1 Introduction and main results

The Sharp Gårding inequality is a powerful tool in the study of systems of PDE. Let \(P=p(x,D_x)=(P_{jk})\) be an \(\ell \times \ell \) matrix of operators \(P_{jk}=p_{jk}(x,D_x)\) with matrix symbol \(p(x,\xi )=(p_{jk}(x,\xi ))\in S^m_{\rho ,\delta }\), \(0\le \delta <\rho \le 1\), that is satisfying

$$\begin{aligned} |\partial _\xi ^\alpha D_x^\beta p(x,\xi )|\le C_{\alpha ,\beta }\langle \xi \rangle ^{m-\rho |\alpha |+\delta |\beta |},\ \ \langle \xi \rangle =\sqrt{1+|\xi |^2}. \end{aligned}$$
(1.1)

Assume that the Hermitian part \(p'=(p+p^*)/2\) of \(p(x,\xi )\) is positive semidefinite. Then, there exists \(C>0\) such that

$$\begin{aligned} \Re (Pu,u)\ge -C\Vert u\Vert ^2_{H^{(m-\mu )/2}},\ \ \mu =\rho -\delta , \end{aligned}$$
(1.2)

for every \(u\in {\mathcal S}\). In particular, for \(p(x,\xi )\in S^m_{1,0}\) we have

$$\begin{aligned} \Re (Pu,u)\ge -C\Vert u\Vert ^2_{H^{(m-1)/2}}. \end{aligned}$$
(1.3)

Hörmander [4] proved inequality (1.3) for scalar operators and Lax-Nirenberg [7] extended this result to systems. Friedrichs [3], Kumano-Go [5] and others improved it and simplified the proof.

For scalar operators, there is the great strengthening \(\mu =2(\rho -\delta )\) in (1.2) due to Fefferman and Phong [2] but for matrix operators with smooth symbol the bound for \(\mu \) remains \(\mu =\rho -\delta \).

In many applications operators with symbol of limited smoothness are involved. Let us consider \(p(x,\xi )\) in the class \(C^sS^m_{1,0}\) of symbols with \(C^s\) regularity in the space variable x defined by

$$\begin{aligned} \Vert \partial _\xi ^\alpha p(x,\xi )\Vert _{C^s}\le C_{\alpha }\langle \xi \rangle ^{m-\alpha |}. \end{aligned}$$
(1.4)

For any fixed \(\delta \in ]0,1[\) one can regularize the symbol obtaining a splitting

$$\begin{aligned} p(x,\xi )=p^\sharp (x,\xi )+p^b(x,\xi ), \ \ p^\sharp (x,\xi )\in S^m_{1,\delta }, \ p^b(x,\xi )\in C^sS^{m-s\delta }_{1,\delta }, \end{aligned}$$
(1.5)

e.g. Taylor [9]. If \(p'\) is positive semidefinite, then the Hermitian part of \(p^\sharp (x,\xi )\) is positive semidefinite as well. Applying (1.2) to \(P^\sharp (x,D_x)\) and using the boundedness

$$\begin{aligned} P^b(x, D_x): H^m\rightarrow H^{m-s\delta }, \end{aligned}$$

the sharp Gårding inequality (1.2) for \(P(x,D_x)\) holds true with a order

$$\begin{aligned} \mu \le 1-\delta , \ \mu \le s\delta . \end{aligned}$$

Negotiating on \(\delta \) as done in [9], one obtains (1.2) for \(p(x,\xi )\in C^sS^m_{1,0}\) with

$$\begin{aligned} \mu = \frac{s}{s+1}. \end{aligned}$$
(1.6)

Taylor’s bound (1.6) gives \(\mu \rightarrow 1\) for \(s\rightarrow \infty \) but it is not optimal. By means of the paradifferential calculus, Bony [1] proved that the best possible bound \(\mu =1\) is achieved already for \(s=2\). For \(0<s<2\) Bony obtained the bound \(\mu =s/2\) which is better than Taylor’s one for \(1<s<2\) but it is worse for \(0<s<1\).

Conjugating the operator with the FBI transform, Taturu [8] proved a generalization of the Sharp Gårding inequality for regular symbols from which he obtained also inequality (1.2) for symbols \(p(x,\xi )\in C^sS^m_{1,0}\) with

$$\begin{aligned} \mu =\mu ^*(s)= {\left\{ \begin{array}{ll} 1,\ s\ge 2,\\ 2s/(s+2), \ 0<s<2. \end{array}\right. } \end{aligned}$$
(1.7)

We believe this one the optimal estimate for \(C^s\) symbols, agreeing with Tataru.

Our aim is to show that a generalization of the Sharp Gårding inequality for regular symbols, sufficient to get \(\mu =\mu ^*(s)\) in the case of \(C^s\) limited smoothness, can be proved directly from Friedrichs symmetrization, that is going back to the original proofs of (1.2) in [3, 5, 6].

As in Tataru’s result, what is really important is the order of \(\partial _x^\beta p(x,\xi )\) with \(|\beta |=2\), let us denote \(m+m_2\) this order. From \(p(x,\xi )\in S^m_{\rho ,\delta }\) clearly we have \(m_2\le 2\delta \). In case of equality one can not obtain better than \(\mu =\rho -\delta \) in (1.2) but we can improve this bound in the case \(m_2<2\delta \). As we will see later on, this is exactly what happens for \(p^\sharp (x,\xi )\in S^m_{1,\delta }\) in the splitting (1.5) of \(p(x,\xi )\in C^sS^m_{1,0}\).

For sake of simplicity, from now on we take \(\rho =1\) which is the case of our interest. Here we prove the following generalization of inequality (1.2) for regular symbols.

Theorem 1.1

Let \(P=p(x,D_x)=(P_{jk})\) be an \(\ell \times \ell \) matrix of operators \(P_{jk}=p_{jk}(x,D_x)\) with matrix symbol \(p(x,\xi )=(p_{jk}(x,\xi ))\in S^m_{1,\delta }\), \(0\le \delta <1\), and such that

$$\begin{aligned} \partial _x^\beta p(x,\xi )\in S^{m+m_1}_{1,\delta }, \ |\beta |=1;\ \ \ \partial _x^\beta p(x,\xi )\in S^{m+m_2}_{1,\delta }, \ |\beta |=2. \end{aligned}$$
(1.8)

Assume that the Hermitian part \(p'=(p+p^*)/2\) of \(p(x,\xi )\) is positive semidefinite.

Then, there exists \(C>0\) such that

$$\begin{aligned} \Re (Pu,u)\ge -C\Vert u\Vert ^2_{(m-\mu ^\sharp )/2} \end{aligned}$$
(1.9)

for every \(u\in {\mathcal S}\), with

$$\begin{aligned} \mu ^\sharp = {\left\{ \begin{array}{ll} \min \{1-m_1, 1-m_2/2\}, \ 2\delta -1\le m_2/2,\\ \min \{1-m_1, 2(1-\delta )\},\ \ \ 2\delta -1> m_2/2. \end{array}\right. } \end{aligned}$$
(1.10)

For the largest possible \(m_2=2\delta \) of course we have \(2\delta -1< m_2/2\) hence the general bound \(\mu ^\sharp =1-\delta \). The same we have with \(m_1=\delta \) and any \(m_2\le 2\delta \).

With \(m_2<2\delta \) and \(m_1<\delta \) there is a gain. For instance, for \(m_1=m_2=0\) we have \(\mu ^\sharp =1\) for \(0\le \delta \le 1/2\) and \(\mu ^\sharp =2(1-\delta )\) for \(1/2<\delta <1\). Spending such a gain we can prove the result for symbols of limited smoothness.

Theorem 1.2

Let \(P=p(x,D_x)=(P_{jk})\) be an \(\ell \times \ell \) matrix of operators with symbol \(p(x,\xi )=(p_{jk}(x,\xi ))\in C^sS^m_{1,0}\). Assume that the Hermitian part \(p'=(p+p^*)/2\) of \(p(x,\xi )\) is positive semidefinite.

Then, there exists \(C>0\) such that

$$\begin{aligned} \Re (Pu,u)\ge -C\Vert u\Vert ^2_{(m-\mu ^*(s))/2} \end{aligned}$$
(1.11)

for every \(u\in {\mathcal S}\), with

$$\begin{aligned} \mu ^*(s)= {\left\{ \begin{array}{ll} 1,\ s\ge 2,\\ 2s/(s+2), \ 0<s<2. \end{array}\right. } \end{aligned}$$
(1.12)

2 Proof of Theorem 1.1

We follow the proof of Friedrichs [3] and Kumano-go [5, 6].

Let \(p(x,\xi )\in S^m_{1,\delta }\) and for \(\delta '\ge 2\delta -1\), \(\tau =(1+\delta ')/2\) \((\ge \delta )\) let us consider

$$\begin{aligned} p_0(x,\xi )=\int p(x,\xi +\sigma \langle \xi \rangle ^\tau )q(\sigma )^2d\sigma \end{aligned}$$
(2.1)

where \(q(\sigma )\ge 0\) is a smooth function of \(\sigma \in {\mathbb R}^n\) with support for \(|\sigma |<1\), \(q(\sigma )=q(-\sigma )\), \(\int q(\sigma )^2d\sigma =1\).

In the original proof \(\tau =(1+\delta )/2\) that is \(\delta '=\delta \) from the beginning. We take some advantage by fixing \(\delta '\in [2\delta -1,1[\) related to \(m_2\) later on.

Performing a change of variable in the integral (2.1) we have

$$\begin{aligned} p_0(x,\xi )=\int p(x,\zeta )F(\xi ,\zeta )^2d\zeta \end{aligned}$$
(2.2)

with

$$\begin{aligned} F(\xi ,\zeta )= q((\zeta -\xi )\langle \xi \rangle ^{-\tau })\langle \xi \rangle ^{-\tau n/2}. \end{aligned}$$
(2.3)

To obtain a symmetric operator, we introduce the double symbol \(p_F(\xi ,x',\xi ')\), such that \(p_F(\xi ,x,\xi )=p_0(x,\xi )\), defined by

$$\begin{aligned} p_F(\xi ,x',\xi ')=\int F(\xi ,\zeta )p(x',\zeta )F(\xi ',\zeta )d\zeta . \end{aligned}$$
(2.4)

We denote again \(p_F(x,\xi )\) the simplified symbol of the operator \(P_F(x, D_x)\).

If the matrix is \(p(x,\xi )\) is positive semidefinite, then \(P_F\) is a positive operator:

$$\begin{aligned} (P_Fu,u)\ge 0, \ u\in {\mathcal S}, \end{aligned}$$

see Theorem 4.3 in [6].

Taking \(\tau =(1+\delta ')/2>\delta \) (this is the case with the original choice \(\delta '=\delta \) of [6]), from the proof of Theorem 4.2 in [6] we have that the simplified symbol \(p_F(x,\xi )\) of the operator \(P_F\) belongs to the class \(S^m_{1,\delta }\) and has an asymptotic expansion

$$\begin{aligned}{} & {} p_F(x,\xi )\sim p(x,\xi )+\sum _{|\beta |=1}\psi _\beta (\xi )p_{(\beta )}(x,\xi )+\sum _{|\alpha +\beta |\ge 2}\psi _{\alpha ,\beta }(\xi )p_{(\beta )}^{(\alpha )}(x,\xi ),\nonumber \\{} & {} \qquad \quad \qquad \quad \psi _\beta \in S^{-1},\ \psi _{\alpha ,\beta }\in S^{\tau (|\alpha |-|\beta |)}. \end{aligned}$$
(2.5)

Looking at the orders of \(\psi _\beta \) and \(\psi _{\alpha ,\beta }\) and at the orders of \(\partial _x^\beta p(x,\xi )\) for \(|\beta |\le 2\) in (1.8), from the above expansion we get

$$\begin{aligned}{} & {} p(x,\xi )=p_F(x,\xi )+p_1(x,\xi )+p_2(x,\xi ),\nonumber \\{} & {} p_1(x,\xi )\in S_{1,\delta }^{m-(1-m_1)},\ p_2(x,\xi )\in S_{1,\delta }^{m-\mu _2},\nonumber \\{} & {} \mu _2=\mu _2(\delta ')=\min \{1-\delta ', 1+\delta '-m_2\}. \end{aligned}$$
(2.6)

In the limit case \(\tau =(1+\delta ')/2=\delta \) the proof of Thoerem 4.2 in [6] still gives (2.6) with the difference \(p_2(x,\xi )\in S_{\delta ,\delta }^{m-\mu _2}\) instead of \(S_{1,\delta }^{m-\mu _2}\) and what we loose in this case is the complete asymptotic expansion (2.5) which is not essential for our aims.

The positivity of the operator \(P_F\) and the orders of \(P_1\), \(P_2\) in the splitting (2.6) yield inequality (1.2) for \(P=P_F+P_1+P_2\) with

$$\begin{aligned} \mu \le \min \{1-m_1,\mu _2\}. \end{aligned}$$

The order of \(P_1\) gives the bound \(\mu ^\sharp \le 1-m_1\) for \(\mu ^\sharp \) in (1.10). Then, we have to maximize \(\mu _2=\mu _2(\delta ')\) in (2.6) for \(2\delta -1\le \delta '<1\) in order to get the best possible second bound. Since

$$\begin{aligned} \max _{2\delta -1\le \delta '<1}\mu _2(\delta ')= {\left\{ \begin{array}{ll} 1-m_2/2,\ 2\delta -1\le m_2/2,\\ 2-2\delta ,\ 2\delta -1>m_2/2, \end{array}\right. } \end{aligned}$$

we complete the proof of Theorem 1.1.

3 Proof of Theorem 1.2

Let us show how Theorem 1.1 implies Theorem 1.2. Coming back to the splitting (1.5) of \(p(x,\xi )\in C^sS^m_{1,0}\), now we have to negotiate between \(\mu ^\sharp =\mu ^\sharp (\delta ,m_1,m_2)\) of Theorem 1.1 for \(p^\sharp \) and \(s\delta \). We obtain the optimal bound \(\mu =\mu ^*(s)\) for

$$\begin{aligned} \mu ^\sharp =s\delta . \end{aligned}$$

We use the more precise estimates for the regularized part \(p^\sharp \)

$$\begin{aligned} \partial _x^\beta p^\sharp \in S^m_{1,\delta }, \ |\beta |\le s; \ \ \partial _x^\beta p^\sharp \in S^{m+\delta (|\beta |-s)}_{1,\delta }, \ |\beta |> s, \end{aligned}$$
(3.1)

given by Proposition 1.3.D in [9]. This means, with our notation,

$$\begin{aligned} m_1=m_1(s)= {\left\{ \begin{array}{ll} 0, \ s\ge 1,\\ \delta (1-s),\ 0< s<1, \end{array}\right. } \end{aligned}$$
(3.2)

and

$$\begin{aligned} m_2=m_2(s)= {\left\{ \begin{array}{ll} 0, \ s\ge 2,\\ \delta (2-s),\ 0< s<2. \end{array}\right. } \end{aligned}$$
(3.3)

We have \(m_1\le m_2/2\) in any case. In particular \(m_1\) does not influence \(\mu ^\sharp \) and (1.10) for \(p^\sharp \) reduces to

$$\begin{aligned} \mu ^\sharp = {\left\{ \begin{array}{ll} 1-m_2/2, \ 2\delta -1\le m_2/2,\\ 2(1-\delta ),\ 2\delta -1> m_2/2. \end{array}\right. } \end{aligned}$$
(3.4)

Here the best choice, if it is possible to fix \(\delta \) such that \(2\delta -1\le m_2/2\), is always \(\mu ^\sharp =1-m_2/2\).

For \(s\ge 2\) we have \(m_2=0\) in (3.3). Choosing \(\delta =1/s\), (\(2\delta -1\le m_2/2\) reads exactly \(s\ge 2\)), we have

$$\begin{aligned} \mu ^\sharp =1-m_2/2=1=s\delta \end{aligned}$$
(3.5)

and the best possible bound \(\mu =\mu ^*(s)=1\) is achieved in (1.12).

For \(0<s< 2\) we have \(m_2=\delta (2-s)\) in (3.3). Choosing \(\delta =2/(s+2)\) we have \(2\delta -1= m_2/2\) and

$$\begin{aligned} \mu ^\sharp =1-m_2/2=2s/(s+2)=s\delta \end{aligned}$$
(3.6)

that leads to \(\mu =\mu ^*(s)=2s/(s+2)\) in (1.12).

This completes the proof of Theorem 1.2.