Subject: Statistics

Paper: Multivariate Analysis

Module: Canonical Correlation

1 / 15
Development Team

Principal investigator: Dr. Bhaswati Ganguli, Professor,

Department of Statistics, University of Calcutta
Paper co-ordinator: Dr. Sugata SenRoy, Professor, Department of
Statistics, University of Calcutta
Content writer: Souvik Bandyopadhyay, Senior Lecturer, Indian
Institute of Public Health, Hyderabad
Content reviewer: Dr. Kalyan Das, Professor, Department of
Statistics, University of Calcutta

2 / 15
Development Team

Principal investigator: Dr. Bhaswati Ganguli, Professor,

Department of Statistics, University of Calcutta
Paper co-ordinator: Dr. Sugata SenRoy, Professor, Department of
Statistics, University of Calcutta
Content writer: Souvik Bandyopadhyay, Senior Lecturer, Indian
Institute of Public Health, Hyderabad
Content reviewer: Dr. Kalyan Das, Professor, Department of
Statistics, University of Calcutta

2 / 15
Development Team

Principal investigator: Dr. Bhaswati Ganguli, Professor,

Department of Statistics, University of Calcutta
Paper co-ordinator: Dr. Sugata SenRoy, Professor, Department of
Statistics, University of Calcutta
Content writer: Souvik Bandyopadhyay, Senior Lecturer, Indian
Institute of Public Health, Hyderabad
Content reviewer: Dr. Kalyan Das, Professor, Department of
Statistics, University of Calcutta

2 / 15
Development Team

Principal investigator: Dr. Bhaswati Ganguli, Professor,

Department of Statistics, University of Calcutta
Paper co-ordinator: Dr. Sugata SenRoy, Professor, Department of
Statistics, University of Calcutta
Content writer: Souvik Bandyopadhyay, Senior Lecturer, Indian
Institute of Public Health, Hyderabad
Content reviewer: Dr. Kalyan Das, Professor, Department of
Statistics, University of Calcutta

2 / 15
Correlations

I Q : What is canonical correlation ?

I It’s a technique which seeks to identify and quantify the
relationship between two sets of variables.
I Special cases
I Simple correlation
I Multiple Correlation
I In simple correlation, we study the relationship between two
variables X and Y .
I In multiple correlation, we study the relationship between a
variable Y and a set of variables (X1 , . . . , Xp ).
← To do this we take a linear combination of X1 , . . . , Xp
and find its correlation with Y .

3 / 15
Correlations

I Q : What is canonical correlation ?

I It’s a technique which seeks to identify and quantify the
relationship between two sets of variables.
I Special cases
I Simple correlation
I Multiple Correlation
I In simple correlation, we study the relationship between two
variables X and Y .
I In multiple correlation, we study the relationship between a
variable Y and a set of variables (X1 , . . . , Xp ).
← To do this we take a linear combination of X1 , . . . , Xp
and find its correlation with Y .

3 / 15
Correlations

I Q : What is canonical correlation ?

I It’s a technique which seeks to identify and quantify the
relationship between two sets of variables.
I Special cases
I Simple correlation
I Multiple Correlation
I In simple correlation, we study the relationship between two
variables X and Y .
I In multiple correlation, we study the relationship between a
variable Y and a set of variables (X1 , . . . , Xp ).
← To do this we take a linear combination of X1 , . . . , Xp
and find its correlation with Y .

3 / 15
Correlations

I Q : What is canonical correlation ?

I It’s a technique which seeks to identify and quantify the
relationship between two sets of variables.
I Special cases
I Simple correlation
I Multiple Correlation
I In simple correlation, we study the relationship between two
variables X and Y .
I In multiple correlation, we study the relationship between a
variable Y and a set of variables (X1 , . . . , Xp ).
← To do this we take a linear combination of X1 , . . . , Xp
and find its correlation with Y .

3 / 15
Correlations

I Q : What is canonical correlation ?

I It’s a technique which seeks to identify and quantify the
relationship between two sets of variables.
I Special cases
I Simple correlation
I Multiple Correlation
I In simple correlation, we study the relationship between two
variables X and Y .
I In multiple correlation, we study the relationship between a
variable Y and a set of variables (X1 , . . . , Xp ).
← To do this we take a linear combination of X1 , . . . , Xp
and find its correlation with Y .

3 / 15
Correlations

I Q : What is canonical correlation ?

I It’s a technique which seeks to identify and quantify the
relationship between two sets of variables.
I Special cases
I Simple correlation
I Multiple Correlation
I In simple correlation, we study the relationship between two
variables X and Y .
I In multiple correlation, we study the relationship between a
variable Y and a set of variables (X1 , . . . , Xp ).
← To do this we take a linear combination of X1 , . . . , Xp
and find its correlation with Y .

3 / 15
Correlations

I Q : What is canonical correlation ?

I It’s a technique which seeks to identify and quantify the
relationship between two sets of variables.
I Special cases
I Simple correlation
I Multiple Correlation
I In simple correlation, we study the relationship between two
variables X and Y .
I In multiple correlation, we study the relationship between a
variable Y and a set of variables (X1 , . . . , Xp ).
← To do this we take a linear combination of X1 , . . . , Xp
and find its correlation with Y .

3 / 15
Canonical Correlation

I In canonical correlation we study the relationship between two

sets of variables (X1 , . . . , Xr ) and (X1∗ , . . . , Xs∗ ).
I Canonical correlation requires that each set of variables be
reduced to a single variable and then finding their correlation.
I Usually these two variables are found by taking linear
combinations of the variables in each set, under certain
pre-fixed criteria.
I The variables obtained by these linear combinations are known
as canonical variables and the correlation between them as
canonical correlation.

4 / 15
Canonical Correlation

I In canonical correlation we study the relationship between two

sets of variables (X1 , . . . , Xr ) and (X1∗ , . . . , Xs∗ ).
I Canonical correlation requires that each set of variables be
reduced to a single variable and then finding their correlation.
I Usually these two variables are found by taking linear
combinations of the variables in each set, under certain
pre-fixed criteria.
I The variables obtained by these linear combinations are known
as canonical variables and the correlation between them as
canonical correlation.

4 / 15
Canonical Correlation

I In canonical correlation we study the relationship between two

sets of variables (X1 , . . . , Xr ) and (X1∗ , . . . , Xs∗ ).
I Canonical correlation requires that each set of variables be
reduced to a single variable and then finding their correlation.
I Usually these two variables are found by taking linear
combinations of the variables in each set, under certain
pre-fixed criteria.
I The variables obtained by these linear combinations are known
as canonical variables and the correlation between them as
canonical correlation.

4 / 15
Canonical Correlation

I In canonical correlation we study the relationship between two

sets of variables (X1 , . . . , Xr ) and (X1∗ , . . . , Xs∗ ).
I Canonical correlation requires that each set of variables be
reduced to a single variable and then finding their correlation.
I Usually these two variables are found by taking linear
combinations of the variables in each set, under certain
pre-fixed criteria.
I The variables obtained by these linear combinations are known
as canonical variables and the correlation between them as
canonical correlation.

4 / 15
The Model

I Let there be r variables in the 1st group X1 = (X1 , . . . , Xr )

and s variables in the 2nd group X2 = (Xr+1 , . . . , Xr+s ).
I Assume, without loss of generality, that r ≤ s, i.e. take the
smaller group of variables as the 1st group.
I Also let
E(X1 ) = µ1 and E(X2 ) = µ2
V ar(X1 ) = Σ11 , V ar(X2 ) = Σ22
and Cov(X1 , X2 ) = Σ12 (or Σ021 )
     
X1 µ1 Σ11 Σ12
I Define X = ,µ= and Σ = .
X2 µ2 Σ21 Σ22

5 / 15
The Model

I Let there be r variables in the 1st group X1 = (X1 , . . . , Xr )

and s variables in the 2nd group X2 = (Xr+1 , . . . , Xr+s ).
I Assume, without loss of generality, that r ≤ s, i.e. take the
smaller group of variables as the 1st group.
I Also let
E(X1 ) = µ1 and E(X2 ) = µ2
V ar(X1 ) = Σ11 , V ar(X2 ) = Σ22
and Cov(X1 , X2 ) = Σ12 (or Σ021 )
     
X1 µ1 Σ11 Σ12
I Define X = ,µ= and Σ = .
X2 µ2 Σ21 Σ22

5 / 15
The Model

I Let there be r variables in the 1st group X1 = (X1 , . . . , Xr )

and s variables in the 2nd group X2 = (Xr+1 , . . . , Xr+s ).
I Assume, without loss of generality, that r ≤ s, i.e. take the
smaller group of variables as the 1st group.
I Also let
E(X1 ) = µ1 and E(X2 ) = µ2
V ar(X1 ) = Σ11 , V ar(X2 ) = Σ22
and Cov(X1 , X2 ) = Σ12 (or Σ021 )
     
X1 µ1 Σ11 Σ12
I Define X = ,µ= and Σ = .
X2 µ2 Σ21 Σ22

5 / 15
The Model

I Let there be r variables in the 1st group X1 = (X1 , . . . , Xr )

and s variables in the 2nd group X2 = (Xr+1 , . . . , Xr+s ).
I Assume, without loss of generality, that r ≤ s, i.e. take the
smaller group of variables as the 1st group.
I Also let
E(X1 ) = µ1 and E(X2 ) = µ2
V ar(X1 ) = Σ11 , V ar(X2 ) = Σ22
and Cov(X1 , X2 ) = Σ12 (or Σ021 )
     
X1 µ1 Σ11 Σ12
I Define X = ,µ= and Σ = .
X2 µ2 Σ21 Σ22

5 / 15
The Problem

I Σ12 contains rs elements which gives the correlations between

each variable of the 1st group with those of the 2nd group
← gives a perfect picture of the relationships
← however, not easily comprehendible.
I Instead, can we look at just a few values, which would make
understanding easier ?

6 / 15
The Problem

I Σ12 contains rs elements which gives the correlations between

each variable of the 1st group with those of the 2nd group
← gives a perfect picture of the relationships
← however, not easily comprehendible.
I Instead, can we look at just a few values, which would make
understanding easier ?

6 / 15
The Method

I Let a and b be respectively r and s dimensional vectors.

I Define the variables,

U = a0 X1 and V = b0 X2 .

Then E(U ) = a0 µ1 , E(V ) = b0 µ2 ,

V ar(U ) = a0 Σ11 a, V ar(V ) = b0 Σ22 b,
with Cov(U, V ) = a0 Σ12 b.

so that
a0 Σ12 b
Corr(U, V ) = √ .
a0 Σ11 a × b0 Σ22 b

7 / 15
The Method

I Let a and b be respectively r and s dimensional vectors.

I Define the variables,

U = a0 X1 and V = b0 X2 .

Then E(U ) = a0 µ1 , E(V ) = b0 µ2 ,

V ar(U ) = a0 Σ11 a, V ar(V ) = b0 Σ22 b,
with Cov(U, V ) = a0 Σ12 b.

so that
a0 Σ12 b
Corr(U, V ) = √ .
a0 Σ11 a × b0 Σ22 b

7 / 15
How to choose a and b

I The coefficients a1 , b1 of the first pair (U1 , V1 ) are so chosen

as to maximize Corr(U, V ).
I The 2nd pair (U2 , V2 ) are chosen to maximize Corr(U, V )
subject to their combining vectors a2 and b2 being orthogonal
to a1 and b1 , respectively.
I In general, the jth pair (Uj , Vj ) are chosen to maximize
Corr(U, V ) subject to aj and bj being orthogonal to ak and
bk , respectively, for all k < j.
I This can be done till j = r.
I Usually for all j = 1, . . . , r,

a0j aj = 1 and b0j bj = 1.

8 / 15
How to choose a and b

I The coefficients a1 , b1 of the first pair (U1 , V1 ) are so chosen

as to maximize Corr(U, V ).
I The 2nd pair (U2 , V2 ) are chosen to maximize Corr(U, V )
subject to their combining vectors a2 and b2 being orthogonal
to a1 and b1 , respectively.
I In general, the jth pair (Uj , Vj ) are chosen to maximize
Corr(U, V ) subject to aj and bj being orthogonal to ak and
bk , respectively, for all k < j.
I This can be done till j = r.
I Usually for all j = 1, . . . , r,

a0j aj = 1 and b0j bj = 1.

8 / 15
How to choose a and b

I The coefficients a1 , b1 of the first pair (U1 , V1 ) are so chosen

as to maximize Corr(U, V ).
I The 2nd pair (U2 , V2 ) are chosen to maximize Corr(U, V )
subject to their combining vectors a2 and b2 being orthogonal
to a1 and b1 , respectively.
I In general, the jth pair (Uj , Vj ) are chosen to maximize
Corr(U, V ) subject to aj and bj being orthogonal to ak and
bk , respectively, for all k < j.
I This can be done till j = r.
I Usually for all j = 1, . . . , r,

a0j aj = 1 and b0j bj = 1.

8 / 15
How to choose a and b

I The coefficients a1 , b1 of the first pair (U1 , V1 ) are so chosen

as to maximize Corr(U, V ).
I The 2nd pair (U2 , V2 ) are chosen to maximize Corr(U, V )
subject to their combining vectors a2 and b2 being orthogonal
to a1 and b1 , respectively.
I In general, the jth pair (Uj , Vj ) are chosen to maximize
Corr(U, V ) subject to aj and bj being orthogonal to ak and
bk , respectively, for all k < j.
I This can be done till j = r.
I Usually for all j = 1, . . . , r,

a0j aj = 1 and b0j bj = 1.

8 / 15
How to choose a and b

I The coefficients a1 , b1 of the first pair (U1 , V1 ) are so chosen

as to maximize Corr(U, V ).
I The 2nd pair (U2 , V2 ) are chosen to maximize Corr(U, V )
subject to their combining vectors a2 and b2 being orthogonal
to a1 and b1 , respectively.
I In general, the jth pair (Uj , Vj ) are chosen to maximize
Corr(U, V ) subject to aj and bj being orthogonal to ak and
bk , respectively, for all k < j.
I This can be done till j = r.
I Usually for all j = 1, . . . , r,

a0j aj = 1 and b0j bj = 1.

8 / 15
Result

I Let λ1 ≥ λ2 ≥ . . . ≥ λr be the eigenvalues of

−1/2 −1/2
Σ11 Σ12 Σ−122 Σ21 Σ11 and p1 , p2 , . . . , pr the corresponding
eigen-vectors.
I In fact, λ1 ≥ λ2 ≥ . . . ≥ λr are also the r largest (out of s)
−1/2 −1/2
eigen-values of Σ22 Σ21 Σ−1 11 Σ12 Σ22 with q1 , q2 , . . . , qr
the corresponding eigen-vectors.
Result
p
Corr(U1 , V1 ) = ρ1 = λ1 ,
−1/2 −1/2
with a1 = Σ11 p1 and b1 = Σ22 q1 .
−1/2 −1/2
I Thus U1 = p01 Σ11 X1 and V2 = q01 Σ22 X2 .

9 / 15
Result

I Let λ1 ≥ λ2 ≥ . . . ≥ λr be the eigenvalues of

−1/2 −1/2
Σ11 Σ12 Σ−122 Σ21 Σ11 and p1 , p2 , . . . , pr the corresponding
eigen-vectors.
I In fact, λ1 ≥ λ2 ≥ . . . ≥ λr are also the r largest (out of s)
−1/2 −1/2
eigen-values of Σ22 Σ21 Σ−1 11 Σ12 Σ22 with q1 , q2 , . . . , qr
the corresponding eigen-vectors.
Result
p
Corr(U1 , V1 ) = ρ1 = λ1 ,
−1/2 −1/2
with a1 = Σ11 p1 and b1 = Σ22 q1 .
−1/2 −1/2
I Thus U1 = p01 Σ11 X1 and V2 = q01 Σ22 X2 .

9 / 15
Result

I Let λ1 ≥ λ2 ≥ . . . ≥ λr be the eigenvalues of

−1/2 −1/2
Σ11 Σ12 Σ−122 Σ21 Σ11 and p1 , p2 , . . . , pr the corresponding
eigen-vectors.
I In fact, λ1 ≥ λ2 ≥ . . . ≥ λr are also the r largest (out of s)
−1/2 −1/2
eigen-values of Σ22 Σ21 Σ−1 11 Σ12 Σ22 with q1 , q2 , . . . , qr
the corresponding eigen-vectors.
Result
p
Corr(U1 , V1 ) = ρ1 = λ1 ,
−1/2 −1/2
with a1 = Σ11 p1 and b1 = Σ22 q1 .
−1/2 −1/2
I Thus U1 = p01 Σ11 X1 and V2 = q01 Σ22 X2 .

9 / 15
Result

I Let λ1 ≥ λ2 ≥ . . . ≥ λr be the eigenvalues of

−1/2 −1/2
Σ11 Σ12 Σ−122 Σ21 Σ11 and p1 , p2 , . . . , pr the corresponding
eigen-vectors.
I In fact, λ1 ≥ λ2 ≥ . . . ≥ λr are also the r largest (out of s)
−1/2 −1/2
eigen-values of Σ22 Σ21 Σ−1 11 Σ12 Σ22 with q1 , q2 , . . . , qr
the corresponding eigen-vectors.
Result
p
Corr(U1 , V1 ) = ρ1 = λ1 ,
−1/2 −1/2
with a1 = Σ11 p1 and b1 = Σ22 q1 .
−1/2 −1/2
I Thus U1 = p01 Σ11 X1 and V2 = q01 Σ22 X2 .

9 / 15
Result (contd.)

Result
In general, for j = 1, . . . , r,
p
Corr(Uj , Vj ) = ρj = λj ,
−1/2 −1/2
with aj = Σ11 pj and bj = Σ22 qj .

I Usually, only the first one or two of the canonical correlations

are looked at.

10 / 15
Result (contd.)

Result
In general, for j = 1, . . . , r,
p
Corr(Uj , Vj ) = ρj = λj ,
−1/2 −1/2
with aj = Σ11 pj and bj = Σ22 qj .

I Usually, only the first one or two of the canonical correlations

are looked at.

10 / 15
Proof

I Since Σ11 and Σ22 are p.d., define

1/2 1/2 −1/2 −1/2
c = Σ11 a, d = Σ22 b so that a = Σ11 c, b = Σ22 d.
−1/2 −1/2
c0 Σ11 Σ12 Σ22 d
Corr(U, V ) = q
−1/2 −1/2 −1/2 −1/2
c0 Σ11 Σ11 Σ11 c × d0 Σ22 Σ22 Σ22 d
−1/2 −1/2
c0 Σ11 Σ12 Σ22 d
i.e. Corr(U, V ) = √
c0 c × d0 d
By Cauchy-Schwarz inequality,
−1/2 −1/2 −1/2 −1/2
c0 Σ11 Σ12 Σ22 d ≤ (c0 Σ11 Σ12 Σ−1
22 Σ21 Σ11 c)1/2(d0 d)1/2

11 / 15
Proof (contd.)

−1/2 −1/2
I Σ11 Σ12 Σ−122 Σ21 Σ11 is symmetric, with largest eigen-value
λ1 and corresponding eigen-vector p1 . Thus
−1/2 −1/2
c0 Σ11 Σ12 Σ−1
22 Σ21 Σ11 c ≤ λ 1 c0 c = λ 1 .

I Equality occurs at c = p1 .
I Also equality in the Cauchy-Schwarz inequality occurs if d is
−1/2 −1/2
proportional to Σ22 Σ21 Σ11 p1 .
I This means b is proportional to
−1/2 −1/2 −1/2 −1/2
Σ22 Σ22 Σ21 Σ11 p1 = Σ22 q1 (say),
−1/2 −1/2
where, q1 = Σ22 Σ21 Σ11 p1 .

12 / 15
Proof (contd.)

−1/2 −1/2
I Σ11 Σ12 Σ−122 Σ21 Σ11 is symmetric, with largest eigen-value
λ1 and corresponding eigen-vector p1 . Thus
−1/2 −1/2
c0 Σ11 Σ12 Σ−1
22 Σ21 Σ11 c ≤ λ 1 c0 c = λ 1 .

I Equality occurs at c = p1 .
I Also equality in the Cauchy-Schwarz inequality occurs if d is
−1/2 −1/2
proportional to Σ22 Σ21 Σ11 p1 .
I This means b is proportional to
−1/2 −1/2 −1/2 −1/2
Σ22 Σ22 Σ21 Σ11 p1 = Σ22 q1 (say),
−1/2 −1/2
where, q1 = Σ22 Σ21 Σ11 p1 .

12 / 15
Proof (contd.)

−1/2 −1/2
I Σ11 Σ12 Σ−122 Σ21 Σ11 is symmetric, with largest eigen-value
λ1 and corresponding eigen-vector p1 . Thus
−1/2 −1/2
c0 Σ11 Σ12 Σ−1
22 Σ21 Σ11 c ≤ λ 1 c0 c = λ 1 .

I Equality occurs at c = p1 .
I Also equality in the Cauchy-Schwarz inequality occurs if d is
−1/2 −1/2
proportional to Σ22 Σ21 Σ11 p1 .
I This means b is proportional to
−1/2 −1/2 −1/2 −1/2
Σ22 Σ22 Σ21 Σ11 p1 = Σ22 q1 (say),
−1/2 −1/2
where, q1 = Σ22 Σ21 Σ11 p1 .

12 / 15
Proof (contd.)

−1/2 −1/2
I Σ11 Σ12 Σ−122 Σ21 Σ11 is symmetric, with largest eigen-value
λ1 and corresponding eigen-vector p1 . Thus
−1/2 −1/2
c0 Σ11 Σ12 Σ−1
22 Σ21 Σ11 c ≤ λ 1 c0 c = λ 1 .

I Equality occurs at c = p1 .
I Also equality in the Cauchy-Schwarz inequality occurs if d is
−1/2 −1/2
proportional to Σ22 Σ21 Σ11 p1 .
I This means b is proportional to
−1/2 −1/2 −1/2 −1/2
Σ22 Σ22 Σ21 Σ11 p1 = Σ22 q1 (say),
−1/2 −1/2
where, q1 = Σ22 Σ21 Σ11 p1 .

12 / 15
Proof (contd.)

I Notice that from properties of eigen-values,

−1/2 −1/2
Σ11 Σ12 Σ−1
22 Σ21 Σ11 p1 = λ 1 p1
−1/2 −1/2
so that premultiplying both sides by Σ22 Σ21 Σ11 yields for
−1/2 −1/2
q1 = Σ22 Σ21 Σ11 p1
−1/2 −1/2
Σ22 Σ21 Σ−1
11 Σ12 Σ22 q1 = λ 1 q1
−1/2 −1/2
i.e. λ1 is an eigen-value of Σ22 Σ21 Σ−1
11 Σ12 Σ22 with
corresponding eigen-vector q1 .
−1/2 −1/2
I Thus with a1 = Σ11 p1 and b1 = Σ22 q1 ,
p
Corr(U1 , V1 ) = maxa,b Corr(a0 X1 , b0 X2 ) = λ1 .

13 / 15
Proof (contd.)

I Notice that from properties of eigen-values,

−1/2 −1/2
Σ11 Σ12 Σ−1
22 Σ21 Σ11 p1 = λ 1 p1
−1/2 −1/2
so that premultiplying both sides by Σ22 Σ21 Σ11 yields for
−1/2 −1/2
q1 = Σ22 Σ21 Σ11 p1
−1/2 −1/2
Σ22 Σ21 Σ−1
11 Σ12 Σ22 q1 = λ 1 q1
−1/2 −1/2
i.e. λ1 is an eigen-value of Σ22 Σ21 Σ−1
11 Σ12 Σ22 with
corresponding eigen-vector q1 .
−1/2 −1/2
I Thus with a1 = Σ11 p1 and b1 = Σ22 q1 ,
p
Corr(U1 , V1 ) = maxa,b Corr(a0 X1 , b0 X2 ) = λ1 .

13 / 15
Properties

−1/2 −1/2
I E(Uj ) = p0j Σ11 µ1 and E(Vj ) = q0j Σ22 µ2 for all j.
−1/2 −1/2
I V ar(Uj ) = p0j Σ11 Σ11 Σ11 pj = p0j pj = 1 for all j.
−1/2 −1/2
I V ar(Vj ) = q0j Σ22 Σ22 Σ22 qj = q0j qj = 1 for all j.
−1/2 −1/2
I Cov(Uj , Uk ) = p0j Σ11 Σ11 Σ11 pk = p0j pk = 0 for all
k 6= j.
−1/2 −1/2
I Cov(Vj , Vk ) = q0j Σ22 Σ22 Σ22 qk = q0j qk = 0 for all k 6= j.
−1/2 −1/2
I Cov(Uj , Vk ) = p0j Σ11 Σ12 Σ22 qk = p0j pk = 0 for all
j 6= k.

14 / 15
Properties

−1/2 −1/2
I E(Uj ) = p0j Σ11 µ1 and E(Vj ) = q0j Σ22 µ2 for all j.
−1/2 −1/2
I V ar(Uj ) = p0j Σ11 Σ11 Σ11 pj = p0j pj = 1 for all j.
−1/2 −1/2
I V ar(Vj ) = q0j Σ22 Σ22 Σ22 qj = q0j qj = 1 for all j.
−1/2 −1/2
I Cov(Uj , Uk ) = p0j Σ11 Σ11 Σ11 pk = p0j pk = 0 for all
k 6= j.
−1/2 −1/2
I Cov(Vj , Vk ) = q0j Σ22 Σ22 Σ22 qk = q0j qk = 0 for all k 6= j.
−1/2 −1/2
I Cov(Uj , Vk ) = p0j Σ11 Σ12 Σ22 qk = p0j pk = 0 for all
j 6= k.

14 / 15
Properties

−1/2 −1/2
I E(Uj ) = p0j Σ11 µ1 and E(Vj ) = q0j Σ22 µ2 for all j.
−1/2 −1/2
I V ar(Uj ) = p0j Σ11 Σ11 Σ11 pj = p0j pj = 1 for all j.
−1/2 −1/2
I V ar(Vj ) = q0j Σ22 Σ22 Σ22 qj = q0j qj = 1 for all j.
−1/2 −1/2
I Cov(Uj , Uk ) = p0j Σ11 Σ11 Σ11 pk = p0j pk = 0 for all
k 6= j.
−1/2 −1/2
I Cov(Vj , Vk ) = q0j Σ22 Σ22 Σ22 qk = q0j qk = 0 for all k 6= j.
−1/2 −1/2
I Cov(Uj , Vk ) = p0j Σ11 Σ12 Σ22 qk = p0j pk = 0 for all
j 6= k.

14 / 15
Properties

−1/2 −1/2
I E(Uj ) = p0j Σ11 µ1 and E(Vj ) = q0j Σ22 µ2 for all j.
−1/2 −1/2
I V ar(Uj ) = p0j Σ11 Σ11 Σ11 pj = p0j pj = 1 for all j.
−1/2 −1/2
I V ar(Vj ) = q0j Σ22 Σ22 Σ22 qj = q0j qj = 1 for all j.
−1/2 −1/2
I Cov(Uj , Uk ) = p0j Σ11 Σ11 Σ11 pk = p0j pk = 0 for all
k 6= j.
−1/2 −1/2
I Cov(Vj , Vk ) = q0j Σ22 Σ22 Σ22 qk = q0j qk = 0 for all k 6= j.
−1/2 −1/2
I Cov(Uj , Vk ) = p0j Σ11 Σ12 Σ22 qk = p0j pk = 0 for all
j 6= k.

14 / 15
Properties

−1/2 −1/2
I E(Uj ) = p0j Σ11 µ1 and E(Vj ) = q0j Σ22 µ2 for all j.
−1/2 −1/2
I V ar(Uj ) = p0j Σ11 Σ11 Σ11 pj = p0j pj = 1 for all j.
−1/2 −1/2
I V ar(Vj ) = q0j Σ22 Σ22 Σ22 qj = q0j qj = 1 for all j.
−1/2 −1/2
I Cov(Uj , Uk ) = p0j Σ11 Σ11 Σ11 pk = p0j pk = 0 for all
k 6= j.
−1/2 −1/2
I Cov(Vj , Vk ) = q0j Σ22 Σ22 Σ22 qk = q0j qk = 0 for all k 6= j.
−1/2 −1/2
I Cov(Uj , Vk ) = p0j Σ11 Σ12 Σ22 qk = p0j pk = 0 for all
j 6= k.

14 / 15
Properties

−1/2 −1/2
I E(Uj ) = p0j Σ11 µ1 and E(Vj ) = q0j Σ22 µ2 for all j.
−1/2 −1/2
I V ar(Uj ) = p0j Σ11 Σ11 Σ11 pj = p0j pj = 1 for all j.
−1/2 −1/2
I V ar(Vj ) = q0j Σ22 Σ22 Σ22 qj = q0j qj = 1 for all j.
−1/2 −1/2
I Cov(Uj , Uk ) = p0j Σ11 Σ11 Σ11 pk = p0j pk = 0 for all
k 6= j.
−1/2 −1/2
I Cov(Vj , Vk ) = q0j Σ22 Σ22 Σ22 qk = q0j qk = 0 for all k 6= j.
−1/2 −1/2
I Cov(Uj , Vk ) = p0j Σ11 Σ12 Σ22 qk = p0j pk = 0 for all
j 6= k.

14 / 15
Summary

I The idea of canonical correlation is discussed.

I The expressions for the canonical correlations are derived.
I Some properties of canonical variables are discussed.

15 / 15
Summary

I The idea of canonical correlation is discussed.

I The expressions for the canonical correlations are derived.
I Some properties of canonical variables are discussed.

15 / 15
Summary

I The idea of canonical correlation is discussed.

I The expressions for the canonical correlations are derived.
I Some properties of canonical variables are discussed.

15 / 15