Integrating of an exponential of a matrix product

Fb.Researcher
Messages
9
Reaction score
0

Homework Statement


I try to solve this integral with with parameter x as a member of this scale:(-∞ , +∞)
I=∫∏dx exp(-0.5XAX + XB)=∫∏dx exp( Ʃ-0.5xa[j]x[j] +Ʃ xb )
In which a[j] and b are components of telated matrix and vector and the first sum is on i and j ranges from 1 to N .Also X and B are two vector with N component and A is a N*N matrix,so the integral is over all x (which denote components of X).

Homework Equations


Gaussian integrals in the same scale obey this equation:
∫dx exp( -ax^2 ) =√(π/a)

The Attempt at a Solution



Using a change in variables like X=Y + CB with CA=Ac=1 should be appropriate:

-0.5(XAX) + XB=(-0.5)(Y+CB)A(Y+CB)+(Y+CB)B=(-0.5)(YAY+YACB+CBAY+(CB)^2)+YB+CB^2
= (-0.5)(YAY+YACB+CBAY+(CB)^2)+YACB+CB^2
=(-0.5)(YAY+CBAY+(CB)^2-YACB + 2CBB)

The problem is there is no YY=Ʃyy to help me change this problem to a Gaussian problem.Another problem is with YACB and CBAY that arenot compatible with this equation:
(A+B)^2=A^2+B^2+AB+BA .

Thank you for noticing.
 
Physics news on Phys.org
Fb.Researcher said:

Homework Statement


I try to solve this integral with with parameter x as a member of this scale:(-∞ , +∞)
I=∫∏dx exp(-0.5XAX + XB)=∫∏dx exp( Ʃ-0.5xa[j]x[j] +Ʃ xb )
In which a[j] and b are components of telated matrix and vector and the first sum is on i and j ranges from 1 to N .Also X and B are two vector with N component and A is a N*N matrix,so the integral is over all x (which denote components of X).

Homework Equations


Gaussian integrals in the same scale obey this equation:
∫dx exp( -ax^2 ) =√(π/a)

The Attempt at a Solution



Using a change in variables like X=Y + CB with CA=Ac=1 should be appropriate:

-0.5(XAX) + XB=(-0.5)(Y+CB)A(Y+CB)+(Y+CB)B=(-0.5)(YAY+YACB+CBAY+(CB)^2)+YB+CB^2
= (-0.5)(YAY+YACB+CBAY+(CB)^2)+YACB+CB^2
=(-0.5)(YAY+CBAY+(CB)^2-YACB + 2CBB)

The problem is there is no YY=Ʃyy to help me change this problem to a Gaussian problem.Another problem is with YACB and CBAY that arenot compatible with this equation:
(A+B)^2=A^2+B^2+AB+BA .

Thank you for noticing.


Your exponent is -(1/2)Q, where Q = <x,Ax> - 2<b,x> (writing the inner product sum u[j]*v[j] as <u,v>). We need to assume A is *symmetric*; if not, we can replace it by a new A that IS symmetric, and work with that. Writing x = y + c (c a constant vector) we have Q = <y+c,A(y+c)> -2<b,y+c> = <y,Ay> + <y,Ac> + <c,Ay> + <c,Ac> - 2<b,y> = <y,Ay> + <c,Ac> +<y,2Ac-2b>. If we take Ac=b we eliminate linear terms and just have Q = <y,Ay> + <c,Ac>.

Now, we must assume A is positive-definite; otherwise, the integral will not be convergent. It follows that A is invertible, so c = Inverse(A)b can be found. Furthermore, we can perform a Cholesky factorization of A, to write A = U^T U, where U^T = transpose of U and U is upper-triangular with all u[i,i] > 0. Thus, if U[1] = u[1,1]y[1] + u[1,2]y[2] + ... + u[1,n]y[n], U[2] = u[2,2]y[2] + u[2,3]y[3] + ... + u[2,n]y[n],..., U[n] = u[n,n]y[n] we have <y,Ay> = U[1]^2 + U[2]^2 + ... + U[n]^2. We can change the variables of integration from y[1], y[2], ..., y[n] to U[1], U[2],...,U[n], to get:
\int_{R^n} \exp{\left[-\frac{1}{2}&lt;x,Ax&gt; + &lt;b,x&gt;\right]} \, dx_1 \cdots dx_n <br /> = C\, \exp{\left(-\frac{1}{2}&lt;c,Ac&gt;\right)} \int_{R^n} \exp{\left[-\frac{1}{2} (U_1^2 + \cdots + U_n^2)\right]} \, dU_1 \cdots dU_n,
where C is the Jacobian you get by changing variables from y to U.

RGV
 
Last edited:
Please explain some points:

1)Using X=Y+C,Q will Change this way:
<X,AX>-2<B,X>=<Y+C,A(Y+C)>-2<B,Y+C>
=<Y,AY>+<Y,AC>+<C,AY>+<C,AC>-2<B,Y>-2<B,C>
I noticed that with the last sentence the final form for exponent -0.5Qwill produce
<C,AC>/2 instead of -<C,AC>/2.

2)It seems to me that :
Q= <y,Ay> + <y,Ac> + <c,Ay> + <c,Ac> - 2<b,y> -2<b,c>
= <y,Ay> + <c,Ac> +<c,Ay> + <y,Ac> - 2<b,y> -2<b,c>
= <y,Ay> + <c,Ac> +<y,2Ac-2b> -2<b,c> - <y,Ac>
By taking Ac=b we obtain Q= <y,Ay> + <c,Ac> - <y,Ac> -2<b,c>

Please explain how you wrote Q = <y,Ay> + <c,Ac>

FBR
 
Fb.Researcher said:
Please explain some points:

1)Using X=Y+C,Q will Change this way:
<X,AX>-2<B,X>=<Y+C,A(Y+C)>-2<B,Y+C>
=<Y,AY>+<Y,AC>+<C,AY>+<C,AC>-2<B,Y>-2<B,C>
I noticed that with the last sentence the final form for exponent -0.5Qwill produce
<C,AC>/2 instead of -<C,AC>/2.

2)It seems to me that :
Q= <y,Ay> + <y,Ac> + <c,Ay> + <c,Ac> - 2<b,y> -2<b,c>
= <y,Ay> + <c,Ac> +<c,Ay> + <y,Ac> - 2<b,y> -2<b,c>
= <y,Ay> + <c,Ac> +<y,2Ac-2b> -2<b,c> - <y,Ac>
By taking Ac=b we obtain Q= <y,Ay> + <c,Ac> - <y,Ac> -2<b,c>

Please explain how you wrote Q = <y,Ay> + <c,Ac>

FBR

You are supposed to know that for real symmetric A we have &lt;c,Ay&gt; = &lt;Ac,y&gt; = &lt;y,Ac&gt; = &lt;Ay,c&gt;. However, you are right in noting that I incorrectly dropped the term 2<b,c>, so we should have Q = &lt;y,Ay&gt; + &lt;c,Ac&gt; +2&lt;y,Ac-b&gt; - 2&lt;b,c&gt; = &lt;y,Ay&gt; + &lt;c,Ac&gt; - 2&lt;b,c&gt; if we take Ac=b. In fact, this gives &lt;c,Ac&gt; - 2&lt;b,c&gt; = -&lt;b,c&gt;, so
\int_{R^n} \exp(-Q/2) \, dx_1 \cdots dx_n = C\, \exp(-&lt;b,c&gt;) \int_{R^n} \exp(-(U_1^2 + \cdots + U_n^2)/2) \, dU_1 \cdots dU_n.

RGV
 
Last edited:
Thread 'Use greedy vertex coloring algorithm to prove the upper bound of χ'
Hi! I am struggling with the exercise I mentioned under "Homework statement". The exercise is about a specific "greedy vertex coloring algorithm". One definition (which matches what my book uses) can be found here: https://people.cs.uchicago.edu/~laci/HANDOUTS/greedycoloring.pdf Here is also a screenshot of the relevant parts of the linked PDF, i.e. the def. of the algorithm: Sadly I don't have much to show as far as a solution attempt goes, as I am stuck on how to proceed. I thought...
Back
Top