================================================================================
IDENTIFYING A GRAPH BY WEISFEILER-LEMAN REFINEMENT
A complete and accessible exposition
================================================================================
This document rewrites the paper "The complexity of identifying a graph
by Weisfeiler-Leman refinement" (September 25, 2026) for a reader who is
comfortable with basic graph theory but has not seen WL refinement
before. Every theorem, lemma, and proof of the original is retained. The
technical statements are unchanged. Proofs are rewritten to lead with
the key idea, then say what the pieces are, then verify the pieces.
Routine case checks are condensed to their essential reason, but every
claim is still justified. Comparisons to prior published work are
folded into a short "Background" note (see Sec. 1.1) rather than
repeated throughout.
The result is: deciding the Weisfeiler-Leman dimension of a graph is
EXPTIME-complete. That is the whole headline. The bulk of this document
explains what the statement means and why it is true.
--------------------------------------------------------------------------------
WHAT YOU SHOULD ALREADY KNOW
--------------------------------------------------------------------------------
We use a small amount of notation without re-defining it every time:
G, H : graphs (finite, undirected, no self-loops)
V(G), E(G) : vertex set, edge set
n : number of vertices in the input graph
m : number of bits of the input word driving the reduction
G ~ H : G and H are isomorphic (equal up to relabeling vertices)
in : "is an element of"
forall : "for every"
exists : "there exists"
<=, >= : the usual order
F_2 : the field {0,1} with arithmetic modulo 2, so 1+1=0
{{ ... }} : multiset ("bag"): order irrelevant, repetition counted
|S| : number of elements in the set S
N(v) : the set of neighbors of vertex v (its "neighborhood")
deg(v) : the degree of v, equal to |N(v)|
==> : "implies"; (a ==> b) means "a implies b"
V(G)^k : the set of all k-tuples of vertices of G, with repeats
allowed (so |V(G)^k| = |V(G)|^k)
If any of those feel unclear, the glossary at the end defines each one in
detail.
--------------------------------------------------------------------------------
CONTENTS
--------------------------------------------------------------------------------
0. Headline
1. The identification problem, in plain words
2. WL refinement, from the ground up
3. Equivalence, identification, and WL dimension
4. The bijective game
5. Why the game equals k-WL at k+1 slots
6. Dimension is monotone
7. Matching vertices have equal degrees
8. The easy half: this all sits inside EXPTIME
9. The hard half: an outline before any details
10. Encoding comparisons as a graph
11. Every equivalent mate is a G(beta)
12. Isomorphism equals a global offset
13. Image consistency: what every winning game guarantees
14. Lifting sums around a cycle
15. The global-offset criterion
16. Small supports and wrapping addresses
17. Translations on a small block collection
18. Translations become a winning strategy
19. Building the computation
20. Acceptance: every equivalent mate is isomorphic
21. Rejection: one small twist
22. Sizes: the reduction is polynomial-time
23. Main theorem
24. Writing the dimension in unary
Glossary
Bibliography
The structure is:
Sections 1-3 say what the problem is.
Sections 4-8 set up the tools and prove the upper bound.
Section 9 is the roadmap of the reduction.
Sections 10-22 build the reduction.
Sections 23-24 state the final result.
If you only want the shape of the argument, read 0, 1, 3, 8, 9, 23.
================================================================================
0. HEADLINE
================================================================================
Given a graph G and a positive integer k (written in binary), decide
whether k is large enough that Weisfeiler-Leman refinement at dimension
k already tells G apart from every nonisomorphic k-equivalent graph.
Write WLdim(G) for the least such k. Then:
The decision problem "is WLdim(G) <= k?" is EXPTIME-complete.
EXPTIME is the set of problems solvable in time 2^(poly(n)). "Complete"
means this identification problem is as hard as any other problem in
EXPTIME, under deterministic polynomial-time reductions.
Even when k is written in unary (as a string of 1's, so k is bounded by
the source input length), the problem remains EXPTIME-complete. So the
hardness is not driven by k being an exponentially large number written
in binary. The hardness is in the graph.
================================================================================
1. THE IDENTIFICATION PROBLEM, IN PLAIN WORDS
================================================================================
You have a graph G. You have a tool, the WL algorithm at dimension k,
that tries to prove two graphs are not isomorphic. Given enough
dimension, that tool can always succeed. The question is: how much
dimension does it need before it can do the job on G?
Wait carefully: "do the job on G" means the tool succeeds at telling G
apart from EVERY graph that is not isomorphic to it. To make the
question finite, we only consider graphs with the same number of
vertices.
For example, a path P_4 on 4 vertices has degrees 1, 2, 2, 1, while
a cycle C_4 has every vertex of degree 2. Same vertex count, different
degree profiles, so 1-WL already separates them. And P_4 is the only
4-vertex graph with degree sequence 2, 2, 1, 1, so no 4-vertex H can be
1-WL equivalent to it without being isomorphic: WLdim(P_4) = 1.
The classical example where 1-WL is helpless: compare
- C_6, the 6-cycle, and
- C_3 + C_3, two disjoint triangles.
Both have six vertices, both are 2-regular, and 1-WL assigns every
vertex in both graphs the same color throughout. So 1-WL cannot tell
them apart. 2-WL, which looks at ordered pairs of vertices, does
separate them, so WLdim(C_6) >= 2.
So WLdim(G) measures how much tuple information you need to isolate G
from every structurally-different graph of the same size. Weakly
symmetric graphs (paths, trees) fall to small dimension; strongly
symmetric ones (regular graphs, especially where a disconnected
counterpart exists) may need more.
The decision problem: given (G, k), decide whether WLdim(G) <= k.
The headline says: this problem is EXPTIME-complete.
-----------------------------------------
1.1 Background
-----------------------------------------
WL refinement originates in Weisfeiler and Leman's 1968 work on graph
canonization [WL1968]. The link between WL, counting logic, and pebble
games made it central in finite model theory [CFI1992, KN2019]. Related
complexity results for identification and pair equivalence:
- Arvind, Kobler, Rattan, Verbitsky proved P-hardness for dimension
ONE identification [AKRV2017].
- Lichter, Rassmann, Schweitzer extended fixed-dimension P-hardness to
every k >= 2, and proved NP-hardness when k is part of the input,
even for simple uncolored graphs [LRS2025].
- Variable-dimension pair equivalence was independently shown
coNP-hard [Seppelt2024, LRS2025].
The present paper completes the picture: for input-dimension
identification, the answer is EXPTIME-complete, not NP and not anything
smaller.
================================================================================
2. WL REFINEMENT, FROM THE GROUND UP
================================================================================
WL refinement is a coloring process. We color the vertices (or tuples of
vertices, for higher dimension), and refine colors round by round until
they stabilize.
-----------------------------------------
2.1 Dimension 1: ordinary neighbor-color refinement
-----------------------------------------
Start with every vertex of G having color RED.
Round 1: give each vertex v the color
(RED, {{color_0(w) : w in N(v)}}).
In words: v's new color records (a) its old color, and (b) the multiset
of colors of its neighbors from the previous round.
Round 2: same rule, using round-1 colors in place of round-0 colors:
color_2(v) = ( color_1(v), {{color_1(w) : w in N(v)}} ).
Continue. Two vertices end up in the same round-s class iff their
"neighborhood descriptions up to round s" are the same. Because color_2
contains color_1 contains color_0, the partition generated by color_s
refines the partition generated by color_{s-1}: no class can expand,
only split.
Colors stabilize after at most n rounds, where n = |V(G)|, because the
partition can split at most n-1 times (one vertex alone in a class).
For two graphs G and H, run the same process on both, SHARING color
names. Write G ==_1 H iff, at every round, the multisets of vertex
colors agree between G and H.
-----------------------------------------
2.2 Dimension k >= 2: tuples with the joint-update rule
-----------------------------------------
For dimension k >= 2, the objects that get colored are k-tuples: ordered
lists v = (v_1, ..., v_k) of vertices of G, repetition allowed. So there
are |V(G)|^k many objects.
Round 0: the initial color chi^G_0(v) records two things:
- which positions hold the same vertex (v_i = v_j or not);
- which pairs of positions are adjacent (v_i v_j in E(G) or not).
Round s+1: the color of v is:
( chi^G_s(v),
{{ ( chi^G_s(v[1<-z]), chi^G_s(v[2<-z]), ...,
chi^G_s(v[k<-z]) ) : z in V(G) }} )
where v[i<-z] means "v with its i-th position replaced by z." The
resulting object is a multiset of k-long LISTS of colors: a multiset
over z of the k-vector of replacement colors.
The words JOINT UPDATE are the technical name for this. The alternative,
which is NOT what we do, would be a "list" of k separate multisets
(one per coordinate). The joint convention is strictly stronger and is
what the game treats below with k+1 slots.
Two graphs G, H of the same order are k-equivalent, written G ==_k H,
iff at every round the multisets of k-tuple colors agree between G and
H.
================================================================================
3. EQUIVALENCE, IDENTIFICATION, AND WL DIMENSION
================================================================================
Two definitions:
k-WL IDENTIFIES G iff for every graph H with |V(H)| = |V(G)|,
G ==_k H ==> G ~ H.
WLdim(G) = least positive k that identifies G.
Some observations:
- If k >= n = |V(G)|, k-WL identifies G. Take the k-tuple that lists
all n vertices once and then arbitrarily repeats the first one for
the remaining positions. Its round-0 color already encodes the whole
graph: it records every equality among vertices and every adjacency.
Two graphs of order n whose color-histogram at dimension k agrees
include that specific tuple in the "same color" comparison. Across
graphs that means an actual structure-preserving bijection. So
WLdim(G) <= n.
- For n = 1, WL_1 identifies the single vertex trivially, so WLdim <= 1.
The WLdim measures how much work is needed to distinguish a particular
graph. Small regular graphs (cycles, strongly regular graphs) can need
larger dimension. Random graphs are typically caught at low dimension.
The theorem says: given G and k, deciding WLdim(G) <= k is EXPTIME-hard.
================================================================================
4. THE BIJECTIVE GAME
================================================================================
The algebraic definition above is awkward to prove things with. We swap
it for a two-player game in which the combinatorial structure is
transparent.
Fix two graphs G and H with |V(G)| = |V(H)| > 0. Fix a positive integer
K = number of "pair slots."
Position: at most K pairs (v, w), one per occupied slot, with v in V(G)
and w in V(H). The same v or the same w can occur in multiple slots.
One ROUND:
1. Spoiler chooses a vacant slot.
2. Duplicator names a bijection f : V(G) -> V(H).
3. Spoiler chooses v in V(G).
4. The chosen slot gets the pair (v, f(v)).
Additionally, Spoiler may discard occupied pairs at any time (making
those slots vacant again).
Winning condition: Duplicator wins when every occupied pair
(v_i, w_i), (v_j, w_j), satisfies
v_i = v_j iff w_i = w_j,
v_i v_j in E(G) iff w_i w_j in E(H).
So the occupied pairs must always exhibit a partial isomorphism. If
Duplicator has a strategy that survives every sequence of Spoiler's
choices, we say "Duplicator wins the game with K slots."
Spoiler tries to build an obstruction: always tries to make the pairs
incompatible. Duplicator has to always hand back a bijection that
extends the currently consistent set.
The whole reason we introduce this game is:
k-equivalence of G and H = Duplicator wins with k+1 slots.
That is Lemma 1 below. Note the "+1."
================================================================================
5. WHY THE GAME EQUALS k-WL AT k+1 SLOTS
================================================================================
LEMMA 1. Let G, H have equal positive order, and let k >= 2. Then:
G ==_k H iff Duplicator wins the bijective game with k+1 slots.
Furthermore, if Duplicator wins with at least 2 slots, then G ==_1 H
(ordinary dimension-1 equivalence).
-----------------------------------------
5.1 Key idea
-----------------------------------------
Run WL refinement "in parallel" on the tuple tables V(G)^k and V(H)^k,
giving matching color names whenever the same "definition" appears in
both.
Two facts about the joint-update rule are used:
(a) Coordinate permutations preserve equality. If chi^G_s(v) =
chi^H_s(w) for k-tuples v, w and sigma is any permutation of
positions, then chi^G_s(sigma v) = chi^H_s(sigma w). This follows
from the initial types and induction: the joint-update multiset is
invariant under reordering the tuple, and the coordinate
permutation just permutes the components of each replacement
vector.
(b) A stable color class carries a matching bijection. If v and w
have equal round-s colors with s a stable round, then within each
vector-of-replacement-colors class, we can pair up the underlying
vertices: the multiset of replacement vectors is the same for
both v and w, so vertices that appear in "the same replacement
vector" can be matched.
With those two facts:
(==>) If G ==_k H, pick a pair v, w of equal-color k-tuples (they
exist because color histograms are equal and both orders are
positive). Use (b) to attach one extra vertex per tuple from the
matching bijection, obtaining (k+1)-tuples v+, w+ whose "delete any
one common slot" projections are all - up to a common coordinate
permutation - equal-color k-tuples. Duplicator maintains exactly that
pair of (k+1)-tuples as the invariant. At every round, to fill the
spare slot, Duplicator announces the bijection from (b). Whichever
vertex Spoiler picks, "delete the spare" is unchanged, "delete any
other slot" is a coordinate permutation of a replacement pair. The
invariant survives. This proves a winning strategy.
(<==) Given a winning strategy with k+1 slots. Induct on the round s
that the refinement has stabilized. At every history ending with k
occupied pairs (i.e. one spare empty), the two k-tuples share a
common ordering of their vertices with equal round-s colors. The base
s=0 = partial isomorphism is the game's winning condition. The step:
fill the spare, Duplicator announces f, Spoiler picks some source
vertex; Spoiler may then discard one of the old slots, which is a
coordinate-deletion of an equal-color pair by induction. The strategy
matches all k replacement colors simultaneously for the same f, so
the whole joint vector matches and the next-round color agrees.
Fill k slots from scratch in a fixed order. The succession of
announced bijections (invert the first one to find the source of any
target anyway) gives a bijection V(G)^k -> V(H)^k that sends every
tuple to one with the same color at every round. So the color
histograms agree, giving G ==_k H.
Final claim (which will also be useful later): if Duplicator wins
with at least 2 slots, run the same inductive argument on
vertex-color refinement at every history with exactly 1 occupied
pair. One more slot is enough to compare the neighbor-color
multisets, giving G ==_1 H.
================================================================================
6. DIMENSION IS MONOTONE
================================================================================
LEMMA 2. For positive integers j <= k, G ==_k H implies G ==_j H.
Consequently, WLdim(G) <= k iff k-WL identifies G.
Proof. From equal histograms at dimension k we get |V(G)|^k = |V(H)|^k,
hence |V(G)| = |V(H)|; if that is zero, the claim is trivial, so assume
positive.
By Lemma 1, G ==_k H iff Duplicator wins with k+1 slots. Restricted to
any j+1 <= k+1 slots, the same strategy still wins (Duplicator ignores
the extra slots). For j = 1, the final clause of Lemma 1 applies.
For "consequently": if some j <= k identifies G, then G ==_k H implies
G ==_j H implies G is isomorphic to H, so k-WL identifies G too.
Conversely, "WLdim(G) <= k" literally means "exists j <= k that
identifies G."
================================================================================
7. MATCHING VERTICES HAVE EQUAL DEGREES
================================================================================
LEMMA 3. In every history following a winning strategy (with at least
2 slots), matched vertices have equal degrees. Consequently, every
bijection announced by the strategy preserves degrees.
Proof. Take any currently matched pair (v, w). Fill one more slot;
Duplicator announces f: V(G) -> V(H). Let u range over V(G). Every pair
(u, f(u)) must be consistent with the fixed (v, w). So u is adjacent to
v iff f(u) is adjacent to w. In other words, f sends N_G(v) onto N_H(w)
bijectively. That is a bijection between two sets, so their cardinalities
are equal: deg(v) = deg(w).
For the "consequently": for each vertex v, its pair (v, f(v)) under the
announced f satisfies the same, so deg(f(v)) = deg(v).
This lemma is what lets us "recover forgotten colors." Section 10 will
encode vertex colors as degrees; Lemma 3 then says that any equivalent
mate automatically respects those colors.
================================================================================
8. THE EASY HALF: THIS ALL SITS INSIDE EXPTIME
================================================================================
PROPOSITION 1. The problem "given a graph G (adjacency matrix) and
positive binary integer k, decide WLdim(G) <= k" belongs to EXPTIME.
Proof.
Step 0. Check the input is a well-formed graph. O(n^2) time on an
n-vertex matrix. Reject bad input.
Step 1. Compare k to n = |V(G)|. If k >= n, accept. Reason: a k-tuple
listing all vertices (with repeats to fill) has a round-0 color that
already determines the whole graph; two such tuples with equal colors
give an isomorphism.
Step 2. Now 1 <= k < n. Enumerate every labeled simple graph H on n
vertices. For each H, check two conditions:
(i) G ==_k H (color histograms equal at every round)
(ii) G is NOT isomorphic to H
If any such H exists, REJECT (WLdim > k). Otherwise ACCEPT (WLdim <= k).
Correctness is Lemma 2 (monotonicity): WLdim(G) <= k iff no
nonisomorphic k-equivalent H exists.
Complexity. The number of labeled simple graphs on n vertices is
2^(C(n,2)), exponential in n. Testing G ~ H is naively n!, exponential
in n log n. Testing G ==_k H: run k-WL refinement on 2 n^k entries,
for at most 2 n^k rounds (colors refine, never merge). Each round's
signature has 1 + n * k entries; sorting and renaming is polynomial in
n^k. So equivalence is poly in n^k, hence poly in n^n, hence
exponential in a polynomial in n.
Total: 2^(C(n,2)) * (n! + poly(n^n)). This is exponential in a
polynomial of the input length (n^2 matrix entries + log k bits). Well
inside EXPTIME.
So the upper bound holds.
================================================================================
9. THE HARD HALF: AN OUTLINE BEFORE ANY DETAILS
================================================================================
The real content is EXPTIME-hardness. Given any EXPTIME language L and
any input word w of length m, we produce, in deterministic polynomial
time in m, a pair (G, r) with G a graph and r a polynomial-in-m
dimension, such that:
w in L ==> WLdim(G) <= r (so r-WL identifies G)
w not in L ==> WLdim(G) > r (some nonisomorphic H with G ==_r H)
Then any EXPTIME language reduces to the WLdim problem in poly time.
The graph G is built so that "one twist on the square bits" corresponds
to "one accepting computation of M on w." There are four main construction
ideas, each a whole section:
A. (Secs. 10-12) The graph encodes BINARY LINEAR CONSTRAINTS on some
"shift bits." A given bit assignment beta yields a graph G(beta).
If G(0) ==_r H, then H is isomorphic to G(beta) for some beta
(Prop. 2). And G(0) is isomorphic to G(beta) iff the shift bits
admit a "global offset" satisfying every square equation (Prop. 3).
So "isomorphism" = "linear system has a solution."
B. (Secs. 13-15) Every equivalent graph gives "image consistency":
nonempty sets Q_{P,a} of values, with one scalar equality per
comparison. This is necessary. It is not, in general, sufficient.
Secs. 16-18 set up the special case where necessary = sufficient:
the special twist where all shift bits live on one row of the
first r wiring rows and one end of each wire.
C. (Sec. 16-18) Exact-projection theorem: on a small collection of
blocks (at most r+1), every old translation has a legal extension
to a slightly larger one. This is what lets Duplicator preserve
consistency of the maintained assignment while answering new
placements.
D. (Secs. 19-22) Fill in the actual computation:
- A polynomial-size monotone circuit simulating the EXPTIME
machine M for T = 2^(p(m)) steps, with "product-of-bits"
address structure at every gate.
- Two-coordinate buffers use the trick that {(0,0),(1,1)} subset
of F_2^2 has full coordinate images but sum set {0}. That lets
"singleton at source" propagate to "singleton at destination"
without requiring matching shifts.
- When M accepts, all equivalent mates are isomorphic.
- When M rejects, one row-r self-connection at the broadcast
type A gives a nonempty-family of shifts whose global-offset
system has no solution.
Combine A-D: correctness. Corollary (Sec. 24): even k in unary suffices.
Everything below is filling in A-D.
================================================================================
10. ENCODING COMPARISONS AS A GRAPH
================================================================================
We now build the graph G used in the reduction. The construction has two
halves:
- A "schema" describes a linear comparison you want to enforce.
- "Wires" and "blocks" translate a whole collection of schemas into an
uncolored graph.
Section 10 says how to build the graph. Sections 11-12 characterize the
graphs that are r-equivalent to it.
-----------------------------------------
10.1 Schemas
-----------------------------------------
Fix the field F = F_2 (arithmetic modulo 2). Choose integers
r >= 32, K := r + 1.
Indices i = 1, ..., r are cyclic: r+1 means 1.
A "real type" P is an abstract object with a dimension d_P in {1, 2}.
A "value at type P and address a in {0,1}^r" is an element of F^{d_P}.
A "comparison schema" e consists of:
- ordered real endpoints P, R;
- a nonempty product domain D_e = prod_{i=1}^r D_{e,i},
each D_{e,i} nonempty in {0,1};
- coordinate injections sigma_{e,i} : D_{e,i} -> {0,1};
- scalar linear forms
lambda_{e,0} : F^{d_P} -> F,
lambda_{e,1} : F^{d_R} -> F.
Write sigma_e(a) for the vector (sigma_{e,1}(a_1), ..., sigma_{e,r}(a_r)).
Meaning: schema e "compares" lambda_{e,0} at (P, a) with lambda_{e,1}
at (R, sigma_e(a)), for every a in D_e.
Loops (P = R), parallel schemas, and zero forms are all allowed.
-----------------------------------------
10.2 Wires
-----------------------------------------
Each schema e : P -> R gets replaced by a "wire" with 4K private internal
types, each of dimension 1. So a wire has 4K+1 links; distinct wires have
no internal types in common.
Real types plus internal types (across all wires) form the "type layout."
For a type t, define:
A_{t,i} = {0,1} if t is real,
D_{e,i} if t is internal to wire e.
These are the "allowed labels at coordinate i." Let d_t be the dimension
of type t.
At each link l (a type pair t-u with t first, u second in wire
orientation), we want per-coordinate allowed matches. Define:
M_{l,i} = { (p, p) : p in D_{e,i} }
if l is not the last link of e,
{ (p, sigma_{e,i}(p)) : p in D_{e,i} }
if l is the last link of e.
At internal endpoints, the scalar form L_{l,*} is the identity on F. At
real endpoints, it is lambda_{e,0} (first link, P side) or lambda_{e,1}
(last link, R side).
-----------------------------------------
10.3 Readouts and blocks
-----------------------------------------
Readouts are the "wires of the graph" -- the named variables we will
assign to F^{d_t} or F.
Horizontal readout (a vector):
x_{i,t,p,p'} in F^{d_t} for p in A_{t,i}, p' in A_{t,i+1}.
Vertical readout (a scalar):
y_{i,l,p,q} in F for (p, q) in M_{l,i}.
Think of them as edges: horizontal readout x_{i,t,p,p'} connects corner
(i, t, p) to corner (i+1, t, p'). Vertical y_{i,l,p,q} connects
(i, t, p) to (i, u, q) where l = t-u.
A "full address at type t" is a vector a with a_i in A_{t,i} for all i.
It selects one horizontal readout per row. The "value represented at
address a" is:
h_t(a) = sum_{i=1}^r x_{i,t,a_i,a_{i+1}} in F^{d_t}.
So the exponentially-many values h_t(a) are determined by the
shorter tables x_{i,t,p,p'} -- the y-readouts are not part of h_t.
There are two kinds of "blocks" (small local consistency conditions):
- Site block: uses just one horizontal readout x_{i,t,p,p'}.
Every value of x is legal. (Site blocks impose nothing.)
- Square block: for each link l = t-u, row i, and pairs
(p, q) in M_{l,i}, (p', q') in M_{l,i+1}, uses FOUR readouts:
x_t = x_{i,t,p,p'},
x_u = x_{i,u,q,q'},
y_i = y_{i,l,p,q},
y_{i+1} = y_{i+1,l,p',q'}.
Choose one bit beta_B in F per square B. A valuation of the block is
LEGAL iff
L_{l,t}(x_t) + L_{l,u}(x_u) + y_i + y_{i+1} = beta_B. (*)
At least one of the two endpoints of any link is internal, so a square
uses d_t + d_u + 2 <= 5 bit coordinates. Equation (*) is a nonzero
linear equation, hence the number of legal valuations of the block is
at most 16; the count is the same whether beta_B = 0 or beta_B = 1.
The reason for long wires: a collection of <= K blocks uses at most K
distinct layout links, because site blocks use none and squares use one.
So such a collection cannot span a whole wire of 4K+1 = 4r+5 links. This
"no short witness wraps a full wire" fact drives Section 17.
-----------------------------------------
10.4 The uncolored graph G(beta)
-----------------------------------------
First, build a COLORED graph:
1. For each bit coordinate of each readout name, create two bit-value
vertices, labeled 0 and 1. Give the pair a fresh color. Pairs are
SHARED across blocks: if the same readout name appears in multiple
blocks, they use the same two bit-value vertices.
2. For each block, create one "valuation vertex" per LEGAL valuation,
all with a fresh color specific to that block. Connect each
valuation vertex to the bit-value vertex for its assigned value in
each coordinate used by the block. No other edges.
The result is bipartite: valuation vertices on one side, bit-value
vertices on the other. Call these "base vertices." Let N = number of
base vertices; this N is independent of the beta choices (by the
"at most 16 per square" count).
Give fresh colors to the base-vertex classes, numbered 1, ..., J.
Finally erase the colors: attach j(N+1) new private leaves at every base
vertex of color j, then "forget" the colors. The resulting uncolored
graph is G(beta). Its order is independent of beta. G(0) is the graph
where all square bits beta_B are zero.
================================================================================
11. EVERY EQUIVALENT MATE IS A G(beta)
================================================================================
PROPOSITION 2 (Reconstruction of any equivalent mate). Fix comparison
data as in Section 10, with r >= 32. Any graph H with H ==_r G(0) is
isomorphic to G(beta) for some choice of one bit beta_B per square block.
-----------------------------------------
11.1 Key idea
-----------------------------------------
Lemma 1 provides a winning strategy from G(0) to H (with K = r+1 slots).
The strategy's degree-preservation (Lemma 3) lets us "recover the
forgotten colors." Once colors are back, we can identify each block and
hence each beta_B globally, because the same two bit-value vertices are
shared by every block that uses that coordinate.
-----------------------------------------
11.2 Details
-----------------------------------------
Base vertices of color j in G(0) have degrees in the interval:
I_j = [ j(N+1), j(N+1) + N - 1 ].
These intervals are pairwise disjoint and above 1. Every other vertex
(private leaves) has degree 1.
Lemma 3 says any announced bijection preserves degrees. So H has exactly
the same number of vertices in each I_j, exactly the same number of
degree-1 vertices, and no vertices of any other degree. Call such
vertices "base of color j."
Every base vertex of H is reachable by a single placement (invert the
first announced bijection). The next announced bijection must preserve
degrees AND adjacency to that vertex, so its degree-1 neighbors are
exactly j(N+1). Summing over all base vertices accounts for all degree-1
vertices of H; hence all degree-1 vertices are private leaves on bases,
none sit in a separate component. (Bases stay base-colored.)
There can also be no base edge between two color classes that are never
adjacent in G(0). Reason: any two target vertices can be reached by two
successive placements using inverse bijections; preimages have the
prescribed colors; partial isomorphism preserves edge tests. In
particular, no edges are within one side of the bipartition and none to
unrelated bit-value vertices.
Now fix a block. Mark every valuation vertex, plus both bit-value
vertices of every coordinate it uses. In a square at most 16 + 2*5 = 26
vertices; a site is smaller. All fit in K >= 33 slots. Their images are
distinct, preserve colors, and exhaust the corresponding color classes
in H (we checked class sizes). So H locally looks like G(0) with colors.
Label the two vertices of every bit pair in H by {0, 1}, once for all
blocks. A local colored isomorphism may flip each bit pair independently,
because flipped-or-not is a degree (and color) preserving operation on
that pair.
A site's valuation vertices are exactly all site valuations under these
chosen labels.
A square's local isomorphism, translated to homogeneous data via the
coordinate flips, produces some right-hand side beta_B; the valuation
vertices of that block are exactly the legal valuations of Equation (*)
with that RHS. Different blocks are permitted to use different local
flips; what is COMMON is the single global labeling of each bit pair.
Combining: H is isomorphic to G(beta) for the beta_B described above.
The subtle assertion being proved: an arbitrary r-equivalent H is not
"some new structure with the same WL color histogram" but literally the
same construction, up to a bit-flip per square.
================================================================================
12. ISOMORPHISM EQUALS A GLOBAL OFFSET
================================================================================
PROPOSITION 3 (Isomorphisms and global offsets). For any choice of
square bits beta, G(0) is isomorphic to G(beta) iff there exists a
"global offset assignment" solving every square equation with its
RHS.
Definition. A global offset assignment gives:
- a vector in F^{d_t} to every horizontal readout x_{i,t,p,p'},
- a scalar in F to every vertical readout y_{i,l,p,q},
with the SAME value used in all blocks that share that readout name.
Such an assignment is "legal for beta" if Equation (*) holds at every
square block.
Note: site blocks impose no condition, so any beta-pinned assignment
satisfying the squares simultaneously is enough.
Proof idea.
(=>) An isomorphism preserves degrees, so it acts on every base color
class. Its action on each bit pair is a bit flip, i.e. a delta in F.
Collecting these flips per readout name gives the offset vector and
offset scalar. The image of the all-zero valuation vertex of a block of
G(0) is adjacent to the bit-value vertices specified by these offsets,
so its coordinates ARE the restriction of the offsets to that block.
Legality of the image forces Equation (*) with RHS beta_B. Because bit
pairs are shared across blocks, all those local flips are part of ONE
global assignment.
(<=) Given such a global offset assignment, we translate each bit pair
by its own offset, and translate each block's valuation by the offsets
on its coordinates. Linearity makes the translation a bijection from
the homogeneous legal valuations to the shifted legal valuations of that
block. Shared offsets preserve every valuation-bit incidence. The
translation extends to the private leaves (whose count depends only on
the preserved base color).
================================================================================
13. IMAGE CONSISTENCY: WHAT EVERY WINNING GAME GUARANTEES
================================================================================
So far, Proposition 2 says any equivalent mate is G(beta). Proposition 3
says "G(beta) is isomorphic to G(0)" is a global-solvability question.
The next three sections extract from a winning game (of G(0) ==_r G(beta))
a NECESSARY condition on the beta bits, and in section 16 the "special"
case where it is also SUFFICIENT.
The necessary condition is called "image consistency."
-----------------------------------------
13.1 Two more useful quantities
-----------------------------------------
Fix a schema e : P -> R and an address a in D_e. Along its wire, use
the address a at the first endpoint and every internal type, and
sigma_e(a) at the last endpoint. For each link l of this wire and row i,
write B(i, l, a) for the square specified by these addresses.
Define:
Delta_l(a) = sum_{i=1}^r beta_{B(i, l, a)}
Delta_e(a) = sum_{l on wire e} Delta_l(a) all in F.
Each summand belonging to row i depends only on (a_i, a_{i+1}). So
Delta_l(a), Delta_e(a) are functions of a. They have "row-sum form"
(preview: sum of functions on adjacent pairs of coordinates), because
each summand depends only on two consecutive coordinates.
-----------------------------------------
13.2 Statement
-----------------------------------------
PROPOSITION 4 (Image consistency). For the construction of Section 10,
suppose G(0) ==_r G(beta). Then there exist nonempty sets
Q_{P,a} subset F^{d_P},
one for each real type P and each address a in {0,1}^r, such that for
every comparison schema e : P -> R and every a in D_e:
lambda_{e,0}( Q_{P,a} ) = Delta_e(a) + lambda_{e,1}( Q_{R, sigma_e(a)} ). (**)
Here the notation lambda(S) means the image set {lambda(s) : s in S},
and the equality is a set equality in F (a scalar set):
{ lambda_{e,0}(v) : v in Q_{P,a} }
= { Delta_e(a) + lambda_{e,1}(w) : w in Q_{R, sigma_e(a)} }.
Interpretation. (**) is necessary for equivalence, and is analogous to
"arc consistency" in constraint satisfaction: each value of one endpoint
has some supporting value at the other endpoint for THIS comparison.
Those supports may be different for different comparisons. In particular,
a "self-comparison" (P = R, with the same address) does NOT force its two
endpoint values to be equal vectors.
-----------------------------------------
13.3 How the sets Q are produced from a winning game
-----------------------------------------
Fix a winning (r+1)-slot strategy for G(0) -> G(beta).
Consider histories in which Spoiler places only the all-zero valuation
vertex of some block of G(0). The response is the valuation of the SAME
block in G(beta); degree preservation guarantees this (base-class sizes
match).
Claim. In any such history, if two marked blocks share a readout name,
their responses agree on that readout.
Proof of claim. If they disagreed, some bit-value coordinate differs.
Discard all pairs except those two blocks, then mark the corresponding
bit pair (whose value is inconsistent). The zero valuation of that bit
in G(0) is adjacent to both source blocks. But no vertex of the target
bit pair is adjacent to both blocks (they disagree). Contradiction.
Uses at most 3 slots.
Now for a type t and a full allowed address a at t, define the "site
ring" to be the r site blocks with readouts x_{i,t,a_i,a_{i+1}}. Mark
their zero valuations. Let h_t be the sum (in F^{d_t}) of the response
vectors on those readouts.
Claim. Starting from this ring, and using one spare slot, we can
transfer the sum across any incident link with matching addresses. The
procedure places the square block of that row before discarding the old
site pair.
After all rows are replaced, using at most r+1 slots:
- Adjacent square responses agree on their shared vertical readout.
- Summing Equation (*) around the square ring cancels every vertical
value twice.
So for the two horizontal sums h_t, h_u at link l = t-u with address a:
L_{l,t}(h_t) + L_{l,u}(h_u) = Delta_l(a). (link identity)
Replace each square by its arrival-side site (place before discarding).
The arrival sum is copied exactly. Works in either direction. Along a
complete wire, the internal sums are scalars and cancel (each vertical
appears once as "y_i" and once as "y_{i+1}" between consecutive links).
So endpoint equality is:
lambda_{e,0}( h_P ) = Delta_e(a) + lambda_{e,1}( h_R[sigma_e(a)] ).
Now define:
Q_{P,a} = { h_P obtained from a zero-block-only history that ends
with exactly the site ring at (P, a) marked }.
Nonempty: the ring can be queried from the empty position.
If we continue any history realizing a value in Q_{P,a} along the wire
(via the same small proof moves), we get one side of (**). Continuing
backwards from Q_{R, sigma_e(a)} gives the reverse inclusion. No
"memorylessness" or common simultaneous choice is used, so this works
even for self-comparisons.
That is the proof of Proposition 4.
================================================================================
14. LIFTING SUMS AROUND A CYCLE
================================================================================
The next fact is purely about F-valued functions on a cyclic product.
LEMMA 4 (Cycle lifting). Let r >= 3, each A_i nonempty, and let
g_i : A_i x A_{i+1} -> F for i = 1, ..., r (cyclically). If
sum_{i=1}^r g_i(a_i, a_{i+1}) = 0 for all a in prod_i A_i, (***)
then there exist functions Y_i : A_i -> F with
g_i(u, v) = Y_i(u) + Y_{i+1}(v) for every i, u in A_i, v in A_{i+1}.
Proof. Fix a base address b. Taking the RECTANGULAR DIFFERENCE in
coordinates i and i+1 of (***) eliminates all but g_i, since r >= 3.
Consequently we can write:
c_i = g_i(b_i, b_{i+1}),
alpha_i(u) = g_i(u, b_{i+1}) + c_i,
gamma_i(v) = g_i(b_i, v) + c_i,
g_i(u, v) = c_i + alpha_i(u) + gamma_i(v).
Varying coordinate i alone from b gives:
alpha_i(u) = gamma_{i-1}(u).
Evaluating at b gives sum_i c_i = 0 (that is exactly (***) at b).
Choose constants C_i in F satisfying C_i + C_{i+1} = c_i cyclically:
choose C_1 arbitrarily and recurse. The final equation holds because the
c_i sum to zero. Set
Y_i(u) = alpha_i(u) + C_i.
Then Y_i(u) + Y_{i+1}(v) = c_i + alpha_i(u) + gamma_i(v) = g_i(u, v).
For a singleton factor the deviations vanish, so the same argument
covers that case.
================================================================================
15. THE GLOBAL-OFFSET CRITERION
================================================================================
This section connects Section 13's necessary condition to Section 12's
isomorphism criterion.
PROPOSITION 5 (Global offsets). For any right-hand sides beta in the
Section 10 construction, all square equations admit a simultaneous
assignment to the readout names if and only if there exist "row-sum
functions":
h_P : {0,1}^r -> F^{d_P}, one per real type P,
satisfying
lambda_{e,0}(h_P(a)) + lambda_{e,1}(h_R(sigma_e(a))) = Delta_e(a)
for all schemas e : P -> R and addresses a in D_e. (GLOBAL)
Equivalently, these functions exist iff G(0) is isomorphic to G(beta).
Here "row-sum form" for h : A_1 x ... x A_r -> V means:
h(a) = sum_{i=1}^r h_i(a_i, a_{i+1})
for functions h_i : A_i x A_{i+1} -> V (with cyclically r+1 = 1).
Note: restriction to a product subdomain preserves row-sum form, as do
linear maps on V and pullback by coordinatewise maps of the domain. In
particular, all functions Delta_l(\cdot), Delta_e(\cdot) defined in
Section 13.1 have row-sum form.
Proof.
(=>) Suppose global offsets exist. For each real type P, sum the
horizontal readouts around each site ring to define h_P. These are by
construction sums of one readout per row, each readout being a function
of the two consecutive coordinates a_i, a_{i+1}. So h_P has row-sum
form.
Sum the square equations around any link l to get the link identity
L_{l,t}(x_t) + L_{l,u}(x_u) = Delta_l(a)
now for the assigned offsets. Sum along a wire: internal scalar sums
cancel and the endpoints satisfy (GLOBAL).
(<=) Suppose row-sum functions exist satisfying (GLOBAL). For each
internal type s of wire e : P -> R, define a scalar function on D_e:
h_s(a) = lambda_{e,0}(h_P(a)) + sum_{l from P through s} Delta_l(a).
This is row-sum form (each term is). The link identity holds immediately
on the first and internal links by telescoping cancellation. It holds on
the last link by (GLOBAL), with h_R evaluated at sigma_e(a); that latter
function is of row-sum form in a because sigma_e is coordinatewise.
Now choose a row decomposition of each function:
- for real-type h_P (defined on {0,1}^r): a sum of adjacent-coordinate
terms on the full product;
- for internal-type h_s (defined on the allowed product): a sum of
adjacent-coordinate terms on that product.
Assign the chosen row-i term to all horizontal readouts of type t at
row i. Each choice is once per type, so readouts shared between
different incident wires already have consistent values across blocks.
For a link l = t-u on wire e with matched addresses, let x_t, x_u be
those assigned horizontal values. The residual
g_i(a_i, a_{i+1}) = beta_{B(i,l,a)} + L_{l,t}(x_t) + L_{l,u}(x_u)
is a function of the two source coordinates (a_i, a_{i+1}), and the
link identity proves the sum around the cycle is zero on D_e. So
Lemma 4 supplies Y_i : D_{e,i} -> F with consecutive sums equal to g_i.
Assign Y_i(a_i) to the vertical readout of link l indexed by the match
corresponding to a_i.
Well-definedness. Every vertical readout is one match; matches are
indexed bijectively by source labels on every link (including the last,
where sigma is an injection on D_{e,i}). Every square is at a pair of
adjacent source labels that extends to a full address (all other product
factors are nonempty). So the assignment really does satisfy Equation
(*) at every square, not just along a selected collection of addresses.
Different links have distinct vertical readout names, so their
assignments are independent. Site blocks impose no additional
constraint. We have obtained the desired global offsets.
The equivalence with isomorphism is Proposition 3.
Finally, an observation used later: a row-sum function on a product
subdomain of {0,1}^r EXTENDS to a row-sum function on the full cube
(extend each row term arbitrarily to the corresponding binary pair).
================================================================================
16. SMALL SUPPORTS AND WRAPPING ADDRESSES
================================================================================
From now on we build "translations" of readouts that correspond to
permitted bit flips. The main tool is a family H_U of admissible
assignments parameterized by a small collection U of blocks. Section 17
shows these translate every legal valuation-for-G(0) into a
legal valuation-for-G(beta), for beta in a "special twist."
-----------------------------------------
16.1 The twist we use
-----------------------------------------
Choose constants b_e in F, one per comparison schema. For a link l on
wire e, let b_l = b_e if l is the LAST link of the wire, else 0. Let
delta_r = 1 and delta_i = 0 for i < r. Define the SQUARE BITS
beta_B = delta_i * b_l for a square B in row i on link l. (T)
Then Delta_e(a) = b_e for every a in D_e (the sum around the wire
collapses: only the row-r end-link bit survives).
-----------------------------------------
16.2 Small supports are trees with short arms
-----------------------------------------
For a set U of blocks, define Gamma_U to be the graph whose vertices are
corners and whose edges are the readout names appearing in U (repeated
names collapse to one edge). Each block belongs to one connected
component of Gamma_U.
The "type support" of a component is the subgraph of the type layout
consisting of its types and links.
LEMMA 5. If |U| <= K, every component of Gamma_U has a type support
that is a tree with at most one real type. A support containing no real
type lies inside a single wire interior. A support containing a real
type is that type plus short arms in its incident wires.
Proof. A type support is connected (it is the image of a connected
corner graph). Each square contributes at most one distinct layout
link, sites contribute none, so a support has at most K links.
Internal types have layout degree exactly 2. A cycle, or a path
between two distinct real types, must traverse a complete wire of
4K+1 links; neither fits. The rest of the description follows.
The point: "small" means "cannot span a full wire."
-----------------------------------------
16.3 Aligned labels and the wrapping marker
-----------------------------------------
To compare labels at different types in a type support, fix for every
link and boundary index an EXTENSION of the label-matching injection
M_{l,i} to a full permutation of {0,1}. Use identity for identity
matches and the inverse permutation when traversing backwards. Such
extensions exist because the injections are on nonempty subsets of a
2-element set.
Choose a reference type in each support tree. Transport along unique
paths to express every label at reference-type coordinates. These are
the "aligned labels." Vertical edges preserve aligned labels.
Markers. A horizontal edge in row i gets marker delta_i. A vertical
edge gets marker zero. A BLOCK COLLECTION is "wrapping" if its corner
graph has a closed walk whose marker sum, mod 2, is 1. Equivalently,
some closed walk crosses the row-r seam an odd number of times.
-----------------------------------------
16.4 Wrapping addresses
-----------------------------------------
LEMMA 6 (Wrapping addresses). Let |U| <= r+1. Then:
1. Every wrapping subcollection of U has at least r blocks.
2. A wrapping collection of exactly r blocks has one block per row,
lies in a single component, and determines one full aligned
address at that component.
3. Every wrapping r-subcollection of U, if any, lies in the same
component of Gamma_U, and they all determine the same aligned
address there.
Proof of (1). For a closed walk, let c_i in F be the number of
horizontal traversals through row i, mod 2. Summing incidences over
every boundary index i (all corners with that boundary index) gives
c_{i-1} = c_i: vertical traversals contribute in pairs. An odd marker
means c_r = 1, hence c_i = 1 for every i. Every row is crossed an odd
number of times. A block supplies horizontal edges in only its own row,
so >= r blocks are needed.
Proof of (2). If there are exactly r blocks, there is one per row and
the walk meets every one. Two blocks in the same component of the
corner graph; that is one component. A site block gives one horizontal
edge, prescribing one pair of aligned labels. A square block gives two
horizontal edges in adjacent rows; the vertical edges preserve
alignment, so the two edges prescribe the SAME adjacent pair. Summing
the walk's contributions at boundary index i leaves only horizontal
contributions (vertical cancels), so the odd horizontal counts force
adjacent rows' labels to agree. Every coordinate gets pinned. Together
these give a full address.
Proof of (3). Two r-subcollections share >= r-1 blocks. For wrapping
subcollections those common blocks are in distinct rows, so the two
subcollections are in the same component. The common rows touch every
boundary index because r >= 3 (one missing row can only fail to reach
one boundary, and the adjacent boundary still is touched). So the two
addresses agree.
================================================================================
17. TRANSLATIONS ON A SMALL BLOCK COLLECTION
================================================================================
-----------------------------------------
17.1 Building the assignment family
-----------------------------------------
Fix U with |U| <= K. In each component C of Gamma_U, choose a vector
h_t in F^{d_t} per type in its support, subject to the "link equation"
L_{l,t}(h_t) + L_{l,u}(h_u) = b_l for every link l = tu in C
of the type support.
Call these h-values SHIFTS. They are chosen independently for distinct
components, even if the same type appears in more than one.
If C contains a wrapping r-subcollection, use its unique aligned
address a (Lemma 6) and impose one additional condition:
- If the support contains a real type P: require h_P in Q_{P,a}.
- If no real type: require its common internal scalar to lie in
Q_{s,a}, where a is the address read in the source coordinates of
that wire.
If no wrapping r-subcollection is present, only the link equations
apply. Such shifts are ADMISSIBLE.
Admissible shifts always exist. In a wire interior all b_l = 0, so
one common scalar (in the indicated nonempty set when required)
suffices. If the support contains a real type P, choose its vector
arbitrarily (or from the indicated nonempty Q); the endpoint forms
propagate to arms, adding the last-link constant on the end-link. The
scalar is constant along the rest of the arm, and distinct arms have
private internal types, so they do not conflict.
Independently choose a potential M_v in F^{d_t} at every corner v of
type t. Define readout offsets by
x_{vw} = M_v + M_w + delta_i * h_t (horizontal edge in row i)
y_{vw} = L_{l,t}(M_v) + L_{l,u}(M_w) (vertical edge on l = tu)
Endpoints of the vertical edge are ordered as in its readout name.
Let H_U be the family of all readout assignments to U produced this
way. Set H_empty = {empty}.
Every assignment in H_U satisfies Equation (*) at every square of U.
Reason: in a square, potential terms cancel pairwise; the sum reduces
to delta_i * (L(h_t) + L(h_u)) = delta_i * b_l = beta_B.
-----------------------------------------
17.2 Exact projection
-----------------------------------------
PROPOSITION 6 (Exact projection). For V subset U with |U| <= r+1,
restriction to the readout names of V maps H_U onto H_V.
The point: every "old" assignment on a smaller set of blocks EXTENDS
to a legal assignment on a larger one. This is exactly the
"preservation of consistency" property that Duplicator will use.
Proof. Components D of Gamma_V lie in components C of Gamma_U, with
D's type support a subtree of C's.
Step A: Restriction of shifts. A nonwrapping D has no Q restriction.
A wrapping D uses all r blocks of V (Lemma 6), so it exhausts V. The
address imposed on C restricts to the same address on D. If D has a
real type, C has the same one and D inherits the same Q condition. If
D lies inside a wire interior but C reaches a real endpoint, the
address a read inside D is in the actual product D_e; its expression
at the real endpoint uses endpoint matches (a at first, sigma_e(a) at
last). By the image equality equation (**), the permitted internal
scalar set is the image of the permitted real-vector set under the
endpoint form (translated by the last-link constant when relevant). So
restriction is admissible.
Step B: Extension of shifts. Retain shifts when possible, or lift an
internal scalar via the preimage under the endpoint form just
described. Extend along remaining arms. Only one arm is constrained in
the preimage case. Since a wrapping old component consumes all r old
blocks, there is never a second wrapping choice to constrain.
Step C: Preservation of old readouts on nonwrapping components. The
marker on any edge vw in D can be written as s_v + s_w for labels
s_v, s_w in F at its corners: choose path sums from a fixed corner
(well-defined because every closed walk in a nonwrapping D has marker
sum 0). Let g_t = (new shift) - (old shift) at type t in this
component.
The two shifts satisfy the same inhomogeneous link equation, so their
difference satisfies the homogeneous one: L(g_t) + L(g_u) = 0.
Replace the old potential at v of type t by M_v + s_v g_t. Then:
- Horizontal edge in row i: change in offset = (s_v + s_w + delta_i)
* g_t = 0 (because the marker equals s_v + s_w).
- Vertical edge on l = tu: s_v = s_w (markers of vertical edges are
zero, so s_v - s_w = 0), so change = s_v(g_t + g_u), which by the
homogeneous link equation summed with s_v is zero.
So every old readout is preserved.
Step D: Compatibility. Different components of Gamma_V have disjoint
corners, so potential adjustments are compatible even if the
components merge in Gamma_U or share a type. Retain the old potentials
on the wrapping component and choose arbitrary new ones elsewhere.
Together: extension from any element of H_V to an element of H_U.
One application: a wrapping (r+1)-block component with no wrapping
r-subcollection is no exception. Every proper subcollection has <= r
blocks and is nonwrapping; exact projection extends any old choice.
================================================================================
18. TRANSLATIONS BECOME A WINNING STRATEGY
================================================================================
THEOREM 1 (Special-twist sufficiency). Fix r >= 32, K = r+1. Let beta
be the special twist (T). Suppose for every real type P and address a
in {0,1}^r there is a nonempty Q_{P,a} subset F^{d_P} such that for
every comparison schema e : P -> R and every a in D_e:
lambda_{e,0}(Q_{P,a}) = b_e + lambda_{e,1}(Q_{R,sigma_e(a)}). (**-twist)
Then G(0) ==_r G(beta).
Setup for the proof. For each internal type s on wire e : P -> R,
define
Q_{s,a} := lambda_{e,0}(Q_{P,a}) = b_e + lambda_{e,1}(Q_{R,sigma_e(a)}).
These are nonempty scalar sets, independent of the internal type on
that wire.
Ownership convention (identical in G(0) and G(beta)):
- A block's valuation vertices are owned by that block.
- Each bit-value pair is owned by one fixed block using that
coordinate.
- A leaf has its parent's owner and a fixed private index.
Duplicator maintains an assignment in H_U, where U is the set of
OWNERS of currently-marked vertices.
Marked base vertices are matched by translation:
- Block valuation: translate by offsets on the block's coordinates.
- Bit-value: translate by the coordinate offsets on the pair.
- Leaf: same private index at translated parent.
|U| <= occupied slots <= K. So H_U is always defined.
Translation is bijective on each base vertex class with owner in U:
it maps homogeneous legal valuations onto beta-legal valuations, and
permutes each bit-value pair.
Parent translation extends to a bijection on the corresponding leaves
(same color of translated parent => same leaf count).
Partial isomorphism. The only possible base incidence is between a
valuation vertex and a bit-value vertex for a coordinate that block
uses. When both are marked, their owners are in U and the common
readout receives one well-defined offset. So bit-equality and
bit-inequality are both preserved. All other base-class pairs have no
edges in either graph. Leaf attachments are preserved by parent
translation and private indices. Translation within each class is
injective, distinct classes stay distinct, leaves preserve their
indices, so EQUALITY of marked vertices is preserved.
When Spoiler discards pairs, the maintained assignment restricts to
a smaller U, allowed by exact projection.
When a slot is vacant, |U| <= r. For EACH possible owner B of the next
vertex, choose an extension of the current assignment to
H_{U + {B}}. This is possible BEFORE any vertex is selected, by
exact projection. (If B in U, current assignment already works.)
The part of the vertex set owned by B is translated using the chosen
extension. These parts partition V(G(0)), and each maps bijectively to
the identically-owned part in G(beta). They combine into one vertex
bijection, which Duplicator announces. Parts for current owners agree
with the maintained assignment, so it extends all currently marked
pairs (including repeated occurrences of one vertex).
After Spoiler selects a vertex, retain the extension chosen for its
owner B. It agrees with every old marked pair, and the new pair is
consistent. The preceding partial-isomorphism check applies. Duplicator
wins with K = r+1 slots. So G(0) ==_r G(beta).
================================================================================
19. BUILDING THE COMPUTATION
================================================================================
We now specify exactly what the circuit encodes, and how a graph-like
constraint system emerges.
Setup. Fix a binary language L in EXPTIME. Fix a deterministic machine
M with a single doubly-infinite tape. Input occupies positions 0, ...,
m-1. Other cells blank; head starts at 0; moves at most one cell per
step.
Running time: T = 2^(p(m)), p polynomial, p(m) >= m+32 (padding).
Any multitape machine can be simulated on one tape; the overhead is
polynomial and gets absorbed in p.
Also arrange a "frozen-halting" convention: on accept or reject, the
tape and head do not move and the state is preserved.
Parameters for input word w of length m:
d = p(m) + 2,
S = 2^d = 4T,
r = 2d.
So r >= 68. In particular r >= 32: Section 11's reconstruction and
Section 18's special-twist theorem both apply.
Addresses. An address in {0,1}^r splits into two groups of d bits,
least-significant-first, representing a pair (tau, j) with tau, j in
[0, S-1]. Time tau is not cyclic (0 <= tau < S). Position j is taken
modulo S.
Gates. Two types:
- Scalar gate: value = OR of incoming values; true if seed or any
incoming true.
- Two-port AND gate: value = (OR of connections to port 1) AND
(OR of connections to port 2).
One gate of each type per address.
Connection schemas. A "connection schema" describes:
- source type, destination type, destination port;
- a nonempty product source domain;
- coordinate injections from source to destination address.
So its address data have exactly the shape of a comparison schema from
Section 10.
Seed and test requests. A seed or test request specifies a scalar type
and a nonempty product of allowed addresses.
PROPOSITION 7 (Computation). For fixed M and each input w with the
parameters above, one can construct in time polynomial in r + m an
acyclic monotone circuit description with:
- O_M(1) gate types;
- O_M((r+1)^2 + m) connection schemas, seed requests, and test
requests;
- total O_M(r * ((r+1)^2 + m)) constant-size coordinate entries.
Every schema is a nonempty product domain with coordinate injections.
All tests are scalar terminal gates. M accepts w iff at least one test
gate is true.
Proof. Apply the uniform product-circuit compilation theorem
[VariableWL, Prop. 3.2] to M, p. Its assumptions match ours:
- one-tape initialization, frozen-halting convention;
- least-significant-bit-first coordinate groups;
- r = 2(p(m) + 2);
- position ring length 4T = S;
- p(m) >= m + 2 (our padding is >= m + 32);
so the theorem gives the stated counts.
Locally, the circuit uses a fixed transition rule on three adjacent
tape cells. Scalar gates record cell symbols. Two-port AND gates
recognize the triples producing each next symbol. Time increment and
position increment/decrement split into carry and borrow cases; each
case fixes some bits and either flips or preserves each coordinate, so
its domain is a product and its coordinate maps are injections.
Combining time and position cases: O_M(r^2) schemas. Input exceptional
cells plus a product decomposition of the blank suffix supply the
seeds. Final-time accepting symbols supply the terminal tests.
The compiler generates all this from short coordinate tables without
listing full addresses. The true-test condition is acceptance of w.
================================================================================
20. ACCEPTANCE: EVERY EQUIVALENT MATE IS ISOMORPHIC
================================================================================
Now add the "identification constraints."
-----------------------------------------
20.1 Real types and the identification constraints
-----------------------------------------
From the circuit:
- Each scalar gate type: dimension 1, with the scalar coordinate as
both input and output form.
- Each AND type: dimension 2, with the input forms the coordinate
projections and the output form their sum.
Add a new scalar type A, the BROADCAST TYPE. Add also a scalar DUMMY
type. Gate types plus A are the PRIMARY TYPES.
Seeds. For each seed request (scalar type + product domain), add a
comparison from that scalar type to the DUMMY on the same product
domain with identity transmission. It compares the scalar form at the
source with the zero form at the dummy.
Additional DIRECTED connections (added to the circuit; NOT part of
its evaluation):
1. From each tested scalar type to A, on its test domain, with
identity transmission.
2. For each coordinate i: from A to A on the full product, flipping
coordinate i. (The broadcast.)
3. From A to every coordinate form of every gate type, on the full
product, identity transmission.
4. From every coordinate form of every primary type to itself, on
the full product, identity transmission. For A's coordinate,
include TWO DISTINCT copies.
These constraints define identification; they do not affect the truth
values the circuit computes.
-----------------------------------------
20.2 Buffering connections with F_2^2
-----------------------------------------
Each connection schema l - original circuit edge OR added directed
connection - is upgraded to a "buffered connection" using a private
real type B_l of dimension 2.
Suppose l goes from form o on type P to form s on type R, with source
domain D_l and address map sigma_l. Denote B_l's two coordinates by
z_1, z_2; it uses source addresses. Add three comparison schemas:
l1: o at P -> z_1 at B_l (address maps: identity on D_l)
l2: o at P -> z_2 at B_l (identity on D_l)
l3: z_1 + z_2 at B_l -> s at R (address: sigma_l)
That is three comparisons replacing one.
The point is made by two subsets of F_2^2:
{(0,0), (1,1)} coordinate images are both F, sum image is {0}.
{(0,1), (1,0)} coordinate images are both F, sum image is {1}.
Consequence: the source can be a full image while the buffer has small
sum image. In the OTHER direction, if both coordinate images are
singletons, the whole nonempty buffer set is a singleton.
-----------------------------------------
20.3 The acceptance theorem
-----------------------------------------
PROPOSITION 8 (Acceptance). If M accepts w, then every graph H with
H ==_r G(0) is isomorphic to G(0).
Proof. By Proposition 2, write H ~ G(beta). Transport equivalence
across the isomorphism; Proposition 4 supplies nonempty Q_{P,a} with
the image-consistency equality.
Step 1: Singleton propagation across a buffer.
Let l be any directed connection at address a in D_l. Suppose the
source image o(Q_{P,a}) is a singleton. Then both coordinate images
of Q_{B_l,a} are singletons (by l1, l2). Since Q_{B_l,a} is nonempty,
it is a single vector. Its sum image is a singleton. The third
comparison (l3) makes s(Q_{R,sigma_l(a)}) a singleton. The specific
"shifts" the singletons sit at do not matter; propagation is about
size.
Step 2: Induction on the circuit.
At each seed, comparison with the dummy's zero form gives a singleton
scalar image (in fact {0}).
Induction in acyclic circuit order: every true gate has a singleton
output image.
- True scalar gate: seed or true incoming source. Both give singleton.
- True AND gate: both ports true, both coordinate images singleton,
their sum singleton.
Step 3: Broadcast.
By Proposition 7, at least one test is true. The test's added connection
(Item 1) makes A have a singleton scalar image at that address. Flipping
each coordinate connects all addresses in {0,1}^r; each A-flip
connection (Item 2) propagates singletons, so A is singleton at every
address.
Step 4: A forces all gate coordinates singleton.
Each A-to-gate connection (Item 3) makes the corresponding coordinate
of the corresponding primary type a singleton. So every primary set is
a singleton:
Q_{P,a} = { h_P(a) } for every primary P and address a.
Step 5: Row-sum form via self-connection.
Fix a coordinate form s of a primary type P, with its identity
self-connection l. At each address, the first two comparisons give
buffer coordinates, which must be
z_1 = s(h_P(a)) + Delta_{l1}(a),
z_2 = s(h_P(a)) + Delta_{l2}(a).
Their sum over F_2 is Delta_{l1} + Delta_{l2} (the two copies of
s(h_P(a)) cancel). The third comparison (l3), with destination the same
coordinate at the same address as P (self, so h_R(a) = h_P(a)), reads:
z_1 + z_2 = s(h_P(a)) + Delta_{l3}(a) (all at R's coords = P's)
Equate:
s(h_P(a)) = Delta_{l1}(a) + Delta_{l2}(a) + Delta_{l3}(a).
The right side is a sum of functions each depending on two consecutive
coordinates: row-sum form. So every coordinate of every primary h_P is
a row-sum function, so h_P itself is.
Step 6: Buffers become row-sum on their active domain.
For each private buffer B_l on the active domain D_l, prescribe:
h_{B_l,j}(a) = o(h_P(a)) + Delta_{lj}(a), j = 1, 2.
These are the unique coordinates of its image-consistency set on that
domain. They are row-sum form there. Extend each row term arbitrarily
from its allowed pair of factors to {0,1}^2; their sums define row-sum
functions on the full address product. This extension changes nothing
(because B_l occurs only in its three schemas on D_l).
Step 7: Extend to non-primary real types.
- The DUMMY gets the zero function.
- All primary types have singleton and row-sum functions.
On each comparison domain, every scalar form in a comparison has the
value required by (**), and the dummy's zero form gives {0}. So the
row-sum functions satisfy (GLOBAL) (Prop 5). Hence global offsets
exist (Prop 5), hence G(beta) ~ G(0) (Prop 3), hence H ~ G(0).
================================================================================
21. REJECTION: ONE SMALL TWIST
================================================================================
PROPOSITION 9 (Rejection). If M does not accept w, there exists a
choice of square bits beta such that G(beta) ==_r G(0) but
G(beta) is NOT isomorphic to G(0). Hence WLdim(G) > r.
Proof.
Choose b_e as follows: all b_e = 0, except the third comparison
(l3) of ONE of the TWO identity self-connections at A, which gets
b_{l3} = 1. Use the special twist (T): b_l = b_e on the last link of
wire e, 0 on all others; beta_B = delta_i * b_l per square in row i,
link l.
(This is well defined: the two A-self-connections are distinct private
buffers, so we choose which one carries the shift.)
Step 1: Non-isomorphism.
Suppose (GLOBAL) had a solution with row-sum functions h_P, in
particular h_A : {0,1}^r -> F (scalar). At any address a:
- First A-self-connection copy (b_e = 0 for its three l1, l2, l3):
h_A(a) copies to both buffer coords: z_1 = z_2 = h_A(a).
Over F_2 their sum is 0. The third comparison gives
0 = h_A(a) + 0, so h_A(a) = 0.
- Second A-self-connection copy (b_e = 1 on its l3, 0 on l1, l2):
First two comparisons still give z_1 = z_2 = h_A(a), sum 0.
Third comparison: 0 = h_A(a) + 1, so h_A(a) = 1.
Contradiction at every address. By Propositions 5 and 3, G(beta) is
not isomorphic to G(0).
Step 2: Equivalence via special-twist sufficiency.
By Theorem 1, it suffices to construct nonempty Q_{P,a} satisfying
image consistency (**-twist) with shifts b_e.
Define sets as follows.
- Scalar gate type: {0} if the gate is true (evaluated in the
ORIGINAL acyclic circuit, not the added constraints), F otherwise.
- AND gate type: product subset of F^2 with coordinate fixed to 0
if the corresponding port is true, free (all of F) otherwise.
Sum image = {0} iff both ports true, F otherwise. So each
coordinate has its own truth-set and the output sum image has
{0} iff true, F iff false.
- A and DUMMY: full set F at every address.
Seed comparisons. A seed is true, and the dummy's zero form has image
{0}. So the seed's scalar source image {0} equals the image of the
zero form at the dummy, modulo b_e = 0 for seeds. Consistent.
Original circuit connections (buffered). At a directed connection l,
suppose the source image is {0} (singleton). The destination image
under s must also be {0}:
- for a circuit edge: true source => true destination scalar or
input port;
- for an identity self-connection: immediate.
Make the buffer a point:
Q_{B_l,a} = { (0,0) }.
Its two coordinate images are {0}, {0}; its sum image is {0}. With
b_{l1} = b_{l2} = 0 and destination image {0}, all three comparisons
hold.
Every other case (source image F):
- Tests are all false; A's broadcasts have full image; so they
never trigger the singleton-source case above.
- Let D' = s(Q_{R,sigma_l(a)}) be the destination image, nonempty.
- Define
Q_{B_l,a} = { (u, v) in F^2 : u + v in D' + b_{l3} }.
Both coordinate projections equal F.
Sum image = D' + b_{l3}.
b_{l1} = b_{l2} = 0, so all three comparisons hold.
This includes the singleton-destination case: buffer exists as a
full-coordinate set with restricted sum, avoiding any compatibility
requirement on the shifts.
Step 3: Extend by convention.
Buffer sets defined only on their active domains; extend to inactive
addresses arbitrarily (they occur in no comparison). So we have a
choice of nonempty Q satisfying (**-twist) with shifts b_e.
Step 4: Apply Theorem 1.
So G(beta) ==_r G(0). By Step 1, G(beta) is not isomorphic to G(0).
Therefore WLdim(G) > r.
================================================================================
22. SIZES: THE REDUCTION IS POLYNOMIAL-TIME
================================================================================
Counts from Proposition 7:
- O_M((r+1)^2 + m) schemas and requests.
- Added connections: O_M(r+1).
- Buffering: 1 real type of dim 2 per connection; constant schemas
per connection.
Combine:
- O_M((r+1)^2 + m) total comparisons. Since r >= m, this is
O_M((r+1)^2).
- O_M((r+1)^2) real types.
Wires:
- 4(r+1) + 1 = 4K + 1 links per schema; each of dim-1 internal
types.
- O_M((r+1)^3) internal types and links.
Rows, labels, squares:
- Per row and type: |A_{t,i}| * |A_{t,i+1}| <= 4 horizontal label
pairs (since dim <= 2).
- Per row and link: |M_{l,i}| * |M_{l,i+1}| <= 4 matched-label pairs.
- Squares per (row, link): <= 4 * 4 = 16.
- Sites per (row, type): <= 4.
- Base vertex count N, base color count J: O_M((r+1)^4) each.
Total order of G(beta):
- Base vertices N.
- Leaves <= N * J * (N + 1).
- n <= O_M(N^2 * J) = O_M((r+1)^12).
Construction time. All local data is generated from coordinate
tables and adjacent pairs of coordinate labels. Enumerating all
legal valuations + all leaves + all edges constructs the full binary
adjacency matrix of G in polynomial time in r + m. No enumeration of
the exponentially-many full addresses occurs.
Also, r = 2 p(m) + 4 is computed in binary in polynomial time in m.
So the reduction is a polynomial-time algorithm.
================================================================================
23. MAIN THEOREM
================================================================================
THEOREM 2 (Main). The following problem is EXPTIME-complete under
deterministic polynomial-time many-one reductions:
Input: a nonempty finite simple undirected uncolored graph G
given by its adjacency matrix, and a positive integer k given in
binary.
Question: is WLdim(G) <= k?
Dimension one uses ordinary neighbor-color refinement; dimensions
at least two use the joint-update rule of Section 2.2.
Proof.
Upper bound: Proposition 1 (Section 8).
Lower bound: Fix any language L in EXPTIME. Fix a deterministic
machine M deciding it as in Section 19. The construction of Propositions
7, 8, 9 is a deterministic polynomial-time map from w to (G, r).
If M accepts w: Proposition 8 says WLdim(G) <= r.
If M rejects w: Proposition 9 supplies H ==_r G(0) but H NOT
~ G(0), so WLdim(G) > r.
By Lemma 2 (monotonicity), WLdim(G) <= k iff k-WL identifies G. So the
construction maps acceptance in L exactly to the YES instances of
WLdim(G) <= r. Since L was arbitrary, WLdim decision is EXPTIME-hard.
Together with the upper bound: EXPTIME-complete.
================================================================================
24. WRITING THE DIMENSION IN UNARY
================================================================================
COROLLARY 1. The identification problem of Theorem 2 remains
EXPTIME-complete when the positive dimension k is encoded in unary.
Proof. Upper bound: convert the unary k to binary and apply
Proposition 1. Lower bound: the reduction output has r = 2 p(m) + 4,
whose numerical value is polynomial in the source input length m. So
writing r in unary is polynomial-time; acceptance equivalence is
unchanged.
So the hardness is not about k being exponentially large in its
encoding. It is in G.
================================================================================
GLOSSARY OF SYMBOLS AND TERMS
================================================================================
F_2
The two-element field {0, 1}, with addition "mod 2" (so 1 + 1 = 0)
and multiplication "AND" (0 anything = 0, 1 any v = v). Every
nonzero element is its own additive inverse; -1 = 1.
Multiset, {{ ... }}
A "bag": order irrelevant, repetitions counted. {{x, x, y}} has
x twice and y once.
Isomorphism G ~ H
A bijection between vertex sets that preserves adjacency (and
nonadjacency).
Degree deg(v)
The number of neighbors of vertex v. |N(v)|.
k-tuple
An ordered list (v_1, ..., v_k) of k vertices of a graph, where
v_i = v_j is allowed. So there are |V(G)|^k many k-tuples in all.
k-equivalence, G ==_k H
Running k-WL on both G and H produces the same color histograms
at every round.
WLdim(G)
The least positive k such that k-WL identifies G (i.e., every
k-equivalent H on the same number of vertices is isomorphic to G).
Bijective game
The two-player game of Section 4. With K slots, it characterizes
k-WL equivalence when K = k+1 (Lemma 1).
Real type
An abstract "address space" for values. Dimension d_P in {1, 2}.
Comparison schema
A single linear comparison between a form at one real endpoint
and a form at another. Includes an address product and coordinate
injections.
Wire
A schema gets expanded to a wire of 4K+1 links with K = r+1. The
long wire length is why small witnesses cannot span it.
Internal type
A dim-1 vertex type private to one wire. Used to subdivide and to
encode linear constraints via Equation (*).
Readout
A named scalar or vector (x_{i,t,p,p'} or y_{i,l,p,q}) that
interacts across blocks. Block shares a readout across "consumers."
Block
A small local constraint on adjacent readouts. Site blocks impose
nothing; square blocks impose Equation (*).
Shift
In H_U, the vector or scalar assigned to all readouts of one
component of Gamma_U per type. Constrained by the link identity.
Potential
In H_U, a per-corner vector M_v that can be adjusted without
changing readouts (path-sum-preserving).
Offset
A per-readout assignment unifying shifts and potentials.
Wrapping
A block collection whose corner graph has a closed walk crossing
the row-r seam an odd number of times. Uses all r rows.
Row-sum form
A function h on {0,1}^r is of row-sum form iff h(a) = sum_i
h_i(a_i, a_{i+1}) for adjacent-coordinate functions h_i.
Broadcast type A
A scalar type that connects all primary types to all addresses;
used in acceptance to force all primary sets singleton.
================================================================================
BIBLIOGRAPHY
================================================================================
[AKRV2017]
V. Arvind, J. Kobler, G. Rattan, O. Verbitsky.
"Graph Isomorphism, Color Refinement, and Compactness."
Computational Complexity 26(3):627-685, 2017.
doi:10.1007/s00037-016-0147-6. arXiv:1502.01255v3 (May 4, 2015).
[CFI1992]
J.-Y. Cai, M. Furer, N. Immerman.
"An optimal lower bound on the number of variables for graph
identification." Combinatorica 12(4):389-410, 1992.
doi:10.1007/BF01305232.
[Grohe1999]
M. Grohe. "Equivalence in Finite-Variable Logics is Complete for
Polynomial Time." Combinatorica 19(4):507-532, 1999.
doi:10.1007/s004939970004.
[GLNS2025]
M. Grohe, M. Lichter, D. Neuen, P. Schweitzer. "Compressing CFI
Graphs and Lower Bounds for the Weisfeiler-Leman Refinements."
Journal of the ACM 72(3), Article 21, 2025.
doi:10.1145/3727978. arXiv:2308.11970v2.
[KN2019]
S. Kiefer, D. Neuen. "The Power of the Weisfeiler-Leman Algorithm
to Decompose Graphs." MFCS 2019, LIPIcs 138, 45:1-45:15.
doi:10.4230/LIPIcs.MFCS.2019.45.
[LRS2025]
M. Lichter, S. Rassmann, P. Schweitzer. "Computational Complexity
of the Weisfeiler-Leman Dimension." CSL 2025, LIPIcs 326,
13:1-13:22. doi:10.4230/LIPIcs.CSL.2025.13.
[LRS2024full]
M. Lichter, S. Rassmann, P. Schweitzer. "Computational Complexity
of the Weisfeiler-Leman Dimension." Full version,
arXiv:2402.11531v2 (Nov 15, 2024). Journal version: ACM
Transactions on Computational Logic 27(2), 2026.
doi:10.1145/3798282.
[Mackworth1977]
A. K. Mackworth. "Consistency in Networks of Relations."
Artificial Intelligence 8(1):99-118, 1977.
doi:10.1016/0004-3702(77)90007-8.
[Seppelt2024]
T. Seppelt. "An Algorithmic Meta Theorem for Homomorphism
Indistinguishability." MFCS 2024, LIPIcs 306, 82:1-82:19.
doi:10.4230/LIPIcs.MFCS.2024.82.
[VariableWL]
OpenAI. "Variable-dimension Weisfeiler-Leman equivalence on
general and subcubic graphs." OpenAI Math Release preprint,
2026.
[WL1968]
B. Weisfeiler, A. Leman. "The Reduction of a Graph to Canonical
Form and the Algebra Which Appears Therein."
Nauchno-Tekhnicheskaya Informatsiya, Series 2, 9:12-16, 1968.
English translation by G. Ryabov:
https://www.iti.zcu.cz/wl2018/pdf/wl_paper_translation.pdf
================================================================================
END OF DOCUMENT
================================================================================