Without explanation, we always assume that Omega subR^(n)\Omega\subset\mathbb{R}^{n} is a nonempty open set, where n⩾1n \geqslant 1. Given the compact set K subR^(n)K\subset\mathbb{R}^{n}, we use C_(K)^(oo)(Omega)C_ {K}^{\infty} (\Omega) to represent the set of smooth functions supported byK sub OmegaK\subset \Omega, i.e
Given a multiple indicator alpha=(alpha_(1),cdots,alpha_(n))\alpha = \left(\alpha_{1}, \cdots, \alpha_{n}\right), where each component is a nonnegative integer, and the symbol del^(alpha)varphi\partial^{\alpha}\varphi represents the following multiple partial derivatives
where del_(x_(i))^(alpha_(i))=(del^(alpha_(i)))/(delx_(i)^(alpha_(i)))\partial_ {x_{i}}^{\alpha_{i}}=\frac{\partial^{\alpha_{i}}}{\partial x_{i}^{\alpha_{i}}}. In addition, we order
We use D(Omega)=C_(0)^(oo)(Omega)\mathcal{D} (\Omega) = C_ {0}^{\infty} (\Omega) represents a set of smooth functions defined on Omega\Omega and with compact support. We call it {\bf test function space}. By definition, for any varphi inD(Omega)\varphi \in \mathcal{D} (\Omega), there is a compact set K sub OmegaK \subset\Omega(this is the compact set in R^(n)\mathbb{R}^{n}), so that f|_(Omega-K)-=0\left. f\right|_{\Omega -K} \equiv 0, that is, for any X in Omega-K,f(x)=0X \in \Omega -K, f (x) = 0.
Definition 1(Test function space): On the space D(Omega)\mathcal{D}(\Omega), we specify the following convergence (topology): given the function sequence {varphi_(p)}_(p⩾1)subD(Omega)\left\{\varphi_{p} \right\}_ {p \geqslant 1} \subset \mathcal{D} (\Omega), the so-called sequence converges to 0inD(Omega)0 \in \mathcal{D}(\Omega)("is recorded as"varphi_(p)longrightarrow^(D(Omega))0)(\mbox {is recorded as} \, \varphi_{p} \stackrel{\mathcal{D}(\Omega)} {\longrightarrow} 0), which refers to
There is a compact set of K sub OmegaK \subset \Omega, so that for each p⩾1p \geqslant 1, there is sup(varphi_(p))sub K\operatorname {sup} \left (\varphi_{p} \right) \subset K;
For each multiple indicator alpha\alpha, function sequence {del^(alpha)varphi_(p)}_(p⩾1)\left\{\partial^{\alpha} \varphi_ {p}\right\}_{p\geqslant 1} converge to 00 uniformly on KK, i.e.,lim_(p rarr oo)||del^(alpha)varphi_(p)||_(L^(oo)(K))=0.\lim_{p \rightarrow\infty}\left\|\partial^{\alpha} \varphi_ {p}\right\|_ {L^{\infty}(K)}=0.
Definition 2(distribution)
A distribution (also known as generalized function) on Omega\Omega refers to a linear functional (linear mapping) on D(Omega)\mathcal{D}(\Omega):
u:D(Omega)rarrC,quad varphi|->(:u,varphi:),u: \mathcal{D}(\Omega) \rightarrow \mathbb{C}, \quad \varphi \mapsto\langle u, \varphi\rangle,
meet the following two conditions
For any varphi,psi inD(Omega)\varphi, \psi \in \mathcal{D} (\Omega) and alpha,beta inC\alpha, \beta \in \mathbb{C}, we have(:u,alpha varphi+beta psi:)=alpha(:u,varphi:)+beta(:u,psi:)\langle u, \alpha \varphi+\beta \psi\rangle=\alpha\langle u, \varphi\rangle+\beta\langle u, \psi\rangle
For any compact set K sub OmegaK \subset \Omega, there are nonnegative integers pp and normal numbers CC (pp and CC depend on KK), so that for any varphi inC_(K)^(oo)(Omega)\varphi \in C_ {K}^{\infty} (\Omega), both|(:u,varphi:)|leqslant Cs u p_(|alpha|leqslant p)||del^(alpha)varphi||_(L^(oo)(K))|\langle u, \varphi\rangle| \leqslant C \sup _ {|\alpha| \leqslant p}\left\|\partial^{\alpha} \varphi\right\|_ {L^{\infty}(K)}
If the above selection of pp does not depend on the selection of compact set kk, then we call the smallest such nonnegative integer pp the order of distribution uu.
We use D^(')(Omega)\mathcal{D}^{\prime} (\Omega) to represent the whole distribution defined on Omega\Omega, and specify the following convergence (topology): the so-called distribution sequence {u_(p)}_(p⩾1)subD^(')(Omega)\left \{u_{p} \right \}_ {p \geqslant 1} \subset \mathcal{D}^{\prime} (\Omega) converges to 0inD^(')(Omega)0 \in \mathcal{D}^{\prime} (\Omega)(this is a linear mapping that maps all functions to 0), which is recorded as u_(p)longrightarrow^(D^(')(Omega))0u_{p} \stackrel {\mathcal{D}^{\prime} (\Omega)}{\longrightarrow}0, which means that we have any test function varphi inD(Omega)\varphi \in \mathcal{D} (\Omega)
We usually say that a distribution can be paired with a (compactly supported) smooth function to obtain a number, that is
D^(')(Omega)xxD(Omega),quad(u,varphi)|->(:u,varphi:)\mathcal{D}^{\prime}(\Omega) \times \mathcal{D}(\Omega), \quad(u, \varphi) \mapsto\langle u, \varphi\rangle
Remark: Any given distribution of u inD^(')(Omega)u \in \mathcal{D}^{\prime}(\Omega), which is actually a {\bf continuous} linear functional on D(Omega)\mathcal{D} (\Omega). We have not discussed the theory of topological linear space related to distribution in detail, so we do not intend to expand this point too much (which has no impact on the understanding of distribution theory). The so-called continuity can be described in the following sequence language: for any test function sequence varphi_(p)longrightarrow^(D)0\varphi_{p} \stackrel{\mathcal{D}}{\longrightarrow} 0, we all have
lim_(p rarr oo)(:u,varphi_(p):)=0\lim_{p\rightarrow \infty}\left\langle u, \varphi_ {p}\right\rangle=0
The proof of this proposition only needs to use the concept of distribution.
2. Examples of distribution
As an example, let's first learn some important distributions:
Example 1 (Dirac function)
For any a in Omegaa \in \Omega, we can define the distribution Delta_(a)inD^(')(Omega)\Delta_ {a} \in \mathcal{D}^{\prime}(\Omega). Where, for any varphi inD(Omega)\varphi \in \mathcal{D} (\Omega), we define
Let's verify Delta_(a)\Delta_ {a} is actually a distribution:
For any compact set K sub OmegaK \subset \Omega, if a!in Ka \notin K, then for any varphi inC_(K)^(oo)(Omega)\varphi\in C_{K}^{\infty}(\Omega), we all have
Therefore, we can take q=0q = 0 and C=1C = 1 in the definition of distribution.
In particular, we also know The order of delta_(a)\delta_{a} is 00.
Example 2 (locally integrable function)
Given the open set Omega\Omega (always equipped with Borel algebra and Lebesgue measure), the so-called locally integrable function refers to the function integrable on each compact local, that is, the function ff (the equivalent class of the corresponding almost everywhere equal function). For any compact set K sub OmegaK \subset \Omega, the function f*1_(K)inL^(1)(Omega)f \cdot \mathbf{1}_{K} \in L^{1}(\Omega). We use L_(loc)^(1)(Omega)L_\mathrm{loc}^{1}(\Omega) represents a locally integrable function on Omega\Omega.
For any f inL_(loc)^(1)(Omega)f \in L_\mathrm{loc}^{1} (\Omega), we define the linear functional on D(Omega)\mathcal{D} (\Omega):
Then we have, in distrubution sense, chi_(epsi)longrightarrow^(D^('))delta_(0)\chi_{\varepsilon} \stackrel{\mathcal{D}^{\prime}}{\longrightarrow} \delta_{0} as epsi rarr0\varepsilon \rightarrow 0.
A New Paradigm for Exploiting Pre-trained Model Hubs
A New Paradigm for Exploiting Pre-trained Model Hubs
It is often overlooked that the number of models in pre-trained model hubs is scaling up very fast. As a result, pre-trained model hubs are popular but under-exploited. Here a new paradigm is advocated to sufficiently exploit pre-trained model hubs.