AI tools and theory building: Gradients on metric measure spaces

Gradient structures on metric measure spaces

This project grew out of curiosity both mathematical and about the use of the latest AI generative tools in mathematical research. It is by now well-established that the latest AI models are very efficient at problem solving in mathematics, but are they already good at theory building ? I was curious to test this on a possible theory that I had in mind for some time. This was not a shot in the dark since I had ideas from the beginning and since the “rank one” case of the theory was already quite well understood and developed in some of my previous works on the Vicsek set and local trees. There was therefore something concrete to extrapolate from. The main idea was to develop a finite rank gradient theory on a large class of metric spaces which might be fractals. This idea is of course not new, and the closest antecedent to the approach I had in mind is probably Hino’s theory of gradient and martingale dimension. However, Hino’s theory is L2L^2 based on Dirichlet forms and I was looking for a LpL^p theory for every pp with a gradient essentially independent from pp.

The way models were used to develop the theory is described in the preface of the document. I think this took about 200 prompts, 40 hours on Astra 6.1, 20h on Sol 6.1 and 20h on Fable 5.1 and 5.2. My role in the project was like an architect, and I am relatively satisfied by the end result and learnt a lot. Though already useful I think, this a first draft which is not intended to be published for now since the document is almost entirely AI written. There will be further digestion of the material and some rewriting but I think the experience was quite productive, interesting and is illustrative of what future work might be for mathematicians. As AI systems become increasingly capable at problem solving and even theory building, one still choose the directions to pursue, advance mathematical knowledge and human understanding along the paths one judge relevant, and, above all, keep asking the right questions.

1. From the Vicsek set to higher rank spaces

With its intrinsic metric, the Vicsek set is a real tree. Fix a root oo and orient the arcs away from it. A Lipschitz function can be differentiated along each arc. If ν\nu is length measure on the skeleton 𝒮\mathcal S, then

u(y)−u(o)=∫[o,y]∂udν. u(y)-u(o)=\int_{[o,y]}\partial u\,d\nu.

The reference measure mm is the Hausdorff measure of dimension log5/log3\log 5/\log 3 and gives zero mass to 𝒮\mathcal S. In my work with Li Chen on the Vicsek set and with Li Chen and Meng Yang on local trees, we used this scalar derivative to develop Sobolev spaces with u∈Lp(X,m)u\in L^p(X,m) and ∂u∈Lp(𝒮,ν)\partial u\in L^p(\mathcal S,\nu).

The level-four Vicsek cable approximation with the arc from the root o to y highlighted in blue.
The level-four Vicsek cable approximation and the arc from o to y.

Having well understood the rank one geometry given by trees, products of trees were at first a natural place to look for higher rank examples. However, the natural product of two trees with singular length measures still has rank one ! Indeed, the two partial derivatives are carried by mutually singular measures, so only one direction is present at almost every point for the gradient measure. A first genuine higher example was obtained by the model Astra which I prompted to find a higher dimensional analogue to cable system limits. In the resulting tree of cubes, the ordinary nn-dimensional gradients on the pieces survive in the limit, and the measure carrying them is singular with respect to mm.

Three stages of a tree of squares, showing successively smaller squares attached to the earlier ones.
Stages X2, X4 and X6 of the tree of cubes in dimension two. Each mesh square has four dyadic children and one new square attached along a segment.

2. Gradient structures

More abstractly, a convenient and usable definition is the following. On a metric space (X,d)(X,d), a gradient structure consists of a Borel set 𝒮⊂X\mathcal S\subset X, a σ\sigma-finite measure ν\nu carried by 𝒮\mathcal S, and a linear map

∇:Lipb⁡(X)⟶L∞(𝒮,ℋ,ν). \nabla:\operatorname{Lip}_b(X)\longrightarrow L^\infty(\mathcal S,\mathcal H,\nu).

Here ℋ=(Hx)x∈𝒮\mathcal H=(H_x)_{x\in\mathcal S} is the smallest measurable field of subspaces of a fixed ℝN\mathbb R^N containing the gradients. The map satisfies the Leibniz rule and |∇u|≤Lip(u)|\nabla u|\le\operatorname{Lip}(u) almost everywhere. It takes pointwise convergence of uniformly bounded, uniformly Lipschitz sequences to weak-** convergence. The dimension of HxH_x is the pointwise rank, and its ν\nu-essential supremum is the rank n*n_*.

The reference measure enters when we close the gradient. We view ∇\nabla as an operator from Lp(X,m)L^p(X,m) to Lp(𝒮,ℋ,ν)L^p(\mathcal S,\mathcal H,\nu) on the compactly supported Lipschitz functions whose gradients belong to the latter space. The structure is pp-admissible if this core is uniformly dense in Cc(X)C_c(X) and the operator is closable. The domain of its closure is then W1,p(X,m)W^{1,p}(X,m). For p=2p=2 this gives a strongly local regular Dirichlet form. When the core gradients span the fibres, n*n_* is the martingale dimension of the associated diffusion.

3. The rank theorem

On the tree of cubes, weak monotonicity at the critical Korevaar–Schoen exponent fails for p=2p=2. This rules out a common scale for the Poincaré inequality and the cutoff energy estimate which is usually assumed in the current literature in Dirichlet spaces. For the classical theory with a common scale, see Barlow, Bass and Kumagai and Grigor’yan, Hu and Lau.

I was surprised by this two-scale feature and spent quite some time to try to construct higher rank examples with the same scale. It turned out to be impossible. The rank theorem explains the obstruction. Suppose a pp-admissible gradient structure, with p>1p>1, has doubling reference measure and a core that spans the fibres. If the pp-Poincaré inequality and the cutoff energy bound hold at a common scale Ψ\Psi whose lower scaling exponent is greater than pp, as for Ψ(r)=rβ\Psi(r)=r^\beta with β>p\beta>p, then

n∗<p. n_*<p.

At p=2p=2, a common scale of order strictly higher than r2r^2 therefore forces rank one for a nontrivial gradient structure. In particular, this applies when two-sided sub-Gaussian heat kernel estimates with walk dimension greater than two give the common Poincaré and capacity scale. The proof uses functions with independent gradients as coordinates and a capacity argument related to Hino’s work on martingale dimension.

4. Heat kernels with two scales

What heat kernel estimates can one obtain when the scales differ? In a geodesic, strongly local regular Dirichlet space with volume doubling, let ΨP\Psi_P be the scale in the Poincaré inequality and ΨS\Psi_S the scale in the cutoff Sobolev inequality in annuli. The latter controls cutoff energy weighted by the square of a test function. See Andres and Barlow.

With the Poincaré inequality on the same ball on both sides, the parabolic Harnack inequality on a cylinder of radius rr and time length 2T2T takes the form

ess supQ−⁡u≤exp⁡[C(1+ΨP(r)T+TΨS(r))]ess infQ+⁡u, \operatorname*{ess\,sup}_{Q_-}u \le \exp\!\left[C\left(1+\frac{\Psi_P(r)}{T}+\frac{T}{\Psi_S(r)}\right)\right] \operatorname*{ess\,inf}_{Q_+}u,

where Q−Q_- and Q+Q_+ are earlier and later interior subcylinders. The first ratio comes from the Poincaré estimate inside the ball. The second comes from localizing the equation with a cutoff.

The two scales also enter the upper bound in different ways. With uniform volume growth m(B(x,r))≍rQm(B(x,r))\asymp r^Q and power scales ΨP(r)=ra\Psi_P(r)=r^a, ΨS(r)=rβ\Psi_S(r)=r^\beta with β≥2\beta\ge2, the small-time bound is

pt(x,y)≤Ct−Q/aexp⁡[−c(d(x,y)βt)1/(β−1)]. p_t(x,y)\le C t^{-Q/a} \exp\!\left[-c\left(\frac{d(x,y)^\beta}{t}\right)^{1/(\beta-1)}\right].

The Poincaré scale determines the prefactor in this bound, and the cutoff Sobolev scale determines the off-diagonal decay. The lower bounds involve both scales. When the scales agree, the estimates recover the usual two-sided bounds. At this stage, I don’t think the theory is complete and I believe we should be able to obtain cleaner results for the heat kernel estimates.

5. Self-improvement of Poincaré inequalities

I think that a feature of a good theory is that one can prove interesting theorems in it, so I tried to test this. Because the same gradient is defined at every exponent on the Lipschitz core where its energy is finite, several natural questions arise. In particular, the so-called self-improvement property of the Poincaré inequalities easily comes to mind. The question comes from the theorem of Keith and Zhong.

On a proper geodesic space with doubling reference measure, start with a (1,p)(1,p)-Poincaré inequality at the cutoff scale and a cutoff estimate at some exponent 1≤b<p1\le b<p. One obtains an (s,q)(s,q)-Poincaré inequality with q<p<sq<p<s, with the gradient and the oscillation integrated over the same ball. Thus, if the scales match at p>1p>1, they also match at all exponents sufficiently close to pp from below under the standing cutoff condition. A set of small mm-measure can have infinite ν\nu-measure, so the proof uses cutoff energy to control the cost of modifying a function on such a set. I believe some extensions might be possible in situations where the scales do not match.

6. Examples

What is a good theory without examples ?

  • Manifolds, finite Riemannian gluings and RCD(K,N)\mathrm{RCD}(K,N) spaces. Under the hypotheses in the draft, smooth or measurable Riemannian metrics, sub-Riemannian manifolds, finite Riemannian gluings and finite-dimensional RCD\mathrm{RCD} spaces give m=νm=\nu and matching quadratic scales r2r^2.
  • Local trees. With uniform ball volume Φ(r)\Phi(r), the length derivative has rank one and the matching scale at exponent pp is rp−1Φ(r)r^{p-1}\Phi(r).
  • Cubical inverse limits of Laakso type. These realize each integer rank n≥1n\ge1 and each volume dimension Q>nQ>n, with m=νm=\nu and Gaussian scales.
  • The harmonic Sierpiński gasket. The harmonic embedding gives a rank-one gradient and critical Korevaar–Schoen exponent 11 for every finite p≥1p\ge1. Ball volumes depend on the centre.
  • Liouville Brownian motion. Here the Euclidean gradient has rank two, ν\nu is planar Lebesgue measure, and mm is Gaussian multiplicative chaos with parameter 0<γ<20<\gamma<2. Almost surely, the gradient is closable exactly for p>γ2/2p>\gamma^2/2 among 1≤p<∞1\le p<\infty.
  • The tree of cubes. The rank is nn, the reference measure is Ahlfors QQ-regular with Q>nQ>n, and m⊥νm\perp\nu. The critical Korevaar–Schoen exponent is 1+(Q−n)/p1+(Q-n)/p for each finite p≥1p\ge1, and the two scales match precisely for p>np>n.
Posted in Global analysis in Dirichlet spaces | Leave a comment

Intermittency for the parabolic Anderson model on spaces with sub-Gaussian heat kernel estimates

With Hongyi Chen, Li Chen and Søren Kjærsgaard, we have posted a new paper, Intermittency for the parabolic Anderson model on spaces with sub-Gaussian heat kernel estimates.

This is a sequel to our earlier work on bounded domains of recurrent metric measure spaces, where we assumed strong recurrence, dh<dwd_h<d_w, and considered Dirichlet and Neumann boundary conditions. The new paper treats unbounded spaces and allows critical and transient diffusion. Even in the strongly recurrent setting, it provides stronger moment lower bounds, Hölder regularity, and the existence of long-time moment growth rates for every real order.

1. Geometry and the equation

We study the stochastic heat equation

∂tu(t,x)=Δu(t,x)+βu(t,x)W˙α(t,x),u(0,⋅)=u0.\partial_tu(t,x)=\Delta u(t,x)+\beta u(t,x)\dot W_\alpha(t,x),\qquad u(0,\cdot)=u_0.

Throughout, β>0\beta>0. The underlying Dirichlet space has volume growth μ(B(x,r))≍rdh\mu(B(x,r))\asymp r^{d_h} and two-sided sub-Gaussian heat kernel estimates with walk dimension dwd_w. The precise hypotheses are stated in the paper.

The noise is white in time. In space, it is white when α=0\alpha=0 and has covariance given by the kernel of (−Δ)−2α(-\Delta)^{-2\alpha} when α>0\alpha>0. The admissible parameter range in the Itô–Walsh framework is

0≤2α<dhdw<1+2α.0\le 2\alpha<\frac{d_h}{d_w}<1+2\alpha.

In particular, space-time white noise is admissible only when dh<dwd_h<d_w. In Euclidean space, dwd_w = 2, so this gives the usual dimensional restriction d = 1.

One property we want to understand is intermittency: as time increases, high peaks concentrate in small regions of the space. A useful way to quantify this behavior is through moments. Heuristically, a moment of order p much larger than the pth power of the first moment indicates large fluctuations. This motivates the study of moment Lyapunov exponents.

2. Real-order moment Lyapunov exponents

For bounded initial conditions, the equation has a unique mild solution with finite moments on bounded time intervals. A central result is the existence of the limit

λp(β)=limt→∞⁡1tlog⁡𝔼[u(t,x)p],p≥1.\lambda_p(\beta)=\lim_{t\to\infty}\frac1t\log\mathbb E[u(t,x)^p],\qquad p\ge1.

For every real p≥1p\ge1, this limit is finite and independent of xx and of the bounded nonnegative initial condition u0u_0, provided u0u_0 is nonzero on a set of positive measure.

For integer moment orders, the Feynman–Kac representation relates the moments to an associated Schrödinger semigroup, and spectral theory can then identify their exponential growth rate. Our argument applies to every real order p ≥ 1. ChatGPT Astra 6 first suggested a rather involved approach, which we subsequently simplified using ideas from The stochastic heat equation with multiplicative Lévy noise: Existence, moments, and intermittency by Quentin Berger, Carsten Chong and Hubert Lacoin.

3. Matching moment bounds and intermittency

For p≥2p\ge2, we prove

λp(β)≍p(β2(p−1))dw(1+2α)dw−dh.\lambda_p(\beta)\asymp p\bigl(\beta^2(p-1)\bigr)^{\frac{d_w}{(1+2\alpha)d_w-d_h}}.

The two comparison constants are independent of pp and β\beta. The lower bound holds for every real p>1p>1, including the white-noise case. For 1<p<21<p<2, the upper estimate is instead obtained by interpolation:

λp(β)≤(p−1)λ2(β).\lambda_p(\beta)\le(p-1)\lambda_2(\beta).

Moreover, λ1(β)=0\lambda_1(\beta)=0 and p↦λp(β)/pp\mapsto\lambda_p(\beta)/p is strictly increasing on (1,∞)(1,\infty). Thus normalized moments of different orders separate exponentially. This is the moment formulation of full intermittency.

The lower-bound argument adapts the Gaussian exponential test of Chen, Deya, Ouyang and Tindel, combined with estimates for confinement of the diffusion to a ball.

4. Hölder regularity

Suppose the heat semigroup satisfies, for some κ∈(0,1]\kappa\in(0,1], the weak Bakry–Émery estimate

|Ptf(x)−Ptf(y)|≤Cd(x,y)κtκ/dw‖f‖∞.|P_tf(x)-P_tf(y)|\le C\frac{d(x,y)^\kappa}{t^{\kappa/d_w}}\|f\|_\infty.

The solution has a modification which is locally Hölder continuous away from time zero, with any time and space exponents satisfying

0<γt<(1+2α)dw−dh2dw,0<γx<min⁡{κ,(1+2α)dw−dh2}.0<\gamma_t<\frac{(1+2\alpha)d_w-d_h}{2d_w},\qquad 0<\gamma_x<\min\left\{\kappa,\frac{(1+2\alpha)d_w-d_h}{2}\right\}.

5. The infinite Sierpiński gasket

Let KK be the compact gasket of diameter 11, rooted at the origin, and consider its unbounded blow-up

X=⋃n≥02nK.X=\bigcup_{n\ge0}2^nK.

Here

dh=log⁡3log⁡2,dw=log⁡5log⁡2.d_h=\frac{\log3}{\log2},\qquad d_w=\frac{\log5}{\log2}.

For space-time white noise, the moment estimate becomes

λp(β)≍p(β2(p−1))log⁡5log⁡(5/3),p≥2.\lambda_p(\beta)\asymp p\bigl(\beta^2(p-1)\bigr)^{\frac{\log5}{\log(5/3)}},\qquad p\ge2.

At large moment orders, the powers of β\beta and pp are approximately 6.306.30 and 4.154.15, respectively. They are determined by the geometry of the gasket.

Since its heat semigroup has exponent κ=dw−dh\kappa=d_w-d_h, the white-noise solution is locally Hölder continuous with any exponents

0<γt<log⁡(5/3)2log⁡5,0<γx<log⁡(5/3)2log⁡2.0<\gamma_t<\frac{\log(5/3)}{2\log5},\qquad 0<\gamma_x<\frac{\log(5/3)}{2\log2}.

In the white-noise case, there is also an exact discrete scaling law. With initial condition u0≡1u_0\equiv1 and noise intensity indicated by a superscript,

uβ(5t,2x)=lawuβ5/3(t,x).u^\beta(5t,2x)\overset{\mathrm{law}}=u^{\beta\sqrt{5/3}}(t,x).

Consequently,

λp(β5/3)=5λp(β).\lambda_p\!\left(\beta\sqrt{5/3}\right)=5\lambda_p(\beta).

This scaling fixes the power-law scale while allowing a modulation periodic in logβ\log\beta.

The pictures below show one realization with space-time white noise, β=1.5\beta=1.5 and u0≡1u_0\equiv1, at times 00, 0.50.5, 11 and 22. The observation cell is D=16KD=16K, of diameter 1616. The computation uses the larger truncation 32K32K, with mesh 1/161/16; there are 98439843 displayed vertices. The two remote corners carry zero Dirichlet conditions, while the origin retains the natural reflecting condition.

The same realization at times t=0t=0 and t=0.5t=0.5. Heights represent the unnormalized solution, with a common scale for all four panels.

The realization at times t=1t=1 and t=2t=2. The numbers below the panels are spatial averages over the observation cell.

These averages approximate

u‾D(t)=1μ(D)∫Du(t,x)dμ(x).\bar u_D(t)=\frac1{\mu(D)}\int_Du(t,x)\,d\mu(x).

At t=2t=2, the largest displayed value is about 34.234.2, while the spatial average is about 1.361.36. The pictures illustrate spatial concentration in a finite graph approximation. The intermittency theorem concerns the asymptotic growth of ensemble moments.

We also simulated the stronger noise intensity β=5\beta=5, with the same geometry, α=0\alpha=0 and u0≡1u_0\equiv1. This is an independent realization. The four panels show actual solution heights on separate vertical scales.

One realization for β=5\beta=5 at times 00, 0.50.5, 11 and 22. Note the powers of ten on the vertical axes.

The additional snapshot at t=0.1t=0.1 shows an early burst: the maximum is approximately 56.556.5, whereas the spatial average is 0.3760.376.

The same β=5\beta=5 realization at t=0.1t=0.1.

The displayed solution subsequently becomes very small, while remaining concentrated: at t=2t=2, its maximum is about 396396 times the spatial average.

Interactive simulation. Open the viewer to choose the noise intensity β from 0 to 5 in steps of 0.1 and play the solution from time 0 to 2. On a fixed mesh, changing β reuses the same Brownian increments.

The new paper contains the full statements and proofs, the scaling argument, and the details of the numerical approximation.

AI tools were only used at the end of the project when we already had a substantial manuscript in our hands. They were mostly used for polishing, proofreading, proof presentations and numerical simulations. However, ChatGPT Astra 6 discovered the approach for the existence of moments of real orders, that we then simplified later; we had a proof based on Feynman-Kac formulas that only worked for moments of integer order. ChatGPT Sol 6.1 was used in assisting the writing of this post.

Posted in Global analysis in Dirichlet spaces, Stochastic PDE | Tagged , | Leave a comment

Riesz transform on eventually Gaussian local trees

With Aobo Chen and Li Chen, we have posted a new paper, Riesz transform on eventually Gaussian local trees which is a companion paper to the paper discussed in the previous post to which we refer for background about Riesz transform on manifolds. In this work, we study spaces which have a fractal structure at small scales but Gaussian diffusion at large scales. On a family of Vicsek fractafolds, we obtain a complete description of the Riesz and reverse Riesz inequalities. The interesting feature is that the ranges of exponents are exactly reversed compared with those on Vicsek manifolds.

More precisely, the starting point for us was the work of Li Chen, Thierry Coulhon, Joseph Feneuil and Emmanuel Russ. They prove weak type (1,1) and boundedness for 1<p\leq2 under volume doubling and heat kernel upper bounds which are Gaussian at small times and sub-Gaussian at large times.

Their examples include Vicsek manifolds: smooth manifolds constructed by replacing the edges of a Vicsek graph by tubes. Locally, diffusion has the usual Gaussian behavior, but it has a sub-Gaussian behavior at large scales. On these manifolds, the Riesz inequality holds exactly for 1<p\leq2, and its reverse exactly for 2\leq p<\infty.

Our spaces have the opposite arrangement of scales. They retain genuine fractal structure at arbitrarily small distances, while their large-scale organization produces Gaussian diffusion. To describe this, let \Psi(r) denote the characteristic diffusion time at distance r. A model case is

\displaystyle \Psi(r)\asymp\begin{cases}r^\beta,&0<r\leq1,\\r^2,&r\geq1,\end{cases}\qquad\beta>2.

The word eventually refers to large distances, or equivalently to long times. The space itself remains singular at small scales.

We work on uniform local trees: geodesic spaces whose sufficiently small balls are real trees, with a radius that can be chosen uniformly over the space. Global loops are allowed. The tree structure gives a concrete derivative along arcs.

There are two measures. The reference measure m (the Hausdorff measure in most examples) describes volume and is used to integrate functions. The length measure \nu is carried by the skeleton S, the union of the open geodesic arcs, and is used to integrate their weak derivatives \partial f. The Dirichlet form is

\displaystyle \mathcal E(f,f)=\int_S|\partial f|^2\,d\nu.

This energy defines the Laplacian on L^2(m). The Riesz transform \partial(-\Delta)^{-1/2} is initially defined on L^2(m) with values in L^2(\nu). The boundedness question asks for which exponents it extends to a bounded map

\displaystyle \mathcal R=\partial(-\Delta)^{-1/2}:L^p(m)\longrightarrow L^p(\nu).

In the fractal examples, m and \nu are mutually singular.

Our result assumes uniform volume growth, two-sided heat kernel estimates with scale function \Psi, and a compatible pointwise gradient estimate for the heat kernel. Under eventual Gaussian scaling, together with a local Dini condition, we prove

\displaystyle \|\mathcal R f\|_{L^p(\nu)}\leq C_p\|f\|_{L^p(m)},\qquad 2\leq p<\infty.

The Dini condition requires, for some q>2,

\displaystyle \int_0^1\left(\frac{\Psi(r)}{r^2}\right)^{1/2-1/q}\frac{dr}{r}<\infty.

It holds whenever \Psi(r)\lesssim r^\beta near zero for some \beta>2, so it includes the model above. Duality then gives the reverse inequality

\displaystyle \|(-\Delta)^{1/2}f\|_{L^p(m)}\leq C_p\|\partial f\|_{L^p(\nu)},\qquad 1<p\leq2,

for functions in the energy domain with finite right-hand side.

The obstruction on the other side of 2 is expressed by a rigidity theorem. Assuming uniform volume growth and two-sided heat kernel estimates, boundedness of \mathcal R for even one exponent 1<p<2, or a reverse inequality for even one exponent p>2, forces

\displaystyle m(B(x,r))\asymp r,\qquad\Psi(r)\asymp r^2

uniformly at small radii. It also forces m and \nu to be comparable as measures. Thus these inequalities rule out the genuinely fractal local geometry of our examples. This necessity result does not require the pointwise heat kernel gradient estimate used in the positive theorem.

We study a concrete family of examples in the paper. The alternating Vicsek fractafold is obtained by joining compact Vicsek sets in a periodic arrangement. For each integer d\geq2, it is locally fractal and has the large-scale geometry of \mathbb Z^d. Write

\displaystyle \alpha_d=\frac{\log(2^d+1)}{\log3}.

Its volume growth and diffusion scale satisfy

\displaystyle m(B(x,r))\asymp\begin{cases}r^{\alpha_d},&0<r\leq1,\\r^d,&r\geq1,\end{cases}

\displaystyle \Psi(r)\asymp\begin{cases}r^{\alpha_d+1},&0<r\leq1,\\r^2,&r\geq1.\end{cases}

Since \alpha_d>1, the positive theorem and the rigidity theorem give the complete answer. The comparison with the Vicsek manifolds is:

PropertyVicsek manifoldsVicsek fractafolds
Small-scale diffusionGaussianSub-Gaussian
Large-scale diffusionSub-GaussianGaussian
Riesz inequality1<p\leq22\leq p<\infty
Reverse Riesz inequality2\leq p<\infty1<p\leq2

The main idea of the proof is to separate the two scales. With P_t=e^{t\Delta}, we write

\displaystyle \mathcal R=\mathcal R(I-P_1)+\mathcal RP_1.

The first term contains the high frequencies and is controlled using the local gradient estimates and the Dini condition. The second term contains the low frequencies. For this part, we discretize the space by a graph and apply a good-lambda argument to local energy norms of the gradient. This is where the large-scale Gaussian behavior enters.

Similarly to the paper, about Riesz transforms on the Vicsek set, AI tools were used during the preliminary phase of the work to explore different directions and strategies and again in the final phase for polishing and proofreading. The writing took a few months. The authors are responsible for the content.

References

Posted in Global analysis in Dirichlet spaces | Tagged | Leave a comment

Riesz transforms on the Vicsek set

This post is about a long paper with Aobo Chen and Li Chen, Riesz transforms on the Vicsek set. I want to explain, without too much technicalities, what we prove and put the results into context.

As a side note, I point out that artificial intelligence tools have contributed in a substantial way in the conception and the architecture of some the arguments. In particular, the proof of the failure of the strong L^p bounds at the critical exponent is essentially due to ChatGPT 5.5. pro. (After research, it appears that the strategy relies on methods previously used by Fefferman). Concretely, AI tools were used during the preliminary phase of the work to explore different directions and strategies and again in the final phase for polishing and proofreading. The writing took a few months. The authors are responsible for the content.

1. Riesz transforms on manifolds

On \mathbb{R}^n the Riesz transform is the operator \nabla(-\Delta)^{-1/2} and its L^p-boundedness for 1<p<\infty is related to the two inequalities,

\displaystyle \|\nabla f\|_{p}\le C\,\|(-\Delta)^{1/2}f\|_{p}\qquad\text{and}\qquad \|(-\Delta)^{1/2}f\|_{p}\le C\,\|\nabla f\|_{p},

the Riesz and the reverse Riesz inequality. For p=2, by spectral theory, both are identities with C=1 on any complete Riemannian manifold, since \int|\nabla f|^2=\int f(-\Delta f). For p\ne2 geometry enters.

Coulhon and Duong proved in 1999 that under volume doubling and a Gaussian upper bound on the heat kernel, the Riesz transform is of weak type (1,1) and bounded on L^p for 1<p\le2 [CD99], and conjectured that this holds on every complete manifold with no assumption at all [CD03, Conjecture 1.1]. After intermediate steps (notably [CCFR17], which only needs sub-Gaussian bounds and thus covers fractal-like manifolds), the conjecture was recently settled in September 2026 by Rui Chen, Renjin Jiang, Bo Li and Hong-Quan Li [CJLL26]: on every complete non-compact Riemannian manifold, the Riesz transform is of weak type (1,1), with constant 2 for real-valued functions, hence bounded on L^p for 1<p\le2.

On the other hand, for p>2 the Riesz inequality genuinely requires some geometry: it holds for all p under non-negative Ricci curvature (Bakry), it is characterised under doubling and Poincaré by gradient bounds for the heat semigroup [ACDH04], it fails for large p on connected sums of Euclidean spaces [CD99], and it fails for every p>2 on spaces with a slow, sub-Gaussian diffusion [Fen26a]. The reverse inequality for p' follows from the Riesz inequality for p by duality, so [CJLL26] gives it for all p\ge2; for p<2 it needs Poincaré-type hypotheses [AC05, DR22] and fails on Vicsek manifolds and graphs [CCFR17].

Therefore: on manifolds the exponent 1/2 is universal, and the range 1<p\le2 is free of any assumption. The situation turns out to be quite different on fractals: In particular the exponent 1/2 has to be replaced by an exponent that depends on p.

2. The Vicsek set

The world of fractal sets is a mathematically interesting world in which some familiar functional analytic results (like Sobolev or isoperimetric inequalities) often take a very different form with the appearance of new critical exponents which reflect the rough nature of the underlying set.. With my co-authors we were interested in looking for analogues of the Riesz inequalities in this world. Thanks to recent works with Li Chen [BC23,BC24], we have a solid understanding of Sobolev theory in a specific fractal, the Vicsek set. It therefore was a convenient framework to understand what the results could look like. The Vicsek set is the self-similar cross made of five copies of itself scaled by 1/3; we use its unbounded version X, of Hausdorff dimension d_f=\log_35, with its geodesic distance and Hausdorff measure m. Notably, it is a tree: two points are joined by a unique arc, and the union of all arcs, the skeleton \mathcal S, is a countable union of segments carrying a length measure \nu. Functions that are absolutely continuous along segments have a derivative \partial f on the skeleton, which gives Sobolev spaces W^{1,p}(X) (f\in L^p(m), \partial f\in L^p(\nu)) [BC23] and the Dirichlet form \mathcal E(f,f)=\int_{\mathcal S}|\partial f|^2\,d\nu on L^2(X,m), with Laplacian \Delta and heat semigroup P_t. The heat kernel has sub-Gaussian estimates with walk dimension d_w=d_f+1.

Two features make the situation different from the smooth setting . The measures m and \nu are mutually singular: the skeleton has m-measure zero, so gradients live on a set of Hausdorff measure zero ! In Dirichlet form language there is no carré du champ, and X falls outside the framework of [CJLL26].

Self-similarity decides the critical exponent for the Riesz inequalities. Dilating a function f with f_n(x)=f(3^{-n}x) multiplies \|\partial f_n\|_{L^p(\nu)} by 3^{n(1/p-1)} and \|(-\Delta)^\gamma f_n\|_{L^p(m)} by 3^{n(d_f/p-d_w\gamma)} (the measure scales by 5^n, the Laplacian by 3^{-nd_w}). An inequality between the two, uniform in n\in\mathbb Z, forces

\displaystyle \gamma=\gamma_p:=\frac{d_f+p-1}{p\,d_w},\qquad\text{i.e.}\qquad \gamma_p-\frac12=\frac{d_f-1}{d_w}\Big(\frac1p-\frac12\Big).

So the natural Riesz transform is \mathcal R_p=\partial(-\Delta)^{-\gamma_p}, with an exponent that depends on p and which deviates from the manifold universal exponent 1/2. We note that the same exponent appears as a Besov critical exponent, in the heat kernel gradient bounds of [BC23, BC24], on the Vicsek cable system [DR26] and on Vicsek graphs [Fen26b].

3. Our results

Off the critical exponent. For the heat regularisation similar to Li Chen’s quasi-Riesz inequalities [Che15], we first prove that the reverse inequality \|(-\Delta)^\gamma P_1f\|_{L^p(m)}\le C\|\partial f\|_{L^p(\nu)} holds for every p\in[1,\infty) and \gamma\in(\gamma_p,1), and fails, even in weak L^p, for \gamma\in(0,\gamma_p).

At the critical exponent. For p\in[1,2) the reverse inequality holds in weak L^p, uniformly in the regularisation,

\displaystyle \sup_{t>0}\|(-\Delta)^{\gamma_p}P_tf\|_{L^{p,\infty}(m)}\le C_p\,\|\partial f\|_{L^p(\nu)},

so that (-\Delta)^{\gamma_p}P_tf has a weak-L^p limit as t\downarrow0 for every f\in W^{1,p}(X), but the strong L^p inequality fails. For p>2 the inequality fails even in weak L^p.

The Riesz transform. \mathcal R_1 is not of weak type (1,1). For 1<p<2, \mathcal R_p is bounded from L^p(m) to L^{p,\infty}(\nu) but not from L^{p,1}(m) to L^p(\nu). For p=2 it is of course an isometry. For p>2 it is bounded from the Lorentz space L^{p,1}(m) to L^p(\nu) but not from L^p(m) to L^p(\nu). In short, \mathcal R_p is bounded L^p(m)\to L^p(\nu) only for p=2.

References

[AC05] P. Auscher, T. Coulhon, Riesz transform on manifolds and Poincaré inequalities, Ann. Sc. Norm. Super. Pisa Cl. Sci. (5) 4 (2005), 531–555.

[ACDH04] P. Auscher, T. Coulhon, X. T. Duong, S. Hofmann, Riesz transform on manifolds and heat kernel regularity, Ann. Sci. École Norm. Sup. (4) 37 (2004), 911–957.

[BC23] F. Baudoin, L. Chen, Sobolev spaces and Poincaré inequalities on the Vicsek fractal, Ann. Fenn. Math. 48 (2023), 3–26.

[BC24] F. Baudoin, L. Chen, Heat kernel gradient estimates for the Vicsek set, Math. Nachr. 297 (2024), 4450–4477.

[CCFR17] L. Chen, T. Coulhon, J. Feneuil, E. Russ, Riesz transform for 1\le p\le2 without Gaussian heat kernel bound, J. Geom. Anal. 27 (2017), 1489–1514.

[CD99] T. Coulhon, X. T. Duong, Riesz transforms for 1\le p\le2, Trans. Amer. Math. Soc. 351 (1999), 1151–1169.

[CD03] T. Coulhon, X. T. Duong, Riesz transform and related inequalities on noncompact Riemannian manifolds, Comm. Pure Appl. Math. 56 (2003), 1728–1751.

[Che15] L. Chen, Sub-Gaussian heat kernel estimates and quasi Riesz transforms for 1\le p\le2, Publ. Mat. 59 (2015), 313–338.

[CJLL26] R. Chen, R. Jiang, B. Li, H.-Q. Li, The Coulhon–Duong conjecture for the Riesz transform on complete Riemannian manifolds, arXiv:2609.23503 (2026).

[DR22] B. Devyver, E. Russ, Reverse inequality for the Riesz transforms on Riemannian manifolds, arXiv:2209.05083 (2022).

[DR26] B. Devyver, E. Russ, Reverse inequality for quasi-Riesz transforms on cable systems, J. Geom. Anal. 36 (2026), Paper No. 2.

[DRY23] B. Devyver, E. Russ, M. Yang, Gradient estimate for the heat kernel on some fractal-like cable systems and quasi-Riesz transforms, Int. Math. Res. Not. IMRN (2023), 15537–15583.

[Fen26a] J. Feneuil, In spaces with a slow diffusion, the Riesz transform is unbounded on L^p, p\in(2,\infty), J. Geom. Anal. 36 (2026), Paper No. 61.

[Fen26b] J. Feneuil, Reverse inequalities for super-Riesz transforms on graphs with a slow diffusion, arXiv:2606.05475 (2026).

Posted in Global analysis in Dirichlet spaces | Tagged , | 1 Comment

Assistant Professor in Probability Theory & Stochastic Analysis


📍 Aarhus University – Department of Mathematics (Ny Munkegade 118, 8000 Aarhus C)
🗓 Position start: 1 August 2026 (3-year full-time fixed-term: 1 Aug 2026–31 Jul 2029)

Join the vibrant Stochastics Group at the Department of Mathematics, Aarhus University, at the faculty of Natural Sciences. Our research spans probability theory, stochastic analysis, stochastic differential geometry, Malliavin calculus, stochastic PDEs, random matrices, and related areas of analysis and geometry. The successful candidate will be affiliated with the “Stochastic Analysis in Aarhus” project, funded by the Villum Foundation and led by myself.

What you’ll bring:
A completed PhD in mathematics, with a strong research record in probability theory, stochastic analysis or closely related fields.
Demonstrated ability to conduct independent research and to collaborate as part of a research group.
Experience in teaching at university level and willingness to contribute to research-based teaching, including examination duties.

What we offer:
A stimulating research environment with an international profile and strong collaborations in stochastic analysis and related fields.
Excellent research infrastructure, including access to travel funds and visitor support.
A collegial, dynamic department with active seminar series, workshops and a broad range of research groups.
Support for international researchers and accompanying families, including relocation services and career counselling for expat partners.

Application deadline: 15 December 2025 – 23:59 CET
—
If you’re passionate about deepening your impact in stochastic analysis in a supportive, internationally engaged academic setting, we would love to hear from you.

💡 Feel free to share this opportunity with suitable colleagues or post-docs.

https://math.au.dk/om/ledige-stillinger/job/assistant-professor-in-probability-theory-and-stochastic-analysis

Posted in Uncategorized | Leave a comment

Aarhus Summer School in Analysis and Probability

The 6–10 July, 2026, Aarhus Summer School in Analysis and Probability aims to bring together leading experts and young researchers in the areas of analysis, probability theory, and their interactions. The school will feature a series of five hours lecture courses delivered by distinguished speakers, offering participants the opportunity to engage with current developments at the forefront of the field. In addition to the main lectures, the program will include short talks by graduate students and early-career researchers. The summer school will  provide a stimulating and friendly environment, ideal for fostering discussions, collaborations, and the exchange of ideas among researchers at all career stages.

Funding for Early Career Researchers

We are pleased to announce that limited funding will be available for early career researchers (PhD students and postdocs) who wish to participate in the summer school and propose to give a talk. Applicants interested in being considered for funding are invited to include a short abstract of their proposed presentation when registering or submitting their application to the organisers.

Registration will begin on December 1st.

https://conferences.au.dk/saa-summer-school-2026

The courses are the following.

Energy measures and applications by Mathav Murugan

I plan to cover some results on energy measures such as singularity/absolute continuity, its role in the attainment problem for conformal walk dimension (and Ahlfors regular conformal dimension), and also applications to martingale dimension.

The Sard conjecture in sub-Riemannian geometry by Luca Rizzi 

Sard’s theorem asserts that the set of critical values of a smooth map between finite-dimensional manifolds has measure zero. It is well-known, however, that when the domain is infinite dimensional and the range is finite dimensional, the result is not true. Counterexamples are known already in the class of polynomial maps from an infinite-dimensional Hilbert space to the Euclidean line. The problem is in particular relevant in sub-Riemannian geometry, where the validity of the above property for the endpoint map is the so-called Sard conjecture, and is one of the main open problems in the field. We present an overview of the Sard problem in sub-Riemannian geometry together with some recent results in collaboration with Lerario and Tiberio.

Gromov hyperbolicity, uniformization, and potential theory by Nageswari Shanmugalingam

In the non smooth setting of metric measure spaces, where related notions of interest are invariant under bi-Lipschitz transformations of a metric space, negative curvature has a counterpart as Gromov hyperbolicity. This is a coarser, more large-scale notion, but appears to be closely associated with bounded uniform domains via a transformation called uniformization. In this series of lectures we will discuss how such a transformation can link potential theory on Gromov hyperbolic spaces with potential theory on bounded uniform domains. We will begin the course by discussing analogs of Sobolev spaces in non-smooth setting.

Some aspects of parabolic Anderson models by Samy Tindel 

 In this mini-course I will introduce some basic elements allowing to understand parabolic Anderson models on geometric structures. The topics covered will be:

(1) Gaussian noises and their regularity.

(2) Existence and uniqueness results for the stochastic heat equation.

(3) Moment estimates.

(4) Extensions from Euclidian spaces to hypoelliptic settings, fractals and other geometric contexts.

Notice that I’m not assuming any knowledge of stochastic analysis from the audience. I will try to introduce the objects I’m manipulating in a self-contained way.

Posted in Uncategorized | Tagged , , , , , | Leave a comment

Postdoctoral Positions in Probability Theory and Analysis

Several postdoctoral positions at Aarhus are opening in my group. The earliest possible start date is February 1st 2026, with flexibility for later starts in the spring.

These positions are supported by a Villum Investigator Grant and the ERC, offering excellent opportunities for travels, research and international collaboration.

https://international.au.dk/about/profile/vacant-positions/job/postdoctoral-positions-in-probability-theory-and-analysis-aarhus-university

Posted in Uncategorized | Leave a comment

Lecture notes: A short introduction to Dirichlet spaces

Here are the lecture notes of the minicourse given at NYU Abu Dhabi.

Posted in Dirichlet forms at NYU, Lecture Notes | Tagged , , , , , | Leave a comment

Lecture 7. Further topics

In this lecture let X be a locally compact and complete metric space equipped with a Radon measure μ supported on X. Let (ℰ, ℱ = dom(ℰ)) be a Dirichlet form on X. We assume throughout that the heat semigroup Pt is stochastically complete, i.e., Pt1 = 1 for every t ≥ 0.

Contents

Regular Dirichlet forms, Energy measures

We denote by Cc(X) the vector space of all continuous functions with compact support in X and C0(X) its closure with respect to the supremum norm.

A core for (X,μ,ℰ,ℱ) is a subset 𝒞 of Cc(X) ∩ ℱ which is dense in Cc(X) in the supremum norm and dense in ℱ in the norm

(||f||^2_{L^2(X,\mu)} + \mathcal{E}(f,f))^{1/2}.

Definition. The Dirichlet form ℰ is called regular if it admits a core.

Recall that for any f,g ∈ ℱ, we have

\mathcal{E}(f,g)=\lim_{t\to 0}\frac1t\langle (I-P_t)f,g\rangle=\lim_{t\to 0}\frac1{t}\int_X \int_X(f(x)-f(y)) g(x) p_t(x,dy) d\mu(x),

where pt(x,·) are the heat kernel measures associated to the Dirichlet form (ℰ, ℱ).
From the symmetry property \eqref{heat kernel measure symmetry} of the heat kernel measure one also has

\mathcal{E}(f, g)=\lim_{t\to 0}\frac1{2t}\int_X \int_X(f(x)-f(y))(g(x)-g(y)) p_t(x,dy) d\mu(x).

Lemma. For f,g ∈ ℱ ∩ L∞(X,μ), fg ∈ ℱ and

\mathcal{E}(fg)^{1/2} \le ||f||_\infty \mathcal{E}(g)^{1/2} +||g||_\infty \mathcal{E}(f)^{1/2}.

Proof. For f, g ∈ ℱ ∩ L∞(X,μ) such that fg ∈ ℱ,

\mathcal{E}(fg)=\lim_{t\to 0}\frac1{2t}\int_X \int_X(f(x)g(x)-f(y)g(y))^2 p_t(x,dy) d\mu(x).

Write f(x)g(x) – f(y)g(y) = f(x)(g(x) – g(y)) + g(y)(f(x) – f(y)), then by Minkowski’s inequality

\left(\int_X \int_X(f(x)g(x)-f(y)g(y))^2 p_t(x,dy) d\mu(x)\right)^{1/2}

\le ||f||_{\infty}\left(\int_X \int_X (g(x)-g(y))^2 p_t(x,dy) d\mu(x)\right)^{1/2}

+||g||_{\infty}\left(\int_X \int_X(f(x)-f(y))^2 p_t(x,dy) d\mu(x)\right)^{1/2}.

We conclude the expected inequality by multiplying by 1/√2t and taking the limit t → 0 for both sides above.

Theorem (Energy measures) Assume that ℰ is regular. For f ∈ ℱ ∩ L∞(X,μ), there exists a unique Radon measure on X, denoted by dΓ(f), so that for every ϕ ∈ ℱ ∩ Cc(X),

\int_X\phi\, d\Gamma(f)=\frac{1}{2}[2\mathcal{E}(\phi f,f)-\mathcal{E}(\phi, f^2)]

= \lim_{t \to 0} \frac{1}{2t} \int_X \int_X \phi(x) (f(x)-f(y))^2 p_t(x,dy) d\mu(x).

The Radon measure dΓ(f) is called the energy measure of f (and is therefore the weak * limit of (1/(2t)) ∫X (f(x) – f(y))2 pt(x,dy)).

Proof. Let f ∈ ℱ ∩ L∞(X,μ). For any ϕ ∈ ℱ ∩ Cc(X),

\frac1{2t} \int_X \int_X \phi(x) (f(x)-f(y))^2 p_t(x,dy) d\mu(x)=-\frac1{2t} \langle (I-P_t)f^2 ,\phi\rangle+\frac1{t} \langle (I-P_t)f ,f\phi\rangle.

Letting t → 0, the right-hand side converges to

\frac{1}{2}[2\mathcal{E}(\phi f,f)-\mathcal{E}(\phi, f^2)] .

On the other hand, observing that

\frac1{2t} \int_X \int_X |\phi(x)| (f(x)-f(y))^2 p_t(x,dy) d\mu(x) \le \|\phi\|_{\infty} \frac1{2t} \int_X \int_X (f(x)-f(y))^2 p_t(x,dy) d\mu(x),

we deduce

\left|\frac{1}{2}[2\mathcal{E}(\phi f,f)-\mathcal{E}(\phi, f^2)]\right|\le \|\phi\|_{\infty}\mathcal E(f).

Therefore we conclude the proof by applying the Riesz-Markov representation theorem.

One can actually define dΓ(f,f) for every f ∈ ℱ using the following lemmas.

Lemma. Let f ∈ ℱ. Then fn = min(n, max(-n, f)) ∈ ℱ and ℰ(f – fn) → 0.

Proof. Let f ∈ ℱ. For every x, y ∈ X, we have

|f_n(x) - f_n(y)| \leq |f(x) - f(y)|

and |fn(x)| ≤ |f(x)|. So fn is a normal contraction of f and ℰ(fn) ≤ ℰ(f).

Recall that

\mathcal{E}(f-f_n)=\lim_{t\to 0}\frac1{2t} \int_X \int_X(f(x)-f_n(x)-f(y)+f_n(y))^2 p_t(x,dy) d\mu(x).

Expanding the square inside the integral gives that

\mathcal{E}(f-f_n)=\mathcal E(f)+\mathcal E(f_n)-2\mathcal E(f_n,f)\le 2\mathcal E(f)-2\mathcal E(f_n,f).

Letting n → ∞, we have ℰ(fn,f) → ℰ(f). Therefore ℰ(f – fn) converges to 0 as n → ∞.

Lemma. For f,g ∈ ℱ ∩ L∞(X,μ) and nonnegative ϕ ∈ ℱ ∩ Cc(X),

\left| \sqrt{\int_X\phi\, d\Gamma(f)} -\sqrt{\int_X\phi\, d\Gamma(g) } \right|^2 \le \int_X\phi\, d\Gamma(f-g) \le ||\phi||_{L^\infty(X,\mu)}\mathcal{E}(f-g)

Proof. The second inequality follows from the proof in Theorem 16.3. For the first inequality, it suffice to show that for any f, g ∈ ℱ ∩ L∞(X,μ) and any nonnegative ϕ ∈ ℱ ∩ Cc(X),

\sqrt{\int_X\phi\, d\Gamma(f)} \le \sqrt{\int_X\phi\, d\Gamma(g) } +\sqrt{ \int_X\phi\, d\Gamma(f-g) }.

Indeed, this inequality follows from a similar proof as in Lemma 16.2 by noting that f = f – g + g and using Minkowski’s inequality.

Thanks to the previous lemmas, by approximation, one can define dΓ(f) for every f ∈ ℱ by

dΓ(f) = sup{dΓ(fn) : fn = min(n, max(-n, f)), n = 1, 2, …}.

For f,g ∈ ℱ, one can define dΓ(f,g) by polarization

dΓ(f,g) = (1/4)(dΓ(f + g) – dΓ(f – g)).

The following representation theorem for regular Dirichlet forms then holds.

Theorem  (Beurling-Deny) Assume that ℰ is regular. For u,v ∈ ℱ, ℰ(u,v) = ∫X dΓ(u,v).

Hunt process associated with a regular Dirichlet form

Definition. A Hunt process with state space X is a family of stochastic process (Xt)t≥0 and probability measures (ℙx)x∈X defined on a measure space (Ω, ℱ), such that (Xt)t≥0 is adapted w.r.t. the right-continuous minimal completed admissible filtration (ℱt)t≥0, X0 = x, ℙx-a.s. and the following hold:

  • (i) x → ℙx(Xt ∈ B) is measurable for all t > 0 and B ∈ ℬ(X),
  • (ii) X is a strong Markov process, i.e. for every stopping time T, XT is ℱT-measurable and for every B ∈ ℬ(X)

    \mathbb{P}_x(X_{T+t} \in B | \mathcal{F}_T) = \mathbb{P}_{X_T}(X_t \in B) \quad \mathbb{P}_x\text{-a.s. on } \{T < \infty\},

  • (iii) X is right-continuous, i.e.

    \lim_{s \downarrow t} X_s = X_t, \quad \forall t \quad \mathbb{P}_x\text{-a.s.}

  • (iv) X is quasi left-continuous, i.e. for all stopping times T and (Tn)n such that Tn ↑ T a.s.

    \lim_{n \to \infty} X_{T_n} = X_T, \quad \mathbb{P}_x\text{-a.s. on } \{T < \infty\}.

Remark 

  1. Note that quasi left-continuity does not necessarily imply left-continuity, because the set

    A = \left\{ \lim_{n \to \infty} X_{s_n} = X_t \right\}

    might depend on the choice of sequence (sn)n, sn ↑ t.

  2. In more generality, one might consider situations where (ℙx(Xt ∈ ·))x,t are sub-probability measures (i.e. all the axioms of probability measures are satisfied but ℙx(Xt ∈ Ω) ≤ 1). In that case we can perform a one-point compactification of X by introducing a cemetery state ∂ ∉ X and redefine ℙx to be a probability measure on X ∪ {∂}.

The following theorem can then be proved, see the book [FOT], Theorem 4.2.1.

Theorem (Fukushima) Assume that ℰ is regular, then there exists a Hunt process ((Xt)t≥0, (ℙx)x∈X) such that for μ-a.e. x ∈ X, A ∈ ℬ(X) and t ≥ 0,

\mathbb{P}_x (X_t \in A)=p_t(x,A)

where pt(x, ·) are the heat kernel measures associated to the Dirichlet form (ℰ, 𝒟(ℰ)).

Intrinsic metric

Definition. The Dirichlet form ℰ is called strongly local if for any two functions f,g ∈ ℱ with compact supports such that f is constant in a neighborhood of the support of g, we have ℰ(f,g) = 0.

With respect to ℰ we can define the following intrinsic metric dℰ on X by

d_{\mathcal{E}}(x,y)=\sup\{u(x)-u(y)\, :\, u\in\mathcal{F}\cap C_0(X)\text{ and } d\Gamma(u,u)\le d\mu\}.

Here the condition dΓ(u,u) ≤ dμ means that Γ(u,u) is absolutely continuous with respect to μ with Radon-Nikodym derivative bounded by 1.

The term “intrinsic metric” is potentially misleading because in general there is no reason why dℰ is a metric on X (it could be infinite for a given pair of points x,y or zero for some distinct pair of points).

Definition. A strongly local regular Dirichlet space is called strictly local if dℰ is a metric on X and the topology induced by dℰ coincides with the topology on X.

Example (Riemannian manifolds) Let (M,g) be a complete n-dimensional Riemannian manifold with Riemannian volume measure μ. We consider the standard Dirichlet form ℰ on M, which is obtained by closing the bilinear form

\mathcal{E}(f,g)=\int_\mathbb{M} \langle \nabla f ,\nabla g \rangle d\mu, \quad f,g \in C_0^\infty(\mathbb M).

Then ℰ is a strictly local Dirichlet form such that

d_\mathcal{E} (x,y)=d_g(x,y).

Example (Carnot groups) Let G be a Carnot group with sub-Laplacian

L=\sum_{i=1}^d V_i^2

and Dirichlet form

\mathcal{E}(f)=\int_{\mathbb G} \sum_{i=1} (V_if)^2 d\mu.

Then ℰ is a strictly local Dirichlet form such that

d_\mathcal{E} (x,y)=d_{CC} (x,y)

where dCC is the so-called Carnot-Carathéodory distance which is defined as follows.

An absolutely continuous curve γ : [0,T] → G is said to be subunit for the operator L if for every smooth function f : G → ℝ we have

\left| \frac{d}{dt} f ( \gamma(t) ) \right| \le \sqrt{ (\Gamma f) (\gamma(t)) }.

We then define the subunit length of γ as ℓs(γ) = T.

Given x,y ∈ G, we indicate with

S(x,y) = {γ : [0,T] → G | γ is subunit for Γ, γ(0) = x, γ(T) = y}.

It is a consequence of the Chow-Rashevskii theorem that

S(x,y) ≠ ∅, for every x, y ∈ G.

One defines then

d_{CC}(x,y) = \inf\{ \ell_s(\gamma) \mid  \gamma \in  S(x,y)\}

Example Consider on the Sierpinski gasket the standard Dirichlet form ℰ. Then ℰ is regular, but unless f is constant, for f ∈ ℱ, dΓ(f) is singular with respect to the Hausdorff measure μ, see [KM19]. As a consequence ℰ is not strictly local.

 Further reading

Far from being exhaustive we mention the following references for further reading:

  1. The book [FOT] is a standard comprehensive reference in the theory of Dirichlet forms and associated Hunt processes, see also [ChenFukushima].
  2. The book [GSC] shows how one can restrict Dirichlet forms to domains (abstract Dirichlet and Neumann boundary conditions).
  3. Parabolic regularity theory for the heat equation can be developed in the setting of abstract strictly local Dirichlet spaces, see [Saloff].
Posted in Dirichlet forms at NYU | Tagged , , | Leave a comment

Lecture 6. Gagliardo-Nirenberg inequalities in Dirichlet spaces

Let (X,μ,ℰ,ℱ) be a Dirichlet space and {Pt}t∈[0,∞) denote the associated Markovian semigroup. Throughout the lecture, we shall assume that P_t admits a measurable heat kernel pt(x,y) satisfying, for some C > 0 and β > 0,

p_{t}(x,y)\leq C t^{-\beta}

for μ × μ-a.e. (x,y) ∈ X × X, and for each t ∈ (0,+∞). We also assume stochastic completeness, which means that Pt1 = 1 for every t ≥ 0.

The goal of the lecture is to prove the Gagliardo-Nirenberg inequalities in that setting. The techniques we use come from a paper by Bakry-Coulhon-Ledoux-Saloff-Coste. For another approach to the Sobolev inequality, we refer to the post.

Contents

Preliminary lemmas

Lemma. For every f ∈ ℱ, t ≥ 0,

|| P_t f -f ||_{L^2(X,\mu)} \le C \sqrt{t} \mathcal{E}(f,f)^{1/2}.

Proof: Let Δ be the generator of the semigroup (Pt)t≥0. We use spectral theorem to represent Δ as a multiplier in some L2(Ω,ν) space:

U^{-1} A U g (x)=-\lambda(x) g(x),

so that

|| P_t f -f ||_{L^2(X,\mu)}^2=\int_{\Omega} \left( 1-e^{-t\lambda(x)}\right)^2(U^{-1} f)^2(x) d\nu(x)

\le C t \int_{\Omega} \lambda(x) (U^{-1} f)^2(x) d\nu(x)

= Ct \mathcal{E}(f,f).

Lemma. For every f ∈ ℱ,

\lim_{t \to 0} t^{-1}\int_X \int_X |f (x)-f(y)|^2 p_t(x,y) d\mu(x)d\mu(y) =2 \mathcal{E}(f,f).

Proof: Using Pt1 = 1 we have

\int_X \int_X |f (x)-f(y)|^2 p_t(x,y) d\mu(x)d\mu(y)

\int_X \int_X( f (x)^2-2f(x)f(y)+f(y)^2) p_t(x,y) d\mu(x)d\mu(y)

=2\int_X f(x)^2 d\mu(x)-2\int_X f(x)(P_t f) (x)d\mu(x)

=2 \int_X (f(x)-P_tf(x)) f(x) d\mu(x)

and we conclude thanks to a previous result.

Sobolev inequality

Lemma. Let 1 ≤ q < +∞. There exists a constant C > 0 such that for every f ∈ ℱ ∩ Lq(X,μ) and s ≥ 0,

\sup_{s \ge 0} s^{1 +\frac{q}{2\beta}} \mu \left( \{ x \in X\, :\, | f(x) | > s \} \right)^{\frac{1}{2}}\le C \mathcal{E}(f,f)^{1/2} || f ||_{L^q(X,\mu)}^{\frac{q}{2\beta}}.

Proof: Let f ∈ ℱ and denote

F(s)=\mu \left( \{ x \in X\, :\, | f(x) | > s \} \right).

We have then

F(s) \le \mu \left( \{ x \in X\, :\, | f(x) -P_t f (x) | > s/2 \} \right)+\mu \left( \{ x \in X\, :\, | P_t f (x) | > s/2 \} \right).

Now, from the heat kernel upper bound pt(x,y) ≤ C t-β, t > 0, one deduces, for g ∈ L1(X,μ), that

|P_t g (x) | \le Ct^{-\beta} \| g \|_{L^1(X,\mu)}.

Since Pt is a contraction in L∞(X,μ), by the Riesz-Thorin interpolation one obtains

| P_t f (x) | \le \frac{C^{1/q}}{t^{\beta /q}} \| f \|_{L^q(X,\mu)}.

Therefore, for

s= 2 \frac{C^{1/q}}{t^{\frac{\beta}{ q}}} \| f \|_{L^q(X,\mu)},

one has

\mu \left( \{ x \in X\, :\, | P_t f (x) | > s/2 \} \right)=0.

On the other hand, from previous  lemma,

\mu \left( \{ x \in X\, :\, | f(x) -P_t f (x) | > s/2 \} \right) \le C s^{-2} t \mathcal{E}(f,f).

We conclude that

F(s)^{1/2} \le C s^{-1 -\frac{q}{2\beta}} \mathcal{E}(f,f)^{1/2} \| f \|_{L^q(X,\mu)}^{\frac{q}{2\beta}}.

Lemma. Assume β > 1. There exists a constant C > 0 such that for every f ∈ ℱ,

\sup_{s \ge 0}\, s\, \mu \left( \{ x \in X\, :\, | f(x) | \ge s \} \right)^{\frac{1}{q}} \le C \mathcal{E}(f,f)^{1/2},

where q = 2β/(β – 1).

Proof: Let f ∈ ℱ be a non-negative function. For k ∈ ℤ, we denote

f_k=(f-2^k)_+ \wedge 2^k.

Observe that fk ∈ L2(X,μ) and ||fk||L2(X,μ) ≤ ||f||L2(X,μ). Moreover, for every x, y ∈ X, |fk(x) – fk(y)| ≤ |f(x) – f(y)| and so ℰ(fk,fk) ≤ ℰ(f,f). We also note that fk ∈ L1(X,μ), with

||f_k||_{L^1(X,\mu)} =\int_X |f_k| d\mu \le 2^k \mu (\{ x \in X\, :\, f(x) \ge 2^k \}).

We now use the previous lemma to deduce:

\sup_{s \ge 0} s^{1 +\frac{1}{2\beta}} \mu \left( \{ x \in X\, :\, f_k(x) > s \} \right)^{\frac{1}{2}} \le C \mathcal{E}(f_k,f_k)^{1/2} || f_k \|_{L^1(X,\mu)}^{\frac{1}{2\beta}}

\le C \mathcal{E}(f_k,f_k)^{1/2} \left( 2^k \mu (\{ x \in X\, :\, f(x) \ge 2^k \})\right)^{\frac{1}{2\beta}}

In particular, by choosing s = 2k we obtain

2^{k \left(1 +\frac{1}{2\beta} \right)} \mu \left( \{ x \in X\, :\, f(x) \ge 2^{k+1} \} \right)^{\frac{1}{2}} \le C \mathcal{E}(f_k,f_k)^{1/2} \left( 2^k \mu (\{ x \in X\, :\, f(x) \ge 2^k \})\right)^{\frac{1}{2\beta}}

Let

M(f)=\sup_{k \in \mathbb{Z}} 2^k \mu (\{ x \in X\, :\, f(x) \ge 2^k \})^{1/q}

where q = 2β/(β – 1). Using the fact that 1/q = 1/2 – 1/(2β) and the previous inequality we obtain:

2^{k} \mu \left( \{ x \in X\, :\, f(x) \ge 2^{k+1} \} \right)^{\frac{1}{2}} \le C 2^{ -\frac{kq}{2\beta}}\mathcal{E}(f,f)^{1/2} M(f)^{\frac{q}{2\beta}}

and

2^k \mu \left( \{ x \in X\, :\, f(x) \ge 2^{k+1} \} \right)^{\frac{1}{q}} \le C^{\frac{2}{q} } \mathcal{E}(f,f)^{1/q}M(f)^{\frac{1}{\beta}}.

Therefore

M(f)^{1-\frac{1}{\beta}} \le 2 C^{\frac{2}{q} } \mathcal{E}(f,f)^{1/q}

and one concludes

M(f) \le 2^{q/2} C \mathcal{E}(f,f)^{1/2}.

This easily yields:

\sup_{s \ge 0} s \mu \left( \{ x \in X\, :\, f(x) \ge s \} \right)^{\frac{1}{q}} \le 2^{1+q/2} C \mathcal{E}(f,f)^{1/2}

Let now f ∈ ℱ, which is not necessarily non-negative. From the previous inequality applied to |f| we deduce

\sup_{s \ge 0}\, s\, \mu \left( \{ x \in X\, :\, |f(x) | \ge s \} \right)^{\frac{1}{q}}

\le 2^{1+q/2} C \mathcal{E}(|f|,|f|)^{1/2}

\le 2^{1+q/2} C \mathcal{E}(f,f)^{1/2}.

Theorem. (Sobolev inequality) Assume β > 1. There exists a constant C > 0 such that for every f ∈ ℱ,

\| f \|_{L^q(X,\mu)} \le C \mathcal{E}(f,f)^{1/2}

where q = 2β/(β – 1).

To show that the weak type inequality implies the desired Sobolev inequality, we will need another slicing argument and the following lemma is needed.

Lemma. For f ∈ ℱ, f ≥ 0, denote fk = (f – 2k)+ ∧ 2k, k ∈ ℤ. There exists a constant C > 0 such that for every f ∈ ℱ,

\sum_{k \in \mathbb{Z}} \mathcal{E} (f_k,f_k) \le C \mathcal{E} (f,f).

Proof: Let pt(x,y) denote the heat kernel of the semigroup Pt. We first observe that, once we prove

\sum_{k \in \mathbb{Z}} \int_X \int_X |f_{k} (x)-f_{k}(y)|^2 p_t(x,y) d\mu(x)d\mu(y) \le C\int_X \int_X |f (x)-f(y)|^2 p_t(x,y) d\mu(x)d\mu(y)

where C > 0 is independent from t, then

\liminf_{t \to 0^+} \sum_{k \in \mathbb{Z}} t^{-1 } \!\!\!\int_X \int_X |f_{\rho} (x)-f_{\rho}(y)|^2 p_t(x,y)d\mu(x)d\mu(y)

\le C\liminf_{t \to 0^+} t^{-1 } \!\!\!\int_X \int_X |f (x)-f(y)|^p p_t(x,y) d\mu(x)d\mu(y),

and, using the superadditivity of the liminf, one concludes

\sum_{k \in \mathbb{Z}} \liminf_{t \to 0^+} t^{-1 } \int_X \int_X |f_{k} (x)-f_{k}(y)|^2 p_t(x,y) d\mu(x)d\mu(y)

\le C \liminf_{t \to 0^+} t^{-1 } \int_X \int_X |f (x)-f(y)|^2 p_t(x,y) d\mu(x)d\mu(y)

which yields

\sum_{k \in \mathbb{Z}} \mathcal{E} (f_k,f_k) \le C \mathcal{E} (f,f).

We therefore aim to prove the inequality above. For each k ∈ ℤ, set Bk = {x ∈ X : 2k < f ≤ 2k+1}. In this way, the external integral on the left-hand side is decomposed it into an integral over Bk and Bkc. For the integrals over Bk, since the mapping f → fk is a contraction, it follows that

\sum_{k\in\mathbb{Z}}\int_{B_k} \int_X |f_{k} (x)-f_{k}(y)|^2 p_t(x,y)d\mu(x)d\mu(y)\leq \int_X \int_X |f (x)-f(y)|^2 p_t(x,y)d\mu(x)d\mu(y).

To perform the integrals over Bkc, we decompose them as

\sum_{k\in\mathbb{Z}}\int_{B_k^c} \int_{B_k} |f_{k} (x)-f_{k}(y)|^2 p_t(x,y) d\mu(x)d\mu(y) +\sum_{k\in\mathbb{Z}}\int_{B^c_k} \int_{B_k^c}|f_{k} (x)-f_{k}(y)|^2 p_t(x,y)d\mu(x)d\mu(y)

=:\sum_{k\in\mathbb{Z}}J_1(k)+\sum_{k\in\mathbb{Z}}J_2(k).

Again, the contraction property of f → fk yields

\sum_{k\in\mathbb{Z}}J_1(k)\leq \sum_{k\in\mathbb{Z}} \int_X \int_{B_k}|f_{k} (x)-f_{k}(y)|^2 p_t(x,y)d\mu(x)d\mu(y)

\leq \int_X \sum_{k\in\mathbb{Z}} \int_{B_k}|f_{k} (x)-f_{k}(y)|^2 p_t(x,y)d\mu(x)d\mu(y)\leq \int_X \int_X |f (x)-f(y)|^2 p_t(x,y) d\mu(x)d\mu(y).

On the other hand, notice that for any (x,y) ∈ Bkc × Bkc we have |fk(x) – fk(y)| ≠ 0 only if

(x,y)\in \{f(x)\leq 2^k<f(y)/2 \}\cup\{f(y)\leq 2^k<f(x)/2\}=:Z_k\cup Z_k^*.

Also, |fk(x) – fk(y)| = 2k for (x,y) ∈ Zk ∪ Zk*. Thus,

\sum_{k\in\mathbb{Z}}J_2(k)\leq \sum_{k\in \mathbb{Z}}\int_X\int_X\big(\mathbf{1}_{Z_k}(x,y)+\mathbf{1}_{Z^*_k}(x,y)\big)|f_k(x)-f_k(y)|^2 p_t(x,y)d\mu(x)d\mu(y)

=\int_X\int_X\sum_{k\in\mathbb{Z}}\big(\mathbf{1}_{Z_k}(x,y)+\mathbf{1}_{Z^*_k}(x,y)\big)2^{2k} p_t(x,y) d\mu(x)d\mu(y).

One can see that

\sum_{k\in\mathbb{Z}}\mathbf{1}_{Z_k}(x,y) 2^{2k}\leq 2 |f(x)-f(y)|^2

and the same holds for Zk*, hence

\sum\limits_{k\in\mathbb{Z}}J_1(k)+\sum\limits_{k\in\mathbb{Z}}J_2(k)\leq 5 \int_X\int_X|f(x)-f(y)|^2 p_t(x,y) d\mu(x)d\mu(y).

Adding to these the term above finally yields the result

We can now conclude the proof of the Sobolev inequality.

Proof of the Sobolev inequality: Let f ∈ ℱ. We can assume f ≥ 0. As before, denote fk = (f – 2k)+ ∧ 2k, k ∈ ℤ. From Lemma 13.2 applied to fk, we see that

\sup_{s \ge 0} s \mu \left( \{ x \in X\, :\, | f_k(x) | \ge s \} \right)^{\frac{1}{q}} \le C \mathcal{E}(f_k,f_k)^{1/2}.

In particular for s = 2k, we get

2^k \mu \left( \{ x \in X\, :\, f (x) \ge 2^{k+1} \} \right)^{\frac{1}{q}} \le C \mathcal{E}(f_k,f_k)^{1/2}

Therefore,

\sum_{k \in \mathbb{Z}} 2^{k q} \mu \left( \{ x \in X\, :\, f (x) \ge 2^{k+1} \} \right) \le C^q \sum_{k \in \mathbb{Z}} \mathcal{E}(f_k,f_k)^{q/2}.

Since q ≥ 2, one has

\sum_{k \in \mathbb{Z}} \mathcal{E}(f_k,f_k)^{q/2} \le \left( \sum_{k \in \mathbb{Z}} \mathcal{E}(f_k,f_k) \right)^{q/2}

Thus, from the previous lemma

\sum_{k \in \mathbb{Z}} 2^{k q} \mu \left( \{ x \in X\, :\, f (x) \ge 2^{k+1} \} \right)\le C \mathcal{E}(f,f)^{q/2}.

Finally, we observe that

\sum_{k \in \mathbb{Z}} 2^{k q} \mu \left( \{ x \in X\, :\, f (x) \ge 2^{k+1} \} \right) \ge \frac{q}{2^{q+1}-2^q} \sum_{k \in \mathbb{Z}}\int_{2^{k+1}}^{2^{k+2}} s^{q-1} \mu \left( \{ x \in X\, :\, f (x) \ge s \} \right)ds

\ge \frac{1}{2^{q+1}-2^q} \| f \|_{L^q(X,\mu)}^q.

The proof is thus complete.

Gagliardo-Nirenberg inequalities

Using similar methods (see the paper) in the general case β > 0 one can get  the family of Gagliardo-Nirenberg inequalities.

Theorem. Let q = 2β/(β – 1) with the convention that q = ∞ if β = 1. Let r,s ∈ (0,+∞] and θ ∈ (0,1] satisfying

\frac{1}{r}=\frac{\theta}{q}+\frac{1-\theta}{s}.

If β = 1, we assume r < +∞. Then, there exists a constant C > 0 such that for every f ∈ ℱ,

\| f \|_{L^r(X,\mu)} \le C \mathcal{E}(f,f)^{\theta/2}\|f\|_{L^s(X,\mu)}^{1-\theta}.

We explicitly point out some particular cases of interest.

  1. Assume that β > 1. If r = s, then r = 2β/(β – 1) and above recovers the Sobolev inequality

    \| f \|_{L^r(X,\mu)} \le C \mathcal{E}(f,f)^{1/2}.

  2. Assume that β > 1. If s = +∞ and r ≥ 2β/(β – 1), then above yields

    \| f \|_{L^r(X,\mu)} \le C \mathcal{E}(f,f)^{\theta/2} \| f \|^{1-\theta}_{L^\infty(X,\mu)}

    with θ = 2β/(r(β – 1)).

  3. If r = 2 and s = 1, then above yields the Nash inequality

    \| f \|_{L^2(X,\mu)} \le C \mathcal{E}(f,f)^{\theta/2} \| f \|^{1-\theta}_{L^1(X,\mu)}

    with θ = β/(1 + β).

In the case β = 1 one obtains the Trudinger-Moser inequalities.

Corollary . Assume that β = 1. Then, there exist constants c,C > 0 such that for every f ∈ ℱ with ℰ(f,f) = 1,

\int_X \left( \exp \left( c |f|^2 \right)-1 \right) d\mu \le C \| f \|^2_{L^2(X,\mu)}.

Posted in Dirichlet forms at NYU | Tagged , , | Leave a comment