Polished-Tail Conditions in Gaussian Sequence Models

Why Control Energy Across Scales?

In a Gaussian sequence model, the unknown signal is represented by coefficients at successive resolution levels. Truncating the expansion at a chosen level makes estimation possible, but it omits every coefficient at finer levels. The omitted signal energy is the squared bias of that truncation. A difficulty arises when a signal has little energy at the levels used to assess its roughness, but substantial energy far beyond them: a finite-scale inspection can then miss a fine-scale tail.

A polished-tail condition rules out this kind of hidden energy by requiring that the total tail beyond a level be controlled by the energy in a nearby range of levels. It is a restriction on the signal, not a conclusion that follows from the Gaussian observation model. Its value for adaptive uncertainty quantification is that it links unresolved fine-scale energy to signal energy at scales that are less remote. It does not, by itself, construct a confidence band or guarantee coverage.

Sequence Model and Definition

Let $j=0,1,2,\ldots$ index resolution levels, and let $k=0,\ldots,2^j-1$ index coefficients within level $j$. In the Gaussian sequence model, suppose

Y_{jk}=\theta_{jk}+n^{-1/2}Z_{jk}, \qquad Z_{jk}\ \text{independent and distributed as }N(0,1).

The unknown sequence $\theta=(\theta_{jk})$ is assumed square-summable, so that its total squared energy is finite. For an orthonormal wavelet expansion, these are the wavelet coefficients of the function, and squared coefficient energy is squared $L^2$ energy. Define the energy at level $j$ and the tail energy from level $J$ by

B_j=\sum_{k=0}^{2^j-1}\theta_{jk}^2, \qquad T_J=\sum_{j=J}^{\infty}B_j.

For an integer $J\geq 0$ and a fixed real number $\rho>1$, the nearby block runs from $J$ through $\lfloor \rho J\rfloor$, inclusive. The condition compares $T_J$ with the energy in that block.

Definition: For fixed constants $C<\infty$, $\rho>1$, and an integer $J_0\geq 0$, a square-summable sequence $\theta$ satisfies the polished-tail condition with parameters $(C,\rho,J_0)$ if, for every integer $J\geq J_0$, $$ T_J\leq C\sum_{j=J}^{\lfloor\rho J\rfloor}B_j. $$ The lower cutoff $J_0$ is part of the definition: the inequality is required eventually, not necessarily at every level from zero.

In applications, one may study a class of sequences for which $C,\rho,J_0$ are fixed uniformly across the class. Allowing constants to vary across individual signals gives a pointwise structural condition, but by itself does not yield uniform guarantees over a class with unbounded constants or cutoffs.

Energy Patterns That Satisfy the Condition

Worked Example: Geometrically Decaying Level Energy

Suppose $B_j=Aq^j$ for every $j\geq 0$, where $A>0$ and $0<q<1$. For any $J$, the tail is the convergent geometric sum $$ T_J=\sum_{j=J}^{\infty}Aq^j=\frac{Aq^J}{1-q}. $$

Write $L_J=\lfloor\rho J\rfloor-J$, which is a nonnegative integer. The energy in the comparison block is $$ \sum_{j=J}^{\lfloor\rho J\rfloor}B_j =Aq^J\sum_{r=0}^{L_J}q^r =\frac{Aq^J(1-q^{L_J+1})}{1-q}. $$

Since $L_J+1\geq 1$ and $0<q<1$, we have $q^{L_J+1}\leq q$, so $1-q^{L_J+1}\geq 1-q$. Therefore $$ \frac{T_J}{\sum_{j=J}^{\lfloor\rho J\rfloor}B_j} =\frac{1}{1-q^{L_J+1}} \leq\frac{1}{1-q}. $$ Thus the polished-tail condition holds for every $J$, with $C=1/(1-q)$, for any $\rho>1$. The amplitude $A$ cancels from the ratio.

This calculation concerns the level energies; such energies can be realized by putting the energy $Aq^j$ in one coefficient at each level and setting the other coefficients to zero. The resulting sequence is square-summable because $\sum_j Aq^j<\infty$.

The condition is not the same as requiring a smoothness class with one particular rate of decay. Rather, it requires energy in the tail to be represented, up to a multiplicative constant, in a prescribed range of levels. The exact constants and scale range matter: a very large $C$ gives a weaker restriction, and increasing $\rho$ makes the comparison block longer.

Separated Spikes and the Eventual Quantifier

A persistent gap in the energy distribution can defeat the condition. To establish failure, however, it is not enough to point to one finite-level spike. The condition only applies for $J\geq J_0$, and $J_0$ may be placed beyond any finite collection of nonzero coefficients. A genuine counterexample must produce violations at arbitrarily large $J$.

Result: Fix $\rho>1$. If there is an unbounded sequence of integer levels $J_r$ such that $$ \sum_{j=J_r}^{\lfloor\rho J_r\rfloor}B_j=0 \quad\text{and}\quad T_{J_r}>0, $$ then the sequence fails the polished-tail condition for every finite $C$ and every finite cutoff $J_0$.

Proof. For any proposed finite $C$, at each such level $J_r$ the right-hand side of the polished-tail inequality is zero, whereas its left-hand side is $T_{J_r}>0$. Hence the inequality fails at $J_r$. Because the $J_r$ are unbounded, at least one such violation occurs above any proposed cutoff $J_0$. No choice of $C<\infty$ and $J_0<\infty$ can make the condition hold. $\square$

Worked Example: Infinitely Many Separated Spikes

Fix $\rho>1$. Choose integer levels $m_1,m_2,\ldots$ recursively so that $m_r>2\rho m_{r-1}$ for every $r\geq 2$. At level $m_r$, set one coefficient equal to $2^{-r}$, and set all other coefficients to zero. Then $B_{m_r}=4^{-r}$, all other $B_j=0$, and the total energy $\sum_{r=1}^{\infty}4^{-r}=1/3$ is finite.

For $r\geq 2$, take $J_r=2m_{r-1}$. The previous spike is below $J_r$, and the next spike satisfies $$ m_r>2\rho m_{r-1}=\rho J_r. $$ All later spikes are above $m_r$, so the block from $J_r$ through $\lfloor\rho J_r\rfloor$ contains no spike. Its energy is therefore zero. But the tail includes the spike at $m_r$, so $T_{J_r}\geq 4^{-r}>0$. Since $J_r$ tends to infinity, the result just proved shows that this sequence fails the polished-tail condition for every finite $C$ and $J_0$.

The issue is not that the coefficients are nonzero at only a few levels by itself. It is that there are arbitrarily large gaps in which the comparison block has no energy even though a later tail still contains positive energy.

Worked Example: Why One Finite Spike Is Not Enough

Suppose the only nonzero coefficient is $\theta_{m,k}=a$, with $a\ne 0$. Then $B_m=a^2$ and $B_j=0$ for $j\ne m$. Choose $J_0=m+1$. For every $J\geq J_0$, the tail $T_J=0$, and the comparison-block energy is also zero. Consequently $T_J\leq C\sum_{j=J}^{\lfloor\rho J\rfloor}B_j$ holds for every finite $C$, for every $\rho>1$.

This sequence can fail an inequality required at a particular earlier $J$ if the spike lies in the tail but outside the comparison block. That does not establish failure of the eventual condition when $J_0$ is allowed to exceed $m$. The quantifier over all sufficiently large $J$ is essential.

How the Condition Controls Truncation Bias

Consider the estimator that retains observations below level $J$ and sets all coefficients at levels $j\geq J$ to zero. Under squared sequence error, its omitted-signal contribution is exactly $T_J$. The following calculation gives a direct link from the polished-tail restriction to the estimator's risk.

Result: Define $\hat\theta^{(J)}_{jk}=Y_{jk}$ for $j<J$, and $\hat\theta^{(J)}_{jk}=0$ for $j\geq J$. If $\theta$ is square-summable, then $$ E\!\left[\sum_{j=0}^{\infty}\sum_{k=0}^{2^j-1} \bigl(\hat\theta^{(J)}_{jk}-\theta_{jk}\bigr)^2\right] =\frac{2^J-1}{n}+T_J. $$ If in addition the polished-tail condition holds at $J$, then this risk is at most $$ \frac{2^J-1}{n} +C\sum_{j=J}^{\lfloor\rho J\rfloor}B_j. $$

Proof. For each included coefficient $j<J$, the estimation error is $n^{-1/2}Z_{jk}$, whose expected square is $1/n$. The number of included coefficients is $$ \sum_{j=0}^{J-1}2^j=2^J-1. $$ For every omitted coefficient $j\geq J$, the estimate is zero, so its squared error is $\theta_{jk}^2$; summing these terms gives $T_J$. Adding the included and omitted contributions proves the equality. Applying the polished-tail inequality $T_J\leq C\sum_{j=J}^{\lfloor\rho J\rfloor}B_j$ gives the stated bound. $\square$

The bound explains the structural role of the restriction: it says that the truncation bias cannot be arbitrarily large relative to energy in the nearby block. It does not say that this block energy is known, or that it can be estimated accurately at every resolution. In the Gaussian model, an observable level-energy quantity can be debiased, but its noise must be accounted for.

Worked Example: Debiasing an Observed Level Energy

At a fixed level $j$, let $d_j=2^j$ and define $$ W_j=\sum_{k=0}^{d_j-1}Y_{jk}^2-\frac{d_j}{n}. $$ Using $Y_{jk}=\theta_{jk}+n^{-1/2}Z_{jk}$, expand each square: $$ E[Y_{jk}^2] =\theta_{jk}^2+2\theta_{jk}n^{-1/2}E[Z_{jk}] +n^{-1}E[Z_{jk}^2] =\theta_{jk}^2+\frac1n. $$ Summing over the $d_j$ coefficients and subtracting $d_j/n$ yields $E[W_j]=B_j$. Thus $W_j$ is an unbiased estimate of the signal energy at level $j$. It need not be nonnegative in a particular sample, and unbiasedness alone does not guarantee precision: as the number of coefficients grows, the noise contribution also grows. Using such quantities to build uncertainty bands requires additional calibration.

Interpretation for Adaptive Uncertainty Quantification

Adaptive procedures must learn enough about the signal's effective resolution to choose how much of its expansion to retain. If substantial energy can be hidden arbitrarily far beyond the levels supplying that information, a data-driven choice can underestimate truncation bias. A polished-tail restriction addresses this particular obstacle by ruling out tail energy that has no corresponding energy in the specified nearby block.

This is a structural assumption about coefficient organization across scales. It does not assert that every function of interest has the property, and it is not equivalent to a universal guarantee of honest adaptive bands. The constants and cutoff must be appropriate to the signal class, and statistical procedures still need to handle observation noise and uncertainty in estimating scale energy. The earlier material on self-similarity conditions gives a related way to constrain scale-wise signal behavior; the present condition is specifically a tail-versus-block inequality for sequence coefficients.

Common Mistakes

Claiming that any finite spike violates the condition. The wrong argument finds a level where tail energy is positive but the comparison block is empty. The correction is to check whether that level is required to be at least $J_0$. A single finite spike has zero tail above its level and can satisfy the eventual condition with a sufficiently large cutoff.
Ignoring the comparison-block endpoint. The block includes levels $J$ through $\lfloor\rho J\rfloor$, inclusive. A spike at its upper endpoint contributes to the block; a spike strictly above it does not.
Treating the condition as a statement about observations alone. It is a restriction on the unknown coefficient sequence. Noisy squared observations do not equal level energy, and their noise contribution must be accounted for.
Calling the condition a confidence-band theorem. It bounds tail energy relative to nearby signal energy. Coverage and width require further statistical arguments and suitable control of the noise and constants.

Key takeaway: A polished-tail condition requires every sufficiently distant tail to be controlled by energy in a specified nearby block. It can limit hidden fine-scale bias, but its eventual quantifier means that isolated finite spikes are not automatically counterexamples; persistent gaps with later tail energy are.

Check Your Understanding

Use the definitions and quantifiers carefully.

For $B_j=Aq^j$, with $A>0$ and $0<q<1$, derive the ratio of tail energy to block energy and give a valid constant $C$.
Why does a single nonzero coefficient at level $m$ not necessarily violate a condition that is required only for $J\geq J_0$?
For fixed $\rho>1$, describe how to choose an unbounded sequence of levels $J_r$ so that each comparison block is empty but its tail contains a later spike.
What are the two terms in the expected squared error of the level-$J$ truncation estimator, and what does each represent?
Why does $E[W_j]=B_j$ not imply that the observed $W_j$ is a reliable estimate of $B_j$ at every resolution?