Milnor Invariants

• Milnor invariants are link invariants defined via the Magnus expansion and nilpotent quotients that measure higher-order linking phenomena in S³.
• They utilize algebraic tools such as free Lie algebras and unipotent matrix embeddings to extract refined invariants from preferred longitudes and linking data.
• Extensions of Milnor invariants include four-dimensional interpretations through Whitney towers and connections to quantum, diagrammatic, and computational frameworks.

Milnor invariants are link invariants that measure higher-order linking phenomena beyond ordinary pairwise linking number. For an ordered, oriented link in $S^3$, they are classically defined from preferred longitudes in nilpotent quotients of the link group via the Magnus expansion, with the length-$2$ case recovering linking numbers and higher lengths detecting phenomena such as the Borromean rings. Subsequent work has recast them in Lie-algebraic, cohomological, diagrammatic, quantum, and four-dimensional terms, and has extended their scope to string links, welded links, clover links, concordances in $4$-space, links in closed orientable $3$-manifolds, transfinite lower-central-series settings, and higher-dimensional spherical links (Conant et al., 2011, Stees, 2023, Komendarczyk et al., 2 Sep 2025).

1. Classical definition and basic structure

Let $L\subset S^3$ be an ordered, oriented $m$-component link, and let $G=\pi_1(S^3\setminus L)$. Its lower central series is defined by

$G_1=G,\qquad G_{k+1}=[G,G_k].$

Choosing meridians $\mu_1,\dots,\mu_m$ and preferred longitudes $\lambda_i$, one applies the Magnus expansion

$2$0

to the longitudes. Writing

$2$1

the coefficients $2$2 are Milnor’s integer coefficients. Their reduced versions are

$2$3

where $2$4 is the greatest common divisor of the lower-length coefficients obtained by deleting at least one index and cyclically permuting the rest. In this formulation, the ordinary linking numbers are exactly the length-$2$5 Milnor invariants, and when the index sequence has no repeated entries, $2$6 is also a link-homotopy invariant (Meilhan et al., 2019, Kotorii et al., 2013, Stees, 2023).

A more structural version packages the first non-vanishing stage into the free Lie algebra. If all longitudes lie in $2$7, then the Magnus expansion identifies

$2$8

where $2$9 is the free group on meridians. The image of each longitude determines $4$0 in the degree-$4$1 piece $4$2 of the free $4$3-Lie algebra, and the universal order-$4$4 invariant is

$4$5

In fact $4$6 lies in the kernel

$4$7

and its coefficients recover exactly the classical Milnor invariants of length $4$8, with repeats allowed (Conant et al., 2011).

For links in $4$9, vanishing of lower-order $3$0-invariants is equivalent to a nilpotent freeness condition on the link group. In the formulation recovered from the homotopy-pushout approach, vanishing of $3$1-invariants of length $3$2 is equivalent to an isomorphism

$3$3

and equivalently to the well-definedness of the next stage of invariants (Stees, 2023).

2. Nilpotent quotients, Magnus embeddings, and refinements

A major algebraic reformulation uses central extensions of nilpotent quotients and a faithful linear representation by unipotent matrices. For a $3$4-component link with free group $3$5, setting

$3$6

one has a central extension

$3$7

Kodani and Nosaka define the unipotent Magnus embedding

$3$8

where $3$9, by sending each generator $L\subset S^3$0 to the upper-triangular unipotent matrix

$L\subset S^3$1

The induced map $L\subset S^3$2 is injective, and the center $L\subset S^3$3 is sent precisely to matrices of the form $L\subset S^3$4. Consequently, the $L\subset S^3$5-entry reads off the central part of the longitude and hence the Milnor data (Kodani et al., 2017).

This matrix model leads to an explicit computational scheme. Labeling the arcs of a link diagram by their images in $L\subset S^3$6, one obtains at each crossing a $L\subset S^3$7-cocycle $L\subset S^3$8 valued in $L\subset S^3$9. Following a longitude around the diagram and multiplying these cocycles yields

$m$0

which is independent of the choices and equals the obstruction to lifting the $m$1-th nilpotent representation to the $m$2-st stage. The $m$3-entry of $m$4 recovers the length-$m$5 Milnor invariant. The same paper defines refined higher $m$6-invariants $m$7 in explicit $m$8-modules; these are universal obstructions to lifting to $m$9, are independent of the choice of section, and can detect links even when the classical $G=\pi_1(S^3\setminus L)$0-invariants vanish (Kodani et al., 2017).

A related algebraic enlargement appears in the theory of $G=\pi_1(S^3\setminus L)$1-reduced groups. For a normally generated group $G=\pi_1(S^3\setminus L)$2, one sets

$G=\pi_1(S^3\setminus L)$3

where $G=\pi_1(S^3\setminus L)$4 is the normal subgroup generated by the $G=\pi_1(S^3\setminus L)$5-th meridian. In this setting, Milnor invariants indexed by sequences in which each index appears at most $G=\pi_1(S^3\setminus L)$6 times are characterized by an Artin-like action on the $G=\pi_1(S^3\setminus L)$7-reduced free group, and for string links equality of all $G=\pi_1(S^3\setminus L)$8 with $G=\pi_1(S^3\setminus L)$9 is equivalent to equality of the induced automorphisms (Audoux et al., 2022).

3. Four-dimensional interpretations

One of the most consequential reinterpretations places Milnor invariants in the topology of immersed surfaces in the $G_1=G,\qquad G_{k+1}=[G,G_k].$0-ball. An order-$G_1=G,\qquad G_{k+1}=[G,G_k].$1 Whitney tower on immersed surfaces $G_1=G,\qquad G_{k+1}=[G,G_k].$2 is a collection of framed, properly immersed disks $G_1=G,\qquad G_{k+1}=[G,G_k].$3 with $G_1=G,\qquad G_{k+1}=[G,G_k].$4. An order-$G_1=G,\qquad G_{k+1}=[G,G_k].$5 tower adds Whitney disks pairing the intersections in $G_1=G,\qquad G_{k+1}=[G,G_k].$6, and higher-order towers continue recursively by pairing intersections among earlier disks. A Whitney disk $G_1=G,\qquad G_{k+1}=[G,G_k].$7 has a relative Euler number $G_1=G,\qquad G_{k+1}=[G,G_k].$8, called its twisting; $G_1=G,\qquad G_{k+1}=[G,G_k].$9 is framed when $\mu_1,\dots,\mu_m$0, and in a twisted Whitney tower of even order $\mu_1,\dots,\mu_m$1 nontrivial twistings are allowed on order-$\mu_1,\dots,\mu_m$2 Whitney disks (Conant et al., 2011).

To each unpaired order-$\mu_1,\dots,\mu_m$3 intersection point one associates a unitrivalent tree $\mu_1,\dots,\mu_m$4, and to each $\mu_1,\dots,\mu_m$5-twisted order-$\mu_1,\dots,\mu_m$6 Whitney disk one associates a twisted tree $\mu_1,\dots,\mu_m$7 of order $\mu_1,\dots,\mu_m$8. Their signed sum defines

$\mu_1,\dots,\mu_m$9

the intersection invariant of the tower. There is a root-sum map

$\lambda_i$0

given on ordinary trees by summing over all choices of root and on twisted trees by

$\lambda_i$1

The central theorem states that if $\lambda_i$2 bounds an order-$\lambda_i$3 twisted Whitney tower $\lambda_i$4, then all $\lambda_i$5 for $\lambda_i$6 and

$\lambda_i$7

Thus the first non-vanishing Milnor invariant is the image of a higher-order intersection obstruction in $\lambda_i$8 (Conant et al., 2011).

This interpretation also organizes the higher-order Arf invariants. At order $\lambda_i$9, all $2$00-invariants vanish on knots, but $2$01 is generated by a single twisted tree, and an order-$2$02 twisted Whitney tower realizes the classical Arf invariant. More generally, in orders $2$03 there is a $2$04-torsion subgroup

$2$05

generated by symmetric trees $2$06, leading to higher-order Arf invariants $2$07 on $2$08 (Conant et al., 2011).

The same framework yields geometric characterizations of vanishing Milnor invariants. Igusa–Orr’s theorem states that a link is $2$09-slice iff $2$10 for all $2$11. A stronger statement says that a link is geometrically $2$12-slice iff $2$13 for $2$14 and $2$15 for $2$16 (Conant et al., 2011).

Milnor-type invariants also arise for embedded concordances between links. For a concordance $2$17 from $2$18 to itself, one defines $2$19-dimensional longitudes $2$20, expands them Magnus-theoretically, and obtains coefficients $2$21. Their reductions

$2$22

are well defined, invariant under ambient isotopy fixing the boundary, unchanged by surface-concordance, and for non-repeated sequences unchanged by link-homotopy. When $2$23 is slice, these become honest $2$24-valued invariants, and for slice links they classify the concordance group $2$25 up to link-homotopy via the reduced Artin-type representation $2$26 (Meilhan et al., 2019).

4. Polynomial, quantum, and diagrammatic characterizations

A striking computational bridge relates Milnor invariants to the HOMFLYPT polynomial. If all $2$27 with $2$28 vanish and $2$29 is a sequence of distinct indices of length $2$30 with $2$31, then after choosing an $2$32-fusion disk and the associated band-sum knots $2$33, Meilhan and Yasuhara proved

$2$34

In the first non-vanishing case $2$35, this becomes an exact integer formula involving $2$36. For the Borromean rings, the only nonzero term comes from the fusion knot $2$37, which is a trefoil, and the formula recovers $2$38 (Meilhan et al., 2010).

At length $2$39, the preceding formula fails in general. Kotorii and Yasuhara showed that one must add correction terms determined by the first nonzero Milnor invariants of length $2$40. Writing $2$41 for a correction term built from HOMFLYPT polynomials of auxiliary knots $2$42 and $2$43, they obtained

$2$44

In the $2$45-component case $2$46, the correction term reduces to a single cubic expression in two linking numbers, yielding an explicit formula for all length-$2$47 Milnor invariants of a $2$48-component link (Kotorii et al., 2013).

Milnor invariants also control reductions of quantum invariants. For an $2$49-component string link, Meilhan and Suzuki defined reductions $2$50 and $2$51 of the universal $2$52 invariant. If $2$53, then

$2$54

where $2$55 is the classical $2$56-weight system. If $2$57, then

$2$58

so one reduction captures Milnor concordance invariants and the stronger reduction captures the link-homotopy invariants indexed by distinct-label trees (Meilhan et al., 2014).

A different line of work gives a purely diagrammatic characterization through welded knot theory. Surgery on $2$59-trees of degree $2$60 does not change the longitudes in the quotient $2$61, so Milnor invariants of length $2$62 are invariants of $2$63-equivalence and welded concordance. Two welded string links are $2$64-concordant iff they have the same Milnor invariants of length $2$65, and a welded link is $2$66-concordant to the unlink iff all Milnor invariants of length $2$67 vanish. The companion $2$68-reduced theory similarly characterizes invariants indexed by sequences in which each index appears at most $2$69 times (Colombari, 2022, Audoux et al., 2022).

5. Extensions beyond classical links

Milnor-type invariants extend to several other link-like objects. For an $2$70-clover link $2$71, one passes to a bottom tangle $2$72 obtained from a disk-band surface $2$73 and defines $2$74. In general this depends on $2$75, but if all Milnor numbers of the leaf link $2$76 of length $2$77 vanish, then $2$78 is independent of $2$79 for $2$80. More generally,

$2$81

is well defined, where $2$82 is the gcd of the $2$83 obtained by removing at least $2$84 indices from $2$85. The same theory gives an explicit description of the possible values at length $2$86 for non-repeated sequences and yields an edge-homotopy classification of $2$87-clover links (Wada et al., 2015).

Covering-link constructions produce further refinements. If $2$88 has $2$89 an unknot with even linking numbers to the other components, then each $2$90 lifts to two components in the double branched cover $2$91, and any choice of lifts gives a covering link $2$92. The set

$2$93

defines covering Milnor invariants. The first nonvanishing covering Milnor invariants are cobordism invariants, and for a Brunnian link the first nonvanishing Milnor invariant is congruent mod $2$94 to a sum of Milnor invariants of covering links; iterated covering links reduce this mod-$2$95 information to sums of linking numbers (Kobayashi et al., 2016).

The theory has also been generalized from links in $2$96 to links in arbitrary closed orientable $2$97-manifolds. For $2$98, letting $2$99 and $4$00, one defines the relative lower-central series

$4$01

A homotopy-pushout tower $4$02 then supports invariants $4$03, $4$04, and $4$05, all concordance invariants. In the special case $4$06 and $4$07 compared to the unlink, the construction recovers Milnor’s original invariants and Orr’s homotopy invariant (Stees, 2023).

Milnor’s original question about transfinite lower-central-series invariants admits a $4$08-manifold solution via Vogel–Levine $4$09-homology localization. For ordinals $4$10, one obtains invariants $4$11 and $4$12 that obstruct lifting an isomorphism one stage further in the transfinite lower central series. In finite length they recover the classical invariants, while at the first infinite ordinal $4$13 there are explicit nontrivial examples, solving Milnor’s transfinite problem in this broader setting (Cha et al., 2020).

A cohomological definition via Massey products extends Milnor invariants verbatim to higher-dimensional spherical links. For a smooth embedding

$4$14

with Alexander dual classes $4$15, a nontrivial $4$16-fold Massey product

$4$17

can occur only in degree $4$18. Under vanishing of all lower-order products, one has

$4$19

and the integer $4$20 is the Milnor invariant. In this setting, these invariants are isotopy and concordance invariants in general, link-homotopy invariants for distinct indices, additive under componentwise connected sum, and subject to sharp polynomial or exponential bounds in terms of thickness (Komendarczyk et al., 2 Sep 2025).

An arithmetic-topological analogue also exists: mod $4$21 Milnor invariants of Galois elements associated to Ihara’s Galois representation recover triple quadratic and cubic residue symbols as mod $4$22 and mod $4$23 triple Milnor invariants of Frobenius elements (Hirano et al., 2019).

6. Examples and structural consequences

Several standard links exhibit the hierarchy of Milnor invariants with exceptional clarity. For the Hopf link $4$24, order-$4$25 disks $4$26 have intersection tree

$4$27

which recovers $4$28. The Borromean rings bound an order-$4$29 Whitney tower with one Whitney disk meeting the third order-$4$30 disk once, producing

$4$31

and hence $4$32. The Whitehead link bounds an order-$4$33 framed tower detecting the Sato–Levine invariant and also an order-$4$34 twisted tower with twisted tree $4$35, detecting the full $4$36. For the Bing-double of the figure-eight knot, the unique order-$4$37 twisted tree is $4$38, and $4$39 detects its nonsliceness (Conant et al., 2011).

The unipotent Magnus formalism reproduces these computations in an explicitly algebraic way. For the Whitehead link, all classical invariants of length $4$40 vanish and the first nonzero term occurs at length $4$41; the cocycle computation yields

$4$42

recovering the well-known value $4$43 up to permutation symmetries. For the Borromean rings, the first nonzero invariants occur at length $4$44, and for the $4$45-component Milnor link the first nonzero length-$4$46 invariant can be written in closed Lie-bracket form (Kodani et al., 2017).

On the level of classification, Milnor invariants organize filtrations and noncommutative structures. For pure welded braids up to homotopy, they are the successive Johnson-type invariants of the Andreadakis filtration on basis-conjugating automorphisms of the reduced free group. In that setting, the Andreadakis filtration coincides with the lower central series, so a welded braid up to homotopy is determined in $4$47 by its Milnor invariants of degrees $4$48 (Darné, 2019). For $4$49-string-link concordance, Milnor invariants define a descending filtration whose associated graded Lie ring is identified with the Lie algebra of special tangential derivations; since this Lie algebra is non-solvable, the concordance group $4$50 is not solvable (Prisa et al., 2023).

These examples and consequences underscore a central feature of the subject: Milnor invariants are simultaneously classical link-group coefficients, obstructions in nilpotent and transfinite towers, higher-order intersection data in $4$51-manifold topology, and classification coordinates in several diagrammatic and algebraic categories (Conant et al., 2011, Stees, 2023).