inNdimensions
C=
{
x
∣
∣
∣x= (x 1 ,...,xN), |xj|≤
1
2
, j= 1,...,N
}
(B.1)
The radius of inertia ofCis
re^2 =E
(
x^2
)
=
∑
E
(
x^2 j
)
=N
∫ 1 / 2
− 1 / 2
x^2 dx=
N
12
(B.2)
Let us now considerr(θ), defined as the maximalrthat is inside the cube for a
given directionθ.
Choose a random directionθ= (θ 1 ,...,θN) by pickingθjto be normal iid with
θj∼N
[
0 ,
1
N
]
, xj=rθj (B.3)
WhenN is large there is a “phase transition” in the sense that
Prob
(
rθ∈C
)
≈
{
1 r < rC
0 r > rC
, rC=
1
2
√
N
logN
(B.4)
To see this we first observe that the probability thatx∼N(0,σ^2 ) takes values
outside the interval [−x 0 ,x 0 ] is given by the complementary error function
Prob(|x|> x 0 ) =
√
2
π
∫∞
x 0 /σ
e−x
(^2) / 2
dx= erfc(x 0 /σ)
Anticipating the result, Eq. (B.4), let us replacerby its re-scaled versionk:
r=krC (B.5)
For the case at hand
σ^2 = 1/N, x 0 =
1
2 r
=
1
2 krC
The probability forr|θj|=|xj|≤ 1 /2 happening for all coordinates simultaneously
is
Prob(rθ∈C) =
(
1 −erfc
(√
logN
k
))N
Since erfc(x)∈[0,1] forx∈[0,∞), the limitN→∞tends to
Prob(r θ∈C)→
0 ifNerfc
(√
logN
k
)
→∞
1 ifNerfc
(√
logN
k
)
→ 0
(B.6)
Since
Nerfc
(√
logN
k
)
→
{
∞ fork > 1
0 fork≤ 1
(B.7)
Eq. (B.4) follows.