I am currently in an online class in genomics, coming in as a wetlab physician, so my statistical knowledge is not the best. Right now we are working on PCA and SVD in R. I got a big matrix:
head(mat)
ALL_GSM330151.CEL ALL_GSM330153.CEL ALL_GSM330154.CEL ALL_GSM330157.CEL ALL_GSM330171.CEL ALL_GSM330174.CEL ALL_GSM330178.CEL ALL_GSM330182.CEL
ENSG00000224137 5.326553 3.512053 3.455480 3.472999 3.639132 3.391880 3.282522 3.682531
ENSG00000153253 6.436815 9.563955 7.186604 2.946697 6.949510 9.095092 3.795587 11.987291
ENSG00000096006 6.943404 8.840839 4.600026 4.735104 4.183136 3.049792 9.736803 3.338362
ENSG00000229807 3.322499 3.263655 3.406379 9.525888 3.595898 9.281170 8.946498 3.473750
ENSG00000138772 7.195113 8.741458 6.109578 5.631912 5.224844 3.260912 8.889246 3.052587
ENSG00000169575 7.853829 10.428492 10.512497 13.041571 10.836815 11.964498 10.786381 11.953912
Those are just the first few columns and rows, it has 60 columns and 1000 rows. Columns are cancer samples, rows are genes
The task is to:
removing the eigenvectors and reconstructing the matrix using SVD, then we need to calculate the reconstruction error as the difference between the original and the reconstructed matrix. HINT: You have to use the svd() function and equalize the eigenvalue to $0$ for the component you want to remove.
I have been all over google, but can't find a way to solve this task, which might be because I don't really get the question itself.
so i performed SVD on my matrix m:
d <- svd(mat)
Which gives me 3 matrices (Eigenassays, Eigenvalues and Eigenvectors), which i can access using d$u and so on.
How do I equalize the eigenvalue and ultimately calculate the error?
https://www.rdocumentation.org/packages/base/versions/3.6.2/topics/svd
the decomposition expresses your matrix mat as a product of 3 matrices
mat = d$u x diag(d$d) x t(d$v)
so first confirm you are able to do the matrix multiplications to get back mat
once you are able to do this, set the last couple of elements of d$d to zero before doing the matrix multiplication
It helps to create a function that handles the singular values.
Here, for instance, is one that zeros out any singular value that is too small compared to the largest singular value:
zap <- function(d, digits = 3) ifelse(d < 10^(-digits) * max(abs(d))), 0, d)
Although mathematically all singular values are guaranteed non-negative, numerical issues with floating point algorithms can--and do--create negative singular values, so I have prophylactically wrapped the singular values in a call to abs.
Apply this function to the diagonal matrix in the SVD of a matrix X and reconstruct the matrix by multiplying the components:
X. <- with(svd(X), u %*% diag(zap(d)) %*% t(v))
There are many ways to assess the reconstruction error. One is the Frobenius norm of the difference,
sqrt(sum((X - X.)^2))
Rao QE is a weighted Euclidian distance matrix. I have the vectors for the elements of the d_ijs in a data table dt, one column per element (say there are x of them). p is the final column. nrow = S. The double sums are for the lower left (or upper right since it is symmetric) elements of the distance matrix.
If I only needed an unweighted distance matrix I could simply do dist() over the x columns. How do I weight the d_ijs by the product of p_i and p_j?
And example data set is at https://github.com/GeraldCNelson/nutmod/blob/master/RaoD_example.csv with the ps in the column called foodQ.ratio.
You still start with dist for the raw Euclidean distance matrix. Let it be D. As you will read from R - How to get row & column subscripts of matched elements from a distance matrix, a "dist" object is not a real matrix, but a 1D array. So first do D <- as.matrix(D) or D <- dist2mat(D) to convert it to a complete matrix before the following.
Now, let p be the vector of weights, the Rao's QE is just a quadratic form q'Dq / 2:
c(crossprod(p, D %*% p)) / 2
Note, I am not doing everything in the most efficient way. I have performed a symmetric matrix-vector multiplication D %*% p using the full D rather than just its lower triangular part. However, R does not have a routine doing triangular matrix-vector multiplication. So I compute the full version than divide 2.
This doubles computation amount that is necessary; also, making D a full matrix doubles memory costs. But if your problem is small to medium size this is absolutely fine. For large problem, if you are R and C wizard, call BLAS routine dtrmv or even dtpmv for the triangular matrix-vector computation.
Update
I just found this simple paper: Rao's quadratic entropy as a measure of functional diversity based on multiple traits for definition and use of Rao's EQ. It mentions that we can replace Euclidean distance with Mahalanobis distance. In case we want to do this, use my code in Mahalanobis distance of each pair of observations for fast computation of Mahalanobis distance matrix.
I have two square matrices A and B. A is symmetric, B is symmetric positive definite. I would like to compute $trace(A.B^{-1})$. For now, I compute the Cholesky decomposition of B, solve for C in the equation $A=C.B$ and sum up the diagonal elements.
Is there a more efficient way of proceeding?
I plan on using Eigen. Could you provide an implementation if the matrices are sparse (A can often be diagonal, B is often band-diagonal)?
If B is sparse, it may be efficient (i.e., O(n), assuming good condition number of B) to solve for x_i in
B x_i = a_i
(sample Conjugate Gradient code is given on Wikipedia). Taking a_i to be the column vectors of A, you get the matrix B^{-1} A in O(n^2). Then you can sum the diagonal elements to get the trace. Generally, it's easier to do this sparse inverse multiplication than to get the full set of eigenvalues. For comparison, Cholesky decomposition is O(n^3). (see Darren Engwirda's comment below about Cholesky).
If you only need an approximation to the trace, you can actually reduce the cost to O(q n) by averaging
r^T (A B^{-1}) r
over q random vectors r. Usually q << n. This is an unbiased estimate provided that the components of the random vector r satisfy
< r_i r_j > = \delta_{ij}
where < ... > indicates an average over the distribution of r. For example, components r_i could be independent gaussian distributed with unit variance. Or they could be selected uniformly from +-1. Typically the trace scales like O(n) and the error in the trace estimate scales like O(sqrt(n/q)), so the relative error scales as O(sqrt(1/nq)).
If generalized eigenvalues are more efficient to compute, you can compute the generalized eigenvalues, A*v = lambda* B *v and then sum up all the lambdas.
How does it actually reduce noise..can you suggest some nice tutorials?
SVD can be understood from a geometric sense for square matrices as a transformation on a vector.
Consider a square n x n matrix M multiplying a vector v to produce an output vector w:
w = M*v
The singular value decomposition M is the product of three matrices M=U*S*V, so w=U*S*V*v. U and V are orthonormal matrices. From a geometric transformation point of view (acting upon a vector by multiplying it), they are combinations of rotations and reflections that do not change the length of the vector they are multiplying. S is a diagonal matrix which represents scaling or squashing with different scaling factors (the diagonal terms) along each of the n axes.
So the effect of left-multiplying a vector v by a matrix M is to rotate/reflect v by M's orthonormal factor V, then scale/squash the result by a diagonal factor S, then rotate/reflect the result by M's orthonormal factor U.
One reason SVD is desirable from a numerical standpoint is that multiplication by orthonormal matrices is an invertible and extremely stable operation (condition number is 1). SVD captures any ill-conditioned-ness in the diagonal scaling matrix S.
One way to use SVD to reduce noise is to do the decomposition, set components that are near zero to be exactly zero, then re-compose.
Here's an online tutorial on SVD.
You might want to take a look at Numerical Recipes.
Singular value decomposition is a method for taking an nxm matrix M and "decomposing" it into three matrices such that M=USV. S is a diagonal square (the only nonzero entries are on the diagonal from top-left to bottom-right) matrix containing the "singular values" of M. U and V are orthogonal, which leads to the geometric understanding of SVD, but that isn't necessary for noise reduction.
With M=USV, we still have the original matrix M with all its noise intact. However, if we only keep the k largest singular values (which is easy, since many SVD algorithms compute a decomposition where the entries of S are sorted in nonincreasing order), then we have an approximation of the original matrix. This works because we assume that the small values are the noise, and that the more significant patterns in the data will be expressed through the vectors associated with larger singular values.
In fact, the resulting approximation is the most accurate rank-k approximation of the original matrix (has the least squared error).
To answer to the tittle question: SVD is a generalization of eigenvalues/eigenvectors to non-square matrices.
Say,
$X \in N \times p$, then the SVD decomposition of X yields X=UDV^T where D is diagonal and U and V are orthogonal matrices.
Now X^TX is a square matrice, and the SVD decomposition of X^TX=VD^2V where V is equivalent to the eigenvectors of X^TX and D^2 contains the eigenvalues of X^TX.
SVD can also be used to greatly ease global (i.e. to all observations simultaneously) fitting of an arbitrary model (expressed in an formula) to data (with respect to two variables and expressed in a matrix).
For example, data matrix A = D * MT where D represents the possible states of a system and M represents its evolution wrt some variable (e.g. time).
By SVD, A(x,y) = U(x) * S * VT(y) and therefore D * MT = U * S * VT
then D = U * S * VT * MT+ where the "+" indicates a pseudoinverse.
One can then take a mathematical model for the evolution and fit it to the columns of V, each of which are a linear combination the components of the model (this is easy, as each column is a 1D curve). This obtains model parameters which generate M? (the ? indicates it is based on fitting).
M * M?+ * V = V? which allows residuals R * S2 = V - V? to be minimized, thus determining D and M.
Pretty cool, eh?
The columns of U and V can also be inspected to glean information about the data; for example each inflection point in the columns of V typically indicates a different component of the model.
Finally, and actually addressing your question, it is import to note that although each successive singular value (element of the diagonal matrix S) with its attendant vectors U and V does have lower signal to noise, the separation of the components of the model in these "less important" vectors is actually more pronounced. In other words, if the data is described by a bunch of state changes that follow a sum of exponentials or whatever, the relative weights of each exponential get closer together in the smaller singular values. In other other words the later singular values have vectors which are less smooth (noisier) but in which the change represented by each component are more distinct.
Is there a way, given a set of values (x,f(x)), to find the polynomial of a given degree that best fits the data?
I know polynomial interpolation, which is for finding a polynomial of degree n given n+1 data points, but here there are a large number of values and we want to find a low-degree polynomial (find best linear fit, best quadratic, best cubic, etc.). It might be related to least squares...
More generally, I would like to know the answer when we have a multivariate function -- points like (x,y,f(x,y)), say -- and want to find the best polynomial (p(x,y)) of a given degree in the variables. (Specifically a polynomial, not splines or Fourier series.)
Both theory and code/libraries (preferably in Python, but any language is okay) would be useful.
Thanks for everyone's replies. Here is another attempt at summarizing them. Pardon if I say too many "obvious" things: I knew nothing about least squares before, so everything was new to me.
NOT polynomial interpolation
Polynomial interpolation is fitting a polynomial of degree n given n+1 data points, e.g. finding a cubic that passes exactly through four given points. As said in the question, this was not want I wanted—I had a lot of points and wanted a small-degree polynomial (which will only approximately fit, unless we've been lucky)—but since some of the answers insisted on talking about it, I should mention them :) Lagrange polynomial, Vandermonde matrix, etc.
What is least-squares?
"Least squares" is a particular definition/criterion/"metric" of "how well" a polynomial fits. (There are others, but this is simplest.) Say you are trying to fit a polynomial
p(x,y) = a + bx + cy + dx2 + ey2 + fxy
to some given data points (xi,yi,Zi) (where "Zi" was "f(xi,yi)" in the question). With least-squares the problem is to find the "best" coefficients (a,b,c,d,e,f), such that what is minimized (kept "least") is the "sum of squared residuals", namely
S = ∑i (a + bxi + cyi + dxi2 + eyi2 + fxiyi - Zi)2
Theory
The important idea is that if you look at S as a function of (a,b,c,d,e,f), then S is minimized at a point at which its gradient is 0. This means that for example ∂S/∂f=0, i.e. that
∑i2(a + … + fxiyi - Zi)xiyi = 0
and similar equations for a, b, c, d, e.
Note that these are just linear equations in a…f. So we can solve them with Gaussian elimination or any of the usual methods.
This is still called "linear least squares", because although the function we wanted was a quadratic polynomial, it is still linear in the parameters (a,b,c,d,e,f). Note that the same thing works when we want p(x,y) to be any "linear combination" of arbitrary functions fj, instead of just a polynomial (= "linear combination of monomials").
Code
For the univariate case (when there is only variable x — the fj are monomials xj), there is Numpy's polyfit:
>>> import numpy
>>> xs = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
>>> ys = [1.1, 3.9, 11.2, 21.5, 34.8, 51, 70.2, 92.3, 117.4, 145.5]
>>> p = numpy.poly1d(numpy.polyfit(xs, ys, deg=2))
>>> print p
2
1.517 x + 2.483 x + 0.4927
For the multivariate case, or linear least squares in general, there is SciPy. As explained in its documentation, it takes a matrix A of the values fj(xi). (The theory is that it finds the Moore-Penrose pseudoinverse of A.) With our above example involving (xi,yi,Zi), fitting a polynomial means the fj are the monomials x()y(). The following finds the best quadratic (or best polynomial of any other degree, if you change the "degree = 2" line):
from scipy import linalg
import random
n = 20
x = [100*random.random() for i in range(n)]
y = [100*random.random() for i in range(n)]
Z = [(x[i]+y[i])**2 + 0.01*random.random() for i in range(n)]
degree = 2
A = []
for i in range(n):
A.append([])
for xd in range(degree+1):
for yd in range(degree+1-xd):
A[i].append((x[i]**xd)*(y[i]**yd)) #f_j(x_i)
c,_,_,_ = linalg.lstsq(A,Z)
j = 0
for xd in range(0,degree+1):
for yd in range(0,degree+1-xd):
print " + (%.2f)x^%dy^%d" % (c[j], xd, yd),
j += 1
prints
+ (0.01)x^0y^0 + (-0.00)x^0y^1 + (1.00)x^0y^2 + (-0.00)x^1y^0 + (2.00)x^1y^1 + (1.00)x^2y^0
so it has discovered that the polynomial is x2+2xy+y2+0.01. [The last term is sometimes -0.01 and sometimes 0, which is to be expected because of the random noise we added.]
Alternatives to Python+Numpy/Scipy are R and Computer Algebra Systems: Sage, Mathematica, Matlab, Maple. Even Excel might be able to do it. Numerical Recipes discusses methods to implement it ourselves (in C, Fortran).
Concerns
It is strongly influenced by how the points are chosen. When I had x=y=range(20) instead of the random points, it always produced 1.33x2+1.33xy+1.33y2, which was puzzling... until I realised that because I always had x[i]=y[i], the polynomials were the same: x2+2xy+y2 = 4x2 = (4/3)(x2+xy+y2). So the moral is that it is important to choose the points carefully to get the "right" polynomial. (If you can chose, you should choose Chebyshev nodes for polynomial interpolation; not sure if the same is true for least squares as well.)
Overfitting: higher-degree polynomials can always fit the data better. If you change the degree to 3 or 4 or 5, it still mostly recognizes the same quadratic polynomial (coefficients are 0 for higher-degree terms) but for larger degrees, it starts fitting higher-degree polynomials. But even with degree 6, taking larger n (more data points instead of 20, say 200) still fits the quadratic polynomial. So the moral is to avoid overfitting, for which it might help to take as many data points as possible.
There might be issues of numerical stability I don't fully understand.
If you don't need a polynomial, you can obtain better fits with other kinds of functions, e.g. splines (piecewise polynomials).
Yes, the way this is typically done is by using least squares. There are other ways of specifying how well a polynomial fits, but the theory is simplest for least squares. The general theory is called linear regression.
Your best bet is probably to start with Numerical Recipes.
R is free and will do everything you want and more, but it has a big learning curve.
If you have access to Mathematica, you can use the Fit function to do a least squares fit. I imagine Matlab and its open source counterpart Octave have a similar function.
For (x, f(x)) case:
import numpy
x = numpy.arange(10)
y = x**2
coeffs = numpy.polyfit(x, y, deg=2)
poly = numpy.poly1d(coeffs)
print poly
yp = numpy.polyval(poly, x)
print (yp-y)
Bare in mind that a polynomial of higher degree ALWAYS fits the data better. Polynomials of higher degree typically leads to highly improbable functions (see Occam's Razor), though (overfitting). You want to find a balance between simplicity (degree of polynomial) and fit (e.g. least square error). Quantitatively, there are tests for this, the Akaike Information Criterion or the Bayesian Information Criterion. These tests give a score which model is to be prefered.
If you want to fit the (xi, f(xi)) to an polynomial of degree n then you would set up a linear least squares problem with the data (1, xi, xi, xi^2, ..., xi^n, f(xi) ). This will return a set of coefficients (c0, c1, ..., cn) so that the best fitting polynomial is *y = c0 + c1 * x + c2 * x^2 + ... + cn * x^n.*
You can generalize this two more than one dependent variable by including powers of y and combinations of x and y in the problem.
Lagrange polynomials (as #j w posted) give you an exact fit at the points you specify, but with polynomials of degree more than say 5 or 6 you can run into numerical instability.
Least squares gives you the "best fit" polynomial with error defined as the sum of squares of the individual errors. (take the distance along the y-axis between the points you have and the function that results, square them, and sum them up) The MATLAB polyfit function does this, and with multiple return arguments, you can have it automatically take care of scaling/offset issues (e.g. if you have 100 points all between x=312.1 and 312.3, and you want a 6th degree polynomial, you're going to want to calculate u = (x-312.2)/0.1 so the u-values are distributed between -1 and +=).
NOTE that the results of least-squares fits are strongly influenced by the distribution of x-axis values. If the x-values are equally spaced, then you'll get larger errors at the ends. If you have a case where you can choose the x values and you care about the maximum deviation from your known function and an interpolating polynomial, then the use of Chebyshev polynomials will give you something that is close to the perfect minimax polynomial (which is very hard to calculate). This is discussed at some length in Numerical Recipes.
Edit: From what I gather, this all works well for functions of one variable. For multivariate functions it is likely to be much more difficult if the degree is more than, say, 2. I did find a reference on Google Books.
at college we had this book which I still find extremely useful: Conte, de Boor; elementary numerical analysis; Mc Grow Hill. The relevant paragraph is 6.2: Data Fitting.
example code comes in FORTRAN, and the listings are not very readable either, but the explanations are deep and clear at the same time. you end up understanding what you are doing, not just doing it (as is my experience of Numerical Recipes).
I usually start with Numerical Recipes but for things like this I quickly have to grab Conte-de Boor.
maybe better posting some code... it's a bit stripped down, but the most relevant parts are there. it relies on numpy, obviously!
def Tn(n, x):
if n==0:
return 1.0
elif n==1:
return float(x)
else:
return (2.0 * x * Tn(n - 1, x)) - Tn(n - 2, x)
class ChebyshevFit:
def __init__(self):
self.Tn = Memoize(Tn)
def fit(self, data, degree=None):
"""fit the data by a 'minimal squares' linear combination of chebyshev polinomials.
cfr: Conte, de Boor; elementary numerical analysis; Mc Grow Hill (6.2: Data Fitting)
"""
if degree is None:
degree = 5
data = sorted(data)
self.range = start, end = (min(data)[0], max(data)[0])
self.halfwidth = (end - start) / 2.0
vec_x = [(x - start - self.halfwidth)/self.halfwidth for (x, y) in data]
vec_f = [y for (x, y) in data]
mat_phi = [numpy.array([self.Tn(i, x) for x in vec_x]) for i in range(degree+1)]
mat_A = numpy.inner(mat_phi, mat_phi)
vec_b = numpy.inner(vec_f, mat_phi)
self.coefficients = numpy.linalg.solve(mat_A, vec_b)
self.degree = degree
def evaluate(self, x):
"""use Clenshaw algorithm
http://en.wikipedia.org/wiki/Clenshaw_algorithm
"""
x = (x-self.range[0]-self.halfwidth) / self.halfwidth
b_2 = float(self.coefficients[self.degree])
b_1 = 2 * x * b_2 + float(self.coefficients[self.degree - 1])
for i in range(2, self.degree):
b_1, b_2 = 2.0 * x * b_1 + self.coefficients[self.degree - i] - b_2, b_1
else:
b_0 = x*b_1 + self.coefficients[0] - b_2
return b_0
Remember, there's a big difference between approximating the polynomial and finding an exact one.
For example, if I give you 4 points, you could
Approximate a line with a method like least squares
Approximate a parabola with a method like least squares
Find an exact cubic function through these four points.
Be sure to select the method that's right for you!
It's rather easy to scare up a quick fit using Excel's matrix functions if you know how to represent the least squares problem as a linear algebra problem. (That depends on how reliable you think Excel is as a linear algebra solver.)
The lagrange polynomial is in some sense the "simplest" interpolating polynomial that fits a given set of data points.
It is sometimes problematic because it can vary wildly between data points.