Professional Documents
Culture Documents
We obtain general, exact formulas for the overlaps between the eigenvectors of large correlated
random matrices, with additive or multiplicative noise. These results have potential applications in
many different contexts, from quantum thermalisation to high dimensional statistics. We apply our
results to the case of empirical correlation matrices, that allow us to estimate reliably the width of
the spectrum of the true underlying correlation matrix, even when the latter is very close to the
identity matrix. We illustrate our results on the example of stock returns correlations, that clearly
reveal a non trivial structure for the bulk eigenvalues.
(1)
+
1
Tr (z S)1 (z 0 S0 )1
,
N
(2)
where z, z 0 C and hiP denotes the average with respect to probability measure associated to S and S0 . Indeed, using the spectral decomposition, we obtain the
following inversion formula, valid when N for any
, 0 supp %, where % is the spectral density of S (that
we assume here to be the same as that of S0 , but see
2
Appendix for more general cases):
(, 0 ) =
(z 0 )g(z) (z)g(z 0 )
,
z 0 (z 0 ) z(z)
(4)
g(z) g(z 0 )
,
(z 0 ) (z)
(5)
2q
(, 0 )
,
0 (, 0 )
(6)
a (, 0 )
2 2 ( 0 )
R () R (0 )
,
a (, 0 )
(8)
2
R ()
,
2
2 () 0 ()2
(10)
u =
d%C () O(, )(, )v ,
(11)
N
where %C is the spectral density of C. Using the orthonormality of the vs, one then finds:
Z
p
1
u u00 =
d%C () O(, )O(, 0 ) (, )(, 0 ).
N
(12)
If we square this last expression and average over the
noise, and make an ergodic hypothesis [3] according to
which all signs (, ) are in fact independent from one
another, one finds the following, rather intuitive convolution result for square overlaps:
Z
0
(, ) = d%C ()O(, ) O(, 0 ).
(13)
3
(0, ), that corresponds to 1/q for Wishart matrices
(see [22] for details). In this case, the function m(z) can
be explicitly computed. This finally leads to:
(, ) =
( + 2q)2
2q 2( + ) 2 + (2q 1)
(14)
with := 1 + q and is within the interval h[ , + ], where the edges i are given by =
p
1 + (2 + 1)(2q + 1) . An interesting limit
corresponds to , where C tends to the identity
matrix, and the overlaps are expected to become all equal
to 1/N . Indeed one finds, for a fixed q:
#
"
( 1)(0 1) 1
2
0
+ O( ) ,
(, )
1+
2q 2
(15)
which is in fact universal in this limit, provided the eigenvalue spectrum of C has a variance given by (2)1 0.
[29] This formula is interesting insofar as it allows one to
estimate the width of the eigenvalue distribution of C,
even when it is close to the identity matrix, i.e. 1.
One could think of directly using information on the empirical spectrum, for example the Marcenko-Pastur prediction Tr C1 = (1 q) Tr S1 , that in principle allows one extract the parameter through 1 + (2)1 =
(1 q) Tr S1 /N . However, this method is numerically
unstable and very imprecise when 1 and finite N s
(for one thing, the RHS can be negative, leading to a
negative variance). Our formula based on overlaps avoid
these difficulties. As an illustration, we check the validity
of Eq. (14) in the inset of Figure 1 in the case = 10,
N = 500 and q = 0.5. We determine the empirical average overlap as follows: we consider 100 independent
realisation of the Wishart noise W. For each pair of
samples we compute a smoothed overlap as:
N
1 X (ui u0 j )2
[(ui u0i )2 ] =
,
Zi j=1 (i 0j )2 + 2
(16)
PN
with Zi = k=1 ((i 0k )2 + 2 )1 the normalization
constant and the width of the Cauchy kernel, that we
choose to be N 1/2 in such a way that N 1 1.
We then average this quantity over all pairs for a given
value of i to obtain [(ui u0i )2 ]e , and plot the resulting
quantity as a function of the average eigenvalue position
[i ]e , see Figure 1, inset. The agreement with Eq. (14)
is excellent, even when the true underlying matrix C is
close to the identity matrix
We now investigate an application to real data, in the
case of stock markets and using a bootstrap technique
to generate different samples. Specifically, we consider
the standardized daily returns of the 450 most liquid US
stocks from 2005 to 2012. We randomly split the total
period of 1800 days into two non-overlapping subsets of
4
a random matrix chosen in the Orthogonal group O(N )
according to the Haar measure (see e.g. [20, 22, 27]).
The case above corresponds to the case where OBO is
a Wishart matrix. We find for this general model that
(4) still holds with (z) replaced by SB (zg(z) 1) where
SB is the so-called Voiculescus S-transform of the B matrix [28]. If B = W, then SB (z) = 1/(1 + qz). Similarly,
the additive model can be generalized to S = C + OBO
with the same definitions for B and O. In that case, the
above result (5) again holds with (z) = z RB (g(z)),
where RB (z) is now the R-transform of the B matrix [28]
which is simply equal to RB (z) = 2 z when B = W
is a Gaussian random matrix, as considered above. Note
that in all these generalized cases, the overlap convolution formula (13) is always valid provided that the noises
are independent.
0
law) W =
W0 + 1 W1 and W0 = W0 + 1 W2 ,
where W1 , W2 are now independent, as above. Since our
formulas do not rely on the common matrix C that can
on overlapping
periods,
S = C [W 0 + W] C and S0 = C W 0 + W 0 C.
This case turns out to be more subtle and is the subject
of ongoing investigations.
In summary, we have provided general, exact formulas for the overlaps between the eigenvectors of large
correlated random matrices, with additive or multiplicative noise. Remarkably, these results do not require the
knowledge of the underlying pure matrix and have a
broad range of applications in different contexts. We
showed that the cross-sample eigenvector overlaps provide unprecedented information about the structure of
the true eigenvalue spectrum, much beyond that contained in the empirical spectrum itself. For example, the
width of the bulk of the spectrum of the true underlying
correlation matrix can be reliably estimated, even when
the latter is very close to the identity matrix. We have
illustrated our results on the example of stock returns
correlations, that clearly reveal a non trivial structure
for the bulk eigenvalues.
Acknowledgments: We want to thank R. Allez, R.
Benichou, R. Chicheportiche, B. Collins, A. Rej, E. Serie
and D. Ullmo for very useful discussions.
5
046202 (2012).
[17] P. Bourgade, L. Erd
os, H.-T. Yau, and J. Yin, Communications on Pure and Applied Mathematics (2015).
[18] R. Allez and J.-P. Bouchaud, Random Matrices: Theory
and Applications 3, 1450010 (2014).
[19] A. Bloemendal, A. Knowles, H.-T. Yau, and J. Yin, Probability Theory and Related Fields pp. 194 (2015).
[20] R. Couillet and F. Benaych-Georges, arXiv preprint
arXiv:1510.03547 (2015).
[21] R. Monasson and D. Villamaina, EPL (Europhysics Letters) 112, 50001 (2015).
[22] J. Bun, R. Allez, J.-P. Bouchaud, and M. Potters, arXiv
preprint arXiv:1502.06736 (2015).
[23] Philippe Biane, Quantum Probability Communications
11, 55 (2003).
[24] R. Allez, J. Bun, and J.-P. Bouchaud, arXiv preprint
arXiv:1412.7108 (2014).
[25] Z. Burda, A. G
orlich, A. Jarosz, and J. Jurkiewicz, Physica A: Statistical Mechanics and its Applications 343, 295
(2004).
[26] A. Knowles and J. Yin, arXiv preprint arXiv:1410.3516
(2014).
[27] Z. Burda, J. Jurkiewicz, and B. Waclaw, Physical Review
E 71, 026111 (2005).
[28] D. Voiculescu, Inventiones mathematicae 104, 201 (1991).
[29] The analysis for the additive case leads to a very similar
result. More precisely, taking C = IN + W0 with W0 a
GOE (independent from W and W0 ) of variance 02
0, one finds that a (, 0 ) is given by exactly the same
formula (15) with the substitution 2q 2 4 /02 .
[30] The calibration of is performed by least squares using
Eq. (14) and where we removed the 10 largest eigenvalues.
The derivation of the inversion formula (3) is pretty straightforward. We start from Eq. (2) that we rewrite using
e as
the spectral decomposition of S and S
*
(z, ze) =
N
1 X
1
1
e )2
(u u
ej i j
N i,j=1 z i ze
+
,
(17)
e
%() %e()
e
e
(, )dd
,
e
z ze
(18)
e i)
(x + i) (y
e
e
e
%()e
%()(,
)dd
2
2
e 2 + 2
(x ) + (y )
Z Z
e 2 + i(y
e (x ))
(x )(y )
e
e
e
%()e
%()(,
)dd
,
e 2 + 2 )
((x )2 + 2 )((y )
(x i, y i)
=
(19)
Z Z
e
%()
%e()
e
e
(, )dd
.
2
2
e 2 + 2
(x ) + (y )
(20)
6
Finally, the inversion formula follows from Sokhotski-Plemelj identity
lim Re (x i, y + i) (x i, y i) 2 2 %(x)e
%(y)(x, y).
(21)
Derivation of (4). We now compute the asymptotic expression of the function for sample covariance matrices.
Note
CW C
that we shall
omit
the
identity
matrix
I
in
our
notations
throughout
the
following.
We
recall
that
S
=
N
f C where W and W
f are two independent Wishart matrices. Here, we allow these two matrices to be
and e
S = CW
parametrized by a possibly different dimensional ratio q and qe. By independence, we have
(z, ze) =
1 X
e 1 ,
(z S)1
z S)
kl P (e
kl P
N
(22)
k,l
then, we use the deterministic estimate like that of [22, 25, 26] to find for any z = x i at global scale that
(z S)1
kl
1
(z) z(z) C
+ O(N 1/2 ),
(23)
kl
1
e z )(e
e z ) C)1 ],
Tr[(z)(z(z) C)1 (e
z (e
N
(24)
1
e z ) C)1 ],
[(z(z) C)1 (e
z (e
e
ze(e
z ) z(z)
(25)
to obtain
(z, ze)
e z) 1
(z) (e
e z ) C)1 ].
Tr[(z(z) C)1 (e
z (e
e
ze(e
z ) z(z) N
(26)
h
i
h
i
e z ) 1 Tr (z)(z(z) C)1 (z) 1 Tr (e
e z (e
e z ) C)1
(z, ze)
(e
e z ) z(z)
N
N
ze(e
!
(27)
(z)
Tr(z(z) C)1
N
and ge(e
z)
e z)
(e
e z ) C)1 .
Tr(e
z (e
N
(28)
e z )g(z) (z)e
(e
g (e
z)
,
e z ) z(z)
ze(e
(29)
7
Derivation of (6). Using the definition of the complex function m and m
e above, we find that (z) = 1/(zm(z)) and
e z ) = 1/(e
(e
z m(e
e z )). Note that we shall omit the arguments z and ze in m and m
e in the following when there are no
confusion. First, we rewrite (4) as
g
ge
1
,
(30)
(z, ze)
1/m
e 1/m zem
e
zm
which is equivalent to
(z, ze)
1
1
[zgm zegem]
e .
ze
zmm
e
.
qe
q ze
z
mm
e
mm
e
qe
z
qze
z
(31)
(32)
m
e
m
+
m
e
m
e
m
e
0
0
0
0 0
0
0
0
qe q 0 0
qe q
+ Im, (33)
+
e
e
qe
q
qe
q m0 m
m0 m
e 0 m0 m
e0
e 0 m0 m
e0
s where we omit the explicit expressions of the imaginary part since this is useless (see Eq. (3)). Then, using the
representation m0 = mR + imI and m
e0 = m
e R + im
e I , one finds
h
h
i
i
m0 m
e 20 m
e0 m
e0 m
e 0 = 2m
e I 2mI m
e R + i(m
e 2R + m
e 2R 2mR m
e R) ,
e 20 + m
e 0m
h
i
m0 m
e0 m
e 0 = 2m
e I mI imR ,
(34)
and
e0
m0 m
m0 m
e 0 = (mR m
e R )2 (m2I m
e 2I ) + 2imI (mR m
e R ).
(35)
(36)
which is exactly the denominator in (6). For the numerator, elementary complex analysis in Eq. (33) yields
h
i
h
i
h
i
e 20 m
e 20 + m
e 0m
e0 m
e0 m
e 0 m0 m
e 0 m0 m
e 0 = 4mI m
e I mR |m
e 0 |2 m
e R |m0 |2 ,
m0 m
(37)
h
i
h
i
m0 m
e0 m
e 0 m0 m
e 0 m0 m
e 0 = 2mI m
e I |m
e 0 |2 |m0 |2 .
(38)
and
e =
By regrouping these last three equations with the prefactors in (33), and recalling that mI () = q%() and m
e I ()
e
e
q %e(), so we obtain by using the inversion formula (3) the general result:
e =
q,eq (, )
e m R |m
2(e
q q )
e 0 |2 m
e R |m0 |2 + (e
q q) |m
e 0 |2 |m0 |2
h
ih
i.
e (mR m
e R )2 + (mI + m
e I )2 (mR m
e R )2 + (mI m
e I )2
(39)
8
e , which is
One easily retrieve Eq. (6) by taking qe = q. Another interesting case if when qe = 0 as we expect
precisely the framework of [15]. In that case, we have m
e R = 1/ and m
e I = 0. Hence, we deduce from (39) that
q
(mR 1/)2 + m2I
q
,
=
|1 m0 ()|2
q,eq=0 (, ) =
(40)
(41)
(42)
and
(, + ) =
i
+ O(3 )
(43)
The result (9) follows by plugging Eqs. (42) and (43) into Eq. (6) and then set = 0.
The derivation of the overlaps (8) for two independent deformed GOEs is very similar to sample covariance matrices.
Hence, we shall omit most details that can be obtained by following the arguments of the above Appendix. Again,
e where there is no confusion.
we shall skip the arguments and
For completeness, we recall some notations. We defined g and ge are respectively the Stieltjes transform of S = C+W
f where the noises are independent. Also, we introduced
and e
S=C+W
(z) = z 2 g(z),
e z ) = ze
(e
e2 g(e
z ),
(44)
(z S)1
kl
1
(z) C
(45)
kl
e
and (e
ze
S)1
kl P is obtained from Eq. (45) by replacing by . Then, plugging this into (2), we obtain,
e + i) ( i,
e i) = g0 ge0 g0 ge0
lim ( i,
0
e
e
g0 e e + ge0 ge0 + g0 e ge0 e
=
e e
(46)
9
We proceed as above (see Eq. (33) and thereafter) to find
g0 e e + ge0 ge0 + g0 e ge0 e = 2 (I geI gI eI ) + i eI (gR geR ) geI (R eR )
(47)
and
e = (R eR )2 + (eI2 I2 ) + 2iI (eR R ).
(48)
Hence, by putting these last two equations into (46) and then using (3), we get after some straightforward computations
2
e2 )(R eR )2 + 2 2
e2 (gR geR )(R eR ) ( 2
e2 )(I2 eI2 )
e = ( +
,
a (, )
(R eR )2 + (I + eI )2 (R eR )2 + (I eI )2
(49)
which is the generalization of Eq. (8) for two different noise parameters 6=
e. If we now consider =
e, we have
(R eR )2 + 2 (gR geR )(R eR )
e = 2 2
a (, )
,
e 2 + (I () + I ())
e 2 (R () R ())
e 2 + (I () I ())
e 2
(R () R ())
e R () ())
e
2 2 ( )(
=
,
e 2 + (I () I ())
e 2
e 2 + (I () + I ())
e 2 (R () R ())
(R () R ())
(50)
e = + to find
that is exactly Eq. (8). As for sample covariance matrices, one can again set
a (, ) =
2I2
2 R ()
,
[ R ()]2 + [ I ()]2
(51)
2
,
(R )2 + I2
(52)
We define
S = C + W1 ,
e
S = C + W2 ,
(53)
(54)
where we denoted by N the Gaussian measure and used the abbreviation 12 = 1 2 . Using the stability of GOE
under addition, let us rewrite the noise terms as
W1 = A + B1 ,
W2 = A + B2 ,
(55)
hA2 iN = 12 ,
hB1 iN = 0, hB2 iN = 0,
hB21 iN = 12 12 , hB22 iN = 22 12 ,
hB1 B2 iN = 0.
(56)
(57)
10
One can check that this parametrization yields exactly the correlation structure of Eq. (54). Therefore, using (55)
into (53), we have the equivalence (in law)
S1 = D + B1 ,
e = D + B2 ,
S
D ..= C + A.
(58)
Since the noises are now independent and that the mean squared overlap a , given in Eq. (49), is independent
from the exact structure of C, we can therefore replace C by D. Hence, we deduce that the overlaps for this model
will again be given by Eq. (49) with 2 = 12 12 , and
e2 = 22 12 . If =
e, then, a is given by Eq. (8) with
2
variance (1 ), as announced.