This is my first post to stackoverflow, so if this isn't the correct area I apologize. I am working on minimizing a L1-Regularized System.
This weekend is my first dive into optimization, I have a basic linear system Y = X*B, X is an n-by-p matrix, B is a p-by-1 vector of model coefficients and Y is a n-by-1 output vector.
I am trying to find the model coefficients, I have implemented both gradient descent and coordinate descent algorithms to minimize the L1 Regularized system. To find my step size I am using the backtracking algorithm, I terminate the algorithm by looking at the norm-2 of the gradient and terminating if it is 'close enough' to zero(for now I'm using 0.001).
The function I am trying to minimize is the following (0.5)*(norm((Y - X*B),2)^2) + lambda*norm(B,1). (Note: By norm(Y,2) I mean the norm-2 value of the vector Y) My X matrix is 150-by-5 and is not sparse.
If I set the regularization parameter lambda to zero I should converge on the least squares solution, I can verify that both my algorithms do this pretty well and fairly quickly.
If I start to increase lambda my model coefficients all tend towards zero, this is what I expect, my algorithms never terminate though because the norm-2 of the gradient is always positive number. For example, a lambda of 1000 will give me coefficients in the 10^(-19) range but the norm2 of my gradient is ~1.5, this is after several thousand iterations, While my gradient values all converge to something in the 0 to 1 range, my step size becomes extremely small (10^(-37) range). If I let the algorithm run for longer the situation does not improve, it appears to have gotten stuck somehow.
Both my gradient and coordinate descent algorithms converge on the same point and give the same norm2(gradient) number for the termination condition. They also work quite well with lambda of 0. If I use a very small lambda(say 0.001) I get convergence, a lambda of 0.1 looks like it would converge if I ran it for an hour or two, a lambda any greater and the convergence rate is so small it's useless.
I had a few questions that I think might relate to the problem?
In calculating the gradient I am using a finite difference method (f(x+h) - f(x-h))/(2h)) with an h of 10^(-5). Any thoughts on this value of h?
Another thought was that at these very tiny steps it is traveling back and forth in a direction nearly orthogonal to the minimum, making the convergence rate so slow it is useless.
My last thought was that perhaps I should be using a different termination method, perhaps looking at the rate of convergence, if the convergence rate is extremely slow then terminate. Is this a common termination method?
The 1-norm isn't differentiable. This will cause fundamental problems with a lot of things, notably the termination test you chose; the gradient will change drastically around your minimum and fail to exist on a set of measure zero.
The termination test you really want will be along the lines of "there is a very short vector in the subgradient."
It is fairly easy to find the shortest vector in the subgradient of ||Ax-b||_2^2 + lambda ||x||_1. Choose, wisely, a tolerance eps and do the following steps:
Compute v = grad(||Ax-b||_2^2).
If x[i] < -eps, then subtract lambda from v[i]. If x[i] > eps, then add lambda to v[i]. If -eps <= x[i] <= eps, then add the number in [-lambda, lambda] to v[i] that minimises v[i].
You can do your termination test here, treating v as the gradient. I'd also recommend using v for the gradient when choosing where your next iterate should be.
Related
I am working on a MR-physic simulation written in Matlab which simulates bloch's equations on an defined object. The magnetisation in the object is updated every time-step with the following functions.
function Mt = evolveMtrans(gamma, delta_B, G, T2, Mt0, delta_t)
% this function calculates precession and relaxation of the
% transversal component, Mt, of M
delta_phi = gamma*(delta_B + G)*delta_t;
Mt = Mt0 .* exp(-delta_t*1./T2 - 1i*delta_phi);
end
This function is a very small part of the entire code but is called upon up to 250.000 times and thus slows down the code and the performance of the entire simulation. I have thought about how I can speed up the calculation but haven't come up with a good solution. There is one line that is VERY time consuming and stands for approximately 50% - 60% of the overall simulation time. This is the line,
Mt = Mt0 .* exp(-delta_t*1./T2 - 1i*delta_phi);
where
Mt0 = 512x512 matrix
delta_t = a scalar
T2 = 512x512 matrix
delta_phi = 512x512 matrix
I would be very grateful for any suggestion to speed up this calculation.
More info below,
The function evovleMtrans is called every timestep during the simulation.
The parameters that are used for calling the function are,
gamma = a constant. (gyramagnetic constant)
delta_B = the magnetic field value
G = gradientstrength
T2 = a 512x512 matrix with T2-values for the object
Mstart.r = a 512x512 matrix with the values M.r had the last timestep
delta_t = a scalar with the difference in time since the last calculated M.r
The only parameters of these that changed during the simulation are,
G, Mstart.r and delta_t. The rest do not change their values during the simulation.
The part below is the part in the main code that calls the function.
% update phase and relaxation to calcTime
delta_t = calcTime - Mstart_t;
delta_B = (d-d0)*B0;
G = Sq.Gx*Sq.xGxref + Sq.Gz*Sq.zGzref;
% Precession around B0 (z-axis) and B1 (+-x-axis or +-y-axis)
% is defined clock-wise in a right hand system x, y, z and
% x', y', z (see the Bloch equation, Bloch 1946 and Levitt
% 1997). The x-axis has angle zero and the y-axis has angle 90.
% For flipping/precession around B1 in the xy-plane, z-axis has
% angle zero.
% For testing of precession direction:
% delta_phi = gamma*((ones(size(d)))*1e-6*B0)*delta_t;
M.r = evolveMtrans(gamma, delta_B, G, T2, Mstart.r, delta_t);
M.l = evolveMlong(T1, M0.l, Mstart.l, delta_t);
This is not a surprise.
That "single line" is a matrix equation. It's really 1,024 simultaneous equations.
Per Jannick, that first term means element-wise division, so "delta_t/T[i,j]". Multiplying a matrix by a scalar is O(N^2). Matrix addition is O(N^2). Evaluating exponential of a matrix will be O(N^2).
I'm not sure if I saw a complex argument in there as well. Does that mean complex matricies with real and imaginary entries? Does your equation simplify to real and imaginary parts? That means twice the number of computations.
Your best hope is to exploit symmetry as much as possible. If all your matricies are symmetric, you cut your calculations roughly in half.
Use parallelization if you can.
Algorithm choice can make a big difference, too. If you're using explicit Euler integration, you may have time step limitations due to stability concerns. Is that why you have 250,000 steps? Maybe a larger time step is possible with a more stable integration schema. Think about a higher order adaptive scheme with error correction, like 5th order Runge Kutta.
There are several possibilities to improve the speed of the code but all that I see come with a caveat.
Numerical ode integration
The first possibility would be to change your analytical solution by numerical differential equation solver. This has several advantages
The analytical solution includes the complex exponential function, which is costly to calculate, while the differential equation contains only multiplication and addition. (d/dt u = -a u => u=exp(-at))
There are plenty of built-in solvers for matlab available and they are typically pretty fast (e.g. ode45). The built-ins however all use a variable step size. This improves speed and accuracy but would be a problem if you really need a fixed equally spaced grid of time points. Here are unofficial fixed step solvers.
As a start you could also try to use just an euler step by replacing
M.r = evolveMtrans(gamma, delta_B, G, T2, Mstart.r, delta_t);
by
delta_phi = gamma*(delta_B + G)*t_step;
M.r += M.r .* (1-t_step*1./T2 - 1i*delta_phi);
You can then further improve that by precalculating all constant values, e.g. one_over_T1=1/T1, moving delta_phi out of the loop.
Caveat:
You are bound to a minimum step size or the accuracy suffers. Therefore this is only a good idea if you time-spacing is quite fine.
Less points in time
You should carfully analyze whether you really need so many points in time. It seems somewhat puzzling to me that you need so many points. As you know the full analytical solution you can freely choose how to sample the time and maybe use this to your advantage.
Going fortran
This might seem like a grand step but in my experience basic (simple loops, matrix operations etc.) matlab code can be relatively easily translated to fortran line-by-line. This would be especially helpful in addition to my first point. If you still want to use the full analytical solution probably there is not much to gain here because exp is already pretty fast in matlab.
I've written some Matlab procedures that evaluate orthogonal polynomials, and as a sanity check I was trying to ensure that their dot product would be zero.
But, while I'm fairly sure there's not much that can go wrong, I'm finding myself with a slightly curious behaviour. My test is quite simple:
x = -1:.01:1;
for i0=0:9
v1 = poly(x, i0);
for i1=0:i0
v2 = poly(x,i1);
fprintf('%d, %d: %g\n', i0, i1, v1*v2');
end
end
(Note the dot product v1*v2' needs to be this way round because x is a horizontal vector.)
Now, to cut to the end of the story, I end up with values close to 0 (order of magnitude about 1e-15) for pairs of degrees that add up to an odd number (i.e., i0+i1=2k+1). When i0==i1 I expect the dot product not to be 0, but this also happens when i0+i1=2k, which I didn't expect.
To give you some more details, I initially did this test with Chebyshev polynomials of first kind. Now, they are orthogonal with respect to the weight
1 ./ sqrt(1-x.^2)
which goes to infinity when x goes to 1. So I thought that leaving this term out could be the cause of non-zero dot products.
But then, I did the same test with Legendre polynomials, and I get exactly the same result: when the sum of the degrees is even, the dot product is definitely far from 0 (order of magnitude 1e2).
One last detail, I used the trigonometric formula cos(n*acos(x)) to evaluate the Chebyshev polynomials, and I tried the recursive formula as well as one of the formulas involving the binomial coefficient to evaluate the Legendre polynomials.
Can anyone explain this odd (pun intended) behaviour?
You're being misled by symmetry. Both Chebyshev and Legendre polynomials are eigenfunctions of the parity operator, which means that they can all be classified as either odd or even functions. I guess the same goes for your custom orthogonal polynomials.
Due to this symmetry, if you multiply a polynomial P_n(x) by P_m(x), then the result will be an odd function if n+m is odd, and it will be even otherwise. You're computing sum_k P_n(x_k)*P_m(x_k) for a symmetric set of x_k values around the origin. This implies that for odd n+m you will always get zero. Try computing sum_k P_n(x_k)*Q_m(x_k) with P a Legendre, and Q a Chebyshev polynomial. My point is that for n+m=odd, the result doesn't tell you anything about orthogonality or the accuracy of your integration.
The problem is that probably you're not integrating accurately enough. These orthogonal polynomials defined on [-1,1] vary quite rapidly on their domain, especially close to the boundaries (x==+-1). Try increasing the points of your integration, using a non-equidistant mesh, or a proper integration using integral.
Final note: I'd advise you against calling your functions poly, since that's a MATLAB built-in. (And so is legendre.)
I am processing a set of data using ridge regression. I found a very interesting phenomenon when apply the learned function to data. Namely, when the ridge parameter increases from zero, the test error keeps increasing. But if we penalize small coefficients(set the parameter <0), the test error can even be smaller.
This is my matlab code:
for i = 1:100
beta = ridgePolyRegression(ty_train,tX_train,lambda(i));
sqridge_train_cost(i) = computePolyCostMSE(ty_train,tX_train,beta);
sqridge_test_cost(i) = computePolyCostMSE(ty_valid,tX_valid,beta);
end
plot(lambda,sqridge_test_cost,'color','b');
lambda is the ridge parameter. ty_train is the output of the training data, tX_train is the input of training data. Also, we use a quadratic function regression here.
function [ beta ] = ridgePolyRegression( y,tX,lambda )
X = tX(:,2:size(tX,2));
tX2 = [tX,X.^2];
beta = (tX2'*tX2 + lambda * eye(size(tX2,2))) \ (tX2'*y);
end
The plotted picture is:
Why the error is minimal when lambda is negative? Is it a sign of under-fitting?
You should not use negative lambdas.
From (probabilistic) theoretic point of view, lambda relates to the inverse of variance of parameter prior distribution, and variance can't be negative.
From computational point of view, it can (given it's less that the smallest eigenvalue of the covariance matrix) turn your positive-definite form into an indefinite form, which means you'll have not a maximum, but a saddle point. It also means there are points where your target function is as small (or as big) as you want, so you can reduce loss indefinitely and no minimum / maximum exists at all.
Your optimization algorithm gives you just a stationary point, which will be a global maximum if and only if the form is positive definite.
Short Answer: When lambda is negative, you're actually overfitting your data. Hence, it's reasonable to get much less error.
Long Answer:
The regularization term (or the penalty term as described by many statisticians) aims to penalize the weights (or the betas as written in the coming Eq.) for going too high (overfitting) and going too low (underfitting). Giving you the power to control how your model behaves, and you usually aim the "right fitting" model.
For mathematical intuition, you can check the following Eq. (P. S. Equation is screenshotted from Elements of Statistical Learning by Trevor Hastie et. al)
When you decide to make your lambda negative, the penalty term is indeed turned into a utility term that helps to increase the weights (i.e., overfitting).
Overfitting is, simply, understanding your data along with the features more than you should, because you do not have the whole population yet; therefore, what you understood so far is possibly wrong on a different dataset.
So, you should never be using negative values of lambdas.
I would like to convolve a time-series containing two spikes (call it Spike) with an exponential kernel (k) in MATLAB. Call the convolved response "calcium1". I would like to recover the original spike ("reconSpike") data using deconvolution with the kernel. I am using the following code.
k1=zeros(1,5000);
k1(1:1000)=(1.1.^((1:1000)/100)-(1.1^0.01))/((1.1^10)-1.1^0.01);
k1(1001:5000)=exp(-((1001:5000)-1001)/1000);
k1(1)=k1(2);
spike = zeros(100000,1);
spike(1000)=1;
spike(1100)=1;
calcium1=conv(k1, spike);
reconSpike1=deconv(calcium1, k1);
The problem is that at the end of reconSpike, I get a chunk of very large, high amplitude waves that was not in the original data. Anyone know why and how to fix it?
Thanks!
It works for me if you keep the spike vector the same length as the k1 vector. i.e.:
k1=zeros(1,5000);
k1(1:1000)=(1.1.^((1:1000)/100)-(1.1^0.01))/((1.1^10)-1.1^0.01);
k1(1001:5000)=exp(-((1001:5000)-1001)/1000);
k1(1)=k1(2);
spike = zeros(5000, 1);
spike(1000)=1;
spike(1100)=1;
calcium1=conv(k1, spike);
reconSpike1=deconv(calcium1, k1);
Any reason you made them different?
You are running into either a problem with MATLAB's deconvolution algorithm, or floating point precision problems (or maybe both). I suspect it's floating point precision due to all the divisions and subtractions that take place during the deconvolution, but it might be worth contacting MathWorks directly to ask what they think.
Per MATLAB documentation, if [q,r] = deconv(v,u), then v = conv(u,q)+r must also hold (i.e., the output of deconv should always satisfy this). In your case this is violently violated. Put the following at the end of your script:
[reconSpike1 rem]=deconv(calcium1, k1);
max(conv(k1, reconSpike1) + rem - calcium1)
I get 6.75e227, which is not zero ;-) Next try changing the length of spike to 6000; you will get a small number (~1e-15). Gradually increase the length of spike; the error will get larger and larger. Note that if you put only one non-zero element into your spike, this behavior doesn't happen: the error is always zero. It makes sense; all MATLAB needs to do is divide everything by the same number.
Here's a simple demonstration using random vectors:
v = random('uniform', 1,2,100,1);
u = random('uniform', 1,2,100,1);
[q r] = deconv(v,u);
fprintf('maximum error for length(v) = 100 is %f\n', max(conv(u, q) + r - v))
v = random('uniform', 1,2,1000,1);
[q r] = deconv(v,u);
fprintf('maximum error for length(v) = 1000 is %f\n', max(conv(u, q) + r - v))
The output is:
maximum error for length(v) = 100 is 0.000000
maximum error for length(v) = 1000 is 14.910770
I don't know what you are really trying to accomplish, so it's hard to give further advice. But I'll just point out that if you have a problem where pulses are piling up and you want to extract information about each pulse, this can be a tricky problem. I know some people who work on things like this, so if you want some references let me know and I will ask them.
You should never expect that a deconvolution can simply undo a convolution. This is because the deconvolution is an ill-posed problem.
The problem comes from the fact that the convolution is an integral operator (in the continuous case you write down an integral int f(x) g(x-t) dx or something similar). Now, the inverse of computing an integral (the de-convolution) is to apply a differentiation. Unfortunately, the differential amplifies noise in the input. Thus, if your integral only has slight errors on it (and floating-point inaccuarcies might already be enough), you end up with a total different outcome after differentiation.
There are some possibilities how this amplification can be mitigated but these have to be tried on a per-application basis.
I want to numerically integrate the following:
where
and a, b and β are constants which for simplicity, can all be set to 1.
Neither Matlab using dblquad, nor Mathematica using NIntegrate can deal with the singularity created by the denominator. Since it's a double integral, I can't specify where the singularity is in Mathematica.
I'm sure that it is not infinite since this integral is based in perturbation theory and without the
has been found before (just not by me so I don't know how it's done).
Any ideas?
(1) It would be helpful if you provide the explicit code you use. That way others (read: me) need not code it up separately.
(2) If the integral exists, it has to be zero. This is because you negate the n(y)-n(x) factor when you swap x and y but keep the rest the same. Yet the integration range symmetry means that amounts to just renaming your variables, hence it must stay the same.
(3) Here is some code that shows it will be zero, at least if we zero out the singular part and a small band around it.
a = 1;
b = 1;
beta = 1;
eps[x_] := 2*(a-b*Cos[x])
n[x_] := 1/(1+Exp[beta*eps[x]])
delta = .001;
pw[x_,y_] := Piecewise[{{1,Abs[Abs[x]-Abs[y]]>delta}}, 0]
We add 1 to the integrand just to avoid accuracy issues with results that are near zero.
NIntegrate[1+Cos[(x+y)/2]^2*(n[x]-n[y])/(eps[x]-eps[y])^2*pw[Cos[x],Cos[y]],
{x,-Pi,Pi}, {y,-Pi,Pi}] / (4*Pi^2)
I get the result below.
NIntegrate::slwcon:
Numerical integration converging too slowly; suspect one of the following:
singularity, value of the integration is 0, highly oscillatory integrand,
or WorkingPrecision too small.
NIntegrate::eincr:
The global error of the strategy GlobalAdaptive has increased more than
2000 times. The global error is expected to decrease monotonically after a
number of integrand evaluations. Suspect one of the following: the
working precision is insufficient for the specified precision goal; the
integrand is highly oscillatory or it is not a (piecewise) smooth
function; or the true value of the integral is 0. Increasing the value of
the GlobalAdaptive option MaxErrorIncreases might lead to a convergent
numerical integration. NIntegrate obtained 39.4791 and 0.459541
for the integral and error estimates.
Out[24]= 1.00002
This is a good indication that the unadulterated result will be zero.
(4) Substituting cx for cos(x) and cy for cos(y), and removing extraneous factors for purposes of convergence assessment, gives the expression below.
((1 + E^(2*(1 - cx)))^(-1) - (1 + E^(2*(1 - cy)))^(-1))/
(2*(1 - cx) - 2*(1 - cy))^2
A series expansion in cy, centered at cx, indicates a pole of order 1. So it does appear to be a singular integral.
Daniel Lichtblau
The integral looks like a Cauchy Principal Value type integral (i.e. it has a strong singularity). That's why you can't apply standard quadrature techniques.
Have you tried PrincipalValue->True in Mathematica's Integrate?
In addition to Daniel's observation about integrating an odd integrand over a symmetric range (so that symmetry indicates the result should be zero), you can also do this to understand its convergence better (I'll use latex, writing this out with pen and paper should make it easier to read; it took a lot longer to write than to do, it's not that complicated):
First, epsilon(x)-\epsilon(y)\propto\cos(y)-\cos(x)=2\sin(\xi_+)\sin(\xi_-) where I have defined \xi_\pm=(x\pm y)/2 (so I've rotated the axes by pi/4). The region of integration then is \xi_+ between \pi/\sqrt{2} and -\pi/\sqrt{2} and \xi_- between \pm(\pi/\sqrt{2}-\xi_-). Then the integrand takes the form \frac{1}{\sin^2(\xi_-)\sin^2(\xi_+)} times terms with no divergences. So, evidently, there are second-order poles, and this isn't convergent as presented.
Perhaps you could email the persons who obtained an answer with the cos term and ask what precisely it is they did. Perhaps there's a physical regularisation procedure being employed. Or you could have given more information on the physical origin of this (some sort of second order perturbation theory for some sort of bosonic system?), had that not been off-topic here...
May be I am missing something here, but the integrand
f[x,y]=Cos^2[(x+y)/2]*(n[x]-n[y])/(eps[x]-eps[y]) with n[x]=1/(1+Exp[Beta*eps[x]]) and eps[x]=2(a-b*Cos[x]) is indeed a symmetric function in x and y: f[x,-y]= f[-x,y]=f[x,y].
Therefore its integral over any domain [-u,u]x[-v,v] is zero. No numerical integration seems to be needed here. The result is just zero.