Optimal control of a tumor growth model with hyperbolic relaxation of the chemical potential

1. Problem Statement

The paper addresses the optimal control of a phase-field model describing tumor growth. The model is a thermodynamically consistent system originally proposed by Hawkins-Daruud et al., but with a crucial modification: the standard parabolic relaxation of the chemical potential is replaced by a hyperbolic relaxation.

The State System:
The system consists of three coupled partial differential equations (PDEs) defined on a bounded domain $\Omega \subset \mathbb{R}^3$ over a time interval $(0, T)$:

1. Chemical Potential Equation (Hyperbolic):
   $\alpha \partial_{tt}\mu + \partial_t \phi - \Delta \mu = P(\phi)(\sigma + \chi(1-\phi) - \mu) - h(\phi)u_1$
   Here, $\alpha > 0$ is the hyperbolic relaxation parameter. The term $\alpha \partial_{tt}\mu$ introduces a wave-like behavior, distinguishing it from standard parabolic models.
2. Phase Field Equation (Viscous Cahn-Hilliard):
   $\tau \partial_t \phi - \Delta \phi + F'(\phi) = \mu + \chi \sigma$
   This equation governs the evolution of the tumor phase variable $\phi$ (where $\phi \approx 1$ is tumor, $\phi \approx -1$ is healthy tissue). It includes a viscosity term $\tau \partial_t \phi$.
3. Nutrient Equation:
   $\partial_t \sigma - \Delta \sigma = -\chi \Delta \phi - P(\phi)(\sigma + \chi(1-\phi) - \mu) + u_2$
   This describes the nutrient concentration $\sigma$.

Variables and Controls:

• State Variables: $\mu$ (chemical potential), $\phi$ (phase field), $\sigma$ (nutrient concentration).
• Control Variables: $u = (u_1, u_2)$.
  • $u_1$: Represents the application of a cytotoxic drug, coupled nonlinearly via a truncation function $h(\phi)$ to restrict action to the tumor region.
  • $u_2$: Represents the supply of nutrients or medication.
• Potential $F(\phi)$: A double-well potential driving phase separation. The paper analyzes two cases:
  • Regular: $F_{reg}(r) = \frac{1}{4}(1-r^2)^2$ (defined on $\mathbb{R}$).
  • Singular (Logarithmic): $F_{log}(r)$ (defined on $(-1, 1)$), which is physically more relevant but mathematically challenging due to singularities at the boundaries.

The Optimal Control Problem (CP):
The goal is to minimize a cost functional $J$ that penalizes:

• Deviation of the tumor phase $\phi$ from a desired target profile $\hat{\phi}_Q$ (in space-time) and $\hat{\phi}_\Omega$ (at final time).
• The magnitude of the control effort ($L^2$ norm of $u$).
• A sparsity term $\kappa G(u)$, where $G(u) = \|u\|_{L^1(Q)}$, encouraging controls that are zero on large sub-regions (sparse controls).

2. Methodology

The authors employ a rigorous functional analytic approach within the framework of infinite-dimensional optimization.

Key Mathematical Tools:

• Well-posedness Analysis: They rely on previous results (from [7]) establishing existence, uniqueness, and regularity for the state system.
• Separation Condition: A critical technical hurdle for singular potentials is ensuring the phase variable $\phi$ stays strictly within the physical range $(-1, 1)$ (uniform separation). The authors prove that for singular potentials, this requires stronger assumptions on the control $u_1$ (specifically $u_1 \in L^\infty(Q) \cap L^2(0,T; V)$) compared to regular potentials.
• Differentiability of the Control-to-State Operator:
  • The core of the methodology is proving that the mapping $S: u \mapsto (\mu, \phi, \sigma)$ is Fréchet differentiable.
  • This is achieved by analyzing the linearized system (perturbation equations) and deriving a priori estimates in suitable Banach spaces.
  • For singular potentials, the analysis is split into specific "Scenarios" (e.g., fixing one control component or using product-type controls) to satisfy the uniform separation condition required for differentiability.
• Adjoint System: To derive optimality conditions, they introduce an adjoint system (backward in time) involving variables $(p, q, r)$. The well-posedness of this adjoint system is proven using Faedo-Galerkin approximations and energy estimates.
• Variational Inequalities: The first-order necessary optimality conditions are derived using the chain rule and subdifferential calculus for convex functionals, leading to variational inequalities involving the adjoint states.

3. Key Contributions

1. Hyperbolic Relaxation in Tumor Models: This is one of the first rigorous optimal control studies for tumor growth models incorporating hyperbolic (wave-like) relaxation of the chemical potential, moving beyond the standard parabolic (diffusive) assumption.
2. Handling Singular Potentials: The paper successfully extends optimal control theory to singular logarithmic potentials. It explicitly addresses the difficulty of the "uniform separation condition" and proposes specific control scenarios (Scenario 1, 2, 3) where the analysis remains valid despite the singularity.
3. Fréchet Differentiability: The authors establish the Fréchet differentiability of the control-to-state operator for both regular and singular potentials in appropriate Banach spaces, a prerequisite for deriving first-order optimality conditions.
4. Sparsity Analysis: The paper derives explicit conditions under which optimal controls exhibit sparsity (i.e., vanish on sub-regions of space-time). This is crucial for medical applications where minimizing drug dosage and treatment duration is vital.

4. Main Results

A. Well-posedness and Regularity (Theorems 2.2 & 2.3):

• Existence and uniqueness of solutions are established for both regular and singular potentials.
• For singular potentials, a uniform separation condition ($r_* \le \phi \le r^*$ strictly inside $(-1, 1)$) is proven to hold if the control $u_1$ satisfies specific regularity ($L^\infty \cap L^2(V)$).

B. Differentiability (Theorem 3.4):

• The control-to-state operator $S$ is Fréchet differentiable from $L^\infty(Q) \times L^2(Q)$ into a Banach space $Y$.
• The derivative is characterized as the solution to the linearized system (3.22)–(3.26).

C. First-Order Necessary Optimality Conditions (Theorem 3.8 & 4.x):

• For a locally optimal control $\bar{u}$, there exist adjoint states $(p, q, r)$ and a subgradient $\lambda \in \partial G(\bar{u})$ such that:
  $$\int_Q (d_0 + \kappa \lambda + b_3 \bar{u}) \cdot (u - \bar{u}) \ge 0 \quad \forall u \in U_{ad}$$
  where $d_0 = (-h(\phi)p, r)$.
• Pointwise Projection Formulas: Under specific assumptions on the sparsity functional, the optimal controls can be expressed via projection formulas:
  $$\bar{u}_1 = \text{proj}_{[u_1, \bar{u}_1]} \left( -b_3^{-1}(-h(\phi)p + \kappa \lambda_1) \right)$$
  $$\bar{u}_2 = \text{proj}_{[u_2, \bar{u}_2]} \left( -b_3^{-1}(r + \kappa \lambda_2) \right)$$

D. Sparsity Results (Theorem 5.1):

• If the sparsity parameter $\kappa$ is sufficiently large, the optimal controls vanish where the adjoint variables are small.
• Specifically, $\bar{u}_1(x,t) = 0 \iff |h(\phi)p| \le \kappa$ and $\bar{u}_2(x,t) = 0 \iff |r| \le \kappa$.
• This implies that for large $\kappa$, the treatment is applied only where the "gradient" of the error (measured by the adjoint state) exceeds a threshold, leading to sparse treatment strategies.

5. Significance

• Mathematical Rigor: The paper provides a complete analytical framework for a complex, coupled, hyperbolic-parabolic system with singular nonlinearities, bridging the gap between theoretical PDE analysis and optimal control.
• Medical Relevance: By incorporating sparsity, the model directly addresses clinical needs. It mathematically justifies treatment strategies that are "on-off" or localized in space and time, potentially reducing side effects and drug resistance.
• Physical Realism: The inclusion of hyperbolic relaxation allows for modeling finite speed of propagation of chemical signals, which may be more physically accurate than infinite-speed diffusion in certain biological contexts.
• Handling Singularities: The specific scenarios developed for singular potentials offer a blueprint for handling similar constraints in other phase-field control problems where physical bounds are strict.

In summary, this work advances the mathematical theory of tumor growth control by tackling the complexities of hyperbolic dynamics and singular potentials, ultimately providing rigorous conditions for designing sparse, efficient therapeutic strategies.