome to a new series for this blog, Low-Rank Approximation Toolbox. As I discussed in a previous post, many matrices we encounter in applications are well-approximated by a matrix with a small rank. Efficiently computing low-rank approximations has been a major area of research, with applications in everything from classical problems in computational physics and signal processing to trendy topics like data science. In this series, I want to explore some broadly useful algorithms and theoretical techniques in the field of low-rank approximation.
I want to begin this series by talking about one of the fundamental types of low-rank approximation, the Nyström approximation of a (real symmetric or complex Hermitian) positive semidefinite (psd) matrix . Given an arbitrary “test matrix” , the Nyström approximation is defined to be
(1)
This formula is sensible whenever is invertible; if is not invertible, then the inverse should be replaced by the Moore–Penrose pseudoinverse . For simplicity, I will assume that is invertible in this post, though everything we discuss will continue to work if this assumption is dropped. I use to denote the conjugate transpose of a matrix, which agrees with the ordinary transpose for real matrices. I will use the word self-adjoint to refer to a matrix which satisfies .
The Nyström approximation (1) answers the question
What is the “best” rank- approximation to the psd matrx provided only with the matrix–matrix product , where is a known matrix ()?
Indeed, if we let , we observe that the Nyström approximation can be written entirely using and :
This is central advantage of the Nyström approximation: to compute it, the only access to the matrix I need is the ability to multiply the matrices and . In particular, I only need a single pass over the entries of to compute the Nyström approximation. This allows the Nyström approximation to be used in settings when other low-rank approximations wouldn’t work, such as when is streamed to me as a sum of matrices that must be processed as they arrive and then discarded.
Choosing the Test Matrix
Every choice of test matrix defines a rank- Nyström approximation1Actually, is only rank at most . For this post, we will use rank- to mean “rank at most “. by (1). Unfortunately, the Nyström approximation won’t be a good low-rank approximation for every choice of . For an example of what can go wrong, if we pick to have columns selected from the eigenvectors of with small eigenvalues, the approximation will be quite poor.
The very best choice of would be the eigenvectors associated with the largest eigenvalues. Unfortunately, computing the eigenvectors to high accuracy is computationally costly. Fortunately, we can get decent low-rank approximations out of much simpler ‘s:
Random: Perhaps surprisingly, we get a fairly low-rank approximation out of just choosing to be a random matrix, say, populated with statistically independentstandard normal random entries. Intuitively, a random matrix is likely to have columns with meaningful overlap with the large-eigenvalue eigenvectors of (and indeed with any fixed orthonormal vectors). One can also pick moreexotickinds of random matrices, which can have computational benefits.
Random then improve: The more similar the test matrix is to the large-eigenvalue eigenvectors of , the better the low-rank approximation will be. Therefore, it makes sense to use the power method (usually called subspace iteration in this context) to improve a random initial test matrix to be closer to the eigenvectors of .2Even better than subspace iteration is block Krylov iteration. See section 11.6 of the following survey for details.
Column selection: If consists of columns of the identity matrix, then just consists of columns of : In MATLAB notation,
This is highly appealing as it allows us to approximate the matrix by only reading a small fraction of its entries (provided )! Producing a good low-rank approximation requires selecting the right column indices (usually under the constraint of reading a small number of entries from ). In my research with Yifan Chen, Joel A. Tropp, and Robert J. Webber, I’ve argued that the most well-rounded algorithm for this task is a randomly pivoted partial Cholesky decomposition.
The Projection Formula
Now that we’ve discussed the choice of test matrix, we shall explore the quality of the Nyström approximation as measured by the size of the residual . As a first step, we shall show that the residual is psd. This means that is an underapproximation to .
The positive semidefiniteness of the residual follows from the following projection formula for the Nyström approximation:
Here, denotes the the orthogonal projection onto the column space of the matrix . To deduce the projection formula, we break down as in (1):
The fact that the paranthesized quantity is can be verified in a variety of ways, such as by QR factorization.3Let , where has orthonormal columns and is square and upper triangular. The orthogonal projection is . The parenthesized expression is .
With the projection formula in hand, we easily obtain the following expression for the residual:
To show that this residual is psd, we make use of the conjugation rule.
Conjugation rule: For a matrix and a self-adjoint matrix , if is psd then is psd. If is invertible, then the converse holds: if is psd, then is psd.
The matrix is an orthogonal projection and therefore psd. Thus, by the conjugation rule, the residual of the is Nyström approximation is psd:
Optimality of the Nyström Approximation
There’s a question we’ve been putting off that can’t be deferred any longer:
Is the Nyström approximation actually a good low-rank approximation?
As we discussed earlier, the answer to this question depends on the test matrix . Different choices for give different approximation errors. See thefollowingpapers for Nyström approximation error bounds with different choices of . While the Nyström approximation can be better or worse depending on the choice of , what is true about Nyström approximation for every choice of is that:
The Nyström approximation is the best possible rank- approximation to given the information .
In precise terms, I mean the following:
Theorem: Out of all self-adjoint matrices spanned by the columns of with a psd residual , the Nyström approximation has the smallest error as measured by either the spectral or Frobeniusnorm (or indeed any unitarily invariant norm, see below).
Let’s break this statement down a bit. This result states that the Nyström approximation is the best approximation to under three conditions:
is self-adjoint.
is spanned by the columns of .
I find these first two requirements to be natural. Since is self-adjoint, it makes sense to require our approximation to be as well. The stipulation that is spanned by the columns seems like a very natural requirement given we want to think of approximations which only use the information . Additionally, requirement 2 ensures that has rank at most , so we are really only considering low-rank approximations to .
The last requirement is less natural:
The residual is psd.
This is not an obvious requirement to impose on our approximation. Indeed, it was a nontrivial calculation using the projection formula to show that the Nyström approximation itself satisfies this requirement! Nevertheless, this third stipulation is required to make the theorem true. The Nyström approximation (1) is the best “underapproximation” to the matrix to in the span of .
Intermezzo: Unitarily Invariant Norms and the Psd Order
To prove our theorem about the optimality of the Nyström approximation, we shall need two ideas from matrix theory: unitarily invariant norms and the psd order. We shall briefly describe each in turn.
A norm defined on the set of matrices is said to be unitarily invariant if the norm of a matrix does not change upon left- or right-multiplication by a unitary matrix:
Recall that a unitary matrix (called a real orthogonal matrix if is real-valued) is one that obeys . Unitary matrices preserve the Euclidean lengths of vectors, which makes the class of unitarily invariant norms highly natural. Important examples include the spectral, Frobenius, and nuclear matrix norms:
The unitarily invariant norm of a matrix depends entirely on its singular values . For instance, the spectral, Frobenius, and nuclear norms take the forms
In addition to being entirely determined by the singular values, unitarily invariant norms are non-decreasing functions of the singular values: If the th singular value of is larger than the th singular value of for , then for every unitarily invariant norm . For more on unitarily invariant norms, see this short and information-packed blog post from Nick Higham.
Our second ingredient is the psd order (also known as the Loewner order). A self-adjoint matrix is larger than a self-adjoint matrix according to the psd order, written , if the difference is psd. As a consequence, if and only if is psd, where here denotes the zero matrix of the same size as . Using the psd order, the positive semidefiniteness of the Nyström residual can be written as .
If and are both psd matrices and is larger than in the psd order, , it seems natural to expect that is larger than in norm. Indeed, this intuitive statement is true, at least when one restricts oneself to unitarily invariant norms.
Psd order and norms. If , then for every unitarily invariant norm .
This fact is a consequence of the following observations:
If , then the eigenvalues of are larger than in the sense that the th largest eigenvalue of is larger than the th largest eigenvalue of .
The singular values of a psd matrix are its eigenvalues.
Unitarily invariant norms are non-decreasing functions of the singular values.
Optimality of the Nyström Approximation: Proof
In this section, we’ll prove our theorem showing the Nyström approximation is the best low-rank approximation satisfying properties 1, 2, and 3. To this end, let be any matrix satisfying properties 1, 2, and 3. Because of properties 1 (self-adjointness) and 2 (spanned by columns of ), can be written in the form
where is a self-adjoint matrix. To make this more similar to the projection formula, we can factor on both sides to obtain
To make this more comparable to the projection formula, we can reparametrize by introducing a matrix with orthonormal columns with the same column space as . Under this parametrization, takes the form
The residual of this approximation is
(2)
We now make use of the of conjugation rule again. To simplify things, we make the assumption that is invertible. (As an exercise, see if you can adapt this argument to the case when this assumption doesn’t hold!) Since is psd (property 3), the conjugation rule tells us that
What does this observation tell us about ? We can apply the conjugation rule again to conclude
(Notice that since has orthonormal columns.)
We are now in a place to show that . Indeed,
The second line is the projection formula together with the observation that and the last line is the conjugation rule combined with the fact that is psd. Thus, having shown
This post is part of a new series for this blog, Note to Self, where I collect together some notes about an idea related to my research. This content may be much more technical than most of the content of this blog and of much less wide interest. My hope in sharing this is that someone will find this interesting and useful for their own work.
This post is about a fundamental tool of high-dimensional probability, the Hanson–Wright inequality. The Hanson–Wright inequality is a concentration inequality for quadratic forms of random vectors—that is, expressions of the form where is a random vector. Many statements of this inequality in the literature have an unspecified constant ; our goal in this post will be to derive a fairly general version of the inequality with only explicit constants.
The core object of the Hanson–Wright inequality is a subgaussian random variable. A random variable is subgaussian if the probability it exceeds a threshold in magnitude decays as
(1)
The name subgaussian is appropriate as the tail probabilities of Gaussian random variables exhibit the same square-exponential decrease .
A (non-obvious) fact is that if is subgaussian in the sense (1) and centered (), then ‘s cumulant generating function (cgf)
is subquadratic: There is a constant (independent of and ), for which
(2)
Moreover,1See Proposition 2.5.2 of Vershynin’sHigh-Dimensional Probability. a subquadratic cgf (2) also implies the subgaussian tail property (1), with a different parameter .
Since properties (1) and (2) are equivalent (up to a change in the parameter ), we are free to fix a version of property (2) as our definition for a (centered) subgaussian random variable.
Definition (subgaussian random variable): A centered random variable is said to be -subgaussian or subgaussian with variance proxy if its cgf is subquadratic:
and is thus subgaussian with variance proxy equal to its variance.
Here is a statement of the Hanson–Wright inequality as it typically appears with unspecified constants (see Theorem 6.2.1 of Vershynin’sHigh-Dimensional Probability):
Theorem (Hanson–Wright): Let be a random vector with independent centered -subgaussian entries and let be a square matrix. Then
where is a constant (not depending on , , , or ).2Here, and denote the Frobenius and spectral norms.
This type of concentration is exactly the same type as provided by Bernstein’s inequality (which I discussed in my post on concentration inequalities). In particular, for small deviations , the tail probabilities decay are subgaussian with variance proxy :
For large deviations , this switches to subexponential tail probabilities with decay rate :
Mediating these two parameter regimes are the size of the matrix , as measured by its Frobenius and spectral norms, and the degree of subgaussianity of , measured by the variance proxy .
Diagonal-Free Hanson–Wright
Now we come to a first version of the Hanson–Wright inequality with explicit constants, first for a matrix which is diagonal-free—that is, having all zeros on the diagonal. I obtained this version of the inequality myself, though I am very sure that this version of the inequality or an improvement thereof appears somewhere in the literature.
Theorem (Hanson–Wright, explicit constants, diagonal-free): Let random vector with independent centered -subguassian entries and let be a diagonal-free square matrix. Then we have the cgf bound
As a consequence, we have the concentration bound
Similarly, we have the lower tail
and the two-sided bound
Let us begin proving this result. Our proof will follow the same steps as Vershynin’s proof in High-Dimensional Probability (which in turn is adapted from an article by Rudelson and Vershynin), but taking care to get explicit constants. Unfortunately, proving all of the relevant tools from first principles would easily triple the length of this post, so I make frequent use of results from the literature.
We begin by the decoupling bound (Theorem 6.1.1 in Vershynin’sHigh-Dimensional Probability), which allows us to replace one with an independent copy at the cost of a factor of four:
(5)
We seek to compare the bilinear form to the Gaussian bilinear form where and are independent standard Gaussian vectors. We begin with the following cgf bound for the Gaussian quadratic form :
We now seek to compare to . To do this, we first evaluate the cgf of only over the randomness in . Since we’re only taking an expectation over the random variable , we can apply the subquadratic tail condition (3) to obtain
(7)
Now we perform a similar computation for the quantity in which has been replaced by the Gaussian vector :
We stress that this is an equality since the cgf of a Gaussian random variable is given by (4). Thus we can substitute the left-hand side of the above display into the right-hand side of (7), yielding
(8)
We now perform this same trick again using the randomness in :
(9)
Packaging up (8) and (9) gives
(10)
Combining all these results (5), (6), and (10), we obtain
Fact (Bernstein concentration from Bernstein cgf bound): Suppose that a random variable satisfies the cgf bound for . Then
To get the bound on the lower tail, apply the result for the upper tail to the matrix to obtain
Finally, to obtain the two-sided bound, use a union bound over the upper and lower tails:
General Hanson–Wright
Now, here’s a more general result (with worse constants) which permits the matrix to possess a diagonal.
Theorem (Hanson–Wright, explicit constants): Let random vector with independent centered -subguassian entries and let be an arbitrary square matrix. Then we have the cgf bound
As a consequence, we have the concentration bound
Left tail and two-sided bounds versions of this bound also hold:
and
Decompose the matrix into its diagonal and off-diagonal portions. For any two random variables and (possibly highly dependent), we can bound the cgf of their sum using the following “union bound”:
(11)
The two equality statements are the definition of the cumulant generating function and the inequality is Cauchy–Schwarz.
Using the “union bound”, it is sufficient to obtain bounds for the cgfs of the diagonal and off-diagonal parts and . We begin with the diagonal part. We compute
For the second inequality, we used the facts that and .
We now look at the off-diagonal part . We use a version of the decoupling bound (5) where we compare to , where we’ve both replaced one copy of with an independent copy and reinstated the diagonal of (see Remark 6.1.3 in Vershynin’sHigh-Dimensional Probability):
We can now just repeat the rest of the argument for the diagonal-free Hanson–Wright inequality, yielding the same conclusion
(14)
Combining (11), (13), and (14), we obtain
As with above, this cgf bound implies the desired probability bound.
Let’s start with a classical problem: connect-the-dots. As we know from geometry, any two points in the plane are connected by one and only one straight line:
But what if we have more than two points? How should we connect them? One natural way is by parabola. Any three points (with distinct coordinates) are connected by one and only one parabola :
And we can keep extending this. Any points1The degree of the polynomial is one less than the number of points because a degree- polynomial is described by coefficients. For instance, a degree-two parabola has three coefficients , , and . (with distinct coordinates) are connected by a unique degree- polynomial :
This game of connect-the-dots with polynomials is known more formally as polynomial interpolation. We can use polynomial interpolation to approximate functions. For instance, we can approximate the function on the interval to visually near-perfect accuracy by connecting the dots between seven points :
But something very peculiar happens when we try and apply this trick to the specially chosen function on the interval :
Unlike , the polynomial interpolant for is terrible! What’s going on? Why doesn’t polynomial interpolation work here? Can we fix it? The answer to the last question is yes and the solution is Chebyshev polynomials.
Reverse-Engineering Chebyshev
The failure of polynomial interpolation for is known as Runge’s phenomenon after Carl Runge who discovered this curious behavior in 1901. The function is called the Runge function. Our goal is to find a fix for polynomial interpolation which crushes the Runge phenomenon, allowing us to reliably approximate every sensible2A famous theorem of Faber states that there does not exist any set of points through which the polynomial interpolants converge for every continuous function. This is not as much of a problem as it may seem. As the famous Weierstrass function shows, arbitrary continuous functions can be very weird. If we restrict ourselves to nicer functions, such as Lipschitz continuous functions, there does exist a set of points through which the polynomial interpolant always converges to the underlying function. Thus, in this senses, it is possible to crush the Runge phenomenon. function with polynomial interpolation.
Carl Runge
Let’s put on our thinking caps and see if we can discover the fix for ourselves. In order to discover a fix, we must first identify the problem. Observe that the polynomial interpolant is fine near the center of the interval; it only fails near the boundary.
This leads us to a guess for what the problem might be; maybe we need more interpolation points near the boundaries of the interval. Indeed, tipping our hand a little bit, this turns out to be the case. For instance, connecting the dots for the following set of “mystery points” clustered at the endpoints works just fine:
Let’s experiment a little and see if we can discover a nice set of interpolation points, which we will call , like this for ourselves. We’ll assume the interpolation points are given by a function so we can form the polynomial interpolant for any desired polynomial degree .3Technically, we should insist on the function being \textit{injective} so that the points are guaranteed to be distinct. For instance, if we pick , the points look like this:
Equally spaced points (shown on vertical axis) give rise to non-equally spaced points (shown on horizontal axis)
How should we pick the function ? First observe that, even for the Runge function, equally spaced interpolation points are fine near the center of the interval. We thus have at least two conditions for our desired interpolation points:
The interior points should maintain their spacing of roughly .
The points must cluster near both boundaries.
As a first attempt let’s divide the interval into thirds and halve the spacing of points except in the middle third. This leads to the function
These interpolation points initially seem promising, even successfully approximating the Runge function itself.
Unfortunately, this set of points fails when we consider other functions. For instance, if we use the Runge-like function , we see that these interpolation points now lead to a failure to approximate the function at the middle of the interval, even if we use a lot of interpolation points!
Maybe the reason this set of interpolation points didn’t work is that the points are too close at the endpoints. Or maybe we should have divided the interval as quarter–half–quarter rather than thirds. There are lots of variations of this strategy for choosing points to explore and all of them eventually lead to failure on some Runge-flavored example. We need a fundamentally different strategy then making the points times closer within distance of the endpoints.
Let’s try a different approach. The closeness of the points at the endpoints is determined by the slope of the function at and . The smaller that and are, the more clustered the points will be. For instance,
When we halved the distance between points, we instead had
So if we want the points to be much more clustered together, it is natural to require
It also makes sense for the function to cluster points equally near both endpoints, since we see no reason to preference one end over the other. Collecting together all the properties we want the function to have, we get the following list:
spans the whole range ,
, and
is symmetric about , .
Mentally scrolling through our Rolodex of friendly functions, a natural one that might come to mind meeting these three criteria is the cosine function, specifically . This function yields points which are more clustered at the endpoints:
The points
we guessed our way into are known as the Chebyshev points.4Some authors refer to these as the “Chebyshev points of the second kind” or use other names. We follow the convention in Approximation Theory and Approximation Practice (Chapter 1) and simply refer to these points simply as the Chebyshev points. The Chebyshev points prove themselves perfectly fine for the Runge function:
As we saw earlier, success on the Runge function alone is not enough to declare victory for the polynomial interpolation problem. However, in this case, there are no other bad examples left to find. For any nice function with no jumps, polynomial interpolation through the Chebyshev points works excellently.5Specifically, for a function which not too rough (i.e., Lipschitz continuous), the degree- polynomial interpolant of through the Chevyshev points converges uniformly to as .
Why the Chebyshev Points?
We’ve guessed our way into a solution to the polynomial interpolation problem, but we still really don’t know what’s going on here. Why are the Chebyshev points much better at polynomial interpolation than equally spaced ones?
Now that we know that the Chebyshev points are a right answer to the interpolation problem,6Indeed, there are other sets of interpolation points through which polynomial interpolation also works well, such as the Legendre points. let’s try and reverse engineer a principled reason for why we would expect them to be effective for this problem. To do this, we ask:
What is special about the cosine function?
From high school trigonometry, we know that gives the coordinate of a point radians along the unit circle. This means that the Chebyshev points are the coordinates of equally spaced points on the unit circle (specifically the top half of the unit circle ).
Chebyshev points are the coordinates of equally spaced points on the unit circle.
This raises the question:
What does the interpolating polynomial look like as a function of the angle ?
To convert between and we simply plug in to :
This new function depending on , which we can call , is a polynomial in the variable . Powers of cosines are not something we encounter every day, so it makes sense to try and simplify things using some trig identities. Here are the first couple powers of cosines:
A pattern has appeared! The powers always take the form7As a fun exercise, you might want to try and prove this using mathematical induction.
The significance of this finding is that, by plugging in each of these formulas for , we see that our polynomial in the variable has morphed into a Fourier cosine series in the variable :
For anyone unfamiliar with Fourier series, we highly encourage the 3Blue1Brown video on the subject, which explains why Fourier series are both mathematically beautiful and practically useful. The basic idea is that almost any function can be expressed as a combination of waves (that is, sines and cosines) with different frequencies.8More precisely, we might call these angular frequencies. In our case, this formula tells us that is equal to units of frequency , plus units of frequency , all the way up to units of frequency . Different types of Fourier series are appropriate in different contexts. Since our Fourier series only possesses cosines, we call it a Fourier cosine series.
We’ve discovered something incredibly cool:
Polynomial interpolation through the Chebyshev points is equivalent to finding a Fourier cosine series for equally spaced angles .
We’ve arrived at an answer to why the Chebyshev points work well for polynomial interpolation.
Polynomial interpolation through the Chebyshev points is effective because Fourier cosine series through equally spaced angles is effective.
Of course, this explanation just raises the further question: Why do Fourier cosine series give effective interpolants through equally spaced angles ? This question has a natural answer as well, involving the convergence theory and aliasing formula (see Section 3 of this paper) for Fourier series. We’ll leave the details to the interested reader for investigation. The success of Fourier cosines series in interpolating equally spaced data is a fundamental observation that underlies the field of digital signal processing. Interpolation through the Chebyshev points effectively hijacks this useful fact and applies it to the seemingly unrelated problem of polynomial interpolation.
Another question this explanation raises is the precise meaning of “effective”. Just how good are polynomial interpolants through the Chebyshev points at approximating functions? As is discussed at length in another post on this blog, the degree to which a function can be effectively approximated is tied to how smooth or rough it is. Chebyshev interpolants approximate nice analytic functions like or with exponentially small errors in the number of interpolation points used. By contrast, functions with kinks like are approximated with errors which decay much more slowly. See theorems 2 and 3 on this webpage for more details.
Chebyshev Polynomials
We’ve now discovered a set of points, the Chebyshev points, through which polynomial interpolation works well. But how should we actually compute the interpolating polynomial
Again, it will be helpful to draw on the connection to Fourier series. Computations with Fourier series are highly accurate and can be made lightning fast using the fast Fourier transform algorithm. By comparison, directly computing with a polynomial through its coefficients is a computational nightmare.
In the variable , the interpolant takes the form
To convert back to , we use the inverse function9One always has to be careful when going from to since multiple values get mapped to a single value by the cosine function. Fortunately, we’re working with variables and , between which the cosine function is one-to-one with the inverse function being given by the arccosine. to obtain:
This is a striking formula. Given all of the trigonometric functions, it’s not even obvious that is a polynomial (it is)!
Despite its seeming peculiarity, this is a very powerful way of representing the polynomial . Rather than expressing using monomials , we’ve instead written as a combination of more exotic polynomials
The polynomials are known as the Chebyshev polynomials,10More precisely, the polynomials are known as the Chebyshev polynomials of the first kind. named after Pafnuty Chebyshev who studied the polynomials intensely.11The letter “T” is used for Chebyshev polynomials since the Russian name “Chebyshev” is often alternately transliterated to English as “Tchebychev”.
Pafnuty Chebyshev
Writing out the first few Chebyshev polynomials shows they are indeed polynomials:
Since and are both polynomials, every Chebyshev polynomial is as well.
We’ve arrived at the following amazing conclusion:
Under the change of variables , the Fourier cosine series
becomes the combination of Chebyshev polynomials
This simple and powerful observations allows us to apply the incredible speed and accuracy of Fourier series to polynomial interpolation.
Beyond being a neat idea with some nice mathematics, this connection between Fourier series and Chebyshev polynomials is a powerful tool for solving computational problems. Once we’ve accurately approximated a function by a polynomial interpolant, many quantities of interest (derivatives, integrals, zeros) become easy to compute—after all, we just have to compute them for a polynomial! We can also use Chebyshev polynomials to solve differential equations with much faster rates of convergence than other methods. Because of the connection to Fourier series, all of these computations can be done to high accuracy and blazingly fast via the fast Fourier transform, as is done in the software package Chebfun.
The Chebyshev polynomials have an array of amazing properties and they appear all over mathematics and its applications in other fields. Indeed, we have only scratched the surface of the surface. Many questions remain:
What is the connection between the Chebyshev points and the Chebyshev polynomials?
The cosine functions are orthogonal to each other; are the Chebyshev polynomials?
Are the Chebyshev points the best points for polynomial interpolation? What does “best” even mean in this context?
Every “nice” even periodic function has an infinite Fourier cosine series which converges to it. Is there a Chebyshev analog? Is there a relation between the infinite Chebyshev series and the (finite) interpolating polynomial through the Chebyshev points?
All of these questions have beautiful and fairly simple answers. The book Approximation Theory and Approximation Practice is a wonderfully written book that answers all of these questions in its first six chapters, which are freely available on the author’s website. We recommend the book highly to the curious reader.
TL;DR: To get an accurate polynomial approximation, interpolate through the Chebyshev points. To compute the resulting polynomial, change variables to , compute the Fourier cosine series interpolant, and obtain your polynomial interpolant as a combination of Chebyshev polynomials.
This post is about randomized algorithms for problems in computational science and a powerful set of tools, known as concentration inequalities, which can be used to analyze why they work. I’ve discussed why randomization can help in solving computational problems in a previous post; this post continues this discussion by presenting an example of a computational problem where, somewhat surprisingly, a randomized algorithm proves effective. We shall then use concentration inequalities to analyze why this method works.
Triangle Counting
Let’s begin our discussion of concentration inequalities by means of an extended example. Consider the following question: How many triangles are there in the Facebook network? That is, how many trios of people are there who are all mutual friends? While seemingly silly at first sight, this is actually a natural and meaningful question about the structure of the Facebook social network and is related to similar questions such as “How likely are two friends of a person to also be friends with each other?”
If there are people on the Facebook graph, then the natural algorithm of iterating over all triplets and checking whether they form a triangle is far too computationally costly for the billions of Facebook accounts. Somehow, we want to do much faster than this, and to achieve this speed we would be willing to settle for an estimate of the triangle count up to some error.
There are many approaches to this problem, but let’s describe a particularly surprising algorithm. Let be an matrix where the th entry of is if users and are friends and otherwise1All of the diagonal entries of are set to zero.; this matrix is called the adjacency matrix of the Facebook graph. A fact from graph theory is that the th entry of the cube of the matrix counts the number of paths from user to user of length three.2By a path of length three, we mean a sequence of users where and , and , and and are all friends. In particular, the th entry of denotes the number of paths from to itself of length , which is twice the number of triangles incident on . (The paths and are both counted as paths of length 3 for a triangle consisting of , , and .) Therefore, the trace of , equal to the sum of its diagonal entries, is six times the number of triangles: The th entry of is twice the number of triangles incident on and each triangle is counted thrice in the th, th, and th entries of . In summary, we have
Therefore, the triangle counting problem is equivalent to computing the trace of . Unfortunately, the problem of computing is, in general, very computationally costly. Therefore, we seek ways of estimating the trace of a matrix without forming it.
Randomized Trace Estimation
Motivated by the triangle counting problem from the previous section, we consider the problem of estimating the trace of a matrix . We assume that we only have access to the matrix through matrix–vector products; that is, we can efficiently compute for a vector . For instance, in the previous example, the Facebook graph has many fewer friend relations (edges) than the maximum possible amount of . Therefore, the matrix is sparse; in particular, matrix–vector multiplications with can be computed in around operations. To compute matrix–vector products with , we simply compute matrix–vector products with three times, .
Here’s a very nifty idea to estimate the trace of using only matrix–vector products, originally due to Didier A. Girard and Michael F. Hutchinson. Choose to be a random vector whose entries are independent -values, where each value and occurs with equal probability. Then if one forms the expression . Since the entries of and are independent, the expectation of is for and for . Consequently, by linearity of expectation, the expected value of is
The average value of is equal to the trace of ! In the language of statistics, we might say that is an unbiased estimator for . Thus, the efficiently computable quantity can serve as a (crude) estimate for .
While the expectation of equals , any random realization of can deviate from by a non-neligible amount. Thus, to reduce the variability of the estimator , it is appropriate to take an average of multiple copies of this random estimate. Specifically, we draw random vectors with independent random entries and compute the averaged trace estimator
(1)
The -sample trace estimator remains an unbiased estimator for , , but with reduced variability. Quantitatively, the variance of is times smaller than the single-sample estimator :
(2)
The Girard–Hutchinson trace estimator gives a natural way of estimating the trace of the matrix , a task which might otherwise be hard without randomness.3To illustrate what randomness is buying us here, it might be instructive to think about how one might try to estimate the trace of via matrix–vector products without the help of randomness. For the trace estimator to be a useful tool, an important question remains: How many samples are needed to compute to a given accuracy? Concentration inequalities answer questions of this nature.
Concentration Inequalities
A concentration inequality provides a bound on the probability a random quantity is significantly larger or smaller than its typical value. Concentration inequalities are useful because they allow us to prove statements like “With at least 99% probability, the randomized trace estimator with 100 samples produces an approximation of the trace which is accurate up to error no larger than .” In other words, concentration inequalities can provide quantitative estimates of the likely size of the error when a randomized algorithm is executed.
In this section, we shall introduce a handful of useful concentration inequalities, which we will apply to the randomized trace estimator in the next section. We’ll then discuss how these and other concentration inequalities can be derived in the following section.
Markov’s Inequality
Markov’s inequality is the most fundamental concentration inequality. When used directly, it is a blunt instrument, requiring little insight to use and producing a crude but sometimes useful estimate. However, as we shall see later, all of the sophisticated concentration inequalities that will follow in this post can be derived from a careful use of Markov’s inequality.
The wide utility of Markov’s inequality is a consequence of the minimal assumptions needed for its use. Let be any nonnegative random variable. Markov’s inequality states that the probability that exceeds a level is bounded by the expected value of over . In equations, we have
(3)
We stress the fact that we make no assumptions on how the random quantity is generated other than that is nonnegative.
As a short example of Markov’s inequality, suppose we have a randomized algorithm which takes one second on average to run. Markov’s inequality then shows that the probability the algorithm takes more than 100 seconds to run is at most . This small example shows both the power and the limitation of Markov’s inequality. On the negative side, our analysis suggests that we might have to wait as much as 100 times the average runtime for the algorithm to complete running with 99% probability; this large huge multiple of 100 seems quite pessimistic. On the other hand, we needed no information whatsoever about how the algorithm works to do this analysis. In general, Markov’s inequality cannot be improved without more assumptions on the random variable .4For instance, imagine an algorithm which 99% of the time completes instantly and 1% of the time takes 100 seconds. This algorithm does have an average runtime of 1 second, but the conclusion of Markov’s inequality that the runtime of the algorithm can be as much as 100 times the average runtime with 1% probability is true.
Chebyshev’s Inequality and Averages
The variance of a random variable describes the expected size of a random variable’s deviation from its expected value. As such, we would expect that the variance should provide a bound on the probability a random variable is far from its expectation. This intuition indeed is correct and is manifested by Chebyshev’s inequality. Let be a random variable (with finite expected value) and . Chebyshev’s inequality states that the probability that deviates from its expected value by more than is at most :
(4)
Chebyshev’s inequality is frequently applied to sums or averages of independent random quantities. Suppose are independent and identically distributed random variables with mean and variance and let denote the average
Since the random variables are independent,5In fact, this calculation works if are only pairwise independent or even pairwise uncorrelated. For algorithmic applications, this means that don’t have to be fully independent of each other; we just need any pair of them to be uncorrelated. This allows many randomized algorithms to be “derandomized“, reducing the amount of “true” randomness needed to execute an algorithm. the properties of variance entail that
where we use the fact that . Therefore, by Chebyshev’s inequality,
(5)
Suppose we want to estimate the mean by up to error and are willing to tolerate a failure probability of . Then setting the right-hand side of (5) to , Chebyshev’s inequality suggests that we need at most
(6)
samples to achieve this goal.
Exponential Concentration: Hoeffding and Bernstein
How happy should we be with the result (6) of applying Chebyshev’s inequality the average ? The central limit theorem suggests that should be approximately normally distributed with mean and variance . Normal random variables have an exponentially small probability of being more than a few standard deviations above their mean, so it is natural to expect this should be true of as well. Specifically, we expect a bound roughly like
(7)
Unfortunately, we don’t have a general result quite this nice without additional assumptions, but there are a diverse array of exponential concentration inequalities available which are quite useful in analyzing sums (or averages) of independent random variables that appear in applications.
Hoeffding’s inequality is one such bound. Let be independent (but not necessarily identically distributed) random variables and consider the average . Hoeffding’s inequality makes the assumption that the summands are bounded, say within an interval .6There are also more general versions of Hoeffding’s inequality where the bound on each random variable is different. Hoeffding’s inequality then states that
(8)
Hoeffding’s inequality is quite similar to the ideal concentration result (7) except with the variance replaced by the potentially much larger quantity7Note that is always smaller than or equal to . .
Bernstein’s inequality fixes this deficit in Hoeffding’s inequality at a small cost. Now, instead of assuming are bounded within the interval , we make the alternate boundedness assumption for every . We continue to denote so that if are identically distributed, denotes the variance of each of . Bernstein’s inequality states that
(9)
For small values of , Bernstein’s inequality yields exactly the kind of concentration that we would hope for from our central limit theorem heuristic (7). However, for large values of , we have
which is exponentially small in rather than . We conclude that Bernstein’s inequality provides sharper bounds then Hoeffding’s inequality for smaller values of but weaker bounds for larger values of .
Chebyshev vs. Hoeffding vs. Bernstein
Let’s return to the situation where we seek to estimate the mean of independent and identically distributed random variables each with variance by using the averaged value . Our goal is to bound how many samples we need to estimate up to error , , except with failure probability at most . Using Chebyshev’s inequality, we showed that (see (7))
Now, let’s try using Hoeffding’s inequality. Suppose that are bounded in the interval . Then Hoeffding’s inequality (8) shows that
Bernstein’s inequality states that if lie in the interval for every , then
(10)
Hoeffding’s and Bernstein’s inequality show that we need roughly proportional to samples are needed rather than proportional to . The fact that we need proportional to samples to achieve error is a consequence of the central limit theorem and is something we would not be able to improve with any concentration inequality. What exponential concentration inequalities allow us to do is to improve the dependence on the failure probability from proportional to to , which is a huge improvement.
Hoeffding’s and Bernstein’s inequalities both have a small drawback. For Hoeffding’s inequality, the constant of proportionality is rather than the true variance of the summands. Bernstein’s inequality gives us the “correct” constant of proportionality but adds a second term proportional to ; for small values of , this term is dominated by the term proportional to but the second term can be relevant for larger values of .
There are a panoply of additional concentration inequalities than the few we’ve mentioned. We give a selected overview in the following optional section.
Other Concentration Inequalities
There are a handful more exponential concentration inequalities for sums of independent random variables such as Chernoff’s inequality (very useful for somes of bounded, positive random variables) and Bennett’s inequality. There are also generalizations of Hoeffding’s, Chernoff’s, and Bernstein’s inequalities for unbounded random variables with subgaussian and subexponential tail decay; these results are documented in Chapter 2 of Roman Vershynin’s excellent book High-Dimensional Probability.
One can also generalize concentration inequalities to so-called martingale sequences, which can be very useful for analyzing adaptive algorithms. These inequalities can often have the advantage of bounding the probability that a martingale sequence ever deviates by some amount from its applications; these results are called maximal inequalities. Maximal analogs of Markov’s and Chebyshev’s inequalities are given by Ville’s inequality and Doob’s inequality. Exponential concentration inequalities include the Hoeffding–Azuma inequality and Freedman’s inequality.
Let us apply some of the concentration inequalities we introduced in last section to analyze the randomized trace estimator. Our goal is not to provide the best possible analysis of the trace estimator,8More precise estimation for trace estimation applied to positive semidefinite matrices was developed by Gratton and Titley-Peloquin; see Theorem 4.5 of the following survey. but to demonstrate how the general concentration inequalities we’ve developed can be useful “out of the box” in analyzing algorithms.
In order to apply Chebyshev’s and Berstein’s inequalities, we shall need to compute or bound the variance of the single-sample trace estimtor , where is a random vector of independent -values. This is a straightforward task using properties of the variance:
where denotes the matrix spectral norm. Therefore, by Bernstein’s inequality (9), we have
In particular, (10) shows that
samples suffice to estimate to error with failure probability at most . Concentration inequalities easily furnish estimates for the number of samples needed for the randomized trace estimator.
We have now accomplished our main goal of using concentration inequalities to analyze the randomized trace estimator, which in turn can be used to solve the triangle counting problem. We leave some additional comments on trace estimation and triangle counting in the following bonus section.
More on Trace Estimation and Triangle Counting
To really complete the analysis of the trace estimator in an application (e.g., triangle counting), we would need to obtain bounds on and . Since we often don’t know good bounds for and , one should really use the trace estimator together with an a posteriori error estimates for the trace estimator, which provide a confidence interval for the trace rather than a point estimate; see sections 4.5 and 4.6 in this survey for details.
Several algorithms have been proposed for triangle counting, many of them randomized. This survey gives a comparison of different methods for the triangle counting problem, and also describes more motivation and applications for the problem.
Deriving Concentration Inequalities
Having introduced concentration inequalities and applied them to the randomized trace estimator, we now turn to the question of how to derive concentration inequalities. Learning how to derive concentration inequalities is more than a matter of mathematical completeness since one can often obtain better results by “hand-crafting” a concentration inequality for a particular application rather than applying a known concentration inequality. (Though standard concentration inequalities like Hoeffding’s and Bernstein’s often give perfectly adequate answers with much less work.)
Markov’s Inequality
At the most fundamental level, concentration inequalities require us to bound a probability by an expectation. In achieving this goal, we shall make a simple observation: The probability that is larger than or equal to is the expectation of a random variable .9More generally, the probability of an event can be written as an expectation of the indicator random variable of that event. Here, is an indicator function which outputs one if its input is larger than or equal to and zero otherwise.
As promised, the probability is larger than is the expectation of :
(11)
We can now obtain bounds on the probability that by bounding its corresponding indicator function. In particular, we have the inequality
(12)
Since is nonnegative, combining equations (11) and (12) gives Markov’s inequality:
Chebyshev’s Inequality
Before we get to Chebyshev’s inequality proper, let’s think about how we can push Markov’s inequality further. Suppose we find a bound on the indicator function of the form
(13)
A bound of this form immediately to bounds on by (11). To obtain sharp and useful bounds on we seek bounding functions in (13) with three properties:
For , should be close to zero,
For , should be close to one, and
We need to be easily computable or boundable.
These three objectives are in tension with each other. To meet criterion 3, we must restrict our attention to pedestrian functions such as powers or exponentials for which we have hopes of computing or bounding for random variables we encounter in practical applications. But these candidate functions have the undesirable property that making the function smaller on (by increasing ) to meet point 1 makes the function larger on , detracting from our ability to achieve point 2. We shall eventually come up with a best-possible resolution to this dilemma by formulating this as an optimization problem to determine the best choice of the parameter to obtain the best possible candidate function of the given form.
Before we get ahead of ourselves, let us use a specific choice for different than we used to prove Markov’s inequality. We readily verify that satisfies the bound (13), and thus by (12),
(14)
This inequality holds for any nonnegative random variable . In particular, now consider a random variable which we do not assume to be nonnegative. Then ‘s deviation from its expectation, , is a nonnegative random variable. Thus applying (14) gives
We have derived Chebyshev’s inequality! Alternatively, one can derive Chebyshev’s inequality by noting that if, and only if, . Therefore, by Markov’s inequality,
The Laplace Transform Method
We shall now realize the plan outlined earlier where we shall choose an optimal bounding function from the family of exponential functions , where is a parameter which we shall optimize over. This method shall allow us to derive exponential concentration inequalities like Hoeffding’s and Bernstein’s. Note that the exponential function bounds the indicator function for all real numbers , so we shall no longer require the random variable to be nonnegative. Therefore, by (11),
The moment generating function coincides with the Laplace transform up to the sign of the parameter , so one name for this approach to deriving concentration inequalities is the Laplace transform method. (This method is also known as the Cramér–Chernoff method.)
The cumulant generating function has an important property for deriving concentration inequalities for sums or averages of independent random variables: If are independent random variables, than the cumulant generating function is additive:11For proof, we compute . Taking logarithms proves the additivity.
(17)
Proving Hoeffding’s Inequality
For us to use the Laplace transform method, we need to either compute or bound the cumulant generating function. Since we are interested in general concentration inequalities which hold under minimal assumptions such as boundedness, we opt for the latter. Suppose and consider the cumulant generating function of . Then one can show the cumulant generating function bound12The bound (18) is somewhat tricky to establish, but we can establish the same result with a larger constant than . We have . Since the function is convex, we have the bound . Taking expectations, we have . One can show by comparing Taylor series that . Therefore, we have .
(18)
Using the additivity of the cumulant generating function (17), we obtain the bound
Plugging this into the probability bound (16), we obtain the concentration bound
(19)
We want to obtain the smallest possible upper bound on this probability, so it behooves us to pick the value of which makes the right-hand side of this inequality as small as possible. To do this, we differentiate the contents of the exponential and set to zero, obtaining
Plugging this value for into the bound (19) gives A bound for being larger than :
(20)
To get the bound on being smaller than , we can apply a small trick. If we apply (20) to the summands instead of , we obtain the bounds
(21)
We can now combine the upper tail bound (19) with the lower tail bound (21) to obtain a “symmetric” bound on the probability that . The means of doing often this goes by the fancy name union bound, but the idea is very simple:
Thus, applying this union bound idea with the upper and lower tail bounds (20) and (21), we obtain Hoeffding’s inequality, exactly as it appeared above as (8):
Voilà! Hoeffding’s inequality has been proven! Bernstein’s inequality is proven essentially the same way except that, instead of (17), we have the cumulant generating function bound
for a random variable with mean zero and satisfying the bound .
Upshot: Randomness can be a very effective tool for solving computational problems, even those which seemingly no connection to probability like triangle counting. Concentration inequalities are a powerful tool for assessing how many samples are needed for an algorithm based on random sampling to work. Some of the most useful concentration inequalities are exponential concentration inequalities like Hoeffding and Bernstein, which show that an average of bounded random quantities are close to their average except with exponentially small probability.
The (ordinary) linear least squares problem is as follows: given an matrix and a vector of length , find the vector such that is as close to as possible, when measured using the two-norm . That is, we seek to
(1)
From this equation, the name “least squares” is self-explanatory: we seek which minimizes the sum of the squared discrepancies between the entries of and .
The least squares problem is ubiquitous in science, engineering, mathematics, and statistics. If we think of each row of as an input and its corresponding entry of as an output, then the solution to the least squares model gives the coefficients of a linear model for the input–output relationship. Given a new previously unseen input , our model predicts the output is approximately . The vector consists of coefficients for this linear model. The least squares solution satisfies the property that the average squared difference between the output and the prediction is as small as it could possibly be for all choices of coefficient vectors .
How do we solve the least squares problem? A classical solution approach, ubiquitous in textbooks, is to solve a system of linear equations known as the normal equations. The normal equations associated with the least squares problem (1) are given by
(2)
This system of equations always has a solution. If has full column-rank, then is invertible and the unique least squares solution to (1) is given by . We assume that has full column-rankQ for the rest of this discussion. To solve the normal equations in software, we compute and and solve (2) using a linear solver like MATLAB’s “\”.1Even better, we could us a Cholesky decomposition since the matrix is positive definite. (As is generally true in matrix computations, it is almost never a good idea to explicitly form the inverse of the matrix , or indeed any matrix.) We also can solve the normal equations using an iterative method like (preconditioned) conjugate gradient.
The purpose of the article is to advocate against the use of the normal equations for solving the least squares problems, at least in most cases. So what’s wrong with the normal equations? The problem is not that the normal equations aren’t mathematically correct. Instead, the problem is that the normal equations often lead to poor accuracy for the least squares solution using computer arithmetic.
Most of the time when using computers, we store real numbers as floating point numbers.2One can represent rational numbers on a computer as fractions of integers and operations can be done exactly. However, this is prone to gross inefficiencies as the number of digits in the rational numbers can grow to be very large, making the storage and time to solve linear algebra problems with rationals dramatically more expensive. For these reasons, the vast majority of numerical computations use floating point numbers which store only a finite number of digits for any given real number. In this model, except for extremely rare circumstances, rounding errors during arithmetic operations are a fact of life. At a coarse level, the right model to have in your head is that real numbers on a computer are stored in scientific notation with only 16 decimal digits after the decimal point.3This is a simplification in multiple ways. First, computers store numbers in binary and thus, rather than storing 16 decimal digits after the decimal point, they store 52 binary digits. This amounts to roughly 16 decimal digits. Secondly, there are different formats for storing real numbers as floating point on a computer with different amounts of stored digits. The widely used IEEE double precision format has about 16 decimal digits of accuracy; the IEEE single precision format has roughly 8. When two numbers are added, subtracted, multiplied, and divided, the answer is computed and then rounded to 16 decimal digits; any extra digits of information are thrown away. Thus, the result of our arithmetic on a computer is the true answer to the arithmetic problem plus a small rounding error. These rounding errors are small individually, but solving an even modestly sized linear algebra problem requires thousands of such operations. Making sure many small errors don’t pile up into a big error is part of the subtle art of numerical computation.
To make a gross simplification, if one solves a system of linear equations on a computer using a well-designed piece of software, one obtains an approximate solution which is, after accounting for the accumulation of rounding errors, close to . But just how close the computed solution and the true solution are depends on how “nice” the matrix is. The “niceness” of a matrix is quantified by a quantity known as the condition number of , which we denote .4In fact, there are multiple definitions of the condition number depending on the norm which one uses the measure the sizes of vectors. Since we use the 2-norm, the appropriate 2-norm condition number is the ratio of the largest and smallest singular values of . As a rough rule of thumb, the relative error between and is roughly bounded as
(3)
The “ corresponds to the fact we have roughly 16 decimal digits of accuracy in double precision floating point arithmetic. Thus, if the condition number of is roughly , then we should expect around digits of accuracy in our computed solution.
The accuracy of the least squares problem is governed by its own condition number . We would hope that we can solve the least squares problem with an accuracy like the rule-of-thumb error bound (3) we had for linear systems of equations, namely a bound like . But this is not the kind of accuracy we get for the least squares problem when we solve it using the normal equations. Instead, we get accuracy like
(4)
By solving the normal equations we effectively square the condition number! Perhaps this is not surprising as the normal equations also more-or-less square the matrix by computing . This squared condition number drastically effects the accuracy of the computed solution. If the condition number of is , then the normal equations give us absolute nonsense for ; we expect to get no digits of the answer correct. Contrast this to above, where we were able to get correct digits in the solution to despite the condition number of being times larger than !
All of this would be just a sad fact of life for the least squares problem if the normal equations and their poor accuracy properties were the best we could do for the least squares problem. But we can do better! One can solve linear least squares problems by computing a so-called QR factorization of the matrix .5In MATLAB, the least squares problem can be solved with QR factorization by calling “A\b”. Without going into details, the upshot is that the least squares solution by a well-designed6One way of computing the QR factorization is by Gram–Schmidt orthogonalization, but the accuracy properties of this are poor too. A gold-standard way of computing the QR factorization by means of Householder reflectors, which has excellent accuracy properties. QR factorization requires a similar amount of time to solving the normal equations and has dramatically improved accuracy properties, achieving the desirable rule-of-thumb behavior7More precisely, the rule of thumb is like . So even if we solve the least squares problem with QR factorization, we still get a squared condition number in our error bound, but this condition number squared is multiplied by the residual , which is small if the least squares fit is good. The least squares solution is usually only interesting when the residual is small, thus justifying dropping it in the rule of thumb.
(5)
I have not described how the QR factorization is accurately computed nor how to use the QR factorization to solve least squares problems nor even what the QR factorization is. All of these topics are explained excellently by thestandardtextbooks in this area, as well as by publicly available resources like Wikipedia. There’s much more that can be said about the many benefits of solving the least squares problem with the QR factorization,8E.g., it can work for sparse matrices while the normal equations often do not, it has superior accuracy to Gaussian elimination with partial pivoting even for solving linear systems, the “” matrix in the QR factorization can be represented implicitly as a product of easy-to-compute-with Householder reflectors which is much more efficient when, etc. but in the interest of brevity let me just say this: TL;DR when presented in the wild with a least squares problem, the solution method one should default to is one based on a well-implemented QR factorization, not solving the normal equations.
Suppose for whatever reason we don’t have a high quality QR factorization algorithm at our disposal. Must we then resort to the normal equations? Even in this case, there is a way we can reduce the problem of solving a least squares problems to a linear system of equations without squaring the condition number! (For those interested, to do this, we recognize the normal equations as a Schur complement of a somewhat larger system of linear equations and then solve that. See Eq. (7) in this post for more discussion of this approach.)
The title of this post Don’t Solve the Normal Equations is deliberately overstated. There are times when solving the normal equations is appropriate. If is well-conditioned with a small condition number, squaring the condition number might not be that bad. If the matrix is too large to store in memory, one might want to solve the least squares problem using the normal equations and the conjugate gradient method.
However, the dramatically reduced accuracy of solving the normal equations should disqualify the approach from being the de-facto way of solving least squares problems. Unless you have good reason to think otherwise, when you see , solve a different way.
Suppose we have a thin, flat (two-dimensional) plate of homogeneous material and we measure the temperature at the border. What is the temperature inside the material? The solution to this problem is described by Laplace’s equation, one of the most ubiquitous partial differential equations in physics. Let denote the temperature of the material at point . Laplace’s equation states that, at any point on the interior of the material,
(1)
Laplace’s equation (1) and the specification of the temperature on the boundary form a well-posed mathematical problem in the sense that the temperature is uniquely determined at each point .1A well-posed problem is also required to depend continuously on the input data which, in this case, are the boundary temperatures. Indeed, the Laplace problem with boundary data is well-posed in this sense. We call this problem the Laplace Dirichlet problem since the boundary conditions
Another area of physics where the Laplace equation (1) appears is the study of electrostatics. In this case, represents the electric potential at the point . The Laplace Dirichlet problem is to find the electric potential in the interior of the region with knowledge of the potential on the boundary.
The electrostatic application motivates a different way of thinking about the Laplace equation. Consider the following question:
How would I place electric charges on the boundary to produce the electric potential at each point on the boundary?
This is a deliciously clever question. If I were able to find an arrangement of charges answering the question, then I could calculate the potential at each point in the interior by adding up the contribution to the electric potential of each element of charge on the boundary. Thus, I can reduce the problem of finding the electric potential at each point in the 2D region to finding a charge distribution on the 1D boundary to that region.
We shall actually use a slight variant of this charge distribution idea which differs in two ways:
Rather than placing simple charges on the boundary of the region, we place charge dipoles.2The reason for why this modification works better is an interesting question, but answering it properly would take us too far afield from the goals of this article.
Since we are considering a two-dimensional problem, we use a different formula for the electric potential than given by Coulomb’s law for charges in 3D. Also, since we are interested in solving the Laplace Dirichlet problem in general, we can choose a convenient dimensionless system of units. We say that the potential at a point induced by a unit “charge” at the origin is given by .
With these modifications, our new question is as follows:
How would I place a density of “charge” dipoles on the boundary to produce the electric potential at each point on the boundary?
We call this function the double layer potential for the Laplace Dirichlet problem. One can show the double layer potential satisfies a certain integral equation. To write down this integral equation, let’s introduce some more notation. Let be the region of interest and its boundary. Denote points concisely as vectors , with the length of denoted . The double layer potential satisfies
(2)
where the integral is taken over the surface of the region ; denotes the directional derivative taken in the direction normal (perpendicular) to the surface at the point . Note we choose a unit system for which hides physical constants particular to the electrostatic context, since we are interested in applying this methodology to the Laplace Dirichlet problem in general (possibly non-electrostatic) applications.
There’s one last ingredient: How do we compute the electric potential at points in the interior of the region? This is answered by the following formula:
(3)
The integral equation (2) is certainly nothing to sneeze at. Rather than trying to comprehend it in its full glory, we shall focus on a special case for the rest of our discussion. Suppose the region is a circular disk with radius centered at . The the partial derivative in the integrand in (2) then is readily computed for points and both on the boundary of the circle:
Substituting in (2) then gives
(4)
The Sherman–Morrison Formula
We are interested in solving the integral equation (4) to obtain an expression for the double-layer potential , as this will give us a solution formula for the Laplace Dirichlet problem. Ultimately, we accomplish this by using a clever trick. In an effort to make this trick seem more self-evident and less of a “rabbit out of a hat”, I want to draw an analogy to a seemingly unrelated problem: rank-one updates to linear systems of equations and the Sherman–Morrison formula.3In accordance with Stigler’s law of eponymy, the Sherman–Morrison formula was actually discovered by William J. Duncan five years before Sherman, Morrison, and Woodbury. For a more general perspective on the Sherman–Morrison formula and its generalization to the Sherman–Morrison–Woodbury formula, you may be interested in the following post of mine on Schur complements and block Gaussian elimination.
Suppose we want to solve the system of linear equations
(5)
where is an square matrix and , , and are length- vectors. We are ultimately interested in finding from . To gain insight into this problem, it will be helpful to first carefully considered the problem in reverse: computing from . We could, of course, perform this computation by forming the matrix in memory and multiplying it with , but there is a more economical way:
Form .
Compute .
Standing back, observe that we now have a system of equations for unknowns and . Specifically, our first equation can be rewritten as
which combined with the second equation
gives the by system4This “state space approach” of systematically writing out a matrix–vector multiply algorithm and then realizing this yields a larger system of linear equations was initially taught to be by my mentor Shiv Chandrasekaran; this approach has much more powerful uses, such as in the theory of rank-structured matrices.
(6)
The original equation for (5) can be derived from the “lifted” equation (6) by applying Gaussian elimination and eliminating the first row of the linear system (6). But now that we have the lifted equation (6), one can naturally wonder what would happen if we instead used Gaussian elimination to eliminate the last rows of (6); this will give us an equation for which we can solved without first computing . Doing this so-called block Gaussian elimination yields
This example shows how it can be conceptually useful to lift a linear system of equations by adding additional variables and equations and then “do Gaussian elimination in a different order”. The same insight shall be useful in solving integral equations like (4).
Solving for the Double Layer Potential
Let’s try repeating the playbook we executed for the rank-one-updated linear system (5) and apply it to the integral equation (4). We are ultimately interested in computing from but, as we did last section, let’s first consider the reverse. To compute from , we first evaluate the integral
Substituting this into (4) gives the system of equations
(7)
(8)
In order to obtain (4) from (7) and (8), we add times equation (8) to equation (7). Following last section, we now instead eliminate from equation (8) using equation (7). To do this, we need to integrate equation (7) in order to cancel the integral in equation (8):
Adding times this integrated equation to equation (8) yields
Thus plugging this expression for into equation (7) yields
We’ve solved for our double layer potential!
As promised, the double layer potential can be used to give a solution formula (known as the Poisson integral formula) for the Laplace Dirichet problem. The details are a mechanical, but also somewhat technical, exercise in vector calculus identities. We plug through the details in the following extra section.
Poisson Integral Formula
Let’s finish this up by using the double layer to derive a solution formula for the electric potential at a point in the interior of the region. To do this, we use equation (3):
(9)
We now need to do a quick calculation which is somewhat technical and not particularly enlightening. We evaluate using the divergence theorem:
Computing the boundary derivative for the spherical region centered at the origin with radius , we obtain the formula
We’ve succeeded at deriving a solution formula for for points in the interior of the disk in terms of for points on the boundary of the disk. This is known as the Poisson integral formula for the disk in two dimensions. This formula can be generalized to balls in higher dimensions, though this proof technique using “Sherman–Morrison” fails to work in more than two dimensions.
Sherman–Morrison for Integral Equations
Having achieved our main goal of deriving a solution formula for the 2D Laplace Dirichlet problem for a circular domain, I want to take a step back to present the approach from two sections ago in more generality. Consider a more general integral equation of the form
(10)
where is some region in space, , , and are functions of one or two arguments on , and is a nonzero constant. Such an integral equation is said to be of the second kind. The integral equation for the Laplace Dirichlet problem (2) is of this form with , , , and . We say the kernel is separable with rank if can be expressed in the form
With the circular domain, the Laplace Dirichlet integral equation (2) is separable with rank .5E.g., set and . We shall focus on the second kind integral equation (10) assuming the kernel is separable with rank (for simplicity, we set ):
(11)
Let’s try and write this equation in a way that’s more similar to the linear system of equation (5). To do this, we make use of linear operators defined on functions:
Let denote the identity operator on functions: It takes as inputs function and outputs the function unchanged.
Let denote the “integration against operator”: It takes as input a function and outputs the number .
With these notations, equation (11) can be written as
Using the same derivation which led to the Sherman–Morrison formula for linear systems of equations, we can apply the Sherman–Morrison formula to this integral equation in operator form, yielding
Therefore, the solution to the integral equation (11) is
This can be interpreted as a kind of Sherman–Morrison formula for the integral equation (11).
One can also generalize to provide a solution formula for the second-kind integral equation (10) for a separable kernel with rank ; in this case, the natural matrix analog is now the Sherman–Morrison–Woodbury identity rather than the Sherman–Morrison formula. Note that this solution formula requires the solution of a system of linear equations. One can use this as a numerical method to solve second-kind integral equations: First, we approximate by a separable kernel of a modest rank and then compute the exact solution of the resulting integral equation with the approximate kernel.6A natural question is why one might want to solve an integral equation formulation of a partial differential equations like the Laplace or Helmholtz equation. An answer is that that formulations based on second-kind integral equations tend to lead to systems of linear equations which much more well-conditioned as compared to other methods like the finite element method. They have a number of computational difficulties as well, as the resulting linear systems of equations are dense and may require elaborate quadrature rules to accurately compute.
My goal in writing this post was to discuss two topics which are both near and dear to my heart, integral equations and the Sherman–Morrison formula. I find the interplay of these two ideas to be highly suggestive. It illustrates the power of the analogy between infinite-dimensional linear equations, like differential and integral equations, and finite-dimensional ones, which are described by matrices. Infinite dimensions certainly do have their peculiarities and technical challenges, but it can be very useful to first pretend infinite-dimensional linear operators (like integral operators) are matrices, do calculations to derive some result, and then justify these computations rigorously post hoc.7The utility of this technique is somewhat of an open secret among some subset of mathematicians and scientists, but such heuristics are usually not communicated to students explicitly, at least in rigorous mathematics classes.
In this post, I want to discuss a beautiful and somewhat subtle matrix factorization known as the Vandermonde decomposition that appears frequently in signal processing and control theory. We’ll begin from the very basics, introducing the controls-and-signals context, how the Vandermonde decomposition comes about, and why it’s useful. By the end, I’ll briefly share how we can push the Vandermonde decomposition beyond matrices to the realm of tensors, which will can allow us to separate mixed signals from multiple measurements. This tensorial generalization plays an important role in my paper -decompositions, sparse component analysis, and the blind separation of sums of exponentials, joint work with Nithin Govindajaran and Lieven De Lathauwer, which recently appeared in the SIAM Journal of Matrix Analysis and Applications.
Finding the Frequencies
Suppose I give you a short recording of a musical chord consisting of three notes. How could you determine which three notes they were? Mathematically, we can represent such a three-note chord as a combination of scaled and shifted cosine functions
(1)
We are interested in obtaining the (angular) frequencies , , and .
In the extreme limit, when we are given the values of the signal for all , both positive and negative, the frequencies are immediately given by taking a Fourier transform of the function . In practice, we only have access to the function at certain times which we assume are equally spaced
Given the samples
we could try to identify , , and using a discrete Fourier transform.1The discrete Fourier transform can be computed very quickly using the fast Fourier transform, as I discussed in a previous post. Unfortunately, this generally requires a large number of samples to identify , , and accurately. (The accuracy scales roughly like , where is the number of samples.) We are interested in finding a better way to identify the frequencies.
Now that we’ve moved from the function , defined for any real input , to a set of samples it will be helpful to rewrite our formula (1) for in a different way. By Euler’s identity, the cosines can be rewritten as
As a consequence, we can rewrite one of the frequency components in (1) as
Here, and are complex coefficients and which contain the same information as the original parameters and . Now notice that we are only interest in values which are multiples of the spacing . Thus, our frequency component can be further rewritten as
where and . Performing these reductions, our samples take the form
(2)
We’ve now reformulated our frequency problems in identifying the parameters and in the relation (2) from a small number of measurements .
Frequency Finding as a Matrix Factorization
We will return to the algorithmic problem of identifying the parameters in the relation (2) from measurements in a little bit. First, we will see that (2) can actually be written as a matrix factorization. Understanding computations by matrix factorization has been an extremely successful paradigm in applied mathematics, and we will see in this post how viewing (2) as a matrix factorization can be very useful.
While it may seem odd at first,2As pointed out to me on math stack exchange, one reason forming the Hankel matrix is sensible is because it effectively augments the sequence of numbers into a sequence of vectors given by the columns of . This can reveal patterns in the sequence which are less obvious when it is represented as given just as numbers. For instance, any seven columns of are linearly dependent, a surprising fact since the columns of have length which can be much larger than seven. In addition, as we will soon effectively exploit later, vectors in the nullspace of (or related Hankel matrices derived from the sequence) give recurrence relations obeyed by the sequence. This speaks to a general phenomenon where properties of sequence (say, arising from snapshots of a dynamical system) can sometimes become more clear by this procedure of delay embedding. it will be illuminating to repackage the measurements as a matrix:
(3)
Here, we have assumed is odd. The matrix is known as the Hankel matrix associated with the sequence . Observe that the entry in position of depends only on the sum of the indices and , . (We use a zero-indexing system to label the rows and columns of where, for instance, the first row of is row .)
Let’s see how we can interpret the frequency decomposition (2) as a factorization of the Hankel matrix . We first write out using (2):
(4)
The power was just begging to be factorized as , which we did. Equation (4) almost looks like the formula for the product of two matrices with entries , so it makes sense to introduce the matrix with entry . This is a so-called Vandermonde matrix associated with and has the form
If we also introduce the diagonal matrix , the formula (4) for can be written as the matrix factorization3In the Vandermonde decomposition , the factor appears transposed even when is populated with complex numbers! This differs from the usual case in linear algebra where we use the conjugate transpose rather than the ordinary transpose when working with complex matrices. As a related issue, observe that if at least one of the measurements is a (non-real) complex number, the Hankel matrix is symmetric but not Hermitian.
(5)
This is the Vandermonde decomposition of the Hankel matrix , a factorization of as a product of the transpose of a Vandermonde matrix, a diagonal matrix, and that same Vandermonde matrix.
The Vandermonde decomposition immediately tells us all the information and describing our sampled recording via (2). Thus, the problem of determining and is equivalent to finding the Vandermonde decomposition (5) of the Hankel matrix .
Computing the Vandermonde Decomposition: Prony’s Method
Computing the Vandermonde decomposition accurately can be a surprisingly hard task, particularly if the measurements are corrupted by even a small amount of measurement error. In view of this, I want to present a very classical way of computing this decomposition (dating back to 1795!) known as Prony’s method. This method is conceptually simple and will be a vehicle to continue exploring frequency finding and its connection with Hankel matrices. It remainsinuse, though it’s accuracy may be significantly worse compared to other methods.
As a first step to deriving Prony’s method, let’s reformulate the frequency finding problem in a different way. Sums of cosines like the ones in our expression (1) for the function often appear as the solution to a (linear) ordinary differential equation (ODE). This means that one way we could find the frequencies comprising would be to find a differential equation which satisfies. Together with the initial condition , determining all the frequencies would be very straightforward.
Since we only have access to samples of at regular time intervals, we will instead look for the “discrete-time” analog of a linear ODE, a linear recurrence relation. This is an expression of the form
(6)
In our case, we’ll have because there are six terms in the formula (2) for . Together with initial conditions , such a recurrence will allow us to determine the parameters and in our formula (2) for our sampled recordings and hence also allow us to compute the Vandermonde decomposition (5).
Observe that the recurrence (6) is a linear equation in the variables . A very good rule of thumb in applied mathematics is to always write down linear equations in matrix–vector notation in see how it looks. Doing this, we obtain
(7)
Observe that the matrix on the right-hand side of this equation is also a Hankel matrix (like in (3)) formed from the samples . Call this Hankel matrix . Unlike in (3), is rectangular. If is much larger than , will be tall, possessing many more rows than columns. We assume going forward.4 would also be fine for our purposes, but we assume to illustrate this highly typical case.
Let’s write (7) a little more compactly as
(8)
where we’ve introduced for the vector on the left-hand side of (7) and collected the recurrence coefficients into a vector . For a typical system of linear equations like (8), we would predict the system to have no solution : Because has more rows than columns (if ), the system equations (8) has more equations than unknowns. Fortunately, we are not in the typical case. Despite the fact that we have more equations than unknowns, the linear equations (8) have a unique solution .5This solution can be computed by solving the system of linear equations In particular, the matrix on the right-hand side of this equation is guaranteed to be nonsingular under our assumptions. Using the Vandermonde decomposition, can you see why? The existence of a unique solution is a consequence of the fact that the samples satisfy the formula (2). As a fun exercise, you might want to verify the existence of a unique satisfying (8)!
As a quick aside, if the measurements are corrupted by small measurement errors, then the equations (8) will usually not possess a solution. In this case, it would be appropriate to find the least squares solution to equation (8) as a way of mitigating these errors.
Hurrah! We’ve found the coefficients providing a recurrence relation (6) for our measurements . All that is left is to find the parameters and in our signal formula (2) and the Vandermonde decomposition (5). Fortunately, this is just a standard computation for linear recurrence relations, nicely paralleling the solution of (homogenous) linear ODEs by means of the so-called “characteristic equation”. I’ll go through fairly quickly since this material is well-explained elsewhere on the internet (like Wikipedia). Let’s guess that our recurrence (6) has a solution of the form ; we seek to find all complex numbers for which this is a bonafide solution. Plugging this solution into the formula (6) for gives
(9)
This is the so-called characteristic equation for the recurrence (6). As a single-variable polynomial equation of degree six, it has six complex solutions . These numbers are precisely those numbers which appear in the sequence formula (2) and the Vandermonde decomposition (5).
Finally, we need to compute the coefficients . But this is easy. Observe that the formula (2) provides the following system of linear equations for :
(10)
Again, this system of equations will have a unique solution if the measurements are uncorrupted by errors (and can be solved in the least squares sense if corrupted). This gives , completing our goal of computing the parameters in the formula (2) or, equivalently, finding the Vandermonde decomposition (5).
We have accomplished our goal of computing the Vandermonde decomposition. The approach by which we did so is known as Prony’s method, as mentioned in the introduction to this section. As suggested, this method may not always give high-accuracy results. There are two obvious culprits that jump out about why this is the case. Prony’s method requires solving for the roots of the polynomial equation (9) expressed “in the monomial basis” and solving a system of linear equations (10) with a (transposed) Vandermonde matrix. Both of theseproblems can be notoriously ill-conditioned and thus challenging to solve accurately and may require the measurements to be done to very high accuracy. Notwithstanding this, Prony’s method does useful results in some cases and forms the basis for potentially more accurate methods, such as those involving generalized eigenvalue problems.
Separating Signals: Extending the Vandermonde Decomposition to Tensors
In our discussion of the frequency identification problem, the Vandermonde decomposition (5) has effectively been an equivalent way of showing the samples are a combination of exponentials . So far, the benefits of the matrix factorization perspective have yet to really reveal themselves.
So what are the benefits of the Vandermonde decompostions? A couple of nice observations related to the Vandermonde decomposition and the “Hankelization” of the signals have already been lurking in the background. For instance, the rank of the Hankel matrix is the number of frequency components needed to describe the samples and the representation of the samples as a mixture of exponentials is uniquely determined only if the matrix does not have full rank; I have a little more to say about this at the very end. There are also benefits to certain computational problems; one can use Vandermonde decompositions to compute super high accuracy singular value decompositions of Hankel matrices.
The power of the Vandermonde decomposition really starts to shine when we go beyond the basic frequency finding problem we discussed by introducing more signals. Suppose now there are three short recordings , , and . (Here, the superscript denotes an index rather than differentiation.) Each signal is a weighted mixture of three sources , , and , each of which plays a musical chord of three notes (thus representable as a sum of cosines as in (1)). One can think of the sources of being produced three different musical instruments at different places in a room and the recordings , , and being taken from different microphones in the room.6This scenario of instruments and microphones ignores the finite propagation speed of sound, which also would introduce time delays in the sources in the recorded signals. We effectively treat the speed of sound as being instantaneous. Our goal is now not just to identify the musical notes in the recordings but also to identify how to assign those notes to reconstruct the source signals , , and .
Taking inspiration from earlier, we record samples for each recording and form each collection of samples into a Hankel matrix
Here comes the crazy part: Stack the Hankelized recordings , , and as slices of a tensor . A tensor, in this context, just means a multidimensional array of numbers. Just as a vector is a one-dimensional array and a matrix is a two-dimensional array, a tensor could have any number of dimensions. In our case, we need just three. If we use a MATLAB-esque indexing notation, is a array given by
The remarkable thing is that the source signals can be determined (under appropriate conditions) by computing a special kind of Vandermonde decomposition of the tensor ! (Specifically, the required decomposition is a Vandermonde-structured -block term decomposition of the tensor .) Even more cool, this decomposition can be computed using general-purpose software like Tensorlab.
If this sounds interesting, I would encourage you to check out my recently published paper -decompositions, sparse component analysis, and the blind separation of sums of exponentials, joint work with Nithin Govindajaran and Lieven De Lathauwer and recently published in the SIAM Journal on Matrix Analysis and Applications. In the paper, we explain what this -decomposition is and how applying it to can be used to separate mixtures of exponentials signals from the resulting Vandermonde structure, an idea originating in the work of De Lathauewer. A very important question for these signal separation problems is that of uniqueness. Given the three sampled recordings (comprising the tensor ), is there just one way of unscrambling the mixtures into different sources or multiple? If there are multiple, then we might have possibly computed the wrong one. If there is just a single unscrambling, though, then we’ve done our job and unmixed the scrambled signals. The uniqueness of these tensor decompositions is fairly complicated math, and we survey existing results and prove new ones in this paper.7One of our main technical contributions is a new notion of uniqueness of -decompositions which we believe is nicely adapted to the signal separation context. Specfically, we prove mathematized versions of the statement “if the source signals are sufficiently different from each others and the measurements of sufficiently high quality, then the signals can uniquely be separated”.
Conclusions, Loose Ends, and Extensions
The central idea that we’ve been discussing is how it can be useful to convert between a sequence of observations and a special matricization of this sequence into a Hankel matrix (either square, as in (3), or rectangular, as in (7)). By manipulating the Hankel matrix, say, by computing its Vandermonde decomposition (5), we learn something about the original signal, namely a representation of the form (2).
This is a powerful idea which appears implicitly or explicitly throughout various subfields of mathematics, engineering, and computation. As with many other useful ideas, this paradigm admits many natural generalizations and extensions. We saw one already in this post, where we extended the Vandermonde decomposition to the realm of tensors to solve signal separation problems. To end this post, I’ll place a few breadcrumbs at the beginning of a few of the trails of these generalizations for any curious to learn more, wrapping up a few loose ends on the way.
Is the Vandermonde Decomposition Unique?
A natural question is whether the Vandermonde decomposition (5) is unique. That is, is it possible that there exists two Vandermonde decompositions
of the same (square) Hankel matrix ? This is equivalent to whether the frequency components can be uniquely determined from the measurements .
Fortunately, the Vandermonde decomposition is unique if (and only if) the matrix is a rank-deficient matrix. Let’s unpack this a little bit. (For those who could use a refresher on rank, I have a blog post on precisely this topic.) Note that the Vandermonde decomposition is a rank factorization8Rank factorizations are sometimes referred to as “rank-revealing factorizations”. I discuss my dispreference for this term in my blog post on low-rank matrices. since has rows, has full (row) rank, and is invertible. This means that if take enough samples of a function which is a (finite) combinations of exponentials, the matrix will be rank-deficient and the Vandermonde decomposition unique.9The uniqueness of the Vandermonde decomposition can be proven by showing that, in our construction by Prony’s method, the ‘s, ‘s, and ‘s are all uniquely determined. If too few samples are taken, then does not contain enough information to determine the frequency components of the signal and thus the Vandermonde decomposition is non-unique.
Does Every Hankel Matrix Have a Vandermonde Decomposition?
This post has exclusively focused on a situation where we are provided with sequence we know to be represented as a mixture of exponentials (i.e., taking the form (2)) from which the existence of the Vandermonde decomposition (5) follows immediately. What if we didn’t know this were the case, and we were just given a (square) Hankel matrix . Is guaranteed to possess a Vandermonde decomposition of the form (5)?
Unfortunately, the answer is no; there exist Hankel matrices which do not possess a Vandermonde decomposition. The issue is related to the fact that the appropriate characteristic equation (analogous to (9)) might possess repeated roots, making the solutions to the recurrence (6) not just take the form but also and perhaps , , etc.
Are There Cases When the Vandermonde Decomposition is Guaranteed To Exist?
There is one natural case when a (square) Hankel matrix is guaranteed to possess a Vandermonde decomposition, namely when the matrix is nonsingular/invertible/full-rank. Despite this being a widely circulated fact, I am unaware of a simple proof for why this is the case. Unfortunately, there is not just one but infinitely many Vandermonde decompositions for a nonsingular Hankel matrix, suggesting these decompositions are not useful for the frequency finding problem that motivated this post.
What If My Hankel Matrix Does Not Possess a Vandermonde Decomposition?
As discussed above, a Hankel matrix may fail to have a Vandermonde decomposition if the characteristic equation (a la (9)) has repeated roots. This is very much analogous to the case of a non-diagonalizable matrix for which the characteristic polynomial has repeated roots. In this case, while diagonalization is not possible, one can “almost-diagonalize” the matrix by reducing it to its Jordan normal form. In total analogy, every Hankel matrix can be “almost Vandermonde decomposed” into a confluent Vandermonde decomposition (a discovery that appears to have been made independentlyseveraltimes). I will leave these links to discuss the exact nature of this decomposition, though I warn any potential reader that these resources introduce the decomposition first for Hankel matrices with infinitely many rows and columns before considering the finite case as we have. One is warned that while the Vandermonde decomposition is always a rank decomposition, the confluent Vandermonde decomposition is not guaranteed to be one.10Rather, the confluent Vandermonde decomposition is a rank decomposition for an infinite extension of a finite Hankel matrix. Consider the Hankel matrix This matrix has rank-two but no rank-two confluent Vandermonde decomposition. The issue is that when extended to an infinite Hankel matrix this (infinite!) matrix has a rank exceeding the size of the original Hankel matrix .
The Toeplitz Vandermonde Decomposition
Just as it proved useful to arrange samples into a Hankel matrix, it can also be useful to form them into a Toeplitz matrix
One can interconvert between Hankel and Toeplitz matrices by reversing the order of the rows. As such, to the extent to which Hankel matrices possess Vandermonde decompositions (with all the asterisks and fine print just discussed), Toeplitz matrices do as well but with the rows of the first factor reversed:
There is a special and important case where more is true. If a Toeplitz matrix is (Hermitian) positive semidefinite, then always possesses a Vandermonde decomposition of the form
where is a Vandermonde matrix associated with parameters which are complex numbers of absolute value one and is a diagonal matrix with real positive entries.12The keen-eyed reader will note that appears conjugate transposed in this formula rather than transposed as in the Hankel Vandermonde decomposition (5). This Vandermonde decomposition is unique if and only if is rank-deficient. Positive semidefinite Toeplitz matrices are important as they occur as autocorrelation matrices which effectively describe the similarity between a signal and different shifts of itself in time. Autocorrelation matrices appear under different names in everything from signal processing to random processes to near-term quantum algorithms (a topic near and dear to my heart). A delightfully simple and linear algebraic derivation of this result is given by Yang and Xie (see Theorem 1).13Unfortunately, Yang and Xie incorrectly claim that every Toeplitz matrix possesses a rank factorization Vandermonde decomposition of the form where is a Vandermonde matrix populated with entries on the unit circle and is a diagonal matrix of possibly *complex* entries. This claim is disproven by the example . This decomposition can be generalized to infinite positive semidefinite Toeplitz matrices (appropriately defined).14Specifically, one can show that an infinite positive semidefinite Toeplitz matrix (appropriately defined) also has a “Vandermonde decomposition” (appropriately defined). This result is often known as Herglotz’s theorem and is generalized by the Bochner–Weil theorem.
In this post, I want to discuss a beautiful and simple geometric picture of the perturbation theory of definite generalized eigenvalue problems. As a culmination, we’ll see a taste of the beautiful perturbation theory of Mathias and Li, which appears to be not widely known in virtue of only being explained in a technical report. Perturbation theory for the generalized eigenvalue problem is a bit of a niche subject, but I hope you will stick around for some elegant arguments. In addition to explaining the mathematics, I hope this post serves as an allegory for the importance of having the right way of thinking about a problem; often, the solution to a seemingly unsolvable problem becomes almost self-evident when one has the right perspective.
What is a Generalized Eigenvalue Problem?
This post is about the definite generalized eigenvalue problem, so it’s probably worth spending a few words talking about what generalized eigenvalue problems are and why you would want to solve them. Slightly simplifying some technicalities, a generalized eigenvalue problem consists of finding nonzero vectors and a (possibly complex) numbers such that .1In an unfortunate choice of naming, there is actually a completely different sense in which it makes sense to talk about generalized eigenvectors, in the context of the Jordan normal form for standard eigenvalue problems. The vector is called an eigenvector and its eigenvalue. For our purposes, and will be real symmetric (or even complex Hermitian) matrices; one can also consider generalized eigenvalue problemss for nonsymmetric and even non-square matrices and , but the symmetric case covers many applications of practical interest. The generalized eigenvalue problem is so-named because it generalizes the standard eigenvalue problem , which is a special case of the generalized eigenvalue problem with .2One can also further generalize the generalized eigenvalue problem to polynomial and nonlinear eigenvalue problems.
Why might we want so solve a generalized eigenvalue problem? The applications are numerous (e.g., in chemistry, quantum computation, systems and control theory, etc.). My interest in perturbation theory for generalized eigenvalue problems arose in the analysis of a quantum algorithm for eigenvalue problems in chemistry, and the theory discussed in this article played a big role in that analysis. To give an application which is more easy to communicate than the quantum computation application which motivated my own interest, let’s discuss an application in classical mechanics.
The Lagrangian formalism is a way of reformulating Newton’s laws of motion in a general coordinate system.3The benefits of the Lagrangian framework are far deeper than working in generalized coordinate systems, but this is beyond the scope of our discussion and mostly beyond the scope of what I am knowledgeable enough to write meaningfully about. If denotes a vector of generalized coordinates describing our system and denotes ‘s time derivative, then the time evolution of a system with Lagrangian functional are given by the Euler–Lagrange equations . If we choose to represent the deviation of our system from equilibrium,4That is, in our generalized coordinate system, is a (static) equilibrium configuration for which whenever and . then our Lagrangian is well-approximated by it’s second order Taylor series:
By the Euler–Lagrange equations, the equations of motion for small deviations from this equillibrium point are described by
A fundamental set of solutions of this system of differential equations is given by , where and are the generalized eigenvalues and eigenvectors of the pair .5That is, all solutions to can be written (uniquely) as linear combinations of solutions of the form . In particular, if all the generalized eigenvalues are positive, then the equillibrium is stable and the square roots of the eigenvalues represent the modes of vibration. In the most simple mechanical systems, such as masses in one dimension connected by springs with the natural coordinate system, the matrix is diagonal with diagonal entries equal to the different masses. In even slightly more complicated “freshman physics” problems, it is quite easy to find examples where, in the natural coordinate system, the matrix is nondiagonal.6Almost always, the matrix is positive definite. As this example shows, generalized eigenvalue problems aren’t crazy weird things since they emerge as natural descriptions of simple mechanical systems like coupled pendulums.
One reason generalized eigenvalue problems aren’t more well-known is that one can easily reduce a generalized eigenvalue problem into a standard one. If the matrix is invertible, then the generalized eigenvalues of are just the eigenvalues of the matrix . For several reasons, this is a less-than-appealing way of reducing a generalized eigenvalue problem to a standard eigenvalue problem. A better way, appropriate when and are both symmetric and is positive definite, is to reduce the generalized eigenvalue problem for to the symmetrically reduced matrix , which also possesses the same eigenvalues as . In particular, the matrix remains symmetric, which shows that has real eigenvalues by the spectral theorem. In the mechanical context, one can think of this reformulation as a change of coordinate system in which the “mass matrix” becomes the identity matrix .
There are several good reasons to not simply reduce a generalized eigenvalue problem to a standard one, and perturbation theory gives a particular good reason. In order for us to change coordinates to change the matrix into an identity matrix, we must first know the matrix. If I presented you with an elaborate mechanical system which you wanted to study, you would need to perform measurements to determine the and matrices. But all measurements are imperfect and the entries of and are inevitably corrupted by measurement errors. In the presence of these measurement errors, we must give up on computing the normal modes of vibration perfectly; we must content ourselves with computing the normal modes of vibration plus-or-minus some error term we hope we can be assured is small if our measurement errors are small. In this setting, reducing the problem to seems less appealing, as I have to understand how the measurement errors in and are amplified in computing the triple product . This also suggests that computing may be a poor algorithmic strategy in practice: if the matrix is ill-conditioned, there might be a great deal of error amplification in the computed product . One might hope that one might be able to devise algorithms with better numerical stability properties if we don’t reduce the matrix pair to a single matrix. This is not to say that reducing a generalized eigenvalue problem to a standard one isn’t a useful tool—it definitely is. However, it is not something one should do reflexively. Sometimes, a generalized eigenvalue problem is best left as is and analyzed in its native form.
The rest of this post will focus on the question if and are real symmetric matrices (satisfying a definiteness condition, to be elaborated upon below), how do the eigenvalues of the pair compare to those of , where and are small real symmetric perturbations? In fact, it shall be no additional toil to handle the complex Hermitian case as well while we’re at it, so we shall do so. (Recall that a Hermitian matrix satisfies , where is the conjugate transpose. Since the complex conjugate does not change a real number, a real Hermitian matrix is necesarily symmetric .) For the remainder of this post, let , , , and be Hermitian matrices of the same size. Let and denote the perturbations.
Symmetric Treatment
As I mentioned at the top of this post, our mission will really be to find the right way of thinking about perturbation theory for the generalized eigenvalue problem, after which the theory will follow much more directly than if we were to take a frontal assault on the problem. As we go, we shall collect nuggets of insight, each of which I hope will follow quite naturally from the last. When we find such an insight, we shall display it on its own line.
The first insight is that we can think of the pair and interchangeably. If is a nonzero eigenvalue of the pair , satisfying , then . That is, is an eigenvalue of the pair . Lack of symmetry is an ugly feature in a mathematical theory, so we seek to smooth it out. After thinking a little bit, notice that we can phrase the generalized eigenvalue condition symmetrically as with the associated eigenvalue being given by . This observation may seem trivial at first, but let us collect it for good measure.
Treat and symmetrically by writing the eigenvalue as with .
Before proceeding, let’s ask a question that, in our new framing, becomes quite natural: what happens when ? The case is problematic because it leads to a division by zero in the expression . However, if we have , this expression still makes sense: we’ve found a vector for which but . It makes sense to consider still an eigenvector of with eigenvalue ! Dividing by zero should justifiably make one squeemish, but it really is quite natural in the case to treat as a genuine eigenvector with eigenvalue .
Things get even worse if we find a vector for which . Then, any can reasonably considered an eigenvalue of since . In such a case, all complex numbers are simultaneously eigenvalues of , in which case we call singular.7More precisely, a pair is singular if the determinant is identically zero for all . For the generalized eigenvalue problem to make sense for a pair , it is natural to require that not be singular. In fact, we shall assume an even stronger “definiteness” condition which ensures that has only real (or infinite) eigenvalues. Let us return to this question of definiteness in a moment and for now assume that is not singular and possesses real eigenvalues.
With this small aside taken care of, let us return to the main thread. By modeling eigenvalues as pairs , we’ve traded one ugliness for another. While reformulating the eigenvalue as a pair treats and symmetrically, it also adds an additional indeterminacy, scale. For instance, if is an eigenvalue of , then so is . Thus, it’s better not to think of so much as a pair of numbers together with all of its possible scalings.8Projective space provides a natural framework for studying such vectors up to scale indeterminacy. For reasons that shall hopefully become more clear as we go forward, it will be helpful to only consider all the possible positive scalings of —e.g., all for . Geometrically, the set of all positive scalings of a point in two-dimensional space is precisely just a ray emanating from the origin.
Represent eigenvalue pairs as rays emanating from the origin to account for scale ambiguity.
Now comes a standard eigenvalue trick. It’s something you would never think to do originally, but once you see it once or twice you learn to try it as a matter of habit. The trick: multiply the eigenvalue-eigenvector relation by the (conjugate) transpose of :9For another example of the trick, try applying it to the standard eigenvalue problem . Multiplying by and rearranging gives —the eigenvalue is equal to the expression , which is so important it is given a name: the Rayleigh quotient. In fact, the largest and smallest eigenvalues of can be found by maximizing and minimizing the Rayleigh quotient.
The above equation is highly suggestive: since and are only determined up to a scaling factor, it shows we can take and . And by different scalings of the eigenvector , we can scale and by any positive factor we want. (This retroactively shows why it makes sense to only consider positive scalings of and .10To make this point more carefully, we shall make great use of the identification between pairs and the pair of quadratic forms . Thus, even though and lead to equivalent eigenvalues since , and don’t necessarily both arise from a pair of quadratic forms: if , this does not mean there exists such that . Therefore, we only consider equivalent to if .) The expression is so important that we give it a name: the quadratic form (associated with and evaluated at ).
The eigenvalue pair can be taken equal to the pair of quadratic forms .
Complexifying Things
Now comes another standard mathematical trick: represent points in two-dimensional space by complex numbers. In particular, we identify the pair with the complex number .11Recall that we are assuming that is real, so we can pick a scaling in which both and are real numbers. Assume we have done this. Similar to the previous trick, it’s not fully clear why this will pay off, but let’s note it as an insight.
Identify the pair with the complex number .
Now, we combine all the previous observations. The eigenvalue is best thought of as a pair which, up to scale, can be taken to be and . But then we represent as the complex number
Let’s stop for a moment and appreciate how far we’ve come. The generalized eigenvalue problem is associated with the expression .If we just went straight from one to the other, this reduction would appear like some crazy stroke of inspiration: why would I ever think to write down ? However, just following our nose lead by a desire to treat and symmetrically and applying a couple standard tricks, this expression appears naturally. The expression will be very useful to us because it is linear in and , and thus for the perturbed problem , we have that : consequently, is a small perturbation of . This observation will be very useful to us.
If is the eigenvector, then the complex number is .
Definiteness and the Crawford Number
With these insights in hand, we can now return to the point we left earlier about what it means for a generalized eigenvalue problem to be “definite”. We know that if there exists a vector for which , then the problem is singular. If we multiply by , we see that this means that as well and thus . It is thus quite natural to assume the following definiteness condition:
The pair is said to be definite if for all complex nonzero vectors .
A definite problem is guaranteed to be not singular, but the reverse is not necessarily true; one can easily find pairs which are not definite and also not singular.12For example, consider . is not definite since for . However, is not singular; the only eigenvalue of the pair is and is not identically zero.. (Note does not imply unless and are both positive (or negative) semidefinite.)
The “natural” symmetric condition for to be “definite” is for for all vectors .
Since the expression is just scaled by a positive factor by scaling the vector , it is sufficient to check the definiteness condition for only complex unit vectors . This leads naturally to a quantitative characterization of the degree of definiteness of a pair :
The Crawford number13The name Crawford number was coined by G. W. Stewart in 1979 in recognition of Crawford’s pioneering work on the perturbation theory of the definite generalized eigenvalue problem. of a pair is the minimum value of over all complex unit vectors .
The Crawford number naturally quantifies the degree of definiteness.14In fact, it has been shown that the Crawford number is, in a sense, the distance from a definite matrix pair to a pair which is not simultaneously diagonalizable by congruence. A problem which has a large Crawford number (relative to a perturbation) will remain definite after perturbation, whereas the pair may become indefinite if the size of the perturbation exceeds the Crawford number. Geometrically, the Crawford number has the following interpretation: must lie on or outside the circle of radius centered at for all (complex) unit vectors .
The “degree of definiteness” can be quantified by the Crawford number .
Now comes another step in our journey which is a bit more challenging. For a matrix (in our case ), the set of complex numbers for all unit vectors has been the subject of considerable study. In fact, this set has a name
The field of values of a matrix is the set .
In particular, the Crawford number is just the absolute value of the closest complex number in the field of values to zero.
It is a very cool and highly nontrivial fact (called the Toeplitz–Hausdorff Theorem) that the field of values is always a convex set, with every two points in the field of values containing the line segment connecting them. Thus, as a consequence, the field of values for a definite matrix pair is always on “one side” of the complex plane (in the sense that there exists a line through zero which lies strictly on one side of15This is a consequence of the hyperplane separation theorem together with the fact that .).
The numbers for unit vectors lie on one half of the complex plane.
From Eigenvalues to Eigenangles
All of this geometry is lovely, but we need some way of relating it to the eigenvalues. As we observed a while ago, each eigenvalue is best thought of as a ray emanating from the origin, owing to the fact that the pair can be scaled by an arbitrary positive factor. A ray is naturally associated with an angle, so it is natural to characterize an eigenvalue pair by the angle describing its arc.
But the angle of a ray is only defined up additions by full rotations ( radians). As such, to associate each ray a unique angle we need to pin down this indeterminacy in some fixed way. Moreover, this indeterminacy should play nice with the field of values and the field of values of the perturbation. But in the last section, we say that each of these field of angles lies (strictly) on one half of the complex plane. Thus, we can find a ray which does not intersect either field of values!
One possible choice is to measure the angle from this ray. We shall make a slightly different choice which plays better when we treat as a complex number . Recall that a number is an argument for if for some real number . The argument is multi-valued since is an argument for as long as is (for all integers ). However, once we exclude our ray , we can assign each complex number not on this ray a unique argument which depends continuously on . Denote this “branch” of the argument by . If represents an eigenvalue , we call an eigenangle.
Represent an eigenvalue pair by its associated eigenangle .
How are these eigenangles related to the eigenvalues? It’s now a trigonometry problem:
The eigenvalues are the cotangents of the eigenangles!
The eigenvalue is the cotangent of the eigenangle .
Variational Characterization
Now comes another difficulty spike in our line of reasoning, perhaps the largest in our whole deduction. To properly motivate things, let us first review some facts about the standard Hermitian/symmetric eigenvalue problem. The big idea is that eigenvalues can be thought of as the solution to a certain optimization problem. The largest eigenvalue of a Hermitian/symmetric matrix is given by the maximization problem
The largest eigenvalue is the maximum of the quadratic form over unit vectors . What about the other eigenvalues? The answer is not obvious, but the famous Courant–Fischer Theorem shows that the th largest eigenvalue can be written as the following minimax optimization problem
The minimum is taken over all subspaces of dimension whereas the maximum is taken over all unit vectors within the subspace . Symmetrically, one can also formulate the eigenvalues as a max-min optimization problem
These variational/minimax characterizations of the eigenvalues of a Hermitian/symmetric matrix are essential to perturbation theory for Hermitian/symmetric eigenvalue problems, so it is only natural to go looking for a variational characterization of the generalized eigenvalue problem. There is one natural way of doing this that works for positive definite: specifically, one can show that
This characterization, while useful in its own right, is tricky to deal with because it is nonlinear in and . It also treats and non-symmetrically, which should set off our alarm bells that there might be a better way. Indeed, the ingenious idea, due to G. W. Stewart in 1979, is to instead provide a variational characterization of the eigenangles! Specifically, Stewart was able to show16Stewart’s original definition of the eigenangles differs from ours; we adopt the definition of Mathias and Li. The result amounts to the same thing.
(1)
for the eigenangles .17Note that since the cotangent is decreasing on , this means that the eigenvalues are now in increasing order, in contrast to our convention from earlier in this section. This shows, in particular, that the field of values is subtended by the smallest and largest eigenangles.
The eigenangles satisfy a minimax variational characterization.
How Big is the Perturbation?
We’re tantalizingly close to our objective. The final piece in our jigsaw puzzle before we’re able to start proving perturbation theorems is to quantify the size of the perturbing matrices and . Based on what we’ve done so far, we see that the eigenvalues are natural associated with the complex number , so it is natural to characterize the size of the perturbing pair by the distance between and . But the difference between these two quantities is just
We’re naturally led to the question: how big can be? If the vector has a large norm, then quite large, so let’s fix to be a unit vector. With this assumption in place, the maximum size of is simple the distance of the farthest point in the field of values from zero. This quantity has a name:
The numerical radius of a matrix (in our case ) is .18This maximum is taken over all complex unit vectors .
The size of the perturbation is the numerical radius .
It is easy to upper-bound the numerical radius by more familiar quantities. For instance, once can straightforwardly show the bound , where is the spectral norm. We prefer to state results using the numerical radius because of its elegance: it is, in some sense, the “correct” measure of the size of the pair in the context of this theory.
Stewart’s Perturbation Theory
Now, after many words of prelude, we finally get to our first perturbation theorem. With the work we’ve done in place, the result is practically effortless.
Let denote the eigenangles of the perturbed pair and consider the th eigenangle. Let be the subspace of dimension achieving the minimum in the first equation of the variational principle (1) for the original unperturbed pair . Then we have
(2)
This is something of a standard trick when dealing with variational problems in matrix analysis: take the solution (in this case the minimizing subspace) for the original problem and plug it in for the perturbed problem. The solution may no longer be optimal, but it at least gives an upper (or lower) bound. The complex number must lie at least a distance from zero and . We’re truly toast if the perturbation is large enough to perturb to be equal to zero, so we should assume that .
For our perturbation theory to work, we must assume .
Making the assumption that , bounding the right-hand side of (2) requires finding the most-counterclockwise angle necessary to subtend a circle of radius centered at , which must lie a distance from the origin. The worst-case scenario is when is exactly a distance from the origin, as is shown in the following diagram.
Solving the geometry problem for the counterclockwise-most subtending angle in this worst-case sitation, we conclude the eigenangle bound . An entirely analogous argument using the max-min variational principle (1) proves an identical lower bound, thus showing
(3)
In the language of eigenvalues, we have19I’m being a little sloppy here. For a result like this to truly hold, I believe all of the perturbed and unperturbed eigenangles should all be contained in one half of the complex plane.
Interpreting Stewart’s Theory
After much work, we have finally proven our first generalized eigenvalue perturbation theorem. After taking a moment to celebrate, let’s ask ourselves: what does this result tell us?
Let’s start with the good. This result shows us that if the perturbation, measured by the numerical radius , is much smaller than the definiteness of the original problem, measured by the Crawford number , then the eigenangles change by a small amount. What does this mean in terms of the eigenvalues? For small eigenvalues (say, less than one in magnitude), small changes in the eigenangles also lead to small changes of the eigenvalues. However, for large eigenangles, small changes in the eigenangle are magnified into potentially large changes in the eigenvalues. One can view this result in a positive or negative framing. On the one hand, large eigenvalues could be subject to dramatic changes by small perturbations; on the other hand, the small eigenvalues aren’t “taken down with the ship” and are much more well-behaved.
Stewart’s theory is beautiful. The variational characterization of the eigenangles (1) is a master stroke and exactly the extension one would want from the standard Hermitian/symmetric theory. From the variational characterization, the perturbation theorem follows almost effortlessly from a little trigonometry. However, Stewart’s theory has one important deficit: the Crawford number. All that Stewart’s theory tells is that all of the eigenangles change by at most roughly “perturbation size over Crawford number”. If the Crawford number is quite small since the problem is nearly indefinite, this becomes a tough pill to swallow.
The Crawford number is in some ways essential: if the perturbation size exceeds the Crawford number, the problem can become indefinite or even singular. Thus, we have no hope of fully removing the Crawford number from our analysis. But might it be the case that some eigenangles change by much less than “pertrubation size over Crawford number”? Could we possibly improve to a result of the form “the eigenangles change by roughly perturbation size over something (potentially) much less than the Crawford number”? Sun improved Stewart’s analysis in 1982, but the scourge of the Crawford number remained.20Sun’s bound does not explicitly have the Crawford number, instead using the quantity and another hard-to-concisely describe quantity. In many cases, one has nothing better to do than to bound , in which case the Crawford number has appeared again. The theory of Mathias and Li, published in a technical report in 2004, finally produced a bound where the Crawford number is replaced.
The Mathias–Li Insight and Reduction to Diagonal Form
Let’s go back to the Stewart theory and look for a possible improvement. Recall in the Stewart theory that we considered the point on the complex plane. We then argued that, in the worst case, this point would lie a distance from the origin and then drew a circle around it with radius . To improve on Stewart’s bound, we must somehow do something better than using the fact that . The insight of the Mathias–Li theory is, in some sense, as simple as this: rather than using the fact that (as in Stewart’s analysis), use how far actually is from zero, where is chosen to be the unit norm eigenvectors of .21This insight is made more nontrivial by the fact that, in the context of generalized eigenvalue problems, it is often not convenient to choose the eigenvectors to have unit norm. As Mathias and Li note, there are often two more popular normalizations for . If is positive definite, one often normalizes such that —the eigenvectors are thus made “-orthonormal”, generalizing the fact that the eigenvectors of a Hermitian/symmetric matrix are orthonormal. Another popular normalization is to scale such that . In this way, just taking the eigenvector to have unit norm is already a nontrivial insight.
Before going further, let us quickly make a small reduction which will simplify our lives greatly. Letting denote a matrix whose columns are the unit-norm eigenvectors of , one can verify that and are diagonal matrices with entries and respectively. With this in mind, it can make our lives a lot easy to just do a change of variables and (which in turn sends and ). The change of variables is very common in linear algebra and is called a congruence transformation.
Perform a change of variables by a congruence transformation with the matrix of eigenvectors.
While this change of variables makes our lives a lot easier, we must first worry about how this change of variables might effect the size of the perturbation matrices . It turns out this change of variables is not totally benign, but it is not maximally harmful either. Specifically, the spectral radius can grow by as much as a factor of .22This is because, in virtue of having unit-norm columns, the spectral norm of the matrix is . Further, note the following variational characterization of the spectral radius . Plugging these two facts together yields . This factor of isn’t great, but it is much better than if the bound were to degrade by a factor of the condition number , which can be arbitrarily large.
This change of variables may increase by at most a factor of .
From now on, we shall tacitly assume that this change of variables has taken place, with and being diagonal and and being such that is at most a factor larger than it was previously. We denote by and the th diagonal element of and , which are given by and where is the th unit-norm eigenvector
Mathias and Li’s Perturbation Theory
We first assume the perturbation is smaller than the Crawford number in the sense , which is required to be assured that the perturbed problem does not lose definiteness. This will be the only place in this analysis where we use the Crawford number.
Draw a circle of radius around .
If is the associated eigenangle, then this circle is subtended by arcs with angles
It would be nice if the perturbed eigenangles were guaranteed to lie in these arcs (i.e., ). Unfortunately this is not necessarily the case. If one is close to the origin, it will have a large arc which may intersect with other arcs; if this happens, we can’t guarantee that each perturbed eigenangle will remain within its individual arc. We can still say something though.
What follows is somewhat technical, so let’s start with the takeaway conclusion: is larger than any of the lower bounds . In particular, this means that is larger than the th largest of all the lower bounds. That is, if we rearrange the lower bounds in decreasing order , we hace . An entirely analogous argument will give an upper bound, yielding
(4)
For those interested in the derivation, read on the in the following optional section:
Derivation of the Mathias–Li Bounds
Since and are diagonal, the eigenvectors of the pair are just the standard basis vectors, the th of which we will denote . The trick will be to use the max-min characterization (1) with the subspace spanned by some collection of basis vectors . Churning through a couple inequalities in quick fashion,23See pg. 17 of the Mathias and Li report. we obtain
Here, denotes the convex hull. Since this holds for every set of indices , it in particular holds for the set of indices which makes the largest. Thus, .
How to Use Mathias–Li’s Perturbation Theory
The eigenangle perturbation bound (4) can be instantiated in a variety of ways. We briefly sketch two. The first is to bound by its minimum over all , which then gives a bound on (and )
Plugging into (4) and simplifying gives
(5)
This improves on Stewart’s bound (3) by replacing the Crawford number by ; as Mathias and Li show is always smaller than or equal to and can be much much smaller.24Recall that Mathias and Li’s bound first requires us to do a change of variables where and both become diagonal, which can increase by a factor of . Thus, for an apples-to-apples comparison with Stewart’s theory where and are non-diagonal, (5) should be interpreted as .
For the second instantiation (4), we recognize that if an eigenangle is sufficiently well-separated from other eigenangles (relative to the size of the perturbation and ), then we have and . (The precise instantiation of “sufficiently well-separated” requires some tedious algebra; if you’re interested, see Footnote 7 in Mathias and Li’s paper.25You may also be interested in Corollary 2.2 in this preprint by myself and coauthors.) Under this separation condition, (4) then reduces to
(6)
This result improves on Stewart’s result (4) by even more, since we have now replaced the Crawford number by for a sufficiently small perturbation. In fact, a result of this form is nearly as good as one could hope for.26Specifically, the condition number of the eigenangle is , so we know for sufficiently small perturbations we have and is the smallest number for which such a relation holds. Mathias and Li’s theory allows for a statement of this form to be made rigorous for a finite-size perturbation. Again, the only small deficit is the additional factor of “” from the change of variables to diagonal form.
The Elegant Geometry of Generalized Eigenvalue Perturbation Theory
As I said at the start of this post, what fascinates me about this generalized eigenvalue perturbation is the beautiful and elegant geometry. When I saw it for the first time, it felt like a magic trick: a definite generalized eigenvalue problem with real eigenvalues was transformed by sleight of hand into a geometry problem on the complex plane, with solutions involving just a little high school geometry and trigonometry. Upon studying the theory, I began to appreciate it for a different reason. Upon closer examination, the magic trick was revealed to be a sequence of deductions, each logically following naturally from the last. To the pioneers of this subject—Stewart, Sun, Mathias, Li, and others—this sequence of logical deductions was not preordained, and their discovery of this theory doubtlessly required careful thought and leaps of insight. Now that this theory has been discovered, however, we get the benefit of retrospection, and can retell a narrative of this theory where each step follows naturally from the last. When told this way, one almost imagines being able to develop this theory by oneself, where at each stage we appeal to some notion of mathematical elegance (e.g., by treating and symmetrically) or by applying a standard trick (e.g., identifying a pair with the complex number ). Since this theory took several decades to fall into place, we should not let this storytelling exercise fool us into thinking the prospective act of developing a new theory will be as straightforward and linear as this retelling, pruned of dead ends and halts in progress, might suggest.
That said, I do think the development of the perturbation theory of the generalized eigenvalue problem does have a lesson for those of us who seek to develop mathematical theories: be guided by mathematical elegance. At several points in the development of the perturbation theory, we obtained great gains by treating quantities which play a symmetric role in the problem symmetrically in the theory or by treating a pair of real numbers as a complex number and asking how to interpret that complex number. My hope is that this perturbation theory serves as a good example for how letting oneself be guided by intuition, a small array of standard tricks, and a search for elegance can lead one to conceptualize a problem in the right way which leads (after a considerable amount of effort and a few lucky breaks) to a natural solution.
Let’s start our discussion of low-rank matrices with an application. Suppose that there are 1000 weather stations spread across the world, and we record the temperature during each of the 365 days in a year.1I borrow the idea for the weather example from Candes and Plan. If we were to store each of the temperature measurements individually, we would need to store 365,000 numbers. However, we have reasons to believe that significant compression is possible. Temperatures are correlated across space and time: If it’s hot in Arizona today, it’s likely it was warm in Utah yesterday.
If we are particularly bold, we might conjecture that the weather approximately experiences a sinusoidal variation over the course of the year:
(1)
For a station , denotes the average temperature of the station and denotes the maximum deviation above or below this station, signed so that it is warmer than average in the Northern hemisphere during June-August and colder-than-average in the Southern hemisphere during these months. The phase shift is chosen so the hottest (or coldest) day in the year occurs at the appropriate time. This model is clearly grossly inexact: The weather does not satisfy a simple sinusoidal model. However, we might plausibly expect it to be fairly informative. Further, we have massively compressed our data, only needing to store the numbers rather than our full data set of 365,000 temperature values.
Let us abstract this approximation procedure in a linear algebraic way. Let’s collect our weather data into a matrix with 1000 rows, one for each station, and 365 columns, one for each day of the year. The entry corresponding to station and day is the temperature at station on day . The approximation Eq. (1) corresponds to the matrix approximation
(2)
Let us call the matrix on the right-hand side of Eq. (2) for ease of discussion. When presented in this linear algebraic form, it’s less obvious in what way is simpler than , but we know from Eq. (1) and our previous discussion that is much more efficient to store than . This leads us naturally to the following question: Linear algebraically, in what way is simpler than ?
The answer is that the matrix has low rank. The rank of the matrix is whereas almost certainly possesses the maximum possible rank of . This example is suggestive that low-rank approximation, where we approximate a general matrix by one of much lower rank, could be a powerful tool. But there any many questions about how to use this tool and how widely applicable it is. How can we compress a low-rank matrix? Can we use this compressed matrix in computations? How good of a low-rank approximation can we find? What even is the rank of a matrix?
What is Rank?
Let’s do a quick review of the foundations of linear algebra. At the core of linear algebra is the notion of a linear combination. A linear combination of vectors is a weighted sum of the form , where are scalars2In our case, matrices will be comprised of real numbers, making scalars real numbers as well.. A collection of vectors is linearly independent if there is no linear combination of them which produces the zero vector, except for the trivial -weighted linear combination . If are not linearly independent, then they’re linearly dependent.
The column rank of a matrix is the size of the largest possible subset of ‘s columns which are linearly independent. So if the column rank of is , then there is some sub-collection of columns of which are linearly independent. There may be some different sub-collections of columns from that are linearly dependent, but every collection of columns is guaranteed to be linearly dependent. Similarly, the row rank is defined to be the maximum size of any linearly independent collection of rows taken from . A remarkable and surprising fact is that the column rank and row rank are equal. Because of this, we refer to the column rank and row rank simply as the rank; we denote the rank of a matrix by .
Linear algebra is famous for its multiple equivalent ways of phrasing the same underlying concept, so let’s mention one more way of thinking about the rank. Define the column space of a matrix to consist of the set of all linear combinations of its columns. A basis for the column space is a linear independent collection of elements of the column space of the largest possible size. Every element of the column space can be written uniquely as a linear combination of the elements in a basis. The size of a basis for the column space is called the dimension of the column space. With these last definitions in place, we note that the rank of is also equal to the dimension of the column space of . Likewise, if we define the row space of to consist of all linear combinations of ‘s rows, then the rank of is equal to the dimension of ‘s row space.
The upshot is that if a matrix has a small rank, its many columns (or rows) can be assembled as linear combinations from a much smaller collection of columns (or rows). It is this fact that allows a low-rank matrix to be compressed for algorithmically useful ends.
Rank Factorizations
Suppose we have an matrix which is of rank much smaller than both and . As we saw in the introduction, we expect that such a matrix can be compressed to be stored with many fewer than entries. How can this be done?
Let’s work backwards and start with the answer to this question and then see why it works. Here’s a fact: a matrix of rank can be factored as , where is an matrix and is an matrix. In other words, can be factored as a “thin” matrix with columns times a “fat” matrix with rows. We use the symbols and for these factors to stand for “left” and “right”; we emphasize that and are general and matrices, not necessarily possessing any additional structure.3Readers familiar with numerical linear algebra may instinctively want to assume that and are lower and upper triangular; we do not make this assumption. The fact that we write the second term in this factorization as a transposed matrix “” is unimportant: We adopt a convention where we write a fat matrix as the transpose of a thin matrix. This notational choice is convenient allows us to easily distinguish between thin and fat matrices in formulas; this choice of notation is far from universal. We call a factorization such as a rank factorization.4Other terms, such as full rank factorization or rank-revealing factorization, have been been used to describe the same concept. A warning is that the term “rank-revealing factorization” can also refer to a factorization which encodes a good low-rank approximation to rather than a genuine factorization of .
Rank factorizations are useful as we can compactly store by storing its factors and . This reduces the storage requirements of to numbers down from numbers. For example, if we store a rank factorization of the low-rank approximation from our weather example, we need only store 2,730 numbers rather than 365,000. In addition to compressing , we shall soon see that one can rapidly perform many calculations from the rank factorization without ever forming itself. For these reasons, whenever performing computations with a low-rank matrix, your first step should almost always be to express it using a rank factorization. From there, most computations can be done faster and using less storage.
Having hopefully convinced ourselves of the usefulness of rank factorizations, let us now convince ourselves that every rank- matrix does indeed possess a rank factorization where and have columns. As we recalled in the previous section, since has rank , there is a basis of ‘s column space consisting of vectors . Collect these vectors as columns of an matrix . But since the columns of comprise a basis of the column space of , every column of can be written as a linear combination of the columns of . For example, the th column of can be written as a linear combination , where we suggestively use the labels for the scalar multiples in our linear combination. Collecting these coefficients into a matrix with th entry , we have constructed a factorization . (Check this!)
This construction gives us a look at what a rank factorization is doing. The columns of comprise a basis for the column space of and the rows of comprise a basis for the row space of . Once we fix a “column basis” , the “row basis” is comprised of linear combination coefficients telling us how to assemble the columns of as linear combinations of the columns in .5It is worth noting here that a slightly more expansive definition of rank factorization has also proved useful. In the more general definition, a rank factorization is a factorization of the form where is , is , and is . With this definition, we can pick an arbitrary column basis and row basis . Then, there exists a unique nonsingular “middle” matrix such that . Note that this means there exist many different rank factorizations of a matrix since one may pick different column bases for .6This non-uniqueness means one should take care to compute a rank factorization which is as “nice” as possible (say, by making sure and are as well-conditioned as is possible). If one modifies a rank factorization during the course of an algorithm, one should take care to make sure that the rank factorization remains nice. (As an example of what can go wrong, “unbalancing” between the left and right factors in a rank factorization can lead to convergence problems for optimization problems.)
Now that we’ve convinced ourselves that every matrix indeed has a rank factorization, how do we compute them in practice? In fact, pretty much any matrix factorization will work. If you can think of a matrix factorization you’re familiar with (e.g., LU, QR, eigenvalue decomposition, singular value decomposition,…), you can almost certainly use it to compute a rank factorization. In addition, manydedicatedmethods have been developed for the specific purpose of computing rank factorizations which can have appealing properties which make them great for certain applications.
Let’s focus on one particular example of how a classic matrix factorization, the singular value decomposition, can be used to get a rank factorization. Recall that the singular value decomposition (SVD) of a (real) matrix is a factorization where and are an and (real) orthogonal matrices and is a (possibly rectangular) diagonal matrix with nonnegative, descending diagonal entries . These diagonal entries are referred to as the singular values of the matrix . From the definition of rank, we can see that the rank of a matrix is equal to its number of nonzero singular values. With this observation in hand, a rank factorization of can be obtained by letting be the first columns of and being the first rows of (note that the remaining rows of are zero).
Computing with Rank Factorizations
Now that we have a rank factorization in hand, what is it good for? A lot, in fact. We’ve already seen that one can store a low-rank matrix expressed as a rank factorization using only numbers, down from numbers by storing all of its entries. Similarly, if we want to compute the matrix-vector product for a vector of length , we can compute this product as . This reduces the operation count down from operations to operations using the rank factorization. As a general rule of thumb, when we have something expressed as a rank factorization, we can usually expect to reduce our operation count (and storage costs) from something proportional to (or worse) down to something proportional to .
Let’s try something more complicated. Say we want to compute an SVD of . In the previous section, we computed a rank factorization of using an SVD, but suppose now we computed in some other way. Our goal is to “upgrade” the general rank factorization into an SVD of . Computing the SVD of a general matrix requires operations (expressed in big O notation). Can we do better? Unfortunately, there’s a big roadblock for us: We need operations even to write down the matrices and , which already prevents us from achieving an operation count proportional to like we’re hoping for. Fortunately, in most applications, only the first columns of and are important. Thus, we can change our goal to compute a so-called economy SVD of , which is a factorization , where and are and matrices with orthonormal columns and is a diagonal matrix listing the nonzero singular values of in decreasing order.
Let’s see how to upgrade a rank factorization into an economy SVD . Let’s break our procedure into steps:
Compute (economy7The economy QR factorization of an thin matrix is a factorization where is an matrix with orthonormal columns and is a upper triangular matrix. The economy QR factorization is sometimes also called a thin or compact QR factorization, and can be computed in operations.) QR factorizations of and : and . Reader beware: We call the “” factor in the QR factorizations of and to be and , as we have already used the letter to denote the second factor in our rank factorization.
Compute the small matrix .
Compute an SVD of .
Set and .
By following the procedure line-by-line, one can check that indeed the matrices and have orthonormal columns and , so this procedure indeed computes an economy SVD of . Let’s see why this approach is also faster. Let’s count operations line-by-line:
Economy QR factorization of an and matrix require and operations.
The product of two matrices requires operations.
The SVD of an matrix requires operations.
The products of a and a matrix by matrices requires and operations.
Accounting for all the operations, we see the operation count is , a significant improvement over the operations for a general matrix.8We can ignore the term of order since so is .
As the previous examples show, many (if not most) things we want to compute from a low-rank matrix can be dramatically more efficiently computed using its rank factorization. The strategy is simple in principle, but can be subtle to execute: Whatever you do, avoid explicitly computing the product at all costs. Instead, compute with the matrices and directly, only operating on , , and matrices.
Another important type of computation one can perform with low-rank matrices are low-rank updates, where we have already solved a problem for a matrix and we want to re-solve it efficiently with the matrix where has low rank. If is expressed in a rank factorization, very often we can do this efficiently as well, as we discuss in the following bonus section. As this is somewhat more niche, the uninterested reader should feel free to skip this and continue to the next section.
Low-rank Updates
Suppose we’ve solved a system of linear equations by computing an LU factorization of the matrix . We now wish to solve the system of linear equations , where is a low-rank matrix expressed as a rank factorization . Our goal is to do this without recomputing a new factorization from scratch.
The first solution uses the Sherman-Morrison-Woodbury formula, which has a nice proof via the Schur complement and block Gaussian elimination which I described here. In our case, the formula yields
(3)
where and denote the and identity matrices. This formula can easily verified by multiplying with and confirming one indeed recovers the identity matrix. This formula suggests the following approach to solving . First, use our already-computed LU factorization for to compute . (This involves solving linear systems of the form to compute each column of from each column of .) We then compute an LU factorization of the much smaller matrix . Finally, we use our factorization of once more to compute , from which our solution is given by
(4)
The net result is we solved our rank--updated linear system using solutions of the original linear system with no need to recompute any factorizations of matrices. We’ve reduced the solution of the system to an operation count of which is dramatically better than the operation count of recomputing the LU factorization from scratch.
This simple example demonstrates a broader pattern: Usually if a matrix problem took to solve originally, one can usually solve the problem after a rank- update in an additional time of only something like operations.9Sometimes, this goal of can be overly optimistic. For symmetric eigenvalue problems, for instance, the operation count may be a bit larger by a (poly)logarithmic factor—say something like . An operation count like this still represents a dramatic improvement over the operation count of recomputing by scratch. For instance, not only can we solve rank--updated linear systems in operations, but we can actually update the LU factorization itself in operations. Similar updates exist for Cholesky, QR, symmetric eigenvalue, and singular value decompositions to update these factorizations in operations.
An important caveat is that, as always with linear algebraic computations, it’s important to read the fine print. There are many algorithms for computing low-rank updates to different matrix factorizations with dramatically different accuracy properties. Just because in principle rank-updated versions of these factorizations can be computed doesn’t mean it’s always advisable. With this qualification stated, these ways of updating matrix computations with low-rank updates can be a powerful tool in practice and reinforce the computational benefits of low-rank matrices expressed via rank factorizations.
Low-rank Approximation
As we’ve seen, computing with low-rank matrices expressed as rank factorizations can yield significant computational savings. Unfortunately, many matrices in application are not low-rank. In fact, even if a matrix in an application is low-rank, the small rounding errors we incur in storing it on a computer may destroy the matrix’s low rank, increasing its rank to the maximum possible value of . The solution in this case is straightforward: approximate our high-rank matrix with a low-rank one, which we express in algorithmically useful form as a rank factorization.
Here’s one simple way of constructing low-rank approximations. Start with a matrix and compute a singular value decomposition of , . Recall from two sections previous that the rank of the matrix is equal to its number of nonzero singular values. But what if ‘s singular values aren’t exactly zero, but they’re very small? It seems reasonable to expect that is nearly low-rank in this case. Indeed, this intuition is true. To approximate a low-rank matrix, we can truncate ‘s singular value decomposition by setting ‘s small singular values to zero. If we zero out all but the largest singular values of , this procedure results in a rank- matrix which approximates . If the singular values that we zeroed out were tiny, then will be very close to and the low-rank approximation is accurate. This matrix is called an -truncated singular value decomposition of , and it is easy to represent it using a rank factorization once we have already computed an SVD of .
It is important to remember that low-rank approximations are, just as the name says, approximations. Not every matrix is well-approximated by one of small rank. A matrix may be excellently approximated by a rank-100 matrix and horribly approximated by a rank-90 matrix. If an algorithm uses a low-rank approximation as a building block, then the approximation error (the difference between and its low-rank approximation ) and its propagations through further steps of the algorithm need to be analyzed and controlled along with other sources of error in the procedure.
Despite this caveat, low-rank approximations can be startlingly effective. Many matrices occurring in practice can be approximated to negligible error by a matrix with very modestly-sized rank. We shall return to this surprising ubiquity of approximately low-rank matrices at the end of the article.
We’ve seen one method for computing low-rank approximations, the truncated singular value decomposition. As we shall see in the next section, the truncated singular value decomposition produces excellent low-rank approximations, the best possible in a certain sense, in fact. As we mentioned above, almost every matrix factorization can be used to compute rank factorizations. Can these matrix factorizations also compute high quality low-rank approximations?
Let’s consider a specific example to see the underlying ideas. Say we want to compute a low-rank approximation to a matrix by a QR factorization. To do this, we want to compute a QR factorization and then throw away all but the first columns of and the first rows of . This will be a good approximation if the rows we discard from are “small” compared to the rows of we keep. Unfortunately, this is not always the case. As a worst case example, if the first columns of are zero, then the first rows of will definitely be zero and the low-rank approximation computed this way is worthless.
We need to modify something to give QR factorization a fighting chance for computing good low-rank approximations. The simplest way to do this is by using column pivoting, where we shuffle the columns of around to bring columns of the largest size “to the front of the line” as we computing the QR factorization. QR factorization with column pivoting produces excellent low-rank approximations in a large number of cases, but it can still give poor-quality approximations for some special examples. For this reason, numerical analysts have developed so-called strong rank-revealing QR factorizations, such as the one developed by Gu and Eisenstat, which are guaranteed to compute quite good low-rank approximations for every matrix . Similarly, there exists a strong rank-revealing LU factorizations which can compute good low-rank approximations using LU factorization.
The upshot is that most matrix factorizations you know and love can be used to compute good-quality low-rank approximations, possibly requiring extra tricks like row or column pivoting. But this simple summary, and the previous discussion, leaves open important questions: what do we mean by good-quality low-rank approximations? How good can a low-rank approximation be?
Best Low-rank Approximation
As we saw in the last section, one way to approximate a matrix by a lower rank matrix is by a truncated singular value decomposition. In fact, in some sense, this is the best way of approximating a matrix by one of lower rank. This fact is encapsulated in a theorem commonly referred to as the Eckart–Young theorem, though the essence of the result is originally due to Schmidt and the modern version of the result to Mirsky.10A nice history of the Eckart–Young theorem is provided in the book Matrix Perturbation Theory by Stewart and Sun.
But what do we mean by best approximation? One ingredient we need is a way of measuring how big the discrepancy between two matrices is. Let’s define a measure of the size of a matrix which we will call ‘s norm, which we denote as . If is a matrix and is a low-rank approximation to it, then is a good approximation to if the norm is small. There might be many different ways of measuring the size of the error, but we have to insist on a couple of properties on our norm for it to really define a sensible measure of size. For instance if the norm of a matrix is , then the norm of should be . A list of the properties we require a norm to have are listed on the Wikipedia page for norms. We shall also insist on one more property for our norm: the norm should be unitarily invariant.11Note that every unitarily invariant norm is a special type of vector norm (called a symmetric gauge function) evaluated on the singular values of the matrix. What this means is the norm of a matrix remains the same if it is multiplied on the left or right by an orthogonal matrix. This property is reasonable since multiplication by orthogonal matrices geometrically represents a rotation or reflection12This is not true in dimensions higher than 2, but it gives the right intuition that orthogonal matrices preserve distances. which preserves distances between points, so it makes sense that we should demand that the size of a matrix as measured by our norm does not change by such multiplications. Two important and popular matrix norms satisfy the unitarily invariant property: the Frobenius norm and the spectral (or operator 2-) norm , which measures the largest singular value.13Both the Frobenius and spectral norms are examples of an important subclass of unitarily invariant norms called Schatten norms. Another example of a Schatten norm, important in matrix completion, is the nuclear norm (sum of the singular values).
With this preliminary out of the way, the Eckart–Young theorem states that the truncated singular value decomposition of truncated to rank is the closest of all rank- matrices when distances are measured using any unitarily invariant norm . If we let denote the -truncated singular value decomposition of , then the Eckart–Young theorem states that
(5)
Less precisely, the -truncated singular value decomposition is the best rank- approximation to a matrix.
Let’s unpack the Eckart–Young theorem using the spectral and Frobenius norms. In this context, a brief calculation and the Eckart–Young theorem proves that for any rank- matrix , we have
(6)
where are the singular values of . This bound is quite intuitive. The error in low-rank approximation will be “small” when we measure the error in the spectral norm when each singular value we zero out is “small”. When we measure error in the Frobenius norm, the error in low-rank approximation is “small” when all of the singular values we zero out are “small” in aggregate when squared and added together.
The Eckart–Young theorem shows that possessing a good low-rank approximation is equivalent to the singular values rapidly decaying.14At least when measured in unitarily invariant norms. A surprising result shows that even the identity matrix, whose singular values are all equal to one, has good low-rank approximations in the maximum entrywise absolute value norm; see, e.g., Theorem 1.0 in this article. If a matrix does not have nice singular value decay, no good low-rank approximation exists, computed by the -truncated SVD or otherwise.
Why Are So Many Matrices (Approximately) Low-rank?
As we’ve seen, we can perform computations with low-rank matrices represented using rank factorizations much faster than general matrices. But all of this would be a moot point if low-rank matrices rarely occurred in practice. But in fact precisely the opposite is true: Approximately low-rank matrices occur all the time in practice.
Sometimes, exact low-rank matrices appear for algebraic reasons. For instance, when we perform one step Gaussian elimination to compute an factorization, the lower right portion of the eliminated matrix, the so-called Schur complement, is a rank-one update to the original matrix. In such cases, a rank- matrix might appear in a computation when one performs steps of some algebraic process: The appearance of low-rank matrices in such cases is unsurprising.
However, often, matrices appearing in applications are (approximately) low-rank for analytic reasons instead. Consider the weather example from the start again. One might reasonably model the temperature on Earth as a smooth function of position and time . If we then let denote the position on Earth of station and the time representing the th day of a given year, then the entries of the matrix are given by . As discussed in my article on smoothness and degree of approximation, a smooth function function of one variable can be excellently approximated by, say, a polynomial of low degree. Analogously, a smooth function depending on two arguments, such as our function , can be excellently be approximated by a separable expansion of rank :
(7)
Similar to functions of a single variable, the degree to which a function can to be approximated by a separable function of small rank depends on the degree smoothness of the function . Assuming the function is quite smooth, then can be approximated has a separable expansion of small rank . This leads immediately to a low-rank approximation to the matrix given by the rank factorization
(8)
Thus, in the context of our weather example, we see that the data matrix can be expected to be low-rank under the reasonable-sounding assumption that the temperature depends smoothly on space and time.
What does this mean in general? Let’s speak informally. Suppose that the th entries of a matrix are samples from a smooth function for points and . Then we can expect that will be approximately low-rank. From a computational point of view, we don’t need to know a separable expansion for the function or even the form of the function itself: If the smooth function exists and is sampled from it, then is approximately low-rank and we can find a low-rank approximation for using the truncated singular value decomposition.15Note here an important subtlety. A more technically precise version of what we’ve stated here is that: if depending on inputs and is sufficiently smooth for in the product of compact regions and , then an matrix with and will be low-rank in the sense that it can be approximated to accuracy by a rank- matrix where grows slowly as and increase and decreases. Note that, phrased this way, the low-rank property of is asymptotic in the size and and the accuracy . If is not smooth on the entirety of the domain or the size of the domains and grow with and , these asymptotic results may no longer hold. And if and are small enough or is large enough, may not be well approximated by a matrix of small rank. Only when there are enough rows and columns will meaningful savings from low-rank approximation be possible.
This “smooth function” explanation for the prevalence of low-rank matrices is the reason for the appearance of low-rank matrices in fast multipole method-type fast algorithms in computational physics and has been proposed16This article considers piecewise analytic functions rather than smooth functions; the principle is more-or-less the same. as a general explanation for the prevalence of low-rank matrices in data science.
(Another explanation for low-rank structure for highly structured matrices like Hankel, Toeplitz, and Cauchy matrices17Computations with these matrices can often also be accelerated with other approaches than low-rank structure; see my post on the fast Fourier transform for a discussion of fast Toeplitz matrix-vector products. which appear in control theory applications has a different explanation involving a certain Sylvester equation; see this lecture for a great explanation.)
Upshot: A matrix is low-rank if it has many fewer linearly independent columns than columns. Such matrices can be efficiently represented using rank-factorizations, which can be used to perform various computations rapidly. Many matrices appearing in applications which are not genuinely low-rank can be well-approximated by low-rank matrices; the best possible such approximation is given by the truncated singular value decomposition. The prevalence of low-rank matrices in diverse application areas can partially be explained by noting that matrices sampled from smooth functions are approximately low-rank.
In this post, I want to answer a simple question: how can randomness help in solving a deterministic (non-random) problem?
Let’s start by defining some terminology. An algorithm is just a precisely defined procedure to solve a problem. For example, one algorithm to compute the integral of a function on the interval is to pick 100 equispaced points on this interval and output the Riemann sum . A randomized algorithm is simply an algorithm which uses, in some form, randomly chosen quantities. For instance, a randomized algorithm1This randomized algorithm is an example of the strategy of Monte Carlo integration, which has proved particularly effective at computing integrals of functions on high-dimensional spaces. for computing integrals would be to pick 100 random points uniformly distributed in the interval and output the average . To help to distinguish, an algorithm which does not use randomness is called a deterministic algorithm.
This may seem puzzling and even paradoxical. There is no randomness in, say, a system of linear equations, so why should introducing randomness help solve such a problem? It can seem very unintuitive that leaving a decision in an algorithm up to chance would be better than making an informed (and non-random) choice. Why do randomized algorithms work so well; what is the randomness actually buying us?
A partial answer to this question is that it can be very costly to choose the best possible option at runtime for an algorithm. Often, a random choice is “good enough”.
A Case Study: Quicksort
Consider the example of quicksort. Given a list of integers to sort in increasing order, quicksort works by selecting an element to be the pivot and then divides the elements of the list into groups larger and smaller than the pivot. One then recurses on the two groups, ending up with a sorted list.
The best choice of pivot is the median of the list, so one might naturally think that one should use the median as the pivot for quicksort. The problem with this reasoning is that finding the median is time-consuming; even using the fastest possible median-finding algorithm,3There is an algorithm guaranteed to find the median in time, which is asymptotically fast as this problem could possibly be solved since any median-finding algorithm must in general look at the entire list.quicksort with exact median pivot selection isn’t very quick. However, one can show quicksort with a pivot selected (uniformly) at random achieves the same expected runtime as quicksort with optimal pivot selection and is much faster in practice.4In fact, is the fastest possible runtime for any comparison sorting algorithm.
But perhaps we have given up on fast deterministic pivot selection in quicksort too soon. What if we just pick the first entry of the list as the pivot? This strategy usually works fairly well too, but it runs into an unfortunate shortfall: if one feeds in a sorted list, the first-element pivot selection is terrible, resulting in a algorithm. If one feeds in a list in a random order, the first-element pivot selection has a expected runtime,5Another way of implementing randomized quicksort is to first randomize the list and then always pick the first entry as the pivot. This is fully equivalent to using quicksort with random pivot selection in the given ordering. Note that randomizing the order of the list before its sorted is still a randomized algorithm because, naturally, randomness is needed to randomize order. but for certain specific orderings (in this case, e.g., the sorted ordering) the runtime is a disappointing . The first-element pivot selection is particularly bad since its nemesis ordering is the common-in-practice already-sorted ordering. But other simple deterministic pivot selections are equally bad. If one selects, for instance, the pivot to be the entry in the position , then we can still come up with an ordering of the input list that makes the algorithm run in time . And because the algorithm is deterministic, it runs in time every time with such-ordered inputs. If the lists we need to sort in our code just happen to be bad for our deterministic pivot selection strategy, quicksort will be slow every time.
We typically analyze algorithms based on their worst-case performance. And this can often be unfair to algorithms. Some algorithms work excellently in practice but are horribly slow on manufactured examples which never occur in “real life”.6The simplex algorithm is an example of algorithm which is often highly effective in practice, but can take a huge amount of time on some pathological cases. But average-case analysis of algorithms, where one measures the average performance of an algorithm over a distribution of inputs, can be equally misleading. If one were swayed by the average-case analysis of the previous section for quicksort with first-element pivot selection, one would say that this should be an effective pivot selection in practice. But sorting an already-sorted or nearly-sorted list occurs all the time in practice: how many times do programmers read in a sorted list from a data source and sort it again just to be safe? In real life, we don’t encounter random inputs to our algorithms: we encounter a very specific distribution of inputs specific to the application we use the algorithm in. An algorithm with excellent average-case runtime analysis is of little use to me if it is consistently and extremely slow every time I run it on my input data on the problem I’m working on. This discussion isn’t meant to disparage average-case analysis (indeed, very often the “random input” model is not too far off), but to explain why worst-case guarantees for algorithms can still be important.
To summarize, often we have a situation where we have an algorithm which makes some choice (e.g. pivot selection). A particular choice usually works well for a randomly chosen input, but can be stymied by certain specific inputs. However, we don’t get to pick which inputs we receive, and it is fully possible the user will always enter inputs which are bad for our algorithms. Here is where randomized algorithms come in. Since we are unable to randomize the input, we instead randomize the algorithm.
A randomized algorithm can be seen as a random selection from a collection of deterministic algorithms. Each individual deterministic algorithm may be confounded by an input, but most algorithms in the collection will do well on any given input. Thus, by picking a random algorithm from our collection, the probability of poor algorithmic performance is small. And if we run our randomized algorithm on an input and it happens to do poorly by chance, if we run it again it is unlikely to do so; there isn’t a specific input we can create which consistently confounds our randomized algorithm.
The Game Theory Framework
Let’s think about algorithm design as a game. The game has two players: you and an opponent. The game is played by solving a certain algorithmic problem, and the game has a score which quantifies how effectively the problem was solved. For example, the score might be the runtime of the algorithm, the amount of space required, or the error of a computed quantity. For simplicity of discussion, let’s use runtime as an example for this discussion. You have to pay your opponent a dollar for every second of runtime, so you want to have the runtime be as low as possible.7In this scenario, we consider algorithms which always produce the correct answer to the problem, and it is only a matter of how long it takes to do so. In the field of algorithms, randomized algorithms of this type are referred to as Las Vegas algorithms. Algorithms which can also give the wrong answer with some (hopefully small) probability are referred to as Monte Carlo algorithms.
Each player makes one move. Your move is to present an algorithm which solves the problem. Your opponent’s move is to provide an input to your algorithm. Your algorithm is then run on your opponents input and you pay your opponent a dollar for every second of runtime.
This setup casts algorithm design as a two-person, zero-sum game. It is a feature of such games that a player’s optimal strategy is often mixed, picking a move at random subject to some probability distribution over all possible moves.
To see why this is the case, let’s consider a classic and very simple game: rock paper scissors. If I use a deterministic strategy, then I am forced to pick one choice, say rock, and throw it every time. But then my savvy opponent could pick paper and be assured victory. By randomizing my strategy and selecting between rock, paper, and scissors (uniformly) at random, I improve my odds to winning a third of the time, losing a third of the time, and drawing a third of the time. Further, there is no way my opponent can improve their luck by adopting a different strategy. This strategy is referred to as minimax optimal: among all (mixed) strategies I could adopt, this strategy has the best performance over all strategies provided my opponent always counters my strategy with the best response strategy they can find.
Algorithm design is totally analogous. For the pivot selection problem, if I pick a the fixed strategy of choosing the first entry to be the pivot (analogous to always picking rock), then my opponent could always give me a sorted list (analogous to always picking paper) and I would always lose. My randomizing my pivot selection, my opponent could input the list in whatever order they choose and my pivot selection will always have the excellent (expected) runtime characteristic of quicksort. In fact, randomized pivot selection is the minimax optimal pivot selection strategy (assuming pivot selection is non-adaptive in that we don’t choose the pivot based on the values in the list).8This is not to say that quicksort with randomized pivot selection is necessarily the minimax optimal sorting algorithm, but to say that once we have fixed quicksort as our sorting algorithm, randomized pivot selection is the minimax optimal non-adaptive pivot selection strategy for quicksort.
How Much Does Randomness Buy?
Hopefully, now, the utility of randomness is less opaque. Randomization, in effect, allows an algorithm designer to trade algorithm which runs fast for most inputs all of the time for an algorithm which runs fast for all inputs most of the time. They do this by introducing randomized decision-making to hedge against particular bad inputs which could confound their algorithm.
Randomness, of course, is not a panacea. Randomness does not allow us to solve every algorithmic problem arbitrarily fast. But how can we quantify this? Are there algorithmic speed limits for computational problems for which no algorithm, randomized or not, can exceed?
The game theoretic framework for randomized algorithms can shed light on this question. Let us return to the framing of last section where you choose an algorithm , your opponent chooses an input , and you pay your opponent a dollar for every second of runtime. Since this cost depends on the algorithm and the input, we can denote the cost .
Suppose you devise a randomized algorithm for this task, which can be interpreted as selecting an algorithm from a probability distribution over the class of all algorithms which solve the problem. (Less formally, we assign each possible algorithm a probability of picking it and pick one subject to these probabilities.) Once your opponent sees your randomized algorithm (equivalently, the distribution ), they can come up with the on-average slowest possible input and present it to your algorithm.
But now let’s switch places to see things from your opponent. Suppose they choose their own strategy of randomly selecting an input from among all inputs with some distribution . Once you see the distribution , you can come up with the best possible deterministic algorithm to counter their strategy.
The next part gets a bit algebraic. Suppose now we apply your randomized algorithm against your opponents strategy. Then, your randomized algorithm could only take longer on average than because, by construction, is the fastest possible algorithm against the input distribution . Symbolically, we have
(1)
Here, and denote the average (expected) value of a random quantity when an input is drawn randomly with distribution or an algorithm is drawn randomly with distribution . But we know that is the slowest input for our randomized algorithm, so, on average, our randomized algorithm will take longer on worst-case input then a random input from . In symbols,
In words, the average performance of any randomized algorithm on its worst-case input can be no better than the average performance of the best possible deterministic algorithm for adistribution of inputs.10In fact, there is a strengthening of Yao’s minimax principle. That Eqs. (1) and (2) are equalities when and are taken to be the minimax optimal randomized algorithm and input distribution, respectively. Specifically, assuming the cost is a natural number and we restrict to a finite class of potential inputs, This nontrivial fact is a consequence of the full power of the von Neumann’s minimax theorem, which itself is a consequence of strong duality in linear programming. My thanks go to Professor Leonard Schulman for teaching me this result.
This, in effect, allows the algorithm analyst seeking to prove “no randomized algorithm can do better than this” to trade randomness in the algorithm to randomness in the input in their analysis. This is a great benefit because randomness in an algorithm can be used in arbitrarily complicated ways, whereas random inputs can be much easier to understand. To prove any randomized algorithm takes at least cost to solve a problem, the algorithm analyst can find a distribution on which every deterministic algorithm takes at least cost . Note that Yao’s minimax principle is an analytical tool, not an algorithmic tool. Yao’s principle establishes speed limits on the best possible randomized algorithm: it does not imply that one can just use deterministic algorithms and assume or make the input to be random.
There is a fundamental question concerning the power of randomized algorithms not answered by Yao’s principle that is worth considering: how much better can randomized algorithms be than deterministic ones? Could infeasible problems for deterministic computations be solved tractably by randomized algorithms?
Questions such as these are considered in the field of computational complexity theory. In this context, one can think of a problem as tractably solvable if it can be solved in polynomial time—that is, in an amount of time proportional to a polynomial function of the size of the input. Very roughly, we call the class of all such problems to be . If one in addition allows randomness, we call this class of problems .11More precisely, is the class of problems solvable in polynomial time by a Monte Carlo randomized algorithm. The class of problems solvable in polynomial time by a Las Vegas randomized algorithm, which we’ve focused on for the duration of this article, is a possibly smaller class called .
It has widely been conjectured that : all problems that can be tractably solved with randomness can tractably be solved without. There is some convincing evidence for this belief. Thus, provided this conjecture turns out to be true, randomness can give us reductions in operation counts by a polynomial amount in the input size for problems already in , but they cannot efficiently solve -hard computational problems like the traveling salesman problem.12Assuming the also quite believable conjecture that .
So let’s return to the question which is also this article’s title: why randomized algorithms? Because randomized algorithms are often faster. Why, intuitively, is this the case? Randomization can us to upgrade simple algorithms that are great for most inputs to randomized algorithms which are great most of the time for all inputs. How much can randomization buy? A randomized algorithm on its worst input can be no better than a deterministic algorithm on a worst-case distribution. Assuming a widely believed and theoretically supported, but not yet proven, conjecture, randomness can’t make intractable problems into tractable ones. Still, there is great empirical evidence that randomness can be an immensely effective algorithm design tool in practice, including computational math and science problems like trace estimation and solving Laplacian linear systems.