Self-Similarity Conditions for Adaptive Confidence Bands

Why Scale-Wise Signal Energy Matters

Adaptive confidence bands face a structural difficulty: a function can look very smooth at the resolutions observed in a sample while hiding variation at finer scales. A bandwidth chosen from the visible, coarse-scale behavior may then be too large, understating the function’s bias and producing a band that misses the truth. Self-similarity conditions address this particular deception by requiring signal energy to recur across scales.

The condition is not a universal guarantee of honest adaptive bands. It restricts the function class so that smoothness cannot be suggested merely by long, uninformative gaps in the signal. Converting that restriction into a confidence-band guarantee still requires appropriate assumptions on the regression model, noise, and band construction.

Wavelet-Scale Energy

For precision, take a periodic wavelet basis on $[0,1]$, with regularity exceeding the smoothness range under consideration. Write the regression function $f$ in its scaling and wavelet coefficients. At resolution $j$, there are $2^j$ wavelet coefficients, denoted $\theta_{jk}$, $0\leq k<2^j$. The energy at that resolution is the sum of their squares.

f=\sum_k \alpha_k\phi_k+\sum_{j\geq 0}\sum_{k=0}^{2^j-1}\theta_{jk}\psi_{jk}, \qquad E_j(f)=\sum_{k=0}^{2^j-1}\theta_{jk}^2.

Orthonormality means $E_j(f)$ is the squared $L^2$ size of the detail contributed by resolution $j$. It is an aggregate: it does not say where on $[0,1]$ the energy occurs. This distinction matters for bands that must cover the function uniformly.

Definition: Fix an integer $K\geq 1$. A function has upper scale regularity of order $s>0$ if, for some finite constant $A$, $E_j(f)\leq A2^{-2sj}$ for every $j$. It is $(s,c,K,j_0)$-self-similar if, for every $j\geq j_0$, $$ \sum_{r=0}^{K-1}E_{j+r}(f)\geq c\,2^{-2sj} $$ for some $c>0$. In an adaptive class, the allowed smoothness range and the constants $A,c,K,j_0$ are taken uniformly over its functions.

The upper bound says that energy cannot decay more slowly than the stated smoothness permits. The self-similarity lower bound says that energy cannot disappear across every scale in a block of $K$ consecutive resolutions. It does not demand a positive lower bound at each individual scale; a few scales in a block may carry little or no energy.

These conditions are scale-based rather than a complete definition of a classical smoothness space. They provide a convenient way to state the relevant upper decay and recurring lower signal. In applications, smoothness classes may be described using Hölder or Besov norms, with wavelet characterizations under suitable basis regularity.

Worked Example: Geometric Energy at Every Resolution

Suppose that at each scale $j$, all $2^j$ coefficients equal $2^{-j(s+1/2)}$, for a fixed $s>0$. Calculate the scale energy and check the block condition.

Each squared coefficient is $2^{-2j(s+1/2)}$. Summing over the $2^j$ coefficients gives $$ E_j=2^j\,2^{-2j(s+1/2)} =2^j\,2^{-2sj-j} =2^{-2sj}. $$

Thus the upper bound holds with $A=1$. For a block of $K$ scales, $$ \sum_{r=0}^{K-1}E_{j+r} =2^{-2sj}\sum_{r=0}^{K-1}2^{-2sr} =2^{-2sj}\frac{1-2^{-2sK}}{1-2^{-2s}}. $$

The multiplier is positive for $s>0$, so the self-similarity lower bound holds with $c=(1-2^{-2sK})/(1-2^{-2s})$. The coefficient sequence is square-summable: $\sum_j E_j=\sum_j2^{-2sj}<\infty$. This example has a fixed, nonvanishing energy pattern after rescaling by the expected smoothness decay.

What the Condition Rules Out

Upper regularity alone allows a function to have zero energy at many successive scales. That can create a misleading impression of exceptional smoothness when the observed resolutions fall in one of those gaps. Self-similarity rules out gaps of arbitrary length: every sufficiently fine block of $K$ scales must contain a definite amount of energy.

Worked Example: A Sparse Sequence of Active Scales

Consider a wavelet expansion whose scale energies are $E_j=2^{-2sj}$ when $j$ is a power of two, and $E_j=0$ otherwise. Verify the upper bound and assess self-similarity for any fixed $K$.

At an active scale, $E_j=2^{-2sj}$; at an inactive scale, $E_j=0$. Therefore $E_j\leq 2^{-2sj}$ for every $j$, so upper scale regularity holds with $A=1$.

The gaps between successive powers of two grow without bound. Given any fixed $K$, choose a power of two $2^r$ so large that $2^{r+1}-2^r>K$. There are then $K$ consecutive indices strictly between $2^r$ and $2^{r+1}$. For a block starting at one of these indices, every energy is zero, and hence $$ \sum_{q=0}^{K-1}E_{j+q}=0<c\,2^{-2sj} $$ for every $c>0$. The sequence satisfies the upper bound but fails self-similarity. Its long empty stretches are exactly what the lower block condition excludes.

This example also clarifies why the lower bound is a restriction on the signal, not a routine consequence of smoothness. A smoothness upper bound limits how large details can be; it does not require details to keep appearing.

Worked Example: One High-Resolution Detail

Let all wavelet coefficients be zero except one coefficient at scale $J$, whose value is $2^{-sJ}$. Find the scale energies and determine whether this function meets the two conditions.

At scale $J$, one coefficient contributes its square; at every other scale, all coefficients vanish. Thus $$ E_j= \begin{cases} 2^{-2sJ},&j=J,\\ 0,&j\ne J. \end{cases} $$

For $j=J$, $E_j=2^{-2sj}$; for $j\ne J$, $E_j=0$. The upper bound therefore holds with $A=1$. But for any fixed block length $K$, a block sufficiently far from $J$ has total energy zero, so no positive self-similarity constant can work for all $j\geq j_0$. Even at the active scale, a block beginning at $J+1$ has no energy.

If observations primarily reveal scales below $J$, this function has no detail energy at those scales and can look like a zero signal there. This is a finite-resolution warning, not a claim that the function is indistinguishable under every observation model. Its point is that an upper smoothness bound does not prevent signal from being concentrated in a single distant scale.

A Consequence for Comparing Scales

A useful implication is that block energies have a controlled log-scale slope. This makes precise how recurring signal can help distinguish smoothness levels. Let $B_j=\sum_{r=0}^{K-1}E_{j+r}(f)$. Suppose the upper scale bound and self-similarity condition hold. Then for all $j\geq j_0$, $$ c\,2^{-2sj}\leq B_j\leq A\,2^{-2sj}\sum_{r=0}^{K-1}2^{-2sr} \leq AK\,2^{-2sj}. $$

Proposition: Under these bounds, set $C=AK$. For any integer $m\geq1$ and $j\geq j_0$, define the block-energy slope $$ \widehat{s}_{j,m}=-\frac{\log_2 B_{j+m}-\log_2 B_j}{2m}. $$ Then $$ \left|\widehat{s}_{j,m}-s\right|\leq\frac{\log_2(C/c)}{2m}. $$

Proof. The bounds ensure $B_j>0$ and $B_{j+m}>0$, so their logarithms exist. Write $\log_2 B_q=-2sq+\delta_q$ for $q\geq j_0$. The energy bounds imply $\log_2 c\leq\delta_q\leq\log_2 C$, so $|\delta_{j+m}-\delta_j|\leq\log_2(C/c)$. Substituting, $$ \widehat{s}_{j,m} =-\frac{-2s(j+m)+\delta_{j+m}+2sj-\delta_j}{2m} =s-\frac{\delta_{j+m}-\delta_j}{2m}. $$ Taking absolute values gives the claimed bound. $\square$

This is a deterministic statement about the signal’s energies. It does not say that noisy empirical energies estimate $B_j$ accurately at every scale. In data, estimation error, the number of coefficients, and the available range of resolutions all matter. The proposition shows why a recurring signal makes scale comparisons informative: the error due to unknown multiplicative constants shrinks when the comparison spans more scales.

Worked Example: Bounding a Two-Scale Slope

Suppose $s=1$, $K=2$, and the energies satisfy the bounds $c2^{-2j}\leq B_j\leq C2^{-2j}$ with $c=1/2$ and $C=3$. How accurately does a comparison across $m=4$ scales constrain the slope?

The proposition gives $$ \left|\widehat{s}_{j,4}-1\right| \leq\frac{\log_2(3/(1/2))}{8} =\frac{\log_2 6}{8}. $$

Since $2^2=4<6<8=2^3$, $2<\log_2 6<3$. Therefore the error is less than $3/8=0.375$. The bound is conservative: it controls every sequence of block energies satisfying the stated constants, including sequences whose scale-to-scale multipliers fluctuate within the allowed range.

Connection to Adaptive Bandwidths

In one-dimensional regression, a common scale for a bandwidth-based estimate is a bias of order $h^s$ and a uniform stochastic fluctuation of order $\sqrt{\log n/(nh)}$. This balance presumes conditions that justify those orders; the self-similarity definition by itself does not establish them. Equating the two terms gives $$ h^s=\sqrt{\frac{\log n}{nh}} \quad\Longrightarrow\quad h^{2s+1}=\frac{\log n}{n} \quad\Longrightarrow\quad h=\left(\frac{\log n}{n}\right)^{1/(2s+1)}. $$

When $s$ is unknown, a procedure must choose a scale without knowing which bias-variance balance applies. Recurring wavelet-scale energy provides a structural reason that comparisons across resolutions can carry information about $s$, rather than encountering arbitrarily long signal-free regions. But an energy condition is not itself a bandwidth-selection rule or a confidence-band theorem.

In particular, the energy $E_j(f)$ is an $L^2$ quantity. Uniform coverage concerns $\sup_x|\hat f(x)-f(x)|$, so aggregate energy alone does not control the location or maximum size of local errors. A valid adaptive-band argument needs further assumptions and a construction that controls uniform bias and stochastic fluctuation, as well as the relevant coverage probability.

Common Pitfall

Wrong reasoning: “The function satisfies an upper smoothness bound, so every scale must contain the amount of energy predicted by that smoothness.” The upper bound is only $E_j\leq A2^{-2sj}$; zero energy satisfies it.
Correction: A recurring lower bound is an additional assumption. The block condition requires positive energy somewhere in every sufficiently fine window of $K$ scales, and thereby excludes arbitrarily long gaps.
Wrong reasoning: “A lower bound on total scale energy guarantees a narrow honest uniform band.” Total energy does not locate the signal, and an $L^2$ bound is not a uniform-error bound.
Correction: Treat self-similarity as a restriction that can support adaptation under additional model and construction conditions, not as a stand-alone coverage guarantee.

Key takeaway: Upper regularity limits how much detail a function can have; self-similarity prevents its detail energy from vanishing across arbitrarily long runs of scales. This recurring signal can make smoothness comparisons informative, but honest adaptive bands require more than scale energy alone.

Check Your Understanding

Use the definitions and derivations above to answer these questions.

Why can a function satisfy the upper scale bound at every resolution while failing self-similarity?
For coefficients $\theta_{jk}=2^{-j(s+1/2)}$ at all $2^j$ locations of scale $j$, derive $E_j$.
What feature of the sparse active-scale example makes it fail every fixed block self-similarity condition?
In the block-energy slope proposition, why does comparing across $m$ scales reduce the bound on slope error?
Why does a lower bound on aggregate wavelet energy not, by itself, imply uniform coverage of a confidence band?