You are on page 1of 25

Continuity of Solutions of Parabolic and Elliptic Equations

J. Nash
American Journal of Mathematics, Vol. 80, No. 4. (Oct., 1958), pp. 931-954.
Stable URL:
American Journal of Mathematics is currently published by The Johns Hopkins University Press.

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained
prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in
the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic
journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers,
and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take
advantage of advances in technology. For more information regarding JSTOR, please contact
Thu Jan 3 15:18:27 2008



Introduction. Successful treatment of non-linear partial differential
equations generally depends on " a priori" estimates controlling the behavior
of solutions. These estimates are themselves theorems about linear equations
with variable coefficients, and they can give a certain compactness to the class
of possible solutions. Some such compactness is necessary for iterative or
fixed-point techniques, such as the Schauder-Leray methods. Alternatively,
the a priori estimates may establish continuity or smoothness of generalized
solutions. The strongest estimates give quantitative information on the continuity of solutions without making quantitative assumptions about the continuity of the coefficients.
The theory of non-linear elliptic equations in two independent variables
is fairly well developed. (See [I] for a survey and bibliography.) An
essential part is the a priori 13b;lder continuity estimate for solutions of uniformly elliptic equations, first proved by Morreg in 1938. All methods used
to obtain this estimate have been quite special to two dimensions, utilizing,
for example, complex analysis and quasi-conformal mappings (see [2]). The
restriction to two variables has been due to this use of such special methods;
except for the crucial a priori estimate, the theory is extensible (and in large
part has been extended) to n dimensions and to parabolic equations. Our
results fill this gap, and it should now be possible to build a general theory of
non-linear parabolic and elliptic equations, free of dimension restrictions.
Strictly speaking, our work needs some generalization to cover equations with
lower order terms, systems, etc. This generalization can probably be accomplished fairly quickly.
I n this paper, we consider linear parabolic equations of the form

* Received May 26, 1958.

and where 1 T I B in R and d(xl. t ) of ( I ) satisfying 1 T 1 5 B and defined for t >= to is 11 where t . [12]. V T ) = O in a region R and the wme bounds c. from the boundary of R. uniqueness and smoothness of solutions of the general equations of flow for a viscous. Also the relationship between this continuum description of a fluid (gas) and the more physically valid statistical mechanical description is not well unclesstood. Detailed proofs are given and all the results presented in [I41 are covered. will be of value. including continuity at the boundary in the Dirichlet problem.) is the lesser of the distances of the points 2. and this field seems poised for rapid development. and other results. reference [lo]. compressible. (See [Ill. > 0 on the eigenvalues of C so that any eigenvalue 0. It became clear that nothing could be done about the continuum description of general fluid flow without the ability to handle . General remarks. each concluding with the attainment of a result significant in itself.). c. Little is known about the existence. The open problems in the area of non-linear partial differential equations are very relevant to applied mathematics and science as a whole. 2 t . If T ( z ) satisfies V . ( C ( x ) . and c. > to.c. and [13]). however. 2 c. An appendix states further results. c. stated without detailed proofs. we obtain a continuity estimate for solutions of elliptic the C. Here A and a are a priori constants which depend only on c. form a symmetric real matrix C(X. This is the standard '' uniform ellipticity " assumption. and the space dimension n. and heat conducting fluid. limit the eigenvalues of C ( z ) . satisfies c. These are a non-linear parabolic system of equations. that fresh methods must be employed.z. 11-e assume there are universal bounds c. perhaps more so than the open problems in any other area of mathematics. The continuity estimate for a solution T(x. As a corollary of our results on parabolic equations. then where a is the a of ( 2 ) and 8' is an a priori constant d'(n. We hope this paper contributes significantly in this way and also that the new methods used in our previous paper. a Harnack inequality. I t seems clear. Our paper is arranged in six parts. An interest in these questions led us to undertake this work... (= 0.. and x.t ) for each point x and time i. and c.

Levinson. (See [3] through [Y]. such as ( 2 ) . max I T (z. such as infinities of temperature or density. the Alfred P. R e are indebted to several persons and institutions in connection with tllis work..(x. conditional on the non-appearance of certain gross types of singularity.. Morrey. We had to work hard to get (13). c ..17. including Bers. . (Kroneclier delta) for 1 x 1 z r . illore than enough is linown about linear parabolic equations with variable (oeliielents to aqsure the existence of well behaved solutions for equations of the form ( 1 ) IC me make strong (qualitative) restrictions on the C. but the ritual of mathematical exposition tends to hide this natural basis. and in applications an elliptic equation typically arises as the description of the steady state of a system which in general is described by a parabolic equation. Sirenberg. This should give existence. Brownian movement.. Part I: The Moment Bound. the Institute for Advanced Study. For parabolic equations.. then the rest followed quickly. We consider only solutions T ( x . t ) = q c . K. Moreover. T. t ) bounded in x for each t for which the solution is defined. 6. and unique continuation (in time) of flows. Carleson.and the Office of S a ~ aR l esearch. and flow of heat or electrical charges all provide heIpful interpretations. I n our work. i. t ) 1 is finite. Only in one dimension would the situation simplify.) A result of this kind woulcl clarify the turbulence problem. e. diffusion. Probably one should first try to prove a conditional existence and uniqueness theorem for the flow equations.) fore we asiume : ( a ) The C. Stein and Wiener. The methods used here were inspired by physical intuition. ]. d . M. from a converging spherical shock wave. to us. smoothness. for example. The key result seems to be the moment bound (13) .. (b) C. Sloan Foundat1011. Reurling. 1.1'. Browder. some large constant. no difference a t all appears between dimensions two and three. It is certainly true in principle that the theory of parabolic equations includes elliptic equations as a specialization. ( A gross singularity could arise. parabolic equations seem more natural than elliptic ones.non-linear parabolic equations ancl that this in turn required an x priori estimate of continuity. and Thererestrict the class of solutions to be consiclered. i t opens the door to the other results. t ) are uniformly C " .(x.

and the a priori estimates then hold for the generalized solutions. where dn: is the volume element in 11-space. for fixed x and t. It is also known Z that fundamental solutions. to) of :c given at an initial time to determines a unique continuation T(x. t ) 1 is non-increasing in t. t. exist and have the general properties we state. any bounded measurable function T(x. t h e fundamental solution concentrates around ite source point . mental solutions. and max I T (x. The final result requires only measurability for the Cij. For fixed Z and t and as a function of n: and t. These are standard relations. [ 7 ] . t o ) almost everywhere as t += to. ( 5 ) reveals a reproductive property of Punda- . Dually. t ) on bounded initial data T ( x . f). T ( z .X] =-aX/i?t. where time runs backwards. L \ S t + to. lim T (:c: t ) is zero unless-x = x..t o ) is expressible through X : in particular. Moreover. S is a fundamental solution of the adjoint equation: V. This duality enables us to use estimates for fundamental solutions in two ways on 8. a passage to the limit can remove the restrictions on the Cdj.. J T ( z . t ) dx = 1 for every t > to. [C (2. Also. V. in which case it is + a . a fundamental solution represents the concentration of a diffusant spreading from an initial concentration of unit weight at xo at time to. Furthermore.) After the a priori results are established. 2. The use of fundamental solutions is very helpful with equations of the lorn1 ( 1 ) . t ) defined for all t 2 to and Cm for t > to. the maximum principle remains valid and with it the unique continuability of solutions bounded in space. Our work is built around step by step control of the properties of fundamental solutions and most of the results concern them directly.. The dependence of a bounded solution T(x. and is defined and positive for t > to. (See [4]. Physically. t ) += T ( z . t ) has a "source point" xo and "starting time " 1. All fundamental solutions are conveniently unified in a "characterizing function" S ( x .Under these restrictions. which we discuss below. This is a standard device in the use of a priori estimates. The Holder continuity ( 2 ) makes the family of solutions equicontinuous and forces a continuous limit (generalized) solution to exist'. t ) . Ll iundamelltal solution T (x. plus the uniform ellipticity condition. X is a fundamental solution of (1) with source point 2 and starting time 5.

Thi? has the familiar property (1 v I2dy= J /u I2dx. If we choose the value of p minimizing the sum of the two bounds ( a ) and (b). lor any p ( .O) with source a t the origin alid starting time zero. Stein gave us the quick proof which follows below. 1 v 1 5 ( ~ T ) . 1 V l2 5 V . we have uriiig the formula for the volume of an 11-sphcrc. we shall be able to bound B above. therefore - For ally vector V .( 1 ?( / dz.936 PARABOLIC AND EI. we assume u is smooth and well behaved a t infinity.f / V u 1' dx. t. this is . obtaining our first a priori estimate. The trailsforin of du/dxk is iykv . E. 011 the other hand.L~YTIC EQUATIONS. V With ( 6 ) and a lower bound for J I V T l 2 dx in terms of E.~1 /e i~. hence and J / du/axk / J /V d~ = j y k 2 1 v l 2 dy. U / ~ ~ Z i = ~ S ( ~ U / ~ . 0. > 0. we obtain a bound on J I ~ / ~ d y = J I u 1 ~ dinx terms of J I u / d x and . Let by integration by parts. Z . ( 11 1 dx == (2x)-42. To bound f / V T I dx we employ a general inequality valid for' any function u (x) in 11-space. 8 J therefore. C . Now consider a special fundamental solution T = T(x. 11. Ic. For our purposes. The Fourier transform of u ( x ) is v(y) = ( 2 ~-n/2Jeix ) %(x) dr. I V / . we have c.= S / ~ Finally. t ) = S ( x . Solved for f 1 V u 1' dx. l~2 )l V~ l 2 ~ d~y .

whence Therefore (8) which is a pointwise bound. (9) From ( 8 ) ) Q 2 S m i n [.~/R.O)~B.~/~)S(~. and c.~. . (E-2/n) 2 k .O. from the above inequality. hence E-2/n 2 kt and We used above the qualitative fact lim E =cc. This is the first use of a convention we now establish that k is a generic symbol for a priori constants which depend only on n." Q=-JTlogTdz.log T I ( T dx) 2 -log G ST dx. The key estimate controls the "moment" of a fundamental solutioll To prove 41 5 kt2 is our first major goal. This is dimensionally the only possible form for a bound on Ill. hence (10) because S T dx = 1.. t+O From this first bound ( 7 ) and the identity ( 5 ) ) we obtain T(x.t) is obtainable from a more sophisticated argument. stronger than ( I ? ) . remembering that f Tdx = 1: me obtain from ( 6 ) -Et 2 kE1+2/11. c.~. Any two instances of lc should be presumed to be different constants. Thus. The sharp result Q 2 i n log(hec. We also define an "entropy.Applying the above inequality with ZL = T..t) =JX(X. The moment bound is essential to all subsequent parts of this paper.

This inequality. or -Q + aM + b 2 - e-b-la-"D n7 where U . V T 1 dxI2. This can be rewritten Since in general V . S T dx = 1. V ( ~ O ~ T ) / ~1 ( T ~ XC . which is N 2 keQln. V T 1 dx..C . gives an optimal constant. is the well known constant 2 n ~ & ( n -[i l ) ( n . C 2 .Q n b 2 . Then . . where V is a vector. *1 Here we used the Schwan inequality in the form f2 0 dzi du 2 [S 1 f du]' with 0 corresponding to T dx. Q log(n/Dn) log (n/&l).937 PARABOLlC AND ELLlPTlU EQUATIONS. an. after integration by parts. obtaining T= IxI and a and b are any constants.( n r + b ) T I dx 2 -e-b-I $ e-ar dx.V = I C . due to L.) . + Let h = ar b. Ry analogous manipulations. v T / d s . First observe that for any fixed A.1) -sphere. Differentiating ( 9 ) . depends only on the facts T 2 0. thus n log JI n Q n log n -log D. Now set a = n/M and e-b = (e/D. + + + + > + + + This ingenious proof.) v l o g T I ( T d x ) I 2 2 [$I C . Carleson. Jft=-$Vr. connecting the rates of change with time of A1 and Q. finally. V 2 V . V have / 2. where orer space.~Tdx and j M t / s $ I V r j / C . The next inequality is a " dynamic" one. we [$I c z Q t 2 $ / C . and integrate $ [T log T $. Our derivation of a bound on &I requires a lower bound on M in terms of 0 as a lemma. /&It 1s $ 1 C .1)] ! related to the gamma-function and the surface of the ( n .1 or n 1 >. hence.C. c.

Combining inequalities. Clearly keR increases faster in R than 25(1+ 4R) so that R must be bourltlcd above. (12) This is a powerful inequality. b/2a4. Applying this result.)i When a and a 4-b are positive (a (1/2t b)a saa Rt)a dt. and we obtain + ktW + 1' + + M 5 (nc. Therefore M / t a is bounded both above and below: . No further reference to the differential equation is needed. Then Qt = nRt n/2t. czQt 2 (Mt) '. as t + 0. From iM(0) = 0 and ( 1 2 ) ) ahence Kow define nR = Q T k -$n log t in such a way that R 2 0 corresponds to (10). as functions of time. Q is defined as it is in order to obtain (12). suffice by themselves to bound above and below both 31 and Q. The three inequalities QZirk++nlogt M 2 IceQln c2Qt 1 (Mt) * and the qualitative fact lim M = 0. hence Here we used integration by parts and R 2 0 in the second and third steps.

Let T be S ( x . we obtain 2 t Q t = H l + H z + H. T. t ) . These tend to make G strongly negative. of two fundamental solutions with nearby source points. Furthermore.T(t&t. where 8 is a small positive constant../cl). This bound limits the possibility for U to be small in a large portion of the region where I. For comparison. the bounds sharpen with increasing a .. Differentiating (16) with respect to time and using (15). G is sensitive to areas where / [ / is not large and U is small.$.O) and let (14) Ti(. = c. This coordinate transformation and renormalization makes $U d [ = 1. we can obtain where and bn = (n/2t)*{d/[*(n- 1 ) ] !)lln 2 2-lIzn + x = -&log(c. Jf = (2nct)a in the simple heat equation where Cij = csBgand c. Part 11: The G Bound. VCT) / 6 /2)10g(li+6)dt. We later obtain a lower bound on G of the form valid for sufliciently small 8. ( C .$ I is not large. indeed.)dx. Here we obtain a result limiting the extent to which a fundamental solution can be very small over a large volume of space near its source point. where d[ is the volume element. 0. starting simultaneously. we can show there is some overlap.t. Thus h is relatively small. 930 If we use best possible constants in (10) and ( l l ) .t ) = t"/. For U.) -log b. (16) G = Sexp(- V7i + 2 V . . equation (1) transforms to ( 15) Let 2tUt=nG+(. = c. we have S / [ I U d[ 5 p./c. From U > 0 the weak lower bound G > anI2logs follows immediately. From this result. Since bn+ 1 as n +oo. 5 (l/2n)log 2 + log(c. defined as Smin(T.PARABOLIC A N D ELLIPTIC EQliATIONS. if p is the constant such that ?ilzpte. they seem surprisingly sharp.

VU)/(U+6)dt =-2SV[exp(-jtj2)/(U+6)l. XASH. finally H3=2Jexp(--j tI2)V. hence. by integration by parts. and H.940 J.Vlog(U+S)dt. ( 4 ~$ n~x . where H3r=4Sexp(-It/2)t. Furthermore. so that H2=-Sex~(-jt1~)(V.C.(C.: we obtain . Hz.C.") a.i n ) . (17) H 3 " 2 2c. H2~-nG+210g6J/~/2exp(-I~12)d~~-nG+nx1L/210gS.=nJexp(-/tj2)U/(U+6)dt20. t l o g ( U + 6 ) d t jt j2)1 t 12[logS+log(l+ U/a)ldt.Vlog(U+6)dt.C.jexp(- j t 12)1 V log(U + 6) l 2 d t Combining the lower bounds available for H I . where H. Hence lH3 /-===k (H.C. 2H3" = 4nc2ni"H/. H3rr=2Sexp(-j tj2)V1og(U+6) . From the Schwarz inequality.t)10g(U+6)dt +Sexp(-jtI2)(2 =--nG+2Jexp(-- It1 V 161) .VUdt tl2)[1t(V I tI.VUl/(U+6)dt =4$(exp(-I It 12)[ V U ' C ' V U l / ( U + s ) 2 d t +2Jexp(= HQ' + H3".

of the form fl Huci.) The equation for the maximum point U ." . quantity has a lower bound of the form E[log ( k 6) -kG] "or Applying this to (19). from which U.~ / ~ 2. say. U as large as possible.-. Hence the U 2_ Uo. where the polynomials are Hermite polynomials. (We know l0g6-a-n/~G < 0. T 5 kt-@. so that [log k -kGI2 < [log (k $6) . .s2)H. 2tGtzH.16 (H. and we can simplify the above inequality on H. and U* = 0 for U 5 U. ( I S ) will yield a lower bound on G. For sufficiently negative G. < Uo = exp (2 T-"/~G).kG)2Jexp(- (20) Let fore = J U* j [ 12)U* d[. rises from zero to a local maximum at U = U.. is log (U. A function f ([) = f ([. (s) obtains. with f= log(U + 6) -T-"/~G."): When we bound H3/' below in terms of Cr.i d t 5 p. is large for very small U.nG + nahnlog 6) + H. tz. in the remaining -region./ [ 1 ? )f d." to the form + H. . af/a& depend very simply on the coefficients in the expansion of f .kGI2. we may say + + + + where U*= U for U > U.. the coefficient of fl Ho(ti) is zero and we obtain Applying the above. we obtain + The quantity U-l [log ( U 6) -T .(s)/ds = (2v)*H. then decreases to zero. Therefore the quantity under discussion is decreasing for U 2 U. If Jexp (. 6) .. dz and observe that J 1 [ ( U* d." 2 (lc.~. The identity dH. say. corresponds to U 5 lc.t n ) may be expanded in products of ( t i ) . and taking the worst case.T . .t 5 $ I 6 1 I. The bound ( a ) . and finally decreases monotonically as U +co for U 2 U.+H2+H.G ]related to the integrand in (19)."-jHij 2 0 + (. There- .PARABOLIC AND ELLIPTIC (18) EQUATIONS.. (s)HA(s) ds = 6.t ='0.~ / ~ G= 2UC/(UC S ) . defined and orthonormalized so that 11" exp (. to ( l 7 ) . the expression log(k 6) -kG will remain positive when 6 is omitted. and the coefficients of these products in the similar expansion of. Thus we are ignoring the contribution to (19) of the region where U 5 U.

Now from (21). 1 U* d[ and so lSIZzs/A S u * d [ ~ x-~A=+x lElS2~/h This result can be applied to (20) and yields H.3) Gz-Ic(-log6)* + . Returning to inequality (18) controlling Gt and applying the above result.. c. n. which implies G+ -co as t + 0. for / t / 5 p and i? = 0 for 1 [ 1 > p. If we had G(tl) = Gs -C.. Of course. we know J 1 [ 1 6 d[ 5 p because i? 5 U. the maximum of d[ is clearly constraint and the constraint realized by having 0= U . (kp/TJo) n/(n+l) or 1 -h 5 .U* so that 0= 0 unless U 5 U. G. for all sufficiently large . n) be the number such that wlien G 5 GI. we have H3" 2 (k -kG) for sufEciently large . c. /2) I -h 5 U. (c. or (2. Let G1(cl. the hypothesis G(t..?~U~l/(~+l). where p is such that 50 This makes 1-A = Jr:' d[ = [ ~ ~ / ~ / ( n !]pnUo. we would have dG/d(logt) z e * for all t 5 t.) = G3. ($A).hence. we can state that f o r sufficiently negative G. Let G. we know G is small enough to make (22) valid..exp (- (21) (2p/A) 2 . Then largest number such that k 1 G l 2 min(Gl. Under the moment 5 U. in which case 0 = U. then 1-h is small and h is bounded below. 6) = -k (.) = G3 is the smallest possible value of G.E is impossible." )= (k . But since G 2 rrnI2log 8. where 6 = U .kG) 2 ..G. and consequently G ( t ) 5 G(tl) -r* log(t. which is exp(2 +n-"I2G). or bound 0d[ = 1-h above.G../t)..log 6)h be the k log8 > 0 for all G < G2. If Uo. Our corlclusion is G 2 G. Thus h )= 4. is small enough. 5 This is not effective unless we can bouncl h below.say.

-k(-logs)$.. = tn/2Tl(t%[. 0) and S(x. j'? log(Umin + 6) d t 5 log 657 d l + mas [f] Slog (1+ lJ. or (24) Jmin(Tl. t.t ) and 11. which n-a5 the source of U in Part 11. x.PARABOLIC A N D ELLIPTIC EQUATIONS. Let [. rather than the origin.. ohtaming It-[. .iin/s) dt 5 w log 6 + 6-I J U. We abbreviate the above to S f " log(U. Part 111: The Overlap Estimate Let 2'.. Here the source of the (renormallzed) fundamental solution Ut is (.exp(- I t-t2 l2)1dt. 2 and 6 must be sufficiently small. T.fficientlysmall values of 6. defining U. For the first integral. be two fundamental solutions S(x. I2). t ) . we observe (assuming 6 5 1) S f * log(Umax + 6)d&5 Jf*(Ui + Uz)dt 5 S (U1+ U2)dt =2 . - Tvhc l. Change coordinates.. and when 6 is small enough. s G.= x. + 6)d[= G.)dx = fmin(U.. Uz)d t 2 -k (- log 6)*. and T . Taking this into account. for all s.(~ w = Jmin[exp(- I [--t./t* and [.>. because G . we apply (23). U2)dt >= 6[- 2 -w log6- k(-log6)$]. d[..ilhcle z = 1 or in which nre Iorm two integrals with sum at least as large as the sun1 of the orjgirial integrals. We may add the inequalities ni~obeand obtain . For the second integral. Therefore we obtain 2 + w log 6 f 6-Ismin (U1. [exp(- /"log(U. t.. 0) wlth nearby sources..u. +S)d[ 2 - is(-log6). x2.. = x2/t*. + 6)d(+ ~flog(U. = tnl2T2(tst..

t)X(x. Also.Z.-t2 1) 2 + ( / xl-2. . t ) -X(x'. Then defining A ( t ) . T. t'. there is a value 6. t ) be the bounded solution in d and t' of (1) defined for t' 2 t and having the initial value T.). = T." = T I-T.+ ( / x1-xz 2min(T1.- TI-T. V e can obtain a stronger inequality by iterative use of (25). t ) = JS (x'. c..t)dxdz.~. Let X(~. Observe that (26) +$I I dx=~j[T. t. we obtain . t'.I .T2)dz2+(I (25) [. t ) T. I and S ( T . t ) I x ( x . t ) . determined only by c.). It is an a priori function.. 2. This inequality (25) is our first estimate on the overlap of fundamental solutions.x. The function is decreasing but always positive." . the right member of (24) is definitely positive. Let T. at t'= t.(x.)dx=O. If we set 6 = & min (a.*(x'.1. t)dx dZ by the superposition principle (T.-[.Ti. the bracketed expression is positive. l/ts) + because zu is a function of I [. t'. Then from (4)..T 2(x'. = I TI-T. x. and TI (z'.~ .(w) such that for 6 < S. t ) = T.T b ) d x = S(Tl-T.. and n. 6. say for 6 5 61. t ) dx = jJs(x'.-T2. x . Z') .(x. + T .* (x'. -T. Tntegrating this over d d ."(z.(w).This is valid for sufficiently small 6. Let T.+T.Tb* are both solutions of (1) for t ' z t. t'. t') = SJ [(x'. To*. Define Thb similarly. t )L=Ta(x)Tb(Z)/A (t).Tb* f . and by definition.O) so that T . Part IV: Continuity in Space.=max(T. 5 . which is increasing but always less than one.0) and Tb=max(T2-T. Its weakness is that we know little about the function 4..t'. and Ta" .T2)]dz l/ti) =$(I $1-x2 l/th) ill which we define the function $. and we may conclude Jmin(Tl.

exists.' = 0. and T.. To begin the iterative argument. as follows : for / x -so / 5 ~ C .') and consequently.t'.) = a. t') JJJI I I dx' S(z'. This is in reference to a specific pair.' > 0. Incidentally.' > 0 and Tbr > 0 SO that both I x-x. 5)) d ~ d Z +(4u-~Jl.' = T. Let Ma ( t ) = J / x -x.T.)].T i )dx 5 +avand ST.. J (Ta. 5. t. we would choose d with l'egard to an explicit formula for ( d ) so as to optimize the result. we know both T. let t. solutions. of fundamental I2/d2. is 4 (x.t)dxd~ by application of (26). we also know that x 2 and $ < 1.Z. (If we were trying to get an explicit formula for the exponent cc in (2). whence A(t') (27) 5 JJ $(I z-5 l/(t'-t)~)x(x..)./(t' -t.x. Z)] dxd5 d~db. (z'.--T. For each integer V . we choose any specific number d and let € = + ( a ) = 1-$(a). / and / Z -so I are 5 2u-Ydl.' and T.-T. that t.)h[{X(x. Now. We decompose T u into nearer and farther parts T. the right member above is thus (t') (= A ( t ) when t' 2 t. otherwise T.t'. M b (t. $ J 1 T. SO that A ( T ) < A (t. This inequality (27) is the key to the iterative argument which strengthens (25) and (26). be the time (or the least time) at which A ( t ) = A (t.t) -S(x'. )(T. the midpoint of the line segment joining the source points x1 and x.) Let a = 1 -4 4 . / x .)a)J Jx: + x:(x.t)dxdZ.) = uV.and .~. for example. (z'. Proceeding further.. T. We know. < T. and consequently.. Z) =U-~T. T.). because when x..e / 4 > 1-E.-x. t') .. x. Define Jib similarly and let i&f. ~V-~M. 1 Tad%( M. applying (27) with t = t..=max[dla(t.' (x) T{ (5).' dx 2 $av.') d x 5 J j z-x. Then d x z J j z-X.t)j d~'~( T1 and T. of the fundamental.~ Mdefine .T d at each time t .Z 1 j4 ~ . we can say + + I -4(t') 5 J J$(( x -2 /)/(t' -t. where x..J(T. where T = 1 2. Tad$. . if t.~ d l .) . Define T{ similarly and define Xi (5.J J {X5 + -Xi(x.5.945 PARABOLIC AND E L L I P T I C EQUATIONS. because A(7) s $ ( / x1-x2 / / ~ h ) = $ ( d ) =I-< and a = 1 .

t') .(tr) 5 J /x-x.) of moments. = uVt1.t)JS(~'. M b ( t v )).-so I = I XZ-xo I = & 1x1-$2 I since x. 0) =max(Ta*(x'.) . + We now set t'= t.x. = (x.t)dx+~(t'-t)*ST.t'. use a similar estimate for iMb and the definition M.t'. t ) .) of times after we obtain a bound on the sequence {&I.(x. and 1 2. t ) dx.t)Ta(x. 0) = 5 T a * ( d .t)dx'dx: or Jf. because T1 and T z concentrate a t x1 and x. t'. t) JS (x'.t)j'l 5'-x I S(x'. 1 6 ~ .t'. Now let t and t' be t.b above becomes d.946 J.1-r.T. Hence tVtl 5 t' (28) = t. t') = max ( T I (x'. and t... ) ~ and d . (x'.) can be bounded: . = max(Ma(t. as t-0. Observe that Ta(xr. t'. 1 T. t ) . Therefore Ma (t') = J I x' -2.~.).t)dx 5 Ma(t) + A (t)p(f-t)'. t') dx' 5 . and obtain. (x'.. I Now to= 0 and Mo =Ma(to) = Ma ( t o ) = 4 x1 -x2 I.(x.JJ[1 2'-x 1 + 1 X-x0 /]S(x'. by (28). With this and (28).Tb*(xr?t'.~ v ( M . NASH. Therefore we have + + Jl. t'). t'.S3 I x1-xz I(1+ 4 ~ / d ) ~ . x.t)d~'d~ + STa(x. hence Ma(tr) J /X-xo I Ta(~.t)dxdx'.~ the .(~. x. + 1 6 ~ -(' M~ y'd-'. t ) T. I T. we obtain A (t') 5 5 ~ [ 3 / 4+ 1/4 ( 1 -e ) ] ='a ( 1 -6 4 ) t. argument of Then since $I (d) -. the sequence {t.x. ) This will bound the sequence {t.

I/(t--t~)~)~. and c. even if we set d = 1.t)/_I 1 ~ [ S ( X ~ . say.PA4RAGOLIC AND ELLIPTIC EQUITIONS. S o w for any time t. $0. 5B J I S (x.. to) -fl(x2. this will make a a function of p and c. to) ( dzo.S ( X ~ ~ ~ ~ X O . Specifically. . X O .t)I SBAI((XI--X. Hence. Also. I n any case.t)--T(~?. we obtain the estimate for the continuity in space of a bounded solution of (1).. d2= c. c. From these observations./cl (proof omitted). then I T(xl. ) be either zero or the (definition of [. If T (x. t. Then tV(t)5 t and A ( t ) 5 A ( t v ( t ) )= d t ) . -x2 I Z < [17v(t)tl if' this integer exists. \Vith ( 3 0 ) . Summing this geometrical series. ~ ) . tz. Both CJ and 7 are determined by d. ~ O ) I T ( X ~ . XO. and a are a priori corlstants depending only on for the dual adjoint equation. Also. ~ . We may choose d arbitrarily as. define ~ ( t to integer such that [rlv(t) 5 t/l 2. t ) -T(x.7). t. a = 1-& ( d ) -and An optimal choice of d i n relation to 4 ( d ) would 7 = [a-2( 1 4/*/d)]" maximize a... (31) 1 T(xl. t ) satisfies (1) and I T I 5 B for t 2 to. we obtain the estimate + where A. we conclude where +a = -log c~/log7.

Therefore.[l(t'-t)/(t + where A.plta = 2p(t'- t ) $ ( t . and if we choose p so as to minimize the sum. . can be derived from (31) and the moment bound (13).). in terms of a radius p . and we obtain (32) I T(x. -t0)1'~'(~'~). (31) gives half of ( 2 ) . where (because JS dy = I ) . time I T (x. A. t ' ) l ~ I l + I . = (1 a ) A. t') I 5 Sow we separate this integral into two parts. Then for t' > t > to we have I since ('8 d? = 1. t ) -T(x. and the two inequalities.T ( Z . t ) . then aA. the remaining part..T ( x . t') I 5 BA. Let T ( z . combined with (31 ) . one where I y 1 5 p a n d o n e . Thus I T ( x . with A =max (A. t ) . ) yields ( 2 ) . ( 2 ~ / a A . This result (32). w h e r e / y j > p .Part V: Time Continuity. t ) be a solution of ( I ) with T I 2 B for t 2 to.

r . The Holder continuity of solutions of unlformly elliptic equations of the form V . V T ) = 0 appears as a corollary of the result for the parabolic case. Define 8 * as the total boundary of 9. Let 9. zo. We can define a boundary value problem for which we know the solution in advance by setting T ( t ) = S ( x ( t ) . also. t ) be determined in this way : (33) T(x. Let 9 be a domain in space-time defined by the constraints 1 x 1 5 and t 0. reference [9]. R. t ( t ) . we can obtain a moment bound for dp.PARABOLIC AKD ELIJIPTIC EQVATIONS. where I x I = U. one in which the coefficients of the equation are time independent and a time independent solution is sought. t ) is a positive measure. the maximuin and minimum principles must hold. Garabedian writes from London of a manuscript by Ennio de Giorgi containing such a result.. t. 2 ) is a poilit of 9. and from (33). Then the solution of the problem is S ( x . These facts require that the solution T ( x . Then 9 is a solid semi-infinite spherical cylinder. I and integrating. i t enables us to convert information on fundamental solutions into information on clp. Call 8 the points of the cylindrical surface or boundary of 9.x. f is any point of 8".ancl dp([. which has S d p = 1 and which vanishes for t ( t ) > t. (r > A "Dirichlet parabolic boundary value problem" is given when values of T are specified on 8" and we ask for a solution of (1) in 9 assuming these specified values on 8. the union 8 U 9.t). x. The time and space coordinates of the point ( are called t ( ( ) and x ( t ) . and in particular. associated with t. be the points of the base of 9.. See de Giorgi's note. t o ) if to < 0. Multiplying (34) by I .r-x. ( C . but refer the reader to the literature. t o ) . We treat elliptlc problems as a special type of parabolic problem. Ilere (x. The solution of the problem must depend linearly on the boundary values. This is a powerful identity.t) =JT(t)dp(t. There inay exist another proof of our result ( 3 ) . P a r t V I : Elliptic Problems. of the base and cylindrical surfaces. P.where t = 0. we have . ST'? cannot pause here for a detailed justification of (33).

1. I) that is. T ' ( x .. (G'(z) . 2. and setting T'(x. also. Thus S o w let T ( 2 ) be a solution in a region aZ of: n-space of V . t ) satisfies V .Another problem with solution T'(x. which is of our form ( 1 ) . F o r any u < d(x. can be defined in the obvious way. I I I (T .). then T ( x . d ( z l .. V T ) where C ( z ) satisfies tlre uniform ellipticity condition wit11 bounds c. 0 ) 5 B ( v ) .) be the smaller of d ( x l ) and d (x. l ) = T ( s ) . If B ( o ) = max T ( x ) over the set of x values where min ( x -z. 1 a. the solution T'(z.). where (x. t ) is in 9. 1) and the boundaries can be regarded as a solution of a parabolic boundary value problem either in 9 . or 9.. Hence we can simplify the above t o This inolnent bound (35) on dp enables us to control the relative sizes of the effects of the two parts of the boundary i n determining T ( x . 0) = 0 for all other x values. (x. and 6.. (of course. x -x. t h a t Since JS dxo = 1. If we introduce time and define T ( z . ( C . and let d(:cl. 9. V T ) = Tt. I I I 1 so. the are two points of clistarices froin the boundary of aZ of x1 and x. etc. $dp = I. T ( x . t ) .. we can define 9. as the set of points (x: t ) in space-time where 2 -x1 1 5 0 and t 2 0 . Suppose z. ) ( the11 j T'(x. furthermore... t ) = T ( x ) when all a where min(1 x-x. x. we obtain Yow d p vanishes unless t ( < ) 5 t. 0 ) = T ( z ) for x-x. a. can be defined for x. t ) satisfies 1 T' 1 = 0. and to can be as near to zero as desired. 1... and z. a. and from the moment bound (13) again. t ) can be defined a t first as an initial value problem in all space by setting T'(z. t ) E BloU 9 .) may be illfinite).

a.n)+ l ] ..2. a sharper estimate for a might take a quite different form. t) = T(x) is time independent. the boundaries being a sequence of spheres centered a t the source of the fundamental solution. for any (x. then .) and obtain ( 3 ) Appendix. This takes the form z = exp [.t. I n this argument. t ( 6 ) ) assumed there*. of the form Choice of the optimal t value gives a n inequality tf I ~ ( x / )5 B in may set E. This. we can bound I T'(xl.. ( p 2 / ~ . and i = 1. where n.ive argument based on (33) and (35) obtains stronger results from (13). The moment bound (13) serves to control the rate of dispersal of funda~iientalsolutions. depends only on the dimension n. t ) = T (x) Therefore. However. Nuinerical calculation of extreinal examples might give a better picture. t ) = T'(x. and by use of (36). with the inequality above yields - valid for any positive t. The methods used a b o ~ ecan give inore explicit results. ( G . t ) as a solution of a boundary ~ a l u eproblem.x. t ) T'(x. We can also regard T'(x. 5B(u) for all t 2 0 by the maximum principle. Now T(x. where dpi is the 'nieasure associated with a)i. .)a]. V T ) = T t in free space.we cr= d(xl. t ) I. .I'ARBUOLIC 951 AND ELLIPTIC EQUATIONS. and on a) $. t ) E a%. such as an explicit lower bound for the Holder exponent a. the largest integer not greater than p/2p(tz. a fundamental solution is treated as the solution of an array of parabolic boundary value problems. where the bou~ldaryvalues are just the values T'(x((). With our I3older continuity estimate ( 2 ) for solutions of V . By (33). either in 9. An iterat. we have T (x. The result is as follows : let v = [p/Zp(t. or i n a).

I 5 p . we where k. from (38) and an analogue of ( 2 5 ) ) . IF-r have . c. c. / Sf. by an argument resembling that which gave (25).=minS(x. Then there are constants D. t 1 ) for I Z .t2. %I. A point 6 on the boundary a3 of a region aZ is called regz~larif there are two positive numbers p and e such that any sphere with radius s p and centered a t [ has at least the fraction e of its volume not within 8 .) t2. For any obtain 6 > 0 . the reproductive identity ( 5 ) . and n such that for any n: in with 1. (T. the lower bound +* where is an a priori function determined by c.With (38). tl ) 2 PaPbPc. where / 2-xl 1 and I 2-x2 1 5 p . and n). ~ 1 .where P. and the bound ( 7 ) . we can estimate the speed of convergence t o assigned boundary values of the solution of an elliptic boundary value problem. P b = m i n S ( 2 .Z.. we obtain a pointwise upper bound of the form On the other hand. The inerluality S (2.7:-6 1 5 ap.&(tl+t2)) for 1 E-X. and n .x . + ( t 1 + t 2 ) . With (38). depend on r (and on cl.8 determined by E. (41) and (35). can be used in a iterative argument to strengthen ( 4 0 ) . provided the boundary is "tame " enough./c. we obtain from ( 5 ) and (23) (or alternatively.... and k.. c.. and . P c = j d 2 .

PARABOLIC A N D E L L I P T I C EQUATIONS. obtaining ultimately the Hiilder continuity of the solution for any typical boundary shape. vol." Jerusalem. 61-70. [2] L. 7 (1940). 104 (1931). Dressel. Journal d'analyse Mathgmatique. I xr I)]-l. pp. pp. 340-354./c. With Hiilder continuous boundary values. 11)) Duke Mathematical Journal. we can fairly easily derive a "Harnack iriequality " for parabolic equations : provided 0 T 5 B for t 2 to. " u b e r die Warmeleitungsgleichung m i t nichtkonstanten Koe5zienten in raumlichen Falle I. vol. [4] F. determined by c. 4 (1954).. Nirenberg. pp. 1-58. and n. (2 represents a variable point on the boundary if?). This result is less (laeily obtained than (43). I 3-x' I / + The a priori function H is determined by c.m a x ( I x I. the solution is Holder continuous in the region and at the boundary. Ahlfors. pp. REFERENCES. From the estimates above. [ I ] L. 11. 186-203. "Estimates and uniqueness of solutions of elliptic equations. Parabolic or elliptic problems with Neumann boundary conditions can aparently be handled by a relatively straightforward rederivation of the estimates of this paper in the context of the Neumann boundary. vol. and n. the result takes the form (44) 1 log(T(x')/T ($1) I S H ( r [ r . pp. 509-530. 354-362. 13 (1940). i t follows that the solution of an elliptic boundary value pl. F is an a priori function. . " O n quasi-conformal mapping. For the elliptic case where T is non-negative in a sphere of radius r centered at the origin. vol. From (42).oblem is continuous a t the boundary if continuous values were assigned on tile boundary and all boundary points are regular. G. 9 (1956). Rothe." Commultications on Pure and Applied Mathematics. c." Mathematische Annalelt." (also ibid.. "The fundamental solution of the parabolic equation. vol. [3] E.

NASH. pp." Proceedings of the National Acadamy of Sciences of the United Btates of America. "The dirichlet problem. vol. 435-441. Nash. pp." Proceedings of the National Academy of Sciences 1141 ----. [7] S." Matematideskii Sbornik. 127-146. vol. [ l o ] J. Browder. [5] 0. pp. p. Ladyzhenskaya. 51-92. [ l l ] J . vol. " Sur le mouvement d'un liquide visqueux emplissant l'espace. " Sull'analiticit& delle estremali degli integrali multipli. vol." Annals of Mathematics. 3 ( 1 9 2 4 ) . Leray. 914-917. pp. "The embedding problem for Riemannian manifolds. [6] F." Communications on Pure and Applied Mathematics. 27 ( 6 9 ) ." Bulletin o f the American Mathematical Society. pp. . (195O). 165. vol. B. vol. 193-248.954 J. D." rlcta Nathematica. "On the uniqueness of the Cauchy problem for linear parabolic equations. pp. Morrey. "results on continuation and uniqueness of fluid flow. pp. pp. de Giorgi. Ser. [12] C. 8. 20-63. [9] E. vol.." MatematiEeskil Sbornik. "On fundamental solutions of parabolic systems. 63 ( 1 9 5 6 ) . 43 ( 1 9 5 7 ) . Eidelman. 42 ( 1 9 5 6 ) . 63 ( 1 9 3 4 ) . Jr." A t t i della Accademia Nasionale dei Lincei. 754-758. 20 ( 1 9 5 6 ) . "On the derivation of the equations of hydrodynamics from statistical mechanics. E. 38 ( 8 0 ) ( 1 9 5 6 ) . 1131 J. " Parabolic equations. 8 ( 1 9 5 5 ) . vol. 279-326. A. 17'5-184. " Parabolic systems of differential equations with time-dependent coefficients. of the United States o f America. [8] N. vol." Journal of Mathematics and Physics. Wiener. pp. 60 ( 1 9 5 4 ) . Nash. vol.