{"id":1809,"date":"2025-05-22T18:53:57","date_gmt":"2025-05-22T18:53:57","guid":{"rendered":"https:\/\/www.ethanepperly.com\/?p=1809"},"modified":"2025-05-22T18:54:21","modified_gmt":"2025-05-22T18:54:21","slug":"markov-musings-5-poincare-inequalities","status":"publish","type":"post","link":"https:\/\/www.ethanepperly.com\/index.php\/2025\/05\/22\/markov-musings-5-poincare-inequalities\/","title":{"rendered":"Markov Musings 5: Poincar\u00e9 Inequalities"},"content":{"rendered":"\n

In the previous<\/a> posts<\/a>, we’ve been using eigenvalues<\/a> to understand the mixing of reversible Markov chains<\/a>. Our main convergence result was<\/a> as follows:

  <\/span>   <\/span>\"\[\chi^2\left(\rho^{(n)} \, \middle|\middle| \, \pi\right) \le \left( \max \{ \lambda_2, -\lambda_n \} \right)^{2n} \chi^2\left(\rho^{(0)} \, \middle|\middle| \, \pi\right).\]\"<\/p>Here, \"\rho^{(n)}\" denotes the distribution of the chain at time \"n\", \"\pi\" denotes the stationary distribution<\/a>, \"\chi^2(\cdot \mid\mid \cdot)\" denotes the \"\chi^2\" divergence<\/a>, and \"1 = \lambda_1 \ge \lambda_2 \ge \cdots \ge \lambda_m \ge -1\" denote the decreasingly ordered eigenvalues of the Markov transition matrix \"P\".<\/p>\n\n\n\n

Bounding the the rate of convergence requires an upper bound<\/em> on \"\lambda_2\" and a lower bound<\/em> on \"\lambda_m\". In this post, we will talk about techniques for bounding \"\lambda_2\". For more on the smallest eigenvalue \"\lambda_m\", see the previous post<\/a>.<\/p>\n\n\n\n

Setting<\/h2>\n\n\n\n

Let’s begin by establishing some notation, mostly the same as previous posts as this series. We work with a reversible Markov chain with transition matrix \"P\" and stationary distribution \"\pi\".<\/p>\n\n\n\n

As in previous posts, we identify vectors \"f \in \real^m\" and functions \"f : \{1,\ldots,m\} \to \real\", treating them as one and the same \"f_i = f(i)\".<\/p>\n\n\n\n

For a vector\/function \"f\", \"\expect_\pi[f]\" and \"\Var_\pi(f)\" denote the variance with respect to the stationary distribution \"\pi\":

  <\/span>   <\/span>\"\[\expect_\pi[f] = \sum_{i=1}^m f(i) \pi_i, \quad \Var_\pi(f) \coloneqq \expect_\pi[(f-\expect_\pi[f])^2].\]\"<\/p>We will make frequent use of the \"\pi\"-inner product

  <\/span>   <\/span>\"\[\langle f, g\rangle \coloneqq \expect_\pi[f\cdot g] = \sum_{i=1}^m f(i) g(i) \pi_i.\]\"<\/p>
We shall also use expressions such as \"\expect_{x \sim \sigma, y\sim \tau} [f(x,y)]\" to denote the expectation of \"f(x,y)\" where \"x\" is drawn from distribution \"\sigma\" and \"y\" is drawn from \"\tau\".<\/p>\n\n\n\n

We denote the eigenvalues of the transition matrix are \"1 = \lambda_1 \ge \lambda_2 \ge \cdots \ge \lambda_m \ge -1\". The associated eigenvectors (eigenfunctions) \"\varphi_1,\ldots,\varphi_m\" are orthonormal in the \"\pi\"-inner product

  <\/span>   <\/span>\"\[\langle \varphi_i ,\varphi_j\rangle = \begin{cases}1, & i = j, \\0, & i \ne j.\end{cases}\]\"<\/p><\/p>\n\n\n\n

Variance and Local Variance<\/h2>\n\n\n\n

To discover methods for bounding \"\lambda_2\", we begin by investigating a seemingly simple question:<\/p>\n\n\n\n

\n

How much variable<\/em> is the output of a function \"f : \{1,\ldots,m\} \to \real\"?<\/p>\n<\/blockquote>\n\n\n\n

There are two natural quantities which provide answers to this question: the variance<\/em> and the local variance<\/em>. Poincar\u00e9 inequalities\u2014the main subject of this post\u2014establish a relation between these two numbers. As a consequence, Poincar\u00e9 inequalities will provide a bound on \"\lambda_2\".<\/p>\n\n\n\n

Variance<\/h3>\n\n\n\n

We begin with the first of our two main characters, the variance<\/em> \"\Var_\pi(f)\". The variance is a very familiar measure of variation, as it is defined for any random variable. It measures the average squared deviation of \"f(x)\" from its mean, where \"x\" is drawn from the stationary distribution \"\pi\".<\/p>\n\n\n\n

Another helpful formula for the variance is the exchangeable pairs formula<\/a><\/em>:

  <\/span>   <\/span>\"\[\Var_\pi(f) = \frac{1}{2} \expect_{x,y \sim \pi} [(f(x) - f(y))^2].\]\"<\/p>The exchangeable pairs formula states that the variance of \"f\" is proportional to the average square difference of \"f\"‘s values when measured at locations \"x\" and \"y\" sampled (independently<\/em>) from the stationary distribution \"\pi\".<\/p>\n\n\n\n

Local Variance<\/h3>\n\n\n\n

The exchangeable pairs formula shows that variance is a measure of the global variability<\/em> of the function: It measures the amount \"f\" varies across locations \"x\" and \"y\" sampled randomly from the entire set of possible states \"\{1,\ldots,m\}\".<\/p>\n\n\n\n

The local variance<\/em> measures how much \"f\" varies between points \"x\" and \"y\" which are separated by just one step of the Markov chain, thus providing a more local measure of variability. Let \"x_0 \sim \pi\" be sampled from the stationary distribution, and let \"x_1\" denote one step of the Markov chain after \"x_0\". The local variance is

  <\/span>   <\/span>\"\[\mathcal{E}(f) = \frac{1}{2} \expect [(f(x_0) - f(x_1))^2].\]\"<\/p>Other names for the local variance include the Dirichlet form and the quadratic form of the Markov chain.<\/p>\n\n\n\n

An important note: The variance of a function \"f\" depends only on the stationary distribution \"\pi\". By contrast, the local variance depends on the Markov transition matrix \"P\".<\/p>\n\n\n\n

Poincar\u00e9 Inequalities<\/h2>\n\n\n\n

If \"f\" does not vary much over a single step of the Markov chain, then it seems reasonable to expect that it doesn\u2019t vary much globally. This intuition is made quantitative using Poincar\u00e9 inequalities.<\/p>\n\n\n\n

\n

Definition (Poincar\u00e9 inequality).<\/strong> A Markov chain is said to satisfy a Poincar\u00e9 inequality with constant \"\alpha\" if

(1) <\/span>   <\/span>\"\[\Var_\pi(f)\le \alpha \cdot \mathcal{E}(f) \quad \text{for every function } f.\]\"<\/p><\/p>\n<\/blockquote>\n\n\n\n

Poincar\u00e9 Inequalities and Mixing<\/h3>\n\n\n\n

Poincar\u00e9 inequalities are intimately related with the speed of mixing for a Markov chain.<\/p>\n\n\n\n

To see why, consider a function \"f\" with small local variance. Because \"f\" has small local variance, \"f(x_0)\" is close to \"f(x_1)\", \"f(x_1)\" is close to \"f(x_2)\", etc.; the function \"f\" does not change much over a single step of the Markov chain. Does this mean that the (global) variance of \"f\" will also be small? Not necessarily. If the Markov chain takes a long time to mix, the small local variance can accumulate to a large global variance over many steps of the Markov chain. Thus, a slowly mixing chain has a large Poincar\u00e9 constant \"\alpha\". Conversely, if the chain mixes rapidly, the Poincar\u00e9 constant \"\alpha\" is small.<\/p>\n\n\n\n

This relation between mixing and Poincar\u00e9 inequalities is quantified by the following theorem:<\/p>\n\n\n\n

\n

Theorem (Poincar\u00e9 inequalities from eigenvalues).<\/strong> The Markov chain satisfies a Poincar\u00e9 inequality with constant

  <\/span>   <\/span>\"\[\alpha= \frac{1}{1-\lambda_2}.\]\"<\/p>This is the smallest possible Poincar\u00e9 inequality for the Markov chain.<\/p>\n<\/blockquote>\n\n\n\n

One way to interpret this result is that the eigenvalue \"\lambda_2\" gives you Poincar\u00e9 inequality (1). But we can flip this result around: Poincar\u00e9 inequalities (1) establish bounds<\/em> on the eigenvalue \"\lambda_2\".<\/p>\n\n\n\n

\n

Corollary (Eigenvalue bounds from Poincar\u00e9 inequalities).<\/strong> If the Markov chain satisfies a Poincar\u00e9 inequality (1) for a certain constant \"\alpha\", then

  <\/span>   <\/span>\"\[\lambda_2 \le \frac{1}{1-\alpha}.\]\"<\/p><\/p>\n<\/blockquote>\n\n\n\n

A View to the Continuous Setting<\/h3>\n\n\n\n

For a particularly vivid example of a Poincar\u00e9 inequality, it will be helpful to take a brief detour to the world of continuous Markov processes<\/em>. This series<\/a> has\u2014to this point\u2014exclusively focused on Markov chains \"x_0,x_1,x_2,\ldots\" that have finitely many possible states<\/em> and are indexed by discrete times<\/em> \"0,1,2,\ldots\".. We can generalize Markov chains by lifting both of these restrictions, considering Markov processes<\/em> \"(x_t)_{t\ge 0}\" which take values in continuous space (such as the real line \"\real\") and are indexed by continuous times \"t\ge 0\".<\/p>\n\n\n

\n
\"\"<\/figure><\/div>\n\n\n

The mathematical details for Markov processes are a lot more complicated than for their Markov chain siblings, so we will keep it light on details.<\/p>\n\n\n\n

For this example, our Markov process will be the Ornstein\u2013Uhlenbeck process<\/a>. This process has the somewhat mysterious form

  <\/span>   <\/span>\"\[x_t = e^{-t}x_0 + e^{-t} B_{e^{2t}-1},\]\"<\/p>where \"(B_s)_{s\ge 0}\" denotes a (standard) Brownian motion<\/a>, independent of the starting state \"x_0\". At time \"s\", the Brownian motion \"B_s\" has a Gaussian distribution<\/a> with variance \"s\". Thus,<\/p>\n\n\n\n

\n

Conditional on its starting value \"x_0\", the Ornstein\u2013Uhlenbeck process \"x_t\" is has a Gaussian distribution with mean \"e^{-t}x_0\" and variance \"1-e^{-2t}\".<\/p>\n<\/blockquote>\n\n\n\n

From this observation, it appears that the stationary distribution of the Ornstein\u2013Uhlenbeck process is the standard Gaussian distribution<\/a>. Indeed, this is the case, and the Ornstein\u2013Uhlenbeck process converges to stationarity exponentially fast.<\/p>\n\n\n\n

Since we have exponential convergence to stationarity,1<\/a><\/sup>And, as can be checked, the Ornstein\u2013Uhlenbeck process is reversible, in the appropriate sense.<\/span> there’s a Poincar\u00e9 inequality lurking in the background, known as the Gaussian Poincar\u00e9 inequality<\/a><\/em>. Letting \"Z\" denote a standard Gaussian random variable, Gaussian Poincar\u00e9 inequality states that

(2) <\/span>   <\/span>\"\[\Var(f(Z)) \le \expect \big[(f'(Z))^2\big].\]\"<\/p>The right-hand side of this inequality is the local variance of the Ornstein\u2013Uhlenbeck process, equal to the expected squared derivative:

  <\/span>   <\/span>\"\[\mathcal{E}(f) = \expect \big[(f'(Z))^2\big].\]\"<\/p><\/p>\n\n\n\n

The Gaussian Poincar\u00e9 inequality presents a very clear demonstration of what a Poincar\u00e9 inequality is<\/em>: The global variance of the function \"f(Z)\" is controlled by its local variability, here quantified by the expected squared derivative:

  <\/span>   <\/span>\"\[\mathcal{E}(f) = \expect \big[(f'(Z))^2\big].\]\"<\/p>For general Markov chains or processes, it can remain helpful to thinking of the local variance \"\mathcal{E}(f)\" as a generalization of the “expected squared derivative” of the function \"f\".<\/p>\n\n\n\n

Our main interest in Poincar\u00e9 inequalities in this post is instrumental<\/em>, we seek to use Poincar\u00e9 inequalities to understand the mixing properties of Markov chains. But the Gaussian Poincar\u00e9 inequality demonstrates that Poincar\u00e9 inequalities are also interesting on their own terms. The inequality (2) is a useful inequality for bounding the variance of a function of a Gaussian random variable. As an immediate example, observe that the function \"f(x) = \tanh(x)\" has derivative bounded by \"1\": \"|f'(x)| \le 1\". Thus,

  <\/span>   <\/span>\"\[\Var(\tanh Z) \le \expect[(f'(Z))^2] \le 1.\]\"<\/p>This inequality is not too difficult to prove directly,2<\/a><\/sup>Indeed, use the fact that \"\tanh\" is \"1\"–Lipschitz continuous<\/a> and use the exchangeable pairs formula for variance: \"\Var(\tanh(Z)) = 0.5\expect[(\tanh Z - \tanh Z')^2] \le 0.5 \expect[(Z-Z')^2] = \Var(Z) = 1\", where \"Z'\" is an independent copy of \"Z\".<\/span> but the point stands that the Gaussian Poincar\u00e9 inequality\u2014and Poincar\u00e9 inequalities in general\u2014can be useful on their own terms.3<\/a><\/sup>The Gaussian Poincar\u00e9 inequality’s multidimensional generalization has especially interesting consequences. See the bottom of this post<\/a> for an example.<\/span>\n\n\n\n

Poincar\u00e9 Inequalities and Eigenvalues<\/h2>\n\n\n\n

For the remainder of this post, we will develop the connection between Poincar\u00e9 inequalities and eigenvalues, leading to a proof of our main theorem:<\/p>\n\n\n\n

\n

Theorem (Poincar\u00e9 inequalities from eigenvalues).<\/strong> The Markov chain satisfies a Poincar\u00e9 inequality with constant

  <\/span>   <\/span>\"\[\alpha= \frac{1}{1-\lambda_2}.\]\"<\/p>That is,

(3) <\/span>   <\/span>\"\[\Var_\pi(f)\le \frac{1}{1-\lambda_2}\cdot\mathcal{E}(f) \quad \text{for all } f\in\real^m.\]\"<\/p>There exists a function \"f\" for which equality is attained.<\/p>\n<\/blockquote>\n\n\n\n

We begin by showing that it suffices to consider mean-zero functions \"\expect[f] = 0\" to prove (3). Next, we derive formulas for \"\Var_\pi(f)\" and \"\mathcal{E}(f)\" using the \"\pi\"-inner product \"\langle\cdot,\cdot\rangle\". We conclude by expanding \"f\" in eigenvectors of \"P\" and deriving the Poincar\u00e9 inequality (3).<\/p>\n\n\n\n

Shift to Mean-Zero<\/h3>\n\n\n\n

To prove the Poincar\u00e9 inequality (3), we are free to assume that \"f\" has mean zero, \"\expect_\pi[f]=0\". Indeed, both the variance \"\Var_\pi(f)\" and local variance \"\mathcal{E}(f)\" don\u2019t change if we shift \"f\" by a constant \"c\". That is, letting \"\mathbb{1}\" denote the function

  <\/span>   <\/span>\"\[\mathbb{1}(i) = 1 \quad\text{for }i =1,\ldots,m,\]\"<\/p>then

  <\/span>   <\/span>\"\[\Var_\pi(f+c\mathbb{1})=\Var_\pi(f)\quad\text{and}\quad\mathcal{E}(f+c\mathbb{1})=\mathcal{E}(f)\]\"<\/p>for every function \"f\" and constant \"c\". Therefore, for proving our Poincar\u00e9 inequality, we can always shift \"f\" so that it is mean-zero:

  <\/span>   <\/span>\"\[\expect_\pi[f] = 0.\]\"<\/p><\/p>\n\n\n\n

Variance<\/h3>\n\n\n\n

Our strategy for proving the main theorem will be to develop a more linear algebraic<\/em> formula for the variance and local variance. Let\u2019s begin with the variance.<\/p>\n\n\n\n

Assume \"\expect_\pi[f] = 0\". Then the variance is

  <\/span>   <\/span>\"\[\Var_\pi(f)=\expect[f^2]=\sum_{i=1}^m f(i)f(i)\pi_i.\]\"<\/p>Using the definition of the \"\pi\"-inner product, we have shown that

  <\/span>   <\/span>\"\[\Var_\pi(f)=\langle f,f\rangle.\]\"<\/p><\/p>\n\n\n\n

Local Variance<\/h3>\n\n\n\n

Now we derive a formula for the local variance:

  <\/span>   <\/span>\"\[\mathcal{E}(f) = \frac{1}{2} \expect[(f(x_0)-f(x_1))^2]\quad \text{where }x_0\sim\pi.\]\"<\/p>The probability that \"x_0=i\" and \"x_1=j\" is \"\pi_iP_{ij}\". Thus,

  <\/span>   <\/span>\"\[\mathcal{E}(f) = \frac{1}{2} \sum_{i,j=1}^m (f(i)-f(j))^2 \pi_iP_{ij}.\]\"<\/p>Expanding the parentheses and regrouping strategically, we obtain

  <\/span>   <\/span>\"\[\mathcal{E}(f) = {\rm A} + {\rm B} + {\rm C}\]\"<\/p>where

  <\/span>   <\/span>\"\begin{align*}{\rm A} &= \frac{1}{2} \sum_{i=1}^n (f(i))^2 \pi_i \left( \sum_{j=1}^m P_{ij}\right),\\{\rm B} &= \frac{1}{2} \sum_{j=1}^m (f(j))^2 \left( \sum_{i=1}^m \pi_i P_{ij} \right), \\{\rm C} &= -\sum_{i=1}^m f(i)\left( \sum_{j=1}^m P_{ij} f(j) \right) \pi_i.\end{align*}\"<\/p><\/p>\n\n\n\n

Let’s take each of these terms one-by-one. For \"{\rm A}\", recognize that \"\sum_{j=1}^m P_{ij} = 1\". Thus,

  <\/span>   <\/span>\"\[{\rm A} = \frac{1}{2}\sum_{i=1}^m (f(i))^2 \pi_i = \frac{1}{2}\langle f, f\rangle.\]\"<\/p>For \"\rm B\", use detailed balance \"\pi_i P_{ij} = \pi_j P_{ji}\". Then, using the condition \"\sum_{i=1}^m P_{ji} = 1\", we obtain

  <\/span>   <\/span>\"\[{\rm B} = \frac{1}{2} \sum_{j=1}^m (f(j))^2 \left(\sum_{i=1}^m \pi_j P_{ji} \right) = \frac{1}{2} \sum_{j=1}^m (f(j))^2 \pi_j = \frac{1}{2} \langle f, f\rangle.\]\"<\/p>Finally, for \"{\rm C}\", recognize that \"\sum_{j=1}^m P_{ij} f(j)\" is the \"i\"th entry of the matrix\u2013vector product \"Pf\". Thus,

  <\/span>   <\/span>\"\[{\rm C} = - \sum_{i=1}^m f(i) Pf(i) \,\pi_i = -\langle f, Pf\rangle.\]\"<\/p>Thus, we conclude

  <\/span>   <\/span>\"\[\mathcal{E}(f) = \langle f, (I-P)f\rangle,\]\"<\/p>where \"I\" denotes the identity matrix.<\/p>\n\n\n\n

Conclusion<\/h3>\n\n\n\n

The Poincar\u00e9 inequality

  <\/span>   <\/span>\"\[\Var_\pi(f) \le \frac{1}{1-\lambda_2} \cdot\mathcal{E}(f) \quad \text{for all $f$ with $\expect_\pi[f] = 0$}.\]\"<\/p>is equivalent to showing

  <\/span>   <\/span>\"\[\frac{\mathcal{E}(f)}{\Var_\pi(f)}\ge 1-\lambda_2 \quad \text{for all $f$ with $\expect_\pi[f] = 0$}.\]\"<\/p>Using our newly derived formulas, this in turn is equivalent to showing

  <\/span>   <\/span>\"\[\frac{\langle f, (I-P)f\rangle}{\langle f, f\rangle} \ge 1-\lambda_2 \quad \text{for all $f$ with $\expect_\pi[f] = 0$}.\]\"<\/p>We shall prove this version of the Poincar\u00e9 inequality by expanding \"f\" as a linear combination of eigenvectors.<\/p>\n\n\n\n

Consider a decomposition of \"f\" as a linear combination of \"P\"‘s eigenvectors:

  <\/span>   <\/span>\"\[f = c_1 \varphi_1 + c_2 \varphi_2 + \cdots + c_m\varphi_m.\]\"<\/p>As we showed in last post, the condition \"\expect[f] = 0\" is equivalent to saying that \"c_1 = 0\".<\/p>\n\n\n\n

Using the orthonormality of \"\varphi_1,\ldots,\varphi_m\" under the \"\pi\"-inner product and the eigenvalue relation \"P\varphi_i = \lambda_i \, \varphi_i\", we have that

  <\/span>   <\/span>\"\begin{align*}\langle f, f\rangle &= c_2^2 + \cdots + c_m^2, \\\langle f, (I-P)f\rangle &= (1-\lambda_2)c_2^2 + \cdots + (1-\lambda_m)c_m^2.\end{align*}\"<\/p>

Thus,<\/p>\n\n\n\n

(4) <\/span>   <\/span>\"\begin{align*}\frac{\langle f, (I-P)f\rangle}{\langle f, f\rangle} &= \frac{(1-\lambda_2)c_2^2 + \cdots + (1-\lambda_m)c_m^2}{c_2^2 + \cdots + c_m^2} \\&= (1-\lambda_2)a_2 + \cdots + (1-\lambda_m)a_m,\end{align*}\"<\/p><\/p>\n\n\n\n

where

  <\/span>   <\/span>\"\[a_i = \frac{c_i^2}{c_2^2 + \cdots + c_m^2}.\]\"<\/p>The coefficients \"a_i\" are nonnegative and add to \"1\":

  <\/span>   <\/span>\"\[a_2+\cdots+a_m = \frac{c_2^2+\cdots+c_m^2}{c_2^2+\cdots+c_m^2} = 1.\]\"<\/p>Therefore, the smallest possible value for (4) is achieved by setting \"a_2 = 1\" and \"a_3 = \cdots = a_m = 0\" (equivalently, setting \"c_3 = \cdots=c_m = 0\"). Thus, we conclude

  <\/span>   <\/span>\"\[\frac{\langle f, (I-P)f\rangle}{\langle f, f\rangle}\ge 1-\lambda_2,\]\"<\/p>with equality when \"f\" is a multiple of \"\varphi_2\".<\/p>\n","protected":false},"excerpt":{"rendered":"

In the previous posts, we’ve been using eigenvalues to understand the mixing of reversible Markov chains. Our main convergence result was as follows:     Here, denotes the distribution of the chain at time , denotes the stationary distribution, denotes the divergence, and denote the decreasingly ordered eigenvalues of the Markov transition matrix . BoundingRead more<\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[11],"tags":[],"class_list":["post-1809","post","type-post","status-publish","format-standard","hentry","category-markov-musings"],"_links":{"self":[{"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/posts\/1809","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/comments?post=1809"}],"version-history":[{"count":11,"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/posts\/1809\/revisions"}],"predecessor-version":[{"id":2090,"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/posts\/1809\/revisions\/2090"}],"wp:attachment":[{"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/media?parent=1809"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/categories?post=1809"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ethanepperly.com\/index.php\/wp-json\/wp\/v2\/tags?post=1809"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}