Two Ways of Measuring

author: Rowan Brad Quni-Gudzinas

ORCID: 0009-0002-4317-5604

ISNI: 0000000526456062

title: Two Ways of Measuring

aliases:

- Two Ways of Measuring

modified: 2026-05-02T11:49:48Z

A Framework for Distance, Memory, and Fault-Tolerant Computation

Author: Rowan Brad Quni-Gudzinas

Contact: [email protected]

ORCID: 0009-0002-4317-5604

ISNI: 0000000526456062

DOI: 10.5281/zenodo.19976945

Date: 2026-05-02

Version: 0.3

Prologue: The Memory of a Pebble

A pebble rests in a shallow depression on a granite outcrop in the Australian outback. It has been there for ten thousand years. Rain has fallen on it. Wind has blown over it. The ground has trembled with distant earthquakes. The pebble has jiggled, rattled, and shifted—but it has never left its depression.

Why?

Because the depression is a container. Its walls rise just high enough that the random jostling of the world—a raindrop’s splash, a gust of wind, the tremor of a far-off quake—cannot lift the pebble over the rim. The pebble is free to move within its container, but it cannot escape without a push that exceeds the rim height. The container remembers the pebble’s rough position across geological time, not because it actively corrects the pebble’s location, but because its geometry passively rejects perturbations below a threshold.

The principle the pebble embodies can be generalized and formalized into a framework for building computers whose memories endure through the geometry of their state space rather than through constant vigilance.

The principle has a name: the threshold principle. The mathematics that underlies it is ultrametric geometry.

Part Zero: Boundaries and Containers

0.1 The Primitive Act of Distinction

Consider the most fundamental cognitive act possible: you draw a line. The line separates the world into two regions—an inside and an outside. Everything on one side of the line is “this.” Everything on the other side is “not this.”

This act—called a distinction—is more primitive than counting, more primitive than naming, more primitive than measurement. Before you can say “one, two, three,” you must be able to say “this, not that.” The boundary creates the very possibility of identity. Without boundaries, everything is everything else; nothing can be singled out.

A distinction requires nothing more than the ability to discriminate—to notice that one thing is not the same as another thing. It is the atom of cognition.

Definition (Boundary). A boundary is a rule that, for any candidate point in a space, answers the question: “inside or outside?” The answer must be unambiguous—a point cannot be both inside and outside, and it cannot be neither.

A boundary is:

• Sharp: there is no “kind of inside.” A point is inside or it is not.
• Closed: the inside is fully enclosed. There is no way to move from inside to outside without crossing the boundary.
• Identity-conferring: the inside is “this container.” It has an identity distinct from its contents.

The space inside a boundary, together with the boundary itself, is called a container.

Definition (Container). A container is a boundary together with the region it encloses. A container is an entity in its own right, distinct from whatever it may contain. An empty container is still a container.

0.2 Containers: Nesting

Having drawn one boundary, you can draw another entirely inside the first. This creates a container within a container—a nesting.

Now consider a point inside the inner container. That point is:

• Inside the inner container (a specific fact)
• Inside the outer container (a less specific fact)
• Not inside any container drawn outside the inner one but inside the outer one (a negative fact)

The point’s address—the complete description of which containers hold it—is a sequence of choices. At the outermost level, it is inside Container A (as opposed to Container B, C, etc., if multiple containers were drawn side by side). At the next level, within Container A, it is inside Container A1 (as opposed to A2, A3, etc.). And so on.

This sequence of choices is the most primitive form of measurement. To measure is to answer: in which container does this thing reside? The precision of a measurement is the depth of nesting—how many levels of “which container?” have been answered.

Definition (Address). The address of a point is the sequence of choices that identifies which container holds it at each level of nesting, from the outermost level inward. If the nesting has $D$ levels, the address is a sequence of length $D$.

Example. Imagine three levels of nesting:

• Level 1: Two large boxes, labeled $0$ and $1$.
• Level 2: Inside each large box, two medium boxes, labeled $0$ and $1$.
• Level 3: Inside each medium box, two small boxes, labeled $0$ and $1$.

A specific small box has the address $(1, 0, 1, \ldots)$: it is in large box $1$, medium box $0$ within that, small box $1$ within that. The address continues as long as the nesting continues.

This is exactly the structure of the infinite tree studied in Part Two.

0.3 Two Ways to Nest

Way 1: Linear nesting. Containers are nested along a single dimension, like Russian dolls. Each container fits inside exactly one larger container of the same shape. There is no branching—at each level of nesting, there is only one “next container inward.” The address is a single real number (the coordinate of the point). Precision is the length of the interval.

Way 2: Branching nesting. At each level, a container contains multiple alternative sub-containers, and the point must be in exactly one of them. The address is a sequence of choices—a path through the tree of containers. Precision is the depth of the path.

These two ways are not merely different. They are the only two ways, in a precise sense established in Part Four.

0.4 Zero and Emptiness

If zero is simply “nothing,” how can nothing be something? How can we speak of it, count with it, use it in calculations? The container perspective resolves this. Zero is an empty container. The boundary exists—the line that defines the container has been drawn. The space inside is empty. The emptiness is a positive fact about the container, not the absence of all facts. Zero is the number that answers the question: “How many items are in this container?” when the answer is “none.”

This is not a semantic trick. It is the operational definition of zero: the count of a container verified to contain nothing. The container is real; its emptiness is a real property. Zero is a second-order concept—a number about a number—and it is only coherent when understood through the container framework.

Part One: The Continuous Way

The number line. The triangle inequality. Why errors accumulate, and why this matters for computation.

1.1 The Number Line as a Continuum of Nested Intervals

Begin with the act of counting: one, two, three. Call these the natural numbers, written $\mathbb{N} = \{1, 2, 3, \ldots\}$. Add zero and the negatives to form the integers $\mathbb{Z} = \{\ldots, -2, -1, 0, 1, 2, \ldots\}$. Between any two integers, insert fractions: $1/2, 3/4, 22/7$. These are the rational numbers $\mathbb{Q}$.

The rationals are dense—between any two, no matter how close, there is always another rational. Yet the rationals have gaps. The diagonal of a unit square has length $\sqrt{2}$, which is not a rational number. To fill all gaps, complete the rationals to obtain the real numbers $\mathbb{R}$. Every point on a continuous line corresponds to a real number. There are no gaps.

The real numbers satisfy the Archimedean property: for any positive numbers $x$ and $y$, no matter how small $x$ and how large $y$, you can add $x$ to itself enough times to exceed $y$:

There are no infinitesimals—no numbers so small that no finite sum can reach a finite amount.

The container interpretation. On the number line, an interval $(a, b) = \{x : a < x < b\}$ is a container. Its length is $b - a$. Nesting intervals—$(0, 1)$, then $(0.2, 0.3)$, then $(0.24, 0.25)$, and so on—gives increasingly precise addresses. A point’s address is its coordinate, a single real number. The precision is the length of the smallest interval that can be specified.

1.2 The Ordinary Distance Formula

On the real line, the distance between two points is simply the absolute value of their difference:

Definition (Absolute value). For a real number $x$, its absolute value $|x|$ is:

• $x$ if $x \geq 0$
• $-x$ if $x < 0$

Geometrically, $|x - y|$ is the length of the interval between $x$ and $y$.

Verification of the metric axioms.

Non-negativity. $|x - y| \geq 0$ by definition. And $|x - y| = 0$ if and only if $x = y$. ✓

Symmetry. $|x - y| = |-(y - x)| = |y - x|$, so $d(x, y) = d(y, x)$. ✓

Triangle inequality. The absolute value satisfies the subadditivity property: $|a + b| \leq |a| + |b|$ for all real $a, b$. Setting $a = x - y$ and $b = y - z$:

✓ All three axioms are satisfied. $(\mathbb{R}, d_{\mathbb{R}})$ is a metric space.

Concrete examples.

• $d(3, 7) = |3 - 7| = 4$
• $d(-2.5, 1.5) = |-2.5 - 1.5| = |-4| = 4$
• $d(0.001, 0.002) = |0.001 - 0.002| = 0.001$

1.3 The Triangle Inequality and Error Accumulation

The triangle inequality $d(x, z) \leq d(x, y) + d(y, z)$ has a direct consequence for any system that stores information as a real number: small errors add up.

Suppose a system maintains a target value $T$. Noise causes perturbations $\varepsilon_1, \varepsilon_2, \varepsilon_3, \ldots$ over time. After the first perturbation, the state is $T + \varepsilon_1$. After the second, $T + \varepsilon_1 + \varepsilon_2$. After $N$ perturbations, the state is:

The total error—the distance from the target—is:

Apply the triangle inequality repeatedly:

If each perturbation is bounded by $\varepsilon$ (that is, $|\varepsilon_i| \leq \varepsilon$ for all $i$), then:

The total error can be as large as the sum of the individual errors. With enough perturbations, the error can exceed any finite bound.

Even in the best-case scenario where perturbations have zero mean and are independent, the root-mean-square error grows without bound:

where $\sigma$ is the standard deviation of each perturbation. The error grows more slowly (as $\sqrt{N}$ rather than $N$), but still grows without limit.

Concrete example. A sensor maintains a voltage at $5.000$ V. Thermal noise causes independent random fluctuations of $\pm 0.001$ V every second. After $1$ hour ($3{,}600$ seconds), the RMS error is $0.001 \times \sqrt{3600} = 0.060$ V—a $1.2\%$ drift. After $1$ day ($86{,}400$ seconds), the RMS error is $0.294$ V—a $5.9\%$ drift. After $1$ month, the error exceeds $1.6$ V—the reading is meaningless. Without active correction, continuous memory inevitably decays.

1.4 The Thermodynamic Wall

The only defense against error accumulation in a continuous space is active error correction: constantly measure the state, compare it to the target, and apply a counter-force to push it back.

Active correction costs energy. The lower bound is given by Landauer’s principle: erasing one bit of information (which measurement inevitably requires) dissipates at least $k_B T \ln 2$ of energy, where $k_B$ is Boltzmann’s constant ($1.38 \times 10^{-23}$ J/K) and $T$ is the temperature.

To maintain an $n$-bit state (precision of one part in $2^n$):

1. Measure to $n$-bit precision, dissipating at least $n \cdot k_B T \ln 2$ per measurement.

1. Compute a correction, which requires additional bit operations.

1. Apply the correction, which requires work against the noise.

The required measurement frequency increases with noise intensity. For a system with noise bandwidth $B$ (roughly, the rate at which independent perturbations occur), measurement must occur at a rate comparable to $B$ to keep up. The total power dissipation scales as:

For a quantum computer maintaining $1{,}000$ logical qubits, with each requiring $n \approx 10^3$ physical qubits (as in surface code architectures) and clock speeds of $B \approx 10^6$ Hz at $T \approx 0.01$ K, the minimum power for measurement alone is on the order of $10^{-10}$ W—small, but this is a lower bound. Practical implementations require orders of magnitude more, and the scaling with qubit count is unfavorable.

This is the thermodynamic wall: the point where the energy cost of active error correction becomes the dominant constraint on computation.

Part Two: The Hierarchical Way

The branching tree. Distance as shared ancestry. The strong triangle inequality. Why errors do not accumulate.

2.1 The Tree as a System of Nested Containers

Recall the branching nesting from Section 0.3. At each level, a container contains not one but $b$ alternative sub-containers. This structure is a rooted tree.

Construction. Begin with a single vertex—the root—at depth $0$. From it, draw $b$ edges outward. At the end of each edge, place a new vertex—these are the $b$ children of the root, at depth $1$. From each depth-$1$ vertex, draw $b$ edges outward to $b$ new vertices—these are the $b^2$ vertices at depth $2$. Continue forever. The result is the infinite $b$-ary tree, denoted $T_b$.

The number $b \geq 2$ is the branching factor. Each vertex (except the root) has exactly one parent and exactly $b$ children.

A vertex at depth $d$ is uniquely identified by the sequence of choices made at each branching from the root downward. Use digits from $\{0, 1, \ldots, b-1\}$ to record these choices. For example, in a binary tree ($b = 2$), the vertex reached by taking branch $0$, then branch $1$, then branch $1$ is identified by the sequence $011$. This sequence is the vertex’s address.

An infinite path from the root is an infinite sequence of digits:

These infinite sequences are the “points” of the space. Each finite vertex is a container—the set of all infinite paths that pass through it. The tree is a system of nested containers: each container at depth $d$ contains exactly $b$ sub-containers at depth $d+1$.

2.2 Shared Ancestry as Distance

How far apart are two points in this tree? There is no line between them—only branching paths. Distance must be defined in terms of the tree structure itself.

The natural measure is shared ancestry. Two points are close if their paths from the root agree for many consecutive steps—that is, if they share a long common prefix. They are far apart if they diverge early.

Definition (Common prefix length). Let $v = a_0 a_1 a_2 \ldots$ and $w = b_0 b_1 b_2 \ldots$ be two points in $T_b$. The common prefix length $k(v, w)$ is the largest integer $k$ such that $a_i = b_i$ for all $i < k$.

• If $v = w$, they agree everywhere, and $k(v, v) = \infty$.
• If they differ at the very first digit, $k(v, w) = 0$.

The common prefix length $k$ is also the depth of the most recent common ancestor—the deepest vertex from which both points descend.

Definition (Tree distance). The tree distance $d_T(v, w)$ is:

with the convention $b^{-\infty} = 0$, so $d_T(v, v) = 0$.

Why $b^{-k}$? The base of the exponent equals the branching factor. This makes the maximum distance between any two distinct points exactly $1$ (achieved when $k = 0$). It also makes the cluster structure especially clean. More fundamentally, this choice makes $T_b$ isometric to the ring of integers of a $p$-adic field for $b = p$ prime—the connection explored in Part Four.

Container interpretation. The quantity $b^{-k}$ is the size of the smallest container that encloses both points. A container at depth $k$ (a subtree rooted at depth $k$) contains $b^{-k}$ of the total space (measured in the natural ultrametric measure). Two points with common prefix length $k$ are together inside exactly those containers at depths $\leq k$, and they separate at depth $k+1$. The distance is the size of the deepest container they share.

2.3 Examples

Work through examples with a binary tree ($b = 2$).

Example 1. $v = 000\ldots$, $w = 001\ldots$.

Compare digit by digit:

• $a_0 = 0$, $b_0 = 0$ ✓
• $a_1 = 0$, $b_1 = 0$ ✓
• $a_2 = 0$, $b_2 = 1$ ✗—differ at position $2$

Common prefix length: $k = 2$.

Distance: $d_T(v, w) = 2^{-2} = 1/4 = 0.25$.

Example 2. $v = 01011\ldots$, $w = 01010\ldots$.

• Agree at positions $0, 1, 2, 3$ (digits: $0, 1, 0, 1$)
• Differ at position $4$

$k = 4$. Distance: $d_T(v, w) = 2^{-4} = 1/16 = 0.0625$.

Example 3. $v = 00000\ldots$, $w = 11111\ldots$.

• Differ at position $0$

$k = 0$. Distance: $d_T(v, w) = 2^0 = 1$. Maximum distance.

Example 4. $v = 1010101010\ldots$, $w = 1010101011\ldots$.

• Agree at positions $0$ through $8$ (nine digits)
• Differ at position $9$

$k = 9$. Distance: $d_T(v, w) = 2^{-9} = 1/512 \approx 0.00195$. Very close.

Pattern. The deeper the shared ancestry (larger $k$), the closer the points (smaller $b^{-k}$). The metric only cares about the first position of disagreement. Everything after that is irrelevant. If two points agree for the first $100$ digits, they are at distance $2^{-100}$, regardless of whether they agree or disagree on all subsequent digits.

2.4 The Strong Triangle Inequality

The metric axioms must be verified. Non-negativity and symmetry are immediate from the definition. The triangle inequality requires proof—and what emerges is stronger than the ordinary version.

Theorem (Strong triangle inequality for $T_b$). For any three points $x, y, z$ in $T_b$:

This is called the ultrametric inequality or the strong triangle inequality. Compare it with the ordinary triangle inequality $d(x, z) \leq d(x, y) + d(y, z)$. Instead of the sum, the strong version uses the maximum. Since the maximum of two non-negative numbers never exceeds their sum, the strong inequality implies the ordinary one. A metric satisfying the strong inequality is called an ultrametric.

Proof. Let $k(x, y)$ denote the common prefix length of $x$ and $y$. The key observation is:

Why is this true? Let $p = k(x, y)$ and $q = k(y, z)$. Then:

• $x$ and $y$ agree for positions $0, 1, \ldots, p-1$.
• $y$ and $z$ agree for positions $0, 1, \ldots, q-1$.

Let $r = \min(p, q)$. For any position $i < r$, we have $i < p$ (so $x_i = y_i$) and $i < q$ (so $y_i = z_i$). By transitivity of equality, $x_i = z_i$ for all $i < r$. Therefore, $x$ and $z$ share a prefix of length at least $r = \min(p, q)$.

Now convert from prefix lengths to distances. The function $t \mapsto b^{-t}$ is strictly decreasing: larger $t$ gives smaller $b^{-t}$. Therefore, taking the minimum in the exponent yields the maximum after applying $b^{-(\cdot)}$:

∎

Numerical verification with Example 1.

Let $x = 000\ldots$, $y = 001\ldots$, $z = 010\ldots$.

• $d_T(x, y) = 2^{-2} = 0.25$
• $d_T(y, z) = 2^{-1} = 0.5$
• $d_T(x, z) = 2^{-1} = 0.5$

Strong inequality: $0.5 \leq \max(0.25, 0.5) = 0.5$. ✓ Equality holds.

Ordinary inequality: $0.5 \leq 0.25 + 0.5 = 0.75$. ✓ But much weaker.

2.5 The No-Accumulation Property

The most important consequence of the strong triangle inequality is immediate:

Theorem (No-accumulation). In an ultrametric space, for any threshold $T > 0$, if $d(x, y) < T$ and $d(y, z) < T$, then $d(x, z) < T$.

Proof. $d(x, z) \leq \max(d(x, y), d(y, z)) < \max(T, T) = T$. ∎

Interpretation. You cannot accumulate small steps to cross a threshold. If each individual perturbation is below the threshold, the combined displacement is guaranteed to remain below the threshold. An error is either below the threshold or above it. There is no middle ground of gradual degradation.

This is the direct opposite of the continuous case:

The no-accumulation property is not an approximation or a statistical tendency. It is an exact theorem—a logical consequence of the ultrametric inequality, which is itself a logical consequence of the tree structure.

2.6 Geometric Consequences

The ultrametric inequality forces a remarkable rigidity on the geometry of the space.

Theorem (Isosceles triangles). In an ultrametric space, every triangle is isosceles—the two largest distances among any three points are equal.

Proof. Let $d_1 = d(x, y)$, $d_2 = d(y, z)$, $d_3 = d(x, z)$. Let $M = \max(d_1, d_2, d_3)$ be the largest distance. Suppose, for contradiction, that $M$ is unique—exactly one of the three distances equals $M$, and the other two are strictly smaller.

Without loss of generality, let $d_1 = M$ be the unique maximum: $d_1 > d_2$ and $d_1 > d_3$.

Apply the strong inequality to the triple $(x, z, y)$ (a reordering of the points):

The left side is $d_1 = M$. The right side is $\max(d_3, d_2)$. But both $d_3$ and $d_2$ are strictly less than $M$, so their maximum is also strictly less than $M$. Thus $M \leq \text{something} < M$, a contradiction.

Therefore, $M$ cannot be unique—it must appear at least twice. Among the three distances, the two largest are equal. Every triangle is isosceles. ∎

Theorem (Transitivity of proximity). In an ultrametric space, “close to” below any fixed threshold is a transitive relation. If $d(x, y) < T$ and $d(y, z) < T$, then $d(x, z) < T$.

Proof. This is exactly the no-accumulation theorem restated. ∎

This is not true in ordinary metric spaces. On the number line, $0$ is close to $0.01$, and $0.01$ is close to $0.02$, but $0$ is not particularly close to $1.00$—yet many small steps can chain together to cross any distance. In an ultrametric space, proximity below any threshold is an equivalence relation—it partitions the space into disjoint, non-overlapping clusters.

Theorem (Nested, disjoint clusters). In an ultrametric space, any two balls of the same radius are either identical or disjoint. Balls of different radii are either disjoint or one is entirely contained within the other.

Proof. Let $B_1 = B(x, r)$ and $B_2 = B(y, r)$ be two balls of the same radius $r$. Suppose they overlap: there exists $z \in B_1 \cap B_2$. Then $d(x, z) < r$ and $d(y, z) < r$. By the strong inequality, $d(x, y) \leq \max(d(x, z), d(z, y)) < r$. Now take any $w \in B_1$: $d(x, w) < r$. Then $d(y, w) \leq \max(d(y, x), d(x, w)) < r$, so $w \in B_2$. Thus $B_1 \subseteq B_2$. By symmetry, $B_2 \subseteq B_1$. Hence $B_1 = B_2$. If balls of the same radius overlap at all, they are identical.

For balls of different radii, say $r_1 < r_2$, if they overlap, the smaller ball is entirely contained in the larger one by a similar argument. ∎

Consequence: the cluster structure of $T_b$. At any threshold $T = b^{-d}$, the relation $d_T(x, y) < b^{-d}$ partitions the space into exactly $b^d$ disjoint clusters, each corresponding to a choice of the first $d$ digits. These clusters are:

• Nested: Clusters at depth $d+1$ are subsets of clusters at depth $d$.
• Disjoint at each depth: Two distinct clusters at the same depth do not overlap.
• Exhaustive: Every point belongs to exactly one cluster at each depth.

This is the geometric basis for error immunity. If logical information is encoded in the choice of cluster at depth $D$, then any perturbation that keeps the system within the same cluster (ultrametric size less than $b^{-D}$) cannot change the logical state.

2.7 Comparison: The Two Geometries

Part Three: The Threshold Principle

The physical consequence of hierarchical distance. Energy landscapes as nested containers. Why errors are exponentially suppressed.

3.1 Energy Landscapes as Nested Containers

The container metaphor has a direct physical realization. A potential energy well is a container in physical space. A particle at the bottom of a well is “inside” the container. The well’s rim is the boundary. The barrier height $\Delta E$ is the energy required to cross the boundary and leave the container.

Now imagine a landscape of nested wells. A large well (the “outer container”) contains several smaller wells (the “inner containers”), each of which contains even smaller wells, and so on. A particle at the bottom of one of the deepest wells is:

• Inside a deep well (specific)
• Inside a medium well (less specific)
• Inside a large well (even less specific)

To change which deep well the particle occupies, it must cross the rim of that deep well. To change which medium well, it must cross a higher rim. To change which large well, it must cross the highest rim of all. The landscape has a hierarchical barrier structure.

Definition (Hierarchical energy landscape). A physical system with state space $T_b$ has a hierarchical energy landscape if the energy barrier $\Delta E_k$ between states that diverge at depth $k$ satisfies:

where:

• $E_0 > 0$ is the energy scale at the shallowest level (the barrier between branches at the root).
• $\alpha > 0$ is the barrier scaling exponent—it controls how quickly barriers shrink with depth.
• $b$ is the branching factor.

At the root ($k = 0$), the barrier is $E_0$—the largest barrier in the system. At depth $k$, the barrier is $b^{-\alpha k}$ times smaller. Deep in the hierarchy (large $k$), barriers are tiny; the system can move freely among states at that depth. Shallow in the hierarchy (small $k$), barriers are large; the system is strongly confined.

This is exactly the pebble-in-a-depression scenario, generalized to arbitrarily many levels. The shallowest levels correspond to the largest, highest-rimmed containers. The deepest levels correspond to the smallest, lowest-rimmed containers.

3.2 Two Kinds of Memory

The distinction between continuous and hierarchical distance translates directly into a distinction between two kinds of physical memory.

Type A: Continuous memory. The stored value is a point on a continuous line—the height of a water column, the voltage on a capacitor, the angle of a needle. Perturbations displace the point continuously. To maintain accuracy, constant measurement and correction are required. Energy cost scales with precision. Error accumulates with time.

Type B: Hierarchical memory. The stored value is the identity of a container at depth $D$ in a hierarchical energy landscape. Perturbations with energy below the rim height $\Delta E_D$ cause the system to jiggle within the container but cannot move it to a different container at the same depth. Only perturbations exceeding $\Delta E_D$ can change the logical state—and these are exponentially rare when $\varepsilon \ll \Delta E_D$. Error does not accumulate with time (except through the rare, discrete error events). Energy cost is near-zero during idle—the barriers do the work of protection.

3.3 Formal Statement

The Threshold Principle. Let $\mathcal{S}$ be a physical system whose stable states form a hierarchical energy landscape with barrier scaling $\Delta E_k = E_0 \cdot b^{-\alpha k}$.

Let the logical information be encoded by the identity of the cluster at depth $D$—that is, the first $D$ digits of the system’s state determine the logical value.

Let the environmental noise have characteristic energy $\varepsilon$. (For thermal noise at temperature $T$, $\varepsilon = k_B T$.)

Then, provided $\varepsilon < \Delta E_D$, the probability of a logical error per unit time satisfies:

where:

• $C$ is a prefactor with units of frequency (typically $10^9$–$10^{13}$ Hz for atomic-scale systems—the attempt frequency for barrier crossing).
• $\beta$ is a dimensionless exponent:

- $\beta = 1$ for classical thermal activation (Arrhenius law).

- $\beta = 2$ for quantum tunneling through a parabolic-like barrier (WKB approximation).

- In general, $1 \leq \beta \leq 2$ depending on barrier shape and temperature.

The key property: the error rate is exponentially suppressed in the ratio $\Delta E_D / \varepsilon$. Double the ratio, and the error rate is squared (for $\beta = 1$). Increase the encoding depth $D$, and $\Delta E_D = E_0 \cdot b^{-\alpha D}$ decreases, so $\Delta E_D / \varepsilon$ increases—but wait: increasing $D$ actually decreases $\Delta E_D$ (barriers shrink with depth), which increases the error rate. The protection improves when $D$ is smaller (shallower encoding) or when $E_0 / \varepsilon$ is larger. This is a crucial design constraint.

3.4 Why the Suppression is Exponential

Why does barrier crossing obey an exponential law? The answer is rooted in the statistics of thermal fluctuations.

Consider a classical particle in a potential well of depth $\Delta E$, in contact with a heat bath at temperature $T$. The particle’s energy fluctuates as it exchanges energy with the bath. The probability that a fluctuation gives the particle energy at least $\Delta E$ is proportional to the Boltzmann factor:

This follows from the fact that in thermal equilibrium, the probability of a state with energy $E$ is proportional to $e^{-E / k_B T}$ (the Boltzmann distribution). The probability of being in any state with energy $\geq \Delta E$ is the sum (integral) of $e^{-E / k_B T}$ over all $E \geq \Delta E$, which is dominated by the lowest such energy and thus scales as $e^{-\Delta E / k_B T}$.

The escape rate is this probability multiplied by an attempt frequency—roughly, how often the particle “tries” to cross the barrier. For atomic-scale systems, attempt frequencies are on the order of vibrational frequencies: $10^{12}$–$10^{14}$ Hz. This gives the Arrhenius rate:

For quantum systems, tunneling provides an alternative escape path that can dominate at low temperatures. The WKB approximation gives a tunneling rate:

For a parabolic-like barrier, the integral scales as $\sqrt{\Delta E}$, giving $\log \Gamma \propto -\sqrt{\Delta E}$, which is $\beta = 1/2$. However, for many realistic barrier shapes (cubic, quartic) and in the presence of dissipation, the effective exponent can be closer to $\beta = 1$ or $\beta = 2$. The framework accommodates any $\beta > 0$; the key is the exponential dependence on $\Delta E / \varepsilon$.

The crucial difference from the continuous case. In a continuous potential, there is no barrier—the potential is flat, or nearly flat on the scale of the noise. The particle drifts freely. The displacement after time $t$ grows as $\sqrt{t}$ (diffusive) or $t$ (ballistic). In a hierarchical landscape, the particle is trapped in a well. It jiggles but stays. The difference is qualitative, not quantitative. It is the difference between a floor and a bowl.

3.5 Engineering Implications

1. No active correction during idle. Between operations, the system protects itself. The energy barriers do the work of error suppression. No measurements, no feedback, no energy expenditure.

2. Only rare global resets. When a logical error does occur (an exponentially rare event), it is detected by a parity check at the encoding depth $D$. A global reset—applying a pulse larger than $\Delta E_D$ to return the system to the intended cluster—corrects it. The reset rate is negligible compared to the system’s operational clock rate.

3. Logarithmic resource scaling. To achieve a logical error rate $p$, the required depth satisfies:

The encoding depth—and therefore the logarithm of the number of physical states—scales logarithmically with the required error suppression $\log(1/p)$. This is exponentially better than the polynomial scaling of conventional quantum error correction.

4. Practical requirements.

• Low temperature: $\varepsilon = k_B T$ must be less than $\Delta E_D$. This is achievable for reasonable $D$ if $E_0 \gg k_B T$.
• Physical realizability: The system must have a genuine hierarchical energy landscape. This is the central engineering challenge.
• Operational access: The system must support write, read, and logic operations without destroying the hierarchical protection. Operations should couple states within the same cluster (low barriers, fast) while avoiding coupling across cluster boundaries (high barriers, to be avoided).

5. Design trade-off. Shallower encoding (smaller $D$) gives better protection (larger $\Delta E_D$) but fewer distinct logical states per physical system ($b^D$ clusters). Deeper encoding gives more logical states but weaker protection at each level. The optimal $D$ balances protection against capacity.

Part Four: The Deep Structure

The mathematical unity behind the two ways of measuring. p-adic numbers. Ostrowski’s theorem. Why there are only two ways.

4.1 P-adic Numbers: The Hierarchical Way Made Algebraic

The tree distance $d_T(v, w) = b^{-k}$ on $T_b$ may seem like an arbitrary construction. It is not. It is the natural distance on the $p$-adic numbers, one of the most important structures in number theory.

Let $p$ be a prime number (e.g., $p = 2, 3, 5, 7, \ldots$). Any non-zero rational number $x$ can be uniquely factored as:

where:

• $k$ is an integer (positive, negative, or zero).
• $a$ and $b$ are integers not divisible by $p$.
• $\frac{a}{b}$ is in lowest terms.

The integer $k$ is the $p$-adic order of $x$, denoted $\text{ord}_p(x)$. It measures “how divisible by $p$” the number $x$ is:

• $\text{ord}_p(8) = 3$ because $8 = 2^3$ (for $p = 2$).
• $\text{ord}_p(12) = 1$ because $12 = 2^2 \cdot 3$, so $\text{ord}_2(12) = 2$ but $\text{ord}_3(12) = 1$.
• $\text{ord}_p(1/2) = -1$ (for $p = 2$) because $1/2 = 2^{-1}$.
• $\text{ord}_p(0) = \infty$ by convention.

Definition (p-adic absolute value). The $p$-adic absolute value of a rational number $x$ is:

with $|0|_p = 0$.

This definition is identical in form to the tree distance. Replace $b$ with $p$, and replace the common prefix length $k$ of two tree points with the $p$-adic order of their difference:

Two rational numbers are $p$-adically close if their difference is divisible by a large power of $p$. In base-$p$ expansion, this means their expansions agree for many digits.

Example. Let $p = 2$. Consider $x = 0$ and $y = 2$. Their difference is $2 = 2^1$. $\text{ord}_2(2) = 1$, so $|2|_2 = 2^{-1} = 1/2$. In binary, $0 = \ldots 000_2$ and $2 = \ldots 010_2$—their $2$-adic expansions differ at the first digit (position $0$). Common prefix length $1$, distance $2^{-1}$. Matches the tree distance exactly.

The $p$-adic numbers $\mathbb{Q}_p$ are obtained by completing $\mathbb{Q}$ under $|\cdot|_p$, just as $\mathbb{R}$ is obtained by completing $\mathbb{Q}$ under the ordinary absolute value $|\cdot|_\infty$.

Key properties of $|\cdot|_p$:

1. $|x|_p \geq 0$, with equality only for $x = 0$. ✓

1. $|x y|_p = |x|_p |y|_p$. ✓ (multiplicativity)

1. $|x + y|_p \leq \max(|x|_p, |y|_p)$. ✓ (strong triangle inequality!)

Property 3 is the ultrametric inequality. The $p$-adic numbers form an ultrametric field—a field equipped with an absolute value satisfying the strong triangle inequality. This is the algebraic counterpart of the tree $T_p$.

4.2 Ostrowski’s Theorem: Why There Are Only Two Ways

Ostrowski’s Theorem (1916). Every non-trivial absolute value on the rational numbers $\mathbb{Q}$ is equivalent either to:

• The ordinary (Archimedean) absolute value $|\cdot|_\infty$, or
• A $p$-adic absolute value $|\cdot|_p$ for some prime $p$.

Two absolute values are equivalent if they define the same notion of convergence—the same topology. Equivalent absolute values differ by a power: $|\cdot|_1 = |\cdot|_2^c$ for some $c > 0$.

What this means. Any coherent notion of distance on the rational numbers that respects the algebraic operations of addition and multiplication must fall into one of two families:

1. The Archimedean family: distances where the triangle inequality uses addition ($|x + y| \leq |x| + |y|$). This gives the continuous number line $\mathbb{R}$.

1. The non-Archimedean (ultrametric) family: distances where the triangle inequality uses the maximum ($|x + y| \leq \max(|x|, |y|)$). This gives the $p$-adic numbers $\mathbb{Q}_p$, which are hierarchical spaces with branching factor $p$.

There is no third way. The two ways of measuring are forced by the structure of the number system itself.

The tree distance $d_T$ on $T_b$ for $b = p$ (a prime) is exactly the $p$-adic distance on the $p$-adic integers. For composite $b$, the distance is related to a product of $p$-adic distances. In either case, the hierarchical way of measuring is not an ad hoc construction—it is one half of the complete classification of distances on $\mathbb{Q}$.

The continuous way (Archimedean) and the hierarchical way (ultrametric) are dual in the sense of Ostrowski: they are the two complementary completions of the same underlying rational structure.

4.3 The P-adic Interpretation of the Threshold Principle

The threshold principle has a clean $p$-adic reformulation.

In $T_p$, the ball of radius $p^{-D}$ centered at a point $x$ consists of all points whose first $D$ digits agree with those of $x$. This ball is the encoding cluster—the set of states that share the same logical value.

A perturbation of $p$-adic size $\delta = |\text{perturbation}|_p$ moves the system within a ball of radius $\delta$. If $\delta < p^{-D}$, the perturbation cannot move the system out of the encoding cluster. The logical state is unchanged.

If $\delta \geq p^{-D}$, the perturbation can change the logical state—but only if the perturbation pushes the system across the cluster boundary, which requires crossing an energy barrier of height at least $\Delta E_D$.

The physical energy $\Delta E_D$ and the $p$-adic distance $p^{-D}$ are related:

The energy barrier is proportional to a power of the $p$-adic cluster radius. Larger clusters (smaller $D$) have higher barriers. Smaller clusters (larger $D$) have lower barriers.

The threshold principle in $p$-adic language:

> Noise with $p$-adic size less than $p^{-D}$ is irrelevant to logical information encoded at depth $D$. The probability of noise exceeding $p^{-D}$ is exponentially suppressed.

4.4 The Convergence of Independent Lines

Independent lines of evidence, from unrelated fields, converge on the same conclusion:

Line 1: Metric space theory. The strong triangle inequality $d(x, z) \leq \max(d(x, y), d(y, z))$ is logically distinct from the ordinary triangle inequality. It implies radically different geometric properties: no error accumulation, isosceles triangles, nested disjoint clusters. Spaces satisfying it (ultrametric spaces) form a distinct category of metric spaces.

Line 2: Number theory (Ostrowski’s theorem). The only absolute values on $\mathbb{Q}$ are Archimedean ($|\cdot|_\infty$) and $p$-adic ($|\cdot|_p$). The $p$-adic absolute values are ultrametric. Thus, ultrametric geometry is forced by the structure of the rational numbers.

Line 3: Statistical mechanics. Barrier crossing rates in thermal equilibrium obey the Arrhenius law: rate $\propto e^{-\Delta E / k_B T}$. When barriers are hierarchical ($\Delta E_k = E_0 \cdot b^{-\alpha k}$), error rates are exponentially suppressed in the encoding depth. The energy landscape must have the hierarchical-nesting structure for this to work—exactly the structure of an ultrametric space.

Line 4: Information theory. The resource cost of error correction scales with the required error suppression. In Archimedean spaces (active correction), the cost is polynomial in the inverse error rate. In ultrametric spaces (passive protection), the cost is logarithmic. The geometry of the state space directly determines the information-theoretic efficiency of error suppression.

All four lines converge on a single unified statement:

> To build a memory that endures without active correction, engineer the state space to be ultrametric.

Part Five: Concrete Architectures

Design principles, four candidate physical substrates, encoding schemes, and a comparison with conventional quantum error correction.

5.1 The Hierarchical Memory Cell: Design Specification

A hierarchical memory cell is a physical system with the following properties:

Required operations:

1. Initialize: Drive the system to a specific state within the desired logical cluster at depth $D$.

1. Hold: Let the system idle. Barriers protect the logical state passively.

1. Read: Measure the cluster identity at depth $D$ without disturbing deeper levels. This is a quantum non-demolition (QND) measurement if the system is quantum.

1. Write: Apply an energetic pulse $> \Delta E_D$ to move the system to a different logical cluster.

1. Reset: Upon detecting a logical error (via parity check at depth $D$), apply a correction pulse.

Design trade-off. Deeper encoding (larger $D$) gives more logical states per physical system but weaker per-level protection. Shallower encoding (smaller $D$) gives stronger protection but fewer logical states. The optimal $D$ balances protection against capacity.

5.2 Physical Substrate 1: Spin Chains with Exponential Coupling Gradient

Physical system. A one-dimensional chain of $N$ spin-1/2 particles with site-dependent Ising couplings.

Hamiltonian:

where:

• $\sigma_i^z$ and $\sigma_i^x$ are Pauli matrices acting on spin $i$.
• $J_i = J_0 \cdot b^{-\alpha i}$: coupling strength decays exponentially along the chain.
• $h$ is a uniform transverse field (provides quantum fluctuations/tunneling).

Hierarchical structure. The ferromagnetic ($\uparrow\uparrow\uparrow\ldots$) and antiferromagnetic ($\uparrow\downarrow\uparrow\ldots$) domains form the tree structure. A domain wall at position $k+1$ separates regions that agree for the first $k$ spins. Moving the domain wall changes the “prefix” of the state. The energy cost to move the domain wall past spin $k$ is proportional to $J_k$, which decays as $b^{-\alpha k}$.

Encoding. The logical bit is determined by the orientation of the first spin (or the first $D$ spins). Deeper spins form the hierarchy that provides protection. A perturbation that flips spin $N$ (deep) costs energy $\sim J_{N-1} \sim b^{-\alpha(N-1)}$—very small. A perturbation that flips spin $1$ (shallow) costs energy $\sim J_0$—very large.

Estimated parameters for a 10-spin chain:

• $b = 2$, $\alpha = 1$, $J_0 / k_B = 10$ K (so $J_0 \approx 1.38 \times 10^{-22}$ J).
• Operating temperature: $T = 0.1$ K (dilution refrigerator).
• $\varepsilon = k_B T = 1.38 \times 10^{-24}$ J.
• $J_9 = J_0 \cdot 2^{-9} \approx 0.02$ K $\approx 2.7 \times 10^{-25}$ J—deep spins are weakly protected.
• $J_0 / (k_B T) = 100$—root barrier is 100× thermal energy.
• Encoding depth $D = 1$: $\Delta E_1 = J_0 \cdot 2^{-1} = 5$ K $\approx 6.9 \times 10^{-23}$ J. Ratio $\Delta E_1 / k_B T = 50$. Error rate (Arrhenius, $\beta = 1$, $\nu_0 = 10^{12}$ Hz): $P_{\text{err}} \sim 10^{12} \cdot e^{-50} \approx 10^{-10}$ per second. Mean time between errors: $\sim 300$ years.

Advantages. Spin chains can be realized in multiple platforms: NMR, trapped ions, superconducting qubits, quantum dots. The coupling gradient can be engineered through physical spacing or tunable couplers. The transverse field provides quantum tunneling ($\beta \approx 2$), further improving error suppression.

Challenges. Long chains suffer from decoherence at the weak-coupling end. Engineering precise exponential coupling gradients requires calibration. Readout of deep spins without disturbing shallow spins is non-trivial.

5.3 Physical Substrate 2: Molecular Conformational Memory

Physical system. An organic molecule with multiple rotatable bonds, each with a torsional energy barrier at a different scale. The set of all conformational states forms a tree: each bond’s orientation (e.g., cis vs. trans, or gauche+ vs. gauche−) is one digit in the address.

Example molecule. Consider a biphenyl derivative with substituents that create a hierarchy of rotational barriers:

• Bond 1 (shallowest): central C–C bond in biphenyl. Barrier $\sim 2$ kcal/mol ($\approx 1000$ K in energy units)—very high, nearly frozen at room temperature.
• Bond 2: rotation of a methyl group. Barrier $\sim 3$ kcal/mol ($\approx 1500$ K)—also high.
• Bond 3: rotation of a hydroxyl group. Barrier $\sim 1$ kcal/mol ($\approx 500$ K).
• Bond 4: rotation of a longer alkyl chain segment. Barrier $\sim 0.5$ kcal/mol ($\approx 250$ K)—nearly free at room temperature.

The barriers form a hierarchy: $E_0 \approx 2000$ K, $b = 2$, $\alpha \approx 1$ (if each level halves the barrier).

Encoding. The logical bit is encoded in the conformation of the highest-barrier bond(s). Thermal fluctuations can rotate lower-barrier bonds (causing local jitter) but cannot rotate the highest-barrier bonds—the logical state is protected.

Readout. Nuclear Magnetic Resonance (NMR) spectroscopy. Different conformations produce different chemical shifts and J-coupling patterns. The NMR spectrum reveals which conformational state the molecule occupies. Modern NMR can resolve conformational states at millikelvin effective temperatures through hyperpolarization techniques.

Advantages. Chemistry naturally provides hierarchical energy landscapes. Molecules can be designed and synthesized with tailored barrier hierarchies. NMR provides non-destructive, high-resolution readout.

Challenges. The hierarchy depth is limited by the number of independently rotatable bonds. Solvent interactions can reduce effective barriers. Quantum coherence across conformational states is difficult to maintain at room temperature.

5.4 Physical Substrate 3: Superconducting Hierarchical Qubits

Physical system. A superconducting flux qubit with an engineered multi-well potential. The potential is shaped by an array of Josephson junctions with carefully chosen critical currents.

Potential engineering. The potential energy as a function of the superconducting phase $\varphi$ can be designed through Fourier synthesis:

where:

• $E_J$ is the Josephson energy scale (typically $10$–$100$ GHz $\times h$, or $0.5$–$5$ K in temperature units).
• $c_m = b^{-\alpha m}$ creates the exponential decay of barrier heights.
• $b^m \varphi$ creates $b$ smaller wells within each larger well.
• $M$ is the number of hierarchical levels.

Physical realization. Each $\cos(b^m \varphi)$ term can be approximated by a loop containing $b^m$ Josephson junctions in parallel. The total potential is the sum of contributions from loops of different sizes. Modern superconducting circuit fabrication can achieve junction critical current uniformity of $\sim 1\%$, which is sufficient for $M \approx 3$–$5$ levels.

Estimated parameters:

• $E_J / h = 50$ GHz ($\approx 2.4$ K).
• Operating temperature: $T = 10$ mK (standard for superconducting qubits).
• $k_B T / h \approx 200$ MHz—very small compared to $E_J / h$.
• With $b = 2$, $\alpha = 1$, $M = 4$, the deepest barrier is $E_J \cdot 2^{-4} \approx 3.1$ GHz $\times h$ $\approx 0.15$ K $\gg 10$ mK.
• All levels are well protected at 10 mK.

Advantages. Superconducting circuits are the most mature quantum computing platform. Precise potential engineering is feasible with current fabrication. Measurement and control infrastructure exists. Millikelvin temperatures are standard.

Challenges. Decoherence from dielectric losses, quasiparticle poisoning, and flux noise limits coherence times. The hierarchy depth is limited by fabrication precision. Engineering potentials with many nested wells while maintaining coherence is an open experimental challenge.

5.5 Physical Substrate 4: Optical Nested-Cavity Memory

Physical system. A series of nested Fabry-Perot optical cavities. A photon is confined in the innermost cavity. To escape, it must tunnel through a series of partially reflective mirrors—each corresponding to a barrier in the hierarchy.

Hierarchical structure.

• Cavity 0 (outermost): large, low finesse, wide linewidth. Easy for a photon to enter or leave.
• Cavity 1: smaller, higher finesse, narrower linewidth.
• ...
• Cavity $M$ (innermost): smallest, highest finesse, narrowest linewidth. The photon is most strongly confined here.

The photon’s “state” is which cavity it occupies, and within that, its polarization or orbital angular momentum.

Encoding. The logical bit is encoded in the photon’s polarization (horizontal = logical 0, vertical = logical 1) while the cavity nesting provides the hierarchical protection. To change the logical bit, the photon must escape its cavity and enter a different branch, requiring it to tunnel through multiple mirrors—a process suppressed by the product of the mirror transmissivities.

Estimated parameters:

• Mirror reflectivities: $R_1 = 0.99$, $R_2 = 0.999$, $R_3 = 0.9999$, $R_4 = 0.99999$ (achievable with modern super-polished mirrors).
• Cavity finesse scales inversely with transmissivity.
• Photon lifetime in the innermost cavity: milliseconds to seconds.
• Protection factor: the product of all mirror reflectivities.

Advantages. Optical cavities have extremely high quality factors. Nested cavities are a mature technology (used in gravitational wave detectors). Photons are naturally robust to many noise sources.

Challenges. Alignment and mechanical stability of nested cavities is demanding. Photon loss is irreversible (unlike spin or charge states, which can relax back to their well). Single-photon detection with high efficiency is required.

5.6 Encoding, Readout, and Reset

Direct encoding. The logical bit is the first digit of the state. For $b = 2$, states with first digit $0$ encode logical $0$; states with first digit $1$ encode logical $1$. The encoding depth $D$ determines the number of physical states within each logical cluster: $b^{D-1}$ states per logical value.

Redundant encoding. Multiple digits at the encoding depth can redundantly encode the same logical bit. For example, at $D = 2$ in a binary tree:

• Logical $0$: states $00\ldots$ and $01\ldots$ (first digit is $0$)
• Logical $1$: states $10\ldots$ and $11\ldots$ (first digit is $1$)

This adds protection: an error at depth $2$ (switching between $00\ldots$ and $01\ldots$) does not change the logical bit, because both are logical $0$. Only an error at depth $1$ (the highest barrier) can flip the logical bit.

Parity-check error detection. At each depth $k < D$, compute a parity from the subtree structure. A change in parity indicates an error that crossed the depth-$k$ barrier. By monitoring parities at multiple depths, errors can be detected and localized. The key difference from conventional QEC: these checks are performed rarely, because errors are exponentially suppressed.

Reset protocol. When a parity check detects a logical error:

1. Identify the depth at which the error occurred.

1. Apply a correction pulse of energy $> \Delta E_k$ to return the system to the correct cluster.

1. The correction rate is negligible compared to the operational rate, because errors are rare.

5.7 Comparison: Hierarchical Encoding vs. Conventional QEC

Part Six: Experimental Protocols

How to test the framework. Verifying ultrametricity, measuring threshold depth, demonstrating no-accumulation, and concrete proposals for first experiments.

6.1 Protocol 1: Verification of Ultrametricity

Prediction. In a hierarchical system, for any three states $A, B, C$, the two largest pairwise distances are equal. Equivalently, all triangles are isosceles.

Rationale. The ultrametric inequality $d(A, C) \leq \max(d(A, B), d(B, C))$ forces this property (proved in Section 2.6). A non-hierarchical system (e.g., a continuous potential) does not satisfy this constraint. Testing the isosceles triangle property is therefore a direct test of ultrametricity.

Procedure:

1. Select three states $A, B, C$ in the candidate hierarchical system. Choose states at different depths in the hierarchy to maximize the expected differences in pairwise distances.

1. Measure the transition rate $r(A \to B)$ from $A$ to $B$. For a thermally activated system, the rate depends on the barrier height: $r \propto e^{-\Delta E_{k(A,B)} / k_B T}$. Thus, larger distance (deeper shared ancestry) corresponds to lower transition rate.

1. Repeat for all three pairs: $r(A, B)$, $r(B, C)$, $r(A, C)$. Compute effective distances: $d_{ij} \propto -\log(r_{ij})$ (the proportionality constant cancels in comparisons).

1. Order the three distances: $d_{(1)} \leq d_{(2)} \leq d_{(3)}$. Compute $\Delta = d_{(3)} - d_{(2)}$.

1. If $\Delta$ is consistent with zero (within measurement uncertainty), the ultrametric prediction is supported. If $\Delta$ is significantly positive, the system is not ultrametric.

Statistical analysis. For each pair $(i, j)$, perform $n$ independent measurements of the transition rate. Compute the mean rate $\bar{r}_{ij}$ and its standard error $\sigma_{ij}$. The effective distance is $d_{ij} = -\log(\bar{r}_{ij})$ with error propagated from $\sigma_{ij}$. Construct a 95% confidence interval for $\Delta$. If the interval includes zero, the data are consistent with ultrametricity.

Example with $n = 100$ measurements per pair. If $\bar{r}_{AB} = 0.01$ s$^{-1}$, $\bar{r}_{BC} = 0.10$ s$^{-1}$, $\bar{r}_{AC} = 0.01$ s$^{-1}$, then $d_{AB} \approx 4.6$, $d_{BC} \approx 2.3$, $d_{AC} \approx 4.6$. The two largest ($d_{AB}$ and $d_{AC}$) are equal—consistent with ultrametricity. If instead $\bar{r}_{AC} = 0.05$ s$^{-1}$ ($d_{AC} \approx 3.0$), then the three distances are all different—inconsistent with ultrametricity.

6.2 Protocol 2: Measurement of Threshold Depth and Error Suppression

Prediction. The logical error rate depends exponentially on the ratio $\Delta E_D / \varepsilon$:

Procedure (temperature variation):

1. Fix the encoding depth $D$.

1. Vary the temperature $T$ (thus varying $\varepsilon = k_B T$) over a range where $\varepsilon$ spans from well below $\Delta E_D$ to comparable to $\Delta E_D$.

1. At each temperature, initialize the system to a known logical state, let it evolve for a fixed time $\tau$, and measure the logical state. Repeat $N_{\text{trials}}$ times to estimate $P(\text{error})$.

1. Plot $\log P(\text{error})$ against $1/T$ (for $\beta = 1$, thermal activation) or against $1/T^{\beta}$ (for other $\beta$).

1. Fit a line to the low-temperature (large $1/T$) region. The slope gives $-\Delta E_D / k_B$ (for $\beta = 1$). Extract $\Delta E_D$.

1. Repeat for different encoding depths $D$. Verify that $\Delta E_D$ decreases with $D$ as $\Delta E_D = E_0 \cdot b^{-\alpha D}$.

Expected plot. The plot of $\log P(\text{error})$ vs. $1/T$ should show:

• Low-temperature regime (large $1/T$): Linear with negative slope. The error rate is exponentially suppressed. The slope magnitude is $\Delta E_D / k_B$.
• High-temperature regime (small $1/T$): Saturation. The barrier is irrelevant; errors occur frequently. The error rate approaches the attempt frequency $C$.
• Transition: A bend where $\varepsilon \approx \Delta E_D$.

Procedure (encoding depth variation):

1. Fix the temperature $T$.

1. Vary the encoding depth $D$ (if the system allows tuning which level encodes the logical bit).

1. Measure $P(\text{error})$ as a function of $D$.

1. Plot $\log P(\text{error})$ against $D$. The slope should be $\alpha \beta \log b$.

6.3 Protocol 3: Demonstration of No-Accumulation

Prediction. In a hierarchical system, $N$ perturbations of strength below the threshold produce a total error bounded by the maximum individual perturbation. In a continuous system, the errors sum.

Procedure (comparative):

1. System H (Hierarchical): A molecular conformational memory with a known barrier hierarchy.

1. System C (Continuous): A classical analog memory—e.g., a low-leakage capacitor storing a voltage, with a known noise floor.

For System H:

• Initialize the system to a known logical state at depth $D$.
• Apply a sequence of $N$ perturbations, each with energy $\varepsilon_{\text{applied}} < \Delta E_D$.
• After every $m$ perturbations (e.g., $m = 100$), measure the logical state.
• Record the cumulative number of logical errors as a function of $N$.

For System C:

• Initialize the voltage to a target $V_0$.
• Apply $N$ perturbations of fixed magnitude $\delta V$.
• After every $m$ perturbations, measure the voltage and compute the error $|V - V_0|$.

Expected results.

• System H: The logical error count should grow linearly with $N$ but with a very small slope $p = P(\text{error per perturbation}) \ll 1$. The total probability of at least one error after $N$ perturbations is approximately $1 - (1-p)^N \approx Np$ for $Np \ll 1$. Critically, this is not accumulation—each perturbation has an independent, small probability of causing an error, but the errors do not build on each other. After an error occurs, the system does not continue to drift further from the target—it is simply in a different valid state.
• System C: The RMS voltage error should grow as $\delta V \sqrt{N}$. After many perturbations, the voltage can be arbitrarily far from the target.

The qualitative signature of hierarchical protection: A flat error rate (or linearly growing error count with very small slope) vs. a growing error magnitude. In the hierarchical system, there is either an error (rare) or there isn’t. There is no “partial” error that grows over time.

6.4 Proposed First Experiment: NMR Molecular Hierarchical Memory

Platform. Liquid-state Nuclear Magnetic Resonance (NMR) on a small fluorinated organic molecule.

Molecule. 1,2,3,4-tetrafluorobenzene ($\text{C}_6\text{H}_2\text{F}_4$). This molecule has four fluorine-19 nuclei coupled through $J$-couplings (scalar couplings) that span a natural hierarchy:

The coupling hierarchy is $J_{12} \approx 2.5 \times J_{13} \approx 10 \times J_{14}$, roughly following an exponential decay.

Encoding scheme.

• The logical bit is encoded in the state of the most strongly coupled spin pair (F1–F2). The two states of this pair (aligned vs. anti-aligned in the $z$-basis) represent logical $0$ and $1$.
• The weaker couplings (F3, F4) form the deep hierarchy. Their states can fluctuate with lower energy cost, absorbing thermal noise without affecting the logical bit.

Experimental protocol:

1. Sample preparation. Dissolve 1,2,3,4-tetrafluorobenzene in a deuterated solvent (e.g., acetone-d6) at $\sim 100$ mM concentration. Degas to remove dissolved oxygen (which causes paramagnetic relaxation).

1. Initialization. Use RF pulses to prepare the spin system in a specific state—e.g., all four $^{19}$F spins polarized along the $+z$ direction (logical $0$). This is a non-equilibrium state; its preparation fidelity depends on the polarization technique.

1. Noise application. Apply a train of weak, off-resonant RF pulses (or continuous low-power irradiation) to simulate environmental noise. Vary the noise power $P_{\text{noise}}$ (proportional to the square of the RF field amplitude).

1. Evolution. Let the system evolve under the noise for a fixed duration $\tau$ (e.g., $\tau = 1$ second).

1. Readout. Apply a readout pulse sequence that measures the state of the F1–F2 spin pair using a spin-echo or gradient-echo sequence. Record whether the logical bit has flipped.

1. Repetition. Repeat steps 2–5 for $N_{\text{trials}} \approx 10^4$ per noise power setting to accumulate statistics.

1. Analysis. Plot logical error probability vs. noise power. Fit to $P(\text{error}) = P_0 + A \cdot \exp(-(P_{\text{thresh}} / P_{\text{noise}})^\beta)$, where $P_0$ is the baseline error rate (preparation/readout infidelity) and $P_{\text{thresh}}$ is the effective threshold noise power.

Expected signature of hierarchical protection. At low noise powers, the error rate should be flat (dominated by $P_0$). At a critical noise power $P_{\text{crit}}$, the error rate should begin to rise—this is the threshold where the noise energy becomes comparable to the effective barrier $\Delta E_D$ protecting the logical encoding. The rise should be steep if $\beta$ is large.

Equipment. Standard NMR spectrometer (400–600 MHz $^1$H frequency, with $^{19}$F capability), temperature control ($\pm 0.1$ K), pulse programmer. This experiment is feasible in a well-equipped university chemistry or physics department.

Estimated signal-to-noise. At $100$ mM concentration in a $0.5$ mL sample, there are $\sim 3 \times 10^{19}$ molecules. With thermal polarization at $300$ K and $9.4$ T (400 MHz), the polarization is $\sim 3 \times 10^{-5}$, giving $\sim 10^{15}$ polarized spins—more than enough for ensemble measurements.

6.5 Proposed Experiment: Trapped Ion Hierarchical Qubit

Platform. A linear chain of $N$ trapped ions (e.g., $^{171}\text{Yb}^+$ or $^{40}\text{Ca}^+$) in a Paul trap.

Hierarchical structure via motional modes. The $N$ ions have $3N$ motional modes (vibrations of the ion chain). The mode frequencies span a range: the center-of-mass mode (all ions moving together) has the highest frequency, and the higher-order “zig-zag” and “buckling” modes have progressively lower frequencies. The mode frequency hierarchy can be tuned by adjusting the trap voltages.

Encoding. The logical bit is encoded in the phonon number (Fock state) of the highest-frequency motional mode. Lower-frequency modes serve as the deep hierarchy. A perturbation that adds phonons to a low-frequency mode (small energy) does not affect the logical encoding. A perturbation that adds phonons to the high-frequency mode requires more energy and constitutes a logical error.

Protocol:

1. Initialization. Cool all motional modes to the ground state ($n = 0$) via sideband cooling. Prepare the highest-frequency mode in $n = 0$ (logical $0$) or $n = 1$ (logical $1$).

1. Noise application. Apply controlled electric field noise to the trap electrodes. The noise spectrum can be shaped to have power primarily at specific frequencies. Vary the noise power spectral density at the frequency of the logical mode.

1. Evolution. Let the system evolve under the noise for time $\tau$.

1. Readout. Use the ion’s internal electronic state as an ancilla: map the phonon number of the logical mode onto the electronic state via a red sideband pulse, then read out the electronic state via state-dependent fluorescence.

1. Analysis. Plot logical error probability vs. noise power. Extract $\Delta E_D$ from the exponential fit.

Advantages over NMR. Trapped ions offer:

• Individual addressing of each motional mode.
• Near-unit fidelity state preparation and readout.
• Tunable mode hierarchy via trap voltages.
• Long coherence times (seconds) for motional states.
• The ability to resolve single quanta (phonons)—true quantum-level precision.

Challenges. Anharmonicities in the trapping potential can mix motional modes. Heating of motional modes from electrode noise is a well-known challenge in ion traps—the same mechanism that would be studied, which makes the experiment both a test of the framework and a probe of the noise environment.

Part Seven: Implications and Open Problems

7.1 For Quantum Computing

The hierarchical framework offers a potential path around the thermodynamic wall of quantum error correction. If hierarchical energy landscapes can be engineered with sufficient depth and fidelity, the resource overhead for fault-tolerant quantum computation could be reduced from polynomial to logarithmic in the inverse error rate.

For this to become practical, several conditions must be met:

1. Engineering depth. The physical substrate must support at least $D \approx 5$–$10$ hierarchical levels with well-controlled barrier scaling. At $D = 10$, $b = 2$, $\alpha = 1$, and $E_0 / k_B T = 100$, the logical error rate is $\sim 10^{-30}$ per second—effectively zero for any practical computation. But fabricating a 10-level hierarchy with precise barriers is non-trivial.

1. Coherent operations. Quantum gates must act on the logical encoding without coupling across high barriers. This requires gates that are selective in the hierarchical depth—coupling states within the same cluster (fast, low-barrier) while avoiding coupling across cluster boundaries (to be avoided, high-barrier). This is a new kind of gate design constraint.

1. Scalability. A single hierarchical memory cell stores one logical bit. To build a quantum computer, many such cells must be entangled. Whether hierarchical cells can be coupled without destroying their individual protection is an open question.

1. Error model validity. The threshold principle assumes that noise is characterized by a single energy scale $\varepsilon$ (e.g., thermal). Real noise has a spectrum—some noise is high-frequency (potentially resonant with shallow barriers), some is low-frequency (causing slow drift). The framework must be extended to handle structured noise.

Comparison with surface codes. Surface codes are the leading approach to fault-tolerant quantum computing. They require a 2D array of physical qubits with nearest-neighbor interactions, and they achieve a logical error rate scaling as $(p/p_{\text{th}})^{d/2}$ where $p$ is the physical error rate, $p_{\text{th}} \approx 0.01$ is the threshold, and $d$ is the code distance. For $p = 10^{-3}$ and $d = 20$, the logical error rate is $\sim 10^{-15}$, using $d^2 = 400$ physical qubits per logical qubit.

Hierarchical encoding, if realizable with $D = 5$ and $b = 2$ (32 physical states per logical bit) at $E_0 / k_B T \approx 50$, could achieve similar or better error rates with far fewer physical resources—but only if the energy landscape can be engineered. The two approaches are not mutually exclusive: hierarchical encoding could provide passive protection at the physical level, with surface codes providing additional logical-level protection, combining the advantages of both.

7.2 For Fundamental Physics

What is the effective metric of physical space at the smallest scales?

All experimental evidence to date is consistent with space being a continuous, Archimedean manifold down to at least $10^{-18}$ meters (the electroweak scale probed by the LHC). But at the Planck scale ($\ell_P = \sqrt{\hbar G / c^3} \approx 1.6 \times 10^{-35}$ meters), quantum gravity effects are expected to become important, and the smooth manifold picture may break down.

Several approaches to quantum gravity suggest that space at the Planck scale may have a discrete, graph-like, or tree-like structure:

• Loop quantum gravity: Space is a spin network—a graph with quantized areas and volumes.
• Causal dynamical triangulations: Spacetime is built from discrete simplices, which can form tree-like branching structures.
• The holographic principle: The information content of a region scales with its boundary area, not its volume—suggestive of a tree-like encoding (as in tensor networks).

If physical space has an ultrametric component at small scales, then the hierarchical protection mechanism might be built into the fabric of reality. Quantum states might be naturally protected against certain classes of errors by the ultrametric structure of spacetime itself.

Testable signatures of ultrametric space:

1. Isosceles triangles at small scales. If space is ultrametric below some length scale $\ell$, then for any three points separated by distances $\lesssim \ell$, the two largest distances should be equal. This could be tested by interferometry at progressively smaller scales.

1. No accumulation of small displacements. In an ultrametric space, applying a sequence of sub-$\ell$ displacements should not result in a super-$\ell$ displacement. This is a violation of the Archimedean property and would be a striking signature.

1. Discrete distance spectra. In an ultrametric space, distances take discrete values ($b^{-k}$ for integer $k$). If small-scale distances are quantized in a geometric progression, this would be evidence of ultrametric structure.

No such signatures have been observed at currently accessible scales, placing an upper bound on any ultrametric length scale: $\ell \lesssim 10^{-18}$ m (from LHC). Future experiments at higher energies or greater precision could improve this bound—or discover the signature.

7.3 For the Theory of Measurement

The container framework developed in Part Zero is a theory of measurement in its own right. To measure is to answer the question: in which container does the thing reside? The precision of measurement is the depth of nesting. The two ways of measuring correspond to two fundamentally different strategies for nesting containers:

• Linear nesting (Archimedean measurement): Precision is a continuous parameter—the length of the interval. Errors are additive. This is the paradigm of classical metrology: measure finer and finer, with error bars that shrink as $1/\sqrt{N}$ for $N$ independent measurements.

• Branching nesting (ultrametric measurement): Precision is a discrete parameter—the depth in the tree. Errors do not accumulate across levels. This is the paradigm of digital measurement: each additional digit (each deeper level of nesting) increases precision by a fixed factor, and errors at one digit do not affect higher digits.

Both paradigms are used in practice. A digital voltmeter uses branching nesting (each digit is a level in a base-10 tree). An analog galvanometer uses linear nesting (the needle’s position on a continuous scale). The framework provides a unified language for understanding both—and for choosing between them when designing measurement systems.

Measurement and memory are the same problem. To measure a quantity is to transfer its value into a memory. To remember a value is to maintain a measurement across time. The same geometric principles that govern the endurance of memory also govern the precision of measurement. An ultrametric measurement—one that uses discrete, nested categories—is as robust against error accumulation as an ultrametric memory.

7.4 Open Problems

1. Engineering hierarchical energy landscapes. What is the maximum achievable depth $D$ in practice, across different physical platforms? What are the dominant sources of imperfection, and how do they affect the error suppression scaling?

1. Coherent operations within the hierarchy. How can quantum gates be designed that respect the hierarchical structure—coupling states within clusters while avoiding cross-cluster coupling? Is there a “hierarchical gate set” that is universal for quantum computation?

1. Entangling hierarchical memory cells. How can multiple hierarchical cells be coupled to create entangled states without destroying their individual protection? Does the hierarchical structure admit a natural notion of multi-cell entanglement?

1. Structured noise. The threshold principle assumes noise characterized by a single energy scale. Real noise has spectral structure (1/f noise, telegraph noise, burst noise). How does the framework extend to structured noise?

1. Optimal encoding depth. For a given noise environment and physical platform, what is the optimal encoding depth $D$ that minimizes the total error rate (including both logical errors from above-threshold perturbations and operational errors from imperfect gates)?

1. Experimental demonstration. Can the NMR or trapped-ion experiments proposed in Part Six demonstrate a statistically significant hierarchical protection effect? What are the minimum requirements on barrier hierarchy, temperature, and measurement fidelity?

1. p-adic quantum mechanics. If the state space of a quantum system is ultrametric, what is the appropriate formulation of quantum mechanics? Standard quantum mechanics is formulated over $\mathbb{R}$ (or $\mathbb{C}$). Is there a consistent $p$-adic or ultrametric formulation, and does it make different experimental predictions?

1. Fundamental length scale. Are there any experimental signatures of an ultrametric structure of space at small scales? What are the best current bounds, and what future experiments could improve them?

Appendices

Appendix A: Glossary

Appendix B: Collected Proofs

B.1 The strong triangle inequality for tree distance.

Let $k(x, y)$ be the longest common prefix length of $x$ and $y$ in $T_b$. Claim:

Proof. Let $p = k(x, y)$ and $q = k(y, z)$. Then $x_i = y_i$ for all $i < p$, and $y_i = z_i$ for all $i < q$. Set $r = \min(p, q)$. For any $i < r$, we have $i < p$ and $i < q$, so $x_i = y_i$ and $y_i = z_i$. By transitivity of equality, $x_i = z_i$ for all $i < r$. Thus $x$ and $z$ share a prefix of length at least $r$, so $k(x, z) \geq r = \min(p, q)$. ∎

Then:

The equality $b^{-\min(p,q)} = \max(b^{-p}, b^{-q})$ follows because $t \mapsto b^{-t}$ is strictly decreasing. ∎

B.2 Every triangle in an ultrametric space is isosceles.

Let $(X, d)$ be an ultrametric space. For any $x, y, z \in X$, let $d_1 = d(x,y)$, $d_2 = d(y,z)$, $d_3 = d(x,z)$. Let $M = \max(d_1, d_2, d_3)$. Claim: $M$ appears at least twice among $\{d_1, d_2, d_3\}$.

Proof. Suppose $M$ is unique. Without loss of generality, $d_1 = M$ and $d_2 < M$, $d_3 < M$. Apply the strong inequality to $(x, z, y)$:

Contradiction. Therefore $M$ cannot be unique. ∎

B.3 Proximity is transitive in an ultrametric space.

If $d(x, y) < T$ and $d(y, z) < T$, then $d(x, z) \leq \max(d(x, y), d(y, z)) < \max(T, T) = T$. Therefore $d(x, z) < T$. ∎

B.4 The triangle inequality for $(\mathbb{R}, d_{\mathbb{R}})$.

For any $x, y, z \in \mathbb{R}$:

The inequality $|a + b| \leq |a| + |b|$ (subadditivity of absolute value) is verified by case analysis:

• If $a, b \geq 0$: $|a + b| = a + b = |a| + |b|$.
• If $a, b \leq 0$: $|a + b| = -(a + b) = (-a) + (-b) = |a| + |b|$.
• If $a \geq 0, b \leq 0$: $|a + b| \leq \max(|a|, |b|) \leq |a| + |b|$.
• If $a \leq 0, b \geq 0$: symmetric. ∎

B.5 Error accumulation bound in $(\mathbb{R}, d_{\mathbb{R}})$.

Let $x_0$ be the target. After $N$ perturbations $\varepsilon_1, \ldots, \varepsilon_N$, the state is $x_N = x_0 + \sum_{i=1}^{N} \varepsilon_i$. The error is:

This follows by induction on $N$ using the triangle inequality:

• Base $N = 1$: $|\varepsilon_1| \leq |\varepsilon_1|$. True.
• Inductive step: $|\sum_{i=1}^{N+1} \varepsilon_i| = |(\sum_{i=1}^{N} \varepsilon_i) + \varepsilon_{N+1}| \leq |\sum_{i=1}^{N} \varepsilon_i| + |\varepsilon_{N+1}| \leq \sum_{i=1}^{N} |\varepsilon_i| + |\varepsilon_{N+1}| = \sum_{i=1}^{N+1} |\varepsilon_i|$. ∎

Appendix C: Numerical Tables

Table C.1: Distance vs. shared prefix length in a binary tree ($b = 2$).

Table C.2: Cluster count and encoding capacity in a binary tree.

Table C.3: Error suppression in a hierarchical memory (illustrative parameters).

Assumptions: $E_0 = 100 \cdot k_B T$, $\alpha = 1$, $\beta = 1$, $b = 2$, prefactor $C = 10^{12}$ Hz (typical vibrational attempt frequency).

At $D = 2$ and below, errors are frequent—the hierarchy is too shallow for these parameters. At $D = 1$ and above, errors are astronomically rare. The crossover (where $P \approx 1$ s$^{-1}$) is near $D \approx 2.5$, corresponding to $\Delta E_D \approx 17.7 \, k_B T$. For robust protection, $\Delta E_D \gtrsim 20 \, k_B T$ is needed, which means $D \leq 2.3$ for these parameters. To reach $P \lesssim 10^{-10}$ s$^{-1}$, $\Delta E_D \gtrsim 50 \, k_B T$ is needed ($D \leq 1$), meaning only depth-1 encoding is viable at this $E_0 / k_B T$ ratio.

Lesson from Table C.3: The ratio $E_0 / k_B T$ is critical. If $E_0 = 100 \, k_B T$, effective protection is achieved only at very shallow depths ($D \leq 1$). To get protection at deeper depths (which allows more logical states per physical system), a larger $E_0 / k_B T$ is required—i.e., lower temperature or higher energy scale. For example, if $E_0 = 1000 \, k_B T$, then at $D = 3$, $\Delta E_D = 125 \, k_B T$, giving $P \approx 10^{-42}$—effectively perfect protection at depth $3$.

Table C.4: Required $E_0 / k_B T$ for given depth and error rate.

Target: $P(\text{error}) \leq 10^{-10}$ s$^{-1}$, with $C = 10^{12}$ Hz, $\alpha = 1$, $\beta = 1$, $b = 2$.

Deeper encoding requires proportionally larger $E_0$ (or lower $T$) to maintain the same barrier at the encoding depth.

Appendix D: Hamiltonian Engineering Cookbook

D.1 General principle. To create a hierarchical potential $U(x)$ with $M$ levels, branching factor $b$, and barrier scaling exponent $\alpha$:

where $f(x)$ is a periodic “unit potential” with period $1$ and minima at integer values. Examples of $f(x)$:

• $f(x) = \cos(2\pi x)$ (sinusoidal, smooth)
• $f(x) = (x \bmod 1 - 0.5)^2$ (piecewise parabolic, sharper wells)
• $f(x) = -\text{triangle}(x)$ (piecewise linear, easiest to analyze)

The term with $m = 0$ creates $1$ well (the global structure). The term with $m = 1$ creates $b$ sub-wells within it. The term with $m = 2$ creates $b^2$ sub-sub-wells. And so on. The amplitude $b^{-\alpha m}$ ensures the barrier hierarchy.

D.2 Fourier synthesis for superconducting circuits. For a Josephson junction array, the potential is a sum of cosines with different periodicities:

where $\varphi$ is the superconducting phase and $E_{J,j}$ is the Josephson energy of junction $j$. To create a hierarchical potential:

• Choose a set of periodicities $\varphi_{0,j}$ that form a geometric progression: $\varphi_{0,j} = \varphi_0 \cdot b^{-j}$.
• Choose Josephson energies $E_{J,j} = E_J \cdot b^{-\alpha j}$.
• The total potential is then $U(\varphi) = -E_J \sum_j b^{-\alpha j} \cos(\varphi / (\varphi_0 \cdot b^{-j})) = -E_J \sum_j b^{-\alpha j} \cos(b^j \varphi / \varphi_0)$.

The different periodicities can be realized by loops containing $b^j$ junctions in series (each loop encloses a different flux). The amplitudes are set by the junction sizes (critical currents).

D.3 Spin chain couplings. For a spin chain with hierarchical Ising couplings:

the couplings must satisfy $J_i = J_0 \cdot b^{-\alpha i}$. This can be achieved by:

• Physical spacing: Place spins at positions $x_i$ with $x_{i+1} - x_i \propto b^{\alpha i}$ (exponentially increasing spacing). The dipolar or exchange coupling decays with distance, creating the exponential gradient.
• Tunable couplers: Use individually controlled couplers (e.g., flux-tunable Josephson junctions between superconducting qubits) and set the coupling strengths electronically.
• Engineered spin chains: In solid-state systems, use graded exchange interactions through compositional grading or strain engineering.

D.4 Molecular design. For molecular conformational hierarchies:

1. Identify rotatable bonds. Each bond has a torsional potential $V(\theta) = \frac{1}{2} \sum_n V_n (1 - \cos(n\theta))$ where $V_n$ are the barrier heights.

1. Rank bonds by barrier height. Use chemical principles to tune barriers:

- Steric hindrance: Bulky substituents increase rotational barriers.

- Conjugation: Partial double-bond character (as in amides) increases barriers dramatically.

- Hyperconjugation: Electron-donating/withdrawing groups modulate barriers.

- Ring strain: Bonds in small rings have different barriers than acyclic analogs.

1. Aim for a geometric progression of barrier heights: $V_k \approx V_0 \cdot 2^{-k}$ (for $b = 2$, $\alpha = 1$). The shallowest level should have the highest barrier (the logical encoding level). The deepest level should have a barrier comparable to but larger than $k_B T$.

1. Verify the hierarchy computationally (DFT or molecular mechanics) before synthesis.

D.5 Verification of the hierarchy. Once a candidate system is built (or simulated), verify the hierarchical structure:

1. Measure (or compute) all pairwise barrier heights. For $N$ distinguishable states, there are $N(N-1)/2$ barriers. These should organize into a tree: the barrier between states that diverge at depth $k$ should be $\Delta E_k$, independent of which specific states they are (within the same subtree).

1. Test the ultrametric inequality on barrier heights. The effective distance derived from barrier heights ($d \propto e^{\Delta E / k_B T}$ in the thermal regime) should satisfy $d(x, z) \leq \max(d(x, y), d(y, z))$ for all triples. This is a stringent test—most random potentials will fail it.

1. Plot $\log(\Delta E_k)$ vs. $k$. The plot should be a straight line with slope $-\alpha \log b$, confirming the exponential barrier scaling.

Version 0.3