Md. Asif Uddin
    Problem I.14.B02

    Where repeated averaging converges, and how fast

    symbolic▲▲▲

    Exact values where possible; six decimal places otherwise.

    STATEMENT

    Find the eigenvalues and the limit of the graph convolution’s averaging on the paw graph exactly, using the graph’s symmetry, and bound the distance to the limit after kk layers.

    GIVEN

    The paw graph of I.14.B01, with degrees d~=(3,3,4,2)\tilde{\vec d}=(3,3,4,2) counted with self-loops, A^=D~−1/2(A+I)D~−1/2\hat{\mat A}=\tilde{\mat D}^{-1/2}(\mat A+\mat I)\tilde{\mat D}^{-1/2} and x=(1,2,3,4)\vec x=(1,2,3,4).

    FIND

    (a) An eigenvector with eigenvalue 11, and one with eigenvalue 00. (b) The other two eigenvalues, exactly. (c) The limit of A^kx\hat{\mat A}^k\vec x. (d) A bound on the distance after kk layers, and its value at k=8k=8.

    STRATEGY

    Guess two eigenvectors from the structure: square roots of degrees, and the difference of the two nodes that swap under the graph’s symmetry. Get the other two from the trace of A^\hat{\mat A} and of A^2\hat{\mat A}^2.

    SOLUTION

    (a) Let s=(3,3,2,2)\vec s=(\sqrt3,\sqrt3,2,\sqrt2), the square roots of the degrees. Then

    A^s=D~−1/2(A+I)1=D~−1/2d~=s,\begin{aligned} \hat{\mat A}\vec s &=\tilde{\mat D}^{-1/2}(\mat A+\mat I)\mathbf 1\\ &=\tilde{\mat D}^{-1/2}\tilde{\vec d}=\vec s, \end{aligned}

    since each row of A+I\mat A+\mat I sums to the degree with self-loop. So s\vec s has eigenvalue 11. Nodes 11 and 22 have identical rows of A^\hat{\mat A}, so A^(1,−1,0,0)T=0\hat{\mat A}(1,-1,0,0)^{\mathsf T}=\vec 0: eigenvalue 00.

    (b) The trace of A^\hat{\mat A} is the sum of its diagonal, ∑v1/d~v=13+13+14+12=1712\sum_v1/\tilde d_v=\tfrac13+\tfrac13+\tfrac14+\tfrac12=\tfrac{17}{12}. The other two eigenvalues therefore add to 1712−1−0=512\tfrac{17}{12}-1-0=\tfrac5{12}. The trace of A^2\hat{\mat A}^2 is the sum of squared entries. The diagonal gives 19+19+116+14=77144\tfrac19+\tfrac19+\tfrac1{16}+\tfrac14=\tfrac{77}{144}, and the four edges, each appearing twice, give 2(19+112+112+18)=1161442\big(\tfrac19+\tfrac1{12}+\tfrac1{12}+\tfrac18\big)=\tfrac{116}{144}, for a total of 193144\tfrac{193}{144}. So the squares of the two add to 193144−1=49144\tfrac{193}{144}-1=\tfrac{49}{144}, and their product is 12((512)2−49144)=−112\tfrac12\big((\tfrac5{12})^2-\tfrac{49}{144}\big)=-\tfrac1{12}. They solve 12λ2−5λ−1=012\lambda^2-5\lambda-1=0:

    λ=5±7324=0.564333,  −0.147667.\begin{aligned} \lambda&=\frac{5\pm\sqrt{73}}{24}\\ &=0.564333,\;-0.147667. \end{aligned}

    (c) All other eigenvalues have absolute value below 11, so only the component along u1=s/∥s∥\vec u_1=\vec s/\|\vec s\| survives, and ∥s∥2=12\|\vec s\|^2=12. With sTx=33+6+42=16.853006\vec s^{\mathsf T}\vec x=3\sqrt3+6+4\sqrt2=16.853006,

    u1u1Tx=16.85300612 s  =1.404417 s  =(2.432522,  2.432522,  2.808834,  1.986146).\begin{aligned} &\vec u_1\vec u_1^{\mathsf T}\vec x=\frac{16.853006}{12}\,\vec s\\ &\;=1.404417\,\vec s\\ &\;=(2.432522,\;2.432522,\\ &\;\qquad2.808834,\;1.986146). \end{aligned}

    (d) Write x\vec x minus its limit in the remaining orthonormal eigenvectors. Each component is multiplied by its eigenvalue at every layer, and none exceeds 0.5643330.564333 in absolute value, so

    ∥A^kx−u1u1Tx∥    ≤0.564333k⋅2.516217,\begin{aligned} &\big\|\hat{\mat A}^k\vec x-\vec u_1\vec u_1^{\mathsf T}\vec x\big\|\\ &\;\;\leq0.564333^k\cdot2.516217, \end{aligned}

    where 2.5162172.516217 is the starting distance. At k=8k=8 the bound is 0.0258840.025884 and the actual distance 0.0247570.024757: almost all of the remainder lies along the second eigenvector.

    Answer

    Eigenvalues 11, (5+73)/24=0.564333(5+\sqrt{73})/24=0.564333, 00 and (5−73)/24=−0.147667(5-\sqrt{73})/24=-0.147667; limit 1.404417 s=(2.432522, 2.432522,1.404417\,\vec s=(2.432522,\,2.432522, 2.808834, 1.986146)2.808834,\,1.986146); distance at most 0.564333k⋅2.5162170.564333^k\cdot2.516217, which is 0.0258840.025884 at k=8k=8, against an actual 0.0247570.024757.

    Check — sanity

    The four eigenvalues add to 1712\tfrac{17}{12}, the trace. Divided by d~v\sqrt{\tilde d_v}, every entry of the limit is 1.4044171.404417: the nodes are indistinguishable except through their degrees.

    Where this breaks

    Without self-loops the graph’s normalised adjacency would have eigenvalue −1-1 if the graph were bipartite; the paw is not, because it contains a triangle. Learned weights between layers change the rate, and the bound then needs their singular values as well.

    Variation

    The zero eigenvalue means one layer removes the difference between nodes 11 and 22 completely, which is why I.14.B01 found them equal. Any graph with two nodes sharing a closed neighbourhood and a degree has such an eigenvector.

    Draws on