Global Lipschitz Bounds for Drift and Diffusion

What a Global Lipschitz Bound Controls

In an SDE, the drift coefficient describes the rate of deterministic change and the diffusion coefficient describes how the Brownian noise enters. Global Lipschitz bounds restrict how rapidly either coefficient can change when the state changes. “Global” means that one constant works for every pair of states, while “uniform in time” means that the same constant works at every time under consideration. These are precise restrictions on state dependence; they do not require the coefficients to be constant in time.

Use a filtered probability space $(\Omega,\mathcal{F},(\mathcal{F}_t)_{t\ge0},\mathbb{P})$ carrying an $m$-dimensional Brownian motion $\mathbf{B}_t$. For a $d$-dimensional state, consider the coefficient dimensions in $dX_t=\mu(t,X_t)\,dt+\sigma(t,X_t)\,d\mathbf{B}_t$: $\mu(t,x)\in\mathbb{R}^d$ and $\sigma(t,x)\in\mathbb{R}^{d\times m}$. Write $|\cdot|$ for the Euclidean vector norm and $\|A\|_F=(\sum_{i,j}A_{ij}^2)^{1/2}$ for the Frobenius matrix norm. The operator norm $\|A\|_{\mathrm{op}}=\sup_{|v|=1}|Av|$ is another possible matrix norm, but the inequalities below use the Frobenius norm unless stated otherwise.

Definition: On a time interval $0\le t\le T$, the drift and diffusion coefficients are globally Lipschitz in the state, uniformly in time if there is a finite constant $L\ge0$ such that, for every $t\in[0,T]$ and every $x,y\in\mathbb{R}^d$, $$ |\mu(t,x)-\mu(t,y)|\le L|x-y|, \qquad \|\sigma(t,x)-\sigma(t,y)\|_F\le L|x-y|. $$ The coefficients may have different best constants; one may use separate constants $L_\mu,L_\sigma$, or take $L=\max(L_\mu,L_\sigma)$. For a condition on all $t\ge0$, the constants must work on that entire time range.

The coefficient functions are usually also assumed measurable in time and continuous in the state, so that their evaluations along adapted processes are meaningful. Lipschitz continuity in the state does not itself require continuity, or even Lipschitz continuity, in time. For example, adding a measurable function of time to a coefficient does not change its state-Lipschitz bound.

The stochastic integral in the SDE is an Ito integral. On $[0,T]$, its integrand must be progressively measurable and belong to $\mathcal{L}^2$, meaning $\mathbb{E}\int_0^T\|H_s\|_F^2\,ds<\infty$ for a matrix-valued integrand $H_s$. A Lipschitz bound compares coefficient values at two states; by itself it does not assert this integrability for a particular process. For instance, conditions on the coefficient values at $x=0$ may be needed separately. Thus the definition is a coefficient regularity condition, not an existence theorem for an SDE solution.

Component Bounds and Matrix Norms

In applications, coefficients are often written component by component. The next result gives the exact implication needed to pass from scalar coordinate estimates to the vector and matrix inequalities in the definition. Conversely, a vector or Frobenius bound immediately bounds each component.

Theorem: Suppose that for every $t,x,y$, each drift coordinate satisfies $|\mu_i(t,x)-\mu_i(t,y)|\le L_i|x-y|$, and each diffusion entry satisfies $|\sigma_{ij}(t,x)-\sigma_{ij}(t,y)|\le K_{ij}|x-y|$. Then $$ |\mu(t,x)-\mu(t,y)|\le\Big(\sum_{i=1}^d L_i^2\Big)^{1/2}|x-y|, \qquad \|\sigma(t,x)-\sigma(t,y)\|_F\le\Big(\sum_{i=1}^d\sum_{j=1}^m K_{ij}^2\Big)^{1/2}|x-y|. $$ In the reverse direction, a vector bound with constant $L$ implies each drift coordinate bound with constant $L$; a Frobenius bound with constant $L$ implies each diffusion-entry bound with constant $L$.

Proof. Put $u=x-y$. By the definition of the Euclidean norm and the coordinate assumptions, $$ |\mu(t,x)-\mu(t,y)|^2 =\sum_{i=1}^d|\mu_i(t,x)-\mu_i(t,y)|^2 \le\sum_{i=1}^d L_i^2|u|^2 =\Big(\sum_{i=1}^dL_i^2\Big)|u|^2. $$ Taking nonnegative square roots proves the drift bound, including the case $u=0$. For the diffusion, the Frobenius norm gives $$ \|\sigma(t,x)-\sigma(t,y)\|_F^2 =\sum_{i=1}^d\sum_{j=1}^m|\sigma_{ij}(t,x)-\sigma_{ij}(t,y)|^2 \le\Big(\sum_{i=1}^d\sum_{j=1}^mK_{ij}^2\Big)|u|^2. $$ Taking square roots proves the asserted matrix bound. Conversely, the absolute value of one coordinate is at most the Euclidean norm of the vector, and the absolute value of one matrix entry is at most its Frobenius norm, since each squared coordinate or entry is one nonnegative term in the corresponding sum of squares. Applying these facts to the proposed vector and matrix bounds proves both reverse implications. $\square$

This result supplies valid bounds, but the constants obtained from coordinate estimates need not be optimal. In particular, a matrix operator norm and its Frobenius norm need not agree. In finite dimensions they are comparable, and either can express a global Lipschitz condition if the chosen norm and the associated constants are stated consistently.

Worked Example: Scalar Coefficients with Time Dependence

Let $\mu(t,x)=3x+\sin t$ and $\sigma(t,x)=2x+\cos t$, for $t\ge0$ and $x\in\mathbb{R}$. Find the smallest state-Lipschitz constants and verify the bounds at two specific states.

For any $x,y$, the time-only terms cancel: $$ |\mu(t,x)-\mu(t,y)|=|3x-3y|=3|x-y|, \qquad |\sigma(t,x)-\sigma(t,y)|=|2x-2y|=2|x-y|. $$ Thus $L_\mu=3$ and $L_\sigma=2$ work for every $t$. They are the smallest constants: whenever $x\ne y$, the corresponding ratios of coefficient difference to $|x-y|$ are exactly $3$ and $2$.

For a direct check, take $x=1$ and $y=-2$. Then $|x-y|=|1-(-2)|=3$, the drift difference is $|\mu(t,1)-\mu(t,-2)|=|3-(-6)|=9=3\cdot3$, and the diffusion difference is $|\sigma(t,1)-\sigma(t,-2)|=|2-(-4)|=6=2\cdot3$. The calculation does not depend on the chosen time, as required by uniformity.

Worked Example: Affine Vector Drift and Matrix Diffusion

For $x=(x_1,x_2)\in\mathbb{R}^2$, set $$ \mu(t,x)= \begin{pmatrix}1&1\\0&2\end{pmatrix}x+ \begin{pmatrix}\sin t\\ \cos t\end{pmatrix}, \qquad \sigma(t,x)= \begin{pmatrix}x_1&x_2\\2x_1&0\end{pmatrix}. $$ Here $d=m=2$. Find uniform Lipschitz constants in the Euclidean and Frobenius norms.

Write $u=x-y=(u_1,u_2)$. The additive time-dependent vector in $\mu$ cancels, leaving $Au$, where $A=\begin{pmatrix}1&1\\0&2\end{pmatrix}$. To compute its Euclidean operator norm, form $$ A^\mathsf{T}A=\begin{pmatrix}1&1\\1&5\end{pmatrix}. $$ Its characteristic polynomial is $\lambda^2-6\lambda+4$, whose roots are $3+\sqrt5$ and $3-\sqrt5$. Therefore the largest eigenvalue is $3+\sqrt5$, and $|Au|\le\sqrt{3+\sqrt5}\,|u|$. This is the exact constant, since a unit eigenvector for the largest eigenvalue attains the norm.

The diffusion difference is $$ \sigma(t,x)-\sigma(t,y)=\begin{pmatrix}u_1&u_2\\2u_1&0\end{pmatrix}, \qquad \|\sigma(t,x)-\sigma(t,y)\|_F^2=5u_1^2+u_2^2\le5(u_1^2+u_2^2). $$ Hence $L_\sigma=\sqrt5$ works, and it is attained when $u=(1,0)$. As a numerical check of the formula, for $x=(1,2)$ and $y=(0,1)$, $u=(1,1)$, so $|u|=\sqrt2$ and the matrix difference is $\begin{pmatrix}1&1\\2&0\end{pmatrix}$, with Frobenius norm $\sqrt6$. The bound gives $\sqrt5\sqrt2=\sqrt{10}$, and $\sqrt6\le\sqrt{10}$, as required.

Proposition: If $\mu(t,x)=A(t)x+b(t)$, where $A(t)$ is a $d\times d$ matrix, then the drift is globally Lipschitz uniformly in $t$ exactly when $\sup_t\|A(t)\|_{\mathrm{op}}<\infty$. If $\sigma(t,x)=C(t)x+D(t)$, where $C(t)$ is a linear map from $\mathbb{R}^d$ to $\mathbb{R}^{d\times m}$, then the diffusion has a uniform Frobenius Lipschitz bound exactly when $\sup_t\sup_{|u|=1}\|C(t)u\|_F<\infty$. The additive terms $b(t)$ and $D(t)$ do not affect these bounds.

Proof. For the drift, subtracting its values at $x$ and $y$ gives $A(t)(x-y)$. If the displayed supremum is finite, the definition of the operator norm gives $|A(t)(x-y)|\le(\sup_s\|A(s)\|_{\mathrm{op}})|x-y|$, which is a uniform global Lipschitz bound. Conversely, suppose a uniform bound $L$ holds. For every $t$ and every unit vector $u$, take $x=u$ and $y=0$. The bound gives $|A(t)u|\le L$. Taking the supremum over unit $u$, and then over $t$, proves $\sup_t\|A(t)\|_{\mathrm{op}}\le L$. For the diffusion, subtraction gives $C(t)(x-y)$. A finite supremum of $\|C(t)u\|_F$ over unit $u$ gives the required inequality by scaling, with the zero-vector case immediate. Conversely, apply a uniform diffusion bound to $x=u,y=0$ for each unit $u$; it yields $\|C(t)u\|_F\le L$, and taking both suprema proves finiteness. The additive terms cancel in both subtractions, proving the final assertion. $\square$

Worked Example: Nonlinear Coefficients with Bounded Slopes

Consider the two-dimensional drift and diffusion $$ \mu(t,x)=\begin{pmatrix}\sin x_1\\2\tanh x_2\end{pmatrix}, \qquad \sigma(t,x)=\begin{pmatrix}\sin(x_1+x_2)&0\\0&3\sin x_1\end{pmatrix}. $$ Show that both satisfy uniform global Lipschitz bounds.

The mean value theorem and the derivative bounds $|\cos z|\le1$ and $0\le(\tanh z)'=1/\cosh^2 z\le1$ give $|\sin a-\sin b|\le|a-b|$ and $|\tanh a-\tanh b|\le|a-b|$ for all real $a,b$. For $u=x-y$, the squared drift difference is at most $u_1^2+4u_2^2$, which is at most $4|u|^2$. Therefore $L_\mu=2$ works.

For the diffusion, its only potentially nonzero difference entries are $\sin(x_1+x_2)-\sin(y_1+y_2)$ and $3(\sin x_1-\sin y_1)$. Their absolute values are at most $|u_1+u_2|$ and $3|u_1|$, respectively. Since $(u_1+u_2)^2\le2(u_1^2+u_2^2)$, the squared Frobenius difference is at most $$ |u_1+u_2|^2+9u_1^2\le 11|u|^2. $$ Thus $L_\sigma=\sqrt{11}$ works for every $t,x,y$. For a check, take $x=(\pi/2,0)$, $y=(0,0)$. The state distance is $\pi/2$; the diffusion difference is $\begin{pmatrix}1&0\\0&3\end{pmatrix}$, with Frobenius norm $\sqrt{10}$. The claimed bound is $\sqrt{11}\,\pi/2$, which exceeds $\sqrt{10}$, consistent with the estimate.

What the Condition Does Not Say

A frequent error is to check a derivative bound only at a fixed state, or over a bounded region, and call the coefficient globally Lipschitz. For example, $\mu(x)=x^2$ has derivative $2x$, which is bounded on any fixed interval $[-R,R]$, but not on all of $\mathbb{R}$. A direct calculation rules out a global constant: for $x=n+1$ and $y=n$, with integer $n\ge0$, $$ \frac{|\mu(n+1)-\mu(n)|}{|(n+1)-n|} =\frac{|(n+1)^2-n^2|}{1}=2n+1. $$ These ratios are unbounded as $n$ increases. On $[-R,R]$, in contrast, factorization gives $|x^2-y^2|=|x-y||x+y|\le2R|x-y|$; that is a bound on the specified region, not a global one.

Another error is to treat a bound in one norm as though it had the same numerical constant in another norm. The coordinate theorem gives a safe conversion to the Euclidean and Frobenius norms, with dimension-dependent constants. It does not claim those converted constants are sharp. In particular, state the matrix norm whenever reporting a diffusion bound.

Global Lipschitz bounds are useful because they give uniform control of coefficient differences when states are compared, including in estimates for the Picard map. They are not conditions on Brownian increments, and they do not by themselves establish strong existence, pathwise uniqueness, or integrability of every stochastic integral. Those conclusions require additional arguments and hypotheses.

Check Your Understanding

For each question, distinguish a uniform bound on all states from a bound restricted to a region.

If each of the three drift coordinates has Lipschitz constant $2$, what Euclidean vector bound follows from the component theorem?
Why does adding a measurable time-only term $b(t)$ to an affine drift leave its state-Lipschitz constant unchanged?
Does $\mu(x)=x^2$ have a global Lipschitz constant on $\mathbb{R}$? Use the ratios for consecutive integers to justify your answer.
For a diffusion matrix, what norm is used in the definition here, and how does a Frobenius bound control each individual matrix entry?
Does a global Lipschitz bound in the state require the coefficient to be continuous in time? Explain what the bound does and does not restrict.

Global Lipschitz Bounds for Drift and Diffusion

What You'll Learn