NPTEL – Mechanical Engineering – Continuum Mechanics

Module-1: Tensor Algebra

Lecture-11: Positive Definite Tensor and
the Polar Decomposition

Here, we discuss the positive definiteness and polar decomposition of second-order tensors.
The polar decomposition plays an important role in defining generalized strain and also
in the applications of principle of material frame-indifferent.
Definition of positive definite tensor:
A second-order tensor S is said to be positive definite if

(u, Su) ≥ 0, ∀u ∈ V with (u, Su) = 0 if and only if u = 0. (1)

Any second-order tensor T can be uniquely decomposed into a symmetric part Ts , and a
skew-symmetric part Tss , i.e., T = Ts + Tss where Ts = 12 (T + T T ) and Tss = 21 (T − T T )
(see Lecture-4). Therefore, we have

(u, T u) = (u, Ts u) + (u, Tss u)

= (u, Ts u) (Since (u, Tss u) = 0 ∀u ∈ V)

Thus, it is sufficient to study the positive definiteness of symmetric tensors.

Problem 1. Given a symmetric tensor S is positive definite if and only if all eigenvalues
are positive.

Proof. Let us assume that the given tensor S is positive definite tensor. Then we have

(u, Su) > 0, for every non-zero vector u ∈ V

Let λ and n be an eigenvalue and its corresponding eigenvector of S. Then we have the
relation Sn = λn. Taking inner product with n on both sides give rise to

(n, Sn)
λ= (2)
(n, n)

Since (n, n) > 0 and S is positive definite tensor (i.e., (n, Sn) > 0), Eq. (2) implies
λ > 0. Thus, positive definite property of tensor S implies positive nature of eigenvalues.
Conversely, let all eigenvalues of symmetric tensor S be positive. Then using spectral
decomposition (see Problem-6 in Lecture-10), we can have

S = λ1 e∗1 ⊗ e∗1 + λ2 e∗2 ⊗ e∗2 + λ3 e∗3 ⊗ e∗3

Joint initiative of IITs and IISc – Funded by MHRD, Government of India 1

NPTEL – Mechanical Engineering – Continuum Mechanics

where λ1 , λ2 , λ3 are eigenvalues and e∗1 , e∗2 , e∗3 are corresponding orthonormal eigenvectors.
Let u be a non-zero vector. Then

Su = (λ1 e∗1 ⊗ e∗1 + λ2 e∗2 ⊗ e∗2 + λ3 e∗3 ⊗ e∗3 )u

= λ1 (e∗1 ⊗ e∗1 )u + λ2 (e∗2 ⊗ e∗2 )u + λ3 (e∗3 ⊗ e∗3 )u
= λ1 (e∗1 · u)e∗1 + λ2 (e∗2 · u)e∗2 + λ3 (e∗3 · u)e∗3
= λ1 u∗1 e∗1 + λ2 u∗2 e∗2 + λ3 u∗3 e∗3

where u∗1 , u∗2 , u∗3 are components vector in the basis {e∗1 , e∗2 , e∗3 }.

Taking inner product of u and Su, we get

(u, Su) = u · (λ1 u∗1 e∗1 + λ2 u∗2 e∗2 + λ3 u∗3 e∗3 )

= λ1 u∗1 (u · e∗1 ) + λ2 u∗2 (u · e∗2 ) + λ3 u∗3 (u · e∗3 )
= λ1 (u∗1 )2 + λ2 (u∗2 )2 + λ3 (u∗3 )2

The positive property of eigenvalues implies (u, Su) > 0. Suppose, (u, Su) = 0 then
from preceding equation u∗i = 0, which implies that u = 0. Thus, S is positive definite
tensor if and only if all eigenvalues are positive.

Problem 2. All principal invariants of symmetric tensor S are positive if and only if all
eigenvalues are positive.

Proof. Let λ1 , λ2 and λ3 be eigenvalues of S and I1 , I2 and I3 be three principal invariants.

Then we have the following relations between eigenvalue and principal invariants.

I1 = λ1 + λ2 + λ3
I2 = λ1 λ2 + λ2 λ3 + λ3 λ1
I3 = λ1 λ2 λ3

It is easy to see that if eigenvalues are positive then principal invariants are also positive.
Conversely, let principal invariants I1 , I2 and I3 be positive. The characteristic equation
of tensor S is given by
λ3 − I1 λ2 + I2 λ − I3 = 0,
where λ is an eigenvalue of tensor S. We know that eigenvalues of symmetric tensor are
real. Hence, if eigenvalue is negative then it cannot satisfy the characteristic equation as
the term λ3 − I1 λ2 + I2 λ − I3 is negative. Therefore, eigenvalue must be positive since the
principal invariants are positive. Thus, all principal invariants are positive if and only if
eigenvalues are positive.

Joint initiative of IITs and IISc – Funded by MHRD, Government of India 2

NPTEL – Mechanical Engineering – Continuum Mechanics

Problem 3 (Square root tensor). Let S be a positive definite symmetric tensor. Then
2
there exist a unique positive definite symmetric tensor U such
√ that U = S. The tensor
U is called square root tensor of S. We usually write U = S.

Proof. If U is not positive definite symmetric (P sym) tensor then there can be many U
satisfying U U = S such that S is P sym. For example, the matrices diag[1, -1, 1] (i.e.
diagonal matrix with entities 1, -1, 1 on the diagonal) and diag[1, -1, -1] are non-positive
definite square roots of the identity tensor I. Therefore, in order to ensure uniqueness,
we stated in the problem that U is the positive definite symmetric tensor. We now show
the relation between the given positive definite symmetric tensor and its positive definite
square root. Consider the spectral decomposition of given symmetric tensor S.

S = λ1 e∗1 ⊗ e∗1 + λ2 e∗2 ⊗ e∗2 + λ3 e∗3 ⊗ e∗3

where λ1 , λ2 and λ3 are eigenvalues and e∗1 , e∗2 and e∗3 are corresponding orthonormal
eigenvectors.
The eigenvalues are positive as S is a positive definite symmetric tensor. Therefore,
we can define a tensor
q q q
U= λ1 e∗1 ⊗ e∗1 + λ2 e∗2 ⊗ e∗2 + λ3 e∗3 ⊗ e∗3 . (3)
√ √
Since the set {e∗1 , e∗2 , e∗3 } is orthonormal, it is easy to see U U = S. Clearly, λ1 , λ2
√
and λ3 are eigenvalues and e∗1 , e∗2 and e∗3 are corresponding orthonormal eigenvectors
of U . Thus, there is a positive definite square root of S. We now show the uniqueness of
positive definite square root.
Let λ and n be the eigenvalue and corresponding eigenvector of S. Then we have

0 = (S − λI)n
= (U 2 − λI)n
√ √
= (U + λI)(U − λI)n.

√
Suppose (U − λI)n = ñ then we have
√
(U + λI)ñ = 0.

This equation√implies ñ = 0. Otherwise, the tensor U violates the positive definite

property as − λ is an eigenvalue. Therefore,
√
(U − λI)n = 0.
√
clear that the vector n is eigenvector corresponding to an eigenvalue λ. Thus,
It is √
U = S is unique as the spectral decomposition of S is unique.

We have discussed the additive decomposition of arbitrary tensor into symmetric and
skew-symmetric parts. We now discuss the multiplicative decomposition of invertible
tensors given by the polar decomposition theorem.

Joint initiative of IITs and IISc – Funded by MHRD, Government of India 3

NPTEL – Mechanical Engineering – Continuum Mechanics

The following polar decomposition is motivated by a representation of complex num-

bers. Every non-zero complex number, z = x + iy, is expressible uniquely as z = reiθ
√
where r > 0 and θ ∈ [0, 2π). We note that |eiθ | = 1 and r = z̄z where z̄ is complex
conjugate of z.
Problem 4 (Polar decomposition theorem). Every invertible tensor F can be uniquely
factored into the following forms

F = QU = V Q, (4)

where Q is an orthogonal tensor and U , V are positive definite symmetric tensors.

Proof. Since given tensor F is invertible, F T F is positive definite symmetric tensor, i.e.,

(u, F T F u) = (F u, F u) > 0 for every non-zero vector u ∈ V

and (F u, F u) = 0 if and only if u = 0.

√
We can define a unique square root tensor, U = F T F (see Problem-3). Furthermore,
the square root tensor U is invertible as it is positive definite tensor. Therefore, we can
have
Q = F U −1 . (5)
It is clear from the following steps that Q is an orthogonal tensor.

QT Q = (F U −1 )T (F U −1 )
= U −T F T F U −1
= U −1 (F T F )U −1 (since U is symmetric)
−1 −1
= U (U U )U
= I

Thus, Eq. (5) implies first part of polar decomposition F = QU , where Q is orthogonal
tensor and U is positive definite tensor.
We now show the uniqueness of the decomposition F = QU . Suppose F has second
polar decomposition such that F = Q̆Ŭ where Q̆ is also orthogonal tensor. We can
obtain F T F = Ŭ T Q̆T Q̆Ŭ = Ŭ 2 and we know F T F = U 2 . Hence, uniqueness of square
root tensor implies Ŭ = U and Q̆ = Q. Furthermore, The det(Q) and det(F ) must have
same sign as det(U ) > 0. Clearly, the tensor Q is rotation as det(F ) is positive.
T
We note that invertibility of F implies invertibility
√ of F . Therefore, similar to pre-
vious case we can define a square root tensor V = F F T since F F T is positive defi-
nite symmetric tensor. We now define an orthogonal tensor Q̃ = V −1 F and it implies
F = V Q̃. It is easy to see that Q̃ is an orthogonal tensor, i.e., Q̃T Q̃ = I. We can have
the relation F = Q̃(Q̃T V Q̃) since Q̃Q̃T = I. We now show that Q̃ = Q. Since F = QU
and F = V Q̃, we can write

QU = V Q̃ =⇒ U = QT V Q̃.

Since U and V are symmetric tensors, taking transpose on both sides of above equation,
we get
U = Q̃T V Q =⇒ Q̃U = V Q

Joint initiative of IITs and IISc – Funded by MHRD, Government of India 4

NPTEL – Mechanical Engineering – Continuum Mechanics

The uniqueness of F = QU factorization implies Q̃ = Q. Thus, we have following unique

decomposition
F = QU = V Q.
We note that the tensors U and V have same eigenvalues since U = QT V Q. If λ1 , λ2 ,
λ3 are eigenvalues and e∗1 , e∗2 , e∗3 are eigenvectors of U respectively then λ1 , λ2 , λ3 are
eigenvalues and Qe∗1 , Qe∗2 , Qe∗3 eigenvectors of V .

Reference

1. C. S. Jog, Continuum Mechanics: Foundations and Applications of Mechanics, Vol-

ume I, Third edition, 2015, Cambridge University Press.

Joint initiative of IITs and IISc – Funded by MHRD, Government of India 5