Introduction
The long-time behavior of solutions to semilinear parabolic equations of the form
∂ t α u + ( − Δ ) θ / 2 u = h ( t ) u p , h ( t ) ∼ t γ ℓ ( t ) , ( t , x ) ∈ ( 0 , ∞ ) × R N (1) \partial_t^\alpha u+(-\Delta)^{\theta/2}u=h(t)u^p,\qquad h(t)\sim t^\gamma\ell(t),\qquad(t,x)\in(0,\infty)\times\mathbb{R}^N \tag*{(1)} ∂ t α u + ( − Δ ) θ /2 u = h ( t ) u p , h ( t ) ∼ t γ ℓ ( t ) , ( t , x ) ∈ ( 0 , ∞ ) × R N (1)
with nonnegative initial data u ( 0 , ⋅ ) = u 0 ≥ 0 u(0,\cdot)=u_0\geq0 u ( 0 , ⋅ ) = u 0 ≥ 0 , u 0 ≢ 0 u_0\not\equiv0 u 0 ≡ 0 , has attracted considerable interest since the pioneering work of Fujita [12 ] , which was motivated in part by the earlier eigenfunction-based blow-up criterion of Kaplan [16 ] . For the autonomous classical case (h = 1 h=1 h = 1 , α = 1 \alpha=1 α = 1 , θ = 2 \theta=2 θ = 2 ), Fujita identified the remarkable threshold
p F = 1 + 2 N . p_F=1+\frac{2}{N}. p F = 1 + N 2 .
If 1 < p < p F 1<p<p_F 1 < p < p F , every nonnegative nontrivial solution blows up in finite time. If p > p F p>p_F p > p F , global positive solutions exist for sufficiently small initial data. The borderline case p = p F p=p_F p = p F belongs to the blow-up regime. This was shown by Hayakawa [14 ] for N = 1 , 2 N=1,2 N = 1 , 2 and by Kobayashi, Sirao and Tanaka [21 ] for general N N N , and Weissler [40 ] later gave a much shorter proof valid in any dimension; see also Key words and phrases. Time-space fractional heat equation; Caputo derivative; fractional Laplacian; Fujita critical exponent; slowly varying function; Karamata theory; finite-time blow-up; global existence; test-function method; mild solution. Aronson and Weinberger [2 ] . This phenomenon, now called the Fujita phenomenon , is robust under many perturbations. It has been extended to systems, to nonlocal diffusion and to various geometries; we refer to the surveys [24 , 8 ] and to the monograph [36 ] for an overview.
A natural and physically relevant generalisation concerns time-dependent source intensities. When h ( t ) h(t) h ( t ) behaves like a power t q t^q t q , or more generally like a regularly varying function, the critical exponent shifts in a predictable way. For the heat equation on a domain D ⊂ R N D \subset\mathbb{R}^N D ⊂ R N with Dirichlet boundary conditions, this question was treated in great generality by Meier [31 ] , the case D = R N D = \mathbb{R}^N D = R N being included; see also his earlier works [29 , 30 ] . Coefficients depending on both t t t and x x x were considered, for instance, by Qi [35 ] and Igarashi [15 ] , while Pinsky [33 ] analysed the spatially inhomogeneous problem u t = Δ u + a ( x ) u p u_t = \Delta u + a(x)u^p u t = Δ u + a ( x ) u p in R N \mathbb{R}^N R N . For the nonlocal equation u t + ( − Δ ) θ / 2 u = h ( t ) u p u_t + (-\Delta)^{\theta/2}u = h(t)u^p u t + ( − Δ ) θ /2 u = h ( t ) u p , nonexistence of global solutions was established by Guedda and Kirane [13 ] , extending the work of Sugitani [38 ] . Meier’s fundamental observation is that the critical exponent admits the universal formula
p ∗ = 1 + q + 1 s ∗ , p^* = 1 + \frac{q+1}{s^*}, p ∗ = 1 + s ∗ q + 1 ,
s ∗ : = sup { s > 0 : ∃ w ≢ 0 : ∂ t w = Δ w with lim sup t → ∞ t s ∥ w ( t ) ∥ ∞ < ∞ } , (2) s^* := \sup\{s > 0 : \exists\, w \not\equiv0 : \partial_t w = \Delta w \text{ with } \limsup_{t\to\infty} t^s\lVert w(t)\rVert_\infty< \infty\}, \tag*{(2)} s ∗ := sup { s > 0 : ∃ w ≡ 0 : ∂ t w = Δ w with t → ∞ lim sup t s ∥ w ( t ) ∥ ∞ < ∞ } , (2)
where the supremum runs over nonnegative solutions of the homogeneous problem. For D = R N D = \mathbb{R}^N D = R N one has s ∗ = N / 2 s^* = N/2 s ∗ = N /2 , so that p ∗ = 1 + 2 ( q + 1 ) / N p^* = 1 + 2(q+1)/N p ∗ = 1 + 2 ( q + 1 ) / N , which reduces to Fujita’s value p F p_F p F when q = 0 q = 0 q = 0 . On conical or sectorial domains, s ∗ s^* s ∗ encodes the geometry through the first Dirichlet eigenvalue of the Laplace–Beltrami operator on the spherical cap; see [31 , 3 , 25 ] .
A closely related line of research concerns additive forcing terms, as in u t − Δ u = ∣ u ∣ p + t σ w ( x ) u_t - \Delta u = |u|^p + t^\sigma w(x) u t − Δ u = ∣ u ∣ p + t σ w ( x ) , for which the Fujita exponent is again shifted by the forcing; see Bandle, Levine and Zhang [4 ] , the second author’s works [26 , 1 ] on Hardy–Hénon and forced parabolic equations, and [5 ] , where a Caputo time derivative and a fractional Laplacian are combined with a forcing term t σ w ( x ) t^\sigma w(x) t σ w ( x ) . Sharp lifespan estimates for fractional Hardy–Hénon parabolic equations are obtained in [27 ] . The present work complements these papers: the time dependence now enters multiplicatively through the intensity h ( t ) h(t) h ( t ) of the nonlinear source.
The purpose of the present paper is twofold.
(a) We lift the Meier framework to the nonlocal setting. The Laplacian is replaced by the fractional Laplacian ( − Δ ) θ / 2 (-\Delta)^{\theta/2} ( − Δ ) θ /2 of order θ ∈ ( 0 , 2 ) \theta\in(0,2) θ ∈ ( 0 , 2 ) and, more delicately, the time derivative is replaced by the Caputo derivative ∂ t α \partial_t^\alpha ∂ t α of order α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . For the space-fractional equation with classical time derivative (α = 1 \alpha= 1 α = 1 , θ ∈ ( 0 , 2 ) \theta\in(0,2) θ ∈ ( 0 , 2 ) ), the autonomous Fujita threshold is p F = 1 + θ / N p_F = 1 + \theta/N p F = 1 + θ / N , and blow-up holds up to and including the borderline; this goes back to Sugitani [38 ] . For the time-fractional equation (α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) ), the autonomous threshold is again p F = 1 + θ / N p_F = 1 + \theta/N p F = 1 + θ / N , independently of α \alpha α . Now, however, the borderline p = p F p = p_F p = p F admits global solutions; see Zhang and Sun [42 ] for θ = 2 \theta= 2 θ = 2 , and Cortázar, Quirós and Wolanski [7 ] for the fully nonlocal case. The memory of the Caputo derivative thus affects the role of α \alpha α in a subtle way. It does not shift the autonomous exponent but, as we shall see, it does shift the exponent of a time-dependent source, replacing Meier’s numerator γ + 1 \gamma+ 1 γ + 1 by γ + α \gamma+ \alpha γ + α . Heuristically, this is the trace of the kernel ( t − s ) α − 1 (t-s)^{\alpha-1} ( t − s ) α − 1 in the Duhamel representation; see §5.6. Moreover, when N > θ N > \theta N > θ , a second and purely local mechanism appears. It is caused by the slow decay t − α t^{-\alpha} t − α of the fundamental solution near the origin and produces blow-up for p < 1 + γ / α p < 1 + \gamma/\alpha p < 1 + γ / α , a range that may exceed p c ( γ , α ) p_c(\gamma,\alpha) p c ( γ , α ) ; see Proposition 3.1.
(b) We incorporate a slowly varying prefactor ℓ \ell ℓ in the source, that is, a measurable function ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) \ell: (0,\infty) \to(0,\infty) ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) such that ℓ ( λ t ) / ℓ ( t ) → 1 \ell(\lambda t)/\ell(t) \to1 ℓ ( λ t ) / ℓ ( t ) → 1 as t → ∞ t \to\infty t → ∞ for every λ > 0 \lambda> 0 λ > 0 . Such functions appear naturally in physics whenever scaling symmetries are only approximate, or when logarithmic corrections arise from boundary layers, mean-field limits or anomalous transport in disordered
media [6 , 37 ] . They strictly include the positive constants, and also logarithms, iterated logarithms and powers of logarithms. Our analysis shows that the slowly varying factor never shifts the value of the critical exponent. At α = 1 \alpha= 1 α = 1 , however, it plays a decisive role at the critical exponent p = p c ( γ ) p = p_c(\gamma) p = p c ( γ ) , for every γ > − 1 \gamma> -1 γ > − 1 : the dichotomy between blow-up and small-data global existence is determined by the convergence or divergence of ∫ ∞ t − 1 ℓ ( t ) d t \int^\infty t^{-1}\ell(t)\,\mathrm{d}t ∫ ∞ t − 1 ℓ ( t ) d t . For α ∈ ( 0 , 1 ) \alpha\in(0, 1) α ∈ ( 0 , 1 ) and N < θ N < \theta N < θ , the same dichotomy is instead governed by the boundedness of ℓ \ell ℓ . Throughout, the source h h h is normalised near the origin so as to be continuous up to t = 0 t = 0 t = 0 (Convention 2.1). This is immaterial for every critical exponent below, but it is needed for the Cauchy problem to be well posed. Indeed, the literal density t ↦ t γ ℓ ( t ) t \mapsto t^\gamma\ell(t) t ↦ t γ ℓ ( t ) need not be integrable at the origin (with ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 , it is not integrable as soon as γ ≤ − 1 \gamma\leq-1 γ ≤ − 1 ), and slow variation constrains only the regime t → ∞ t \to\infty t → ∞ .
Main results (informal). For α = 1 \alpha= 1 α = 1 and θ ∈ ( 0 , 2 ] \theta\in(0, 2] θ ∈ ( 0 , 2 ] we obtain the complete classification
p c ( γ ) = { ∞ , if γ ≤ − 1 , 1 + θ ( γ + 1 ) N , if γ > − 1. p_c(\gamma) =
\begin{cases}
\infty, & \text{if } \gamma\leq-1,\\
1 + \dfrac{\theta(\gamma+ 1)}{N}, & \text{if } \gamma> -1.
\end{cases} p c ( γ ) = ⎩ ⎨ ⎧ ∞ , 1 + N θ ( γ + 1 ) , if γ ≤ − 1 , if γ > − 1.
Moreover, for every γ > − 1 \gamma> -1 γ > − 1 , the behavior at p = p c ( γ ) p = p_c(\gamma) p = p c ( γ ) is governed by the integrability of t ↦ t − 1 ℓ ( t ) t \mapsto t^{-1}\ell(t) t ↦ t − 1 ℓ ( t ) at infinity. Divergence forces blow-up of every nontrivial solution, whereas convergence allows small-data global existence. For α ∈ ( 0 , 1 ) \alpha\in(0, 1) α ∈ ( 0 , 1 ) we prove small-data global existence for every p > 1 p > 1 p > 1 when γ ≤ − α \gamma\leq-\alpha γ ≤ − α . When γ > − α \gamma> -\alpha γ > − α , we prove blow-up of every nontrivial solution for 1 < p < p c ( γ , α ) : = 1 + θ ( γ + α ) / ( N α ) 1 < p < p_c(\gamma,\alpha) := 1 + \theta(\gamma+\alpha)/(N\alpha) 1 < p < p c ( γ , α ) := 1 + θ ( γ + α ) / ( N α ) in every dimension, and also at p = p c ( γ , α ) p = p_c(\gamma,\alpha) p = p c ( γ , α ) when ℓ \ell ℓ is unbounded. When N < θ N < \theta N < θ , i.e. N = 1 N = 1 N = 1 and θ ∈ ( 1 , 2 ] \theta\in(1, 2] θ ∈ ( 1 , 2 ] , small global solutions exist for p > p c ( γ , α ) p > p_c(\gamma,\alpha) p > p c ( γ , α ) , and also at p = p c ( γ , α ) p = p_c(\gamma,\alpha) p = p c ( γ , α ) when ℓ \ell ℓ is bounded; the classification is therefore complete. When N ≥ θ N \geq\theta N ≥ θ , we reach the threshold 1 + ( γ + α ) / α 1 + (\gamma+\alpha)/\alpha 1 + ( γ + α ) / α , and we show that, for N > θ N > \theta N > θ , blow-up also occurs for 1 < p < 1 + γ / α 1 < p < 1 + \gamma/\alpha 1 < p < 1 + γ / α ; we conjecture that the Fujita exponent is then max { p c ( γ , α ) , 1 + γ / α } \max\{p_c(\gamma,\alpha), 1 + \gamma/\alpha\} max { p c ( γ , α ) , 1 + γ / α } (Conjecture 3.1). Note that p c ( γ , 1 ) = p c ( γ ) p_c(\gamma,1) = p_c(\gamma) p c ( γ , 1 ) = p c ( γ ) and p c ( 0 , α ) = 1 + θ / N p_c(0,\alpha) = 1 + \theta/N p c ( 0 , α ) = 1 + θ / N for every α \alpha α , in agreement with [42 , 7 ] .
The paper is organised as follows. Section 2 collects the necessary definitions and classical results: slowly varying functions (§ 2.1 ), the Caputo derivative and the fractional Laplacian (§§§ 2.2 –§ 2.3 ), the time–space fractional solution operators and their decay (§ 2.4 ), mild solutions and the standing convention on the source (§ 2.5 ), the classical Fujita theory (§ 2.6 ) and Meier’s theorem (§2.7). Section 3 contains the precise statements of the main results for α = 1 \alpha= 1 α = 1 and for α < 1 \alpha< 1 α < 1 , the accompanying conjecture, and a discussion of their consequences. Section 4 surveys related classes of h ( t ) h(t) h ( t ) (exponential, integrable, oscillatory) and their Fujita exponents. Section 5 contains the proofs: a fixed-point argument in weighted spaces for global existence, a Kaplan-type functional for blow-up at α = 1 \alpha= 1 α = 1 , a nonlinear iteration of Cortázar–Quiros–Wolanski type for blow-up when α < 1 \alpha< 1 α < 1 , and a local iteration exhibiting the second blow-up mechanism. We close with a few open problems in Section 6.
Preliminaries
Slowly varying functions
Definition 2.1. A measurable function ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) \ell: (0,\infty) \to(0,\infty) ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) is called slowly varying at infinity if
lim t → ∞ ℓ ( λ t ) ℓ ( t ) = 1 for every λ > 0. (3) \lim_{t \to\infty} \frac{\ell(\lambda t)}{\ell(t)} = 1 \qquad\text{for every } \lambda> 0. \tag*{(3)} t → ∞ lim ℓ ( t ) ℓ ( λ t ) = 1 for every λ > 0. (3)
A measurable function g : ( 0 , ∞ ) → ( 0 , ∞ ) g : (0,\infty) \to(0,\infty) g : ( 0 , ∞ ) → ( 0 , ∞ ) is regularly varying of index ρ ∈ R \rho\in\mathbb{R} ρ ∈ R if g ( t ) = t ρ ℓ ( t ) g(t) = t^\rho\ell(t) g ( t ) = t ρ ℓ ( t ) for some slowly varying ℓ \ell ℓ .
Typical examples are
ℓ ( t ) = ( log ( 1 + t ) ) a \ell(t) = (\log(1+t))^{a} ℓ ( t ) = ( log ( 1 + t ) ) a , ℓ ( t ) = ( log log ( 2 + t ) ) b \ell(t) = (\log\log(2+t))^{b} ℓ ( t ) = ( log log ( 2 + t ) ) b , ℓ ( t ) = exp ( ( log t ) δ ) \ell(t) = \exp((\log t)^{\delta}) ℓ ( t ) = exp (( log t ) δ ) (0 < δ < 1 0 < \delta< 1 0 < δ < 1 , t ≥ 1 t \ge1 t ≥ 1 ),
as well as products and quotients of such functions. Any function that converges to a positive constant at infinity is trivially slowly varying.
By Karamata’s representation theorem [6 Theorem 1.3.1] , a slowly varying function can be written as
ℓ ( t ) = c ( t ) exp ( ∫ a t ε ( s ) s d s ) , t ≥ a , (4) \ell(t) = c(t) \exp\left(\int_a^t \frac{\varepsilon(s)}{s}\,ds\right), \qquad t \ge a, \tag*{(4)} ℓ ( t ) = c ( t ) exp ( ∫ a t s ε ( s ) d s ) , t ≥ a , (4)
for some a > 0 a > 0 a > 0 , where c ( t ) → c 0 ∈ ( 0 , ∞ ) c(t) \to c_0 \in(0,\infty) c ( t ) → c 0 ∈ ( 0 , ∞ ) and ε ( t ) → 0 \varepsilon(t) \to0 ε ( t ) → 0 as t → ∞ t \to\infty t → ∞ . We shall repeatedly use the following four consequences.
Lemma 2.1. Let ℓ \ell ℓ be slowly varying at infinity and locally bounded on [ a , ∞ ) [a,\infty) [ a , ∞ ) for some a > 0 a > 0 a > 0 (this holds if ℓ \ell ℓ is continuous). Then the following properties hold.
(K1) (Sub-polynomial growth and decay.) For every ε > 0 \varepsilon> 0 ε > 0 , t − ε ℓ ( t ) → 0 t^{-\varepsilon}\ell(t) \to0 t − ε ℓ ( t ) → 0 and t ε ℓ ( t ) → ∞ t^{\varepsilon}\ell(t) \to\infty t ε ℓ ( t ) → ∞ as t → ∞ t \to\infty t → ∞ [6 Proposition 1.3.6] .
*(K2) (Karamata’s integral theorem.) For σ > − 1 \sigma> -1 σ > − 1 ,
∫ a t s σ ℓ ( s ) d s ∼ t σ + 1 ℓ ( t ) σ + 1 \int_a^t s^\sigma\ell(s)\,ds \sim\frac{t^{\sigma+1}\ell(t)}{\sigma+1} ∫ a t s σ ℓ ( s ) d s ∼ σ + 1 t σ + 1 ℓ ( t ) as t → ∞ t \to\infty t → ∞ .
For σ < − 1 \sigma< -1 σ < − 1 , we have ∫ a ∞ s σ ℓ ( s ) d s < ∞ \int_a^\infty s^\sigma\ell(s)\,ds < \infty ∫ a ∞ s σ ℓ ( s ) d s < ∞ and
∫ t ∞ s σ ℓ ( s ) d s ∼ t σ + 1 ℓ ( t ) − ( σ + 1 ) \int_t^\infty s^\sigma\ell(s)\,ds \sim\frac{t^{\sigma+1}\ell(t)}{-(\sigma+1)} ∫ t ∞ s σ ℓ ( s ) d s ∼ − ( σ + 1 ) t σ + 1 ℓ ( t ) [6 Propositions 1.5.8 and 1.5.10] .
Both statements involve a fixed lower limit a > 0 a > 0 a > 0 . They describe the tail t → ∞ t \to\infty t → ∞ and say nothing about integrability at the origin.
(K3) (Borderline case σ = − 1 \sigma=-1 σ = − 1 .) The integral ∫ a ∞ s − 1 ℓ ( s ) d s \int_a^\infty s^{-1}\ell(s)\,ds ∫ a ∞ s − 1 ℓ ( s ) d s may converge or diverge, depending on ℓ \ell ℓ . For example, ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 gives divergence, ℓ ( t ) = ( log t ) − 2 \ell(t)=(\log t)^{-2} ℓ ( t ) = ( log t ) − 2 gives convergence, and ℓ ( t ) = ( log t ) − 1 \ell(t)=(\log t)^{-1} ℓ ( t ) = ( log t ) − 1 gives divergence (like log log t \log\log t log log t ).
(K4) (Decay under a convergent borderline integral.) If ∫ a ∞ s − 1 ℓ ( s ) d s < ∞ \int_a^\infty s^{-1}\ell(s)\,ds < \infty ∫ a ∞ s − 1 ℓ ( s ) d s < ∞ , then ℓ ( t ) → 0 \ell(t) \to0 ℓ ( t ) → 0 as t → ∞ t \to\infty t → ∞ . Indeed, by the Uniform Convergence Theorem [6 Theorem 1.2.1] , ℓ ( s ) ≥ ( 1 + o ( 1 ) ) ℓ ( t ) \ell(s) \ge(1+o(1))\ell(t) ℓ ( s ) ≥ ( 1 + o ( 1 )) ℓ ( t ) uniformly for s ∈ [ t , 2 t ] s \in[t,2t] s ∈ [ t , 2 t ] . We deduce that ∫ t 2 t s − 1 ℓ ( s ) d s ≥ ( log 2 ) ( 1 + o ( 1 ) ) ℓ ( t ) \int_t^{2t} s^{-1}\ell(s)\,ds \ge(\log2)(1+o(1))\ell(t) ∫ t 2 t s − 1 ℓ ( s ) d s ≥ ( log 2 ) ( 1 + o ( 1 )) ℓ ( t ) , and the left-hand side tends to 0 0 0 as t → ∞ t \to\infty t → ∞ since the tail integral converges.
We shall also use the Uniform Convergence Theorem in the following form: for every compact interval [ λ 1 , λ 2 ] ⊂ ( 0 , ∞ ) [\lambda_1,\lambda_2] \subset(0,\infty) [ λ 1 , λ 2 ] ⊂ ( 0 , ∞ ) ,
sup λ ∈ [ λ 1 , λ 2 ] ∣ ℓ ( λ t ) ℓ ( t ) − 1 ∣ ⟶ 0 ( t → ∞ ) , (5) \sup_{\lambda\in[\lambda_1,\lambda_2]} \left|\frac{\ell(\lambda t)}{\ell(t)}-1\right| \longrightarrow0 \qquad(t\to\infty), \tag*{(5)} λ ∈ [ λ 1 , λ 2 ] sup ℓ ( t ) ℓ ( λ t ) − 1 ⟶ 0 ( t → ∞ ) , (5)
see [6 Theorem 1.2.1] .
Caputo fractional derivative
Definition 2.2. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . For an absolutely continuous function f : [ 0 , T ] → R f:[0,T]\to\mathbb{R} f : [ 0 , T ] → R , the Caputo derivative of order α \alpha α is defined by
∂ t α f ( t ) = 1 Γ ( 1 − α ) ∫ 0 t ( t − s ) − α f ′ ( s ) d s , t > 0. (6) \partial_t^\alpha f(t) = \frac{1}{\Gamma(1-\alpha)}\int_0^t (t-s)^{-\alpha}f'(s)\,ds, \qquad t>0. \tag*{(6)} ∂ t α f ( t ) = Γ ( 1 − α ) 1 ∫ 0 t ( t − s ) − α f ′ ( s ) d s , t > 0. (6)
For α = 1 \alpha=1 α = 1 we set ∂ t 1 f = f ′ \partial_t^1 f=f' ∂ t 1 f = f ′ in the classical sense.
The Caputo derivative is the natural notion of fractional time derivative for Cauchy problems with classical initial conditions. Since ∂ t α c = 0 \partial_t^\alpha c=0 ∂ t α c = 0 for every constant c c c , the initial value is consistent with the equation. Equivalently, on absolutely continuous functions, ∂ t α f ( t ) = D α [ f ( t ) − f ( 0 ) ] \partial_t^\alpha f(t)=D^\alpha[f(t)-f(0)] ∂ t α f ( t ) = D α [ f ( t ) − f ( 0 )] , where D α D^\alpha D α is the Riemann–Liouville derivative; see [18 , 34 ] .
Fractional Laplacian
For θ ∈ ( 0 , 2 ) \theta\in(0,2) θ ∈ ( 0 , 2 ) , the fractional Laplacian ( − Δ ) θ / 2 (-\Delta)^{\theta/2} ( − Δ ) θ /2 acts on Schwartz functions φ ∈ S ( R N ) \varphi\in\mathcal{S}(\mathbb{R}^{N}) φ ∈ S ( R N ) by
( − Δ ) θ / 2 φ ( ξ ) = ∣ ξ ∣ θ φ ^ ( ξ ) , (7) (-\Delta)^{\theta/2}\varphi(\xi) = |\xi|^{\theta}\widehat{\varphi}(\xi), \tag*{(7)} ( − Δ ) θ /2 φ ( ξ ) = ∣ ξ ∣ θ φ ( ξ ) , (7)
or, equivalently, by the singular integral
( − Δ ) θ / 2 φ ( x ) = C N , θ P.V. ∫ R N φ ( x ) − φ ( y ) ∣ x − y ∣ N + θ d y , (8) (-\Delta)^{\theta/2}\varphi(x) = C_{N,\theta}\operatorname{P.V.}\int_{\mathbb{R}^{N}}\frac{\varphi(x)-\varphi(y)}{|x-y|^{N+\theta}}\,\mathrm{d}y, \tag*{(8)} ( − Δ ) θ /2 φ ( x ) = C N , θ P.V. ∫ R N ∣ x − y ∣ N + θ φ ( x ) − φ ( y ) d y , (8)
with the explicit constant C N , θ = θ 2 θ − 1 Γ ( ( N + θ ) / 2 ) / ( π N / 2 Γ ( 1 − θ / 2 ) ) C_{N,\theta} = \theta2^{\theta-1}\Gamma((N+\theta)/2)/(\pi^{N/2}\Gamma(1-\theta/2)) C N , θ = θ 2 θ − 1 Γ (( N + θ ) /2 ) / ( π N /2 Γ ( 1 − θ /2 )) ; see [9 , 10 , 22 ] . The operator extends to a nonnegative self-adjoint operator on L 2 ( R N ) L^{2}(\mathbb{R}^{N}) L 2 ( R N ) with form domain H θ / 2 ( R N ) H^{\theta/2}(\mathbb{R}^{N}) H θ /2 ( R N ) . For θ = 2 \theta= 2 θ = 2 we recover − Δ -\Delta − Δ .
The time–space fractional solution operators and kernel estimates
Consider the homogeneous Cauchy problem
∂ t α w + ( − Δ ) θ / 2 w = 0 , w ( 0 , ⋅ ) = w 0 . (9) \partial_{t}^{\alpha}w+(-\Delta)^{\theta/2}w=0,\qquad w(0,\cdot)=w_{0}. \tag*{(9)} ∂ t α w + ( − Δ ) θ /2 w = 0 , w ( 0 , ⋅ ) = w 0 . (9)
Using the Fourier transform in x x x and the Laplace transform in t t t , we obtain w ( t , x ) = [ S α , θ ( t ) w 0 ] ( x ) w(t,x)=[S_{\alpha,\theta}(t)w_{0}](x) w ( t , x ) = [ S α , θ ( t ) w 0 ] ( x ) , where
S α , θ ( t ) w 0 ( x ) = ∫ R N Z ( t , x − y ) w 0 ( y ) d y , (10) S_{\alpha,\theta}(t)w_{0}(x)=\int_{\mathbb{R}^{N}}Z(t,x-y)w_{0}(y)\,\mathrm{d}y, \tag*{(10)} S α , θ ( t ) w 0 ( x ) = ∫ R N Z ( t , x − y ) w 0 ( y ) d y , (10)
and the fundamental solution Z Z Z is characterised by Z ^ ( t , ξ ) = E α ( − t α ∣ ξ ∣ θ ) \widehat{Z}(t,\xi)=E_{\alpha}(-t^{\alpha}|\xi|^{\theta}) Z ( t , ξ ) = E α ( − t α ∣ ξ ∣ θ ) . Here E α E_{\alpha} E α is the Mittag-Leffler function E α ( z ) = ∑ k = 0 ∞ z k Γ ( α k + 1 ) E_{\alpha}(z)=\sum_{k=0}^{\infty}\frac{z^{k}}{\Gamma(\alpha k+1)} E α ( z ) = ∑ k = 0 ∞ Γ ( α k + 1 ) z k ; see [28 , 19 , 17 ] . For every t > 0 t>0 t > 0 the kernel Z ( t , ⋅ ) Z(t,\cdot) Z ( t , ⋅ ) is a positive probability density, so that ∫ R N Z ( t , x ) d x = 1 \int_{\mathbb{R}^{N}}Z(t,x)\,\mathrm{d}x=1 ∫ R N Z ( t , x ) d x = 1 . Since Z ^ ( t , ξ ) \widehat{Z}(t,\xi) Z ( t , ξ ) depends only on t α ∣ ξ ∣ θ t^{\alpha}|\xi|^{\theta} t α ∣ ξ ∣ θ , the kernel is self-similar:
Z ( t , x ) = t − N α / θ F ( x t − α / θ ) , ( t , x ) ∈ ( 0 , ∞ ) × R N , (11) Z(t,x)=t^{-N\alpha/\theta}F(xt^{-\alpha/\theta}),\qquad(t,x)\in(0,\infty)\times\mathbb{R}^{N}, \tag*{(11)} Z ( t , x ) = t − N α / θ F ( x t − α / θ ) , ( t , x ) ∈ ( 0 , ∞ ) × R N , (11)
where the profile F = Z ( 1 , ⋅ ) F=Z(1,\cdot) F = Z ( 1 , ⋅ ) is radial, positive and continuous on R N ∖ { 0 } \mathbb{R}^{N}\setminus\{0\} R N ∖ { 0 } . Writing f ≍ g f\asymp g f ≍ g when c − 1 g ≤ f ≤ c g c^{-1}g\leq f\leq cg c − 1 g ≤ f ≤ c g for some constant c ≥ 1 c\geq1 c ≥ 1 , the two-sided estimates of [17 Lemma 3.3] give, for 0 < ∣ y ∣ ≤ 1 0<|y|\leq1 0 < ∣ y ∣ ≤ 1 ,
F ( y ) ≍ { ∣ y ∣ − N + θ if N > θ and 0 < α < 1 , 1 + ∣ log ( ∣ y ∣ θ ) ∣ if N = θ and 0 < α < 1 , 1 if α = 1 , or if N < θ and 0 < α < 1. F(y)\asymp
\begin{cases}
|y|^{-N+\theta} & \text{if }N>\theta\text{ and }0<\alpha<1,\\
1+|\log(|y|^{\theta})| & \text{if }N=\theta\text{ and }0<\alpha<1,\\
1 & \text{if }\alpha=1,\text{ or if }N<\theta\text{ and }0<\alpha<1.
\end{cases} F ( y ) ≍ ⎩ ⎨ ⎧ ∣ y ∣ − N + θ 1 + ∣ log ( ∣ y ∣ θ ) ∣ 1 if N > θ and 0 < α < 1 , if N = θ and 0 < α < 1 , if α = 1 , or if N < θ and 0 < α < 1.
In particular, since F F F is bounded from below by a positive constant on the unit ball in all three cases,
Z ( t , x ) ≥ C t − N α / θ , if ∣ x ∣ ≤ t α / θ , (12) Z(t,x)\geq Ct^{-N\alpha/\theta},\qquad\text{if }|x|\leq t^{\alpha/\theta}, \tag*{(12)} Z ( t , x ) ≥ C t − N α / θ , if ∣ x ∣ ≤ t α / θ , (12)
and, if N > θ N>\theta N > θ and 0 < α < 1 0<\alpha<1 0 < α < 1 ,
Z ( t , x ) ≥ C t − α ∣ x ∣ θ − N , if 0 < ∣ x ∣ ≤ t α / θ . (13) Z(t,x)\geq Ct^{-\alpha}|x|^{\theta-N},\qquad\text{if }0<|x|\leq t^{\alpha/\theta}. \tag*{(13)} Z ( t , x ) ≥ C t − α ∣ x ∣ θ − N , if 0 < ∣ x ∣ ≤ t α / θ . (13)
For α = 1 \alpha=1 α = 1 and θ = 2 \theta=2 θ = 2 we recover the Gaussian kernel, and for α = 1 \alpha=1 α = 1 , θ ∈ ( 0 , 2 ) \theta\in(0,2) θ ∈ ( 0 , 2 ) , the θ \theta θ -stable kernel. The validity of L p L^{p} L p –L q L^{q} L q smoothing estimates depends crucially on α \alpha α . The following statement records the sharp ranges; see [19 , 17 , 7 ] .
Proposition 2.1. Let θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] and 1 ≤ p ≤ q ≤ ∞ 1\leq p\leq q\leq\infty 1 ≤ p ≤ q ≤ ∞ .
(i) If α = 1 \alpha=1 α = 1 , then for all ϕ ∈ L p ( R N ) \phi\in L^{p}(\mathbb{R}^{N}) ϕ ∈ L p ( R N ) ,
∥ S α , θ ( t ) ϕ ∥ L q ( R N ) ≤ C t − N α / θ ( 1 / p − 1 / q ) ∥ ϕ ∥ L p ( R N ) , t > 0 , (14) \|S_{\alpha,\theta}(t)\phi\|_{L^{q}(\mathbb{R}^{N})}\leq Ct^{-N\alpha/\theta(1/p-1/q)}\|\phi\|_{L^{p}(\mathbb{R}^{N})},\qquad t>0, \tag*{(14)} ∥ S α , θ ( t ) ϕ ∥ L q ( R N ) ≤ C t − N α / θ ( 1/ p − 1/ q ) ∥ ϕ ∥ L p ( R N ) , t > 0 , (14) with no restriction on ( p , q ) (p,q) ( p , q ) .
(ii) If α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , then (14) holds whenever 1 p − 1 q < θ N \frac{1}{p}-\frac{1}{q}<\frac{\theta}{N} p 1 − q 1 < N θ and fails whenever 1 p − 1 q > θ N \frac{1}{p}-\frac{1}{q}>\frac{\theta}{N} p 1 − q 1 > N θ .
When it is valid, the decay rate in (14) is sharp. In both cases S α , θ ( t ) S_{\alpha,\theta}(t) S α , θ ( t ) is a contraction on every L r ( R N ) L^r(\mathbb{R}^N) L r ( R N ) , 1 ≤ r ≤ ∞ 1 \le r \le\infty 1 ≤ r ≤ ∞ .
Part (ii) follows from Young’s inequality, since by (11) we have ∥ Z ( t , ⋅ ) ∥ L m = t − N α θ ( 1 − 1 m ) ∥ F ∥ L m \lVert Z(t,\cdot)\rVert_{L^m}=t^{-\frac{N\alpha}{\theta}(1-\frac{1}{m})}\lVert F\rVert_{L^m} ∥ Z ( t , ⋅ ) ∥ L m = t − θ N α ( 1 − m 1 ) ∥ F ∥ L m , and F ∈ L m ( R N ) F\in L^m(\mathbb{R}^N) F ∈ L m ( R N ) if and only if ( N − θ ) m < N (N-\theta)m<N ( N − θ ) m < N , that is, 1 − 1 m < θ N 1-\frac{1}{m}<\frac{\theta}{N} 1 − m 1 < N θ .
Remark 2.1.
For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , consider the critical line 1 p − 1 q = θ N \frac{1}{p}-\frac{1}{q}=\frac{\theta}{N} p 1 − q 1 = N θ , which is non-void only when N ≥ θ N\ge\theta N ≥ θ . There, the estimate (14) holds for 1 < p < q < ∞ 1<p<q<\infty 1 < p < q < ∞ . This is a consequence of the global pointwise bound Z ( t , x ) ≤ C t − α ∣ x ∣ θ − N Z(t,x)\le Ct^{-\alpha}|x|^{\theta-N} Z ( t , x ) ≤ C t − α ∣ x ∣ θ − N for N > θ N>\theta N > θ and of the Hardy–Littlewood–Sobolev inequality. It fails exactly at the endpoint pairs ( p , q ) = ( 1 , N N − θ ) (p,q)=(1,\frac{N}{N-\theta}) ( p , q ) = ( 1 , N − θ N ) and ( p , q ) = ( N θ , ∞ ) (p,q)=(\frac{N}{\theta},\infty) ( p , q ) = ( θ N , ∞ ) (at ( p , q ) = ( 1 , ∞ ) (p,q)=(1,\infty) ( p , q ) = ( 1 , ∞ ) when N = θ N=\theta N = θ ). In particular, (14) holds for all 1 ≤ p ≤ q ≤ ∞ 1\le p\le q\le\infty 1 ≤ p ≤ q ≤ ∞ when N < θ N<\theta N < θ , whereas the endpoint L 1 → L ∞ L^1\to L^\infty L 1 → L ∞ estimate fails whenever N ≥ θ N\ge\theta N ≥ θ .
The restriction in Proposition 2.1(ii) reflects a genuine spatial singularity of the fundamental solution. For α < 1 \alpha<1 α < 1 , the profile in (11) satisfies F ( ξ ) ≍ ∣ ξ ∣ θ − N F(\xi)\asymp|\xi|^{\theta-N} F ( ξ ) ≍ ∣ ξ ∣ θ − N as ξ → 0 \xi\to0 ξ → 0 when N > θ N>\theta N > θ , with a logarithmic singularity when N = θ N=\theta N = θ , so that Z ( t , ⋅ ) ∉ L ∞ ( R N ) Z(t,\cdot)\notin L^\infty(\mathbb{R}^N) Z ( t , ⋅ ) ∈ / L ∞ ( R N ) ; see [19 , 17 ] . Consequently, for α < 1 \alpha<1 α < 1 and N > θ N>\theta N > θ , the sup-norm of S α , θ ( t ) w 0 S_{\alpha,\theta}(t)w_0 S α , θ ( t ) w 0 with 0 ≤ w 0 ∈ L 1 ∩ L ∞ 0\le w_0\in L^1\cap L^\infty 0 ≤ w 0 ∈ L 1 ∩ L ∞ , w 0 ≢ 0 w_0\not\equiv0 w 0 ≡ 0 , decays exactly at the rate t − α t^{-\alpha} t − α , which is slower than the formal L 1 → L ∞ L^1\to L^\infty L 1 → L ∞ rate t − N α θ t^{-\frac{N\alpha}{\theta}} t − θ N α . This slow decay is the origin of the local blow-up mechanism of Proposition 3.1. The operator family P α , θ P_{\alpha,\theta} P α , θ of §2.5 obeys analogous estimates on its own admissible ranges, namely with θ \theta θ replaced by 2 θ 2\theta 2 θ in the admissibility condition, because its kernel behaves like ∣ x ∣ 2 θ − N |x|^{2\theta-N} ∣ x ∣ 2 θ − N near the origin when N > 2 θ N>2\theta N > 2 θ ; see [7 , 17 ] .
At α = 1 \alpha=1 α = 1 the exponent N / θ N/\theta N / θ plays the role of s ∗ s^* s ∗ in Meier’s formula (2) . For α < 1 \alpha<1 α < 1 the correct analogue is subtler; see §5.6.
Mild solutions and the standing convention on the source
Following the standard approach [20 , 41 ] , see also [42 , 7 ] , the Cauchy problem (1) is written in the integral form
u ( t , x ) = S α , θ ( t ) u 0 ( x ) + ∫ 0 t ( t − s ) α − 1 P α , θ ( t − s ) [ h ( s ) u ( s , ⋅ ) p ] ( x ) d s (15) u(t,x)=S_{\alpha,\theta}(t)u_0(x)+\int_0^t(t-s)^{\alpha-1}P_{\alpha,\theta}(t-s)\left[h(s)u(s,\cdot)^p\right](x)\,\mathrm{d}s \tag*{(15)} u ( t , x ) = S α , θ ( t ) u 0 ( x ) + ∫ 0 t ( t − s ) α − 1 P α , θ ( t − s ) [ h ( s ) u ( s , ⋅ ) p ] ( x ) d s (15)
where, for each t > 0 t>0 t > 0 , P α , θ ( t ) P_{\alpha,\theta}(t) P α , θ ( t ) is the operator with Fourier symbol P α , θ ( t ) f ^ ( ξ ) = α E α ′ ( − t α ∣ ξ ∣ θ ) f ^ ( ξ ) \widehat{P_{\alpha,\theta}(t)f}(\xi)=\alpha E'_\alpha(-t^\alpha|\xi|^\theta)\widehat{f}(\xi) P α , θ ( t ) f ( ξ ) = α E α ′ ( − t α ∣ ξ ∣ θ ) f ( ξ ) . Since α E α ′ = E α , α \alpha E'_\alpha=E_{\alpha,\alpha} α E α ′ = E α , α , where E α , α ( z ) = ∑ k ≥ 0 z k / Γ ( α k + α ) E_{\alpha,\alpha}(z)=\sum_{k\ge0}z^k/\Gamma(\alpha k+\alpha) E α , α ( z ) = ∑ k ≥ 0 z k /Γ ( α k + α ) , this symbol equals E α , α ( − t α ∣ ξ ∣ θ ) E_{\alpha,\alpha}(-t^\alpha|\xi|^\theta) E α , α ( − t α ∣ ξ ∣ θ ) . For α = 1 \alpha=1 α = 1 we have P 1 , θ = S 1 , θ P_{1,\theta}=S_{1,\theta} P 1 , θ = S 1 , θ , and (15) reduces to the standard Duhamel formula. The kernel of P α , θ ( t ) P_{\alpha,\theta}(t) P α , θ ( t ) is nonnegative with mass 1 / Γ ( α ) 1/\Gamma(\alpha) 1/Γ ( α ) . Hence P α , θ ( t ) P_{\alpha,\theta}(t) P α , θ ( t ) is bounded on every L r ( R N ) L^r(\mathbb{R}^N) L r ( R N ) , uniformly in t t t , and smoothing estimates of the form (14) hold on the corresponding admissible ranges; see [7 ] .
Definition 2.3. Let 0 ≤ u 0 ∈ L 1 ( R N ) ∩ L ∞ ( R N ) 0\le u_0\in L^1(\mathbb{R}^N)\cap L^\infty(\mathbb{R}^N) 0 ≤ u 0 ∈ L 1 ( R N ) ∩ L ∞ ( R N ) and T ∈ ( 0 , ∞ ] T\in(0,\infty] T ∈ ( 0 , ∞ ] . A nonnegative measurable function u u u with u ∈ L ∞ ( ( 0 , T ′ ) ; L 1 ( R N ) ∩ L ∞ ( R N ) ) u\in L^\infty((0,T');L^1(\mathbb{R}^N)\cap L^\infty(\mathbb{R}^N)) u ∈ L ∞ (( 0 , T ′ ) ; L 1 ( R N ) ∩ L ∞ ( R N )) for every T ′ < T T'<T T ′ < T is a mild solution of (1) on [ 0 , T ) [0,T) [ 0 , T ) if (15) holds for almost every ( t , x ) ∈ ( 0 , T ) × R N (t,x)\in(0,T)\times\mathbb{R}^N ( t , x ) ∈ ( 0 , T ) × R N . For q ∈ [ 1 , ∞ ] q\in[1,\infty] q ∈ [ 1 , ∞ ] we say, following [7 ] , that u u u is an L q L^q L q -solution if u ∈ L l o c ∞ ( [ 0 , T ) ; L q ( R N ) ) u\in L^\infty_{\mathrm{loc}}([0,T);L^q(\mathbb{R}^N)) u ∈ L loc ∞ ([ 0 , T ) ; L q ( R N )) and (15) holds in L q ( R N ) L^q(\mathbb{R}^N) L q ( R N ) for almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) . A mild solution in the above sense is both an L 1 L^1 L 1 - and an L ∞ L^\infty L ∞ -solution.
Convention 2.1. The function ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) \ell:(0,\infty)\to(0,\infty) ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) is continuous and slowly varying at infinity, and the source intensity h h h is understood as a large-time profile . Following Meier, we assume that
h ∈ C ( [ 0 , ∞ ) ) , h ( t ) > 0 for t > 0 , and h ( t ) = t ℓ ℓ ( t ) for t ≥ 1. h\in C([0,\infty)),\qquad h(t)>0\ \text{for }t>0,\qquad\text{and}\qquad h(t)=t^\ell\ell(t)\ \text{for }t\ge1. h ∈ C ([ 0 , ∞ )) , h ( t ) > 0 for t > 0 , and h ( t ) = t ℓ ℓ ( t ) for t ≥ 1.
The threshold 1 is arbitrary and can be replaced by any t 0 > 0 t_0>0 t 0 > 0 . In particular, h ∈ L l o c 1 ( [ 0 , ∞ ) ) h\in L^1_{\mathrm{loc}}([0,\infty)) h ∈ L loc 1 ([ 0 , ∞ )) . This normalisation is necessary for the Duhamel formulation (15) to make sense. Indeed, for u u u bounded near t = 0 t=0 t = 0 , the contribution of small times s s s to the nonlinear term behaves like t α − 1 ∥ u 0 ∥ ∞ p ∫ 0 t h ( s ) d s t^{\alpha-1}\lVert u_0\rVert_\infty^p\int_0^t h(s)\,\mathrm{d}s t α − 1 ∥ u 0 ∥ ∞ p ∫ 0 t h ( s ) d s . The literal density t ℓ ℓ ( t ) t^\ell\ell(t) t ℓ ℓ ( t ) , however, need not be integrable at the origin (for ℓ = 1 \ell=1 ℓ = 1 it is not integrable as soon as γ ≤ − 1 \gamma\le-1 γ ≤ − 1 , because slow variation constrains only the behavior as t → ∞ t \to\infty t → ∞ . The completion on [ 0 , 1 ] [0,1] [ 0 , 1 ] is immaterial: every Fujita exponent below is governed solely by the behavior of h h h as t → ∞ t \to\infty t → ∞ . For definiteness one may take h ( t ) = ( 1 + t ) γ ℓ ( 1 + t ) h(t)=(1+t)^\gamma\ell(1+t) h ( t ) = ( 1 + t ) γ ℓ ( 1 + t ) . This function is continuous and positive on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) and asymptotic to t γ ℓ ( t ) t^\gamma\ell(t) t γ ℓ ( t ) . Strictly speaking, it does not coincide with t γ ℓ ( t ) t^\gamma\ell(t) t γ ℓ ( t ) for t ≥ 1 t \ge1 t ≥ 1 , but it can be written as t γ ℓ ~ ( t ) t^\gamma\widetilde{\ell}(t) t γ ℓ ( t ) with ℓ ~ ( t ) : = ( 1 + 1 / t ) γ ℓ ( 1 + t ) \widetilde{\ell}(t):=(1+1/t)^\gamma\ell(1+t) ℓ ( t ) := ( 1 + 1/ t ) γ ℓ ( 1 + t ) , which is continuous, positive and slowly varying by (5) . Moreover, ℓ ~ / ℓ → 1 \widetilde{\ell}/\ell\to1 ℓ / ℓ → 1 , so ℓ ~ \widetilde{\ell} ℓ and ℓ \ell ℓ share all the properties used below (boundedness, unboundedness, and convergence of ∫ 0 ∞ t − γ ℓ ( t ) d t \int_0^\infty t^{-\gamma}\ell(t)\,dt ∫ 0 ∞ t − γ ℓ ( t ) d t ). In the subintegrable regime γ < − 1 \gamma<-1 γ < − 1 , this realisation satisfies ∫ 0 ∞ h ( t ) d t < ∞ \int_0^\infty h(t)\,dt<\infty ∫ 0 ∞ h ( t ) d t < ∞ .
Under Convention 2.1, the source is continuous on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . A standard contraction argument in L ∞ ( ( 0 , T ) ; L 1 ∩ L ∞ ) L^\infty((0,T);L^1\cap L^\infty) L ∞ (( 0 , T ) ; L 1 ∩ L ∞ ) then provides, for every 0 ≤ u 0 ∈ L 1 ∩ L ∞ 0\le u_0\in L^1\cap L^\infty 0 ≤ u 0 ∈ L 1 ∩ L ∞ , a unique maximal mild solution on [ 0 , T ∗ ) [0,T^*) [ 0 , T ∗ ) , which is nonnegative. If T ∗ < ∞ T^*<\infty T ∗ < ∞ , then lim sup t ↑ T ∗ ∥ u ( t ) ∥ L ∞ = ∞ \limsup_{t\uparrow T^*}\|u(t)\|_{L^\infty}=\infty lim sup t ↑ T ∗ ∥ u ( t ) ∥ L ∞ = ∞ ; see [39 ] for α = 1 \alpha=1 α = 1 and [42 , 7 ] for α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , whose arguments apply verbatim to a bounded continuous coefficient h h h on [ 0 , T ] [0,T] [ 0 , T ] .
Classical Fujita theory
Consider the autonomous Cauchy problem
∂ t u − Δ u = u p , u ( 0 , x ) = u 0 ( x ) ≥ 0 , x ∈ R N . (16) \partial_tu-\Delta u=u^p,\qquad u(0,x)=u_0(x)\ge0,\qquad x\in\mathbb{R}^N. \tag*{(16)} ∂ t u − Δ u = u p , u ( 0 , x ) = u 0 ( x ) ≥ 0 , x ∈ R N . (16)
Fujita [12 ] proved the following dichotomy.
If 1 < p < 1 + 2 / N 1<p<1+2/N 1 < p < 1 + 2/ N , every nonnegative nontrivial solution blows up in finite time.
If p > 1 + 2 / N p>1+2/N p > 1 + 2/ N , there exist small initial data for which the solution exists globally and decays to zero.
The borderline exponent p F = 1 + 2 / N p_F=1+2/N p F = 1 + 2/ N belongs to the blow-up case [14 , 21 , 40 ] .
For the autonomous nonlocal equation ∂ t α u + ( − Δ ) θ / 2 u = u p \partial_t^\alpha u+(-\Delta)^{\theta/2}u=u^p ∂ t α u + ( − Δ ) θ /2 u = u p in R N \mathbb{R}^N R N , the threshold is
p F = 1 + θ N , (17) p_F=1+\frac{\theta}{N}, \tag*{(17)} p F = 1 + N θ , (17)
independently of α ∈ ( 0 , 1 ] \alpha\in(0,1] α ∈ ( 0 , 1 ] . The status of the borderline, however, depends on α \alpha α .
For α = 1 \alpha=1 α = 1 and θ ∈ ( 0 , 2 ) \theta\in(0,2) θ ∈ ( 0 , 2 ) , Sugitani [38 ] proved blow-up of every nontrivial nonnegative solution for 1 < p ≤ 1 + θ / N 1<p\le1+\theta/N 1 < p ≤ 1 + θ / N , while small-data global solutions exist above the threshold.
For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , Zhang and Sun [42 ] (θ = 2 \theta=2 θ = 2 ) and Cortázar, Quirós and Wolanski [7 ] (θ ∈ ( 0 , 2 ) \theta\in(0,2) θ ∈ ( 0 , 2 ) ; the details are written there for N > 2 θ N>2\theta N > 2 θ ) proved blow-up for 1 < p < 1 + θ / N 1<p<1+\theta/N 1 < p < 1 + θ / N . In sharp contrast with the case α = 1 \alpha=1 α = 1 , they also proved the existence of global solutions at the borderline p = 1 + θ / N p=1+\theta/N p = 1 + θ / N for suitable small data.
Test-function (nonexistence) results for related problems with spatio-temporal fractional derivatives were obtained by Kirane, Laskri and Tatar [20 ] , using the method of Mitidieri and Pohozaev [32 ] . As is typical of pure scaling arguments, the ranges they obtain are not sharp for α < 1 \alpha<1 α < 1 (compare with §5.3).
Meier's theorem and its s ∗ s^* s ∗ -characterisation
We now state Meier’s results in a form suited to our adaptation. Throughout this subsection, α = 1 \alpha=1 α = 1 , θ = 2 \theta=2 θ = 2 , and h ∈ C ( [ 0 , ∞ ) ) h\in C([0,\infty)) h ∈ C ([ 0 , ∞ )) with h ≥ 0 h\ge0 h ≥ 0 .
Theorem 2.1 (Meier [31 Theorem 2] ). Consider the Cauchy problem ∂ t u − Δ u = h ( t ) u p \partial_tu-\Delta u=h(t)u^p ∂ t u − Δ u = h ( t ) u p in R N \mathbb{R}^N R N with u ( 0 , ⋅ ) = u 0 ≥ 0 u(0,\cdot)=u_0\ge0 u ( 0 , ⋅ ) = u 0 ≥ 0 .
(i) (Global existence.) If there exists a nonnegative solution w ≢ 0 w\not\equiv0 w ≡ 0 of the homogeneous heat equation such that
∫ 0 ∞ h ( t ) ∥ w ( t , ⋅ ) ∥ L ∞ ( R N ) p − 1 d t < ∞ , (18) \int_0^\infty h(t)\|w(t,\cdot)\|_{L^\infty(\mathbb{R}^N)}^{p-1}\,dt<\infty, \tag*{(18)} ∫ 0 ∞ h ( t ) ∥ w ( t , ⋅ ) ∥ L ∞ ( R N ) p − 1 d t < ∞ , (18) then the problem admits global positive solutions u ≤ C w u\le Cw u ≤ C w , with ∥ u ( t ) ∥ ∞ → 0 \|u(t)\|_\infty\to0 ∥ u ( t ) ∥ ∞ → 0 as t → ∞ t\to\infty t → ∞ whenever ∥ w ( t ) ∥ ∞ → 0 \|w(t)\|_\infty\to0 ∥ w ( t ) ∥ ∞ → 0 .
(ii) (Blow-up.) If, for every nonnegative nontrivial solution w w w of the homogeneous heat equation,
lim sup t → ∞ ∥ w ( t , ⋅ ) ∥ L ∞ ( R N ) p − 1 ∫ 0 t h ( s ) d s = ∞ , (19) \limsup_{t\to\infty}\|w(t,\cdot)\|_{L^\infty(\mathbb{R}^N)}^{p-1}\int_0^t h(s)\,ds=\infty, \tag*{(19)} t → ∞ lim sup ∥ w ( t , ⋅ ) ∥ L ∞ ( R N ) p − 1 ∫ 0 t h ( s ) d s = ∞ , (19)
then every nontrivial nonnegative solution blows up in finite time.
Theorem 2.2 (Meier’s universal formula [31 Theorem 1] ). Assume hypotheses ( h 1 ) (h1) ( h 1 ) –( h 2 ) (h2) ( h 2 ) of [31 ] , that is, α 0 t q ≤ h ( t ) ≤ α 1 t q \alpha_0t^q\le h(t)\le\alpha_1t^q α 0 t q ≤ h ( t ) ≤ α 1 t q for t t t large and some q > − 1 q>-1 q > − 1 . Then the critical exponent for the Cauchy problem ∂ t u − Δ u = h ( t ) u p \partial_tu-\Delta u=h(t)u^p ∂ t u − Δ u = h ( t ) u p is
p ∗ = 1 + q + 1 s ∗ , p^*=1+\frac{q+1}{s^*}, p ∗ = 1 + s ∗ q + 1 , where s ∗ s^* s ∗ is given by (2) . For D = R N D=\mathbb{R}^N D = R N , s ∗ = N / 2 s^*=N/2 s ∗ = N /2 , hence p ∗ = 1 + 2 ( q + 1 ) / N p^*=1+2(q+1)/N p ∗ = 1 + 2 ( q + 1 ) / N .
The proof of Theorem 2.1(i) relies on an explicit supersolution u ˉ ( x , t ) = a ( t ) w ( x , t ) \bar{u}(x,t)=a(t)w(x,t) u ˉ ( x , t ) = a ( t ) w ( x , t ) , where a a a solves the ODE
a ′ ( t ) = h ( t ) ∥ w ( t , ⋅ ) ∥ L ∞ p − 1 a ( t ) p , a ( 0 ) = a 0 > 0. a'(t)=h(t)\|w(t,\cdot)\|_{L^\infty}^{p-1}a(t)^p,\qquad a(0)=a_0>0. a ′ ( t ) = h ( t ) ∥ w ( t , ⋅ ) ∥ L ∞ p − 1 a ( t ) p , a ( 0 ) = a 0 > 0.
Indeed, since w p − 1 ≤ ∥ w ∥ ∞ p − 1 w^{p-1}\le\|w\|_\infty^{p-1} w p − 1 ≤ ∥ w ∥ ∞ p − 1 , we get u ˉ t − Δ u ˉ = a ′ w ≥ h a p w p = h u ˉ p \bar{u}_t-\Delta\bar{u}=a'w\ge ha^pw^p=h\bar{u}^p u ˉ t − Δ u ˉ = a ′ w ≥ h a p w p = h u ˉ p . Separating variables, we obtain a ( t ) 1 − p = a 0 1 − p − ( p − 1 ) ∫ 0 t h ( s ) ∥ w ( s ) ∥ ∞ p − 1 d s a(t)^{1-p}=a_0^{1-p}-(p-1)\int_0^t h(s)\|w(s)\|_\infty^{p-1}\,ds a ( t ) 1 − p = a 0 1 − p − ( p − 1 ) ∫ 0 t h ( s ) ∥ w ( s ) ∥ ∞ p − 1 d s . This quantity stays bounded away from zero for all t > 0 t>0 t > 0 , so that a a a remains bounded, provided a 0 a_0 a 0 is small enough compared with the finite integral (18) . The blow-up part is proved by a subsolution argument. One constructs u ( x , t ) = z ( v ( t ) ; w ( x , t ) ) u(x,t)=z(v(t);w(x,t)) u ( x , t ) = z ( v ( t ) ; w ( x , t )) , where v ( t ) = ∫ 0 t h ( s ) d s v(t)=\int_0^t h(s)\,ds v ( t ) = ∫ 0 t h ( s ) d s and z ( ⋅ ; ξ ) z(\cdot;\xi) z ( ⋅ ; ξ ) solves the autonomous ODE z ′ = z p z'=z^p z ′ = z p with z ( 0 ; ξ ) = ξ z(0;\xi)=\xi z ( 0 ; ξ ) = ξ . Since z z z is convex in ξ \xi ξ , a direct computation gives u t − Δ u = h u p − z ξ ξ ∣ ∇ w ∣ 2 ≤ h u p u_t-\Delta u=hu^p-z_{\xi\xi}|\nabla w|^2\le hu^p u t − Δ u = h u p − z ξ ξ ∣∇ w ∣ 2 ≤ h u p .
Remark 2.2. Comparison principles for mild solutions remain valid for the operator ∂ t α + ( − Δ ) θ / 2 \partial_t^\alpha+(-\Delta)^{\theta/2} ∂ t α + ( − Δ ) θ /2 , because all the kernels involved are nonnegative (see e.g. [41 , 18 , 7 ] ). What fails for α < 1 \alpha<1 α < 1 is the product ansatz . The Caputo derivative does not satisfy a Leibniz rule, so the supersolution u ˉ = a ( t ) w ( x , t ) \bar{u}=a(t)w(x,t) u ˉ = a ( t ) w ( x , t ) is no longer compatible with the nonlocal time operator. The proofs must therefore be modified in the fractional setting. Global existence is established through fixed-point arguments for mild solutions, and blow-up through Kaplan-type functionals (α = 1 \alpha=1 α = 1 ) and nonlinear iterations on suitable weighted averages (α < 1 \alpha<1 α < 1 ); see § 5 . The fate of Meier’s universal formula under this modification is discussed in §5.6. At α = 1 \alpha=1 α = 1 it persists verbatim with s ∗ = N / θ s^*=N/\theta s ∗ = N / θ , while for α < 1 \alpha<1 α < 1 the memory kernel shifts the numerator q + 1 q+1 q + 1 to q + α q+\alpha q + α .
Main results
We consider the Cauchy problem (1) with the source h h h as in Convention 2.1 and nonnegative initial data u 0 ∈ L 1 ( R N ) ∩ L ∞ ( R N ) u_0\in L^1(\mathbb{R}^N)\cap L^\infty(\mathbb{R}^N) u 0 ∈ L 1 ( R N ) ∩ L ∞ ( R N ) , u 0 ≢ 0 u_0\not\equiv0 u 0 ≡ 0 . In the global-existence statements, smallness is always measured in ∥ u 0 ∥ L 1 + ∥ u 0 ∥ L ∞ \|u_0\|_{L^1}+\|u_0\|_{L^\infty} ∥ u 0 ∥ L 1 + ∥ u 0 ∥ L ∞ , while the blow-up statements hold for every nontrivial nonnegative datum in this class. Under Convention 2.1 the source is continuous up to t = 0 t=0 t = 0 , so a mild solution in the sense of Definition 2.3 exists locally and is unique. It is called global if it can be extended to all of [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) and is bounded on every [ 0 , T ] [0,T] [ 0 , T ] . Otherwise we say that the solution blows up in finite time , which means that there exists T ∗ < ∞ T^*<\infty T ∗ < ∞ such that lim sup t ↑ T ∗ ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) = ∞ \limsup_{t\uparrow T^*}\|u(t,\cdot)\|_{L^\infty(\mathbb{R}^N)}=\infty lim sup t ↑ T ∗ ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) = ∞ . Our first main result settles the case of the classical time derivative completely.
Theorem 3.1. Let α = 1 \alpha=1 α = 1 , θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] and p > 1 p>1 p > 1 , and let ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) \ell:(0,\infty)\to(0,\infty) ℓ : ( 0 , ∞ ) → ( 0 , ∞ ) be continuous and slowly varying at infinity. Set h ( t ) = t γ ℓ ( t ) h(t)=t^\gamma\ell(t) h ( t ) = t γ ℓ ( t ) for large t t t , in the sense of Convention 2.1.
(No Fujita regime: γ ≤ − 1 \gamma\le-1 γ ≤ − 1 .) For every p > 1 p>1 p > 1 and all sufficiently small nontrivial nonnegative data u 0 ∈ L 1 ∩ L ∞ u_0\in L^1\cap L^\infty u 0 ∈ L 1 ∩ L ∞ , the mild solution of (1) is global and satisfies ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) → 0 \|u(t,\cdot)\|_{L^\infty(\mathbb{R}^N)}\to0 ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) → 0 as t → ∞ t\to\infty t → ∞ . There is no Fujita threshold in this regime; formally, p c ( γ ) = ∞ p_c(\gamma)=\infty p c ( γ ) = ∞ .
(ii) (Fujita regime: γ > − 1 \gamma>-1 γ > − 1 .) The critical exponent is
p c ( γ ) = 1 + θ ( γ + 1 ) N . p_c(\gamma)=1+\frac{\theta(\gamma+1)}{N}. p c ( γ ) = 1 + N θ ( γ + 1 ) .
If 1 < p < p c ( γ ) 1<p<p_c(\gamma) 1 < p < p c ( γ ) , every nontrivial nonnegative solution blows up in finite time. If p > p c ( γ ) p>p_c(\gamma) p > p c ( γ ) , small global solutions exist and satisfy ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) ≤ C ( 1 + t ) − N / θ \lVert u(t,\cdot)\rVert_{L^\infty(\mathbb{R}^N)}\le C(1+t)^{-N/\theta} ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) ≤ C ( 1 + t ) − N / θ .
(iii) (Critical case: γ > − 1 \gamma>-1 γ > − 1 , p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) .) The behavior at the critical exponent is determined by the slowly varying function:
∫ 1 ∞ t − 1 ℓ ( t ) d t = ∞ ⟹ every nontrivial solution blows up in finite time, \int_1^\infty t^{-1}\ell(t)\,dt=\infty\quad\Longrightarrow\quad\text{every nontrivial solution blows up in finite time,} ∫ 1 ∞ t − 1 ℓ ( t ) d t = ∞ ⟹ every nontrivial solution blows up in finite time,
∫ 1 ∞ t − 1 ℓ ( t ) d t < ∞ ⟹ small global solutions exist. \int_1^\infty t^{-1}\ell(t)\,dt<\infty\quad\Longrightarrow\quad\text{small global solutions exist.} ∫ 1 ∞ t − 1 ℓ ( t ) d t < ∞ ⟹ small global solutions exist.
For the genuinely fractional time derivative the exponents change. For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and γ > − α \gamma>-\alpha γ > − α we set
p c ( γ , α ) : = 1 + θ ( γ + α ) N α . (20) p_c(\gamma,\alpha):=1+\frac{\theta(\gamma+\alpha)}{N\alpha}. \tag*{(20)} p c ( γ , α ) := 1 + N α θ ( γ + α ) . (20)
and note that p c ( γ , 1 ) = p c ( γ ) p_c(\gamma,1)=p_c(\gamma) p c ( γ , 1 ) = p c ( γ ) and p c ( 0 , α ) = 1 + θ / N p_c(0,\alpha)=1+\theta/N p c ( 0 , α ) = 1 + θ / N .
Theorem 3.2. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] and p > 1 p>1 p > 1 , and let h h h be as in Convention 2.1.
(a) (No Fujita regime: γ ≤ − α \gamma\le-\alpha γ ≤ − α .) For every p > 1 p>1 p > 1 and every N ≥ 1 N\ge1 N ≥ 1 , small data in L 1 ∩ L ∞ L^1\cap L^\infty L 1 ∩ L ∞ generate global mild solutions. In particular, global existence for all p > 1 p>1 p > 1 extends from γ ≤ − 1 \gamma\le-1 γ ≤ − 1 to the strictly larger range γ ≤ − α \gamma\le-\alpha γ ≤ − α .
(b) (Global existence: γ > − α \gamma>-\alpha γ > − α .) Small-data global existence holds in the following cases.
N < θ N<\theta N < θ (since N ≥ 1 N\ge1 N ≥ 1 , this means N = 1 N=1 N = 1 and θ ∈ ( 1 , 2 ] \theta\in(1,2] θ ∈ ( 1 , 2 ] ) and p > p c ( γ , α ) p>p_c(\gamma,\alpha) p > p c ( γ , α ) . In this case the solutions satisfy ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) ≤ C ( 1 + t ) − N α / θ \lVert u(t,\cdot)\rVert_{L^\infty(\mathbb{R}^N)}\le C(1+t)^{-N\alpha/\theta} ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) ≤ C ( 1 + t ) − N α / θ . Moreover, if N < θ N<\theta N < θ and ℓ \ell ℓ is bounded on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , small-data global existence also holds at p = p c ( γ , α ) p=p_c(\gamma,\alpha) p = p c ( γ , α ) .
Note that 1 + ( γ + α ) / α ≥ p c ( γ , α ) 1+(\gamma+\alpha)/\alpha\ge p_c(\gamma,\alpha) 1 + ( γ + α ) / α ≥ p c ( γ , α ) when N ≥ θ N\ge\theta N ≥ θ , with equality if and only if N = θ N=\theta N = θ .
(c) (Blow-up: γ > − α \gamma>-\alpha γ > − α , every N ≥ 1 N\ge1 N ≥ 1 .) If 1 < p < p c ( γ , α ) 1<p<p_c(\gamma,\alpha) 1 < p < p c ( γ , α ) , every nontrivial nonnegative solution blows up in finite time. The same conclusion holds at p = p c ( γ , α ) p=p_c(\gamma,\alpha) p = p c ( γ , α ) provided that lim sup t → ∞ ℓ ( t ) = ∞ \limsup_{t\to\infty}\ell(t)=\infty lim sup t → ∞ ℓ ( t ) = ∞ .
Proposition 3.1. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] , N > θ N>\theta N > θ , γ > 0 \gamma>0 γ > 0 , and let h h h be as in Convention 2.1. Assume that either 1 < p < 1 + γ / α 1<p<1+\gamma/\alpha 1 < p < 1 + γ / α , or p = 1 + γ / α p=1+\gamma/\alpha p = 1 + γ / α and lim sup t → ∞ ℓ ( t ) = ∞ \limsup_{t\to\infty}\ell(t)=\infty lim sup t → ∞ ℓ ( t ) = ∞ . Then every nontrivial nonnegative mild solution of (1) blows up in finite time.
Combining parts (b) and (c) of Theorem 3.2, we obtain a complete picture in the one-dimensional regime N < θ N<\theta N < θ .
Corollary 3.1. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , N = 1 N=1 N = 1 , θ ∈ ( 1 , 2 ] \theta\in(1,2] θ ∈ ( 1 , 2 ] and γ > − α \gamma>-\alpha γ > − α . Then p c ( γ , α ) p_c(\gamma,\alpha) p c ( γ , α ) is the Fujita exponent of (1) . At p = p c ( γ , α ) p=p_c(\gamma,\alpha) p = p c ( γ , α ) , small global solutions exist if ℓ \ell ℓ is bounded on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) , and every nontrivial solution blows up if ℓ \ell ℓ is unbounded on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) .
Since ℓ \ell ℓ is continuous, it is unbounded on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) if and only if lim sup t → ∞ ℓ ( t ) = ∞ \limsup_{t\to\infty}\ell(t)=\infty lim sup t → ∞ ℓ ( t ) = ∞ , so the two alternatives in Corollary 3.1 are exhaustive.
For N > θ N>\theta N > θ , Theorem 3.2 and Proposition 3.1 leave a gap between the blow-up range 1 < p < max { p c ( γ , α ) , 1 + γ / α } 1<p<\max\{p_c(\gamma,\alpha),1+\gamma/\alpha\} 1 < p < max { p c ( γ , α ) , 1 + γ / α } and the global range p > 1 + ( γ + α ) / α p>1+(\gamma+\alpha)/\alpha p > 1 + ( γ + α ) / α . Note that 1 + γ / α > p c ( γ , α ) 1+\gamma/\alpha>p_c(\gamma,\alpha) 1 + γ / α > p c ( γ , α ) if and only if γ > α θ / ( N − θ ) \gamma>\alpha\theta/(N-\theta) γ > α θ / ( N − θ ) ; in this regime p c ( γ , α ) p_c(\gamma,\alpha) p c ( γ , α ) is not the Fujita exponent.
Conjecture 3.1. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , N > θ N > \theta N > θ and γ > − α \gamma> -\alpha γ > − α . Then the Fujita exponent of (1) is
p ∗ ( γ , α ) : = max { p c ( γ , α ) , 1 + γ α } , p_*(\gamma,\alpha) := \max\left\{p_c(\gamma,\alpha), 1+\frac{\gamma}{\alpha}\right\}, p ∗ ( γ , α ) := max { p c ( γ , α ) , 1 + α γ } , that is, small global solutions exist for every p > p ∗ ( γ , α ) p > p_*(\gamma,\alpha) p > p ∗ ( γ , α ) . For N = θ N = \theta N = θ , small global solutions exist at p = p c ( γ , α ) p = p_c(\gamma,\alpha) p = p c ( γ , α ) whenever ℓ \ell ℓ is bounded.
The first term in p ∗ p_* p ∗ comes from the scaling (mass) mechanism of Theorem 3.2 (c), and the second one from the local mechanism of Proposition 3.1. The conjecture is supported only by heuristics: on balls of fixed size, the time-fractional dynamics behaves like the damped fractional ODE ∂ t α v + λ v = t γ v p \partial_t^\alpha v+\lambda v=t^\gamma v^p ∂ t α v + λ v = t γ v p , whose small solutions decay like t − α t^{-\alpha} t − α and remain global precisely when p > 1 + γ / α p>1+\gamma/\alpha p > 1 + γ / α . For γ = 0 \gamma=0 γ = 0 it reduces to the results of [42 , 7 ] .
A few remarks are in order.
Remark 3.1. Meier’s classification cannot persist verbatim for α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , and Theorem 3.2 (a) alone already shows it. Fix α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and γ ∈ ( − 1 , − α ] \gamma\in(-1,-\alpha] γ ∈ ( − 1 , − α ] , which is a non-empty range. Any classification driven by the numerator γ + 1 \gamma+1 γ + 1 , that is, any candidate critical exponent of the form 1 + c ( γ + 1 ) 1+c(\gamma+1) 1 + c ( γ + 1 ) with c > 0 c>0 c > 0 as produced by Meier’s formula (2) for any choice of s ∗ s^* s ∗ , would be finite and larger than 1 for such γ \gamma γ . It would therefore predict a non-empty blow-up window 1 < p < 1 + c ( γ + 1 ) 1<p<1+c(\gamma+1) 1 < p < 1 + c ( γ + 1 ) . Theorem 3.2 (a) excludes this: for every γ ≤ − α \gamma\leq-\alpha γ ≤ − α , small-data global solutions exist for every p > 1 p>1 p > 1 , although the source t γ ℓ ( t ) t^\gamma\ell(t) t γ ℓ ( t ) is not integrable at infinity when γ ∈ ( − 1 , − α ] \gamma\in(-1,-\alpha] γ ∈ ( − 1 , − α ] . Hence, for α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , the Fujita regime starts at γ = − α \gamma=-\alpha γ = − α and not at γ = − 1 \gamma=-1 γ = − 1 . Conversely, the blow-up range 1 < p < p c ( γ , α ) 1<p<p_c(\gamma,\alpha) 1 < p < p c ( γ , α ) of Theorem 3.2 (c) shows that the shift also occurs on the blow-up side, for every γ > − α \gamma>-\alpha γ > − α and every N ≥ 1 N\geq1 N ≥ 1 . The memory of the Caputo derivative thus provably shifts the numerator from γ + 1 \gamma+1 γ + 1 to γ + α \gamma+\alpha γ + α .
Remark 3.2. When θ = 2 \theta=2 θ = 2 (and α = 1 \alpha=1 α = 1 ), Theorem 3.1(i)–(ii) recovers the results of Meier [31 ] on R N \mathbb{R}^N R N , with p c ( γ ) = 1 + 2 ( γ + 1 ) / N p_c(\gamma)=1+2(\gamma+1)/N p c ( γ ) = 1 + 2 ( γ + 1 ) / N . The critical-case dichotomy (iii) can be read off the integral ∫ 1 ∞ h ( t ) t − N ( p − 1 ) / θ d t \int_1^\infty h(t)t^{-N(p-1)/\theta}\,dt ∫ 1 ∞ h ( t ) t − N ( p − 1 ) / θ d t , which, for h ( t ) = t γ ℓ ( t ) h(t)=t^\gamma\ell(t) h ( t ) = t γ ℓ ( t ) and p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) , reduces exactly to ∫ 1 ∞ t − 1 ℓ ( t ) d t \int_1^\infty t^{-1}\ell(t)\,dt ∫ 1 ∞ t − 1 ℓ ( t ) d t . Its divergence is the blow-up criterion of Guedda and Kirane [13 ] , while its convergence is precisely condition (18) for w = S 1 , θ ( t ) w 0 w=S_{1,\theta}(t)w_0 w = S 1 , θ ( t ) w 0 , because ∥ S 1 , θ ( t ) w 0 ∥ ∞ ≍ t − N / θ \|S_{1,\theta}(t)w_0\|_\infty\asymp t^{-N/\theta} ∥ S 1 , θ ( t ) w 0 ∥ ∞ ≍ t − N / θ for large t t t . Meier’s supersolution a ( t ) w a(t)w a ( t ) w is also a supersolution when − Δ -\Delta − Δ is replaced by ( − Δ ) θ / 2 (-\Delta)^{\theta/2} ( − Δ ) θ /2 . Consequently, at α = 1 \alpha=1 α = 1 the individual ingredients of Theorem 3.1 are essentially available in the literature [31 , 13 , 38 ] . Our contribution there is a unified treatment of all θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] and the formulation of the critical case in terms of Karamata’s theory. The genuinely new results concern α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , namely Theorem 3.2 , Proposition 3.1 and Corollary 3.1.
Remark 3.3. The dependence of the critical exponent on the time order α \alpha α is more subtle than the autonomous case suggests. Writing p c ( γ , α ) − 1 = θ N ( 1 + γ α ) p_c(\gamma,\alpha)-1=\frac{\theta}{N}\left(1+\frac{\gamma}{\alpha}\right) p c ( γ , α ) − 1 = N θ ( 1 + α γ ) , we see the following. At γ = 0 \gamma=0 γ = 0 , the exponent 1 + θ / N 1+\theta/N 1 + θ / N does not depend on α \alpha α , in agreement with [42 , 7 ] . For γ > 0 \gamma>0 γ > 0 , a smaller α \alpha α (more memory) yields a larger exponent, and hence a larger blow-up window. For γ ∈ ( − α , 0 ) \gamma\in(-\alpha,0) γ ∈ ( − α , 0 ) the monotonicity is reversed. Finally, the onset of the Fujita regime itself moves from γ = − 1 \gamma=-1 γ = − 1 (at α = 1 \alpha=1 α = 1 ) to γ = − α \gamma=-\alpha γ = − α . Heuristically, the memory kernel ( t − s ) α − 1 (t-s)^{\alpha-1} ( t − s ) α − 1 in (15) converts the cumulative source ∫ 0 t h ( s ) d s \int_0^t h(s)\,ds ∫ 0 t h ( s ) d s of Meier’s criterion into the weighted average ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≍ B ( α , γ + 1 ) t γ + α ℓ ( t ) \int_0^t(t-s)^{\alpha-1}h(s)\,ds\asymp B(\alpha,\gamma+1)t^{\gamma+\alpha}\ell(t) ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≍ B ( α , γ + 1 ) t γ + α ℓ ( t ) (valid for γ > − 1 \gamma>-1 γ > − 1 ), which replaces γ + 1 \gamma+1 γ + 1 by γ + α \gamma+\alpha γ + α throughout. For N > θ N>\theta N > θ , the local mechanism of Proposition 3.1 adds a second, θ \theta θ -independent constraint p > 1 + γ / α p>1+\gamma/\alpha p > 1 + γ / α .
Remark 3.4. The slowly varying factor ℓ \ell ℓ has no effect on the value of the critical exponent. At α = 1 \alpha=1 α = 1 , it is decisive exactly at the critical exponent p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) , for every γ > − 1 \gamma>-1 γ > − 1 . There, the dichotomy is controlled by the single Karamata-type integral ∫ 1 ∞ t − 1 ℓ ( t ) d t \int_1^\infty t^{-1}\ell(t)\,dt ∫ 1 ∞ t − 1 ℓ ( t ) d t , and Lemma 2.1(K3) yields the following examples.
ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 : divergence, hence blow-up at criticality (for γ = 0 \gamma= 0 γ = 0 , θ = 2 \theta= 2 θ = 2 this is the classical critical-case blow-up of Hayakawa [14 ] and Kobayashi, Sirao and Tanaka [21 ] );
ℓ ( t ) = ( log t ) a \ell(t) = (\log t)^a ℓ ( t ) = ( log t ) a for t ≥ e t \ge e t ≥ e : divergence if and only if a ≥ − 1 a \ge-1 a ≥ − 1 , convergence if and only if a < − 1 a < -1 a < − 1 ;
ℓ ( t ) = ( log t ) − 1 ( log log t ) b \ell(t) = (\log t)^{-1}(\log\log t)^b ℓ ( t ) = ( log t ) − 1 ( log log t ) b for large t t t : divergence if and only if b ≥ − 1 b \ge-1 b ≥ − 1 .
For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and N < θ N < \theta N < θ , Corollary 3.1 shows that the relevant property is the boundedness of ℓ \ell ℓ , a weaker requirement than the convergence of the Karamata integral. For instance, for ℓ ( t ) = ( log t ) − 1 / 2 \ell(t) = (\log t)^{-1/2} ℓ ( t ) = ( log t ) − 1/2 (for t ≥ e t \ge e t ≥ e ), solutions blow up at the critical exponent when α = 1 \alpha= 1 α = 1 , whereas small global solutions exist at the critical exponent when α < 1 \alpha< 1 α < 1 . The memory of the Caputo derivative thus turns critical blow-up into critical global existence for every bounded ℓ \ell ℓ , in agreement with the autonomous case ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 , γ = 0 \gamma= 0 γ = 0 [42 , 7 ] .
Remark 3.5. The dichotomy in Theorem 3.1(iii) is sharp. At p = p c ( γ ) p = p_c(\gamma) p = p c ( γ ) , any slowly varying ℓ \ell ℓ with a finite tail integral ∫ 1 ∞ t − 1 ℓ ( t ) d t \int_1^\infty t^{-1}\ell(t)\,dt ∫ 1 ∞ t − 1 ℓ ( t ) d t allows small global solutions, while any ℓ \ell ℓ with a divergent tail integral forces blow-up of every nontrivial solution. In particular, an arbitrarily mild logarithmic damping, for instance ℓ ( t ) = ( log t ) − 1 − δ \ell(t) = (\log t)^{-1-\delta} ℓ ( t ) = ( log t ) − 1 − δ with δ > 0 \delta> 0 δ > 0 , suffices to move the critical exponent to the global-existence side.
Example 3.1. Let α = 1 \alpha= 1 α = 1 , fix any γ > − 1 \gamma> -1 γ > − 1 , and take ℓ ( t ) = ( log t ) − 2 \ell(t) = (\log t)^{-2} ℓ ( t ) = ( log t ) − 2 for t ≥ e t \ge e t ≥ e (extended to ( 0 , e ) (0,e) ( 0 , e ) as a positive continuous function). Then
∫ e ∞ 1 t ( log t ) 2 d t = [ − 1 log t ] e ∞ = 1 < ∞ . \int_e^\infty\frac{1}{t(\log t)^2}\,dt = \left[-\frac{1}{\log t}\right]_e^\infty= 1 < \infty. ∫ e ∞ t ( log t ) 2 1 d t = [ − log t 1 ] e ∞ = 1 < ∞.
By Theorem 3.1(iii), small global solutions exist at the critical exponent p = p c ( γ ) = 1 + θ ( γ + 1 ) / N p = p_c(\gamma) = 1+\theta(\gamma+1)/N p = p c ( γ ) = 1 + θ ( γ + 1 ) / N . This is impossible for ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 , where the critical case always belongs to the blow-up regime [14 , 21 , 38 ] .
Example 3.2. Let α = 1 \alpha= 1 α = 1 , γ = 0 \gamma= 0 γ = 0 and ℓ ( t ) = ( log ( 2 + t ) ) a \ell(t) = (\log(2+t))^a ℓ ( t ) = ( log ( 2 + t ) ) a with a ∈ R a \in\mathbb{R} a ∈ R . This corresponds to h ( t ) ∼ ( log t ) a h(t) \sim(\log t)^a h ( t ) ∼ ( log t ) a , a logarithmic perturbation of the autonomous source. Theorem 3.1(ii) gives p c ( 0 ) = 1 + θ / N p_c(0) = 1+\theta/N p c ( 0 ) = 1 + θ / N independently of a a a , so the logarithmic correction does not shift the threshold. It does, however, decide the critical case. At p = 1 + θ / N p = 1+\theta/N p = 1 + θ / N , every nontrivial solution blows up when a ≥ − 1 a \ge-1 a ≥ − 1 , whereas small global solutions exist when a < − 1 a < -1 a < − 1 .
Example 3.3. Let α = 1 \alpha= 1 α = 1 and γ > − 1 \gamma> -1 γ > − 1 , and consider the doubly logarithmic correction ℓ ( t ) = ( log ( 2 + t ) ) a ( log log ( 16 + t ) ) b \ell(t) = (\log(2+t))^a(\log\log(16+t))^b ℓ ( t ) = ( log ( 2 + t ) ) a ( log log ( 16 + t ) ) b with a , b ∈ R a,b \in\mathbb{R} a , b ∈ R . The critical exponent is still p c ( γ ) = 1 + θ ( γ + 1 ) / N p_c(\gamma) = 1+\theta(\gamma+1)/N p c ( γ ) = 1 + θ ( γ + 1 ) / N , independently of ( a , b ) (a,b) ( a , b ) . At p = p c ( γ ) p = p_c(\gamma) p = p c ( γ ) , blow-up occurs precisely when a > − 1 a > -1 a > − 1 , or when a = − 1 a = -1 a = − 1 and b ≥ − 1 b \ge-1 b ≥ − 1 .
Other classes of time-dependent coefficients
We briefly comment on classes of source intensities h h h that are not covered by the Karamata framework h ( t ) = t γ ℓ ( t ) h(t) = t^\gamma\ell(t) h ( t ) = t γ ℓ ( t ) .
Exponential growth
Let h ( t ) = e μ t h(t) = e^{\mu t} h ( t ) = e μ t with μ > 0 \mu> 0 μ > 0 ; the growth rate μ \mu μ is unrelated to the decay exponent β = N α / θ \beta= N\alpha/\theta β = N α / θ of § 5.1 . The ODE a ′ ( t ) = e μ t a ( t ) p a'(t) = e^{\mu t}a(t)^p a ′ ( t ) = e μ t a ( t ) p , which is the analogue of Meier’s auxiliary equation, has the explicit solution
a ( t ) 1 − p = a 0 1 − p − p − 1 μ ( e μ t − 1 ) , a(t)^{1-p} = a_0^{1-p} - \frac{p-1}{\mu}\left(e^{\mu t}-1\right), a ( t ) 1 − p = a 0 1 − p − μ p − 1 ( e μ t − 1 ) ,
which blows up at
T ∗ = μ − 1 log ( 1 + μ a 0 1 − p p − 1 ) = μ − 1 log ( 1 + μ ( p − 1 ) a 0 p − 1 ) . T^* = \mu^{-1}\log\left(1+\frac{\mu a_0^{1-p}}{p-1}\right) = \mu^{-1}\log\left(1+\frac{\mu}{(p-1)a_0^{p-1}}\right). T ∗ = μ − 1 log ( 1 + p − 1 μ a 0 1 − p ) = μ − 1 log ( 1 + ( p − 1 ) a 0 p − 1 μ ) .
Thus T ∗ T^* T ∗ is finite for every a 0 > 0 a_0 > 0 a 0 > 0 , and T ∗ → ∞ T^* \to\infty T ∗ → ∞ as a 0 → 0 + a_0 \to0^+ a 0 → 0 + . Hence, even without diffusion, the source alone forces blow-up in finite time for every p > 1 p > 1 p > 1 . In the PDE setting on the whole space R N \mathbb{R}^N R N with α = 1 \alpha= 1 α = 1 , the extreme growth of h h h overwhelms the dissipation. The Kaplan-type functional of §5.2 yields the differential inequality ψ ′ ≥ e μ t ψ p \psi' \ge e^{\mu t}\psi^p ψ ′ ≥ e μ t ψ p , and the resulting necessary condition for global existence, sup T T − N ( p − 1 ) / θ e μ T < ∞ \sup_T T^{-N(p-1)/\theta}e^{\mu T} < \infty sup T T − N ( p − 1 ) / θ e μ T < ∞ , fails for every p > 1 p > 1 p > 1 . Consequently, every nontrivial nonnegative solution blows up in finite time for every p > 1 p > 1 p > 1 , regardless of θ \theta θ . The same behavior is expected for α < 1 \alpha< 1 α < 1 , although the memory term precludes the simple adjoint argument. On a bounded domain the situation is more interesting; see Meier [31, Theorem 4] for the sharp threshold p μ ∗ = 1 + μ / λ 1 p_\mu^* = 1 + \mu/\lambda_1 p μ ∗ = 1 + μ / λ 1 , where λ 1 \lambda_1 λ 1 is the first Dirichlet eigenvalue of the Laplacian.
Pure power behavior
If h ( t ) = t q h(t) = t^q h ( t ) = t q , with no slowly varying correction, we are in the regularly varying case with ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 . For α = 1 \alpha= 1 α = 1 and q > − 1 q > -1 q > − 1 , Theorem 3.1 yields p c = 1 + θ ( q + 1 ) / N p_c = 1 + \theta(q + 1)/N p c = 1 + θ ( q + 1 ) / N , with blow-up including the critical exponent , since ∫ 1 ∞ t − 1 d t = ∞ \int_1^\infty t^{-1}\,dt = \infty ∫ 1 ∞ t − 1 d t = ∞ . Taking θ = 2 \theta= 2 θ = 2 recovers Meier’s value p c = 1 + 2 ( q + 1 ) / N p_c = 1 + 2(q + 1)/N p c = 1 + 2 ( q + 1 ) / N [31 ] , while θ ∈ ( 0 , 2 ) \theta\in(0,2) θ ∈ ( 0 , 2 ) gives its natural θ \theta θ -stable analogue, which reduces to Sugitani’s theorem [38 ] at q = 0 q = 0 q = 0 . For α < 1 \alpha< 1 α < 1 , see Theorem 3.2 , Proposition 3.1 and Conjecture 3.1.
Oscillatory coefficients
Slowly varying functions may exhibit persistent oscillations. A canonical example is ℓ ( t ) = 2 + sin ( log log t ) \ell(t) = 2 + \sin(\log\log t) ℓ ( t ) = 2 + sin ( log log t ) for large t t t . It is slowly varying because
log log ( λ t ) − log log t = log ( 1 + log λ log t ) ⟶ 0 ( t → ∞ ) , \log\log(\lambda t) - \log\log t = \log\left(1 + \frac{\log\lambda}{\log t}\right) \longrightarrow0 \qquad(t \to\infty), log log ( λ t ) − log log t = log ( 1 + log t log λ ) ⟶ 0 ( t → ∞ ) ,
so that sin ( log log ( λ t ) ) − sin ( log log t ) → 0 \sin(\log\log(\lambda t)) - \sin(\log\log t) \to0 sin ( log log ( λ t )) − sin ( log log t ) → 0 for every λ > 0 \lambda> 0 λ > 0 , and yet ℓ \ell ℓ does not converge as t → ∞ t \to\infty t → ∞ . Note that the often-quoted candidate 2 + sin ( log t ) 2 + \sin(\log t) 2 + sin ( log t ) is not slowly varying. Indeed, sin ( log ( λ t ) ) − sin ( log t ) = 2 cos ( log t + log λ 2 ) sin ( log λ 2 ) \sin(\log(\lambda t)) - \sin(\log t) = 2\cos\left(\log t + \frac{\log\lambda}{2}\right)\sin\left(\frac{\log\lambda}{2}\right) sin ( log ( λ t )) − sin ( log t ) = 2 cos ( log t + 2 l o g λ ) sin ( 2 l o g λ ) oscillates between ± 2 ∣ sin ( log λ / 2 ) ∣ \pm2\lvert\sin(\log\lambda/2)\rvert ± 2 ∣ sin ( log λ /2 )∣ as t → ∞ t \to\infty t → ∞ , for every fixed λ ≠ 1 \lambda\ne1 λ = 1 with sin ( log λ / 2 ) ≠ 0 \sin(\log\lambda/2) \ne0 sin ( log λ /2 ) = 0 . Theorem 3.1 still applies. The value of p c ( γ ) p_c(\gamma) p c ( γ ) does not see such oscillations, and at p = p c ( γ ) p = p_c(\gamma) p = p c ( γ ) the dichotomy depends only on the Karamata integral ∫ 1 ∞ t − 1 ℓ ( t ) d t \int_1^\infty t^{-1}\ell(t)\,dt ∫ 1 ∞ t − 1 ℓ ( t ) d t . For ℓ ( t ) = 2 + sin ( log log t ) \ell(t) = 2 + \sin(\log\log t) ℓ ( t ) = 2 + sin ( log log t ) this integral diverges (since ℓ ≥ 1 \ell\ge1 ℓ ≥ 1 ), so blow-up persists at criticality. For α < 1 \alpha< 1 α < 1 and N < θ N < \theta N < θ , this ℓ \ell ℓ is bounded, and Corollary 3.1 gives global existence at the critical exponent.
Integrable sources of non-power decay
Assume that α = 1 \alpha= 1 α = 1 , that h ∈ C ( [ 0 , ∞ ) ) h \in C([0,\infty)) h ∈ C ([ 0 , ∞ )) is nonnegative and h ∈ L 1 ( 0 , ∞ ) h \in L^1(0,\infty) h ∈ L 1 ( 0 , ∞ ) , but that h h h is not regularly varying; for example h ( t ) = e − t h(t) = e^{-t} h ( t ) = e − t or h ( t ) = ( 1 + t ) − 2 ( 2 + sin t ) h(t) = (1+t)^{-2}(2+\sin t) h ( t ) = ( 1 + t ) − 2 ( 2 + sin t ) . Integrability alone then forces global existence for small data, for every p > 1 p > 1 p > 1 . Indeed, in the Duhamel fixed-point argument of § 5.1 , the contribution of [ 0 , t / 2 ] [0,t/2] [ 0 , t /2 ] is bounded by C t − β p ∫ 0 ∞ h Ct^{-\beta p}\int_0^\infty h C t − β p ∫ 0 ∞ h , and that of [ t / 2 , t ] [t/2,t] [ t /2 , t ] by C ( 1 + t ) − β p ∫ 0 ∞ h C(1+t)^{-\beta p}\int_0^\infty h C ( 1 + t ) − β p ∫ 0 ∞ h . For α < 1 \alpha< 1 α < 1 , integrability alone is not sufficient for this argument, since the near-diagonal term ∫ t / 2 t ( t − s ) α − 1 h ( s ) d s \int_{t/2}^t(t-s)^{\alpha-1}h(s)\,ds ∫ t /2 t ( t − s ) α − 1 h ( s ) d s need not decay. A pointwise bound such as h ( t ) ≤ C ( 1 + t ) − α h(t) \le C(1+t)^{-\alpha} h ( t ) ≤ C ( 1 + t ) − α restores the conclusion through Theorem 3.2 (a) and its proof. The autonomous-case analysis is then irrelevant.
More general regularly varying functions
If h ( t ) = t γ g ( t ) h(t) = t^\gamma g(t) h ( t ) = t γ g ( t ) with g g g regularly varying of index ρ ≠ 0 \rho\ne0 ρ = 0 , i.e. g ( λ t ) / g ( t ) → λ ρ g(\lambda t)/g(t) \to\lambda^\rho g ( λ t ) / g ( t ) → λ ρ for every λ > 0 \lambda> 0 λ > 0 , then h h h is itself regularly varying of index γ + ρ \gamma+ \rho γ + ρ . At α = 1 \alpha= 1 α = 1 , the critical exponent formula becomes
p c ( γ + ρ ) = 1 + θ ( γ + ρ + 1 ) N , p_c(\gamma+\rho) = 1 + \frac{\theta(\gamma+\rho+1)}{N}, p c ( γ + ρ ) = 1 + N θ ( γ + ρ + 1 ) ,
provided that γ + ρ > − 1 \gamma+ \rho> -1 γ + ρ > − 1 . The critical-case dichotomy is then governed by the slowly varying function g ( t ) t − ρ g(t)t^{-\rho} g ( t ) t − ρ , through the integral ∫ 1 ∞ t − 1 g ( t ) t − ρ d t \int_1^\infty t^{-1}g(t)t^{-\rho}\,dt ∫ 1 ∞ t − 1 g ( t ) t − ρ d t .
Two-sided power bounds (Meier's assumption (h2))
Meier [31 ] works under the weaker assumption
α 0 t q ≤ h ( t ) ≤ α 1 t q for t large, \alpha_0 t^q \le h(t) \le\alpha_1 t^q \qquad\text{for } t \text{ large,} α 0 t q ≤ h ( t ) ≤ α 1 t q for t large,
with constants 0 < α 0 ≤ α 1 0 < \alpha_0 \le\alpha_1 0 < α 0 ≤ α 1 (here α 0 \alpha_0 α 0 , α 1 \alpha_1 α 1 are unrelated to the Caputo order α \alpha α ). Our framework covers this case with γ = q \gamma= q γ = q , since such an h h h is sandwiched between two regularly varying functions with the same index. The global-existence estimates of § 5.1 use only the upper bound, and the blow-up arguments of §§5.2–5.4 use only the lower bound. Hence the proofs adapt directly away from the critical exponent. At the critical exponent, the two-sided bound does not determine the dichotomy, which is genuinely a property of ℓ \ell ℓ .
Proofs of the main results
This section contains the proofs. Global existence is obtained by a fixed-point argument in weighted spaces (§ 5.1 ). For blow-up at α = 1 \alpha= 1 α = 1 we use a Kaplan-type functional built on the kernel Z Z Z , together with the integral criterion of Guedda and Kirane [13 ] at the critical exponent (§5.2). For α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) we adapt the nonlinear iteration of Cortázar, Quirós and Wolanski [7 ] to the time-dependent source (§5.3), and we prove the local blow-up mechanism of Proposition 3.1 in §5.4. The regime γ ≤ − α \gamma\le-\alpha γ ≤ − α and the fate of Meier’s formula are discussed in §§5.5–5.6. Throughout, the kernels and solutions are written as functions of ( t , x ) (t,x) ( t , x ) , in this order.
Global existence
5.1. Global existence. Throughout this subsection we work under the standing hypothesis
(S) α = 1 (any N ≥ 1 ) or α ∈ ( 0 , 1 ) and N < θ . (21) \text{(S)} \qquad\alpha= 1 \text{ (any } N \ge1\text{)} \qquad\text{or} \qquad\alpha\in(0,1) \text{ and } N < \theta. \tag*{(21)} (S) α = 1 (any N ≥ 1 ) or α ∈ ( 0 , 1 ) and N < θ . (21)
By Proposition 2.1 and Remark 2.1, this is exactly the regime in which the L 1 → L ∞ L^1 \to L^\infty L 1 → L ∞ smoothing estimate ∥ S α , θ ( t ) ∥ L 1 → L ∞ ≤ C t − N α / θ \lVert S_{\alpha,\theta}(t)\rVert_{L^1\to L^\infty} \le Ct^{-N\alpha/\theta} ∥ S α , θ ( t ) ∥ L 1 → L ∞ ≤ C t − N α / θ holds. Under (S), the analogous estimate ∥ P α , θ ( t ) ∥ L 1 → L ∞ ≤ C t − N α / θ \lVert P_{\alpha,\theta}(t)\rVert_{L^1\to L^\infty} \le Ct^{-N\alpha/\theta} ∥ P α , θ ( t ) ∥ L 1 → L ∞ ≤ C t − N α / θ also holds, because N < θ < 2 θ N < \theta< 2\theta N < θ < 2 θ when α < 1 \alpha< 1 α < 1 (and P 1 , θ = S 1 , θ P_{1,\theta} = S_{1,\theta} P 1 , θ = S 1 , θ ). The case α < 1 \alpha< 1 α < 1 , N ≥ θ N \ge\theta N ≥ θ is treated in Remark 5.1 below.
Set-up. Set β : = N α / θ \beta:= N\alpha/\theta β := N α / θ and let
X β : = { u : ( 0 , ∞ ) × R N → R measurable : ∥ u ∥ X β < ∞ } , X_\beta:= \{u : (0,\infty) \times\mathbb{R}^N \to\mathbb{R} \text{ measurable} : \lVert u\rVert_{X_\beta} < \infty\}, X β := { u : ( 0 , ∞ ) × R N → R measurable : ∥ u ∥ X β < ∞ } ,
endowed with the norm
∥ u ∥ X β : = e s s s u p t > 0 [ ∥ u ( t , ⋅ ) ∥ L 1 ( R N ) + ( 1 + t ) β ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) ] . \lVert u\rVert_{X_\beta} := \mathop{\mathrm{ess\,sup}}_{t>0} \left[\lVert u(t,\cdot)\rVert_{L^1(\mathbb{R}^N)} + (1+t)^\beta\lVert u(t,\cdot)\rVert_{L^\infty(\mathbb{R}^N)}\right]. ∥ u ∥ X β := ess sup t > 0 [ ∥ u ( t , ⋅ ) ∥ L 1 ( R N ) + ( 1 + t ) β ∥ u ( t , ⋅ ) ∥ L ∞ ( R N ) ] .
Then X β X_\beta X β is a Banach space, and its nonnegative elements form a closed subset X β + X_\beta^+ X β + . For nonnegative u 0 ∈ L 1 ( R N ) ∩ L ∞ ( R N ) u_0 \in L^1(\mathbb{R}^N) \cap L^\infty(\mathbb{R}^N) u 0 ∈ L 1 ( R N ) ∩ L ∞ ( R N ) , we define the map
Λ [ u ] ( t ) = S α , θ ( t ) u 0 + ∫ 0 t ( t − s ) α − 1 P α , θ ( t − s ) [ h ( s ) u ( s ) p ] d s = : L [ u ] ( t ) + N [ u ] ( t ) . \Lambda[u](t) = S_{\alpha,\theta}(t)u_0 + \int_0^t (t-s)^{\alpha-1}P_{\alpha,\theta}(t-s)[h(s)u(s)^p]\,\mathrm{d}s =: \mathcal{L}[u](t) + \mathcal{N}[u](t). Λ [ u ] ( t ) = S α , θ ( t ) u 0 + ∫ 0 t ( t − s ) α − 1 P α , θ ( t − s ) [ h ( s ) u ( s ) p ] d s =: L [ u ] ( t ) + N [ u ] ( t ) .
A mild solution of (1) is a fixed point of Λ \Lambda Λ . Since the kernels of S α , θ S_{\alpha,\theta} S α , θ and P α , θ P_{\alpha,\theta} P α , θ and the function h h h are nonnegative, Λ \Lambda Λ maps X β + X_\beta^+ X β + into nonnegative functions.
Decay of the linear part. Under (S), Proposition 2.1 with ( p , q ) = ( 1 , ∞ ) (p,q) = (1,\infty) ( p , q ) = ( 1 , ∞ ) gives ∥ L [ u ] ( t ) ∥ ∞ ≤ C t − N α / θ ∥ u 0 ∥ 1 \lVert\mathcal{L}[u](t)\rVert_\infty\le Ct^{-N\alpha/\theta}\lVert u_0\rVert_1 ∥ L [ u ] ( t ) ∥ ∞ ≤ C t − N α / θ ∥ u 0 ∥ 1 for t ≥ 1 t \ge1 t ≥ 1 , while ∥ L [ u ] ( t ) ∥ ∞ ≤ ∥ u 0 ∥ ∞ \lVert\mathcal{L}[u](t)\rVert_\infty\le\lVert u_0\rVert_\infty ∥ L [ u ] ( t ) ∥ ∞ ≤ ∥ u 0 ∥ ∞ for t ∈ ( 0 , 1 ] t \in(0,1] t ∈ ( 0 , 1 ] . Hence
∥ L [ u ] ( t ) ∥ ∞ ≤ C 0 ( 1 + t ) − N α / θ ( ∥ u 0 ∥ 1 + ∥ u 0 ∥ ∞ ) . \lVert\mathcal{L}[u](t)\rVert_\infty\le C_0(1+t)^{-N\alpha/\theta}(\lVert u_0\rVert_1+\lVert u_0\rVert_\infty). ∥ L [ u ] ( t ) ∥ ∞ ≤ C 0 ( 1 + t ) − N α / θ (∥ u 0 ∥ 1 + ∥ u 0 ∥ ∞ ) .
The mass bound ∥ L [ u ] ( t ) ∥ 1 ≤ ∥ u 0 ∥ 1 \lVert\mathcal{L}[u](t)\rVert_1 \le\lVert u_0\rVert_1 ∥ L [ u ] ( t ) ∥ 1 ≤ ∥ u 0 ∥ 1 follows at once from the L 1 L^1 L 1 contractivity of S α , θ ( t ) S_{\alpha,\theta}(t) S α , θ ( t ) .
Nonlinear estimates. For u ∈ X β + u \in X_\beta^+ u ∈ X β + , we write M : = ∥ u ∥ X β M := \lVert u\rVert_{X_\beta} M := ∥ u ∥ X β . Then, for almost every s > 0 s > 0 s > 0 ,
∥ ( u ( s ) ) p ∥ L 1 ( R N ) ≤ ∥ u ( s ) ∥ L ∞ p − 1 ∥ u ( s ) ∥ L 1 ≤ M p ( 1 + s ) − β ( p − 1 ) , (22) \lVert(u(s))^p\rVert_{L^1(\mathbb{R}^N)} \le\lVert u(s)\rVert_{L^\infty} ^{p-1}\lVert u(s)\rVert_{L^1} \le M^p(1+s)^{-\beta(p-1)}, \tag*{(22)} ∥( u ( s ) ) p ∥ L 1 ( R N ) ≤ ∥ u ( s ) ∥ L ∞ p − 1 ∥ u ( s ) ∥ L 1 ≤ M p ( 1 + s ) − β ( p − 1 ) , (22)
and
∥ ( u ( s ) ) p ∥ L ∞ ( R N ) ≤ ∥ u ( s ) ∥ L ∞ p ≤ M p ( 1 + s ) − β p . (23) \left\|(u(s))^p\right\|_{L^\infty(\mathbb{R}^N)} \le\|u(s)\|_{L^\infty}^p \le M^p(1+s)^{-\beta p}. \tag*{(23)} ∥ ( u ( s ) ) p ∥ L ∞ ( R N ) ≤ ∥ u ( s ) ∥ L ∞ p ≤ M p ( 1 + s ) − β p . (23)
The decisive quantity is
κ : = γ + α − β ( p − 1 ) , (24) \kappa:= \gamma+\alpha-\beta(p-1), \tag*{(24)} κ := γ + α − β ( p − 1 ) , (24)
which measures the competition between the cumulative source, weighted by the memory kernel, and the dispersive decay. With β = N α / θ \beta=N\alpha/\theta β = N α / θ , we have
κ < 0 ⟺ p > 1 + γ + α β = 1 + θ ( γ + α ) N α . \kappa<0 \Longleftrightarrow p>1+\frac{\gamma+\alpha}{\beta}=1+\frac{\theta(\gamma+\alpha)}{N\alpha}. κ < 0 ⟺ p > 1 + β γ + α = 1 + N α θ ( γ + α ) .
At α = 1 \alpha=1 α = 1 this is exactly p > p c ( γ ) p>p_c(\gamma) p > p c ( γ ) , and it holds for every p > 1 p>1 p > 1 when γ ≤ − α \gamma\le-\alpha γ ≤ − α . For α < 1 \alpha<1 α < 1 and γ > − α \gamma>-\alpha γ > − α it reads p > p c ( γ , α ) p>p_c(\gamma,\alpha) p > p c ( γ , α ) .
The strictly supercritical regime κ < 0 \kappa<0 κ < 0 . Fix ε ∈ ( 0 , − κ ) \varepsilon\in(0,-\kappa) ε ∈ ( 0 , − κ ) . By Lemma 2.1(K1) there exists t 0 ≥ 2 t_0\ge2 t 0 ≥ 2 such that
ℓ ( s ) ≤ C s ε , for all s ≥ t 0 . (25) \ell(s)\le Cs^\varepsilon,\qquad\text{for all }s\ge t_0. \tag*{(25)} ℓ ( s ) ≤ C s ε , for all s ≥ t 0 . (25)
Case t ≤ 2 t 0 t\le2t_0 t ≤ 2 t 0 . Using the uniform bound ∥ P α , θ ( t − s ) f ∥ ∞ ≤ C ∥ f ∥ ∞ \|P_{\alpha,\theta}(t-s)f\|_\infty\le C\|f\|_\infty ∥ P α , θ ( t − s ) f ∥ ∞ ≤ C ∥ f ∥ ∞ , we get
∥ N [ u ] ( t ) ∥ ∞ ≤ C M p ∫ 0 t ( t − s ) α − 1 h ( s ) ( 1 + s ) − β p d s ≤ C M p ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≤ C M p ( 2 t 0 ) α . \begin{aligned}
\|\mathcal{N}[u](t)\|_\infty
&\le CM^p\int_0^t(t-s)^{\alpha-1}h(s)(1+s)^{-\beta p}\,ds\\
&\le CM^p\int_0^t(t-s)^{\alpha-1}h(s)\,ds\\
&\le CM^p(2t_0)^\alpha.
\end{aligned} ∥ N [ u ] ( t ) ∥ ∞ ≤ C M p ∫ 0 t ( t − s ) α − 1 h ( s ) ( 1 + s ) − β p d s ≤ C M p ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≤ C M p ( 2 t 0 ) α .
where we used the boundedness of h h h on [ 0 , 2 t 0 ] [0,2t_0] [ 0 , 2 t 0 ] . Since ( 1 + t ) − β ≥ ( 1 + 2 t 0 ) − β (1+t)^{-\beta}\ge(1+2t_0)^{-\beta} ( 1 + t ) − β ≥ ( 1 + 2 t 0 ) − β for t ≤ 2 t 0 t\le2t_0 t ≤ 2 t 0 , it follows that
∥ N [ u ] ( t ) ∥ ∞ ≤ C M p ( 1 + t ) − β . (26) \|\mathcal{N}[u](t)\|_\infty\le CM^p(1+t)^{-\beta}. \tag*{(26)} ∥ N [ u ] ( t ) ∥ ∞ ≤ C M p ( 1 + t ) − β . (26)
Case t ≥ 2 t 0 t\ge2t_0 t ≥ 2 t 0 . We split the integral defining N [ u ] ( t ) \mathcal{N}[u](t) N [ u ] ( t ) into the intervals [ 0 , t / 2 ] [0,t/2] [ 0 , t /2 ] and [ t / 2 , t ] [t/2,t] [ t /2 , t ] . On [ 0 , t / 2 ] [0,t/2] [ 0 , t /2 ] we apply the L 1 L^1 L 1 –L ∞ L^\infty L ∞ estimate for P α , θ P_{\alpha,\theta} P α , θ (cf. [7, Theorem 3.2}),
∥ P α , θ ( t − s ) f ∥ ∞ ≤ C ( t − s ) − N α / θ ∥ f ∥ 1 , \|P_{\alpha,\theta}(t-s)f\|_\infty\le C(t-s)^{-N\alpha/\theta}\|f\|_1, ∥ P α , θ ( t − s ) f ∥ ∞ ≤ C ( t − s ) − N α / θ ∥ f ∥ 1 ,
which is valid under assumption (S). On [ t / 2 , t ] [t/2,t] [ t /2 , t ] we use the uniform L ∞ L^\infty L ∞ bound. We obtain
∥ N [ u ] ( t ) ∥ ∞ ≤ C ∫ 0 t / 2 ( t − s ) α − 1 − β h ( s ) ∥ ( u ( s ) ) p ∥ L 1 d s + C ∫ t / 2 t ( t − s ) α − 1 h ( s ) ∥ ( u ( s ) ) p ∥ ∞ d s ≤ C M p ( I 1 ( t ) + I 2 ( t ) ) , \begin{aligned}
\|\mathcal{N}[u](t)\|_\infty
&\le C\int_0^{t/2}(t-s)^{\alpha-1-\beta}h(s)\|(u(s))^p\|_{L^1}\,ds\\
&\quad+C\int_{t/2}^t(t-s)^{\alpha-1}h(s)\|(u(s))^p\|_\infty\,ds\\
&\le CM^p(I_1(t)+I_2(t)),
\end{aligned} ∥ N [ u ] ( t ) ∥ ∞ ≤ C ∫ 0 t /2 ( t − s ) α − 1 − β h ( s ) ∥ ( u ( s ) ) p ∥ L 1 d s + C ∫ t /2 t ( t − s ) α − 1 h ( s ) ∥ ( u ( s ) ) p ∥ ∞ d s ≤ C M p ( I 1 ( t ) + I 2 ( t )) ,
where
I 1 ( t ) : = ∫ 0 t / 2 ( t − s ) α − 1 − β h ( s ) ( 1 + s ) − β ( p − 1 ) d s and I 2 ( t ) : = ∫ t / 2 t ( t − s ) α − 1 h ( s ) ( 1 + s ) − β p d s . I_1(t):=\int_0^{t/2}(t-s)^{\alpha-1-\beta}h(s)(1+s)^{-\beta(p-1)}\,ds
\quad\text{and}\quad
I_2(t):=\int_{t/2}^t(t-s)^{\alpha-1}h(s)(1+s)^{-\beta p}\,ds. I 1 ( t ) := ∫ 0 t /2 ( t − s ) α − 1 − β h ( s ) ( 1 + s ) − β ( p − 1 ) d s and I 2 ( t ) := ∫ t /2 t ( t − s ) α − 1 h ( s ) ( 1 + s ) − β p d s .
We first estimate I 1 ( t ) I_1(t) I 1 ( t ) . Since t − s ≥ t / 2 t-s\ge t/2 t − s ≥ t /2 on [ 0 , t / 2 ] [0,t/2] [ 0 , t /2 ] and α − 1 − β < 0 \alpha-1-\beta<0 α − 1 − β < 0 , we have
I 1 ( t ) ≤ C t α − 1 − β ∫ 0 t / 2 h ( s ) ( 1 + s ) − β ( p − 1 ) d s . I_1(t)\le Ct^{\alpha-1-\beta}\int_0^{t/2}h(s)(1+s)^{-\beta(p-1)}\,ds. I 1 ( t ) ≤ C t α − 1 − β ∫ 0 t /2 h ( s ) ( 1 + s ) − β ( p − 1 ) d s .
We split the last integral at t 0 t_0 t 0 . As h h h is bounded on [ 0 , t 0 ] [0,t_0] [ 0 , t 0 ] and t 0 t_0 t 0 is fixed,
∫ 0 t 0 h ( s ) ( 1 + s ) − β ( p − 1 ) d s ≤ C ( t 0 ) = : C . \int_0^{t_0}h(s)(1+s)^{-\beta(p-1)}\,ds\le C(t_0)=:C. ∫ 0 t 0 h ( s ) ( 1 + s ) − β ( p − 1 ) d s ≤ C ( t 0 ) =: C .
On the other hand, h ( s ) = s γ ℓ ( s ) h(s)=s^\gamma\ell(s) h ( s ) = s γ ℓ ( s ) for s ≥ t 0 s\ge t_0 s ≥ t 0 , and γ + ε − β ( p − 1 ) = κ + ε − α < − α \gamma+\varepsilon-\beta(p-1)=\kappa+\varepsilon-\alpha<-\alpha γ + ε − β ( p − 1 ) = κ + ε − α < − α . Using (25) , we infer that
∫ t 0 t / 2 h ( s ) ( 1 + s ) − β ( p − 1 ) d s ≤ C ∫ t 0 t / 2 s γ + ε − β ( p − 1 ) d s ≤ C ( 1 + t 1 − α ) . \int_{t_0}^{t/2} h(s)(1+s)^{-\beta(p-1)}\,\mathrm{d}s \le C\int_{t_0}^{t/2}s^{\gamma+\varepsilon-\beta(p-1)}\,\mathrm{d}s \le C(1+t^{1-\alpha}). ∫ t 0 t /2 h ( s ) ( 1 + s ) − β ( p − 1 ) d s ≤ C ∫ t 0 t /2 s γ + ε − β ( p − 1 ) d s ≤ C ( 1 + t 1 − α ) .
Indeed, set σ : = γ + ε − β ( p − 1 ) < − α \sigma:=\gamma+\varepsilon-\beta(p-1)<-\alpha σ := γ + ε − β ( p − 1 ) < − α . If σ < − 1 \sigma<-1 σ < − 1 , the integral is bounded. If σ = − 1 \sigma=-1 σ = − 1 , which can happen only when α < 1 \alpha<1 α < 1 , it is bounded by log t ≤ C t 1 − α \log t\le Ct^{1-\alpha} log t ≤ C t 1 − α . If − 1 < σ < − α -1<\sigma<-\alpha − 1 < σ < − α , it is bounded by C t σ + 1 ≤ C t 1 − α Ct^{\sigma+1}\le Ct^{1-\alpha} C t σ + 1 ≤ C t 1 − α . For α = 1 \alpha=1 α = 1 the integral is in fact convergent and the bound is a constant. In either case
I 1 ( t ) ≤ C t α − 1 − β ( 1 + t 1 − α ) ≤ C t − β ≤ C ( 1 + t ) − β , I_1(t)\le Ct^{\alpha-1-\beta}(1+t^{1-\alpha})\le Ct^{-\beta}\le C(1+t)^{-\beta}, I 1 ( t ) ≤ C t α − 1 − β ( 1 + t 1 − α ) ≤ C t − β ≤ C ( 1 + t ) − β ,
where we used α ≤ 1 \alpha\le1 α ≤ 1 and t ≥ 1 t\ge1 t ≥ 1 . Next, we estimate I 2 ( t ) I_2(t) I 2 ( t ) , again using h ( s ) = s γ ℓ ( s ) h(s)=s^\gamma\ell(s) h ( s ) = s γ ℓ ( s ) and (25) for s ≥ t / 2 ≥ t 0 s\ge t/2\ge t_0 s ≥ t /2 ≥ t 0 , together with s ≍ t s\asymp t s ≍ t on [ t / 2 , t ] [t/2,t] [ t /2 , t ] :
I 2 ( t ) = ∫ t / 2 t ( t − s ) α − 1 s γ ℓ ( s ) ( 1 + s ) − β p d s ≤ C ∫ t / 2 t ( t − s ) α − 1 s γ + ε − β p d s ≤ C t γ + ε − β p ∫ t / 2 t ( t − s ) α − 1 d s ≤ C t γ + ε − β p + α = C t κ + ε t − β ≤ C t − β ≤ C ( 1 + t ) − β , \begin{aligned}
I_2(t)&=\int_{t/2}^{t}(t-s)^{\alpha-1}s^\gamma\ell(s)(1+s)^{-\beta p}\,\mathrm{d}s\\
&\le C\int_{t/2}^{t}(t-s)^{\alpha-1}s^{\gamma+\varepsilon-\beta p}\,\mathrm{d}s\\
&\le Ct^{\gamma+\varepsilon-\beta p}\int_{t/2}^{t}(t-s)^{\alpha-1}\,\mathrm{d}s\\
&\le Ct^{\gamma+\varepsilon-\beta p+\alpha}=Ct^{\kappa+\varepsilon}t^{-\beta}\\
&\le Ct^{-\beta}\le C(1+t)^{-\beta},
\end{aligned} I 2 ( t ) = ∫ t /2 t ( t − s ) α − 1 s γ ℓ ( s ) ( 1 + s ) − β p d s ≤ C ∫ t /2 t ( t − s ) α − 1 s γ + ε − β p d s ≤ C t γ + ε − β p ∫ t /2 t ( t − s ) α − 1 d s ≤ C t γ + ε − β p + α = C t κ + ε t − β ≤ C t − β ≤ C ( 1 + t ) − β ,
since κ + ε < 0 \kappa+\varepsilon<0 κ + ε < 0 . Combining the estimates of I 1 ( t ) I_1(t) I 1 ( t ) and I 2 ( t ) I_2(t) I 2 ( t ) with (26) , we conclude that, for all t > 0 t>0 t > 0 ,
∥ N [ u ] ( t ) ∥ ∞ ≤ C M p ( 1 + t ) − β . (27) \left\|\mathcal{N}[u](t)\right\|_{\infty}\le CM^p(1+t)^{-\beta}. \tag*{(27)} ∥ N [ u ] ( t ) ∥ ∞ ≤ C M p ( 1 + t ) − β . (27)
To estimate ∥ N [ u ] ( t ) ∥ L 1 \left\|\mathcal{N}[u](t)\right\|_{L^1} ∥ N [ u ] ( t ) ∥ L 1 , we use the L 1 L^1 L 1 bound ∥ P α , θ ( t − s ) f ∥ 1 ≤ C ∥ f ∥ 1 \|P_{\alpha,\theta}(t-s)f\|_1\le C\|f\|_1 ∥ P α , θ ( t − s ) f ∥ 1 ≤ C ∥ f ∥ 1 and get
∥ N [ u ] ( t ) ∥ 1 ≤ C M p ∫ 0 t ( t − s ) α − 1 h ( s ) ( 1 + s ) − β ( p − 1 ) d s . \left\|\mathcal{N}[u](t)\right\|_1\le CM^p\int_0^t(t-s)^{\alpha-1}h(s)(1+s)^{-\beta(p-1)}\,\mathrm{d}s. ∥ N [ u ] ( t ) ∥ 1 ≤ C M p ∫ 0 t ( t − s ) α − 1 h ( s ) ( 1 + s ) − β ( p − 1 ) d s .
For t ≤ 2 t 0 t\le2t_0 t ≤ 2 t 0 the right-hand side is bounded by C M p CM^p C M p , since h h h is bounded on [ 0 , 2 t 0 ] [0,2t_0] [ 0 , 2 t 0 ] . For t ≥ 2 t 0 t\ge2t_0 t ≥ 2 t 0 we split at t / 2 t/2 t /2 exactly as above. The contribution of [ 0 , t / 2 ] [0,t/2] [ 0 , t /2 ] is at most C t α − 1 ( 1 + t 1 − α ) ≤ C Ct^{\alpha-1}(1+t^{1-\alpha})\le C C t α − 1 ( 1 + t 1 − α ) ≤ C , while that of [ t / 2 , t ] [t/2,t] [ t /2 , t ] is at most C t γ + ε − β ( p − 1 ) + α = C t κ + ε ≤ C Ct^{\gamma+\varepsilon-\beta(p-1)+\alpha}=Ct^{\kappa+\varepsilon}\le C C t γ + ε − β ( p − 1 ) + α = C t κ + ε ≤ C . Hence
∥ N [ u ] ( t ) ∥ L 1 ≤ C M p . (28) \left\|\mathcal{N}[u](t)\right\|_{L^1}\le CM^p. \tag*{(28)} ∥ N [ u ] ( t ) ∥ L 1 ≤ C M p . (28)
The marginal regime κ = 0 \kappa=0 κ = 0 . Here there is no room left for ε \varepsilon ε , and the slowly varying factor itself decides whether the scheme closes. We treat the classical and the genuinely fractional cases separately.
α = 1 \alpha=1 α = 1 , p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) , under ∫ t 0 ∞ t − 1 ℓ ( t ) d t < ∞ \int_{t_0}^{\infty}t^{-1}\ell(t)\,\mathrm{d}t<\infty ∫ t 0 ∞ t − 1 ℓ ( t ) d t < ∞ . Since β ( p − 1 ) = γ + 1 > 0 \beta(p-1)=\gamma+1>0 β ( p − 1 ) = γ + 1 > 0 , the large-time density satisfies h ( s ) ( 1 + s ) − β ( p − 1 ) ≤ s − 1 ℓ ( s ) h(s)(1+s)^{-\beta(p-1)}\le s^{-1}\ell(s) h ( s ) ( 1 + s ) − β ( p − 1 ) ≤ s − 1 ℓ ( s ) , so that
∫ t 0 t / 2 h ( s ) ( 1 + s ) − β ( p − 1 ) d s ≤ ∫ t 0 ∞ s − 1 ℓ ( s ) d s < ∞ , \int_{t_0}^{t/2}h(s)(1+s)^{-\beta(p-1)}\,\mathrm{d}s\le\int_{t_0}^{\infty}s^{-1}\ell(s)\,\mathrm{d}s<\infty, ∫ t 0 t /2 h ( s ) ( 1 + s ) − β ( p − 1 ) d s ≤ ∫ t 0 ∞ s − 1 ℓ ( s ) d s < ∞ ,
and I 1 ( t ) ≤ C t α − 1 − β = C t − β I_1(t)\le Ct^{\alpha-1-\beta}=Ct^{-\beta} I 1 ( t ) ≤ C t α − 1 − β = C t − β . For I 2 I_2 I 2 we cannot afford any positive power. Instead, by Lemma 2.1(K4), the convergence of the borderline integral forces ℓ ( t ) → 0 \ell(t)\to0 ℓ ( t ) → 0 . Hence ℓ \ell ℓ is bounded on [ t 0 , ∞ ) [t_0,\infty) [ t 0 , ∞ ) and
I 2 ( t ) ≤ C ( sup [ t / 2 , t ] ℓ ) t γ − β p t α = C ( sup [ t / 2 , t ] ℓ ) t κ t − β ≤ C t − β . I_2(t)\le C\left(\sup_{[t/2,t]}\ell\right)t^{\gamma-\beta p}t^\alpha=C\left(\sup_{[t/2,t]}\ell\right)t^\kappa t^{-\beta}\le Ct^{-\beta}. I 2 ( t ) ≤ C ( [ t /2 , t ] sup ℓ ) t γ − β p t α = C ( [ t /2 , t ] sup ℓ ) t κ t − β ≤ C t − β .
The L 1 L^1 L 1 estimate is handled in the same way. Thus the scheme closes at the critical exponent, which proves the global-existence half of Theorem 3.1(iii).
α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , N < θ N < \theta N < θ , p = p c ( γ , α ) p = p_c(\gamma,\alpha) p = p c ( γ , α ) , with ℓ \ell ℓ bounded. Now β ( p − 1 ) = γ + α \beta(p-1)=\gamma+\alpha β ( p − 1 ) = γ + α , and the large-time density is bounded by s − α ℓ ( s ) ≤ C s − α s^{-\alpha}\ell(s)\le Cs^{-\alpha} s − α ℓ ( s ) ≤ C s − α . Therefore ∫ t 0 t / 2 s − α ℓ ( s ) d s ≤ C t 1 − α \int_{t_0}^{t/2}s^{-\alpha}\ell(s)\,ds\le Ct^{1-\alpha} ∫ t 0 t /2 s − α ℓ ( s ) d s ≤ C t 1 − α and I 1 ( t ) ≤ C t − β I_1(t)\le Ct^{-\beta} I 1 ( t ) ≤ C t − β , while I 2 ( t ) ≤ C ( sup ℓ ) t − β = C t − β I_2(t)\le C(\sup\ell)t^{-\beta}=Ct^{-\beta} I 2 ( t ) ≤ C ( sup ℓ ) t − β = C t − β . The L 1 L^1 L 1 estimate is again identical. This proves the critical-case global existence in Theorem 3.2 (b), which is consistent with the autonomous critical-case global existence of [42 , 7 ] (γ = 0 \gamma=0 γ = 0 , ℓ ≡ 1 \ell\equiv1 ℓ ≡ 1 ).
Fixed-point argument. The estimates above show that there is a constant C ∗ C_* C ∗ , independent of u 0 u_0 u 0 , such that ∥ N [ u ] ∥ X β ≤ C ∗ ∥ u ∥ X β p \|\mathcal{N}[u]\|_{X_\beta}\le C_*\|u\|_{X_\beta}^p ∥ N [ u ] ∥ X β ≤ C ∗ ∥ u ∥ X β p for every u ∈ X β + u\in X_\beta^+ u ∈ X β + . Moreover, the elementary inequality ∣ a p − b p ∣ ≤ p ( a ∨ b ) p − 1 ∣ a − b ∣ |a^p-b^p|\le p(a\vee b)^{p-1}|a-b| ∣ a p − b p ∣ ≤ p ( a ∨ b ) p − 1 ∣ a − b ∣ for a , b ≥ 0 a,b\ge0 a , b ≥ 0 gives ∥ u ( s ) p − v ( s ) p ∥ L 1 ≤ p M p − 1 ( 1 + s ) − β ( p − 1 ) ∥ u − v ∥ X β \|u(s)^p-v(s)^p\|_{L^1}\le pM^{p-1}(1+s)^{-\beta(p-1)}\|u-v\|_{X_\beta} ∥ u ( s ) p − v ( s ) p ∥ L 1 ≤ p M p − 1 ( 1 + s ) − β ( p − 1 ) ∥ u − v ∥ X β and ∥ u ( s ) p − v ( s ) p ∥ L ∞ ≤ p M p − 1 ( 1 + s ) − β p ∥ u − v ∥ X β \|u(s)^p-v(s)^p\|_{L^\infty}\le pM^{p-1}(1+s)^{-\beta p}\|u-v\|_{X_\beta} ∥ u ( s ) p − v ( s ) p ∥ L ∞ ≤ p M p − 1 ( 1 + s ) − β p ∥ u − v ∥ X β whenever ∥ u ∥ X β , ∥ v ∥ X β ≤ M \|u\|_{X_\beta},\|v\|_{X_\beta}\le M ∥ u ∥ X β , ∥ v ∥ X β ≤ M . These are the same densities as above, so that ∥ N [ u ] − N [ v ] ∥ X β ≤ p C ∗ M p − 1 ∥ u − v ∥ X β \|\mathcal{N}[u]-\mathcal{N}[v]\|_{X_\beta}\le pC_*M^{p-1}\|u-v\|_{X_\beta} ∥ N [ u ] − N [ v ] ∥ X β ≤ p C ∗ M p − 1 ∥ u − v ∥ X β . Let M : = 2 C 0 ( ∥ u 0 ∥ 1 + ∥ u 0 ∥ ∞ ) M:=2C_0(\|u_0\|_1+\|u_0\|_\infty) M := 2 C 0 ( ∥ u 0 ∥ 1 + ∥ u 0 ∥ ∞ ) , and assume that u 0 u_0 u 0 is so small that p C ∗ M p − 1 ≤ 1 / 2 pC_*M^{p-1}\le1/2 p C ∗ M p − 1 ≤ 1/2 . Then, for u u u in the closed set B M + : = { u ∈ X β + : ∥ u ∥ X β ≤ M } B_M^+:=\{u\in X_\beta^+:\|u\|_{X_\beta}\le M\} B M + := { u ∈ X β + : ∥ u ∥ X β ≤ M } , we have ∥ Λ [ u ] ∥ X β ≤ M / 2 + C ∗ M p ≤ M \|\Lambda[u]\|_{X_\beta}\le M/2+C_*M^p\le M ∥Λ [ u ] ∥ X β ≤ M /2 + C ∗ M p ≤ M . Thus Λ \Lambda Λ maps B M + B_M^+ B M + into itself and is a contraction there, with constant 1 / 2 1/2 1/2 . The Banach fixed-point theorem produces a unique fixed point. By uniqueness of mild solutions, it coincides with the maximal mild solution, which is therefore global and satisfies ∥ u ( t ) ∥ ∞ ≤ M ( 1 + t ) − β → 0 \|u(t)\|_\infty\le M(1+t)^{-\beta}\to0 ∥ u ( t ) ∥ ∞ ≤ M ( 1 + t ) − β → 0 as t → ∞ t\to\infty t → ∞ .
Summary of the closure region. The scheme closes, and hence small-data global existence holds with decay ( 1 + t ) − N α / θ (1+t)^{-N\alpha/\theta} ( 1 + t ) − N α / θ under hypothesis (S), precisely when
κ < 0 , or κ = 0 with { ∫ ∞ t − 1 ℓ ( t ) d t < ∞ , α = 1 , ℓ bounded , α < 1. \kappa<0,\qquad\text{or}\qquad\kappa=0\ \text{with}\ \begin{cases}
\int^\infty t^{-1}\ell(t)\,dt<\infty,&\alpha=1,\\
\ell\ \text{bounded},&\alpha<1.
\end{cases} κ < 0 , or κ = 0 with { ∫ ∞ t − 1 ℓ ( t ) d t < ∞ , ℓ bounded , α = 1 , α < 1.
At α = 1 \alpha=1 α = 1 this gives exactly all p > 1 p>1 p > 1 for γ ≤ − 1 \gamma\le-1 γ ≤ − 1 , p > p c ( γ ) p>p_c(\gamma) p > p c ( γ ) for γ > − 1 \gamma>-1 γ > − 1 , and p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) under the convergent Karamata integral; these are the global halves of Theorem 3.1. For α < 1 \alpha<1 α < 1 (with N < θ N<\theta N < θ ) it gives all p > 1 p>1 p > 1 for γ ≤ − α \gamma\le-\alpha γ ≤ − α , and p > p c ( γ , α ) p>p_c(\gamma,\alpha) p > p c ( γ , α ) together with the marginal case with bounded ℓ \ell ℓ for γ > − α \gamma>-\alpha γ > − α ; these are Theorem 3.2 (a)–(b) in the regime N < θ N<\theta N < θ . For α < 1 \alpha<1 α < 1 and N ≥ θ N\ge\theta N ≥ θ , the L r L^r L r variant of Remark 5.1 below closes for every p > 1 + ( γ + α ) / α p>1+(\gamma+\alpha)/\alpha p > 1 + ( γ + α ) / α , with decay ( 1 + t ) − β r (1+t)^{-\beta_r} ( 1 + t ) − β r , β r = N α / ( θ r ) \beta_r=N\alpha/(\theta r) β r = N α / ( θ r ) , which yields the remaining range in Theorem 3.2 (b).
Remark 5.1. When α < 1 \alpha<1 α < 1 and N ≥ θ N\ge\theta N ≥ θ , the endpoint smoothing fails (Proposition 2.1(ii)), and the choice β = N α / θ \beta=N\alpha/\theta β = N α / θ is not available. Instead, we fix a Lebesgue exponent r > max { 1 , N / θ } r>\max\{1,N/\theta\} r > max { 1 , N / θ } and run the same scheme with the norm ess sup t [ ∥ u ( t ) ∥ L r + ( 1 + t ) β r ∥ u ( t ) ∥ L ∞ ] \operatorname*{ess\,sup}_t[\|u(t)\|_{L^r}+(1+t)^{\beta_r}\|u(t)\|_{L^\infty}] ess sup t [ ∥ u ( t ) ∥ L r + ( 1 + t ) β r ∥ u ( t ) ∥ L ∞ ] , where β r : = N α / ( θ r ) < α \beta_r:=N\alpha/(\theta r)<\alpha β r := N α / ( θ r ) < α . The smoothing L r → L ∞ L^r\to L^\infty L r → L ∞ is admissible for both S α , θ S_{\alpha,\theta} S α , θ and P α , θ P_{\alpha,\theta} P α , θ since 1 / r < θ / N < 1 1/r<\theta/N<1 1/ r < θ / N < 1 , and ∥ u p ∥ L r ≤ ∥ u ∥ ∞ p − 1 ∥ u ∥ L r \|u^p\|_{L^r}\le\|u\|_\infty^{p-1}\|u\|_{L^r} ∥ u p ∥ L r ≤ ∥ u ∥ ∞ p − 1 ∥ u ∥ L r . The computation of §5.1 applies verbatim, with L 1 L^1 L 1 replaced by L r L^r L r and β \beta β by β r \beta_r β r ; data in L 1 ∩ L ∞ L^1\cap L^\infty L 1 ∩ L ∞ belong to L r ∩ L ∞ L^r\cap L^\infty L r ∩ L ∞ . The closure condition becomes β r ( p − 1 ) > γ + α \beta_r(p-1)>\gamma+\alpha β r ( p − 1 ) > γ + α . The constraint r ≥ 1 r\ge1 r ≥ 1 is decisive, and two regimes must be distinguished according to the position of N / θ N/\theta N / θ relative to 1 1 1 . If N ≥ θ N\ge\theta N ≥ θ , the admissible exponents approach N / θ ≥ 1 N/\theta\ge1 N / θ ≥ 1 from above. Letting r ↓ N / θ r\downarrow N/\theta r ↓ N / θ , so that β r ↑ α \beta_r\uparrow\alpha β r ↑ α , we obtain small-data global existence for every p > 1 + ( γ + α ) / α p>1+(\gamma+\alpha)/\alpha p > 1 + ( γ + α ) / α . If instead N < θ N<\theta N < θ , the constraint r ≥ 1 r\ge1 r ≥ 1 is binding. The smallest admissible exponent is r = 1 r=1 r = 1 , for which β 1 = N α / θ \beta_1=N\alpha/\theta β 1 = N α / θ , and the scheme reproduces exactly the range p > p c ( γ , α ) p>p_c(\gamma,\alpha) p > p c ( γ , α ) of Theorem 3.2 (b). The threshold 1 + ( γ + α ) / α 1+(\gamma+\alpha)/\alpha 1 + ( γ + α ) / α , which is strictly smaller than p c ( γ , α ) p_c(\gamma,\alpha) p c ( γ , α ) when N < θ N<\theta N < θ , is then out of reach. In fact it could not be reached, since global existence below p c ( γ , α ) p_c(\gamma,\alpha) p c ( γ , α ) would contradict the blow-up range of Theorem 3.2 (c). For instance, for γ = 0 \gamma=0 γ = 0 , N = 1 N=1 N = 1 , θ = 2 \theta=2 θ = 2 and any α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , it would give global existence for p > 2 p>2 p > 2 , while Theorem 3.2 (c) and the autonomous results of [42 , 7 ] force blow-up for 1 < p < p c ( 0 , α ) = 3 1<p<p_c(0,\alpha)=3 1 < p < p c ( 0 , α ) = 3 .
Finally, when γ ≤ − α \gamma\le-\alpha γ ≤ − α , the closure condition β r ( p − 1 ) > γ + α \beta_r(p-1) > \gamma+ \alpha β r ( p − 1 ) > γ + α holds for every p > 1 p > 1 p > 1 , for every admissible r r r and in every dimension. This proves Theorem 3.2 (a) without any restriction. For N > θ N > \theta N > θ , reaching the conjectured threshold p ∗ ( γ , α ) p_*(\gamma,\alpha) p ∗ ( γ , α ) of Conjecture 3.1 seems to require a finer functional framework, for instance the critical Lebesgue space approach of [42 , 7 ] combined with a control of the local dynamics described in §5.4. We do not pursue this here.
Blow-up: the case α = 1 \alpha= 1 α = 1
In this subsection we investigate the blow-up of nontrivial nonnegative solutions when α = 1 \alpha= 1 α = 1 . We begin with the Kaplan functional method. It gives a simple proof in the subcritical case and illustrates the difficulty that arises in the critical regime.
Let θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] and γ > − 1 \gamma> -1 γ > − 1 , and let u ≥ 0 u \ge0 u ≥ 0 be a global mild solution with u 0 ≥ 0 u_0 \ge0 u 0 ≥ 0 , u 0 ≢ 0 u_0 \not\equiv0 u 0 ≡ 0 . After a translation in x x x , we may assume that
m 0 : = ∫ ∣ x ∣ ≤ 1 u 0 d x > 0. m_0 := \int_{\lvert x\rvert\le1} u_0\,dx > 0. m 0 := ∫ ∣ x ∣ ≤ 1 u 0 d x > 0.
For T > 1 T > 1 T > 1 we define the adjoint weight
φ T ( t , x ) : = Z ( T + 1 − t , x ) , 0 ≤ t ≤ T , \varphi_T(t,x) := Z(T+1-t,x), \qquad0 \le t \le T, φ T ( t , x ) := Z ( T + 1 − t , x ) , 0 ≤ t ≤ T ,
which solves the backward equation ∂ t φ T = ( − Δ ) θ / 2 φ T \partial_t\varphi_T = (-\Delta)^{\theta/2}\varphi_T ∂ t φ T = ( − Δ ) θ /2 φ T and satisfies ∫ R N φ T ( t , x ) d x = 1 \int_{\mathbb{R}^N}\varphi_T(t,x)\,dx = 1 ∫ R N φ T ( t , x ) d x = 1 . We set
ψ ( t ) : = ∫ R N u ( t , x ) φ T ( t , x ) d x , 0 ≤ t ≤ T . \psi(t) := \int_{\mathbb{R}^N} u(t,x)\varphi_T(t,x)\,dx, \qquad0 \le t \le T. ψ ( t ) := ∫ R N u ( t , x ) φ T ( t , x ) d x , 0 ≤ t ≤ T .
Since Z ( t , ⋅ ) Z(t,\cdot) Z ( t , ⋅ ) is even and Z ( t ) ∗ Z ( s ) = Z ( t + s ) Z(t) * Z(s) = Z(t+s) Z ( t ) ∗ Z ( s ) = Z ( t + s ) , Fubini’s theorem and the mild formulation (15) give
ψ ( t ) = ∫ R N u 0 ( x ) Z ( T + 1 , x ) d x + ∫ 0 t h ( s ) ∫ R N u ( s , x ) p φ T ( s , x ) d x d s . \psi(t) = \int_{\mathbb{R}^N} u_0(x)Z(T+1,x)\,dx + \int_0^t h(s)\int_{\mathbb{R}^N} u(s,x)^p\varphi_T(s,x)\,dx\,ds. ψ ( t ) = ∫ R N u 0 ( x ) Z ( T + 1 , x ) d x + ∫ 0 t h ( s ) ∫ R N u ( s , x ) p φ T ( s , x ) d x d s .
Hence ψ \psi ψ is absolutely continuous on [ 0 , T ] [0,T] [ 0 , T ] , and Jensen’s inequality (the measure φ T ( t , x ) d x \varphi_T(t,x)\,dx φ T ( t , x ) d x is a probability measure) yields
ψ ′ ( t ) = h ( t ) ∫ R N ( u ( t , x ) ) p φ T ( t , x ) d x ≥ h ( t ) ψ ( t ) p for a.e. t ∈ ( 0 , T ) . \psi'(t) = h(t)\int_{\mathbb{R}^N} \bigl(u(t,x)\bigr)^p\varphi_T(t,x)\,dx \ge h(t)\psi(t)^p \qquad\text{for a.e. }t \in(0,T). ψ ′ ( t ) = h ( t ) ∫ R N ( u ( t , x ) ) p φ T ( t , x ) d x ≥ h ( t ) ψ ( t ) p for a.e. t ∈ ( 0 , T ) .
Since u u u is global and ψ > 0 \psi> 0 ψ > 0 , integrating the inequality − 1 p − 1 d d t ψ 1 − p ≥ h -\frac{1}{p-1}\frac{d}{dt}\psi^{1-p} \ge h − p − 1 1 d t d ψ 1 − p ≥ h over [ 0 , T ] [0,T] [ 0 , T ] gives
( p − 1 ) ∫ 0 T h ( s ) d s ≤ ψ ( 0 ) 1 − p . (29) (p-1)\int_0^T h(s)\,ds \le\psi(0)^{1-p}. \tag*{(29)} ( p − 1 ) ∫ 0 T h ( s ) d s ≤ ψ ( 0 ) 1 − p . (29)
The θ \theta θ -stable kernel satisfies the lower estimate (12) , namely
Z ( s , x ) ≥ c s − N / θ , ∣ x ∣ ≤ s 1 / θ . Z(s,x) \ge c\,s^{-N/\theta}, \qquad\lvert x\rvert\le s^{1/\theta}. Z ( s , x ) ≥ c s − N / θ , ∣ x ∣ ≤ s 1/ θ .
Since ∣ x ∣ ≤ 1 ≤ ( T + 1 ) 1 / θ \lvert x\rvert\le1 \le(T+1)^{1/\theta} ∣ x ∣ ≤ 1 ≤ ( T + 1 ) 1/ θ on the support of the integral defining m 0 m_0 m 0 , we deduce that
ψ ( 0 ) = ∫ R N u 0 ( x ) Z ( T + 1 , x ) d x ≥ c m 0 ( T + 1 ) − N / θ ≥ c 2 − N / θ m 0 T − N / θ . \psi(0) = \int_{\mathbb{R}^N} u_0(x)Z(T+1,x)\,dx \ge c\,m_0(T+1)^{-N/\theta} \ge c\,2^{-N/\theta}m_0T^{-N/\theta}. ψ ( 0 ) = ∫ R N u 0 ( x ) Z ( T + 1 , x ) d x ≥ c m 0 ( T + 1 ) − N / θ ≥ c 2 − N / θ m 0 T − N / θ .
On the other hand, by Lemma 2.1(K2) there is T 1 > 1 T_1 > 1 T 1 > 1 such that, for T ≥ T 1 T \ge T_1 T ≥ T 1 ,
∫ 0 T h ( t ) d t ≥ ∫ 1 T t γ ℓ ( t ) d t ≥ T γ + 1 ℓ ( T ) 2 ( γ + 1 ) . \int_0^T h(t)\,dt \ge\int_1^T t^\gamma\ell(t)\,dt \ge\frac{T^{\gamma+1}\ell(T)}{2(\gamma+1)}. ∫ 0 T h ( t ) d t ≥ ∫ 1 T t γ ℓ ( t ) d t ≥ 2 ( γ + 1 ) T γ + 1 ℓ ( T ) .
Consequently, the necessary condition (29) becomes
sup T ≥ T 1 T γ + 1 − N ( p − 1 ) θ ℓ ( T ) ≤ C ( m 0 ) < ∞ . (30) \sup_{T \ge T_1} T^{\gamma+1-\frac{N(p-1)}{\theta}}\ell(T) \le C(m_0) < \infty. \tag*{(30)} T ≥ T 1 sup T γ + 1 − θ N ( p − 1 ) ℓ ( T ) ≤ C ( m 0 ) < ∞. (30)
For 1 < p < p c ( γ ) 1 < p < p_c(\gamma) 1 < p < p c ( γ ) , the exponent in (30) is positive. Hence Lemma 2.1(K1) shows that (30) cannot hold, and every nontrivial nonnegative solution blows up in finite time. This proves the blow-up part of Theorem 3.1(ii). In the critical case p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) , the power in (30) vanishes, and the argument yields a contradiction only when
lim sup T → ∞ ℓ ( T ) = ∞ . \limsup_{T\to\infty}\ell(T)=\infty. T → ∞ lim sup ℓ ( T ) = ∞.
Therefore the Kaplan functional method does not cover all critical cases. The remaining critical cases, as well as an alternative proof of the subcritical case, follow from [13 Corollary 3.1] .
Theorem 5.1. Let α = 1 \alpha=1 α = 1 , θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] , γ > − 1 \gamma>-1 γ > − 1 , p ∈ ( 1 , p c ( γ ) ] p\in(1,p_c(\gamma)] p ∈ ( 1 , p c ( γ )] and q ∈ [ 1 , ∞ ] q\in[1,\infty] q ∈ [ 1 , ∞ ] . In the critical case p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) , assume in addition that
∫ 1 ∞ s − 1 ℓ ( s ) d s = ∞ . \int_1^\infty s^{-1}\ell(s)\,ds=\infty. ∫ 1 ∞ s − 1 ℓ ( s ) d s = ∞. Let u u u be a nontrivial nonnegative L q L^q L q -solution of (1) . Then there exists a finite time T ∈ ( 0 , ∞ ) T\in(0,\infty) T ∈ ( 0 , ∞ ) such that
lim sup t → T ∥ u ( t , ⋅ ) ∥ L q ( R N ) = ∞ . \limsup_{t\to T}\|u(t,\cdot)\|_{L^q(\mathbb{R}^N)}=\infty. t → T lim sup ∥ u ( t , ⋅ ) ∥ L q ( R N ) = ∞.
Proof. Suppose first that p < p c ( γ ) p<p_c(\gamma) p < p c ( γ ) . Then − N ( p − 1 ) / θ + γ > − 1 -N(p-1)/\theta+\gamma>-1 − N ( p − 1 ) / θ + γ > − 1 , and Lemma 2.1(K2) gives
∫ 1 t s − N ( p − 1 ) θ h ( s ) d s = ∫ 1 t s − N ( p − 1 ) θ + γ ℓ ( s ) d s ∼ t − N ( p − 1 ) θ + γ + 1 ℓ ( t ) − N ( p − 1 ) θ + γ + 1 , t → ∞ . \int_1^t s^{-\frac{N(p-1)}{\theta}}h(s)\,ds=\int_1^t s^{-\frac{N(p-1)}{\theta}+\gamma}\ell(s)\,ds\sim\frac{t^{-\frac{N(p-1)}{\theta}+\gamma+1}\ell(t)}{-\frac{N(p-1)}{\theta}+\gamma+1},\qquad t\to\infty. ∫ 1 t s − θ N ( p − 1 ) h ( s ) d s = ∫ 1 t s − θ N ( p − 1 ) + γ ℓ ( s ) d s ∼ − θ N ( p − 1 ) + γ + 1 t − θ N ( p − 1 ) + γ + 1 ℓ ( t ) , t → ∞.
Since − N ( p − 1 ) θ + γ + 1 > 0 -\frac{N(p-1)}{\theta}+\gamma+1>0 − θ N ( p − 1 ) + γ + 1 > 0 , Lemma 2.1(K1) yields t − N ( p − 1 ) θ + γ + 1 ℓ ( t ) → ∞ t^{-\frac{N(p-1)}{\theta}+\gamma+1}\ell(t)\to\infty t − θ N ( p − 1 ) + γ + 1 ℓ ( t ) → ∞ as t → ∞ t\to\infty t → ∞ . Hence
lim t → ∞ ∫ 1 t s − N ( p − 1 ) θ h ( s ) d s = ∞ , \lim_{t\to\infty}\int_1^t s^{-\frac{N(p-1)}{\theta}}h(s)\,ds=\infty, t → ∞ lim ∫ 1 t s − θ N ( p − 1 ) h ( s ) d s = ∞ ,
and the conclusion follows from [13 Corollary 3.1] . In the critical case p = p c ( γ ) p=p_c(\gamma) p = p c ( γ ) we have − N ( p − 1 ) / θ + γ = − 1 -N(p-1)/\theta+\gamma=-1 − N ( p − 1 ) / θ + γ = − 1 , so that
lim t → ∞ ∫ 1 t s − N ( p − 1 ) θ h ( s ) d s = ∫ 1 ∞ s − 1 ℓ ( s ) d s = ∞ \lim_{t\to\infty}\int_1^t s^{-\frac{N(p-1)}{\theta}}h(s)\,ds=\int_1^\infty s^{-1}\ell(s)\,ds=\infty t → ∞ lim ∫ 1 t s − θ N ( p − 1 ) h ( s ) d s = ∫ 1 ∞ s − 1 ℓ ( s ) d s = ∞
by assumption. The desired conclusion again follows from [13 Corollary 3.1] . □
Remark 5.2. The critical half of Theorem 3.1(iii) rests entirely on [13 Corollary 3.1] . Before submission, one should check against the original paper that this corollary (a) covers every θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] and a continuous nonnegative coefficient h ( t ) h(t) h ( t ) , (b) is formulated under the divergence condition ∫ 1 ∞ s − N ( p − 1 ) / θ h ( s ) d s = ∞ \int_1^\infty s^{-N(p-1)/\theta}h(s)\,ds=\infty ∫ 1 ∞ s − N ( p − 1 ) / θ h ( s ) d s = ∞ (and not, for instance, under a lim sup condition of Meier type), and (c) applies to the solutions of the integral equation (15) with data in L 1 ∩ L ∞ L^1\cap L^\infty L 1 ∩ L ∞ (or in L q L^q L q ). If any of these points fails, a self-contained proof of the critical case is needed. A natural route combines the Kaplan inequality on [ T / 2 , T ] [T/2,T] [ T /2 , T ] with the lower bound ψ ( T / 2 ) ≥ c T − N / θ ( m 0 + c m 0 p ∫ 1 T / 2 s − 1 ℓ ( s ) d s ) \psi(T/2)\ge cT^{-N/\theta}(m_0+cm_0^p\int_1^{T/2}s^{-1}\ell(s)\,ds) ψ ( T /2 ) ≥ c T − N / θ ( m 0 + c m 0 p ∫ 1 T /2 s − 1 ℓ ( s ) d s ) , followed by a bootstrap in the spirit of Weissler [40 ] .
Blow-up: the case 0 < α < 1 0<\alpha<1 0 < α < 1
Our proof follows that of [7 ] closely, with two main modifications. First, the detailed analysis in [7 ] is written for N > 2 θ N>2\theta N > 2 θ ; by relying only on a positive lower bound for the kernel Y Y Y on balls, which is provided by [17 Lemma 3.6] (Lemma 4.6 in the arXiv version), the argument works in every dimension N ≥ 1 N\ge1 N ≥ 1 . Second, we allow a time-dependent source h ( t ) = t γ ℓ ( t ) h(t)=t^\gamma\ell(t) h ( t ) = t γ ℓ ( t ) by working with the functional J J J defined in (33) , whose weight is adapted to the source.
Theorem 5.2. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] , γ ∈ R \gamma\in\mathbb{R} γ ∈ R , p ∈ ( 1 , ∞ ) p\in(1,\infty) p ∈ ( 1 , ∞ ) and q ∈ [ 1 , ∞ ] q\in[1,\infty] q ∈ [ 1 , ∞ ] . There exist R 0 ≥ 1 R_0\ge1 R 0 ≥ 1 , depending only on γ \gamma γ and ℓ \ell ℓ , and a constant C = C ( N , α , θ , γ , p ) > 0 C=C(N,\alpha,\theta,\gamma,p)>0 C = C ( N , α , θ , γ , p ) > 0 with the following property. If u u u is a nontrivial nonnegative L q L^q L q -solution of (1) whose initial datum satisfies
R α + γ p − 1 − N α θ ( ℓ ( R ) ) 1 p − 1 ∫ ∣ y ∣ < R α θ / 2 u 0 ( y ) d y > C for some R ≥ R 0 , (31) R^{\frac{\alpha+\gamma}{p-1}-\frac{N\alpha}{\theta}}\bigl(\ell(R)\bigr)^{\frac{1}{p-1}}\int_{|y|<R^{\frac{\alpha}{\theta}}/2}u_0(y)\,dy>C\qquad\text{for some }R\ge R_0, \tag*{(31)} R p − 1 α + γ − θ N α ( ℓ ( R ) ) p − 1 1 ∫ ∣ y ∣ < R θ α /2 u 0 ( y ) d y > C for some R ≥ R 0 , (31)
then there exists T ∈ ( 0 , ∞ ) T \in(0,\infty) T ∈ ( 0 , ∞ ) (in fact T ≤ 4 R T \le4R T ≤ 4 R ) such that
lim sup t → T ∥ u ( t , ⋅ ) ∥ L q ( R N ) = ∞ . \limsup_{t\to T}\|u(t,\cdot)\|_{L^q(\mathbb{R}^N)}=\infty. t → T lim sup ∥ u ( t , ⋅ ) ∥ L q ( R N ) = ∞.
Corollary 5.1. Let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] , γ > − α \gamma> -\alpha γ > − α , p ∈ ( 1 , p c ( γ , α ) ] p \in(1,p_c(\gamma,\alpha)] p ∈ ( 1 , p c ( γ , α )] and q ∈ [ 1 , ∞ ] q \in[1,\infty] q ∈ [ 1 , ∞ ] . In the critical case p = p c ( γ , α ) p=p_c(\gamma,\alpha) p = p c ( γ , α ) , assume in addition that lim sup t → ∞ ℓ ( t ) = ∞ \limsup_{t\to\infty}\ell(t)=\infty lim sup t → ∞ ℓ ( t ) = ∞ . Let u u u be a nontrivial nonnegative L q L^q L q -solution of (1) . Then there exists a finite time T ∈ ( 0 , ∞ ) T \in(0,\infty) T ∈ ( 0 , ∞ ) such that
lim sup t → T ∥ u ( t , ⋅ ) ∥ L q ( R N ) = ∞ . \limsup_{t\to T}\|u(t,\cdot)\|_{L^q(\mathbb{R}^N)}=\infty. t → T lim sup ∥ u ( t , ⋅ ) ∥ L q ( R N ) = ∞.
Proof. Since u 0 u_0 u 0 is nontrivial and nonnegative, there exists R 1 > 0 R_1>0 R 1 > 0 such that
∫ ∣ y ∣ < R 1 θ 2 u 0 ( y ) d y = B > 0 , \int_{|y|<R_1^{\frac{\theta}{2}}}u_0(y)\,dy=B>0, ∫ ∣ y ∣ < R 1 2 θ u 0 ( y ) d y = B > 0 ,
and the integral in (31) is at least B B B for every R ≥ R 1 R\ge R_1 R ≥ R 1 . First, assume that p < p c ( γ , α ) p<p_c(\gamma,\alpha) p < p c ( γ , α ) . Then
α + γ p − 1 − N α θ > 0 , \frac{\alpha+\gamma}{p-1}-\frac{N\alpha}{\theta}>0, p − 1 α + γ − θ N α > 0 ,
and Lemma 2.1(K1) gives
R α + γ p − 1 − N α θ ( ℓ ( R ) ) 1 p − 1 → ∞ as R → ∞ . R^{\frac{\alpha+\gamma}{p-1}-\frac{N\alpha}{\theta}}\bigl(\ell(R)\bigr)^{\frac{1}{p-1}}\to\infty
\quad\text{as }R\to\infty. R p − 1 α + γ − θ N α ( ℓ ( R ) ) p − 1 1 → ∞ as R → ∞.
Next, suppose that p = p c ( γ , α ) p=p_c(\gamma,\alpha) p = p c ( γ , α ) . In this case
α + γ p − 1 − N α θ = 0 , \frac{\alpha+\gamma}{p-1}-\frac{N\alpha}{\theta}=0, p − 1 α + γ − θ N α = 0 ,
and, by assumption, lim sup R → ∞ ℓ ( R ) = ∞ \limsup_{R\to\infty}\ell(R)=\infty lim sup R → ∞ ℓ ( R ) = ∞ . In both cases, if C C C denotes the constant of Theorem 5.2, we can choose R ≥ max { R 0 , R 1 } R\ge\max\{R_0,R_1\} R ≥ max { R 0 , R 1 } such that
R α + γ p − 1 − N α θ ( ℓ ( R ) ) 1 p − 1 B > C . R^{\frac{\alpha+\gamma}{p-1}-\frac{N\alpha}{\theta}}\bigl(\ell(R)\bigr)^{\frac{1}{p-1}}B>C. R p − 1 α + γ − θ N α ( ℓ ( R ) ) p − 1 1 B > C .
Thus condition (31) holds, and the conclusion follows from Theorem 5.2. □
Corollary 5.1 with q = ∞ q=\infty q = ∞ proves Theorem 3.2 (c), since mild solutions in the sense of Definition 2.3 are L ∞ L^\infty L ∞ -solutions. The rest of this subsection is devoted to the proof of Theorem 5.2.
Let Y Y Y denote the kernel of P α , θ ( t ) P_{\alpha,\theta}(t) P α , θ ( t ) , that is,
P α , θ ( t ) v ( x ) = ∫ R N Y ( t , x − y ) v ( y ) d y . P_{\alpha,\theta}(t)v(x)=\int_{\mathbb{R}^N}Y(t,x-y)v(y)\,dy. P α , θ ( t ) v ( x ) = ∫ R N Y ( t , x − y ) v ( y ) d y .
As for Z Z Z , the kernel Y Y Y is self-similar,
Y ( t , x ) = t − N α θ G ( x t − α θ ) , ( t , x ) ∈ ( 0 , ∞ ) × R N . (32) Y(t,x)=t^{-\frac{N\alpha}{\theta}}G\left(xt^{-\frac{\alpha}{\theta}}\right),\qquad(t,x)\in(0,\infty)\times\mathbb{R}^N. \tag*{(32)} Y ( t , x ) = t − θ N α G ( x t − θ α ) , ( t , x ) ∈ ( 0 , ∞ ) × R N . (32)
where the profile G G G , the inverse Fourier transform of E α , α ( − ∣ ξ ∣ θ ) E_{\alpha,\alpha}(-|\xi|^\theta) E α , α ( − ∣ ξ ∣ θ ) , is radial, positive and continuous on R N ∖ { 0 } \mathbb{R}^N\setminus\{0\} R N ∖ { 0 } . By [17 Lemma 3.6] (whose kernel is t α − 1 Y ( t , x ) t^{\alpha-1}Y(t,x) t α − 1 Y ( t , x ) in our notation), for 0 < ∣ y ∣ ≤ 1 0<|y|\le1 0 < ∣ y ∣ ≤ 1 ,
G ( y ) ≍ { ∣ y ∣ − N + 2 θ , if N > 2 θ , ∣ log ( ∣ y ∣ / 2 θ ) ∣ , if N = 2 θ , 1 , if N < 2 θ . G(y)\asymp
\begin{cases}
|y|^{-N+2\theta}, & \text{if }N>2\theta,\\
|\log(|y|/2^\theta)|, & \text{if }N=2\theta,\\
1, & \text{if }N<2\theta.
\end{cases} G ( y ) ≍ ⎩ ⎨ ⎧ ∣ y ∣ − N + 2 θ , ∣ log ( ∣ y ∣/ 2 θ ) ∣ , 1 , if N > 2 θ , if N = 2 θ , if N < 2 θ .
For ∣ y ∣ ≥ 1 |y|\ge1 ∣ y ∣ ≥ 1 , G ( y ) ≍ ∣ y ∣ − N − θ G(y)\asymp|y|^{-N-\theta} G ( y ) ≍ ∣ y ∣ − N − θ when θ < 2 \theta<2 θ < 2 , and G G G is positive and continuous in all cases. Consequently, for every b > 0 b>0 b > 0 ,
G ‾ ( b ) : = inf 0 < ∣ ξ ∣ < b G ( ξ ) > 0. \overline{G}(b):=\inf_{0<|\xi|<b}G(\xi)>0. G ( b ) := 0 < ∣ ξ ∣ < b inf G ( ξ ) > 0.
For b > 0 b>0 b > 0 and T > 0 T>0 T > 0 we define
J ( T ; b ) = ∫ T / 4 T ∫ R N u ( t , x ) Y b ( T − t , x ) d x d t (33) J(T;b)=\int_{T/4}^{T}\int_{\mathbb{R}^{N}}u(t,x)Y_b(T-t,x)\,dx\,dt \tag*{(33)} J ( T ; b ) = ∫ T /4 T ∫ R N u ( t , x ) Y b ( T − t , x ) d x d t (33)
where
Y b ( T − t , x ) = χ { ∣ x ∣ < ( T − t ) α / θ } ( T − t ) α p − 1 G ‾ ( b ) . Y_b(T-t,x)=\chi_{\{|x|<(T-t)^{\alpha/\theta}\}}(T-t)^{\frac{\alpha}{p-1}}\overline{G}(b). Y b ( T − t , x ) = χ { ∣ x ∣ < ( T − t ) α / θ } ( T − t ) p − 1 α G ( b ) .
We first show that J ( T ; b ) J(T;b) J ( T ; b ) is finite whenever ∥ u ( t , ⋅ ) ∥ L q \|u(t,\cdot)\|_{L^q} ∥ u ( t , ⋅ ) ∥ L q remains bounded.
Proposition 5.1. Let p > 1 p>1 p > 1 and b > 0 b>0 b > 0 . If ∥ u ( t , ⋅ ) ∥ L q ( R N ) ≤ C \|u(t,\cdot)\|_{L^q(\mathbb{R}^N)}\le C ∥ u ( t , ⋅ ) ∥ L q ( R N ) ≤ C for t ∈ [ 0 , T ) t\in[0,T) t ∈ [ 0 , T ) , then J ( T ; b ) J(T;b) J ( T ; b ) is finite.
Proof. Let r ∈ [ 1 , ∞ ] r\in[1,\infty] r ∈ [ 1 , ∞ ] be the conjugate exponent of q q q , so that 1 r + 1 q = 1 \frac{1}{r}+\frac{1}{q}=1 r 1 + q 1 = 1 . Hölder’s inequality yields
J ( T ; b ) ≤ ∫ T / 4 T ∥ u ( t , ⋅ ) ∥ L q ∥ Y b ( T − t , ⋅ ) ∥ L r d t ≤ C ∫ T / 4 T ∥ Y b ( T − t , ⋅ ) ∥ L r d t . J(T;b)\le\int_{T/4}^{T}\|u(t,\cdot)\|_{L^q}\|Y_b(T-t,\cdot)\|_{L^r}\,dt\le C\int_{T/4}^{T}\|Y_b(T-t,\cdot)\|_{L^r}\,dt. J ( T ; b ) ≤ ∫ T /4 T ∥ u ( t , ⋅ ) ∥ L q ∥ Y b ( T − t , ⋅ ) ∥ L r d t ≤ C ∫ T /4 T ∥ Y b ( T − t , ⋅ ) ∥ L r d t .
Since
∥ Y b ( T − t , ⋅ ) ∥ L r ( R N ) = G ‾ ( b ) C ( N , r ) ( T − t ) N α θ r ( T − t ) α p − 1 , \|Y_b(T-t,\cdot)\|_{L^r(\mathbb{R}^N)}=\overline{G}(b)C(N,r)(T-t)^{\frac{N\alpha}{\theta r}}(T-t)^{\frac{\alpha}{p-1}}, ∥ Y b ( T − t , ⋅ ) ∥ L r ( R N ) = G ( b ) C ( N , r ) ( T − t ) θ r N α ( T − t ) p − 1 α ,
we conclude that J ( T ; b ) < ∞ J(T;b)<\infty J ( T ; b ) < ∞ .
The next step is an upper bound for J ( T ; b ) J(T;b) J ( T ; b ) , whenever this quantity is finite, which does not depend on the initial data. By (5) with [ λ 1 , λ 2 ] = [ 1 , 4 ] [\lambda_1,\lambda_2]=[1,4] [ λ 1 , λ 2 ] = [ 1 , 4 ] , there exists T 0 ≥ 4 T_0\ge4 T 0 ≥ 4 , depending only on ℓ \ell ℓ , such that ℓ ( s ) ≥ 1 2 ℓ ( T / 4 ) \ell(s)\ge\frac{1}{2}\ell(T/4) ℓ ( s ) ≥ 2 1 ℓ ( T /4 ) for all s ∈ [ T / 4 , T ] s\in[T/4,T] s ∈ [ T /4 , T ] and all T ≥ T 0 T\ge T_0 T ≥ T 0 . Since h ( s ) = s γ ℓ ( s ) h(s)=s^\gamma\ell(s) h ( s ) = s γ ℓ ( s ) for s ≥ 1 s\ge1 s ≥ 1 and s γ ≥ min { 1 , 4 − γ } T γ s^\gamma\ge\min\{1,4^{-\gamma}\}T^\gamma s γ ≥ min { 1 , 4 − γ } T γ for s ∈ [ T / 4 , T ] s\in[T/4,T] s ∈ [ T /4 , T ] , we get
h ( s ) ≥ c γ T γ ℓ ( T / 4 ) , s ∈ [ T / 4 , T ] , T ≥ T 0 , c γ : = 1 2 min { 1 , 4 − γ } . (34) h(s)\ge c_\gamma T^\gamma\ell(T/4),\qquad s\in[T/4,T],\quad T\ge T_0,\qquad c_\gamma:=\frac{1}{2}\min\{1,4^{-\gamma}\}. \tag*{(34)} h ( s ) ≥ c γ T γ ℓ ( T /4 ) , s ∈ [ T /4 , T ] , T ≥ T 0 , c γ := 2 1 min { 1 , 4 − γ } . (34)
Proposition 5.2. Let b ≥ 2 θ / α + 1 b\ge2^{\theta/\alpha}+1 b ≥ 2 θ / α + 1 . There exists a constant
C 1 = C 1 ( N , α , θ , γ , p , b ) > 0 C_1=C_1(N,\alpha,\theta,\gamma,p,b)>0 C 1 = C 1 ( N , α , θ , γ , p , b ) > 0 such that, if T ≥ T 0 T\ge T_0 T ≥ T 0 and J ( T ; b ) < ∞ J(T;b)<\infty J ( T ; b ) < ∞ , then
J ( T ; b ) ≤ C 1 T 1 + N α θ − γ p − 1 ( ℓ ( T / 4 ) ) − 1 p − 1 . (35) J(T;b)\le C_1T^{1+\frac{N\alpha}{\theta}-\frac{\gamma}{p-1}}\bigl(\ell(T/4)\bigr)^{-\frac{1}{p-1}}. \tag*{(35)} J ( T ; b ) ≤ C 1 T 1 + θ N α − p − 1 γ ( ℓ ( T /4 ) ) − p − 1 1 . (35)
Proof. Since u u u solves the integral equation (15) and S α , θ ( t ) u 0 ≥ 0 S_{\alpha,\theta}(t)u_0\ge0 S α , θ ( t ) u 0 ≥ 0 , we have
u ( t , x ) ≥ ∫ 0 t ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) h ( s ) u p ( s , y ) d y d s . u(t,x)\ge\int_0^t\int_{\mathbb{R}^N}(t-s)^{\alpha-1}Y(t-s,x-y)h(s)u^p(s,y)\,dy\,ds. u ( t , x ) ≥ ∫ 0 t ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) h ( s ) u p ( s , y ) d y d s .
Splitting the domain { T / 4 < t < T , 0 < s < t } \{T/4<t<T,\ 0<s<t\} { T /4 < t < T , 0 < s < t } according to s < T / 4 s<T/4 s < T /4 or s > T / 4 s>T/4 s > T /4 , and using Fubini’s theorem, we obtain
J ( T ; b ) ≥ ∫ T / 4 T ∫ R N ∫ 0 t ∫ R N h ( s ) u p ( s , y ) ( t − s ) α − 1 Y ( t − s , x − y ) d y d s Y b ( T − t , x ) d x d t = ∫ 0 T / 4 ∫ R N h ( s ) u p ( s , y ) ∫ T / 4 T ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) Y b ( T − t , x ) d x d t d y d s + ∫ T / 4 T ∫ R N h ( s ) u p ( s , y ) ∫ s T ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) Y b ( T − t , x ) d x d t d y d s ≥ ∫ T / 4 T ∫ R N h ( s ) u p ( s , y ) ∫ s T ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) Y b ( T − t , x ) d x d t d y d s . \begin{aligned}
J(T;b)&\ge\int_{T/4}^{T}\int_{\mathbb{R}^N}\int_0^t\int_{\mathbb{R}^N}h(s)u^p(s,y)(t-s)^{\alpha-1}Y(t-s,x-y)\,dy\,ds\,Y_b(T-t,x)\,dx\,dt\\
&=\int_0^{T/4}\int_{\mathbb{R}^N}h(s)u^p(s,y)\int_{T/4}^{T}\int_{\mathbb{R}^N}(t-s)^{\alpha-1}Y(t-s,x-y)Y_b(T-t,x)\,dx\,dt\,dy\,ds\\
&\quad+\int_{T/4}^{T}\int_{\mathbb{R}^N}h(s)u^p(s,y)\int_s^{T}\int_{\mathbb{R}^N}(t-s)^{\alpha-1}Y(t-s,x-y)Y_b(T-t,x)\,dx\,dt\,dy\,ds\\
&\ge\int_{T/4}^{T}\int_{\mathbb{R}^N}h(s)u^p(s,y)\int_s^{T}\int_{\mathbb{R}^N}(t-s)^{\alpha-1}Y(t-s,x-y)Y_b(T-t,x)\,dx\,dt\,dy\,ds.
\end{aligned} J ( T ; b ) ≥ ∫ T /4 T ∫ R N ∫ 0 t ∫ R N h ( s ) u p ( s , y ) ( t − s ) α − 1 Y ( t − s , x − y ) d y d s Y b ( T − t , x ) d x d t = ∫ 0 T /4 ∫ R N h ( s ) u p ( s , y ) ∫ T /4 T ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) Y b ( T − t , x ) d x d t d y d s + ∫ T /4 T ∫ R N h ( s ) u p ( s , y ) ∫ s T ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) Y b ( T − t , x ) d x d t d y d s ≥ ∫ T /4 T ∫ R N h ( s ) u p ( s , y ) ∫ s T ∫ R N ( t − s ) α − 1 Y ( t − s , x − y ) Y b ( T − t , x ) d x d t d y d s .
Set λ b : = ( 2 / b ) θ / α \lambda_b:=(2/b)^{\theta/\alpha} λ b := ( 2/ b ) θ / α . Since b ≥ 2 θ / α + 1 b\ge2^{\theta/\alpha}+1 b ≥ 2 θ / α + 1 , we have λ b ≤ 1 2 \lambda_b\le\frac{1}{2} λ b ≤ 2 1 . If
s < T , t > s + λ b ( T − s ) and ∣ x ∣ , ∣ y ∣ < ( T − s ) α / θ , s<T,\qquad t>s+\lambda_b(T-s)\qquad\text{and}\qquad|x|,|y|<(T-s)^{\alpha/\theta}, s < T , t > s + λ b ( T − s ) and ∣ x ∣ , ∣ y ∣ < ( T − s ) α / θ ,
then
∣ x − y ∣ < 2 ( T − s ) α θ < b ( t − s ) α θ , that is, ∣ x − y ∣ ( t − s ) − α θ < b . |x-y|<2(T-s)^{\frac{\alpha}{\theta}}<b(t-s)^{\frac{\alpha}{\theta}}, \qquad\text{that is,}\qquad|x-y|(t-s)^{-\frac{\alpha}{\theta}}<b. ∣ x − y ∣ < 2 ( T − s ) θ α < b ( t − s ) θ α , that is, ∣ x − y ∣ ( t − s ) − θ α < b .
Moreover, ∣ x ∣ < ( T − t ) α θ |x|<(T-t)^{\frac{\alpha}{\theta}} ∣ x ∣ < ( T − t ) θ α implies ∣ x ∣ < ( T − s ) α θ |x|<(T-s)^{\frac{\alpha}{\theta}} ∣ x ∣ < ( T − s ) θ α for t > s t>s t > s . Hence, using the self-similar structure (32) of Y Y Y and the fact that α − 1 − N α θ < 0 \alpha-1-\frac{N\alpha}{\theta}<0 α − 1 − θ N α < 0 (so that ( t − s ) α − 1 − N α θ ≥ ( T − s ) α − 1 − N α θ (t-s)^{\alpha-1-\frac{N\alpha}{\theta}}\ge(T-s)^{\alpha-1-\frac{N\alpha}{\theta}} ( t − s ) α − 1 − θ N α ≥ ( T − s ) α − 1 − θ N α ), we obtain
J ( T ; b ) ≥ G ˉ ( b ) ∫ T / 4 T ∫ ∣ y ∣ < ( T − s ) α θ h ( s ) u p ( s , y ) ∫ s + λ b ( T − s ) T ( t − s ) α − 1 × ∫ ∣ x ∣ < ( T − t ) α θ Y ( t − s , x − y ) ( T − t ) α p − 1 d x d t d y d s ≥ ( G ˉ ( b ) ) 2 ∫ T / 4 T ∫ ∣ y ∣ < ( T − s ) α θ h ( s ) u p ( s , y ) ∫ s + λ b ( T − s ) T ( t − s ) α − 1 − N α θ × ∫ ∣ x ∣ < ( T − t ) α θ ( T − t ) α p − 1 d x d t d y d s ≥ C ( N ) G ˉ ( b ) 2 ∫ T / 4 T ∫ ∣ y ∣ < ( T − s ) α θ h ( s ) u p ( s , y ) × ( T − s ) α − 1 − N α θ ∫ s + λ b ( T − s ) T ( T − t ) α p − 1 + N α θ d t d y d s . \begin{aligned}
J(T;b)&\ge\bar{G}(b)\int_{T/4}^{T}\int_{|y|<(T-s)^{\frac{\alpha}{\theta}}}h(s)u^{p}(s,y)\int_{s+\lambda_b(T-s)}^{T}(t-s)^{\alpha-1}\\
&\qquad{}\times\int_{|x|<(T-t)^{\frac{\alpha}{\theta}}}Y(t-s,x-y)(T-t)^{\frac{\alpha}{p-1}}\,dx\,dt\,dy\,ds\\
&\ge\bigl(\bar{G}(b)\bigr)^2\int_{T/4}^{T}\int_{|y|<(T-s)^{\frac{\alpha}{\theta}}}h(s)u^{p}(s,y)\int_{s+\lambda_b(T-s)}^{T}(t-s)^{\alpha-1-\frac{N\alpha}{\theta}}\\
&\qquad{}\times\int_{|x|<(T-t)^{\frac{\alpha}{\theta}}}(T-t)^{\frac{\alpha}{p-1}}\,dx\,dt\,dy\,ds\\
&\ge C(N)\bar{G}(b)^2\int_{T/4}^{T}\int_{|y|<(T-s)^{\frac{\alpha}{\theta}}}h(s)u^{p}(s,y)\\
&\qquad{}\times(T-s)^{\alpha-1-\frac{N\alpha}{\theta}}\int_{s+\lambda_b(T-s)}^{T}(T-t)^{\frac{\alpha}{p-1}+\frac{N\alpha}{\theta}}\,dt\,dy\,ds.
\end{aligned} J ( T ; b ) ≥ G ˉ ( b ) ∫ T /4 T ∫ ∣ y ∣ < ( T − s ) θ α h ( s ) u p ( s , y ) ∫ s + λ b ( T − s ) T ( t − s ) α − 1 × ∫ ∣ x ∣ < ( T − t ) θ α Y ( t − s , x − y ) ( T − t ) p − 1 α d x d t d y d s ≥ ( G ˉ ( b ) ) 2 ∫ T /4 T ∫ ∣ y ∣ < ( T − s ) θ α h ( s ) u p ( s , y ) ∫ s + λ b ( T − s ) T ( t − s ) α − 1 − θ N α × ∫ ∣ x ∣ < ( T − t ) θ α ( T − t ) p − 1 α d x d t d y d s ≥ C ( N ) G ˉ ( b ) 2 ∫ T /4 T ∫ ∣ y ∣ < ( T − s ) θ α h ( s ) u p ( s , y ) × ( T − s ) α − 1 − θ N α ∫ s + λ b ( T − s ) T ( T − t ) p − 1 α + θ N α d t d y d s .
Writing a : = α p − 1 + N α θ a:=\frac{\alpha}{p-1}+\frac{N\alpha}{\theta} a := p − 1 α + θ N α and using λ b ≤ 1 2 \lambda_b\le\frac{1}{2} λ b ≤ 2 1 , we have
∫ s + λ b ( T − s ) T ( T − t ) a d t = ( ( 1 − λ b ) ( T − s ) ) a + 1 a + 1 ≥ ( T − s ) a + 1 2 a + 1 ( a + 1 ) , \int_{s+\lambda_b(T-s)}^{T}(T-t)^a\,dt=\frac{\bigl((1-\lambda_b)(T-s)\bigr)^{a+1}}{a+1}\ge\frac{(T-s)^{a+1}}{2^{a+1}(a+1)}, ∫ s + λ b ( T − s ) T ( T − t ) a d t = a + 1 ( ( 1 − λ b ) ( T − s ) ) a + 1 ≥ 2 a + 1 ( a + 1 ) ( T − s ) a + 1 ,
and α − 1 − N α θ + a + 1 = α p p − 1 \alpha-1-\frac{N\alpha}{\theta}+a+1=\frac{\alpha p}{p-1} α − 1 − θ N α + a + 1 = p − 1 α p . Introducing
Y ^ b ( T − s , y ) : = χ { ∣ y ∣ < ( T − s ) α θ } G ˉ ( b ) , \widehat{Y}_b(T-s,y):=\chi_{\{|y|<(T-s)^{\frac{\alpha}{\theta}}\}}\bar{G}(b), Y b ( T − s , y ) := χ { ∣ y ∣ < ( T − s ) θ α } G ˉ ( b ) ,
and using (34) , we deduce that, for all T ≥ T 0 T\ge T_0 T ≥ T 0 ,
J ( T ; b ) ≥ C ( N , α , θ , p , b ) ∫ T / 4 T ∫ R N h ( s ) u p ( s , y ) Y ^ b ( T − s , y ) ( T − s ) α p p − 1 d y d s ≥ C ( N , α , θ , γ , p , b ) T γ ℓ ( T / 4 ) ∫ T / 4 T ∫ R N u p ( s , y ) Y ^ b ( T − s , y ) ( T − s ) α p p − 1 d y d s . (36) \begin{aligned}
J(T;b)&\ge C(N,\alpha,\theta,p,b)\int_{T/4}^{T}\int_{\mathbb{R}^{N}}h(s)u^{p}(s,y)\widehat{Y}_b(T-s,y)(T-s)^{\frac{\alpha p}{p-1}}\,dy\,ds\\
&\ge C(N,\alpha,\theta,\gamma,p,b)T^\gamma\ell(T/4)\int_{T/4}^{T}\int_{\mathbb{R}^{N}}u^{p}(s,y)\widehat{Y}_b(T-s,y)(T-s)^{\frac{\alpha p}{p-1}}\,dy\,ds. \tag*{(36)}
\end{aligned} J ( T ; b ) ≥ C ( N , α , θ , p , b ) ∫ T /4 T ∫ R N h ( s ) u p ( s , y ) Y b ( T − s , y ) ( T − s ) p − 1 α p d y d s ≥ C ( N , α , θ , γ , p , b ) T γ ℓ ( T /4 ) ∫ T /4 T ∫ R N u p ( s , y ) Y b ( T − s , y ) ( T − s ) p − 1 α p d y d s . (36)
On the other hand, Hölder’s (or Jensen’s) inequality with respect to the measure Y ^ b d y d s \widehat{Y}_b\,dy\,ds Y b d y d s on ( T / 4 , T ) × R N (T/4,T)\times\mathbb{R}^{N} ( T /4 , T ) × R N gives
∫ T / 4 T ∫ R N u p ( s , y ) Y ^ b ( T − s , y ) ( T − s ) α p p − 1 d y d s ≥ ( ∫ T / 4 T ∫ R N u ( s , y ) Y ^ b ( T − s , y ) ( T − s ) α p − 1 d y d s ) p × ∥ Y ^ b ∥ L 1 ( ( T / 4 , T ) × R N ) − ( p − 1 ) = ( J ( T ; b ) ) p ∥ Y ^ b ∥ L 1 ( ( T / 4 , T ) × R N ) − ( p − 1 ) . (37) \begin{aligned}
\int_{T/4}^{T}\int_{\mathbb{R}^{N}}u^{p}(s,y)\widehat{Y}_b(T-s,y)(T-s)^{\frac{\alpha p}{p-1}}\,dy\,ds
&\ge\left(\int_{T/4}^{T}\int_{\mathbb{R}^{N}}u(s,y)\widehat{Y}_b(T-s,y)(T-s)^{\frac{\alpha}{p-1}}\,dy\,ds\right)^p\\
&\qquad{}\times\|\widehat{Y}_b\|_{L^1((T/4,T)\times\mathbb{R}^{N})}^{-(p-1)}\\
&=(J(T;b))^p\|\widehat{Y}_b\|_{L^1((T/4,T)\times\mathbb{R}^{N})}^{-(p-1)}. \tag*{(37)}
\end{aligned} ∫ T /4 T ∫ R N u p ( s , y ) Y b ( T − s , y ) ( T − s ) p − 1 α p d y d s ≥ ( ∫ T /4 T ∫ R N u ( s , y ) Y b ( T − s , y ) ( T − s ) p − 1 α d y d s ) p × ∥ Y b ∥ L 1 (( T /4 , T ) × R N ) − ( p − 1 ) = ( J ( T ; b ) ) p ∥ Y b ∥ L 1 (( T /4 , T ) × R N ) − ( p − 1 ) . (37)
From (36) and (37) , we infer that
J ( T ; b ) ≥ C ( N , α , θ , γ , p , b ) T γ ℓ ( T / 4 ) ( J ( T ; b ) ) p ∥ Y ^ b ∥ L 1 ( ( T / 4 , T ) × R N ) − ( p − 1 ) . J(T;b)\ge C(N,\alpha,\theta,\gamma,p,b)T^\gamma\ell(T/4)(J(T;b))^p\|\widehat{Y}_b\|_{L^1((T/4,T)\times\mathbb{R}^{N})}^{-(p-1)}. J ( T ; b ) ≥ C ( N , α , θ , γ , p , b ) T γ ℓ ( T /4 ) ( J ( T ; b ) ) p ∥ Y b ∥ L 1 (( T /4 , T ) × R N ) − ( p − 1 ) .
A direct computation gives
∥ Y ^ b ∥ L 1 ( ( T / 4 , T ) × R N ) = G ‾ ( b ) ∫ T / 4 T ∫ ∣ x ∣ < ( T − t ) α θ d x d t = C ( N ) G ‾ ( b ) ∫ T / 4 T ( T − t ) N α θ d t = C ( N , α , θ ) G ‾ ( b ) T 1 + N α θ . \|\widehat{Y}_b\|_{L^1((T/4,T)\times\mathbb{R}^N)}=\overline{G}(b)\int_{T/4}^{T}\int_{|x|<(T-t)^{\frac{\alpha}{\theta}}}dx\,dt=C(N)\overline{G}(b)\int_{T/4}^{T}(T-t)^{\frac{N\alpha}{\theta}}dt
=C(N,\alpha,\theta)\overline{G}(b)T^{1+\frac{N\alpha}{\theta}}. ∥ Y b ∥ L 1 (( T /4 , T ) × R N ) = G ( b ) ∫ T /4 T ∫ ∣ x ∣ < ( T − t ) θ α d x d t = C ( N ) G ( b ) ∫ T /4 T ( T − t ) θ N α d t = C ( N , α , θ ) G ( b ) T 1 + θ N α .
Therefore,
J ( T ; b ) ≥ C ‾ ( N , α , θ , γ , p , b ) ( J ( T ; b ) ) p T − ( 1 + N α θ ) ( p − 1 ) + γ ℓ ( T / 4 ) . J(T;b)\geq\overline{C}(N,\alpha,\theta,\gamma,p,b)(J(T;b))^pT^{-(1+\frac{N\alpha}{\theta})(p-1)+\gamma}\ell(T/4). J ( T ; b ) ≥ C ( N , α , θ , γ , p , b ) ( J ( T ; b ) ) p T − ( 1 + θ N α ) ( p − 1 ) + γ ℓ ( T /4 ) .
If J ( T ; b ) = 0 J(T;b)=0 J ( T ; b ) = 0 there is nothing to prove. Otherwise, dividing by J ( T ; b ) ∈ ( 0 , ∞ ) J(T;b)\in(0,\infty) J ( T ; b ) ∈ ( 0 , ∞ ) and taking the power 1 / ( p − 1 ) 1/(p-1) 1/ ( p − 1 ) , we obtain (35) . □
Proposition 5.3. There exists a constant C 2 = C 2 ( N , α , θ , p , b ) > 0 C_2=C_2(N,\alpha,\theta,p,b)>0 C 2 = C 2 ( N , α , θ , p , b ) > 0 such that
J ( T ; b ) ≥ C 2 A ( T ) T 1 + N α θ , where A ( T ) : = T α p − 1 − N α θ ∫ ∣ y ∣ < ( T / 4 ) α θ / 2 u 0 ( y ) d y . J(T;b)\geq C_2A(T)T^{1+\frac{N\alpha}{\theta}},\qquad\text{where}\qquad A(T):=T^{\frac{\alpha}{p-1}-\frac{N\alpha}{\theta}}\int_{|y|<(T/4)^{\frac{\alpha}{\theta}}/2}u_0(y)\,dy. J ( T ; b ) ≥ C 2 A ( T ) T 1 + θ N α , where A ( T ) := T p − 1 α − θ N α ∫ ∣ y ∣ < ( T /4 ) θ α /2 u 0 ( y ) d y . Proof. If ∣ x ∣ , ∣ y ∣ < t α / θ / 2 |x|,|y|<t^{\alpha/\theta}/2 ∣ x ∣ , ∣ y ∣ < t α / θ /2 , then ∣ x − y ∣ < t α / θ |x-y|<t^{\alpha/\theta} ∣ x − y ∣ < t α / θ . Hence, by (2.10),
u ( t , x ) ≥ ∫ R N u 0 ( y ) Z ( t , x − y ) d y ≥ ∫ ∣ y ∣ < t α / θ / 2 u 0 ( y ) Z ( t , x − y ) d y u(t,x)\geq\int_{\mathbb{R}^N}u_0(y)Z(t,x-y)\,dy\geq\int_{|y|<t^{\alpha/\theta}/2}u_0(y)Z(t,x-y)\,dy u ( t , x ) ≥ ∫ R N u 0 ( y ) Z ( t , x − y ) d y ≥ ∫ ∣ y ∣ < t α / θ /2 u 0 ( y ) Z ( t , x − y ) d y ≥ C ( N , α , θ ) t − N α θ ∫ ∣ y ∣ < t α / θ / 2 u 0 ( y ) d y \geq C(N,\alpha,\theta)t^{-\frac{N\alpha}{\theta}}\int_{|y|<t^{\alpha/\theta}/2}u_0(y)\,dy ≥ C ( N , α , θ ) t − θ N α ∫ ∣ y ∣ < t α / θ /2 u 0 ( y ) d y for all ∣ x ∣ < t α / θ / 2 |x|<t^{\alpha/\theta}/2 ∣ x ∣ < t α / θ /2 . Therefore,
J ( T ; b ) ≥ C ( N , α , θ ) G ‾ ( b ) ∫ T / 4 T / 2 ∫ ∣ y ∣ < t α / θ / 2 u 0 ( y ) d y ∫ ∣ x ∣ < min { t α / θ / 2 , ( T − t ) α / θ } t − N α θ ( T − t ) α p − 1 d x d t . J(T;b)\geq C(N,\alpha,\theta)\overline{G}(b)\int_{T/4}^{T/2}\int_{|y|<t^{\alpha/\theta}/2}u_0(y)\,dy\int_{|x|<\min\{t^{\alpha/\theta}/2,(T-t)^{\alpha/\theta}\}}t^{-\frac{N\alpha}{\theta}}(T-t)^{\frac{\alpha}{p-1}}\,dx\,dt. J ( T ; b ) ≥ C ( N , α , θ ) G ( b ) ∫ T /4 T /2 ∫ ∣ y ∣ < t α / θ /2 u 0 ( y ) d y ∫ ∣ x ∣ < m i n { t α / θ /2 , ( T − t ) α / θ } t − θ N α ( T − t ) p − 1 α d x d t . For t ∈ [ T / 4 , T / 2 ] t\in[T/4,T/2] t ∈ [ T /4 , T /2 ] we have T − t ≥ t T-t\geq t T − t ≥ t , hence
min { t α / θ / 2 , ( T − t ) α / θ } = t α / θ / 2 ≥ ( T / 4 ) α / θ / 2. \min\left\{t^{\alpha/\theta}/2,(T-t)^{\alpha/\theta}\right\}=t^{\alpha/\theta}/2\geq(T/4)^{\alpha/\theta}/2. min { t α / θ /2 , ( T − t ) α / θ } = t α / θ /2 ≥ ( T /4 ) α / θ /2. Moreover, ( T − t ) α p − 1 ≥ ( T / 2 ) α p − 1 (T-t)^{\frac{\alpha}{p-1}}\geq(T/2)^{\frac{\alpha}{p-1}} ( T − t ) p − 1 α ≥ ( T /2 ) p − 1 α , t − N α θ ≥ ( T / 2 ) − N α θ t^{-\frac{N\alpha}{\theta}}\geq(T/2)^{-\frac{N\alpha}{\theta}} t − θ N α ≥ ( T /2 ) − θ N α , and the integral of u 0 u_0 u 0 over { ∣ y ∣ < t α / θ / 2 } \{|y|<t^{\alpha/\theta}/2\} { ∣ y ∣ < t α / θ /2 } is at least its integral over { ∣ y ∣ < ( T / 4 ) α / θ / 2 } \{|y|<(T/4)^{\alpha/\theta}/2\} { ∣ y ∣ < ( T /4 ) α / θ /2 } . Since the interval [ T / 4 , T / 2 ] [T/4,T/2] [ T /4 , T /2 ] has length T / 4 T/4 T /4 , we obtain
J ( T ; b ) ≥ C 2 T 1 + α p − 1 ∫ ∣ y ∣ < ( T / 4 ) α / θ / 2 u 0 ( y ) d y = C 2 A ( T ) T 1 + N α θ . J(T;b)\geq C_2T^{1+\frac{\alpha}{p-1}}\int_{|y|<(T/4)^{\alpha/\theta}/2}u_0(y)\,dy=C_2A(T)T^{1+\frac{N\alpha}{\theta}}. J ( T ; b ) ≥ C 2 T 1 + p − 1 α ∫ ∣ y ∣ < ( T /4 ) α / θ /2 u 0 ( y ) d y = C 2 A ( T ) T 1 + θ N α .
□
We now have all the ingredients to prove Theorem 5.2.
Proof of Theorem 5.2. We fix b = 2 θ α + 1 b=2^{\frac{\theta}{\alpha}}+1 b = 2 α θ + 1 and set R 0 : = T 0 / 4 R_0:=T_0/4 R 0 := T 0 /4 . Let T ≥ T 0 T\geq T_0 T ≥ T 0 . If ∥ u ( t , ⋅ ) ∥ L q ( R N ) \|u(t,\cdot)\|_{L^q(\mathbb{R}^N)} ∥ u ( t , ⋅ ) ∥ L q ( R N ) is bounded on [ 0 , T ) [0,T) [ 0 , T ) , then J ( T ; b ) J(T;b) J ( T ; b ) is well defined and finite by Proposition 5.1. Hence, by Proposition 5.2,
J ( T ; b ) ≤ C 1 T 1 + N α θ − γ p − 1 ( ℓ ( T / 4 ) ) − 1 p − 1 , C 1 = C 1 ( N , α , θ , γ , p , b ) . J(T;b)\leq C_1T^{1+\frac{N\alpha}{\theta}-\frac{\gamma}{p-1}}(\ell(T/4))^{-\frac{1}{p-1}},\qquad C_1=C_1(N,\alpha,\theta,\gamma,p,b). J ( T ; b ) ≤ C 1 T 1 + θ N α − p − 1 γ ( ℓ ( T /4 ) ) − p − 1 1 , C 1 = C 1 ( N , α , θ , γ , p , b ) .
On the other hand, Proposition 5.3 gives
J ( T ; b ) ≥ C 2 A ( T ) T 1 + N α θ , C 2 = C 2 ( N , α , θ , p , b ) > 0. J(T;b)\geq C_2A(T)T^{1+\frac{N\alpha}{\theta}},\qquad C_2=C_2(N,\alpha,\theta,p,b)>0. J ( T ; b ) ≥ C 2 A ( T ) T 1 + θ N α , C 2 = C 2 ( N , α , θ , p , b ) > 0.
Therefore, if
A ( T ) > C 1 C 2 T − γ p − 1 ( ℓ ( T / 4 ) ) − 1 p − 1 , A(T)>\frac{C_1}{C_2}T^{-\frac{\gamma}{p-1}}(\ell(T/4))^{-\frac{1}{p-1}}, A ( T ) > C 2 C 1 T − p − 1 γ ( ℓ ( T /4 ) ) − p − 1 1 ,
we reach a contradiction. Consequently, whenever this inequality holds, the L q L^q L q norm of u u u cannot remain bounded on [ 0 , T ) [0,T) [ 0 , T ) , so that the maximal existence time is at most T T T . For T = 4 R T=4R T = 4 R with R ≥ R 0 R\ge R_0 R ≥ R 0 , this inequality is exactly condition (31) , with
C = 4 N α θ − α + 1 p − 1 C 1 C 2 . C=4^{\frac{N\alpha}{\theta}-\frac{\alpha+1}{p-1}}\frac{C_1}{C_2}. C = 4 θ N α − p − 1 α + 1 C 2 C 1 .
This completes the proof. □
A local blow-up mechanism for N > θ N>\theta N > θ
In this subsection we prove Proposition 3.1. The idea is that, for α < 1 \alpha<1 α < 1 and N > θ N>\theta N > θ , the solution decays on the unit ball only like t − α t^{-\alpha} t − α (Remark 2.1(ii)), while on a short time interval the Duhamel term acts like the fractional ODE ∂ t α v = h ( t ) v p \partial_t^\alpha v=h(t)v^p ∂ t α v = h ( t ) v p .
Proof of Proposition 3.1. Let u u u be a nontrivial nonnegative mild solution, and assume, for contradiction, that it is global, so that u u u is bounded on [ 0 , T ] × R N [0,T]\times\mathbb{R}^N [ 0 , T ] × R N for every T > 0 T>0 T > 0 . After a translation in x x x we may assume that m 0 : = ∫ ∣ y ∣ ≤ 1 u 0 ( y ) d y > 0 m_0:=\int_{|y|\le1}u_0(y)\,dy>0 m 0 := ∫ ∣ y ∣ ≤ 1 u 0 ( y ) d y > 0 . For a.e. t > 0 t>0 t > 0 we set
μ ( t ) : = ess inf ∣ x ∣ ≤ 1 u ( t , x ) . \mu(t):=\mathop{\operatorname{ess\,inf}}_{|x|\le1}u(t,x). μ ( t ) := ess inf ∣ x ∣ ≤ 1 u ( t , x ) .
Step 1: linear lower bound. Let t ≥ t 1 : = 2 θ / α t\ge t_1:=2^{\theta/\alpha} t ≥ t 1 := 2 θ / α . If ∣ x ∣ ≤ 1 |x|\le1 ∣ x ∣ ≤ 1 and ∣ y ∣ ≤ 1 |y|\le1 ∣ y ∣ ≤ 1 , then ∣ x − y ∣ ≤ 2 ≤ t α / θ |x-y|\le2\le t^{\alpha/\theta} ∣ x − y ∣ ≤ 2 ≤ t α / θ , and (13) gives Z ( t , x − y ) ≥ C 2 θ − N t − α Z(t,x-y)\ge C2^{\theta-N}t^{-\alpha} Z ( t , x − y ) ≥ C 2 θ − N t − α . Since u ≥ S α , θ ( t ) u 0 u\ge S_{\alpha,\theta}(t)u_0 u ≥ S α , θ ( t ) u 0 , we obtain
μ ( t ) ≥ c 1 m 0 t − α , t ≥ t 1 . (38) \mu(t)\ge c_1m_0t^{-\alpha},\qquad t\ge t_1. \tag*{(38)} μ ( t ) ≥ c 1 m 0 t − α , t ≥ t 1 . (38)
Step 2: local Duhamel lower bound. Let
c 2 : = inf ∣ x ∣ ≤ 1 , 0 < τ ≤ 1 ∫ ∣ y ∣ ≤ 1 Y ( τ , x − y ) d y . c_2:=\inf_{\substack{|x|\le1,\ 0<\tau\le1}}\int_{|y|\le1}Y(\tau,x-y)\,dy. c 2 := ∣ x ∣ ≤ 1 , 0 < τ ≤ 1 inf ∫ ∣ y ∣ ≤ 1 Y ( τ , x − y ) d y .
We claim that c 2 > 0 c_2>0 c 2 > 0 . By (32) , the integral equals ∫ E G ( ξ ) d ξ \int_EG(\xi)\,d\xi ∫ E G ( ξ ) d ξ with E = { ξ : ∣ x − τ α / θ ξ ∣ ≤ 1 } E=\{\xi:|x-\tau^{\alpha/\theta}\xi|\le1\} E = { ξ : ∣ x − τ α / θ ξ ∣ ≤ 1 } . Put ρ = τ α / θ ∈ ( 0 , 1 ] \rho=\tau^{\alpha/\theta}\in(0,1] ρ = τ α / θ ∈ ( 0 , 1 ] . If ∣ x ∣ ≤ 1 2 |x|\le\frac12 ∣ x ∣ ≤ 2 1 , then E ⊃ { ∣ ξ ∣ ≤ 1 2 } E\supset\{|\xi|\le\frac12\} E ⊃ { ∣ ξ ∣ ≤ 2 1 } . If 1 2 < ∣ x ∣ ≤ 1 \frac12<|x|\le1 2 1 < ∣ x ∣ ≤ 1 , then E E E contains the cone sector C x : = { ∣ ξ ∣ ≤ 1 2 , ξ ⋅ x ≥ 1 2 ∣ ξ ∣ ∣ x ∣ } C_x:=\{|\xi|\le\frac12,\ \xi\cdot x\ge\frac12|\xi||x|\} C x := { ∣ ξ ∣ ≤ 2 1 , ξ ⋅ x ≥ 2 1 ∣ ξ ∣∣ x ∣ } , because, for ξ ∈ C x \xi\in C_x ξ ∈ C x , we have ∣ x − ρ ξ ∣ 2 ≤ ∣ x ∣ 2 − ρ ∣ ξ ∣ ( ∣ x ∣ − ρ ∣ ξ ∣ ) ≤ 1 |x-\rho\xi|^2\le|x|^2-\rho|\xi|(|x|-\rho|\xi|)\le1 ∣ x − ρ ξ ∣ 2 ≤ ∣ x ∣ 2 − ρ ∣ ξ ∣ ( ∣ x ∣ − ρ ∣ ξ ∣ ) ≤ 1 since ρ ∣ ξ ∣ ≤ 1 2 < ∣ x ∣ \rho|\xi|\le\frac12<|x| ρ ∣ ξ ∣ ≤ 2 1 < ∣ x ∣ . Since G G G is radial and positive, the integrals of G G G over { ∣ ξ ∣ ≤ 1 2 } \{|\xi|\le\frac12\} { ∣ ξ ∣ ≤ 2 1 } and over C x C_x C x are positive and independent of x x x , which proves the claim.
Now let δ ∈ ( 0 , 1 ] \delta\in(0,1] δ ∈ ( 0 , 1 ] and t ≥ 1 + δ t\ge1+\delta t ≥ 1 + δ . Keeping in (15) only the times s ∈ [ t − δ , t − δ / 2 ] s\in[t-\delta,t-\delta/2] s ∈ [ t − δ , t − δ /2 ] and the points ∣ y ∣ ≤ 1 |y|\le1 ∣ y ∣ ≤ 1 , and using ( t − s ) α − 1 ≥ δ α − 1 (t-s)^{\alpha-1}\ge\delta^{\alpha-1} ( t − s ) α − 1 ≥ δ α − 1 there, we get, for a.e. ∣ x ∣ ≤ 1 |x|\le1 ∣ x ∣ ≤ 1 ,
u ( t , x ) ≥ c 2 δ α − 1 ∫ t − δ t − δ / 2 h ( s ) μ ( s ) p d s ≥ c 2 2 δ α inf s ∈ [ t − δ , t − δ / 2 ] h ( s ) μ ( s ) p . (39) u(t,x)\ge c_2\delta^{\alpha-1}\int_{t-\delta}^{t-\delta/2}h(s)\mu(s)^p\,ds\ge\frac{c_2}{2}\delta^\alpha\inf_{s\in[t-\delta,t-\delta/2]}h(s)\mu(s)^p. \tag*{(39)} u ( t , x ) ≥ c 2 δ α − 1 ∫ t − δ t − δ /2 h ( s ) μ ( s ) p d s ≥ 2 c 2 δ α s ∈ [ t − δ , t − δ /2 ] inf h ( s ) μ ( s ) p . (39)
Step 3: renormalisation. By (5) and since s γ ≥ 2 − γ t γ s^\gamma\ge2^{-\gamma}t^\gamma s γ ≥ 2 − γ t γ for s ∈ [ t / 2 , t ] s\in[t/2,t] s ∈ [ t /2 , t ] , there are t 2 ≥ t 1 + 1 t_2\ge t_1+1 t 2 ≥ t 1 + 1 and c 3 ∈ ( 0 , 1 ] c_3\in(0,1] c 3 ∈ ( 0 , 1 ] such that c 3 h ( t ) ≤ h ( s ) ≤ c 3 − 1 h ( t ) c_3h(t)\le h(s)\le c_3^{-1}h(t) c 3 h ( t ) ≤ h ( s ) ≤ c 3 − 1 h ( t ) for all s ∈ [ t − 1 , t ] s\in[t-1,t] s ∈ [ t − 1 , t ] and t ≥ t 2 t\ge t_2 t ≥ t 2 . We set
W ( t ) : = h ( t ) 1 p − 1 μ ( t ) . W(t):=h(t)^{\frac{1}{p-1}}\mu(t). W ( t ) := h ( t ) p − 1 1 μ ( t ) .
For s ∈ [ t − 1 , t ] s\in[t-1,t] s ∈ [ t − 1 , t ] and t ≥ t 2 t\ge t_2 t ≥ t 2 we have h ( s ) μ ( s ) p = h ( s ) − 1 p − 1 W ( s ) p ≥ c 3 1 p − 1 h ( t ) − 1 p − 1 W ( s ) p h(s)\mu(s)^p=h(s)^{-\frac{1}{p-1}}W(s)^p\ge c_3^{\frac{1}{p-1}}h(t)^{-\frac{1}{p-1}}W(s)^p h ( s ) μ ( s ) p = h ( s ) − p − 1 1 W ( s ) p ≥ c 3 p − 1 1 h ( t ) − p − 1 1 W ( s ) p . Multiplying (39) by h ( t ) 1 / ( p − 1 ) h(t)^{1/(p-1)} h ( t ) 1/ ( p − 1 ) , we infer that
W ( t ) ≥ A δ α inf s ∈ [ t − δ , t − δ / 2 ] W ( s ) p , t ≥ t 2 , δ ∈ ( 0 , 1 ] , (40) W(t)\ge A\delta^\alpha\inf_{s\in[t-\delta,t-\delta/2]}W(s)^p,\qquad t\ge t_2,\ \delta\in(0,1], \tag*{(40)} W ( t ) ≥ A δ α s ∈ [ t − δ , t − δ /2 ] inf W ( s ) p , t ≥ t 2 , δ ∈ ( 0 , 1 ] , (40)
with A : = c 2 2 c 3 1 p − 1 > 0 A:=\frac{c_2}{2}c_3^{\frac{1}{p-1}}>0 A := 2 c 2 c 3 p − 1 1 > 0 . On the other hand, (38) gives
W ( t ) ≥ c 1 m 0 ( t γ − α ( p − 1 ) ℓ ( t ) ) 1 p − 1 , t ≥ t 2 . (41) W(t)\ge c_1m_0\left(t^{\gamma-\alpha(p-1)}\ell(t)\right)^{\frac{1}{p-1}},\qquad t\ge t_2. \tag*{(41)} W ( t ) ≥ c 1 m 0 ( t γ − α ( p − 1 ) ℓ ( t ) ) p − 1 1 , t ≥ t 2 . (41)
If p < 1 + γ / α p < 1 + \gamma/\alpha p < 1 + γ / α , then γ − α ( p − 1 ) > 0 \gamma- \alpha(p - 1) > 0 γ − α ( p − 1 ) > 0 , and the right-hand side of (41) tends to infinity by (K1). If p = 1 + γ / α p = 1 + \gamma/\alpha p = 1 + γ / α and lim sup ℓ = ∞ \limsup\ell= \infty lim sup ℓ = ∞ , it is unbounded. In both cases, using (5) once more, we deduce that for every Λ > 0 \Lambda> 0 Λ > 0 there is t 0 ≥ t 2 + 1 t_0 \ge t_2 + 1 t 0 ≥ t 2 + 1 with
inf s ∈ [ t 0 − 1 , t 0 ] W ( s ) ≥ Λ . (42) \inf_{s\in[t_0-1,t_0]} W(s) \ge\Lambda. \tag*{(42)} s ∈ [ t 0 − 1 , t 0 ] inf W ( s ) ≥ Λ. (42)
Step 4: iteration. Let d k : = 2 − k − 2 d_k := 2^{-k-2} d k := 2 − k − 2 for k ≥ 0 k \ge0 k ≥ 0 , and define τ 0 : = t 0 \tau_0 := t_0 τ 0 := t 0 and τ k + 1 : = τ k + d k \tau_{k+1} := \tau_k + d_k τ k + 1 := τ k + d k , so that τ k < t 0 + 1 2 \tau_k < t_0 + \frac{1}{2} τ k < t 0 + 2 1 for all k k k . Set Λ 0 : = inf [ t 0 − 1 , t 0 ] W \Lambda_0 := \inf_{[t_0-1,t_0]} W Λ 0 := inf [ t 0 − 1 , t 0 ] W and Λ k + 1 : = inf [ τ k , τ k + 1 ] W \Lambda_{k+1} := \inf_{[\tau_k,\tau_{k+1}]} W Λ k + 1 := inf [ τ k , τ k + 1 ] W for k ≥ 0 k \ge0 k ≥ 0 . For t ∈ [ τ k , τ k + 1 ] t \in[\tau_k,\tau_{k+1}] t ∈ [ τ k , τ k + 1 ] we apply (40) with δ = 2 d k ≤ 1 2 \delta= 2d_k \le\frac{1}{2} δ = 2 d k ≤ 2 1 . Since 2 d k = d k − 1 2d_k = d_{k-1} 2 d k = d k − 1 for k ≥ 1 k \ge1 k ≥ 1 , we have [ t − 2 d k , t − d k ] ⊂ [ τ k − 2 d k , τ k + 1 − d k ] = [ τ k − 1 , τ k ] [t-2d_k,t-d_k] \subset[\tau_k-2d_k,\tau_{k+1}-d_k] = [\tau_{k-1},\tau_k] [ t − 2 d k , t − d k ] ⊂ [ τ k − 2 d k , τ k + 1 − d k ] = [ τ k − 1 , τ k ] for k ≥ 1 k \ge1 k ≥ 1 , and [ t − 2 d 0 , t − d 0 ] ⊂ [ t 0 − 1 2 , t 0 ] [t-2d_0,t-d_0] \subset[t_0-\frac{1}{2},t_0] [ t − 2 d 0 , t − d 0 ] ⊂ [ t 0 − 2 1 , t 0 ] for k = 0 k = 0 k = 0 . In both cases the infimum of W W W over this window is at least Λ k \Lambda_k Λ k , and we get
Λ k + 1 ≥ A ( 2 d k ) α Λ k p , k ≥ 0. \Lambda_{k+1} \ge A(2d_k)^\alpha\Lambda_k^p,\qquad k \ge0. Λ k + 1 ≥ A ( 2 d k ) α Λ k p , k ≥ 0.
Let L k : = log Λ k L_k := \log\Lambda_k L k := log Λ k and a : = log A a := \log A a := log A . Since 2 d k = 2 − k − 1 2d_k = 2^{-k-1} 2 d k = 2 − k − 1 , we obtain L k + 1 ≥ p L k + a − α ( k + 1 ) log 2 L_{k+1} \ge pL_k + a - \alpha(k+1)\log2 L k + 1 ≥ p L k + a − α ( k + 1 ) log 2 , and an induction gives
L k ≥ p k ( L 0 + ∑ j = 0 k − 1 p − j − 1 ( a − α ( j + 1 ) log 2 ) ) ≥ p k ( L 0 − K ) , L_k \ge p^k\left(L_0+\sum_{j=0}^{k-1}p^{-j-1}\left(a-\alpha(j+1)\log2\right)\right) \ge p^k(L_0-K), L k ≥ p k ( L 0 + j = 0 ∑ k − 1 p − j − 1 ( a − α ( j + 1 ) log 2 ) ) ≥ p k ( L 0 − K ) ,
where K : = ∑ j ≥ 0 p − j − 1 ( ∣ a ∣ + α ( j + 1 ) log 2 ) < ∞ K := \sum_{j\ge0}p^{-j-1}(|a|+\alpha(j+1)\log2)<\infty K := ∑ j ≥ 0 p − j − 1 ( ∣ a ∣ + α ( j + 1 ) log 2 ) < ∞ . Choosing Λ \Lambda Λ in (42) so large that L 0 ≥ K + 1 L_0 \ge K+1 L 0 ≥ K + 1 , we conclude that L k ≥ p k → ∞ L_k \ge p^k \to\infty L k ≥ p k → ∞ . Hence W W W is unbounded on [ t 0 − 1 , t 0 + 1 2 ] [t_0-1,t_0+\frac{1}{2}] [ t 0 − 1 , t 0 + 2 1 ] . This contradicts the boundedness of u u u on [ 0 , t 0 + 1 ] × R N [0,t_0+1]\times\mathbb{R}^N [ 0 , t 0 + 1 ] × R N , since h h h is bounded there. Therefore u u u blows up in finite time. □
Remark 5.3. For N = θ N = \theta N = θ , the kernel satisfies Z ( t , x ) ≥ c t − α Z(t,x) \ge ct^{-\alpha} Z ( t , x ) ≥ c t − α for ∣ x ∣ ≤ t α / θ |x| \le t^{\alpha/\theta} ∣ x ∣ ≤ t α / θ , and the same proof gives blow-up for 1 < p < 1 + γ / α 1 < p < 1+\gamma/\alpha 1 < p < 1 + γ / α . This range, however, is contained in 1 < p < p c ( γ , α ) 1 < p < p_c(\gamma,\alpha) 1 < p < p c ( γ , α ) when N = θ N=\theta N = θ . For N < θ N<\theta N < θ , the linear decay on the unit ball is t − N α / θ t^{-N\alpha/\theta} t − N α / θ , and the argument gives only p < 1 + θ γ / ( N α ) < p c ( γ , α ) p < 1+\theta\gamma/(N\alpha) < p_c(\gamma,\alpha) p < 1 + θ γ / ( N α ) < p c ( γ , α ) . Thus the local mechanism is new information only when N > θ N>\theta N > θ .
The regime γ ≤ − α \gamma\le-\alpha γ ≤ − α : global existence for every p > 1 p > 1 p > 1
For γ ≤ − α \gamma\le-\alpha γ ≤ − α we have κ = γ + α − β ( p − 1 ) < 0 \kappa= \gamma+\alpha-\beta(p-1)<0 κ = γ + α − β ( p − 1 ) < 0 for every p > 1 p>1 p > 1 . Hence the fixed-point scheme of § 5.1 closes for every p > 1 p>1 p > 1 , in its X β X_\beta X β form under hypothesis (S) and in its L r L^r L r form of Remark 5.1 in general. This proves Theorem 3.1(i) and Theorem 3.2 (a). Two remarks complete the picture.
First, when γ < − 1 \gamma<-1 γ < − 1 , the source is integrable on all of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) . The tail ∫ 1 ∞ t γ ℓ ( t ) d t \int_1^\infty t^\gamma\ell(t)\,dt ∫ 1 ∞ t γ ℓ ( t ) d t converges by Lemma 2.1(K2), while ∫ 0 1 h ( t ) d t < ∞ \int_0^1 h(t)\,dt<\infty ∫ 0 1 h ( t ) d t < ∞ by Convention 2.1; without the convention, the literal density t γ ℓ ( t ) t^\gamma\ell(t) t γ ℓ ( t ) could fail to be integrable at the origin. At α = 1 \alpha=1 α = 1 one may then argue alternatively via Theorem 2.1(i), extended verbatim to the θ \theta θ -stable semigroup. The constant caloric function w ≡ c w\equiv c w ≡ c satisfies ∫ 0 ∞ h ( t ) ∥ w ∥ ∞ p − 1 d t = c p − 1 ∫ 0 ∞ h ( t ) d t < ∞ \int_0^\infty h(t)\|w\|_\infty^{p-1}\,dt=c^{p-1}\int_0^\infty h(t)\,dt<\infty ∫ 0 ∞ h ( t ) ∥ w ∥ ∞ p − 1 d t = c p − 1 ∫ 0 ∞ h ( t ) d t < ∞ . To obtain decaying solutions, one may instead take w = S 1 , θ ( t ) w 0 w=S_{1,\theta}(t)w_0 w = S 1 , θ ( t ) w 0 with 0 ≤ w 0 ∈ L 1 ∩ L ∞ 0\le w_0\in L^1\cap L^\infty 0 ≤ w 0 ∈ L 1 ∩ L ∞ , w 0 ≢ 0 w_0\not\equiv0 w 0 ≡ 0 ; then ∥ w ( t ) ∥ ∞ ≤ ∥ w 0 ∥ ∞ \|w(t)\|_\infty\le\|w_0\|_\infty ∥ w ( t ) ∥ ∞ ≤ ∥ w 0 ∥ ∞ , the same integral is finite, and ∥ w ( t ) ∥ ∞ → 0 \|w(t)\|_\infty\to0 ∥ w ( t ) ∥ ∞ → 0 .
Second, for γ ∈ ( − 1 , − α ] \gamma\in(-1,-\alpha] γ ∈ ( − 1 , − α ] , a non-empty range only when α < 1 \alpha<1 α < 1 , the source is not integrable in time, and yet no Fujita threshold exists. The memory of the Caputo derivative, encoded in the kernel ( t − s ) α − 1 (t-s)^{\alpha-1} ( t − s ) α − 1 , weakens the cumulative effect of the source. Instead of ∫ 0 t h ( s ) d s ∼ t γ + 1 ℓ ( t ) / ( γ + 1 ) \int_0^t h(s)\,ds\sim t^{\gamma+1}\ell(t)/(\gamma+1) ∫ 0 t h ( s ) d s ∼ t γ + 1 ℓ ( t ) / ( γ + 1 ) , the relevant quantity is the fractional average ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≲ t γ + α ℓ ( t ) \int_0^t(t-s)^{\alpha-1}h(s)\,ds\lesssim t^{\gamma+\alpha}\ell(t) ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≲ t γ + α ℓ ( t ) , which tends to 0 0 0 when γ < − α \gamma<-\alpha γ < − α and is of the order of ℓ ( t ) \ell(t) ℓ ( t ) when γ = − α \gamma=-\alpha γ = − α . In either case κ < 0 \kappa<0 κ < 0 for every p > 1 p>1 p > 1 ; the ε \varepsilon ε -room of § 5.1 absorbs the slowly varying factor even when ℓ \ell ℓ is unbounded at the endpoint γ = − α \gamma=-\alpha γ = − α . Hence the fixed-point scheme closes, and no Fujita threshold arises.
Meier’s s ∗ s^* s ∗ in the nonlocal setting
At α = 1 \alpha= 1 α = 1 , the decay rate ∥ S 1 , θ ( t ) w 0 ∥ ∞ ≍ t − N / θ \|S_{1,\theta}(t)w_0\|_\infty\asymp t^{-N/\theta} ∥ S 1 , θ ( t ) w 0 ∥ ∞ ≍ t − N / θ (for 0 ≤ w 0 ∈ L 1 ∩ L ∞ 0 \le w_0 \in L^1 \cap L^\infty 0 ≤ w 0 ∈ L 1 ∩ L ∞ , w 0 ≢ 0 w_0 \not\equiv0 w 0 ≡ 0 ) identifies s ∗ = N / θ s^* = N/\theta s ∗ = N / θ , and Meier’s formula (2) persists verbatim:
p c ( γ ) = 1 + γ + 1 s ∗ = 1 + θ ( γ + 1 ) N , s ∗ = N θ . p_c(\gamma) = 1 + \frac{\gamma+1}{s^*} = 1 + \frac{\theta(\gamma+1)}{N}, \qquad s^* = \frac{N}{\theta}. p c ( γ ) = 1 + s ∗ γ + 1 = 1 + N θ ( γ + 1 ) , s ∗ = θ N .
For α < 1 \alpha< 1 α < 1 , the naive substitution s ∗ = N α / θ s^* = N\alpha/\theta s ∗ = N α / θ is not correct. It would predict the autonomous exponent 1 + θ / ( N α ) 1 + \theta/(N\alpha) 1 + θ / ( N α ) , which contradicts the α \alpha α -independent threshold 1 + θ / N 1 + \theta/N 1 + θ / N of [42 , 7 ] . Moreover, for N > θ N > \theta N > θ , the sup-norm decay saturates at the slower rate t − α t^{-\alpha} t − α (Remark 2.1), so no single sup-norm exponent captures the dynamics. In the scaling (mass) regime, the correct modification acts on the numerator rather than on s ∗ s^* s ∗ . The memory kernel ( t − s ) α − 1 (t-s)^{\alpha-1} ( t − s ) α − 1 of the Duhamel formula replaces the cumulative source ∫ 0 t h ( s ) d s ∼ t γ + 1 ℓ ( t ) / ( γ + 1 ) \int_0^t h(s)\,ds \sim t^{\gamma+1}\ell(t)/(\gamma+1) ∫ 0 t h ( s ) d s ∼ t γ + 1 ℓ ( t ) / ( γ + 1 ) of Meier’s criterion by the fractional average ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≍ t γ + α ℓ ( t ) \int_0^t(t-s)^{\alpha-1}h(s)\,ds \asymp t^{\gamma+\alpha}\ell(t) ∫ 0 t ( t − s ) α − 1 h ( s ) d s ≍ t γ + α ℓ ( t ) , which leads to
p c ( γ , α ) = 1 + γ + α α s ∗ , s ∗ = N θ . p_c(\gamma,\alpha) = 1 + \frac{\gamma+\alpha}{\alpha s^*}, \qquad s^* = \frac{N}{\theta}. p c ( γ , α ) = 1 + α s ∗ γ + α , s ∗ = θ N .
This formula reduces to Meier’s at α = 1 \alpha= 1 α = 1 and to the α \alpha α -independent value 1 + θ / N 1+\theta/N 1 + θ / N at γ = 0 \gamma= 0 γ = 0 . When N > θ N > \theta N > θ , however, the saturated decay t − α t^{-\alpha} t − α on bounded sets produces the second constraint of Proposition 3.1. In terms of decay rates, it reads p > 1 + γ / s l o c p > 1+\gamma/s_{\mathrm{loc}} p > 1 + γ / s loc with s l o c : = α s_{\mathrm{loc}} := \alpha s loc := α , the exponent of local decay, so a single s ∗ s^* s ∗ no longer suffices. The persistence of s ∗ = N / θ s^* = N/\theta s ∗ = N / θ as the geometric ingredient of the scaling mechanism suggests that the same framework should govern problems on cones, exterior domains and weighted spaces, with s ∗ s^* s ∗ encoding the geometry-dependent decay rate of the underlying linear evolution.
Conclusion and open problems
We have classified the Fujita phenomenon for the reaction–diffusion equation (1) with source intensity h ( t ) = t γ ℓ ( t ) h(t) = t^\gamma\ell(t) h ( t ) = t γ ℓ ( t ) , where ℓ \ell ℓ is slowly varying. The classification is complete for the classical time derivative α = 1 \alpha= 1 α = 1 and the full range of stable generators θ ∈ ( 0 , 2 ] \theta\in(0,2] θ ∈ ( 0 , 2 ] . For the Caputo derivative α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , it is complete when N < θ N < \theta N < θ ; for N ≥ θ N \ge\theta N ≥ θ , the blow-up range 1 < p < p c ( γ , α ) 1 < p < p_c(\gamma,\alpha) 1 < p < p c ( γ , α ) is established in every dimension, and a precise conjecture is formulated. Three features stand out. First, the slowly varying factor never shifts the critical exponent but decides the behavior at it. At α = 1 \alpha= 1 α = 1 this happens through the single Karamata integral ∫ 1 ∞ t − 1 ℓ ( t ) d t \int_1^\infty t^{-1}\ell(t)\,dt ∫ 1 ∞ t − 1 ℓ ( t ) d t ; in particular, global existence may occur at the critical exponent, in contrast with the autonomous theory. For α < 1 \alpha< 1 α < 1 and N < θ N < \theta N < θ , the boundedness of ℓ \ell ℓ plays this role. Second, the memory of the Caputo derivative shifts the exponent of a time-dependent source, replacing γ + 1 \gamma+ 1 γ + 1 by γ + α \gamma+ \alpha γ + α , even though it leaves the autonomous threshold 1 + θ / N 1 + \theta/N 1 + θ / N untouched. Third, for N > θ N > \theta N > θ and large γ \gamma γ , a purely local blow-up mechanism, driven by the singularity of the fundamental solution, overrides the scaling exponent.
Several directions remain open.
Prove Conjecture 3.1, that is, global existence above p ∗ ( γ , α ) = max { p c ( γ , α ) , 1 + γ / α } p_*(\gamma,\alpha) = \max\{p_c(\gamma,\alpha), 1+\gamma/\alpha\} p ∗ ( γ , α ) = max { p c ( γ , α ) , 1 + γ / α } for N > θ N > \theta N > θ , and determine the behavior at p = p ∗ ( γ , α ) p = p_*(\gamma,\alpha) p = p ∗ ( γ , α ) . The critical-space approach of [42 , 7 ] is a natural starting point.
Determine the precise blow-up rate and profile in the presence of slowly varying coefficients, especially at the critical exponent p = p c ( γ ) p = p_c(\gamma) p = p c ( γ ) with a divergent integral ∫ 1 ∞ t − 1 ℓ ( t ) d t \int_1^\infty t^{-1}\ell(t)\,dt ∫ 1 ∞ t − 1 ℓ ( t ) d t . For the classical autonomous case the self-similar blow-up profile is well understood. For slowly varying corrections, logarithmic-time refinements should be expected.
Investigate the influence of a slowly varying factor on the second critical exponent , in the sense of Lee and Ni [23 ] , which governs the long-time behavior for initial data with prescribed spatial decay.
Extend the analysis to spatially localised sources h ( t , x ) = t γ ℓ ( t ) ∣ x ∣ σ h(t,x) = t^\gamma\ell(t)|x|^\sigma h ( t , x ) = t γ ℓ ( t ) ∣ x ∣ σ of Hardy–Hénon type, in the spirit of [20 , 11 , 26 , 27 ] , and to equations with an additional forcing term, as in [4 , 5 ] .
Treat the corresponding problem on sectorial or conical domains, or on exterior domains, where s ∗ s^* s ∗ is modified by the boundary geometry, thereby generalising [3 , 25 ] .
Investigate non-monotone slowly varying functions, such as ℓ ( t ) = 2 + sin ( log log t ) \ell(t)=2+\sin(\log\log t) ℓ ( t ) = 2 + sin ( log log t ) . The critical-case integral condition ∫ 1 ∞ t − 1 ℓ ( t ) d t \int_{1}^{\infty}t^{-1}\ell(t)\,dt ∫ 1 ∞ t − 1 ℓ ( t ) d t still determines the dichotomy, but finer information on the blow-up time may depend on the oscillations.
Declarations
Funding. This research received no external funding.
Conflict of interest. The authors declare that they have no conflicts of interest.
Data availability. No data were generated or analyzed in this study.
References
(Ahmad Z. Fino) DEPARTMENT OF MATHEMATICS, FACULTY OF SCIENCES, LEBANESE UNIVERSITY, P.O. BOX 1352 TRIPOLI, LEBANON.
(M. Majdoub) DEPARTMENT OF MATHEMATICS, COLLEGE OF SCIENCE, IMAM ABDULRAHMAN BIN FAISAL UNIVERSITY, P.O. BOX 1982, DAMMAM, SAUDI ARABIA.
BASIC AND APPLIED SCIENTIFIC RESEARCH CENTER, IMAM ABDULRAHMAN BIN FAISAL UNIVERSITY, P.O. BOX 1982, 31441, DAMMAM, SAUDI ARABIA.
Email address: afino@ul.edu.lb
Email address: mmajdoub@iau.edu.sa