Minimum Protected-Depth Envelope

• Minimum Protected-Depth Envelope is defined as the critical depth (α(n)) below which encoding good codes requires superlinear wire complexity.
• The work shows that linear-wire encoders are achievable at depth α(n)±O(1), with the inverse-Ackermann functions governing the trade-off between depth and size.
• It establishes a deep connection between circuit topology and superconcentrator-induced codes, yielding tight upper and lower bounds on the number of wires.

Searching arXiv for the cited paper and related prior work. The depth complexity of encoding good error-correcting codes can be studied in the arbitrary-gates, unbounded fan-in Boolean circuit model by measuring size in wires rather than gates. In that setting, the central quantity is $S_d(n)$, the minimum number of wires in a depth-$d$ circuit that encodes a code $C:\{0,1\}^n \to \{0,1\}^{32n}$ with distance at least $4n$. The paper "On the Minimum Depth of Circuits with Linear Number of Wires Encoding Good Codes" establishes that $S_d(n)$ is governed by inverse-Ackermann-type functions $\lambda_d(n)$, proves that linear-wire encoders become possible at depth $\alpha(n)$, and shows that this depth is necessary up to an additive constant $2$. It also isolates a subclass of MDS codes, called superconcentrator-induced codes, whose encoders are tightly characterized by superconcentrator graphs (Drucker et al., 2024).

1. Circuit model and the coding problem

The circuit model is strictly stronger than standard $AC^0/AC^1$. Inputs are $n$ Boolean bits; gates have unbounded fan-in and fan-out; and a gate of fan-in $d$0 may compute any Boolean function $d$1. Size is the number of wires, namely directed edges between inputs, gates, and outputs. Depth is the length of the longest directed path from an input to an output (Drucker et al., 2024).

Within this model, the paper studies encoders for good error-correcting codes. A code is called good when it has constant rate $d$2 and constant relative distance $d$3. More generally, the paper considers codes

$d$4

with relative distance at least $d$5, under the condition $d$6, where $d$7 is the binary entropy function. The normalized form used to define $d$8 fixes output length $d$9 and distance at least $C:\{0,1\}^n \to \{0,1\}^{32n}$0, corresponding to rate $C:\{0,1\}^n \to \{0,1\}^{32n}$1 and relative distance $C:\{0,1\}^n \to \{0,1\}^{32n}$2 (Drucker et al., 2024).

Formally,

$C:\{0,1\}^n \to \{0,1\}^{32n}$3

In the formal upper-bound statement the construction is linear, meaning that all gates are XOR gates, so the code is linear over $C:\{0,1\}^n \to \{0,1\}^{32n}$4. The broader lower bound, however, applies even when arbitrary Boolean-function gates of unrestricted fan-in are allowed (Drucker et al., 2024).

The underlying question is a depth-size tradeoff: for a fixed depth $C:\{0,1\}^n \to \{0,1\}^{32n}$5, how many wires are required to encode a constant-rate, constant-distance code? The paper sharpens an earlier result of Gál, Hansen, Koucký, Pudlák, and Viola by removing the depth dependence from the hidden constant in the upper bound and by extending the lower-bound method to super-constant depths (Drucker et al., 2024).

2. Inverse-Ackermann-type functions and the upper bound

The main asymptotic scale is given by a family of slowly growing functions $C:\{0,1\}^n \to \{0,1\}^{32n}$6. For a function $C:\{0,1\}^n \to \{0,1\}^{32n}$7 with $C:\{0,1\}^n \to \{0,1\}^{32n}$8 for all $C:\{0,1\}^n \to \{0,1\}^{32n}$9, let

$4n$0

The paper defines

$4n$1

(Drucker et al., 2024).

A few representative cases are listed below.

Depth parameter	Function	Order stated in the paper
$4n$2	$4n$3	$4n$4
$4n$5	$4n$6	$4n$7
$4n$8	$4n$9	$S_d(n)$0

These functions decrease with depth and become extremely small well before any conventional logarithmic scale. The paper defines an inverse-Ackermann depth threshold by

$S_d(n)$1

This quantity grows extremely slowly and is used as the critical depth at which linear wire complexity becomes attainable (Drucker et al., 2024).

The main upper bound states that for any $S_d(n)$2 and $S_d(n)$3 satisfying $S_d(n)$4, for sufficiently large $S_d(n)$5 and any $S_d(n)$6, not necessarily constant, there exists a linear circuit

$S_d(n)$7

of size $S_d(n)$8 and depth $S_d(n)$9 that encodes an error-correcting code with relative distance at least $\lambda_d(n)$0. In particular, when $\lambda_d(n)$1, the circuit has size $\lambda_d(n)$2 (Drucker et al., 2024).

Specialized to the normalization underlying $\lambda_d(n)$3, this yields

$\lambda_d(n)$4

with an absolute constant hidden in the $\lambda_d(n)$5, independent of $\lambda_d(n)$6. This improves the earlier $\lambda_d(n)$7 bound, where the constant deteriorated rapidly with $\lambda_d(n)$8 because of the recursive analysis used there (Drucker et al., 2024).

A more explicit constructive statement is given in terms of $\lambda_d(n)$9-partial good codes and a quantity $\alpha(n)$0. The paper proves that there exist absolute constants $\alpha(n)$1 such that, for suitable parameter ranges,

$\alpha(n)$2

with output fan-in at most $\alpha(n)$3, and

$\alpha(n)$4

Taking $\alpha(n)$5 yields

$\alpha(n)$6

which is the stated $\alpha(n)$7 behavior for all $\alpha(n)$8 (Drucker et al., 2024).

The passage to linear size at depth $\alpha(n)$9 follows from inverse-Ackermann properties proved in the appendix. In particular, the paper records that if $2$0, then $2$1, and by definition $2$2 is the smallest even $2$3 for which $2$4. Consequently,

$2$5

stated as Corollary 3.9 (Drucker et al., 2024).

3. Lower bounds and near-optimal depth

The principal lower bound shows that the inverse-Ackermann threshold is not merely sufficient but essentially necessary. Let $2$6, $2$7, and $2$8 be constants, and let

$2$9

be a family of circuits of size at most $AC^0/AC^1$0 encoding error-correcting codes with relative distance at least $AC^0/AC^1$1. Then, for sufficiently large $AC^0/AC^1$2, the depth of $AC^0/AC^1$3 is at least

$AC^0/AC^1$4

even when arbitrary Boolean-function gates of unrestricted fan-in are allowed (Drucker et al., 2024).

Combined with the upper bound, this yields a necessary-and-sufficient depth threshold for linear-wire encoders of good Boolean codes: $AC^0/AC^1$5 More precisely, depth $AC^0/AC^1$6 suffices, while depth below $AC^0/AC^1$7 is impossible for asymptotically linear-wire encoders (Drucker et al., 2024).

The argument is graph-theoretic. A circuit is converted into a directed acyclic graph, and the resulting graph must satisfy a dense regularity condition. The paper defines a layered depth-$AC^0/AC^1$8 DAG with $AC^0/AC^1$9 inputs and $n$0 outputs to be $n$1-densely regular if, for every $n$2, there exist distributions on $n$3-element subsets of inputs and outputs such that each individual vertex is selected with probability at most $n$4, while the expected number of vertex-disjoint paths from the sampled input set to the sampled output set is at least $n$5 (Drucker et al., 2024).

Let $n$6 denote the minimum number of edges of such a layered DAG. Pudlák’s original theorem gave

$n$7

for fixed $n$8. The paper refines Pudlák’s method and makes the dependence on $n$9 explicit: $d$00 with an absolute hidden constant (Drucker et al., 2024).

A key input from earlier work is that circuits encoding good Boolean codes must have densely regular underlying graphs. In the form quoted in the paper: if $d$01 encodes a code

$d$02

with relative distance at least $d$03, then after extending the graph with $d$04 dummy inputs, the underlying graph is $d$05-densely regular (Drucker et al., 2024).

To conclude the circuit lower bound, the graph is first layered, increasing size by a factor $d$06, and then the refined Pudlák bound is applied with $d$07. Comparison between the lower bound on edges and the assumed $d$08 wire count yields

$d$09

The inverse-Ackermann properties then force $d$10, which is equivalent to the stated lower bound $d$11 (Drucker et al., 2024).

For fixed depth $d$12, the upper and lower bounds match up to constant factors: $d$13 For super-constant depths, the paper presents this as a nearly complete characterization, while also noting that some intermediate regimes remain open (Drucker et al., 2024).

4. The depth threshold as a minimum protected-depth envelope

The paper’s results admit a threshold interpretation. If the objective is to preserve linear wire complexity while encoding good codes, then the depth scale $d$14 functions as a minimum protected-depth envelope. Below that envelope, the connectivity requirements imposed by constant rate and constant relative distance force superlinear size; at the envelope, linear size becomes achievable (Drucker et al., 2024).

For constant depth, the tradeoff is already explicit. Since

$d$15

the corresponding wire bounds are of the form

$d$16

up to constant factors. Thus constant depth cannot support linear-wire encoding of good codes (Drucker et al., 2024).

The same conclusion persists for slowly growing depths below the inverse-Ackermann scale. The paper explicitly notes that even depths such as $d$17, which are already very small relative to $d$18, remain below the point at which $d$19 becomes constant. Consequently, the wire complexity is still superlinear there (Drucker et al., 2024).

At depth $d$20, however, $d$21, so the upper bound becomes linear: $d$22 The lower bound shows that this is best possible up to an additive constant $d$23 in the depth parameter. This yields the sharp statement that the necessary and sufficient depth for linear wires is $d$24 (Drucker et al., 2024).

A plausible implication is that the obstruction is not tied to gate functionality but to graph structure. The lower bound allows arbitrary Boolean gates, yet still forces depth $d$25 for linear-wire encoding. This suggests that the controlling phenomenon is the depth required to realize sufficient disjoint-path connectivity with only $d$26 edges (Drucker et al., 2024).

5. Superconcentrator-induced codes

The second major theme is a special class of codes over finite fields. A code

$d$27

is called a superconcentrator-induced code if for every distinct $d$28,

$d$29

Since $d$30, this implies $d$31, so such codes achieve the Singleton bound and are MDS. The paper describes them as a subclass of MDS codes with the stronger property that output distance increases as input distance decreases (Drucker et al., 2024).

For linear codes, this condition has a precise matrix characterization. A linear code $d$32 is a superconcentrator-induced code if and only if every square submatrix of its generator matrix is nonsingular. The paper identifies these matrices as totally invertible or totally nonsingular (Drucker et al., 2024).

The graph-theoretic counterpart is the notion of an $d$33-superconcentrator: a DAG with $d$34 inputs and $d$35 outputs such that for every $d$36, and every $d$37-element input set $d$38 and $d$39-element output set $d$40, there exist $d$41 vertex-disjoint paths from $d$42 to $d$43. The paper notes that classical results of Valiant and of Dolev, Dwork, Pippenger, and Wigderson show that superconcentrators exist with size $d$44 at depth $d$45 (Drucker et al., 2024).

The central observation is a two-way connection between these algebraic codes and superconcentrator graphs.

Direction	Statement	Consequence
Code $d$46 graph	Any unrestricted arithmetic circuit encoding a superconcentrator-induced code must be a superconcentrator graph	The code inequality forces disjoint-path connectivity
Graph $d$47 code	Any superconcentrator graph can be converted to a linear arithmetic circuit over a sufficiently large field that encodes a superconcentrator-induced code	Superconcentrator structure is sufficient for the code property

The first direction is proved by contradiction via Menger’s theorem. If the underlying graph were not a superconcentrator, there would be equal-size input and output sets $d$48 separated by a cut $d$49 of size smaller than $d$50. Fixing inputs outside $d$51 and varying inputs on $d$52, the outputs on $d$53 would be determined by fewer than $d$54 field elements, so two distinct inputs would collide on $d$55. That would force

$d$56

while $d$57, contradicting

$d$58

(Drucker et al., 2024).

The converse direction is probabilistic. Given an $d$59-superconcentrator $d$60, one replaces each internal vertex by an addition gate over $d$61 and assigns a random coefficient $d$62 to each edge. If $d$63 is the depth of $d$64, then with probability at least

$d$65

the resulting circuit encodes a superconcentrator-induced code (Drucker et al., 2024).

The proof studies the $d$66 generator matrix $d$67, where

$d$68

For each equal-size pair $d$69, $d$70, the determinant $d$71 is a polynomial in the variables $d$72. Because $d$73 is a superconcentrator, there exist $d$74 vertex-disjoint paths from $d$75 to $d$76, which implies that this polynomial is nonzero. Its degree is at most $d$77, so Schwartz–Zippel yields

$d$78

and a union bound gives the stated failure probability (Drucker et al., 2024).

The paper therefore obtains a tight structural equivalence: encoding a superconcentrator-induced code forces superconcentrator topology, and superconcentrator topology can be turned into such a code over sufficiently large fields. This shows that the $d$79 tradeoff is not merely an artifact of one proof technique but is intrinsic to the combinatorial structure of this class of codes (Drucker et al., 2024).

6. Consequences, scope, and open questions

One consequence is a sharp answer to the question of how much depth is required to encode good Boolean codes with linear resources. For every depth $d$80, the paper frames the wire complexity as $d$81 for fixed $d$82, and shows that linear wires are achievable exactly when depth reaches the inverse-Ackermann scale up to an additive constant (Drucker et al., 2024).

The work also distinguishes circuit models. In the unbounded fan-in, arbitrary-gate Boolean model, the critical depth is $d$83. By contrast, for bounded fan-in models such as AND/OR circuits, the depth-time tradeoff is different; the paper notes that Sipser and Spielman show depth $d$84 is optimal and sufficient for linear-size encoders (Drucker et al., 2024).

For arithmetic circuits over finite fields, the superconcentrator-induced construction gives linear encoders over large fields. The paper states that exponential field size in the size of the graph suffices for the probabilistic conversion from superconcentrators to codes. It further notes that reducing the field size while maintaining the superconcentrator-code property connects to the classical MDS conjecture (Drucker et al., 2024).

The construction itself is non-explicit. The paper emphasizes that, as in the earlier work of Gál, Hansen, Koucký, Pudlák, and Viola, the resulting families are probabilistic rather than explicit. An open problem is to construct explicit families of good codes whose encoding circuits match the optimal depth-size tradeoff, especially at inverse-Ackermann depth (Drucker et al., 2024).

The paper also identifies unresolved depth regimes. It states that the precise asymptotic complexity of encoding good codes remains open for

$d$85

This leaves a gap between the fixed-depth regime, where the $d$86 behavior is exact, and the inverse-Ackermann threshold, where linear wires become possible (Drucker et al., 2024).

Finally, the parameter range treated by the upper bound covers any constant rate and relative distance within the Gilbert–Varshamov bound, via the condition $d$87. The paper notes natural extensions, including sub-constant rate, very high rate with small distance, and nonbinary alphabets under Boolean encoders. These are presented as directions rather than resolved cases (Drucker et al., 2024).