Want to learn Machine Learning (ML), Deep Learning (DL), 5G, 6G, and PYTHON? Check out our schools below! https://www.iitk.ac.in/mwn/MLDL/index.html https://www.iitk.ac.in/mwn/IITK6G/index.html IIT Kanpur Intensive Training School (ITS) on PYTHON for Machine Learning, Neural Networks and Deep Learning 2nd to 22nd December 2023 https://www.iitk.ac.in/mwn/MLDL/index.html Welcome to the IIT Kanpur Intensive Training School (ITS) on PYTHON for Machine Learning, Neural Networks and Deep Learning. This school, spanning over 3 weeks, is an intensive program for scholars, students, faculty members, industry professionals and R&D staff aspiring to get an in-depth exposure and learn hands-on implementation of the cutting edge algorithms and software for Machine Learning (ML), Neural Networks (NN) and Deep Learning (DL) toward projects, research and placements to further their career. The school will include extensive assisted PYTHON programming sessions, and numerous important packages such as NUMPY, LINALG, MATPLOTLIB, SCIKIT-LEARN, TENSORFLOW, KERAS. The school also includes problem solving sessions to prepare for tests/ job interviews in ML and DL. For an extensive list of lecture modules and hands-on programming projects, please check the PROGRAM IIT Kanpur PYTHON + MATLAB Training School on 5G/ 6G Wireless Technologies 2nd to 22nd March 2024 https://www.iitk.ac.in/mwn/IITK6G/index.html Welcome to the PYTHON + MATLAB Training School on 5G/ 6G Wireless Technologies: MIMO, massive MIMO, mmWave MIMO, NOMA, OTFS and IRS. This is a unique cutting edge project based training school that will feature exhaustive lecture modules and several PYTHON/ MATLAB projects/ case-studies on the latest MIMO, Massive MIMO, mmWave MIMO, NOMA technologies for 5G. Also, with 5G systems being rapidly deployed around the globe, there is accelerated development of Beyond 5G (B5G) and 6G wireless technologies that break the barriers of 5G. OTFS is a cutting-edge 2D modulation technique that operates in the Delay-Doppler domain, while Intelligent reflecting surface (IRS) is another novel technology that enables dynamically altering the radio propagation environment to achieve a phenomenal boost in the capacities. Together, these dual technologies are expected to power the shift to 6G. This training school will feature exhaustive tutorial style lecture modules and several intensive supervised PYTHON/ MATLAB/ OCTAVE programming modules so that participants are able to gain practical hands-on experience of working on state-of-the-art 5G/ 6G projects All modules will be held on evenings and weekends for the convenience of participants. The extensive projects and PYTHON programming for MACHINE LEARNING (ML), NEURAL NETWORKS (NN), DEEP LEARNING (DL), 5G and 6G are of significant value to participants of all backgrounds. How do these programs benefit YOU? UG/ PG students: Learn the latest programming techniques in PYTHON and various packages such as NUMPY, LINALG, MATPLOTLIB, PANDAS, SCIKIT-LEARN, SEABORN, TENSORFLOW, KERAS together with practical ML/ NN / DL/ 5G/ 6G knowledge for projects/ thesis and also to gain an unbeatable edge in placements! PhD Scholars/ Faculty members: Use PYTHON, and various packages such as NUMPY, LINALG, MATPLOTLIB, PANDAS, SCIKIT-LEARN, SEABORN, TENSORFLOW, KERAS for research and also to establish virtual labs or for project guidance in ML/ NN / DL/ 5G/ 6G Technologies with real world datasets! Industry Professionals: Take your skills to the next level by learning PYTHON, and various packages such as NUMPY, LINALG, MATPLOTLIB, PANDAS, SCIKITLEARN, SEABORN, TENSORFLOW, KERAS, together with principles of Module Design and Analysis using Practical Data Sets for ML/ NN / DL/ 5G/ 6G Technologies! Learn the latest ML/ NN/ DL/ 5G/ 6G Algorithms, PYTHON Programming & Libraries from experts at IIT Kanpur PYTHON, Machine Learning, Neural Networks, Deep Learning, 5G and 6G are the top skills for placements in leading global MNCs such as Google, Amazon, facebook, Microsoft, Qualcomm, Broadcom, Samsung, Mediatek and many others! Our industry participants have come from top global MNCs, Indian R&D Labs below and many more! Our academic participants have come from IITs, NITs, top national and international educational institutions and more! Contact Us: e-mail: [email protected] Join/ follow us on our social media sites https://www.facebook.com/people/MWN-Lab-IIT-Kanpur/100063786175713/ https://twitter.com/LabMwn https://www.linkedin.com/in/mwn-lab-7236231ba/?originalSubdomain=in
Hello, welcome to another module in this massive open online course. So, we are looking at
optimal MIMO power allocation, alright, and we have demonstrated the solution by using
the KKT conditions, alright. Now, let us do an example to understand this better, alright.
So, what we want to do is, we want to do an example to understand this concept of optimal.
And now, consider the channel matrix. So, let us do an example. So, for this, consider
the MIMO channel. This is our MIMO channel. In fact, this we have already seen, this is
the channel matrix. This is your MIMO, this is your MIMO channel matrix. And let us say,
we have a total power P equal to 4 and noise power sigma square equal to 2. And we have
to come up with the optimal power allocation, alright. So, this 2 can also be written as
3 dB. So, noise power is 3 dB, which means, sigma square equals 2, alright.
And now, we have to optimally allocate this power. This total power is 4. We have to optimally
allocate this power. Remember, once you decompose this MIMO channel as a set of parallel channels,
which we obtain using the singular value decomposition, alright. So, for optimal power allocation,
which maximizes the sum rate, alright, which achieves the capacity of the MIMO channel,
we have to first start with the singular value decomposition.
And the singular value decomposition can be obtained as follows. So, first find the singular
value decomposition of H. Find the singular value decomposition of H. And remember, we
have H equals channel matrix. What I am going to do now, remember, we have to write this
as U sigma V Hermitian, where U contains U orthonormal columns. And these unit norm columns,
that are orthogonal to each other. So, this contains orthonormal content. Now, we can
already see that, these 2 columns are orthogonal to each other. You can see, these 2 columns,
1 1, 2 minus 2, these are already orthogonal. So, all we have to do is, we have to simply
normalize them to unit norm, alright, which means, divide them by their norm.
And that I can simply do as follows. Remember, each column I have to divide it by its norm.
So, the first column, you can see, norm equals square root of 1 plus 1 equals square root
of 2. And second column, the norm is square root of 2 square, that is, 4 plus 4, square
root of 4 plus 4, that is, square root of 8, which is 2 square root of 2. So, the corresponding
orthonormal columns will be, so I can write H as 1 divided by square root of 2, 1 divided
by square root of 2. So, each column I am simply dividing by the norm, 2 divided by
2 square root of 2 minus 2 divided by 2 square root of 2 times. Obviously, since I am dividing
by the norm, I have to also multiply each column by the norm, which is basically diagonal
matrix comprising of the norms, square root of 2, 2 square root of 2, 0 comma.
Now, obviously, you can simplify further. I can simply write this as 1 over square root
of 2 minus 1 over square root of 2, I am sorry, 1 over square root of 2, 1 over square root
of 2 minus 1 over square root of 2 times the product square root of 2, 0. So, we are done
to some extent. So, we have orthonormal columns. Orthonormal columns means, you have columns
which are basically orthogonal to each other and also have unit norm. And therefore, now
you can see, if you look at this matrix, you can see, this satisfies the property of U,
because it has orthonormal columns. Therefore, U Hermitian, U you can see is identity.
And in fact, if you look at this, this can be possibly sigma, because this is a diagonal
matrix and these are non-negative. Remember, these are greater than equal to. So, these
are possible singular values. And now, we have the V matrix. Remember, V is a unitary
matrix. All I need is a unitary matrix and I can simply use the identity matrix in this
case as a unitary matrix, simply multiplied by the 2 cross 2 identity matrix. That is
a unitary matrix. So, I can simply write this as, since I already have U and sigma, I can
simply obtain V by multiplying by the identity matrix. So, this will be 1 minus 1 over square
root of 2, square root of 2. And here, I can simply use the identity matrix. I can simply
use the 2 cross 2 identity matrix. This is the 2 cross 2. And this can be your V Hermitian.
That is it. So, this can be your V Hermitian. Now, the only problem is the following. Now,
if you look at this, this is your U. But, is this valid for sigma? That is, this can
be sigma 1. We can say this can be singular values, sigma 2. Is this valid? Now, the problem
is... Now, of course, you will realize that identity matrix is V, because you can have
V V Hermitian. This satisfies V V Hermitian is V Hermitian V is identity. So, identity
matrix is unitary matrix. Therefore, it satisfies V V Hermitian equals V Hermitian V equals
identity. Now, the problem with this matrix is, if you
look at sigma 1, we have sigma 1 equal to square root of 2. Sigma 2 is equal to square
root of 2, square root of 2. Sigma 1 is less than sigma 2. So, here the problem is, sigma
1 is less than sigma 2. And therefore, not a valid SVD, because this is a unitary matrix.
Because remember, for a singular value decomposition, we have the singular values have to be ordered
in decreasing order. That is, we have sigma 1 greater than equal to sigma 2, greater than
equal to sigma 3 and so on. So, we need... The singular values have to be ordered in
decreasing order. So, we have to somehow switch these two singular values. So, we have everything.
We have the basic structure of the singular value decomposition. Somehow, we have to switch
the singular values. We have to interchange the singular values. And, you will realize
that, that is possible, if I basically interchange the columns of U. And here, I interchange
the rows of V Hermitian. And then, I can flip the singular values. And therefore, I can
equivalently write it as, first I will, as I told you, I will interchange the columns.
So, this will become 1 over square root of 2, minus 1 over square root of 2, 1 over square
root of 2. I can now switch the singular values, 2 square root of 2, 0, diagonal matrix, 2
square root of 2, square root of 2. And finally, I switch the rows of V. So, that becomes 0,
1, 1, 0. And, this is now your matrix U. This is now the matrix V. This is now the matrix
V Hermitian. And, you can see, now it satisfies all the properties. For instance, U Hermitian
U equals identity. V Hermitian V equals V, V Hermitian equals identity. And, sigma 1
equals 2 square root 2, which is greater than 2, equals sigma. So, singular values are arranged
in decreasing order. And, this is the diagonal matrix of, remember, diagonal matrix of singular
values. This is the diagonal matrix of singular values.
And, now we have to do optimal power allocation. Now, perform optimal. Remember, what this
means is, you can decompose this using precoding as the combination of two parallel channels.
So, the first, remember you have this tilde. So, you have x 1 tilde. You are transmitting
through the first channel, that has gain sigma 1 and noise n 1 tilde to give y 1 tilde. And,
what is sigma 1? Sigma 1 is 2 square root of 2. And similarly, through the second channel,
you have gain sigma 2 followed by the addition of the noise n 2 tilde followed by. So, these
are the parallel channels, correct. And, sigma 2 equals square root of 2. So, these are the
parallel channels. These are the parallel channels for a given MIMO channel. In fact,
you will realize that, the given MIMO channel, this is a 2 cross 2 MIMO channel, which we
have not mentioned it, 2 cross 2. And, this is very easy to see. This is a 2 cross 2 MIMO
channel, which means, r equal to t equal to 2 in this scenario. This is r cross t. 2 cross
2 is r cross t. So, we have r equal to t equal to 2. So, this is a 2 cross 2 MIMO channel
with r equal to t equal to 2, fantastic. So, these are the two parallel channels. So, gain
of the first channel is sigma 1 equal to 2 root 2, gain of the second channel. So, sigma
1 equals gain of first channel or you can say, gain of first sub channel. And, sigma
2, this is the gain of the second channel. Now, what is the optimal power allocation?
We have seen optimal power allocation P i for the MIMO channel equals, well, what is
it? 1 over nu minus sigma square, sigma i square. This is the optimal power. This is
the optimal power for channel i. And now, we have, well, what we have? We have sigma
square. Remember, we have seen sigma square equal to 2 and we have sigma 1 equals 2 square
root of 2, sigma 2 equals square root of 2. So, P 1 equals 1 over nu minus sigma square
by sigma 1 square equals 1 over nu minus 2 divided by 2 square root 2 whole square, which
is 8 plus, of course, which means, it is this quantity, which is greater than equal to 0
or it is 0 otherwise, which is basically 1 over nu minus.
Similarly, P 2 equals 1 over nu minus sigma square by sigma 2 square equals, which equals
1 over nu minus 2 divided by 2. Of course, there is always this function plus, which
is 1 over nu minus 1. So, this is optimal power for channel 2 or you can say optimal
power for x tilde. This is optimal power for channel 1, optimal power for the parallel
channel, component parallel channel. So, optimal power for channel 1 and now, how to find nu,
1 over nu or nu? We have to use the total power. To find nu, the Lagrange multiplier,
we have to use the total power. So, to find nu, that becomes, what does that become? That
becomes 1 over nu minus 1 plus 1 over nu minus, I am sorry, 1 over nu minus 1 by 4 plus 1
over nu minus 1 plus equals 4. Now, observe, we do not know, if each of
this quantity is, what each of this quantity is, because remember, we said, if 1 over nu
depends if 1 over nu is, that is, each quantity, each power is equal to this quantity, that
is, this plus indicates that, this quantity plus indicates that, it is equal to this quantity,
which is greater than equal to 0 and 0 otherwise. So, we do not know. Now, that depends again
on the value of 1 over nu, which we are again trying to find. So, this is basically a non-linear
equation. Although, it might appear as a, because of this plus, because of this function
x plus, this makes it a non-linear equation. So, to solve this, we start with an assumption.
Therefore, we have to start with an assumption. Assume that, 1 over nu is greater than equal
to 1. So, start with the assumption. This implies, 1 over nu is greater than equal to
1. This implies that, 1 over nu is greater than equal to 1 by 4. So, p 1 greater than
0, p 2 greater than or greater than equal to 0. So, this implies, 1 over nu minus 1
over 4 plus equals 1 over nu minus 1 over 4, 1 over nu minus 1 plus equals 1 over nu
minus 1. So, therefore, this implies, we must have, now substitute this in the total power,
1 over nu minus 1 by 4 plus 1 over nu minus 1 equals 4.
Now, remember, this is all based on the assumption that, 1 over nu is greater than equal to 1.
If that assumption does not hold, then the entire procedure here, so the values derived
will not hold. So, this is condition of the assumption that, 1 over nu is greater than
equal to 1. So, this implies that, 2 over nu is 4 plus 1 plus 1 over 4 plus 1 over 4
which is 5 plus 1 over 4, which is 21 over 4. So, that is your 2 over, 2 over nu is 21
over 4, which implies 1 over nu is 21 over 8. So, this is the value of 1 over nu. We
do not need to really compute nu, because we only need 1 over nu. So, that is what we
need, which implies, if you look at p 2, this is equal to 1 over nu minus 1, which is 21
by 8 minus 1. So, that will be 13 by 8 and p 1 is 1 over nu minus 1 over 4, which is
21 by 8 minus 1 over 4, which is 19 over 8. Now, you can see that, both p 2 is greater
than 0, p 1 greater than 0, which implies and you can see, p 2 equals 13 by 8 greater
than 1, greater than 0. So, both powers p 2 greater than 0 and p 1 greater than 0. So,
and you can also see that, 21 over 8, this is greater than your 1. So, the original assumption
holds, 1 over nu is greater than 1, greater than 1. So, original assumption holds, p 2
greater than 0, p 1 greater than 0, implies that, optimal powers are indeed p 1 star equals,
well, p 1 star equals 19 by 8 and p 2 star equals, p 2 star equals 1.
So, 13 by 8. If one of the powers would have been negative, for instance, let us say, p
2 star would have been negative, then that implies, original assumption is incorrect.
So, therefore, we have to assume something else and then redo it. So, here we are able
to show that, the powers p 1 star and p 2 star are indeed non-negative, indeed they
are positive. So, therefore, the original assumption is correct. Non-negative powers
are indeed allocated to both the channels. Otherwise, if one of the power is negative,
it implies that, that power is, that is less than, remember, there is this concept of notion
of water level. So, if the power is negative, it implies that, that corresponding channel
is above the water level. So, power is not allocated. So, power, so the corresponding
channel, the power has to be set to 0 and the problem has to be repeated with the total
power constant. So, this is the procedure.
And now, you also observe that, observe p 1 star greater than equal to p 2 star, implies
there is more power, more power to the stronger channel. Remember, this is what our water
filling algorithm tells us. For instance, if you remember our water filling algorithm,
you have these different bars, correct and the water level indicates the power. So, if
this is the level, this is 1 over nu, the power. So, this is the power and this is the
strongest channel. And, what you see is that, which is indeed true is that, more power,
because there is a bar, corresponding bar is smaller, more power. So, what this shows
is, to maximize capacity, because of the nature of water filling power allocation, more power
is allocated to the stronger channel. In this case, the stronger channel will be
the one with the larger singular value. So, sigma 1 is greater than equal to sigma 2 is
greater than equal to sigma 3, if there is 1 and so on. So, which means, more power will
be allocated to the channel with gain sigma 1, in comparison to the channel with gain
sigma 2 and so on and so forth. And, that is what indeed we observe in this exercise
of or example of optimal power allocation, all right. So, we will stop here and continue
in the subsequent modules.